13848 lines
432 KiB
Plaintext
13848 lines
432 KiB
Plaintext
//
|
|
// 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_50
|
|
.address_size 64
|
|
|
|
.const .align 16 .b8 kernel_params[4416];
|
|
.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};
|
|
|
|
.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+2232];
|
|
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+2176];
|
|
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+2328];
|
|
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__BB0_6;
|
|
|
|
mov.u64 %rd180, %rd199;
|
|
mov.u32 %r71, %r76;
|
|
mov.u16 %rs8, %rs10;
|
|
mov.u32 %r72, %r77;
|
|
|
|
$L__BB0_2:
|
|
setp.eq.s64 %p2, %rd180, 0;
|
|
@%p2 bra $L__BB0_4;
|
|
|
|
add.s32 %r72, %r72, 2;
|
|
bra.uni $L__BB0_5;
|
|
|
|
$L__BB0_4:
|
|
setp.eq.s16 %p3, %rs8, 0;
|
|
mov.u32 %r75, -1;
|
|
mov.u32 %r72, %r71;
|
|
@%p3 bra $L__BB0_7;
|
|
|
|
$L__BB0_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__BB0_2;
|
|
|
|
$L__BB0_6:
|
|
setp.eq.s16 %p5, %rs8, 0;
|
|
selp.b32 %r75, -1, %r71, %p5;
|
|
|
|
$L__BB0_7:
|
|
ld.const.u64 %rd133, [kernel_params+2248];
|
|
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__BB0_9;
|
|
bra.uni $L__BB0_8;
|
|
|
|
$L__BB0_9:
|
|
ld.const.u64 %rd142, [kernel_params+2272];
|
|
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__BB0_10;
|
|
|
|
$L__BB0_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+2352];
|
|
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__BB0_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__BB0_12;
|
|
bra.uni $L__BB0_11;
|
|
|
|
$L__BB0_12:
|
|
ld.const.u64 %rd152, [kernel_params+2272];
|
|
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__BB0_13;
|
|
|
|
$L__BB0_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+2352];
|
|
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__BB0_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__BB0_18;
|
|
|
|
and.b32 %r56, %r20, 2;
|
|
setp.eq.s32 %p11, %r56, 0;
|
|
@%p11 bra $L__BB0_16;
|
|
bra.uni $L__BB0_15;
|
|
|
|
$L__BB0_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__BB0_17;
|
|
|
|
$L__BB0_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__BB0_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__BB0_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__BB0_20;
|
|
bra.uni $L__BB0_19;
|
|
|
|
$L__BB0_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__BB0_21;
|
|
|
|
$L__BB0_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__BB0_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__BB0_36;
|
|
|
|
setp.eq.s64 %p14, %rd199, 12;
|
|
@%p14 bra $L__BB0_28;
|
|
|
|
mov.u32 %r80, -1;
|
|
|
|
$L__BB0_24:
|
|
setp.eq.s64 %p15, %rd199, 0;
|
|
@%p15 bra $L__BB0_26;
|
|
|
|
add.s32 %r77, %r77, 2;
|
|
bra.uni $L__BB0_27;
|
|
|
|
$L__BB0_26:
|
|
setp.eq.s16 %p16, %rs10, 0;
|
|
mov.u32 %r77, %r76;
|
|
@%p16 bra $L__BB0_29;
|
|
|
|
$L__BB0_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__BB0_24;
|
|
|
|
$L__BB0_28:
|
|
setp.eq.s16 %p18, %rs10, 0;
|
|
selp.b32 %r80, -1, %r76, %p18;
|
|
|
|
$L__BB0_29:
|
|
@%p6 bra $L__BB0_31;
|
|
bra.uni $L__BB0_30;
|
|
|
|
$L__BB0_31:
|
|
ld.const.u64 %rd165, [kernel_params+2240];
|
|
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__BB0_32;
|
|
|
|
$L__BB0_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+2352];
|
|
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__BB0_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__BB0_34;
|
|
bra.uni $L__BB0_33;
|
|
|
|
$L__BB0_34:
|
|
ld.const.u64 %rd175, [kernel_params+2240];
|
|
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__BB0_35;
|
|
|
|
$L__BB0_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+2352];
|
|
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__BB0_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__BB0_36:
|
|
ret;
|
|
|
|
}
|
|
// .globl __miss__kernel_optix_miss
|
|
.visible .entry __miss__kernel_optix_miss()
|
|
{
|
|
.reg .f32 %f<2>;
|
|
.reg .b32 %r<5>;
|
|
|
|
|
|
// begin inline asm
|
|
call (%f1), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
mov.b32 %r2, %f1;
|
|
mov.u32 %r4, 0;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r4, %r2);
|
|
// end inline asm
|
|
mov.u32 %r3, 5;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r3, %r4);
|
|
// end inline asm
|
|
ret;
|
|
|
|
}
|
|
// .globl __anyhit__kernel_optix_local_hit
|
|
.visible .entry __anyhit__kernel_optix_local_hit()
|
|
{
|
|
.reg .pred %p<13>;
|
|
.reg .f32 %f<39>;
|
|
.reg .b32 %r<47>;
|
|
.reg .b64 %rd<36>;
|
|
|
|
|
|
// begin inline asm
|
|
call (%r9), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p1, %r9, 254;
|
|
@%p1 bra $L__BB2_3;
|
|
|
|
// begin inline asm
|
|
call (%r10), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p2, %r10, 255;
|
|
@%p2 bra $L__BB2_3;
|
|
bra.uni $L__BB2_2;
|
|
|
|
$L__BB2_3:
|
|
mov.u32 %r11, 0;
|
|
// begin inline asm
|
|
call (%rd3), _optix_get_transform_list_handle, (%r11);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r12), _optix_get_instance_id_from_handle, (%rd3);
|
|
// end inline asm
|
|
mov.u32 %r14, 4;
|
|
// begin inline asm
|
|
call (%r13), _optix_get_payload, (%r14);
|
|
// end inline asm
|
|
setp.eq.s32 %p3, %r12, %r13;
|
|
@%p3 bra $L__BB2_5;
|
|
bra.uni $L__BB2_4;
|
|
|
|
$L__BB2_5:
|
|
// begin inline asm
|
|
call (%r15), _optix_read_primitive_idx, ();
|
|
// end inline asm
|
|
mov.u32 %r17, 7;
|
|
// begin inline asm
|
|
call (%r16), _optix_get_payload, (%r17);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd5, %r16;
|
|
mov.u32 %r19, 6;
|
|
// begin inline asm
|
|
call (%r18), _optix_get_payload, (%r19);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd6, %r18;
|
|
bfi.b64 %rd7, %rd5, %rd6, 32, 32;
|
|
ld.u32 %r20, [%rd7+36];
|
|
setp.eq.s32 %p4, %r20, %r15;
|
|
@%p4 bra $L__BB2_21;
|
|
bra.uni $L__BB2_6;
|
|
|
|
$L__BB2_21:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB2_22;
|
|
|
|
$L__BB2_4:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB2_22;
|
|
|
|
$L__BB2_2:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB2_22;
|
|
|
|
$L__BB2_6:
|
|
mov.u32 %r22, 5;
|
|
// begin inline asm
|
|
call (%r21), _optix_get_payload, (%r22);
|
|
// end inline asm
|
|
setp.eq.s32 %p5, %r21, 0;
|
|
@%p5 bra $L__BB2_20;
|
|
|
|
// begin inline asm
|
|
call (%r23), _optix_get_payload, (%r11);
|
|
// end inline asm
|
|
mov.u32 %r26, 1;
|
|
// begin inline asm
|
|
call (%r25), _optix_get_payload, (%r26);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd8, %r25;
|
|
cvt.u64.u32 %rd9, %r23;
|
|
bfi.b64 %rd1, %rd8, %rd9, 32, 32;
|
|
mov.u32 %r28, 2;
|
|
// begin inline asm
|
|
call (%r27), _optix_get_payload, (%r28);
|
|
// end inline asm
|
|
mov.u32 %r30, 3;
|
|
// begin inline asm
|
|
call (%r29), _optix_get_payload, (%r30);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd10, %r29;
|
|
cvt.u64.u32 %rd11, %r27;
|
|
bfi.b64 %rd2, %rd10, %rd11, 32, 32;
|
|
setp.eq.s64 %p6, %rd1, 0;
|
|
ld.u32 %r46, [%rd2];
|
|
@%p6 bra $L__BB2_15;
|
|
|
|
min.u32 %r45, %r21, %r46;
|
|
// begin inline asm
|
|
call (%f1), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
|
|
$L__BB2_9:
|
|
add.s32 %r45, %r45, -1;
|
|
setp.gt.s32 %p7, %r45, -1;
|
|
@%p7 bra $L__BB2_13;
|
|
bra.uni $L__BB2_10;
|
|
|
|
$L__BB2_13:
|
|
mul.wide.s32 %rd12, %r45, 24;
|
|
add.s64 %rd13, %rd2, %rd12;
|
|
ld.f32 %f2, [%rd13+4];
|
|
setp.neu.ftz.f32 %p10, %f1, %f2;
|
|
@%p10 bra $L__BB2_9;
|
|
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB2_22;
|
|
|
|
$L__BB2_20:
|
|
mov.u32 %r44, 1;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r22, %r44);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call _optix_terminate_ray, ();
|
|
// end inline asm
|
|
bra.uni $L__BB2_22;
|
|
|
|
$L__BB2_15:
|
|
setp.eq.s32 %p11, %r46, 0;
|
|
@%p11 bra $L__BB2_18;
|
|
|
|
// begin inline asm
|
|
call (%f3), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
ld.f32 %f4, [%rd2+4];
|
|
setp.leu.ftz.f32 %p12, %f3, %f4;
|
|
@%p12 bra $L__BB2_18;
|
|
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB2_22;
|
|
|
|
$L__BB2_10:
|
|
add.s32 %r31, %r46, 1;
|
|
st.u32 [%rd2], %r31;
|
|
setp.le.u32 %p8, %r31, %r21;
|
|
@%p8 bra $L__BB2_19;
|
|
|
|
ld.u32 %r32, [%rd1];
|
|
mad.lo.s32 %r33, %r32, 1103515245, 12345;
|
|
st.u32 [%rd1], %r33;
|
|
ld.u32 %r34, [%rd2];
|
|
rem.u32 %r46, %r33, %r34;
|
|
setp.lt.u32 %p9, %r46, %r21;
|
|
@%p9 bra $L__BB2_19;
|
|
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB2_22;
|
|
|
|
$L__BB2_18:
|
|
st.u32 [%rd2], %r26;
|
|
mov.u32 %r46, %r11;
|
|
|
|
$L__BB2_19:
|
|
// begin inline asm
|
|
call (%f5), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
mul.wide.s32 %rd16, %r46, 24;
|
|
add.s64 %rd17, %rd2, %rd16;
|
|
st.f32 [%rd17+4], %f5;
|
|
st.u32 [%rd17+16], %r15;
|
|
// begin inline asm
|
|
call (%rd14), _optix_get_transform_list_handle, (%r11);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r38), _optix_get_instance_id_from_handle, (%rd14);
|
|
// end inline asm
|
|
st.u32 [%rd17+20], %r38;
|
|
ld.const.u64 %rd18, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd19, %rd18;
|
|
mul.wide.s32 %rd20, %r38, 208;
|
|
add.s64 %rd21, %rd19, %rd20;
|
|
ld.global.u32 %r39, [%rd21+200];
|
|
st.u32 [%rd17+24], %r39;
|
|
// begin inline asm
|
|
call (%f6, %f7), _optix_get_triangle_barycentrics, ();
|
|
// end inline asm
|
|
st.f32 [%rd17+8], %f6;
|
|
st.f32 [%rd17+12], %f7;
|
|
ld.const.u64 %rd22, [kernel_params+2248];
|
|
cvta.to.global.u64 %rd23, %rd22;
|
|
mul.wide.s32 %rd24, %r15, 16;
|
|
add.s64 %rd25, %rd23, %rd24;
|
|
ld.const.u64 %rd26, [kernel_params+2272];
|
|
cvta.to.global.u64 %rd27, %rd26;
|
|
ld.global.u32 %r40, [%rd25+12];
|
|
mul.wide.u32 %rd28, %r40, 12;
|
|
add.s64 %rd29, %rd27, %rd28;
|
|
add.s32 %r41, %r40, 1;
|
|
mul.wide.u32 %rd30, %r41, 12;
|
|
add.s64 %rd31, %rd27, %rd30;
|
|
add.s32 %r42, %r40, 2;
|
|
mul.wide.u32 %rd32, %r42, 12;
|
|
add.s64 %rd33, %rd27, %rd32;
|
|
ld.global.f32 %f8, [%rd31];
|
|
ld.global.f32 %f9, [%rd29];
|
|
sub.ftz.f32 %f10, %f8, %f9;
|
|
ld.global.f32 %f11, [%rd31+4];
|
|
ld.global.f32 %f12, [%rd29+4];
|
|
sub.ftz.f32 %f13, %f11, %f12;
|
|
ld.global.f32 %f14, [%rd31+8];
|
|
ld.global.f32 %f15, [%rd29+8];
|
|
sub.ftz.f32 %f16, %f14, %f15;
|
|
ld.global.f32 %f17, [%rd33];
|
|
sub.ftz.f32 %f18, %f17, %f9;
|
|
ld.global.f32 %f19, [%rd33+4];
|
|
sub.ftz.f32 %f20, %f19, %f12;
|
|
ld.global.f32 %f21, [%rd33+8];
|
|
sub.ftz.f32 %f22, %f21, %f15;
|
|
mul.ftz.f32 %f23, %f13, %f22;
|
|
mul.ftz.f32 %f24, %f16, %f20;
|
|
sub.ftz.f32 %f25, %f23, %f24;
|
|
mul.ftz.f32 %f26, %f16, %f18;
|
|
mul.ftz.f32 %f27, %f10, %f22;
|
|
sub.ftz.f32 %f28, %f26, %f27;
|
|
mul.ftz.f32 %f29, %f10, %f20;
|
|
mul.ftz.f32 %f30, %f13, %f18;
|
|
sub.ftz.f32 %f31, %f29, %f30;
|
|
mul.ftz.f32 %f32, %f28, %f28;
|
|
fma.rn.ftz.f32 %f33, %f25, %f25, %f32;
|
|
fma.rn.ftz.f32 %f34, %f31, %f31, %f33;
|
|
sqrt.approx.ftz.f32 %f35, %f34;
|
|
div.approx.ftz.f32 %f36, %f25, %f35;
|
|
div.approx.ftz.f32 %f37, %f28, %f35;
|
|
div.approx.ftz.f32 %f38, %f31, %f35;
|
|
mul.wide.s32 %rd34, %r46, 12;
|
|
add.s64 %rd35, %rd2, %rd34;
|
|
st.f32 [%rd35+100], %f36;
|
|
st.f32 [%rd35+104], %f37;
|
|
st.f32 [%rd35+108], %f38;
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
|
|
$L__BB2_22:
|
|
ret;
|
|
|
|
}
|
|
// .globl __anyhit__kernel_optix_shadow_all_hit
|
|
.visible .entry __anyhit__kernel_optix_shadow_all_hit()
|
|
{
|
|
.reg .pred %p<34>;
|
|
.reg .b16 %rs<13>;
|
|
.reg .f32 %f<33>;
|
|
.reg .b32 %r<101>;
|
|
.reg .b64 %rd<98>;
|
|
|
|
|
|
// begin inline asm
|
|
call (%r90), _optix_read_primitive_idx, ();
|
|
// end inline asm
|
|
mov.u32 %r33, 0;
|
|
// begin inline asm
|
|
call (%rd15), _optix_get_transform_list_handle, (%r33);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r34), _optix_get_instance_id_from_handle, (%rd15);
|
|
// end inline asm
|
|
mov.u32 %r36, 4;
|
|
// begin inline asm
|
|
call (%r35), _optix_get_payload, (%r36);
|
|
// end inline asm
|
|
ld.const.u64 %rd17, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd1, %rd17;
|
|
cvt.u64.u32 %rd2, %r34;
|
|
mul.wide.u32 %rd18, %r34, 208;
|
|
add.s64 %rd19, %rd1, %rd18;
|
|
ld.global.u32 %r37, [%rd19+196];
|
|
and.b32 %r38, %r37, %r35;
|
|
setp.eq.s32 %p1, %r38, 0;
|
|
@%p1 bra $L__BB3_41;
|
|
|
|
// begin inline asm
|
|
call (%r39), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p2, %r39, 254;
|
|
mov.u16 %rs10, 1;
|
|
@%p2 bra $L__BB3_3;
|
|
|
|
// begin inline asm
|
|
call (%r40), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p3, %r40, 255;
|
|
selp.u16 %rs10, 1, 0, %p3;
|
|
|
|
$L__BB3_3:
|
|
setp.eq.s16 %p4, %rs10, 0;
|
|
@%p4 bra $L__BB3_5;
|
|
|
|
// begin inline asm
|
|
call (%f29, %f28), _optix_get_triangle_barycentrics, ();
|
|
// end inline asm
|
|
mul.lo.s64 %rd20, %rd2, 208;
|
|
add.s64 %rd21, %rd1, %rd20;
|
|
add.s64 %rd95, %rd21, 200;
|
|
bra.uni $L__BB3_9;
|
|
|
|
$L__BB3_41:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB3_42;
|
|
|
|
$L__BB3_5:
|
|
// begin inline asm
|
|
call (%r41), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
and.b32 %r42, %r41, -65;
|
|
setp.eq.s32 %p5, %r42, 8;
|
|
@%p5 bra $L__BB3_8;
|
|
bra.uni $L__BB3_6;
|
|
|
|
$L__BB3_8:
|
|
mul.lo.s64 %rd26, %rd2, 208;
|
|
add.s64 %rd27, %rd1, %rd26;
|
|
add.s64 %rd95, %rd27, 200;
|
|
mov.f32 %f28, 0f00000000;
|
|
mov.f32 %f29, %f28;
|
|
bra.uni $L__BB3_9;
|
|
|
|
$L__BB3_6:
|
|
// begin inline asm
|
|
call (%r43), _optix_get_attribute_0, ();
|
|
// end inline asm
|
|
mov.b32 %f29, %r43;
|
|
// begin inline asm
|
|
call (%r44), _optix_get_attribute_1, ();
|
|
// end inline asm
|
|
mov.b32 %f28, %r44;
|
|
ld.const.u64 %rd22, [kernel_params+2296];
|
|
cvta.to.global.u64 %rd23, %rd22;
|
|
mul.wide.s32 %rd24, %r90, 8;
|
|
add.s64 %rd25, %rd23, %rd24;
|
|
ld.global.u32 %r90, [%rd25];
|
|
add.s64 %rd95, %rd25, 4;
|
|
setp.neu.ftz.f32 %p6, %f29, 0f00000000;
|
|
setp.neu.ftz.f32 %p7, %f29, 0f3F800000;
|
|
and.pred %p8, %p6, %p7;
|
|
@%p8 bra $L__BB3_9;
|
|
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB3_42;
|
|
|
|
$L__BB3_9:
|
|
ld.global.u32 %r5, [%rd95];
|
|
mov.u32 %r46, 7;
|
|
// begin inline asm
|
|
call (%r45), _optix_get_payload, (%r46);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd28, %r45;
|
|
mov.u32 %r48, 6;
|
|
// begin inline asm
|
|
call (%r47), _optix_get_payload, (%r48);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd29, %r47;
|
|
bfi.b64 %rd7, %rd28, %rd29, 32, 32;
|
|
ld.u32 %r49, [%rd7+36];
|
|
setp.ne.s32 %p9, %r49, %r90;
|
|
@%p9 bra $L__BB3_11;
|
|
|
|
ld.u32 %r50, [%rd7+40];
|
|
setp.eq.s32 %p10, %r50, %r34;
|
|
@%p10 bra $L__BB3_40;
|
|
|
|
$L__BB3_11:
|
|
ld.u32 %r51, [%rd7+44];
|
|
setp.ne.s32 %p11, %r51, %r90;
|
|
@%p11 bra $L__BB3_13;
|
|
|
|
ld.u32 %r52, [%rd7+48];
|
|
setp.eq.s32 %p12, %r52, %r34;
|
|
@%p12 bra $L__BB3_40;
|
|
bra.uni $L__BB3_13;
|
|
|
|
$L__BB3_40:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB3_42;
|
|
|
|
$L__BB3_13:
|
|
mov.u32 %r54, 3;
|
|
// begin inline asm
|
|
call (%r53), _optix_get_payload, (%r54);
|
|
// end inline asm
|
|
mov.u32 %r56, 2;
|
|
// begin inline asm
|
|
call (%r55), _optix_get_payload, (%r56);
|
|
// end inline asm
|
|
shr.u32 %r99, %r55, 16;
|
|
and.b32 %r9, %r55, 65535;
|
|
setp.ge.u32 %p13, %r9, %r53;
|
|
@%p13 bra $L__BB3_39;
|
|
|
|
and.b32 %r57, %r5, 1;
|
|
setp.eq.b32 %p14, %r57, 1;
|
|
mov.pred %p15, 0;
|
|
xor.pred %p16, %p14, %p15;
|
|
not.pred %p17, %p16;
|
|
@%p17 bra $L__BB3_16;
|
|
bra.uni $L__BB3_15;
|
|
|
|
$L__BB3_16:
|
|
and.b32 %r58, %r5, 8;
|
|
setp.eq.s32 %p18, %r58, 0;
|
|
@%p18 bra $L__BB3_18;
|
|
bra.uni $L__BB3_17;
|
|
|
|
$L__BB3_18:
|
|
and.b32 %r60, %r5, 6;
|
|
setp.eq.s32 %p19, %r60, 0;
|
|
mov.u32 %r91, 0;
|
|
@%p19 bra $L__BB3_20;
|
|
|
|
ld.const.u64 %rd38, [kernel_params+2280];
|
|
cvta.to.global.u64 %rd39, %rd38;
|
|
mul.wide.s32 %rd40, %r90, 16;
|
|
add.s64 %rd41, %rd39, %rd40;
|
|
ld.global.u32 %r91, [%rd41];
|
|
bra.uni $L__BB3_20;
|
|
|
|
$L__BB3_15:
|
|
ld.const.u64 %rd30, [kernel_params+2232];
|
|
cvta.to.global.u64 %rd31, %rd30;
|
|
mul.wide.s32 %rd32, %r90, 4;
|
|
add.s64 %rd33, %rd31, %rd32;
|
|
ld.global.u32 %r91, [%rd33];
|
|
bra.uni $L__BB3_20;
|
|
|
|
$L__BB3_17:
|
|
ld.const.u64 %rd34, [kernel_params+2320];
|
|
cvta.to.global.u64 %rd35, %rd34;
|
|
mul.wide.s32 %rd36, %r90, 4;
|
|
add.s64 %rd37, %rd35, %rd36;
|
|
ld.global.u32 %r91, [%rd37];
|
|
|
|
$L__BB3_20:
|
|
ld.const.u64 %rd42, [kernel_params+2424];
|
|
cvta.to.global.u64 %rd43, %rd42;
|
|
shl.b32 %r61, %r91, 5;
|
|
cvt.u64.u32 %rd44, %r61;
|
|
and.b64 %rd45, %rd44, 134217696;
|
|
add.s64 %rd46, %rd43, %rd45;
|
|
ld.global.u8 %rs8, [%rd46+18];
|
|
and.b16 %rs9, %rs8, 2;
|
|
setp.eq.s16 %p20, %rs9, 0;
|
|
@%p20 bra $L__BB3_39;
|
|
bra.uni $L__BB3_21;
|
|
|
|
$L__BB3_39:
|
|
mov.u32 %r88, 5;
|
|
mov.u32 %r89, 1;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r88, %r89);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call _optix_terminate_ray, ();
|
|
// end inline asm
|
|
|
|
$L__BB3_42:
|
|
ret;
|
|
|
|
$L__BB3_21:
|
|
and.b32 %r62, %r5, 6;
|
|
setp.eq.s32 %p21, %r62, 0;
|
|
@%p21 bra $L__BB3_33;
|
|
|
|
mov.u32 %r64, 1;
|
|
// begin inline asm
|
|
call (%r63), _optix_get_payload, (%r64);
|
|
// end inline asm
|
|
mul.wide.s32 %rd47, %r34, 208;
|
|
add.s64 %rd48, %rd1, %rd47;
|
|
ld.const.u64 %rd49, [kernel_params+2328];
|
|
cvta.to.global.u64 %rd8, %rd49;
|
|
ld.global.u32 %r93, [%rd48+164];
|
|
mul.wide.u32 %rd50, %r93, 16;
|
|
add.s64 %rd51, %rd8, %rd50;
|
|
ld.global.u32 %r92, [%rd51+8];
|
|
ld.global.u16 %rs11, [%rd51+12];
|
|
ld.global.u64 %rd96, [%rd51];
|
|
setp.eq.s64 %p22, %rd96, 31;
|
|
@%p22 bra $L__BB3_27;
|
|
|
|
$L__BB3_23:
|
|
setp.eq.s64 %p23, %rd96, 0;
|
|
@%p23 bra $L__BB3_25;
|
|
|
|
add.s32 %r93, %r93, 2;
|
|
bra.uni $L__BB3_26;
|
|
|
|
$L__BB3_25:
|
|
setp.eq.s16 %p24, %rs11, 0;
|
|
mov.u32 %r96, -1;
|
|
mov.u32 %r93, %r92;
|
|
@%p24 bra $L__BB3_28;
|
|
|
|
$L__BB3_26:
|
|
mul.wide.u32 %rd52, %r93, 16;
|
|
add.s64 %rd53, %rd8, %rd52;
|
|
ld.global.u32 %r92, [%rd53+8];
|
|
ld.global.u16 %rs11, [%rd53+12];
|
|
ld.global.u64 %rd96, [%rd53];
|
|
setp.ne.s64 %p25, %rd96, 31;
|
|
@%p25 bra $L__BB3_23;
|
|
|
|
$L__BB3_27:
|
|
setp.eq.s16 %p26, %rs11, 0;
|
|
selp.b32 %r96, -1, %r92, %p26;
|
|
|
|
$L__BB3_28:
|
|
setp.eq.s32 %p27, %r96, -1;
|
|
mov.f32 %f30, 0f00000000;
|
|
@%p27 bra $L__BB3_30;
|
|
|
|
ld.const.u64 %rd54, [kernel_params+2280];
|
|
cvta.to.global.u64 %rd55, %rd54;
|
|
mul.wide.s32 %rd56, %r90, 16;
|
|
add.s64 %rd57, %rd55, %rd56;
|
|
shr.s32 %r66, %r5, 7;
|
|
ld.global.u32 %r67, [%rd57+4];
|
|
add.s32 %r68, %r67, %r66;
|
|
ld.const.u64 %rd58, [kernel_params+2336];
|
|
cvta.to.global.u64 %rd59, %rd58;
|
|
add.s32 %r69, %r68, %r96;
|
|
mul.wide.s32 %rd60, %r69, 4;
|
|
add.s64 %rd61, %rd59, %rd60;
|
|
mov.f32 %f18, 0f3F800000;
|
|
sub.ftz.f32 %f19, %f18, %f29;
|
|
ld.global.f32 %f20, [%rd61];
|
|
ld.global.f32 %f21, [%rd61+4];
|
|
mul.ftz.f32 %f22, %f29, %f21;
|
|
fma.rn.ftz.f32 %f30, %f19, %f20, %f22;
|
|
|
|
$L__BB3_30:
|
|
mov.b32 %f23, %r63;
|
|
mul.ftz.f32 %f24, %f30, %f23;
|
|
mov.b32 %r71, %f24;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r64, %r71);
|
|
// end inline asm
|
|
and.b32 %r74, %r55, -65536;
|
|
add.s32 %r75, %r9, 1;
|
|
or.b32 %r73, %r74, %r75;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r56, %r73);
|
|
// end inline asm
|
|
setp.lt.ftz.f32 %p28, %f24, 0f3A83126F;
|
|
@%p28 bra $L__BB3_32;
|
|
bra.uni $L__BB3_31;
|
|
|
|
$L__BB3_32:
|
|
mov.u32 %r76, 5;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r76, %r64);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call _optix_terminate_ray, ();
|
|
// end inline asm
|
|
bra.uni $L__BB3_42;
|
|
|
|
$L__BB3_33:
|
|
add.s32 %r82, %r9, 1;
|
|
add.s32 %r83, %r55, 65536;
|
|
and.b32 %r84, %r83, -65536;
|
|
or.b32 %r79, %r84, %r82;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r56, %r79);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r80), _optix_get_payload, (%r33);
|
|
// end inline asm
|
|
min.u32 %r25, %r53, 4;
|
|
setp.lt.u32 %p29, %r99, %r25;
|
|
cvt.s64.s32 %rd12, %r80;
|
|
@%p29 bra $L__BB3_38;
|
|
|
|
ld.const.u64 %rd62, [kernel_params+3568];
|
|
cvta.to.global.u64 %rd63, %rd62;
|
|
shl.b64 %rd64, %rd12, 2;
|
|
add.s64 %rd65, %rd63, %rd64;
|
|
ld.global.f32 %f32, [%rd65];
|
|
setp.lt.u32 %p30, %r25, 2;
|
|
mov.u32 %r99, %r33;
|
|
@%p30 bra $L__BB3_37;
|
|
|
|
mov.u32 %r99, 0;
|
|
mov.u32 %r97, 1;
|
|
mov.u64 %rd97, kernel_params;
|
|
|
|
$L__BB3_36:
|
|
ld.const.u64 %rd67, [%rd97+3616];
|
|
cvta.to.global.u64 %rd68, %rd67;
|
|
add.s64 %rd70, %rd68, %rd64;
|
|
ld.global.f32 %f25, [%rd70];
|
|
setp.gt.ftz.f32 %p31, %f25, %f32;
|
|
selp.f32 %f32, %f25, %f32, %p31;
|
|
selp.b32 %r99, %r97, %r99, %p31;
|
|
add.s64 %rd97, %rd97, 48;
|
|
add.s32 %r97, %r97, 1;
|
|
setp.lt.u32 %p32, %r97, %r25;
|
|
@%p32 bra $L__BB3_36;
|
|
|
|
$L__BB3_37:
|
|
// begin inline asm
|
|
call (%f26), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
setp.ge.ftz.f32 %p33, %f26, %f32;
|
|
@%p33 bra $L__BB3_42;
|
|
|
|
$L__BB3_38:
|
|
mul.wide.u32 %rd71, %r99, 48;
|
|
mov.u64 %rd72, kernel_params;
|
|
add.s64 %rd73, %rd72, %rd71;
|
|
ld.const.u64 %rd74, [%rd73+3576];
|
|
cvta.to.global.u64 %rd75, %rd74;
|
|
shl.b64 %rd76, %rd12, 2;
|
|
add.s64 %rd77, %rd75, %rd76;
|
|
st.global.f32 [%rd77], %f29;
|
|
ld.const.u64 %rd78, [%rd73+3584];
|
|
cvta.to.global.u64 %rd79, %rd78;
|
|
add.s64 %rd80, %rd79, %rd76;
|
|
st.global.f32 [%rd80], %f28;
|
|
// begin inline asm
|
|
call (%f27), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
ld.const.u64 %rd81, [%rd73+3568];
|
|
cvta.to.global.u64 %rd82, %rd81;
|
|
add.s64 %rd83, %rd82, %rd76;
|
|
st.global.f32 [%rd83], %f27;
|
|
ld.const.u64 %rd84, [%rd73+3592];
|
|
cvta.to.global.u64 %rd85, %rd84;
|
|
add.s64 %rd86, %rd85, %rd76;
|
|
st.global.u32 [%rd86], %r90;
|
|
ld.const.v2.u64 {%rd87, %rd88}, [%rd73+3600];
|
|
cvta.to.global.u64 %rd91, %rd87;
|
|
add.s64 %rd92, %rd91, %rd76;
|
|
st.global.u32 [%rd92], %r34;
|
|
cvta.to.global.u64 %rd93, %rd88;
|
|
add.s64 %rd94, %rd93, %rd76;
|
|
st.global.u32 [%rd94], %r5;
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB3_42;
|
|
|
|
$L__BB3_31:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB3_42;
|
|
|
|
}
|
|
// .globl __anyhit__kernel_optix_volume_test
|
|
.visible .entry __anyhit__kernel_optix_volume_test()
|
|
{
|
|
.reg .pred %p<7>;
|
|
.reg .b16 %rs<3>;
|
|
.reg .b32 %r<17>;
|
|
.reg .b64 %rd<15>;
|
|
|
|
|
|
// begin inline asm
|
|
call (%r1), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p1, %r1, 254;
|
|
@%p1 bra $L__BB4_3;
|
|
|
|
// begin inline asm
|
|
call (%r2), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p2, %r2, 255;
|
|
@%p2 bra $L__BB4_3;
|
|
bra.uni $L__BB4_2;
|
|
|
|
$L__BB4_3:
|
|
mov.u32 %r3, 0;
|
|
// begin inline asm
|
|
call (%rd3), _optix_get_transform_list_handle, (%r3);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r4), _optix_get_instance_id_from_handle, (%rd3);
|
|
// end inline asm
|
|
mov.u32 %r6, 4;
|
|
// begin inline asm
|
|
call (%r5), _optix_get_payload, (%r6);
|
|
// end inline asm
|
|
ld.const.u64 %rd5, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd6, %rd5;
|
|
cvt.u64.u32 %rd1, %r4;
|
|
mul.wide.u32 %rd7, %r4, 208;
|
|
add.s64 %rd8, %rd6, %rd7;
|
|
ld.global.u32 %r7, [%rd8+196];
|
|
and.b32 %r8, %r7, %r5;
|
|
setp.eq.s32 %p3, %r8, 0;
|
|
@%p3 bra $L__BB4_9;
|
|
|
|
ld.const.u64 %rd9, [kernel_params+2200];
|
|
cvta.to.global.u64 %rd10, %rd9;
|
|
shl.b64 %rd11, %rd1, 2;
|
|
add.s64 %rd12, %rd10, %rd11;
|
|
ld.global.u8 %rs1, [%rd12];
|
|
and.b16 %rs2, %rs1, 16;
|
|
setp.eq.s16 %p4, %rs2, 0;
|
|
@%p4 bra $L__BB4_8;
|
|
bra.uni $L__BB4_5;
|
|
|
|
$L__BB4_8:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB4_10;
|
|
|
|
$L__BB4_9:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB4_10;
|
|
|
|
$L__BB4_2:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB4_10;
|
|
|
|
$L__BB4_5:
|
|
// begin inline asm
|
|
call (%r9), _optix_read_primitive_idx, ();
|
|
// end inline asm
|
|
mov.u32 %r11, 7;
|
|
// begin inline asm
|
|
call (%r10), _optix_get_payload, (%r11);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd13, %r10;
|
|
mov.u32 %r13, 6;
|
|
// begin inline asm
|
|
call (%r12), _optix_get_payload, (%r13);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd14, %r12;
|
|
bfi.b64 %rd2, %rd13, %rd14, 32, 32;
|
|
ld.u32 %r14, [%rd2+36];
|
|
setp.ne.s32 %p5, %r14, %r9;
|
|
@%p5 bra $L__BB4_10;
|
|
|
|
ld.u32 %r15, [%rd2+40];
|
|
cvt.u32.u64 %r16, %rd1;
|
|
setp.ne.s32 %p6, %r15, %r16;
|
|
@%p6 bra $L__BB4_10;
|
|
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
|
|
$L__BB4_10:
|
|
ret;
|
|
|
|
}
|
|
// .globl __anyhit__kernel_optix_visibility_test
|
|
.visible .entry __anyhit__kernel_optix_visibility_test()
|
|
{
|
|
.reg .pred %p<16>;
|
|
.reg .f32 %f<2>;
|
|
.reg .b32 %r<32>;
|
|
.reg .b64 %rd<14>;
|
|
|
|
|
|
// begin inline asm
|
|
call (%r7), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r8), _optix_get_primitive_type_from_hit_kind, (%r7);
|
|
// end inline asm
|
|
setp.ne.s32 %p1, %r8, 9474;
|
|
@%p1 bra $L__BB5_2;
|
|
|
|
// begin inline asm
|
|
call (%r10), _optix_get_attribute_0, ();
|
|
// end inline asm
|
|
mov.b32 %f1, %r10;
|
|
setp.eq.ftz.f32 %p2, %f1, 0f00000000;
|
|
setp.eq.ftz.f32 %p3, %f1, 0f3F800000;
|
|
or.pred %p4, %p2, %p3;
|
|
@%p4 bra $L__BB5_18;
|
|
bra.uni $L__BB5_2;
|
|
|
|
$L__BB5_18:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB5_19;
|
|
|
|
$L__BB5_2:
|
|
mov.u32 %r11, 0;
|
|
// begin inline asm
|
|
call (%rd2), _optix_get_transform_list_handle, (%r11);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r12), _optix_get_instance_id_from_handle, (%rd2);
|
|
// end inline asm
|
|
mov.u32 %r14, 4;
|
|
// begin inline asm
|
|
call (%r13), _optix_get_payload, (%r14);
|
|
// end inline asm
|
|
ld.const.u64 %rd4, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd5, %rd4;
|
|
mul.wide.u32 %rd6, %r12, 208;
|
|
add.s64 %rd7, %rd5, %rd6;
|
|
ld.global.u32 %r15, [%rd7+196];
|
|
and.b32 %r16, %r15, %r13;
|
|
setp.eq.s32 %p5, %r16, 0;
|
|
@%p5 bra $L__BB5_17;
|
|
|
|
// begin inline asm
|
|
call (%r31), _optix_read_primitive_idx, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r18), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p6, %r18, 254;
|
|
@%p6 bra $L__BB5_7;
|
|
|
|
// begin inline asm
|
|
call (%r19), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p7, %r19, 255;
|
|
@%p7 bra $L__BB5_7;
|
|
|
|
// begin inline asm
|
|
call (%r20), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
and.b32 %r21, %r20, -65;
|
|
setp.eq.s32 %p8, %r21, 8;
|
|
@%p8 bra $L__BB5_7;
|
|
|
|
ld.const.u64 %rd8, [kernel_params+2296];
|
|
cvta.to.global.u64 %rd9, %rd8;
|
|
mul.wide.s32 %rd10, %r31, 8;
|
|
add.s64 %rd11, %rd9, %rd10;
|
|
ld.global.u32 %r31, [%rd11];
|
|
|
|
$L__BB5_7:
|
|
mov.u32 %r23, 7;
|
|
// begin inline asm
|
|
call (%r22), _optix_get_payload, (%r23);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd12, %r22;
|
|
mov.u32 %r25, 6;
|
|
// begin inline asm
|
|
call (%r24), _optix_get_payload, (%r25);
|
|
// end inline asm
|
|
cvt.u64.u32 %rd13, %r24;
|
|
bfi.b64 %rd1, %rd12, %rd13, 32, 32;
|
|
and.b32 %r26, %r13, 256;
|
|
setp.eq.s32 %p9, %r26, 0;
|
|
ld.u32 %r6, [%rd1+36];
|
|
@%p9 bra $L__BB5_14;
|
|
bra.uni $L__BB5_8;
|
|
|
|
$L__BB5_14:
|
|
setp.ne.s32 %p14, %r6, %r31;
|
|
@%p14 bra $L__BB5_19;
|
|
|
|
ld.u32 %r30, [%rd1+40];
|
|
setp.ne.s32 %p15, %r30, %r12;
|
|
@%p15 bra $L__BB5_19;
|
|
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB5_19;
|
|
|
|
$L__BB5_17:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB5_19;
|
|
|
|
$L__BB5_8:
|
|
setp.ne.s32 %p10, %r6, %r31;
|
|
@%p10 bra $L__BB5_10;
|
|
|
|
ld.u32 %r27, [%rd1+40];
|
|
setp.eq.s32 %p11, %r27, %r12;
|
|
@%p11 bra $L__BB5_13;
|
|
|
|
$L__BB5_10:
|
|
ld.u32 %r28, [%rd1+44];
|
|
setp.ne.s32 %p12, %r28, %r31;
|
|
@%p12 bra $L__BB5_12;
|
|
|
|
ld.u32 %r29, [%rd1+48];
|
|
setp.eq.s32 %p13, %r29, %r12;
|
|
@%p13 bra $L__BB5_13;
|
|
bra.uni $L__BB5_12;
|
|
|
|
$L__BB5_13:
|
|
// begin inline asm
|
|
call _optix_ignore_intersection, ();
|
|
// end inline asm
|
|
bra.uni $L__BB5_19;
|
|
|
|
$L__BB5_12:
|
|
// begin inline asm
|
|
call _optix_terminate_ray, ();
|
|
// end inline asm
|
|
|
|
$L__BB5_19:
|
|
ret;
|
|
|
|
}
|
|
// .globl __closesthit__kernel_optix_hit
|
|
.visible .entry __closesthit__kernel_optix_hit()
|
|
{
|
|
.reg .pred %p<4>;
|
|
.reg .f32 %f<4>;
|
|
.reg .b32 %r<40>;
|
|
.reg .b64 %rd<15>;
|
|
|
|
|
|
mov.u32 %r6, 0;
|
|
// begin inline asm
|
|
call (%rd1), _optix_get_transform_list_handle, (%r6);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r4), _optix_get_instance_id_from_handle, (%rd1);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r5), _optix_read_primitive_idx, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f1), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
mov.b32 %r7, %f1;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r6, %r7);
|
|
// end inline asm
|
|
mov.u32 %r8, 4;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r8, %r4);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r10), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p1, %r10, 254;
|
|
@%p1 bra $L__BB6_5;
|
|
|
|
// begin inline asm
|
|
call (%r11), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
setp.eq.s32 %p2, %r11, 255;
|
|
@%p2 bra $L__BB6_5;
|
|
bra.uni $L__BB6_2;
|
|
|
|
$L__BB6_5:
|
|
// begin inline asm
|
|
call (%f2, %f3), _optix_get_triangle_barycentrics, ();
|
|
// end inline asm
|
|
mov.b32 %r33, %f2;
|
|
mov.u32 %r32, 1;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r32, %r33);
|
|
// end inline asm
|
|
mov.b32 %r35, %f3;
|
|
mov.u32 %r34, 2;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r34, %r35);
|
|
// end inline asm
|
|
mov.u32 %r36, 3;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r36, %r5);
|
|
// end inline asm
|
|
ld.const.u64 %rd11, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd12, %rd11;
|
|
mul.wide.s32 %rd13, %r4, 208;
|
|
add.s64 %rd14, %rd12, %rd13;
|
|
ld.global.u32 %r39, [%rd14+200];
|
|
mov.u32 %r38, 5;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r38, %r39);
|
|
// end inline asm
|
|
|
|
$L__BB6_6:
|
|
ret;
|
|
|
|
$L__BB6_2:
|
|
// begin inline asm
|
|
call (%r12), _optix_get_hit_kind, ();
|
|
// end inline asm
|
|
and.b32 %r13, %r12, -65;
|
|
setp.eq.s32 %p3, %r13, 8;
|
|
@%p3 bra $L__BB6_4;
|
|
bra.uni $L__BB6_3;
|
|
|
|
$L__BB6_4:
|
|
mov.u32 %r24, 1;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r24, %r6);
|
|
// end inline asm
|
|
mov.u32 %r26, 2;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r26, %r6);
|
|
// end inline asm
|
|
mov.u32 %r28, 3;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r28, %r5);
|
|
// end inline asm
|
|
ld.const.u64 %rd7, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd8, %rd7;
|
|
mul.wide.s32 %rd9, %r4, 208;
|
|
add.s64 %rd10, %rd8, %rd9;
|
|
ld.global.u32 %r31, [%rd10+200];
|
|
mov.u32 %r30, 5;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r30, %r31);
|
|
// end inline asm
|
|
bra.uni $L__BB6_6;
|
|
|
|
$L__BB6_3:
|
|
ld.const.u64 %rd3, [kernel_params+2296];
|
|
cvta.to.global.u64 %rd4, %rd3;
|
|
mul.wide.s32 %rd5, %r5, 8;
|
|
add.s64 %rd6, %rd4, %rd5;
|
|
ld.global.u32 %r21, [%rd6];
|
|
ld.global.u32 %r23, [%rd6+4];
|
|
// begin inline asm
|
|
call (%r14), _optix_get_attribute_0, ();
|
|
// end inline asm
|
|
mov.u32 %r15, 1;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r15, %r14);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r17), _optix_get_attribute_1, ();
|
|
// end inline asm
|
|
mov.u32 %r18, 2;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r18, %r17);
|
|
// end inline asm
|
|
mov.u32 %r20, 3;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r20, %r21);
|
|
// end inline asm
|
|
mov.u32 %r22, 5;
|
|
// begin inline asm
|
|
call _optix_set_payload, (%r22, %r23);
|
|
// end inline asm
|
|
bra.uni $L__BB6_6;
|
|
|
|
}
|
|
// .globl __intersection__curve_ribbon
|
|
.visible .entry __intersection__curve_ribbon()
|
|
{
|
|
.reg .pred %p<38>;
|
|
.reg .b16 %rs<7>;
|
|
.reg .f32 %f<789>;
|
|
.reg .b32 %r<75>;
|
|
.reg .b64 %rd<66>;
|
|
|
|
|
|
ld.const.u64 %rd7, [kernel_params+2296];
|
|
cvta.to.global.u64 %rd8, %rd7;
|
|
// begin inline asm
|
|
call (%r27), _optix_read_primitive_idx, ();
|
|
// end inline asm
|
|
mul.wide.u32 %rd9, %r27, 8;
|
|
add.s64 %rd1, %rd8, %rd9;
|
|
ld.global.u32 %r1, [%rd1+4];
|
|
and.b32 %r28, %r1, 4;
|
|
setp.eq.s32 %p3, %r28, 0;
|
|
@%p3 bra $L__BB7_33;
|
|
|
|
ld.global.u32 %r2, [%rd1];
|
|
mov.u32 %r29, 0;
|
|
// begin inline asm
|
|
call (%rd10), _optix_get_transform_list_handle, (%r29);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r30), _optix_get_instance_id_from_handle, (%rd10);
|
|
// end inline asm
|
|
mov.u32 %r32, 4;
|
|
// begin inline asm
|
|
call (%r31), _optix_get_payload, (%r32);
|
|
// end inline asm
|
|
ld.const.u64 %rd12, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd13, %rd12;
|
|
mul.wide.s32 %rd14, %r30, 208;
|
|
add.s64 %rd15, %rd13, %rd14;
|
|
add.s64 %rd2, %rd15, 196;
|
|
ld.global.u32 %r33, [%rd15+196];
|
|
and.b32 %r34, %r33, %r31;
|
|
setp.eq.s32 %p4, %r34, 0;
|
|
@%p4 bra $L__BB7_33;
|
|
|
|
// begin inline asm
|
|
call (%f241), _optix_get_object_ray_origin_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f242), _optix_get_object_ray_origin_y, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f243), _optix_get_object_ray_origin_z, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f244), _optix_get_object_ray_direction_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f245), _optix_get_object_ray_direction_y, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f246), _optix_get_object_ray_direction_z, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f247), _optix_get_ray_tmin, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f248), _optix_get_ray_time, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f788), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
and.b32 %r35, %r1, 64;
|
|
setp.eq.s32 %p5, %r35, 0;
|
|
ld.const.u64 %rd16, [kernel_params+2280];
|
|
cvta.to.global.u64 %rd17, %rd16;
|
|
mul.wide.s32 %rd18, %r2, 16;
|
|
add.s64 %rd19, %rd17, %rd18;
|
|
shr.s32 %r36, %r1, 7;
|
|
ld.global.u32 %r37, [%rd19+4];
|
|
add.s32 %r3, %r37, %r36;
|
|
add.s32 %r38, %r3, -1;
|
|
max.s32 %r4, %r38, %r37;
|
|
add.s32 %r39, %r3, 2;
|
|
ld.global.u32 %r40, [%rd19+8];
|
|
add.s32 %r41, %r37, %r40;
|
|
add.s32 %r42, %r41, -1;
|
|
min.s32 %r5, %r39, %r42;
|
|
@%p5 bra $L__BB7_16;
|
|
bra.uni $L__BB7_3;
|
|
|
|
$L__BB7_16:
|
|
ld.const.u64 %rd57, [kernel_params+2288];
|
|
cvta.to.global.u64 %rd58, %rd57;
|
|
mul.wide.s32 %rd59, %r4, 16;
|
|
add.s64 %rd60, %rd58, %rd59;
|
|
ld.global.v4.f32 {%f767, %f766, %f765, %f764}, [%rd60];
|
|
mul.wide.s32 %rd61, %r3, 16;
|
|
add.s64 %rd62, %rd58, %rd61;
|
|
ld.global.v4.f32 {%f763, %f762, %f761, %f760}, [%rd62];
|
|
ld.global.v4.f32 {%f759, %f758, %f757, %f756}, [%rd62+16];
|
|
mul.wide.s32 %rd63, %r5, 16;
|
|
add.s64 %rd64, %rd58, %rd63;
|
|
ld.global.v4.f32 {%f755, %f754, %f753, %f752}, [%rd64];
|
|
bra.uni $L__BB7_17;
|
|
|
|
$L__BB7_3:
|
|
ld.global.u32 %r6, [%rd2+-48];
|
|
ld.global.u32 %r7, [%rd2+-44];
|
|
shl.b32 %r43, %r7, 1;
|
|
cvt.rn.f32.s32 %f250, %r43;
|
|
mul.ftz.f32 %f10, %f248, %f250;
|
|
cvt.rzi.ftz.s32.f32 %r44, %f10;
|
|
add.s32 %r45, %r43, -1;
|
|
min.s32 %r8, %r44, %r45;
|
|
ld.const.u64 %rd20, [kernel_params+2328];
|
|
cvta.to.global.u64 %rd3, %rd20;
|
|
ld.global.u32 %r68, [%rd2+-32];
|
|
mul.wide.u32 %rd21, %r68, 16;
|
|
add.s64 %rd22, %rd3, %rd21;
|
|
ld.global.u16 %rs5, [%rd22+12];
|
|
ld.global.u32 %r67, [%rd22+8];
|
|
ld.global.u64 %rd65, [%rd22];
|
|
setp.eq.s64 %p6, %rd65, 11;
|
|
@%p6 bra $L__BB7_8;
|
|
|
|
$L__BB7_4:
|
|
setp.eq.s64 %p7, %rd65, 0;
|
|
@%p7 bra $L__BB7_6;
|
|
|
|
add.s32 %r68, %r68, 2;
|
|
bra.uni $L__BB7_7;
|
|
|
|
$L__BB7_6:
|
|
setp.eq.s16 %p8, %rs5, 0;
|
|
mov.u32 %r71, -1;
|
|
mov.u32 %r68, %r67;
|
|
@%p8 bra $L__BB7_9;
|
|
|
|
$L__BB7_7:
|
|
mul.wide.u32 %rd23, %r68, 16;
|
|
add.s64 %rd24, %rd3, %rd23;
|
|
ld.global.u16 %rs5, [%rd24+12];
|
|
ld.global.u32 %r67, [%rd24+8];
|
|
ld.global.u64 %rd65, [%rd24];
|
|
setp.ne.s64 %p9, %rd65, 11;
|
|
@%p9 bra $L__BB7_4;
|
|
|
|
$L__BB7_8:
|
|
setp.eq.s16 %p10, %rs5, 0;
|
|
selp.b32 %r71, -1, %r67, %p10;
|
|
|
|
$L__BB7_9:
|
|
setp.eq.s32 %p11, %r8, %r7;
|
|
@%p11 bra $L__BB7_11;
|
|
bra.uni $L__BB7_10;
|
|
|
|
$L__BB7_11:
|
|
ld.const.u64 %rd33, [kernel_params+2288];
|
|
cvta.to.global.u64 %rd34, %rd33;
|
|
mul.wide.s32 %rd35, %r4, 16;
|
|
add.s64 %rd36, %rd34, %rd35;
|
|
ld.global.v4.f32 {%f735, %f734, %f733, %f732}, [%rd36];
|
|
mul.wide.s32 %rd37, %r3, 16;
|
|
add.s64 %rd38, %rd34, %rd37;
|
|
ld.global.v4.f32 {%f731, %f730, %f729, %f728}, [%rd38];
|
|
ld.global.v4.f32 {%f727, %f726, %f725, %f724}, [%rd38+16];
|
|
mul.wide.s32 %rd39, %r5, 16;
|
|
add.s64 %rd40, %rd34, %rd39;
|
|
ld.global.v4.f32 {%f723, %f722, %f721, %f720}, [%rd40];
|
|
bra.uni $L__BB7_12;
|
|
|
|
$L__BB7_10:
|
|
setp.gt.s32 %p12, %r8, %r7;
|
|
selp.b32 %r47, -1, 0, %p12;
|
|
add.s32 %r48, %r8, %r47;
|
|
mad.lo.s32 %r49, %r48, %r6, %r71;
|
|
ld.const.u64 %rd25, [kernel_params+2360];
|
|
cvta.to.global.u64 %rd26, %rd25;
|
|
add.s32 %r50, %r49, %r4;
|
|
mul.wide.s32 %rd27, %r50, 16;
|
|
add.s64 %rd28, %rd26, %rd27;
|
|
ld.global.v4.f32 {%f735, %f734, %f733, %f732}, [%rd28];
|
|
add.s32 %r51, %r49, %r3;
|
|
mul.wide.s32 %rd29, %r51, 16;
|
|
add.s64 %rd30, %rd26, %rd29;
|
|
ld.global.v4.f32 {%f731, %f730, %f729, %f728}, [%rd30];
|
|
ld.global.v4.f32 {%f727, %f726, %f725, %f724}, [%rd30+16];
|
|
add.s32 %r52, %r49, %r5;
|
|
mul.wide.s32 %rd31, %r52, 16;
|
|
add.s64 %rd32, %rd26, %rd31;
|
|
ld.global.v4.f32 {%f723, %f722, %f721, %f720}, [%rd32];
|
|
|
|
$L__BB7_12:
|
|
add.s32 %r19, %r8, 1;
|
|
setp.eq.s32 %p13, %r19, %r7;
|
|
@%p13 bra $L__BB7_14;
|
|
bra.uni $L__BB7_13;
|
|
|
|
$L__BB7_14:
|
|
ld.const.u64 %rd49, [kernel_params+2288];
|
|
cvta.to.global.u64 %rd50, %rd49;
|
|
mul.wide.s32 %rd51, %r4, 16;
|
|
add.s64 %rd52, %rd50, %rd51;
|
|
ld.global.v4.f32 {%f751, %f750, %f749, %f748}, [%rd52];
|
|
mul.wide.s32 %rd53, %r3, 16;
|
|
add.s64 %rd54, %rd50, %rd53;
|
|
ld.global.v4.f32 {%f747, %f746, %f745, %f744}, [%rd54];
|
|
ld.global.v4.f32 {%f743, %f742, %f741, %f740}, [%rd54+16];
|
|
mul.wide.s32 %rd55, %r5, 16;
|
|
add.s64 %rd56, %rd50, %rd55;
|
|
ld.global.v4.f32 {%f739, %f738, %f737, %f736}, [%rd56];
|
|
bra.uni $L__BB7_15;
|
|
|
|
$L__BB7_13:
|
|
setp.ge.s32 %p14, %r8, %r7;
|
|
selp.b32 %r53, -1, 0, %p14;
|
|
add.s32 %r54, %r19, %r53;
|
|
mad.lo.s32 %r55, %r54, %r6, %r71;
|
|
ld.const.u64 %rd41, [kernel_params+2360];
|
|
cvta.to.global.u64 %rd42, %rd41;
|
|
add.s32 %r56, %r55, %r4;
|
|
mul.wide.s32 %rd43, %r56, 16;
|
|
add.s64 %rd44, %rd42, %rd43;
|
|
ld.global.v4.f32 {%f751, %f750, %f749, %f748}, [%rd44];
|
|
add.s32 %r57, %r55, %r3;
|
|
mul.wide.s32 %rd45, %r57, 16;
|
|
add.s64 %rd46, %rd42, %rd45;
|
|
ld.global.v4.f32 {%f747, %f746, %f745, %f744}, [%rd46];
|
|
ld.global.v4.f32 {%f743, %f742, %f741, %f740}, [%rd46+16];
|
|
add.s32 %r58, %r55, %r5;
|
|
mul.wide.s32 %rd47, %r58, 16;
|
|
add.s64 %rd48, %rd42, %rd47;
|
|
ld.global.v4.f32 {%f739, %f738, %f737, %f736}, [%rd48];
|
|
|
|
$L__BB7_15:
|
|
cvt.rn.f32.s32 %f315, %r8;
|
|
sub.ftz.f32 %f316, %f10, %f315;
|
|
mov.f32 %f317, 0f3F800000;
|
|
sub.ftz.f32 %f318, %f317, %f316;
|
|
mul.ftz.f32 %f319, %f316, %f751;
|
|
fma.rn.ftz.f32 %f767, %f318, %f735, %f319;
|
|
mul.ftz.f32 %f320, %f316, %f750;
|
|
fma.rn.ftz.f32 %f766, %f318, %f734, %f320;
|
|
mul.ftz.f32 %f321, %f316, %f749;
|
|
fma.rn.ftz.f32 %f765, %f318, %f733, %f321;
|
|
mul.ftz.f32 %f322, %f316, %f748;
|
|
fma.rn.ftz.f32 %f764, %f318, %f732, %f322;
|
|
mul.ftz.f32 %f323, %f316, %f747;
|
|
fma.rn.ftz.f32 %f763, %f318, %f731, %f323;
|
|
mul.ftz.f32 %f324, %f316, %f746;
|
|
fma.rn.ftz.f32 %f762, %f318, %f730, %f324;
|
|
mul.ftz.f32 %f325, %f316, %f745;
|
|
fma.rn.ftz.f32 %f761, %f318, %f729, %f325;
|
|
mul.ftz.f32 %f326, %f316, %f744;
|
|
fma.rn.ftz.f32 %f760, %f318, %f728, %f326;
|
|
mul.ftz.f32 %f327, %f316, %f743;
|
|
fma.rn.ftz.f32 %f759, %f318, %f727, %f327;
|
|
mul.ftz.f32 %f328, %f316, %f742;
|
|
fma.rn.ftz.f32 %f758, %f318, %f726, %f328;
|
|
mul.ftz.f32 %f329, %f316, %f741;
|
|
fma.rn.ftz.f32 %f757, %f318, %f725, %f329;
|
|
mul.ftz.f32 %f330, %f316, %f740;
|
|
fma.rn.ftz.f32 %f756, %f318, %f724, %f330;
|
|
mul.ftz.f32 %f331, %f316, %f739;
|
|
fma.rn.ftz.f32 %f755, %f318, %f723, %f331;
|
|
mul.ftz.f32 %f332, %f316, %f738;
|
|
fma.rn.ftz.f32 %f754, %f318, %f722, %f332;
|
|
mul.ftz.f32 %f333, %f316, %f737;
|
|
fma.rn.ftz.f32 %f753, %f318, %f721, %f333;
|
|
mul.ftz.f32 %f334, %f316, %f736;
|
|
fma.rn.ftz.f32 %f752, %f318, %f720, %f334;
|
|
|
|
$L__BB7_17:
|
|
mul.ftz.f32 %f351, %f245, %f245;
|
|
fma.rn.ftz.f32 %f352, %f244, %f244, %f351;
|
|
fma.rn.ftz.f32 %f353, %f246, %f246, %f352;
|
|
rsqrt.approx.ftz.f32 %f155, %f353;
|
|
mov.f32 %f354, 0f3F800000;
|
|
mul.ftz.f32 %f355, %f244, %f155;
|
|
mul.ftz.f32 %f356, %f245, %f155;
|
|
mul.ftz.f32 %f357, %f246, %f155;
|
|
neg.ftz.f32 %f358, %f356;
|
|
neg.ftz.f32 %f359, %f357;
|
|
fma.rn.ftz.f32 %f360, %f357, %f357, 0f00000000;
|
|
fma.rn.ftz.f32 %f361, %f356, %f356, %f360;
|
|
fma.rn.ftz.f32 %f362, %f355, %f355, %f360;
|
|
setp.gt.ftz.f32 %p15, %f361, %f362;
|
|
selp.f32 %f363, 0f00000000, %f359, %p15;
|
|
selp.f32 %f364, %f357, 0f00000000, %p15;
|
|
selp.f32 %f365, %f358, %f355, %p15;
|
|
mul.ftz.f32 %f366, %f364, %f364;
|
|
fma.rn.ftz.f32 %f367, %f363, %f363, %f366;
|
|
fma.rn.ftz.f32 %f368, %f365, %f365, %f367;
|
|
sqrt.approx.ftz.f32 %f369, %f368;
|
|
div.approx.ftz.f32 %f370, %f363, %f369;
|
|
div.approx.ftz.f32 %f371, %f364, %f369;
|
|
div.approx.ftz.f32 %f372, %f365, %f369;
|
|
mul.ftz.f32 %f373, %f356, %f372;
|
|
mul.ftz.f32 %f374, %f357, %f371;
|
|
sub.ftz.f32 %f375, %f373, %f374;
|
|
mul.ftz.f32 %f376, %f357, %f370;
|
|
mul.ftz.f32 %f377, %f355, %f372;
|
|
sub.ftz.f32 %f378, %f376, %f377;
|
|
mul.ftz.f32 %f379, %f355, %f371;
|
|
mul.ftz.f32 %f380, %f356, %f370;
|
|
sub.ftz.f32 %f381, %f379, %f380;
|
|
mul.ftz.f32 %f382, %f378, %f378;
|
|
fma.rn.ftz.f32 %f383, %f375, %f375, %f382;
|
|
fma.rn.ftz.f32 %f384, %f381, %f381, %f383;
|
|
sqrt.approx.ftz.f32 %f385, %f384;
|
|
div.approx.ftz.f32 %f386, %f375, %f385;
|
|
div.approx.ftz.f32 %f387, %f378, %f385;
|
|
div.approx.ftz.f32 %f388, %f381, %f385;
|
|
mul.ftz.f32 %f389, %f155, %f355;
|
|
mul.ftz.f32 %f390, %f155, %f356;
|
|
mul.ftz.f32 %f391, %f155, %f357;
|
|
sub.ftz.f32 %f392, %f767, %f241;
|
|
sub.ftz.f32 %f393, %f766, %f242;
|
|
mul.ftz.f32 %f394, %f393, %f371;
|
|
fma.rn.ftz.f32 %f395, %f392, %f370, %f394;
|
|
sub.ftz.f32 %f396, %f765, %f243;
|
|
fma.rn.ftz.f32 %f156, %f396, %f372, %f395;
|
|
mul.ftz.f32 %f397, %f393, %f387;
|
|
fma.rn.ftz.f32 %f398, %f392, %f386, %f397;
|
|
fma.rn.ftz.f32 %f157, %f396, %f388, %f398;
|
|
mul.ftz.f32 %f399, %f393, %f390;
|
|
fma.rn.ftz.f32 %f400, %f392, %f389, %f399;
|
|
fma.rn.ftz.f32 %f158, %f396, %f391, %f400;
|
|
sub.ftz.f32 %f401, %f763, %f241;
|
|
sub.ftz.f32 %f402, %f762, %f242;
|
|
mul.ftz.f32 %f403, %f402, %f371;
|
|
fma.rn.ftz.f32 %f404, %f401, %f370, %f403;
|
|
sub.ftz.f32 %f405, %f761, %f243;
|
|
fma.rn.ftz.f32 %f159, %f405, %f372, %f404;
|
|
mul.ftz.f32 %f406, %f402, %f387;
|
|
fma.rn.ftz.f32 %f407, %f401, %f386, %f406;
|
|
fma.rn.ftz.f32 %f160, %f405, %f388, %f407;
|
|
mul.ftz.f32 %f408, %f402, %f390;
|
|
fma.rn.ftz.f32 %f409, %f401, %f389, %f408;
|
|
fma.rn.ftz.f32 %f161, %f405, %f391, %f409;
|
|
sub.ftz.f32 %f410, %f759, %f241;
|
|
sub.ftz.f32 %f411, %f758, %f242;
|
|
mul.ftz.f32 %f412, %f411, %f371;
|
|
fma.rn.ftz.f32 %f413, %f410, %f370, %f412;
|
|
sub.ftz.f32 %f414, %f757, %f243;
|
|
fma.rn.ftz.f32 %f162, %f414, %f372, %f413;
|
|
mul.ftz.f32 %f415, %f411, %f387;
|
|
fma.rn.ftz.f32 %f416, %f410, %f386, %f415;
|
|
fma.rn.ftz.f32 %f163, %f414, %f388, %f416;
|
|
mul.ftz.f32 %f417, %f411, %f390;
|
|
fma.rn.ftz.f32 %f418, %f410, %f389, %f417;
|
|
fma.rn.ftz.f32 %f164, %f414, %f391, %f418;
|
|
sub.ftz.f32 %f419, %f755, %f241;
|
|
sub.ftz.f32 %f420, %f754, %f242;
|
|
mul.ftz.f32 %f421, %f420, %f371;
|
|
fma.rn.ftz.f32 %f422, %f419, %f370, %f421;
|
|
sub.ftz.f32 %f423, %f753, %f243;
|
|
fma.rn.ftz.f32 %f165, %f423, %f372, %f422;
|
|
mul.ftz.f32 %f424, %f420, %f387;
|
|
fma.rn.ftz.f32 %f425, %f419, %f386, %f424;
|
|
fma.rn.ftz.f32 %f166, %f423, %f388, %f425;
|
|
mul.ftz.f32 %f426, %f420, %f390;
|
|
fma.rn.ftz.f32 %f427, %f419, %f389, %f426;
|
|
fma.rn.ftz.f32 %f167, %f423, %f391, %f427;
|
|
abs.ftz.f32 %f428, %f156;
|
|
abs.ftz.f32 %f429, %f157;
|
|
abs.ftz.f32 %f430, %f158;
|
|
abs.ftz.f32 %f431, %f159;
|
|
abs.ftz.f32 %f432, %f160;
|
|
abs.ftz.f32 %f433, %f161;
|
|
max.ftz.f32 %f434, %f428, %f431;
|
|
max.ftz.f32 %f435, %f429, %f432;
|
|
max.ftz.f32 %f436, %f430, %f433;
|
|
abs.ftz.f32 %f437, %f760;
|
|
abs.ftz.f32 %f438, %f764;
|
|
max.ftz.f32 %f439, %f438, %f437;
|
|
abs.ftz.f32 %f440, %f162;
|
|
abs.ftz.f32 %f441, %f163;
|
|
abs.ftz.f32 %f442, %f164;
|
|
abs.ftz.f32 %f443, %f165;
|
|
abs.ftz.f32 %f444, %f166;
|
|
abs.ftz.f32 %f445, %f167;
|
|
max.ftz.f32 %f446, %f440, %f443;
|
|
max.ftz.f32 %f447, %f441, %f444;
|
|
max.ftz.f32 %f448, %f442, %f445;
|
|
abs.ftz.f32 %f449, %f752;
|
|
abs.ftz.f32 %f450, %f756;
|
|
max.ftz.f32 %f451, %f450, %f449;
|
|
max.ftz.f32 %f452, %f434, %f446;
|
|
max.ftz.f32 %f453, %f435, %f447;
|
|
max.ftz.f32 %f454, %f436, %f448;
|
|
max.ftz.f32 %f455, %f439, %f451;
|
|
max.ftz.f32 %f456, %f452, %f453;
|
|
max.ftz.f32 %f457, %f454, %f455;
|
|
max.ftz.f32 %f458, %f456, %f457;
|
|
mul.ftz.f32 %f168, %f458, 0f35000000;
|
|
ld.const.u32 %r20, [kernel_params+1148];
|
|
cvt.rn.f32.s32 %f459, %r20;
|
|
rcp.approx.ftz.f32 %f169, %f459;
|
|
sub.ftz.f32 %f460, %f354, 0f00000000;
|
|
mul.ftz.f32 %f461, %f460, 0f00000000;
|
|
mul.ftz.f32 %f462, %f461, %f460;
|
|
mul.ftz.f32 %f463, %f156, %f462;
|
|
mul.ftz.f32 %f464, %f157, %f462;
|
|
mul.ftz.f32 %f465, %f158, %f462;
|
|
add.ftz.f32 %f466, %f159, %f159;
|
|
add.ftz.f32 %f467, %f160, %f160;
|
|
add.ftz.f32 %f468, %f161, %f161;
|
|
sub.ftz.f32 %f469, %f466, %f463;
|
|
sub.ftz.f32 %f470, %f467, %f464;
|
|
sub.ftz.f32 %f471, %f468, %f465;
|
|
add.ftz.f32 %f472, %f760, %f760;
|
|
mul.ftz.f32 %f473, %f764, %f462;
|
|
sub.ftz.f32 %f474, %f472, %f473;
|
|
mul.ftz.f32 %f475, %f460, %f460;
|
|
mul.ftz.f32 %f476, %f460, 0f40400000;
|
|
add.ftz.f32 %f477, %f476, 0fC0A00000;
|
|
fma.rn.ftz.f32 %f478, %f475, %f477, 0f40000000;
|
|
fma.rn.ftz.f32 %f479, %f162, %f478, %f469;
|
|
fma.rn.ftz.f32 %f480, %f163, %f478, %f470;
|
|
fma.rn.ftz.f32 %f481, %f164, %f478, %f471;
|
|
fma.rn.ftz.f32 %f482, %f756, %f478, %f474;
|
|
mul.ftz.f32 %f483, %f461, 0f00000000;
|
|
mul.ftz.f32 %f484, %f165, %f483;
|
|
mul.ftz.f32 %f485, %f166, %f483;
|
|
mul.ftz.f32 %f486, %f167, %f483;
|
|
sub.ftz.f32 %f487, %f479, %f484;
|
|
sub.ftz.f32 %f488, %f480, %f485;
|
|
sub.ftz.f32 %f170, %f481, %f486;
|
|
mul.ftz.f32 %f489, %f752, %f483;
|
|
sub.ftz.f32 %f490, %f482, %f489;
|
|
mul.ftz.f32 %f773, %f487, 0f3F000000;
|
|
mul.ftz.f32 %f772, %f488, 0f3F000000;
|
|
mul.ftz.f32 %f770, %f490, 0f3F000000;
|
|
add.ftz.f32 %f491, %f460, %f460;
|
|
mul.ftz.f32 %f492, %f491, 0f00000000;
|
|
sub.ftz.f32 %f493, %f492, %f475;
|
|
mul.ftz.f32 %f494, %f159, 0f00000000;
|
|
mul.ftz.f32 %f495, %f160, 0f00000000;
|
|
mul.ftz.f32 %f496, %f161, 0f00000000;
|
|
fma.rn.ftz.f32 %f497, %f156, %f493, %f494;
|
|
fma.rn.ftz.f32 %f498, %f157, %f493, %f495;
|
|
fma.rn.ftz.f32 %f499, %f158, %f493, %f496;
|
|
mul.ftz.f32 %f500, %f476, %f460;
|
|
add.ftz.f32 %f501, %f491, %f491;
|
|
sub.ftz.f32 %f502, %f501, %f500;
|
|
fma.rn.ftz.f32 %f503, %f162, %f502, %f497;
|
|
fma.rn.ftz.f32 %f504, %f163, %f502, %f498;
|
|
fma.rn.ftz.f32 %f505, %f164, %f502, %f499;
|
|
mul.ftz.f32 %f506, %f460, 0fC0000000;
|
|
fma.rn.ftz.f32 %f507, %f506, 0f00000000, 0f00000000;
|
|
fma.rn.ftz.f32 %f508, %f165, %f507, %f503;
|
|
fma.rn.ftz.f32 %f509, %f166, %f507, %f504;
|
|
fma.rn.ftz.f32 %f510, %f167, %f507, %f505;
|
|
mul.ftz.f32 %f768, %f508, 0f3F000000;
|
|
mul.ftz.f32 %f769, %f509, 0f3F000000;
|
|
mul.ftz.f32 %f511, %f510, 0f3F000000;
|
|
abs.ftz.f32 %f512, %f768;
|
|
abs.ftz.f32 %f513, %f769;
|
|
abs.ftz.f32 %f514, %f511;
|
|
max.ftz.f32 %f515, %f512, %f513;
|
|
max.ftz.f32 %f516, %f515, %f514;
|
|
setp.geu.ftz.f32 %p16, %f516, %f168;
|
|
@%p16 bra $L__BB7_19;
|
|
|
|
sub.ftz.f32 %f518, %f354, %f169;
|
|
mul.ftz.f32 %f519, %f169, %f518;
|
|
mul.ftz.f32 %f520, %f518, %f519;
|
|
fma.rn.ftz.f32 %f521, %f169, 0f40400000, 0fC0A00000;
|
|
mul.ftz.f32 %f522, %f169, %f169;
|
|
fma.rn.ftz.f32 %f523, %f522, %f521, 0f40000000;
|
|
mul.ftz.f32 %f524, %f518, %f518;
|
|
fma.rn.ftz.f32 %f525, %f518, 0f40400000, 0fC0A00000;
|
|
fma.rn.ftz.f32 %f526, %f524, %f525, 0f40000000;
|
|
mul.ftz.f32 %f527, %f169, %f519;
|
|
mul.ftz.f32 %f528, %f156, %f520;
|
|
mul.ftz.f32 %f529, %f157, %f520;
|
|
mul.ftz.f32 %f530, %f523, %f159;
|
|
mul.ftz.f32 %f531, %f523, %f160;
|
|
sub.ftz.f32 %f532, %f530, %f528;
|
|
sub.ftz.f32 %f533, %f531, %f529;
|
|
fma.rn.ftz.f32 %f534, %f162, %f526, %f532;
|
|
fma.rn.ftz.f32 %f535, %f526, %f163, %f533;
|
|
mul.ftz.f32 %f536, %f165, %f527;
|
|
mul.ftz.f32 %f537, %f166, %f527;
|
|
sub.ftz.f32 %f538, %f534, %f536;
|
|
sub.ftz.f32 %f539, %f535, %f537;
|
|
mul.ftz.f32 %f540, %f538, 0f3F000000;
|
|
mul.ftz.f32 %f541, %f539, 0f3F000000;
|
|
sub.ftz.f32 %f768, %f540, %f773;
|
|
sub.ftz.f32 %f769, %f541, %f772;
|
|
|
|
$L__BB7_19:
|
|
mul.ftz.f32 %f542, %f769, %f769;
|
|
fma.rn.ftz.f32 %f180, %f768, %f768, %f542;
|
|
setp.lt.s32 %p18, %r20, 1;
|
|
mov.pred %p37, -1;
|
|
@%p18 bra $L__BB7_31;
|
|
|
|
mov.u32 %r72, 0;
|
|
add.ftz.f32 %f544, %f180, 0f00000000;
|
|
mov.f32 %f545, 0f00000000;
|
|
sqrt.approx.ftz.f32 %f546, %f544;
|
|
mul.ftz.f32 %f771, %f170, 0f3F000000;
|
|
div.approx.ftz.f32 %f547, %f545, %f546;
|
|
mul.ftz.f32 %f774, %f770, %f547;
|
|
neg.ftz.f32 %f548, %f768;
|
|
div.approx.ftz.f32 %f549, %f548, %f546;
|
|
mul.ftz.f32 %f775, %f770, %f549;
|
|
div.approx.ftz.f32 %f550, %f769, %f546;
|
|
mul.ftz.f32 %f776, %f770, %f550;
|
|
mov.pred %p37, -1;
|
|
mov.pred %p30, 0;
|
|
|
|
$L__BB7_21:
|
|
mov.f32 %f191, %f776;
|
|
mov.f32 %f190, %f775;
|
|
mov.f32 %f189, %f774;
|
|
mov.f32 %f188, %f773;
|
|
mov.f32 %f187, %f772;
|
|
mov.f32 %f186, %f771;
|
|
mov.f32 %f185, %f770;
|
|
cvt.rn.f32.s32 %f551, %r72;
|
|
mul.ftz.f32 %f195, %f169, %f551;
|
|
add.ftz.f32 %f552, %f169, %f195;
|
|
sub.ftz.f32 %f554, %f354, %f552;
|
|
mul.ftz.f32 %f555, %f552, %f554;
|
|
mul.ftz.f32 %f556, %f554, %f555;
|
|
mul.ftz.f32 %f557, %f552, %f552;
|
|
mul.ftz.f32 %f558, %f552, 0f40400000;
|
|
add.ftz.f32 %f559, %f558, 0fC0A00000;
|
|
fma.rn.ftz.f32 %f560, %f557, %f559, 0f40000000;
|
|
mul.ftz.f32 %f561, %f554, %f554;
|
|
mul.ftz.f32 %f562, %f554, 0f40400000;
|
|
add.ftz.f32 %f563, %f562, 0fC0A00000;
|
|
fma.rn.ftz.f32 %f564, %f561, %f563, 0f40000000;
|
|
mul.ftz.f32 %f565, %f552, %f555;
|
|
mul.ftz.f32 %f566, %f156, %f556;
|
|
mul.ftz.f32 %f567, %f157, %f556;
|
|
mul.ftz.f32 %f568, %f158, %f556;
|
|
mul.ftz.f32 %f569, %f764, %f556;
|
|
mul.ftz.f32 %f570, %f159, %f560;
|
|
mul.ftz.f32 %f571, %f160, %f560;
|
|
mul.ftz.f32 %f572, %f161, %f560;
|
|
mul.ftz.f32 %f573, %f760, %f560;
|
|
sub.ftz.f32 %f574, %f570, %f566;
|
|
sub.ftz.f32 %f575, %f571, %f567;
|
|
sub.ftz.f32 %f576, %f572, %f568;
|
|
sub.ftz.f32 %f577, %f573, %f569;
|
|
fma.rn.ftz.f32 %f578, %f162, %f564, %f574;
|
|
fma.rn.ftz.f32 %f579, %f163, %f564, %f575;
|
|
fma.rn.ftz.f32 %f580, %f164, %f564, %f576;
|
|
fma.rn.ftz.f32 %f581, %f756, %f564, %f577;
|
|
mul.ftz.f32 %f582, %f165, %f565;
|
|
mul.ftz.f32 %f583, %f166, %f565;
|
|
mul.ftz.f32 %f584, %f167, %f565;
|
|
mul.ftz.f32 %f585, %f752, %f565;
|
|
sub.ftz.f32 %f586, %f578, %f582;
|
|
sub.ftz.f32 %f587, %f579, %f583;
|
|
sub.ftz.f32 %f588, %f580, %f584;
|
|
sub.ftz.f32 %f589, %f581, %f585;
|
|
mul.ftz.f32 %f773, %f586, 0f3F000000;
|
|
mul.ftz.f32 %f772, %f587, 0f3F000000;
|
|
mul.ftz.f32 %f771, %f588, 0f3F000000;
|
|
mul.ftz.f32 %f770, %f589, 0f3F000000;
|
|
sub.ftz.f32 %f200, %f773, %f188;
|
|
sub.ftz.f32 %f201, %f772, %f187;
|
|
mul.ftz.f32 %f590, %f201, %f201;
|
|
fma.rn.ftz.f32 %f202, %f200, %f200, %f590;
|
|
add.ftz.f32 %f591, %f554, %f554;
|
|
mul.ftz.f32 %f592, %f552, %f591;
|
|
sub.ftz.f32 %f593, %f592, %f561;
|
|
add.ftz.f32 %f594, %f552, %f552;
|
|
mul.ftz.f32 %f595, %f594, %f559;
|
|
fma.rn.ftz.f32 %f596, %f552, %f558, %f595;
|
|
add.ftz.f32 %f597, %f558, 0f40000000;
|
|
mul.ftz.f32 %f598, %f591, %f597;
|
|
mul.ftz.f32 %f599, %f554, %f562;
|
|
sub.ftz.f32 %f600, %f598, %f599;
|
|
mul.ftz.f32 %f601, %f554, 0fC0000000;
|
|
fma.rn.ftz.f32 %f602, %f552, %f601, %f557;
|
|
mul.ftz.f32 %f603, %f159, %f596;
|
|
mul.ftz.f32 %f604, %f160, %f596;
|
|
mul.ftz.f32 %f605, %f161, %f596;
|
|
fma.rn.ftz.f32 %f606, %f156, %f593, %f603;
|
|
fma.rn.ftz.f32 %f607, %f157, %f593, %f604;
|
|
fma.rn.ftz.f32 %f608, %f158, %f593, %f605;
|
|
fma.rn.ftz.f32 %f609, %f162, %f600, %f606;
|
|
fma.rn.ftz.f32 %f610, %f163, %f600, %f607;
|
|
fma.rn.ftz.f32 %f611, %f164, %f600, %f608;
|
|
fma.rn.ftz.f32 %f612, %f165, %f602, %f609;
|
|
fma.rn.ftz.f32 %f613, %f166, %f602, %f610;
|
|
fma.rn.ftz.f32 %f614, %f167, %f602, %f611;
|
|
mul.ftz.f32 %f780, %f612, 0f3F000000;
|
|
mul.ftz.f32 %f781, %f613, 0f3F000000;
|
|
mul.ftz.f32 %f205, %f614, 0f3F000000;
|
|
abs.ftz.f32 %f615, %f780;
|
|
abs.ftz.f32 %f616, %f781;
|
|
abs.ftz.f32 %f617, %f205;
|
|
max.ftz.f32 %f618, %f615, %f616;
|
|
max.ftz.f32 %f619, %f618, %f617;
|
|
setp.lt.ftz.f32 %p19, %f619, %f168;
|
|
@%p19 bra $L__BB7_22;
|
|
bra.uni $L__BB7_23;
|
|
|
|
$L__BB7_22:
|
|
mov.f32 %f780, %f200;
|
|
mov.f32 %f781, %f201;
|
|
|
|
$L__BB7_23:
|
|
max.ftz.f32 %f620, %f185, %f770;
|
|
mul.ftz.f32 %f621, %f188, %f201;
|
|
mul.ftz.f32 %f622, %f187, %f200;
|
|
sub.ftz.f32 %f623, %f622, %f621;
|
|
neg.ftz.f32 %f624, %f780;
|
|
mul.ftz.f32 %f625, %f780, %f780;
|
|
fma.rn.ftz.f32 %f626, %f781, %f781, %f625;
|
|
add.ftz.f32 %f627, %f626, 0f00000000;
|
|
sqrt.approx.ftz.f32 %f629, %f627;
|
|
div.approx.ftz.f32 %f630, %f781, %f629;
|
|
div.approx.ftz.f32 %f631, %f624, %f629;
|
|
div.approx.ftz.f32 %f632, %f545, %f629;
|
|
mul.ftz.f32 %f776, %f770, %f630;
|
|
mul.ftz.f32 %f775, %f770, %f631;
|
|
mul.ftz.f32 %f774, %f770, %f632;
|
|
mul.ftz.f32 %f633, %f620, %f620;
|
|
mul.ftz.f32 %f634, %f202, %f633;
|
|
mul.ftz.f32 %f635, %f623, %f623;
|
|
setp.gtu.ftz.f32 %p20, %f635, %f634;
|
|
@%p20 bra $L__BB7_30;
|
|
|
|
add.ftz.f32 %f636, %f191, %f188;
|
|
sub.ftz.f32 %f637, %f188, %f191;
|
|
add.ftz.f32 %f638, %f773, %f776;
|
|
sub.ftz.f32 %f639, %f638, %f637;
|
|
sub.ftz.f32 %f640, %f187, %f190;
|
|
add.ftz.f32 %f641, %f772, %f775;
|
|
sub.ftz.f32 %f642, %f641, %f640;
|
|
sub.ftz.f32 %f643, %f186, %f189;
|
|
add.ftz.f32 %f644, %f771, %f774;
|
|
sub.ftz.f32 %f645, %f644, %f643;
|
|
mul.ftz.f32 %f646, %f640, %f645;
|
|
mul.ftz.f32 %f647, %f643, %f642;
|
|
sub.ftz.f32 %f648, %f646, %f647;
|
|
mul.ftz.f32 %f649, %f643, %f639;
|
|
mul.ftz.f32 %f650, %f637, %f645;
|
|
sub.ftz.f32 %f651, %f649, %f650;
|
|
mul.ftz.f32 %f652, %f637, %f642;
|
|
mul.ftz.f32 %f653, %f640, %f639;
|
|
sub.ftz.f32 %f654, %f652, %f653;
|
|
mul.ftz.f32 %f655, %f651, 0f00000000;
|
|
fma.rn.ftz.f32 %f656, %f648, 0f00000000, %f655;
|
|
add.ftz.f32 %f218, %f654, %f656;
|
|
setp.gtu.ftz.f32 %p22, %f218, 0f00000000;
|
|
sub.ftz.f32 %f657, %f773, %f776;
|
|
selp.f32 %f219, %f657, %f636, %p22;
|
|
sub.ftz.f32 %f658, %f772, %f775;
|
|
add.ftz.f32 %f659, %f190, %f187;
|
|
selp.f32 %f220, %f658, %f659, %p22;
|
|
sub.ftz.f32 %f660, %f771, %f774;
|
|
add.ftz.f32 %f661, %f189, %f186;
|
|
selp.f32 %f221, %f660, %f661, %p22;
|
|
selp.f32 %f662, %f637, %f638, %p22;
|
|
selp.f32 %f663, %f640, %f641, %p22;
|
|
selp.f32 %f664, %f643, %f644, %p22;
|
|
selp.f32 %f665, %f638, %f637, %p22;
|
|
selp.f32 %f666, %f641, %f640, %p22;
|
|
selp.f32 %f667, %f644, %f643, %p22;
|
|
sub.ftz.f32 %f222, %f665, %f219;
|
|
sub.ftz.f32 %f223, %f666, %f220;
|
|
sub.ftz.f32 %f224, %f667, %f221;
|
|
sub.ftz.f32 %f225, %f219, %f662;
|
|
sub.ftz.f32 %f226, %f220, %f663;
|
|
sub.ftz.f32 %f227, %f221, %f664;
|
|
mul.ftz.f32 %f668, %f220, %f224;
|
|
mul.ftz.f32 %f669, %f221, %f223;
|
|
sub.ftz.f32 %f670, %f668, %f669;
|
|
mul.ftz.f32 %f671, %f221, %f222;
|
|
mul.ftz.f32 %f672, %f219, %f224;
|
|
sub.ftz.f32 %f673, %f671, %f672;
|
|
mul.ftz.f32 %f674, %f219, %f223;
|
|
mul.ftz.f32 %f675, %f220, %f222;
|
|
sub.ftz.f32 %f676, %f674, %f675;
|
|
mul.ftz.f32 %f677, %f673, 0f00000000;
|
|
fma.rn.ftz.f32 %f678, %f670, 0f00000000, %f677;
|
|
add.ftz.f32 %f228, %f676, %f678;
|
|
mul.ftz.f32 %f679, %f663, %f227;
|
|
mul.ftz.f32 %f680, %f664, %f226;
|
|
sub.ftz.f32 %f681, %f679, %f680;
|
|
mul.ftz.f32 %f682, %f664, %f225;
|
|
mul.ftz.f32 %f683, %f662, %f227;
|
|
sub.ftz.f32 %f684, %f682, %f683;
|
|
mul.ftz.f32 %f685, %f662, %f226;
|
|
mul.ftz.f32 %f686, %f663, %f225;
|
|
sub.ftz.f32 %f687, %f685, %f686;
|
|
mul.ftz.f32 %f688, %f684, 0f00000000;
|
|
fma.rn.ftz.f32 %f689, %f681, 0f00000000, %f688;
|
|
add.ftz.f32 %f229, %f687, %f689;
|
|
max.ftz.f32 %f690, %f228, %f229;
|
|
setp.gtu.ftz.f32 %p23, %f690, 0f00000000;
|
|
mov.pred %p36, %p37;
|
|
@%p23 bra $L__BB7_27;
|
|
|
|
mul.ftz.f32 %f691, %f226, %f224;
|
|
mul.ftz.f32 %f692, %f227, %f223;
|
|
sub.ftz.f32 %f693, %f691, %f692;
|
|
mul.ftz.f32 %f694, %f225, %f224;
|
|
mul.ftz.f32 %f695, %f227, %f222;
|
|
sub.ftz.f32 %f696, %f695, %f694;
|
|
mul.ftz.f32 %f697, %f226, %f222;
|
|
mul.ftz.f32 %f698, %f225, %f223;
|
|
sub.ftz.f32 %f699, %f698, %f697;
|
|
mul.ftz.f32 %f700, %f696, 0f00000000;
|
|
fma.rn.ftz.f32 %f701, %f693, 0f00000000, %f700;
|
|
add.ftz.f32 %f702, %f699, %f701;
|
|
rcp.approx.ftz.f32 %f230, %f702;
|
|
mul.ftz.f32 %f703, %f220, %f696;
|
|
fma.rn.ftz.f32 %f704, %f219, %f693, %f703;
|
|
fma.rn.ftz.f32 %f705, %f221, %f699, %f704;
|
|
mul.ftz.f32 %f231, %f705, %f230;
|
|
setp.ltu.ftz.f32 %p25, %f231, %f247;
|
|
setp.gtu.ftz.f32 %p26, %f231, %f788;
|
|
or.pred %p27, %p26, %p25;
|
|
setp.eq.ftz.f32 %p28, %f702, 0f00000000;
|
|
or.pred %p29, %p28, %p27;
|
|
mov.pred %p36, %p37;
|
|
@%p29 bra $L__BB7_27;
|
|
|
|
mul.ftz.f32 %f706, %f228, %f230;
|
|
sub.ftz.f32 %f708, %f354, %f706;
|
|
setp.le.ftz.f32 %p31, %f218, 0f00000000;
|
|
selp.f32 %f787, %f706, %f708, %p31;
|
|
mul.ftz.f32 %f709, %f229, %f230;
|
|
sub.ftz.f32 %f710, %f354, %f709;
|
|
selp.f32 %f786, %f709, %f710, %p31;
|
|
mov.f32 %f785, %f231;
|
|
mov.pred %p36, %p30;
|
|
|
|
$L__BB7_27:
|
|
@%p36 bra $L__BB7_30;
|
|
|
|
sub.ftz.f32 %f711, %f770, %f185;
|
|
fma.rn.ftz.f32 %f712, %f711, %f787, %f185;
|
|
add.ftz.f32 %f713, %f712, %f712;
|
|
mul.ftz.f32 %f714, %f155, %f713;
|
|
setp.leu.ftz.f32 %p32, %f785, %f714;
|
|
@%p32 bra $L__BB7_30;
|
|
bra.uni $L__BB7_29;
|
|
|
|
$L__BB7_30:
|
|
add.s32 %r72, %r72, 1;
|
|
setp.lt.s32 %p35, %r72, %r20;
|
|
@%p35 bra $L__BB7_21;
|
|
bra.uni $L__BB7_31;
|
|
|
|
$L__BB7_29:
|
|
cvt.rn.f32.s32 %f719, %r72;
|
|
mul.ftz.f32 %f718, %f169, %f719;
|
|
fma.rn.ftz.f32 %f715, %f786, 0f40000000, 0fBF800000;
|
|
fma.rn.ftz.f32 %f716, %f169, %f787, %f718;
|
|
mov.b32 %r74, %f716;
|
|
mov.b32 %r73, %f715;
|
|
mov.pred %p37, 0;
|
|
mov.f32 %f788, %f785;
|
|
|
|
$L__BB7_31:
|
|
@%p37 bra $L__BB7_33;
|
|
|
|
and.b32 %r63, %r1, 127;
|
|
// begin inline asm
|
|
call (%r62), _optix_report_intersection_2, (%f788, %r63, %r74, %r73);
|
|
// end inline asm
|
|
|
|
$L__BB7_33:
|
|
ret;
|
|
|
|
}
|
|
// .globl __intersection__point
|
|
.visible .entry __intersection__point()
|
|
{
|
|
.reg .pred %p<19>;
|
|
.reg .b16 %rs<7>;
|
|
.reg .f32 %f<94>;
|
|
.reg .b32 %r<44>;
|
|
.reg .b64 %rd<42>;
|
|
|
|
|
|
// begin inline asm
|
|
call (%r17), _optix_read_primitive_idx, ();
|
|
// end inline asm
|
|
mov.u32 %r18, 0;
|
|
// begin inline asm
|
|
call (%rd12), _optix_get_transform_list_handle, (%r18);
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r19), _optix_get_instance_id_from_handle, (%rd12);
|
|
// end inline asm
|
|
ld.const.u64 %rd14, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd15, %rd14;
|
|
mul.wide.s32 %rd16, %r19, 208;
|
|
add.s64 %rd17, %rd15, %rd16;
|
|
add.s64 %rd1, %rd17, 200;
|
|
ld.global.u32 %r2, [%rd17+200];
|
|
mov.u32 %r21, 4;
|
|
// begin inline asm
|
|
call (%r20), _optix_get_payload, (%r21);
|
|
// end inline asm
|
|
ld.global.u32 %r22, [%rd17+196];
|
|
and.b32 %r23, %r22, %r20;
|
|
setp.eq.s32 %p3, %r23, 0;
|
|
@%p3 bra $L__BB8_20;
|
|
|
|
// begin inline asm
|
|
call (%f33), _optix_get_object_ray_origin_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f34), _optix_get_object_ray_origin_y, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f35), _optix_get_object_ray_origin_z, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f36), _optix_get_object_ray_direction_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f37), _optix_get_object_ray_direction_y, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f38), _optix_get_object_ray_direction_z, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f39), _optix_get_ray_tmin, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f40), _optix_get_ray_time, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%f93), _optix_get_ray_tmax, ();
|
|
// end inline asm
|
|
and.b32 %r24, %r2, 64;
|
|
setp.eq.s32 %p4, %r24, 0;
|
|
@%p4 bra $L__BB8_15;
|
|
bra.uni $L__BB8_2;
|
|
|
|
$L__BB8_15:
|
|
ld.const.u64 %rd35, [kernel_params+2312];
|
|
cvta.to.global.u64 %rd36, %rd35;
|
|
mul.wide.s32 %rd37, %r17, 16;
|
|
add.s64 %rd38, %rd36, %rd37;
|
|
ld.global.v4.f32 {%f89, %f90, %f91, %f92}, [%rd38];
|
|
bra.uni $L__BB8_16;
|
|
|
|
$L__BB8_2:
|
|
ld.global.u32 %r3, [%rd1+-52];
|
|
ld.global.u32 %r4, [%rd1+-48];
|
|
shl.b32 %r25, %r4, 1;
|
|
cvt.rn.f32.s32 %f42, %r25;
|
|
mul.ftz.f32 %f10, %f40, %f42;
|
|
cvt.rzi.ftz.s32.f32 %r26, %f10;
|
|
add.s32 %r27, %r25, -1;
|
|
min.s32 %r5, %r26, %r27;
|
|
ld.const.u64 %rd18, [kernel_params+2328];
|
|
cvta.to.global.u64 %rd2, %rd18;
|
|
ld.global.u32 %r40, [%rd1+-36];
|
|
mul.wide.u32 %rd19, %r40, 16;
|
|
add.s64 %rd20, %rd2, %rd19;
|
|
ld.global.u16 %rs5, [%rd20+12];
|
|
ld.global.u32 %r39, [%rd20+8];
|
|
ld.global.u64 %rd39, [%rd20];
|
|
setp.eq.s64 %p5, %rd39, 11;
|
|
@%p5 bra $L__BB8_7;
|
|
|
|
$L__BB8_3:
|
|
setp.eq.s64 %p6, %rd39, 0;
|
|
@%p6 bra $L__BB8_5;
|
|
|
|
add.s32 %r40, %r40, 2;
|
|
bra.uni $L__BB8_6;
|
|
|
|
$L__BB8_5:
|
|
setp.eq.s16 %p7, %rs5, 0;
|
|
mov.u32 %r43, -1;
|
|
mov.u32 %r40, %r39;
|
|
@%p7 bra $L__BB8_8;
|
|
|
|
$L__BB8_6:
|
|
mul.wide.u32 %rd21, %r40, 16;
|
|
add.s64 %rd22, %rd2, %rd21;
|
|
ld.global.u16 %rs5, [%rd22+12];
|
|
ld.global.u32 %r39, [%rd22+8];
|
|
ld.global.u64 %rd39, [%rd22];
|
|
setp.ne.s64 %p8, %rd39, 11;
|
|
@%p8 bra $L__BB8_3;
|
|
|
|
$L__BB8_7:
|
|
setp.eq.s16 %p9, %rs5, 0;
|
|
selp.b32 %r43, -1, %r39, %p9;
|
|
|
|
$L__BB8_8:
|
|
setp.eq.s32 %p10, %r5, %r4;
|
|
@%p10 bra $L__BB8_10;
|
|
bra.uni $L__BB8_9;
|
|
|
|
$L__BB8_10:
|
|
ld.const.u64 %rd26, [kernel_params+2312];
|
|
cvta.to.global.u64 %rd27, %rd26;
|
|
mul.wide.s32 %rd28, %r17, 16;
|
|
add.s64 %rd40, %rd27, %rd28;
|
|
bra.uni $L__BB8_11;
|
|
|
|
$L__BB8_9:
|
|
setp.gt.s32 %p11, %r5, %r4;
|
|
selp.b32 %r29, -1, 0, %p11;
|
|
add.s32 %r30, %r5, %r29;
|
|
ld.const.u64 %rd23, [kernel_params+2360];
|
|
cvta.to.global.u64 %rd24, %rd23;
|
|
mad.lo.s32 %r31, %r30, %r3, %r17;
|
|
add.s32 %r32, %r31, %r43;
|
|
mul.wide.s32 %rd25, %r32, 16;
|
|
add.s64 %rd40, %rd24, %rd25;
|
|
|
|
$L__BB8_11:
|
|
ld.global.v4.f32 {%f43, %f44, %f45, %f46}, [%rd40];
|
|
add.s32 %r16, %r5, 1;
|
|
setp.eq.s32 %p12, %r16, %r4;
|
|
@%p12 bra $L__BB8_13;
|
|
bra.uni $L__BB8_12;
|
|
|
|
$L__BB8_13:
|
|
ld.const.u64 %rd32, [kernel_params+2312];
|
|
cvta.to.global.u64 %rd33, %rd32;
|
|
mul.wide.s32 %rd34, %r17, 16;
|
|
add.s64 %rd41, %rd33, %rd34;
|
|
bra.uni $L__BB8_14;
|
|
|
|
$L__BB8_12:
|
|
setp.ge.s32 %p13, %r5, %r4;
|
|
selp.b32 %r33, -1, 0, %p13;
|
|
add.s32 %r34, %r16, %r33;
|
|
ld.const.u64 %rd29, [kernel_params+2360];
|
|
cvta.to.global.u64 %rd30, %rd29;
|
|
mad.lo.s32 %r35, %r34, %r3, %r17;
|
|
add.s32 %r36, %r35, %r43;
|
|
mul.wide.s32 %rd31, %r36, 16;
|
|
add.s64 %rd41, %rd30, %rd31;
|
|
|
|
$L__BB8_14:
|
|
cvt.rn.f32.s32 %f47, %r5;
|
|
sub.ftz.f32 %f48, %f10, %f47;
|
|
ld.global.v4.f32 {%f49, %f50, %f51, %f52}, [%rd41];
|
|
mov.f32 %f57, 0f3F800000;
|
|
sub.ftz.f32 %f58, %f57, %f48;
|
|
mul.ftz.f32 %f59, %f48, %f49;
|
|
mul.ftz.f32 %f60, %f48, %f50;
|
|
mul.ftz.f32 %f61, %f48, %f51;
|
|
mul.ftz.f32 %f62, %f48, %f52;
|
|
fma.rn.ftz.f32 %f92, %f58, %f46, %f62;
|
|
fma.rn.ftz.f32 %f91, %f58, %f45, %f61;
|
|
fma.rn.ftz.f32 %f90, %f58, %f44, %f60;
|
|
fma.rn.ftz.f32 %f89, %f58, %f43, %f59;
|
|
|
|
$L__BB8_16:
|
|
mul.ftz.f32 %f67, %f37, %f37;
|
|
fma.rn.ftz.f32 %f68, %f36, %f36, %f67;
|
|
fma.rn.ftz.f32 %f69, %f38, %f38, %f68;
|
|
rcp.approx.ftz.f32 %f27, %f69;
|
|
sub.ftz.f32 %f70, %f89, %f33;
|
|
sub.ftz.f32 %f71, %f90, %f34;
|
|
sub.ftz.f32 %f72, %f91, %f35;
|
|
mul.ftz.f32 %f73, %f37, %f71;
|
|
fma.rn.ftz.f32 %f74, %f36, %f70, %f73;
|
|
fma.rn.ftz.f32 %f75, %f38, %f72, %f74;
|
|
mul.ftz.f32 %f28, %f27, %f75;
|
|
mul.ftz.f32 %f76, %f36, %f28;
|
|
mul.ftz.f32 %f77, %f37, %f28;
|
|
mul.ftz.f32 %f78, %f38, %f28;
|
|
sub.ftz.f32 %f79, %f70, %f76;
|
|
sub.ftz.f32 %f80, %f71, %f77;
|
|
sub.ftz.f32 %f81, %f72, %f78;
|
|
mul.ftz.f32 %f82, %f80, %f80;
|
|
fma.rn.ftz.f32 %f83, %f79, %f79, %f82;
|
|
fma.rn.ftz.f32 %f29, %f81, %f81, %f83;
|
|
mul.ftz.f32 %f30, %f92, %f92;
|
|
setp.gtu.ftz.f32 %p15, %f29, %f30;
|
|
mov.pred %p18, -1;
|
|
@%p15 bra $L__BB8_18;
|
|
|
|
sub.ftz.f32 %f84, %f30, %f29;
|
|
mul.ftz.f32 %f85, %f27, %f84;
|
|
sqrt.approx.ftz.f32 %f86, %f85;
|
|
sub.ftz.f32 %f87, %f28, %f86;
|
|
setp.ltu.ftz.f32 %p16, %f87, %f39;
|
|
setp.gtu.ftz.f32 %p17, %f87, %f93;
|
|
or.pred %p18, %p17, %p16;
|
|
selp.f32 %f93, %f93, %f87, %p18;
|
|
|
|
$L__BB8_18:
|
|
@%p18 bra $L__BB8_20;
|
|
|
|
and.b32 %r38, %r2, 127;
|
|
// begin inline asm
|
|
call (%r37), _optix_report_intersection_0, (%f93, %r38);
|
|
// end inline asm
|
|
|
|
$L__BB8_20:
|
|
ret;
|
|
|
|
}
|
|
// .globl __raygen__kernel_optix_integrator_intersect_closest
|
|
.visible .entry __raygen__kernel_optix_integrator_intersect_closest()
|
|
{
|
|
.local .align 4 .b8 __local_depot9[60];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<273>;
|
|
.reg .b16 %rs<88>;
|
|
.reg .f32 %f<353>;
|
|
.reg .b32 %r<604>;
|
|
.reg .b64 %rd<805>;
|
|
|
|
|
|
mov.u64 %SPL, __local_depot9;
|
|
cvta.local.u64 %SP, %SPL;
|
|
// begin inline asm
|
|
call (%r3), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
ld.const.u64 %rd1, [kernel_params];
|
|
setp.eq.s64 %p13, %rd1, 0;
|
|
@%p13 bra $L__BB9_2;
|
|
|
|
cvta.to.global.u64 %rd136, %rd1;
|
|
mul.wide.s32 %rd137, %r3, 4;
|
|
add.s64 %rd138, %rd136, %rd137;
|
|
ld.global.u32 %r3, [%rd138];
|
|
|
|
$L__BB9_2:
|
|
add.u64 %rd140, %SPL, 0;
|
|
add.s64 %rd3, %rd140, 36;
|
|
mov.u32 %r129, 0;
|
|
st.local.u32 [%rd140+36], %r129;
|
|
st.local.u32 [%rd140+40], %r129;
|
|
st.local.u32 [%rd140+44], %r129;
|
|
st.local.u32 [%rd140+48], %r129;
|
|
ld.const.u64 %rd4, [kernel_params+2648];
|
|
cvta.to.global.u64 %rd5, %rd4;
|
|
cvt.s64.s32 %rd6, %r3;
|
|
mul.wide.s32 %rd141, %r3, 12;
|
|
add.s64 %rd7, %rd5, %rd141;
|
|
ld.global.f32 %f1, [%rd7];
|
|
ld.global.f32 %f2, [%rd7+4];
|
|
ld.global.f32 %f3, [%rd7+8];
|
|
st.local.f32 [%rd140], %f1;
|
|
st.local.f32 [%rd140+4], %f2;
|
|
st.local.f32 [%rd140+8], %f3;
|
|
ld.const.v2.u64 {%rd142, %rd143}, [kernel_params+2656];
|
|
cvta.to.global.u64 %rd10, %rd142;
|
|
add.s64 %rd11, %rd10, %rd141;
|
|
ld.global.f32 %f4, [%rd11];
|
|
ld.global.f32 %f5, [%rd11+4];
|
|
ld.global.f32 %f6, [%rd11+8];
|
|
st.local.f32 [%rd140+12], %f4;
|
|
st.local.f32 [%rd140+16], %f5;
|
|
st.local.f32 [%rd140+20], %f6;
|
|
cvta.to.global.u64 %rd12, %rd143;
|
|
mul.wide.s32 %rd144, %r3, 4;
|
|
add.s64 %rd13, %rd12, %rd144;
|
|
ld.global.f32 %f7, [%rd13];
|
|
st.local.f32 [%rd140+24], %f7;
|
|
ld.const.v2.u64 {%rd145, %rd146}, [kernel_params+2672];
|
|
cvta.to.global.u64 %rd16, %rd145;
|
|
add.s64 %rd17, %rd16, %rd144;
|
|
ld.global.f32 %f328, [%rd17];
|
|
st.local.f32 [%rd140+28], %f328;
|
|
cvta.to.global.u64 %rd18, %rd146;
|
|
add.s64 %rd19, %rd18, %rd144;
|
|
ld.global.f32 %f9, [%rd19];
|
|
st.local.f32 [%rd140+32], %f9;
|
|
ld.const.v2.u64 {%rd147, %rd148}, [kernel_params+2688];
|
|
cvta.to.global.u64 %rd22, %rd147;
|
|
add.s64 %rd23, %rd22, %rd144;
|
|
ld.global.f32 %f102, [%rd23];
|
|
st.local.f32 [%rd140+52], %f102;
|
|
cvta.to.global.u64 %rd24, %rd148;
|
|
add.s64 %rd25, %rd24, %rd144;
|
|
ld.global.f32 %f103, [%rd25];
|
|
st.local.f32 [%rd140+56], %f103;
|
|
ld.const.u64 %rd26, [kernel_params+2560];
|
|
cvta.to.global.u64 %rd27, %rd26;
|
|
add.s64 %rd28, %rd27, %rd144;
|
|
ld.global.u32 %r4, [%rd28];
|
|
ld.const.u64 %rd29, [kernel_params+2728];
|
|
cvta.to.global.u64 %rd30, %rd29;
|
|
add.s64 %rd31, %rd30, %rd144;
|
|
ld.global.u32 %r5, [%rd31];
|
|
ld.const.u64 %rd32, [kernel_params+2736];
|
|
cvta.to.global.u64 %rd33, %rd32;
|
|
add.s64 %rd34, %rd33, %rd144;
|
|
ld.global.u32 %r6, [%rd34];
|
|
ld.const.u32 %r7, [kernel_params+1552];
|
|
setp.eq.s32 %p14, %r7, 0;
|
|
mov.u16 %rs77, 0;
|
|
@%p14 bra $L__BB9_4;
|
|
|
|
ld.const.v2.u64 {%rd149, %rd150}, [kernel_params+2480];
|
|
cvta.to.global.u64 %rd152, %rd149;
|
|
shl.b64 %rd153, %rd6, 1;
|
|
add.s64 %rd154, %rd152, %rd153;
|
|
ld.global.u16 %r130, [%rd154];
|
|
ld.const.u64 %rd155, [kernel_params+2512];
|
|
cvta.to.global.u64 %rd156, %rd155;
|
|
add.s64 %rd157, %rd156, %rd153;
|
|
ld.global.u16 %r131, [%rd157];
|
|
sub.s32 %r132, %r130, %r131;
|
|
ld.const.v2.u64 {%rd158, %rd159}, [kernel_params+2496];
|
|
cvta.to.global.u64 %rd161, %rd159;
|
|
add.s64 %rd162, %rd161, %rd153;
|
|
ld.global.u16 %rs26, [%rd162];
|
|
setp.ne.s16 %p15, %rs26, 0;
|
|
selp.b32 %r133, -1, 0, %p15;
|
|
add.s32 %r134, %r132, %r133;
|
|
setp.ge.s32 %p16, %r134, %r7;
|
|
selp.u16 %rs77, 1, 0, %p16;
|
|
|
|
$L__BB9_4:
|
|
setp.eq.s16 %p17, %rs77, 0;
|
|
@%p17 bra $L__BB9_8;
|
|
|
|
ld.const.f32 %f328, [kernel_params+1556];
|
|
st.local.f32 [%rd3+-8], %f328;
|
|
setp.eq.s32 %p18, %r6, -1;
|
|
@%p18 bra $L__BB9_8;
|
|
|
|
cvt.s64.s32 %rd787, %r6;
|
|
ld.const.u64 %rd163, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd164, %rd163;
|
|
mul.lo.s64 %rd165, %rd787, 208;
|
|
add.s64 %rd166, %rd164, %rd165;
|
|
ld.global.f32 %f11, [%rd166+188];
|
|
setp.eq.ftz.f32 %p19, %f11, 0f00000000;
|
|
@%p19 bra $L__BB9_8;
|
|
|
|
add.u64 %rd789, %SPL, 0;
|
|
add.s64 %rd788, %rd789, 36;
|
|
st.local.f32 [%rd788+-8], %f11;
|
|
mov.f32 %f328, %f11;
|
|
|
|
$L__BB9_8:
|
|
add.u64 %rd792, %SPL, 0;
|
|
add.s64 %rd791, %rd792, 36;
|
|
st.local.u32 [%rd791+4], %r6;
|
|
st.local.u32 [%rd791], %r5;
|
|
mov.u32 %r135, -1;
|
|
st.local.u32 [%rd791+12], %r135;
|
|
st.local.u32 [%rd791+8], %r135;
|
|
mov.b32 %r8, %f1;
|
|
setp.nan.ftz.f32 %p20, %f1, %f1;
|
|
mov.u16 %rs80, 0;
|
|
mov.u16 %rs78, %rs80;
|
|
@%p20 bra $L__BB9_11;
|
|
|
|
or.b32 %r136, %r8, -2147483648;
|
|
setp.ne.s32 %p21, %r136, -2147483648;
|
|
add.ftz.f32 %f104, %f1, %f1;
|
|
setp.eq.ftz.f32 %p22, %f104, %f1;
|
|
and.pred %p23, %p22, %p21;
|
|
mov.u16 %rs78, %rs80;
|
|
@%p23 bra $L__BB9_11;
|
|
|
|
shl.b32 %r137, %r8, 1;
|
|
setp.lt.u32 %p24, %r137, -16777215;
|
|
selp.u16 %rs78, 1, 0, %p24;
|
|
|
|
$L__BB9_11:
|
|
setp.eq.s16 %p25, %rs78, 0;
|
|
@%p25 bra $L__BB9_17;
|
|
|
|
mov.b32 %r9, %f4;
|
|
setp.nan.ftz.f32 %p26, %f4, %f4;
|
|
mov.u16 %rs80, 0;
|
|
mov.u16 %rs79, %rs80;
|
|
@%p26 bra $L__BB9_15;
|
|
|
|
or.b32 %r138, %r9, -2147483648;
|
|
setp.ne.s32 %p27, %r138, -2147483648;
|
|
add.ftz.f32 %f105, %f4, %f4;
|
|
setp.eq.ftz.f32 %p28, %f105, %f4;
|
|
and.pred %p29, %p28, %p27;
|
|
mov.u16 %rs79, %rs80;
|
|
@%p29 bra $L__BB9_15;
|
|
|
|
shl.b32 %r139, %r9, 1;
|
|
setp.lt.u32 %p30, %r139, -16777215;
|
|
selp.u16 %rs79, 1, 0, %p30;
|
|
|
|
$L__BB9_15:
|
|
setp.eq.s16 %p31, %rs79, 0;
|
|
@%p31 bra $L__BB9_17;
|
|
|
|
mul.ftz.f32 %f106, %f5, %f5;
|
|
fma.rn.ftz.f32 %f107, %f4, %f4, %f106;
|
|
fma.rn.ftz.f32 %f108, %f6, %f6, %f107;
|
|
setp.neu.ftz.f32 %p32, %f108, 0f00000000;
|
|
selp.u16 %rs80, 1, 0, %p32;
|
|
|
|
$L__BB9_17:
|
|
mov.u32 %r575, 0;
|
|
add.u64 %rd786, %SP, 0;
|
|
and.b32 %r211, %r4, 4;
|
|
setp.eq.s32 %p33, %r211, 0;
|
|
selp.b32 %r212, 2047, 2023, %p33;
|
|
and.b32 %r213, %r212, %r4;
|
|
and.b32 %r214, %r4, 128;
|
|
shr.u32 %r215, %r214, 4;
|
|
or.b32 %r216, %r213, %r215;
|
|
and.b32 %r217, %r4, 1073741824;
|
|
setp.eq.s32 %p34, %r217, 0;
|
|
shl.b32 %r218, %r216, 16;
|
|
selp.b32 %r183, %r216, %r218, %p34;
|
|
and.b32 %r219, %r183, 255;
|
|
setp.eq.s32 %p35, %r219, 0;
|
|
setp.gt.u32 %p36, %r183, 255;
|
|
and.pred %p37, %p36, %p35;
|
|
ld.const.u64 %rd168, [kernel_params+2096];
|
|
setp.eq.s16 %p38, %rs80, 0;
|
|
selp.b64 %rd167, 0, %rd168, %p38;
|
|
cvt.u32.u64 %r185, %rd786;
|
|
shr.u64 %rd170, %rd786, 32;
|
|
cvt.u32.u64 %r186, %rd170;
|
|
selp.b32 %r173, 255, %r219, %p37;
|
|
and.b32 %r220, %r183, 256;
|
|
setp.eq.s32 %p39, %r220, 0;
|
|
or.pred %p40, %p39, %p37;
|
|
selp.b32 %r174, 2, 6, %p40;
|
|
mov.u32 %r178, 8;
|
|
// begin inline asm
|
|
call(%r140,%r141,%r142,%r143,%r585,%r69,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167,%r168,%r169,%r170,%r171),_optix_trace_typed_32,(%r575,%rd167,%f1,%f2,%f3,%f4,%f5,%f6,%f7,%f328,%f9,%r173,%r174,%r575,%r575,%r575,%r178,%r575,%r575,%r575,%r575,%r183,%r575,%r185,%r186,%r221,%r222,%r223,%r224,%r225,%r226,%r227,%r228,%r229,%r230,%r231,%r232,%r233,%r234,%r235,%r236,%r237,%r238,%r239,%r240,%r241,%r242,%r243,%r244);
|
|
// end inline asm
|
|
mov.b32 %f342, %r140;
|
|
mov.b32 %f341, %r141;
|
|
mov.b32 %f340, %r142;
|
|
setp.ne.s32 %p271, %r69, 0;
|
|
selp.b32 %r71, %r143, -1, %p271;
|
|
ld.global.u32 %r44, [%rd28];
|
|
ld.const.u32 %r45, [kernel_params+1608];
|
|
setp.eq.s32 %p41, %r45, 0;
|
|
@%p41 bra $L__BB9_24;
|
|
|
|
and.b32 %r245, %r44, 1;
|
|
setp.eq.b32 %p43, %r245, 1;
|
|
setp.eq.s32 %p44, %r6, -1;
|
|
mov.u16 %rs81, 0;
|
|
mov.pred %p267, 0;
|
|
or.pred %p45, %p44, %p43;
|
|
@%p45 bra $L__BB9_20;
|
|
|
|
cvt.s64.s32 %rd779, %r6;
|
|
ld.const.u64 %rd171, [kernel_params+2200];
|
|
cvta.to.global.u64 %rd172, %rd171;
|
|
shl.b64 %rd173, %rd779, 2;
|
|
add.s64 %rd174, %rd172, %rd173;
|
|
ld.global.u32 %r246, [%rd174];
|
|
bfe.u32 %r247, %r246, 10, 1;
|
|
cvt.u16.u32 %rs81, %r247;
|
|
shr.u32 %r248, %r246, 9;
|
|
and.b32 %r249, %r248, 1;
|
|
setp.eq.b32 %p267, %r249, 1;
|
|
|
|
$L__BB9_20:
|
|
ld.const.u64 %rd175, [kernel_params+2568];
|
|
cvta.to.global.u64 %rd176, %rd175;
|
|
add.s64 %rd36, %rd176, %rd6;
|
|
ld.global.u8 %rs11, [%rd36];
|
|
and.b16 %rs34, %rs11, 2;
|
|
setp.eq.s16 %p46, %rs34, 0;
|
|
and.b16 %rs82, %rs11, 251;
|
|
st.global.u8 [%rd36], %rs82;
|
|
not.pred %p47, %p267;
|
|
or.pred %p48, %p47, %p46;
|
|
@%p48 bra $L__BB9_22;
|
|
|
|
or.b16 %rs82, %rs11, 4;
|
|
st.global.u8 [%rd36], %rs82;
|
|
|
|
$L__BB9_22:
|
|
setp.eq.s16 %p49, %rs81, 0;
|
|
@%p49 bra $L__BB9_24;
|
|
|
|
or.b16 %rs35, %rs82, 2;
|
|
st.global.u8 [%rd36], %rs35;
|
|
|
|
$L__BB9_24:
|
|
ld.const.u32 %r250, [kernel_params+1604];
|
|
setp.eq.s32 %p50, %r250, 0;
|
|
ld.const.u64 %rd37, [kernel_params+2744];
|
|
@%p50 bra $L__BB9_61;
|
|
|
|
ld.const.u32 %r47, [kernel_params+1512];
|
|
setp.lt.s32 %p51, %r47, 1;
|
|
mov.u32 %r584, %r69;
|
|
@%p51 bra $L__BB9_60;
|
|
|
|
add.u64 %rd781, %SPL, 0;
|
|
add.s64 %rd780, %rd781, 36;
|
|
mov.u32 %r580, 0;
|
|
cvta.to.global.u64 %rd177, %rd37;
|
|
shl.b64 %rd178, %rd6, 2;
|
|
add.s64 %rd179, %rd177, %rd178;
|
|
ld.global.u32 %r48, [%rd179];
|
|
ld.const.u64 %rd180, [kernel_params+2384];
|
|
cvta.to.global.u64 %rd181, %rd180;
|
|
and.b32 %r49, %r44, 1;
|
|
and.b32 %r50, %r44, 1073741824;
|
|
ld.const.u64 %rd182, [kernel_params+2568];
|
|
cvta.to.global.u64 %rd183, %rd182;
|
|
add.s64 %rd38, %rd183, %rd6;
|
|
ld.local.f32 %f16, [%rd780+-36];
|
|
ld.local.f32 %f17, [%rd780+-32];
|
|
ld.local.f32 %f18, [%rd780+-28];
|
|
ld.local.f32 %f19, [%rd780+-24];
|
|
ld.local.f32 %f20, [%rd780+-20];
|
|
ld.local.f32 %f21, [%rd780+-16];
|
|
ld.local.f32 %f22, [%rd780+-12];
|
|
ld.local.f32 %f23, [%rd780+-8];
|
|
add.s64 %rd799, %rd181, 144;
|
|
neg.s32 %r579, %r5;
|
|
setp.eq.s32 %p52, %r49, 0;
|
|
setp.ne.s32 %p57, %r50, 0;
|
|
mov.pred %p83, -1;
|
|
setp.eq.s32 %p104, %r48, 32;
|
|
mov.u32 %r584, %r69;
|
|
|
|
$L__BB9_27:
|
|
ld.global.u32 %r57, [%rd799+-128];
|
|
@%p52 bra $L__BB9_29;
|
|
bra.uni $L__BB9_28;
|
|
|
|
$L__BB9_29:
|
|
and.b32 %r253, %r57, 268435456;
|
|
setp.eq.s32 %p54, %r253, 0;
|
|
@%p54 bra $L__BB9_59;
|
|
|
|
ld.global.u8 %rs36, [%rd38];
|
|
and.b16 %rs37, %rs36, 4;
|
|
setp.eq.s16 %p55, %rs37, 0;
|
|
@%p55 bra $L__BB9_32;
|
|
|
|
ld.global.u32 %r254, [%rd799+-104];
|
|
setp.ne.s32 %p56, %r254, 0;
|
|
@%p56 bra $L__BB9_59;
|
|
bra.uni $L__BB9_32;
|
|
|
|
$L__BB9_28:
|
|
and.b32 %r252, %r57, 16777216;
|
|
setp.eq.s32 %p53, %r252, 0;
|
|
@%p53 bra $L__BB9_32;
|
|
bra.uni $L__BB9_59;
|
|
|
|
$L__BB9_32:
|
|
shr.u32 %r255, %r57, 22;
|
|
and.b32 %r256, %r255, 1;
|
|
setp.eq.b32 %p58, %r256, 1;
|
|
and.pred %p59, %p57, %p58;
|
|
@%p59 bra $L__BB9_59;
|
|
|
|
ld.global.u32 %r257, [%rd799+-144];
|
|
setp.eq.s32 %p60, %r257, 0;
|
|
@%p60 bra $L__BB9_44;
|
|
|
|
setp.eq.s32 %p61, %r257, 4;
|
|
@%p61 bra $L__BB9_50;
|
|
|
|
setp.ne.s32 %p62, %r257, 3;
|
|
@%p62 bra $L__BB9_59;
|
|
|
|
ld.global.f32 %f27, [%rd799+12];
|
|
abs.ftz.f32 %f118, %f27;
|
|
setp.eq.ftz.f32 %p63, %f118, 0f00000000;
|
|
@%p63 bra $L__BB9_59;
|
|
|
|
ld.global.v4.f32 {%f119, %f120, %f121, %f122}, [%rd799+32];
|
|
mul.ftz.f32 %f123, %f120, %f20;
|
|
fma.rn.ftz.f32 %f124, %f119, %f19, %f123;
|
|
fma.rn.ftz.f32 %f31, %f121, %f21, %f124;
|
|
setp.ge.ftz.f32 %p64, %f31, 0f00000000;
|
|
@%p64 bra $L__BB9_59;
|
|
|
|
mul.ftz.f32 %f128, %f120, %f17;
|
|
fma.rn.ftz.f32 %f129, %f119, %f16, %f128;
|
|
fma.rn.ftz.f32 %f130, %f121, %f18, %f129;
|
|
ld.global.f32 %f32, [%rd799+-140];
|
|
ld.global.f32 %f33, [%rd799+-136];
|
|
mul.ftz.f32 %f131, %f120, %f33;
|
|
fma.rn.ftz.f32 %f132, %f119, %f32, %f131;
|
|
ld.global.f32 %f34, [%rd799+-132];
|
|
fma.rn.ftz.f32 %f133, %f121, %f34, %f132;
|
|
sub.ftz.f32 %f134, %f130, %f133;
|
|
neg.ftz.f32 %f135, %f134;
|
|
div.approx.ftz.f32 %f35, %f135, %f31;
|
|
setp.leu.ftz.f32 %p66, %f35, %f22;
|
|
setp.geu.ftz.f32 %p67, %f35, %f23;
|
|
or.pred %p68, %p67, %p66;
|
|
mov.pred %p268, -1;
|
|
mov.f32 %f332, 0f00000000;
|
|
mov.f32 %f333, %f332;
|
|
mov.f32 %f334, %f332;
|
|
@%p68 bra $L__BB9_43;
|
|
|
|
ld.global.v4.f32 {%f139, %f140, %f141, %f142}, [%rd799];
|
|
fma.rn.ftz.f32 %f146, %f19, %f35, %f16;
|
|
fma.rn.ftz.f32 %f147, %f20, %f35, %f17;
|
|
fma.rn.ftz.f32 %f148, %f21, %f35, %f18;
|
|
sub.ftz.f32 %f36, %f146, %f32;
|
|
sub.ftz.f32 %f37, %f147, %f33;
|
|
sub.ftz.f32 %f38, %f148, %f34;
|
|
mul.ftz.f32 %f149, %f140, %f37;
|
|
fma.rn.ftz.f32 %f150, %f139, %f36, %f149;
|
|
fma.rn.ftz.f32 %f151, %f141, %f38, %f150;
|
|
mul.ftz.f32 %f152, %f140, %f140;
|
|
fma.rn.ftz.f32 %f153, %f139, %f139, %f152;
|
|
fma.rn.ftz.f32 %f154, %f141, %f141, %f153;
|
|
div.approx.ftz.f32 %f39, %f151, %f154;
|
|
setp.lt.ftz.f32 %p70, %f39, 0fBF000000;
|
|
setp.gt.ftz.f32 %p71, %f39, 0f3F000000;
|
|
or.pred %p72, %p70, %p71;
|
|
mov.f32 %f333, %f332;
|
|
mov.f32 %f334, %f332;
|
|
@%p72 bra $L__BB9_43;
|
|
|
|
mov.pred %p268, -1;
|
|
mov.f32 %f332, 0f00000000;
|
|
ld.global.v4.f32 {%f158, %f159, %f160, %f161}, [%rd799+16];
|
|
mul.ftz.f32 %f165, %f159, %f37;
|
|
fma.rn.ftz.f32 %f166, %f158, %f36, %f165;
|
|
fma.rn.ftz.f32 %f167, %f160, %f38, %f166;
|
|
mul.ftz.f32 %f168, %f159, %f159;
|
|
fma.rn.ftz.f32 %f169, %f158, %f158, %f168;
|
|
fma.rn.ftz.f32 %f170, %f160, %f160, %f169;
|
|
div.approx.ftz.f32 %f40, %f167, %f170;
|
|
setp.lt.ftz.f32 %p74, %f40, 0fBF000000;
|
|
setp.gt.ftz.f32 %p75, %f40, 0f3F000000;
|
|
or.pred %p76, %p74, %p75;
|
|
mov.f32 %f333, %f332;
|
|
mov.f32 %f334, %f332;
|
|
@%p76 bra $L__BB9_43;
|
|
|
|
mov.pred %p268, -1;
|
|
mov.f32 %f332, 0f00000000;
|
|
setp.lt.ftz.f32 %p78, %f27, 0f00000000;
|
|
mul.ftz.f32 %f174, %f40, %f40;
|
|
fma.rn.ftz.f32 %f175, %f39, %f39, %f174;
|
|
setp.gt.ftz.f32 %p79, %f175, 0f3E800000;
|
|
and.pred %p80, %p78, %p79;
|
|
mov.f32 %f333, %f332;
|
|
mov.f32 %f334, %f332;
|
|
@%p80 bra $L__BB9_43;
|
|
|
|
add.ftz.f32 %f333, %f40, 0f3F000000;
|
|
neg.ftz.f32 %f176, %f39;
|
|
sub.ftz.f32 %f332, %f176, %f40;
|
|
mov.pred %p268, 0;
|
|
mov.f32 %f334, %f35;
|
|
|
|
$L__BB9_43:
|
|
@%p268 bra $L__BB9_59;
|
|
bra.uni $L__BB9_57;
|
|
|
|
$L__BB9_44:
|
|
ld.global.f32 %f46, [%rd799+-140];
|
|
ld.global.f32 %f47, [%rd799+-136];
|
|
ld.global.f32 %f48, [%rd799+-132];
|
|
ld.global.f32 %f49, [%rd799];
|
|
setp.eq.ftz.f32 %p82, %f49, 0f00000000;
|
|
@%p82 bra $L__BB9_59;
|
|
|
|
sub.ftz.f32 %f178, %f16, %f46;
|
|
sub.ftz.f32 %f179, %f17, %f47;
|
|
mul.ftz.f32 %f180, %f179, %f179;
|
|
fma.rn.ftz.f32 %f181, %f178, %f178, %f180;
|
|
sub.ftz.f32 %f182, %f18, %f48;
|
|
fma.rn.ftz.f32 %f183, %f182, %f182, %f181;
|
|
sqrt.approx.ftz.f32 %f184, %f183;
|
|
div.approx.ftz.f32 %f185, %f178, %f184;
|
|
div.approx.ftz.f32 %f186, %f179, %f184;
|
|
div.approx.ftz.f32 %f187, %f182, %f184;
|
|
mul.ftz.f32 %f188, %f186, %f179;
|
|
fma.rn.ftz.f32 %f189, %f185, %f178, %f188;
|
|
fma.rn.ftz.f32 %f50, %f187, %f182, %f189;
|
|
mul.ftz.f32 %f190, %f185, %f19;
|
|
mul.ftz.f32 %f191, %f186, %f20;
|
|
neg.ftz.f32 %f192, %f191;
|
|
sub.ftz.f32 %f193, %f192, %f190;
|
|
mul.ftz.f32 %f194, %f187, %f21;
|
|
sub.ftz.f32 %f51, %f193, %f194;
|
|
mul.ftz.f32 %f195, %f50, %f51;
|
|
setp.leu.ftz.f32 %p84, %f195, 0f00000000;
|
|
mov.f32 %f334, 0f00000000;
|
|
mov.pred %p269, %p83;
|
|
@%p84 bra $L__BB9_49;
|
|
|
|
mov.f32 %f334, 0f00000000;
|
|
div.approx.ftz.f32 %f52, %f50, %f51;
|
|
setp.lt.ftz.f32 %p86, %f52, 0f00000000;
|
|
mov.pred %p269, %p83;
|
|
@%p86 bra $L__BB9_49;
|
|
|
|
mov.f32 %f334, 0f00000000;
|
|
fma.rn.ftz.f32 %f198, %f19, %f52, %f16;
|
|
fma.rn.ftz.f32 %f199, %f20, %f52, %f17;
|
|
fma.rn.ftz.f32 %f200, %f21, %f52, %f18;
|
|
sub.ftz.f32 %f201, %f198, %f46;
|
|
sub.ftz.f32 %f202, %f199, %f47;
|
|
sub.ftz.f32 %f203, %f200, %f48;
|
|
mul.ftz.f32 %f204, %f202, %f202;
|
|
fma.rn.ftz.f32 %f205, %f201, %f201, %f204;
|
|
fma.rn.ftz.f32 %f206, %f203, %f203, %f205;
|
|
mul.ftz.f32 %f207, %f49, %f49;
|
|
setp.geu.ftz.f32 %p88, %f206, %f207;
|
|
mov.pred %p269, %p83;
|
|
@%p88 bra $L__BB9_49;
|
|
|
|
setp.gt.ftz.f32 %p89, %f52, %f22;
|
|
setp.lt.ftz.f32 %p90, %f52, %f23;
|
|
and.pred %p91, %p89, %p90;
|
|
selp.f32 %f334, %f52, 0f00000000, %p91;
|
|
not.pred %p269, %p91;
|
|
|
|
$L__BB9_49:
|
|
mov.f32 %f332, 0f00000000;
|
|
mov.f32 %f333, %f332;
|
|
@%p269 bra $L__BB9_59;
|
|
bra.uni $L__BB9_57;
|
|
|
|
$L__BB9_50:
|
|
ld.global.f32 %f55, [%rd799+-140];
|
|
ld.global.f32 %f56, [%rd799+-136];
|
|
ld.global.f32 %f57, [%rd799+-132];
|
|
ld.global.f32 %f58, [%rd799];
|
|
setp.eq.ftz.f32 %p92, %f58, 0f00000000;
|
|
@%p92 bra $L__BB9_59;
|
|
|
|
sub.ftz.f32 %f59, %f16, %f55;
|
|
sub.ftz.f32 %f60, %f17, %f56;
|
|
mul.ftz.f32 %f210, %f60, %f60;
|
|
fma.rn.ftz.f32 %f211, %f59, %f59, %f210;
|
|
sub.ftz.f32 %f61, %f18, %f57;
|
|
fma.rn.ftz.f32 %f212, %f61, %f61, %f211;
|
|
sqrt.approx.ftz.f32 %f213, %f212;
|
|
div.approx.ftz.f32 %f62, %f59, %f213;
|
|
div.approx.ftz.f32 %f63, %f60, %f213;
|
|
div.approx.ftz.f32 %f64, %f61, %f213;
|
|
mul.ftz.f32 %f65, %f62, %f19;
|
|
mul.ftz.f32 %f66, %f63, %f20;
|
|
add.ftz.f32 %f214, %f65, %f66;
|
|
mul.ftz.f32 %f67, %f64, %f21;
|
|
add.ftz.f32 %f215, %f214, %f67;
|
|
setp.ge.ftz.f32 %p93, %f215, 0f00000000;
|
|
@%p93 bra $L__BB9_59;
|
|
|
|
mul.ftz.f32 %f217, %f63, %f60;
|
|
fma.rn.ftz.f32 %f218, %f62, %f59, %f217;
|
|
fma.rn.ftz.f32 %f68, %f64, %f61, %f218;
|
|
neg.ftz.f32 %f219, %f66;
|
|
sub.ftz.f32 %f220, %f219, %f65;
|
|
sub.ftz.f32 %f69, %f220, %f67;
|
|
mul.ftz.f32 %f221, %f68, %f69;
|
|
setp.leu.ftz.f32 %p95, %f221, 0f00000000;
|
|
mov.pred %p270, -1;
|
|
mov.f32 %f332, 0f00000000;
|
|
mov.f32 %f334, %f332;
|
|
@%p95 bra $L__BB9_56;
|
|
|
|
div.approx.ftz.f32 %f70, %f68, %f69;
|
|
setp.lt.ftz.f32 %p97, %f70, 0f00000000;
|
|
mov.f32 %f334, %f332;
|
|
@%p97 bra $L__BB9_56;
|
|
|
|
fma.rn.ftz.f32 %f224, %f19, %f70, %f16;
|
|
fma.rn.ftz.f32 %f225, %f20, %f70, %f17;
|
|
fma.rn.ftz.f32 %f226, %f21, %f70, %f18;
|
|
sub.ftz.f32 %f227, %f224, %f55;
|
|
sub.ftz.f32 %f228, %f225, %f56;
|
|
sub.ftz.f32 %f229, %f226, %f57;
|
|
mul.ftz.f32 %f230, %f228, %f228;
|
|
fma.rn.ftz.f32 %f231, %f227, %f227, %f230;
|
|
fma.rn.ftz.f32 %f232, %f229, %f229, %f231;
|
|
mul.ftz.f32 %f233, %f58, %f58;
|
|
setp.geu.ftz.f32 %p99, %f232, %f233;
|
|
mov.f32 %f334, %f332;
|
|
@%p99 bra $L__BB9_56;
|
|
|
|
setp.gt.ftz.f32 %p100, %f70, %f22;
|
|
setp.lt.ftz.f32 %p101, %f70, %f23;
|
|
and.pred %p102, %p100, %p101;
|
|
selp.f32 %f334, %f70, 0f00000000, %p102;
|
|
not.pred %p270, %p102;
|
|
|
|
$L__BB9_56:
|
|
mov.f32 %f333, %f332;
|
|
@%p270 bra $L__BB9_59;
|
|
|
|
$L__BB9_57:
|
|
setp.geu.ftz.f32 %p103, %f334, %f342;
|
|
@%p103 bra $L__BB9_59;
|
|
|
|
setp.eq.s32 %p105, %r579, 0;
|
|
setp.eq.s32 %p106, %r6, -1;
|
|
and.pred %p107, %p106, %p105;
|
|
and.pred %p108, %p104, %p107;
|
|
selp.b32 %r584, %r584, 32, %p108;
|
|
selp.b32 %r585, %r585, -1, %p108;
|
|
selp.b32 %r71, %r71, %r580, %p108;
|
|
selp.f32 %f340, %f340, %f332, %p108;
|
|
selp.f32 %f341, %f341, %f333, %p108;
|
|
selp.f32 %f342, %f342, %f334, %p108;
|
|
|
|
$L__BB9_59:
|
|
add.s64 %rd799, %rd799, 192;
|
|
add.s32 %r579, %r579, 1;
|
|
add.s32 %r580, %r580, 1;
|
|
setp.lt.s32 %p109, %r580, %r47;
|
|
@%p109 bra $L__BB9_27;
|
|
|
|
$L__BB9_60:
|
|
setp.ne.s32 %p110, %r71, -1;
|
|
setp.ne.s32 %p111, %r69, 0;
|
|
or.pred %p271, %p111, %p110;
|
|
mov.u32 %r69, %r584;
|
|
|
|
$L__BB9_61:
|
|
ld.const.u64 %rd798, [kernel_params+2744];
|
|
ld.const.v2.u64 {%rd184, %rd185}, [kernel_params+2704];
|
|
cvta.to.global.u64 %rd45, %rd184;
|
|
shl.b64 %rd186, %rd6, 2;
|
|
add.s64 %rd46, %rd45, %rd186;
|
|
st.global.f32 [%rd46], %f342;
|
|
cvta.to.global.u64 %rd47, %rd185;
|
|
add.s64 %rd48, %rd47, %rd186;
|
|
st.global.f32 [%rd48], %f341;
|
|
ld.const.u64 %rd49, [kernel_params+2720];
|
|
cvta.to.global.u64 %rd50, %rd49;
|
|
add.s64 %rd51, %rd50, %rd186;
|
|
st.global.f32 [%rd51], %f340;
|
|
st.global.u32 [%rd34], %r585;
|
|
st.global.u32 [%rd31], %r71;
|
|
cvta.to.global.u64 %rd52, %rd798;
|
|
add.s64 %rd53, %rd52, %rd186;
|
|
st.global.u32 [%rd53], %r69;
|
|
ld.const.u32 %r258, [kernel_params+16];
|
|
and.b32 %r72, %r258, 131072;
|
|
setp.eq.s32 %p112, %r72, 0;
|
|
mov.u16 %rs83, 1;
|
|
@%p112 bra $L__BB9_63;
|
|
|
|
ld.const.u64 %rd187, [kernel_params+2792];
|
|
cvta.to.global.u64 %rd188, %rd187;
|
|
add.s64 %rd190, %rd188, %rd186;
|
|
ld.global.u32 %r259, [%rd190];
|
|
setp.eq.s32 %p113, %r259, -1;
|
|
selp.u16 %rs83, 1, 0, %p113;
|
|
|
|
$L__BB9_63:
|
|
setp.eq.s16 %p114, %rs83, 0;
|
|
@%p114 bra $L__BB9_248;
|
|
|
|
@%p271 bra $L__BB9_66;
|
|
bra.uni $L__BB9_65;
|
|
|
|
$L__BB9_66:
|
|
and.b32 %r262, %r69, 32;
|
|
setp.eq.s32 %p115, %r262, 0;
|
|
@%p115 bra $L__BB9_68;
|
|
bra.uni $L__BB9_67;
|
|
|
|
$L__BB9_68:
|
|
and.b32 %r73, %r69, 1;
|
|
setp.eq.s32 %p116, %r73, 0;
|
|
@%p116 bra $L__BB9_70;
|
|
bra.uni $L__BB9_69;
|
|
|
|
$L__BB9_70:
|
|
and.b32 %r265, %r69, 8;
|
|
setp.eq.s32 %p117, %r265, 0;
|
|
@%p117 bra $L__BB9_72;
|
|
bra.uni $L__BB9_71;
|
|
|
|
$L__BB9_72:
|
|
and.b32 %r267, %r69, 6;
|
|
setp.eq.s32 %p118, %r267, 0;
|
|
mov.u32 %r593, 0;
|
|
@%p118 bra $L__BB9_74;
|
|
|
|
ld.const.u64 %rd215, [kernel_params+2280];
|
|
cvta.to.global.u64 %rd216, %rd215;
|
|
mul.wide.s32 %rd217, %r71, 16;
|
|
add.s64 %rd218, %rd216, %rd217;
|
|
ld.global.u32 %r593, [%rd218];
|
|
bra.uni $L__BB9_74;
|
|
|
|
$L__BB9_248:
|
|
and.b32 %r446, %r69, 32;
|
|
setp.eq.s32 %p225, %r446, 0;
|
|
and.pred %p11, %p225, %p271;
|
|
mov.u64 %rd804, -1;
|
|
not.pred %p226, %p11;
|
|
@%p226 bra $L__BB9_256;
|
|
|
|
and.b32 %r447, %r69, 1;
|
|
setp.eq.b32 %p227, %r447, 1;
|
|
mov.pred %p228, 0;
|
|
xor.pred %p229, %p227, %p228;
|
|
not.pred %p230, %p229;
|
|
@%p230 bra $L__BB9_251;
|
|
bra.uni $L__BB9_250;
|
|
|
|
$L__BB9_251:
|
|
and.b32 %r448, %r69, 8;
|
|
setp.eq.s32 %p231, %r448, 0;
|
|
@%p231 bra $L__BB9_253;
|
|
bra.uni $L__BB9_252;
|
|
|
|
$L__BB9_253:
|
|
and.b32 %r450, %r69, 6;
|
|
setp.eq.s32 %p232, %r450, 0;
|
|
mov.u32 %r602, 0;
|
|
@%p232 bra $L__BB9_255;
|
|
|
|
ld.const.u64 %rd633, [kernel_params+2280];
|
|
cvta.to.global.u64 %rd634, %rd633;
|
|
mul.wide.s32 %rd635, %r71, 16;
|
|
add.s64 %rd636, %rd634, %rd635;
|
|
ld.global.u32 %r602, [%rd636];
|
|
bra.uni $L__BB9_255;
|
|
|
|
$L__BB9_65:
|
|
ld.const.u64 %rd191, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd192, %rd191;
|
|
add.s64 %rd193, %rd192, 8;
|
|
atom.global.add.u32 %r260, [%rd193], -1;
|
|
add.s64 %rd194, %rd192, 24;
|
|
atom.global.add.u32 %r261, [%rd194], 1;
|
|
ld.const.u64 %rd195, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd196, %rd195;
|
|
shl.b64 %rd197, %rd6, 1;
|
|
add.s64 %rd198, %rd196, %rd197;
|
|
mov.u16 %rs39, 6;
|
|
st.global.u16 [%rd198], %rs39;
|
|
bra.uni $L__BB9_283;
|
|
|
|
$L__BB9_67:
|
|
ld.const.u64 %rd199, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd200, %rd199;
|
|
add.s64 %rd201, %rd200, 8;
|
|
atom.global.add.u32 %r263, [%rd201], -1;
|
|
add.s64 %rd202, %rd200, 28;
|
|
atom.global.add.u32 %r264, [%rd202], 1;
|
|
ld.const.u64 %rd203, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd204, %rd203;
|
|
shl.b64 %rd205, %rd6, 1;
|
|
add.s64 %rd206, %rd204, %rd205;
|
|
mov.u16 %rs40, 7;
|
|
st.global.u16 [%rd206], %rs40;
|
|
bra.uni $L__BB9_283;
|
|
|
|
$L__BB9_250:
|
|
ld.const.u64 %rd625, [kernel_params+2232];
|
|
cvta.to.global.u64 %rd626, %rd625;
|
|
mul.wide.s32 %rd627, %r71, 4;
|
|
add.s64 %rd628, %rd626, %rd627;
|
|
ld.global.u32 %r602, [%rd628];
|
|
bra.uni $L__BB9_255;
|
|
|
|
$L__BB9_69:
|
|
ld.const.u64 %rd207, [kernel_params+2232];
|
|
cvta.to.global.u64 %rd208, %rd207;
|
|
mul.wide.s32 %rd209, %r71, 4;
|
|
add.s64 %rd210, %rd208, %rd209;
|
|
ld.global.u32 %r593, [%rd210];
|
|
bra.uni $L__BB9_74;
|
|
|
|
$L__BB9_252:
|
|
ld.const.u64 %rd629, [kernel_params+2320];
|
|
cvta.to.global.u64 %rd630, %rd629;
|
|
mul.wide.s32 %rd631, %r71, 4;
|
|
add.s64 %rd632, %rd630, %rd631;
|
|
ld.global.u32 %r602, [%rd632];
|
|
|
|
$L__BB9_255:
|
|
and.b32 %r451, %r602, 4194303;
|
|
cvt.u64.u32 %rd804, %r451;
|
|
|
|
$L__BB9_256:
|
|
mov.u32 %r603, 0;
|
|
@%p226 bra $L__BB9_258;
|
|
|
|
ld.const.u64 %rd637, [kernel_params+2424];
|
|
cvta.to.global.u64 %rd638, %rd637;
|
|
shl.b64 %rd639, %rd804, 5;
|
|
add.s64 %rd640, %rd638, %rd639;
|
|
ld.global.u32 %r603, [%rd640+16];
|
|
|
|
$L__BB9_258:
|
|
ld.const.u32 %r577, [kernel_params+1552];
|
|
setp.eq.s32 %p264, %r577, 0;
|
|
@%p264 bra $L__BB9_265;
|
|
|
|
ld.const.u32 %r572, [kernel_params+1552];
|
|
ld.const.v2.u64 {%rd641, %rd642}, [kernel_params+2480];
|
|
cvta.to.global.u64 %rd644, %rd641;
|
|
shl.b64 %rd645, %rd6, 1;
|
|
add.s64 %rd646, %rd644, %rd645;
|
|
ld.global.u16 %r453, [%rd646];
|
|
ld.const.u64 %rd647, [kernel_params+2512];
|
|
cvta.to.global.u64 %rd648, %rd647;
|
|
add.s64 %rd649, %rd648, %rd645;
|
|
ld.global.u16 %r454, [%rd649];
|
|
sub.s32 %r455, %r453, %r454;
|
|
ld.const.v2.u64 {%rd650, %rd651}, [kernel_params+2496];
|
|
cvta.to.global.u64 %rd653, %rd651;
|
|
add.s64 %rd654, %rd653, %rd645;
|
|
ld.global.u16 %rs72, [%rd654];
|
|
setp.ne.s16 %p235, %rs72, 0;
|
|
selp.b32 %r456, -1, 0, %p235;
|
|
add.s32 %r457, %r455, %r456;
|
|
setp.lt.s32 %p236, %r457, %r572;
|
|
@%p236 bra $L__BB9_265;
|
|
|
|
and.b32 %r458, %r603, 537001984;
|
|
setp.eq.s32 %p237, %r458, 0;
|
|
@%p237 bra $L__BB9_262;
|
|
|
|
ld.global.u32 %r459, [%rd28];
|
|
or.b32 %r460, %r459, 131072;
|
|
st.global.u32 [%rd28], %r460;
|
|
|
|
$L__BB9_265:
|
|
ld.const.v2.u64 {%rd659, %rd660}, [kernel_params+2544];
|
|
cvta.to.global.u64 %rd663, %rd659;
|
|
add.s64 %rd665, %rd663, %rd186;
|
|
ld.global.u32 %r121, [%rd665];
|
|
cvta.to.global.u64 %rd666, %rd660;
|
|
shl.b64 %rd667, %rd6, 1;
|
|
add.s64 %rd668, %rd666, %rd667;
|
|
ld.global.u16 %r122, [%rd668];
|
|
ld.const.u64 %rd669, [kernel_params+2472];
|
|
cvta.to.global.u64 %rd670, %rd669;
|
|
add.s64 %rd671, %rd670, %rd186;
|
|
ld.global.u32 %r123, [%rd671];
|
|
ld.global.u8 %rs73, [%rd28];
|
|
and.b16 %rs74, %rs73, 64;
|
|
setp.eq.s16 %p242, %rs74, 0;
|
|
@%p242 bra $L__BB9_267;
|
|
bra.uni $L__BB9_266;
|
|
|
|
$L__BB9_267:
|
|
ld.const.u64 %rd676, [kernel_params+2480];
|
|
cvta.to.global.u64 %rd677, %rd676;
|
|
add.s64 %rd679, %rd677, %rd667;
|
|
ld.global.u16 %r466, [%rd679];
|
|
ld.const.u32 %r467, [kernel_params+1528];
|
|
setp.ge.u32 %p244, %r467, %r466;
|
|
mov.f32 %f351, 0f3F800000;
|
|
@%p244 bra $L__BB9_269;
|
|
bra.uni $L__BB9_268;
|
|
|
|
$L__BB9_266:
|
|
ld.const.u64 %rd672, [kernel_params+2488];
|
|
cvta.to.global.u64 %rd673, %rd672;
|
|
add.s64 %rd675, %rd673, %rd667;
|
|
ld.global.u16 %r464, [%rd675];
|
|
ld.const.u32 %r465, [kernel_params+1568];
|
|
setp.lt.u32 %p243, %r465, %r464;
|
|
mov.f32 %f351, 0f3F800000;
|
|
@%p243 bra $L__BB9_268;
|
|
bra.uni $L__BB9_269;
|
|
|
|
$L__BB9_268:
|
|
ld.const.u64 %rd680, [kernel_params+2600];
|
|
cvta.to.global.u64 %rd681, %rd680;
|
|
mul.lo.s64 %rd682, %rd6, 12;
|
|
add.s64 %rd683, %rd681, %rd682;
|
|
ld.global.f32 %f308, [%rd683];
|
|
abs.ftz.f32 %f309, %f308;
|
|
ld.global.f32 %f310, [%rd683+4];
|
|
abs.ftz.f32 %f311, %f310;
|
|
ld.global.f32 %f312, [%rd683+8];
|
|
abs.ftz.f32 %f313, %f312;
|
|
max.ftz.f32 %f314, %f309, %f311;
|
|
max.ftz.f32 %f315, %f314, %f313;
|
|
sqrt.approx.ftz.f32 %f316, %f315;
|
|
mov.f32 %f317, 0f3F800000;
|
|
min.ftz.f32 %f351, %f316, %f317;
|
|
|
|
$L__BB9_269:
|
|
ld.const.u64 %rd684, [kernel_params+2592];
|
|
cvta.to.global.u64 %rd685, %rd684;
|
|
add.s64 %rd687, %rd685, %rd186;
|
|
st.global.f32 [%rd687], %f351;
|
|
setp.eq.ftz.f32 %p246, %f351, 0f3F800000;
|
|
mov.pred %p245, -1;
|
|
mov.pred %p272, %p245;
|
|
@%p246 bra $L__BB9_280;
|
|
|
|
ld.const.u32 %r468, [kernel_params+1612];
|
|
setp.eq.s32 %p247, %r468, 0;
|
|
brev.b32 %r469, %r123;
|
|
mul.lo.s32 %r470, %r469, 1025551850;
|
|
xor.b32 %r124, %r470, %r469;
|
|
@%p247 bra $L__BB9_273;
|
|
|
|
ld.const.v2.u32 {%r471, %r472}, [kernel_params+1616];
|
|
mov.b32 %f97, %r472;
|
|
setp.geu.ftz.f32 %p248, %f97, 0f3F800000;
|
|
setp.lt.ftz.f32 %p249, %f97, 0f3F800000;
|
|
ld.const.u32 %r475, [kernel_params+1592];
|
|
selp.b32 %r476, %r475, %r121, %p249;
|
|
and.b32 %r477, %r122, 255;
|
|
xor.b32 %r478, %r476, %r477;
|
|
mul.lo.s32 %r479, %r478, -512718531;
|
|
shr.u32 %r480, %r476, 16;
|
|
xor.b32 %r481, %r479, %r480;
|
|
bfe.u32 %r482, %r481, 4, 4;
|
|
shr.u32 %r483, %r476, 8;
|
|
xor.b32 %r484, %r481, %r483;
|
|
xor.b32 %r485, %r484, %r482;
|
|
mul.lo.s32 %r486, %r485, 153742143;
|
|
shr.u32 %r487, %r476, 23;
|
|
xor.b32 %r488, %r486, %r487;
|
|
bfe.u32 %r489, %r488, 1, 7;
|
|
xor.b32 %r490, %r489, %r488;
|
|
shr.u32 %r491, %r476, 27;
|
|
or.b32 %r492, %r491, 1;
|
|
mul.lo.s32 %r493, %r492, %r490;
|
|
mul.lo.s32 %r494, %r493, -1074439621;
|
|
bfe.u32 %r495, %r494, 2, 6;
|
|
xor.b32 %r496, %r495, %r494;
|
|
mul.lo.s32 %r497, %r496, -1638916925;
|
|
bfe.u32 %r498, %r497, 2, 6;
|
|
xor.b32 %r499, %r498, %r497;
|
|
mul.lo.s32 %r500, %r499, -933190689;
|
|
and.b32 %r501, %r500, 255;
|
|
bfe.u32 %r502, %r500, 5, 3;
|
|
xor.b32 %r503, %r502, %r501;
|
|
add.s32 %r504, %r471, -1;
|
|
xor.b32 %r125, %r122, 61;
|
|
xor.b32 %r505, %r476, %r125;
|
|
mul.lo.s32 %r506, %r505, 9;
|
|
shr.u32 %r507, %r506, 4;
|
|
xor.b32 %r508, %r507, %r506;
|
|
mul.lo.s32 %r509, %r508, 668265261;
|
|
add.s32 %r510, %r509, %r124;
|
|
shr.u32 %r511, %r509, 16;
|
|
or.b32 %r512, %r511, 1;
|
|
mul.lo.s32 %r513, %r512, %r510;
|
|
mul.lo.s32 %r514, %r513, 89287766;
|
|
xor.b32 %r515, %r514, %r513;
|
|
mul.lo.s32 %r516, %r515, 1403136100;
|
|
xor.b32 %r517, %r516, %r515;
|
|
brev.b32 %r518, %r517;
|
|
neg.s32 %r519, %r471;
|
|
and.b32 %r520, %r123, %r519;
|
|
and.b32 %r521, %r518, %r504;
|
|
or.b32 %r522, %r521, %r520;
|
|
mad.lo.s32 %r523, %r503, %r471, %r522;
|
|
shl.b32 %r524, %r471, 8;
|
|
rem.u32 %r525, %r523, %r524;
|
|
ld.const.u64 %rd688, [kernel_params+2440];
|
|
cvta.to.global.u64 %rd689, %rd688;
|
|
shl.b32 %r526, %r525, 1;
|
|
mul.wide.u32 %rd690, %r526, 4;
|
|
add.s64 %rd691, %rd689, %rd690;
|
|
ld.global.f32 %f352, [%rd691];
|
|
@%p248 bra $L__BB9_274;
|
|
|
|
xor.b32 %r527, %r125, %r121;
|
|
mul.lo.s32 %r528, %r527, 9;
|
|
shr.u32 %r529, %r528, 4;
|
|
xor.b32 %r530, %r529, %r528;
|
|
mul.lo.s32 %r531, %r530, 668265261;
|
|
cvt.rn.f32.u32 %f318, %r531;
|
|
mul.ftz.f32 %f319, %f318, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f320, %f319, %f97, %f352;
|
|
cvt.rmi.ftz.f32.f32 %f321, %f320;
|
|
sub.ftz.f32 %f352, %f320, %f321;
|
|
bra.uni $L__BB9_274;
|
|
|
|
$L__BB9_273:
|
|
mul.lo.s32 %r532, %r122, 569420461;
|
|
shr.u32 %r533, %r532, 15;
|
|
xor.b32 %r534, %r533, %r532;
|
|
mul.lo.s32 %r535, %r534, -749064809;
|
|
shr.u32 %r536, %r535, 15;
|
|
xor.b32 %r537, %r121, %r535;
|
|
xor.b32 %r538, %r537, %r536;
|
|
xor.b32 %r539, %r538, 1493655573;
|
|
add.s32 %r540, %r539, %r124;
|
|
shr.u32 %r541, %r539, 16;
|
|
or.b32 %r542, %r541, 1;
|
|
mul.lo.s32 %r543, %r542, %r540;
|
|
mul.lo.s32 %r544, %r543, 89287766;
|
|
xor.b32 %r545, %r544, %r543;
|
|
mul.lo.s32 %r546, %r545, 1403136100;
|
|
xor.b32 %r547, %r546, %r545;
|
|
xor.b32 %r548, %r538, -2052961194;
|
|
brev.b32 %r549, %r547;
|
|
mul.lo.s32 %r550, %r549, 1025551850;
|
|
xor.b32 %r551, %r550, %r549;
|
|
add.s32 %r552, %r551, %r548;
|
|
shr.u32 %r553, %r548, 16;
|
|
or.b32 %r554, %r553, 1;
|
|
mul.lo.s32 %r555, %r552, %r554;
|
|
mul.lo.s32 %r556, %r555, 89287766;
|
|
xor.b32 %r557, %r556, %r555;
|
|
mul.lo.s32 %r558, %r557, 1403136100;
|
|
xor.b32 %r559, %r558, %r557;
|
|
brev.b32 %r560, %r559;
|
|
cvt.rn.f32.u32 %f322, %r560;
|
|
mul.ftz.f32 %f352, %f322, 0f2F7FFFFE;
|
|
|
|
$L__BB9_274:
|
|
setp.ltu.ftz.f32 %p251, %f352, %f351;
|
|
setp.neu.ftz.f32 %p252, %f351, 0f00000000;
|
|
and.pred %p253, %p252, %p251;
|
|
mov.pred %p272, %p245;
|
|
@%p253 bra $L__BB9_280;
|
|
|
|
and.b32 %r561, %r603, 536870912;
|
|
setp.eq.s32 %p254, %r561, 0;
|
|
@%p254 bra $L__BB9_277;
|
|
bra.uni $L__BB9_276;
|
|
|
|
$L__BB9_277:
|
|
mov.pred %p272, 0;
|
|
@%p112 bra $L__BB9_280;
|
|
|
|
ld.const.u64 %rd692, [kernel_params+2792];
|
|
cvta.to.global.u64 %rd693, %rd692;
|
|
add.s64 %rd695, %rd693, %rd186;
|
|
ld.global.u32 %r564, [%rd695];
|
|
setp.eq.s32 %p259, %r564, -1;
|
|
@%p259 bra $L__BB9_280;
|
|
|
|
ld.global.u32 %r565, [%rd28];
|
|
or.b32 %r566, %r565, 65536;
|
|
st.global.u32 [%rd28], %r566;
|
|
mov.pred %p272, %p245;
|
|
bra.uni $L__BB9_280;
|
|
|
|
$L__BB9_276:
|
|
ld.global.u32 %r562, [%rd28];
|
|
or.b32 %r563, %r562, 32768;
|
|
st.global.u32 [%rd28], %r563;
|
|
mov.pred %p272, %p245;
|
|
bra.uni $L__BB9_280;
|
|
|
|
$L__BB9_262:
|
|
mov.pred %p272, 0;
|
|
@%p112 bra $L__BB9_280;
|
|
|
|
ld.const.u64 %rd655, [kernel_params+2792];
|
|
cvta.to.global.u64 %rd656, %rd655;
|
|
add.s64 %rd658, %rd656, %rd186;
|
|
ld.global.u32 %r461, [%rd658];
|
|
setp.eq.s32 %p241, %r461, -1;
|
|
@%p241 bra $L__BB9_280;
|
|
|
|
ld.global.u32 %r462, [%rd28];
|
|
or.b32 %r463, %r462, 262144;
|
|
st.global.u32 [%rd28], %r463;
|
|
bra.uni $L__BB9_265;
|
|
|
|
$L__BB9_280:
|
|
ld.const.u64 %rd696, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd135, %rd696;
|
|
add.s64 %rd697, %rd135, 8;
|
|
atom.global.add.u32 %r567, [%rd697], -1;
|
|
@%p272 bra $L__BB9_282;
|
|
bra.uni $L__BB9_281;
|
|
|
|
$L__BB9_282:
|
|
add.s64 %rd702, %rd135, 44;
|
|
atom.global.add.u32 %r568, [%rd702], 1;
|
|
ld.const.u64 %rd703, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd704, %rd703;
|
|
shl.b64 %rd705, %rd6, 1;
|
|
add.s64 %rd706, %rd704, %rd705;
|
|
mov.u16 %rs76, 11;
|
|
st.global.u16 [%rd706], %rs76;
|
|
bra.uni $L__BB9_283;
|
|
|
|
$L__BB9_281:
|
|
ld.const.u64 %rd698, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd699, %rd698;
|
|
shl.b64 %rd700, %rd6, 1;
|
|
add.s64 %rd701, %rd699, %rd700;
|
|
mov.u16 %rs75, 0;
|
|
st.global.u16 [%rd701], %rs75;
|
|
|
|
$L__BB9_283:
|
|
ret;
|
|
|
|
$L__BB9_71:
|
|
ld.const.u64 %rd211, [kernel_params+2320];
|
|
cvta.to.global.u64 %rd212, %rd211;
|
|
mul.wide.s32 %rd213, %r71, 4;
|
|
add.s64 %rd214, %rd212, %rd213;
|
|
ld.global.u32 %r593, [%rd214];
|
|
|
|
$L__BB9_74:
|
|
ld.const.u32 %r576, [kernel_params+1552];
|
|
setp.eq.s32 %p263, %r576, 0;
|
|
ld.const.u64 %rd219, [kernel_params+2424];
|
|
cvta.to.global.u64 %rd54, %rd219;
|
|
and.b32 %r78, %r593, 4194303;
|
|
shl.b32 %r268, %r593, 5;
|
|
cvt.u64.u32 %rd220, %r268;
|
|
and.b64 %rd221, %rd220, 134217696;
|
|
add.s64 %rd222, %rd54, %rd221;
|
|
ld.global.u32 %r79, [%rd222+16];
|
|
mov.u16 %rs84, 0;
|
|
@%p263 bra $L__BB9_76;
|
|
|
|
ld.const.u32 %r570, [kernel_params+1552];
|
|
ld.const.v2.u64 {%rd223, %rd224}, [kernel_params+2480];
|
|
cvta.to.global.u64 %rd226, %rd223;
|
|
shl.b64 %rd227, %rd6, 1;
|
|
add.s64 %rd228, %rd226, %rd227;
|
|
ld.global.u16 %r269, [%rd228];
|
|
ld.const.u64 %rd229, [kernel_params+2512];
|
|
cvta.to.global.u64 %rd230, %rd229;
|
|
add.s64 %rd231, %rd230, %rd227;
|
|
ld.global.u16 %r270, [%rd231];
|
|
sub.s32 %r271, %r269, %r270;
|
|
ld.const.v2.u64 {%rd232, %rd233}, [kernel_params+2496];
|
|
cvta.to.global.u64 %rd235, %rd233;
|
|
add.s64 %rd236, %rd235, %rd227;
|
|
ld.global.u16 %rs42, [%rd236];
|
|
setp.ne.s16 %p120, %rs42, 0;
|
|
selp.b32 %r272, -1, 0, %p120;
|
|
add.s32 %r273, %r271, %r272;
|
|
setp.ge.s32 %p121, %r273, %r570;
|
|
selp.u16 %rs84, 1, 0, %p121;
|
|
|
|
$L__BB9_76:
|
|
setp.eq.s16 %p122, %rs84, 0;
|
|
@%p122 bra $L__BB9_83;
|
|
|
|
and.b32 %r274, %r79, 537001984;
|
|
setp.eq.s32 %p123, %r274, 0;
|
|
@%p123 bra $L__BB9_79;
|
|
|
|
ld.global.u32 %r275, [%rd28];
|
|
or.b32 %r594, %r275, 131072;
|
|
st.global.u32 [%rd28], %r594;
|
|
bra.uni $L__BB9_84;
|
|
|
|
$L__BB9_83:
|
|
ld.global.u32 %r594, [%rd28];
|
|
|
|
$L__BB9_84:
|
|
ld.const.v2.u64 {%rd241, %rd242}, [kernel_params+2544];
|
|
cvta.to.global.u64 %rd57, %rd241;
|
|
add.s64 %rd58, %rd57, %rd186;
|
|
ld.global.u32 %r84, [%rd58];
|
|
cvta.to.global.u64 %rd59, %rd242;
|
|
shl.b64 %rd244, %rd6, 1;
|
|
add.s64 %rd60, %rd59, %rd244;
|
|
ld.global.u16 %r85, [%rd60];
|
|
ld.const.u64 %rd61, [kernel_params+2472];
|
|
cvta.to.global.u64 %rd62, %rd61;
|
|
add.s64 %rd63, %rd62, %rd186;
|
|
ld.global.u32 %r86, [%rd63];
|
|
and.b32 %r278, %r594, 64;
|
|
setp.eq.s32 %p127, %r278, 0;
|
|
@%p127 bra $L__BB9_86;
|
|
bra.uni $L__BB9_85;
|
|
|
|
$L__BB9_86:
|
|
ld.const.u64 %rd249, [kernel_params+2480];
|
|
cvta.to.global.u64 %rd250, %rd249;
|
|
add.s64 %rd252, %rd250, %rd244;
|
|
ld.global.u16 %r281, [%rd252];
|
|
ld.const.u32 %r282, [kernel_params+1528];
|
|
setp.ge.u32 %p129, %r282, %r281;
|
|
mov.f32 %f349, 0f3F800000;
|
|
@%p129 bra $L__BB9_88;
|
|
bra.uni $L__BB9_87;
|
|
|
|
$L__BB9_85:
|
|
ld.const.u64 %rd245, [kernel_params+2488];
|
|
cvta.to.global.u64 %rd246, %rd245;
|
|
add.s64 %rd248, %rd246, %rd244;
|
|
ld.global.u16 %r279, [%rd248];
|
|
ld.const.u32 %r280, [kernel_params+1568];
|
|
setp.lt.u32 %p128, %r280, %r279;
|
|
mov.f32 %f349, 0f3F800000;
|
|
@%p128 bra $L__BB9_87;
|
|
bra.uni $L__BB9_88;
|
|
|
|
$L__BB9_87:
|
|
ld.const.u64 %rd253, [kernel_params+2600];
|
|
cvta.to.global.u64 %rd254, %rd253;
|
|
mul.lo.s64 %rd255, %rd6, 12;
|
|
add.s64 %rd256, %rd254, %rd255;
|
|
ld.global.f32 %f238, [%rd256];
|
|
abs.ftz.f32 %f239, %f238;
|
|
ld.global.f32 %f240, [%rd256+4];
|
|
abs.ftz.f32 %f241, %f240;
|
|
ld.global.f32 %f242, [%rd256+8];
|
|
abs.ftz.f32 %f243, %f242;
|
|
max.ftz.f32 %f244, %f239, %f241;
|
|
max.ftz.f32 %f245, %f244, %f243;
|
|
sqrt.approx.ftz.f32 %f246, %f245;
|
|
mov.f32 %f247, 0f3F800000;
|
|
min.ftz.f32 %f349, %f246, %f247;
|
|
|
|
$L__BB9_88:
|
|
ld.const.u64 %rd64, [kernel_params+2592];
|
|
cvta.to.global.u64 %rd65, %rd64;
|
|
add.s64 %rd66, %rd65, %rd186;
|
|
st.global.f32 [%rd66], %f349;
|
|
setp.eq.ftz.f32 %p130, %f349, 0f3F800000;
|
|
@%p130 bra $L__BB9_101;
|
|
|
|
ld.const.u32 %r283, [kernel_params+1612];
|
|
setp.eq.s32 %p131, %r283, 0;
|
|
brev.b32 %r284, %r86;
|
|
mul.lo.s32 %r285, %r284, 1025551850;
|
|
xor.b32 %r87, %r285, %r284;
|
|
@%p131 bra $L__BB9_92;
|
|
|
|
ld.const.v2.u32 {%r286, %r287}, [kernel_params+1616];
|
|
mov.b32 %f90, %r287;
|
|
setp.geu.ftz.f32 %p132, %f90, 0f3F800000;
|
|
setp.lt.ftz.f32 %p133, %f90, 0f3F800000;
|
|
ld.const.u32 %r290, [kernel_params+1592];
|
|
selp.b32 %r291, %r290, %r84, %p133;
|
|
and.b32 %r292, %r85, 255;
|
|
xor.b32 %r293, %r291, %r292;
|
|
mul.lo.s32 %r294, %r293, -512718531;
|
|
shr.u32 %r295, %r291, 16;
|
|
xor.b32 %r296, %r294, %r295;
|
|
bfe.u32 %r297, %r296, 4, 4;
|
|
shr.u32 %r298, %r291, 8;
|
|
xor.b32 %r299, %r296, %r298;
|
|
xor.b32 %r300, %r299, %r297;
|
|
mul.lo.s32 %r301, %r300, 153742143;
|
|
shr.u32 %r302, %r291, 23;
|
|
xor.b32 %r303, %r301, %r302;
|
|
bfe.u32 %r304, %r303, 1, 7;
|
|
xor.b32 %r305, %r304, %r303;
|
|
shr.u32 %r306, %r291, 27;
|
|
or.b32 %r307, %r306, 1;
|
|
mul.lo.s32 %r308, %r307, %r305;
|
|
mul.lo.s32 %r309, %r308, -1074439621;
|
|
bfe.u32 %r310, %r309, 2, 6;
|
|
xor.b32 %r311, %r310, %r309;
|
|
mul.lo.s32 %r312, %r311, -1638916925;
|
|
bfe.u32 %r313, %r312, 2, 6;
|
|
xor.b32 %r314, %r313, %r312;
|
|
mul.lo.s32 %r315, %r314, -933190689;
|
|
and.b32 %r316, %r315, 255;
|
|
bfe.u32 %r317, %r315, 5, 3;
|
|
xor.b32 %r318, %r317, %r316;
|
|
add.s32 %r319, %r286, -1;
|
|
xor.b32 %r88, %r85, 61;
|
|
xor.b32 %r320, %r291, %r88;
|
|
mul.lo.s32 %r321, %r320, 9;
|
|
shr.u32 %r322, %r321, 4;
|
|
xor.b32 %r323, %r322, %r321;
|
|
mul.lo.s32 %r324, %r323, 668265261;
|
|
add.s32 %r325, %r324, %r87;
|
|
shr.u32 %r326, %r324, 16;
|
|
or.b32 %r327, %r326, 1;
|
|
mul.lo.s32 %r328, %r327, %r325;
|
|
mul.lo.s32 %r329, %r328, 89287766;
|
|
xor.b32 %r330, %r329, %r328;
|
|
mul.lo.s32 %r331, %r330, 1403136100;
|
|
xor.b32 %r332, %r331, %r330;
|
|
brev.b32 %r333, %r332;
|
|
neg.s32 %r334, %r286;
|
|
and.b32 %r335, %r86, %r334;
|
|
and.b32 %r336, %r333, %r319;
|
|
or.b32 %r337, %r336, %r335;
|
|
mad.lo.s32 %r338, %r318, %r286, %r337;
|
|
shl.b32 %r339, %r286, 8;
|
|
rem.u32 %r340, %r338, %r339;
|
|
ld.const.u64 %rd258, [kernel_params+2440];
|
|
cvta.to.global.u64 %rd259, %rd258;
|
|
shl.b32 %r341, %r340, 1;
|
|
mul.wide.u32 %rd260, %r341, 4;
|
|
add.s64 %rd261, %rd259, %rd260;
|
|
ld.global.f32 %f350, [%rd261];
|
|
@%p132 bra $L__BB9_93;
|
|
|
|
xor.b32 %r342, %r88, %r84;
|
|
mul.lo.s32 %r343, %r342, 9;
|
|
shr.u32 %r344, %r343, 4;
|
|
xor.b32 %r345, %r344, %r343;
|
|
mul.lo.s32 %r346, %r345, 668265261;
|
|
cvt.rn.f32.u32 %f248, %r346;
|
|
mul.ftz.f32 %f249, %f248, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f250, %f249, %f90, %f350;
|
|
cvt.rmi.ftz.f32.f32 %f251, %f250;
|
|
sub.ftz.f32 %f350, %f250, %f251;
|
|
bra.uni $L__BB9_93;
|
|
|
|
$L__BB9_79:
|
|
mov.u16 %rs85, 1;
|
|
@%p112 bra $L__BB9_81;
|
|
|
|
ld.const.u64 %rd237, [kernel_params+2792];
|
|
cvta.to.global.u64 %rd238, %rd237;
|
|
add.s64 %rd240, %rd238, %rd186;
|
|
ld.global.u32 %r276, [%rd240];
|
|
setp.eq.s32 %p125, %r276, -1;
|
|
selp.u16 %rs85, 1, 0, %p125;
|
|
|
|
$L__BB9_81:
|
|
setp.ne.s16 %p126, %rs85, 0;
|
|
@%p126 bra $L__BB9_99;
|
|
|
|
ld.global.u32 %r277, [%rd28];
|
|
or.b32 %r594, %r277, 262144;
|
|
st.global.u32 [%rd28], %r594;
|
|
bra.uni $L__BB9_84;
|
|
|
|
$L__BB9_92:
|
|
mul.lo.s32 %r347, %r85, 569420461;
|
|
shr.u32 %r348, %r347, 15;
|
|
xor.b32 %r349, %r348, %r347;
|
|
mul.lo.s32 %r350, %r349, -749064809;
|
|
shr.u32 %r351, %r350, 15;
|
|
xor.b32 %r352, %r84, %r350;
|
|
xor.b32 %r353, %r352, %r351;
|
|
xor.b32 %r354, %r353, 1493655573;
|
|
add.s32 %r355, %r354, %r87;
|
|
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;
|
|
xor.b32 %r363, %r353, -2052961194;
|
|
brev.b32 %r364, %r362;
|
|
mul.lo.s32 %r365, %r364, 1025551850;
|
|
xor.b32 %r366, %r365, %r364;
|
|
add.s32 %r367, %r366, %r363;
|
|
shr.u32 %r368, %r363, 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 %f252, %r375;
|
|
mul.ftz.f32 %f350, %f252, 0f2F7FFFFE;
|
|
|
|
$L__BB9_93:
|
|
setp.ltu.ftz.f32 %p134, %f350, %f349;
|
|
setp.neu.ftz.f32 %p135, %f349, 0f00000000;
|
|
and.pred %p136, %p135, %p134;
|
|
@%p136 bra $L__BB9_101;
|
|
|
|
and.b32 %r376, %r79, 536870912;
|
|
setp.eq.s32 %p137, %r376, 0;
|
|
@%p137 bra $L__BB9_96;
|
|
bra.uni $L__BB9_95;
|
|
|
|
$L__BB9_96:
|
|
mov.u16 %rs86, 1;
|
|
@%p112 bra $L__BB9_98;
|
|
|
|
ld.const.u64 %rd262, [kernel_params+2792];
|
|
cvta.to.global.u64 %rd263, %rd262;
|
|
add.s64 %rd265, %rd263, %rd186;
|
|
ld.global.u32 %r379, [%rd265];
|
|
setp.eq.s32 %p139, %r379, -1;
|
|
selp.u16 %rs86, 1, 0, %p139;
|
|
|
|
$L__BB9_98:
|
|
setp.eq.s16 %p140, %rs86, 0;
|
|
@%p140 bra $L__BB9_100;
|
|
|
|
$L__BB9_99:
|
|
ld.const.u64 %rd266, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd267, %rd266;
|
|
add.s64 %rd268, %rd267, 8;
|
|
atom.global.add.u32 %r380, [%rd268], -1;
|
|
ld.const.u64 %rd269, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd270, %rd269;
|
|
shl.b64 %rd271, %rd6, 1;
|
|
add.s64 %rd272, %rd270, %rd271;
|
|
mov.u16 %rs45, 0;
|
|
st.global.u16 [%rd272], %rs45;
|
|
bra.uni $L__BB9_283;
|
|
|
|
$L__BB9_95:
|
|
ld.global.u32 %r377, [%rd28];
|
|
or.b32 %r378, %r377, 32768;
|
|
st.global.u32 [%rd28], %r378;
|
|
bra.uni $L__BB9_101;
|
|
|
|
$L__BB9_100:
|
|
ld.global.u32 %r381, [%rd28];
|
|
or.b32 %r382, %r381, 65536;
|
|
st.global.u32 [%rd28], %r382;
|
|
|
|
$L__BB9_101:
|
|
ld.const.u32 %r571, [kernel_params+1608];
|
|
ld.const.u64 %rd273, [kernel_params+2200];
|
|
cvta.to.global.u64 %rd274, %rd273;
|
|
mul.wide.s32 %rd275, %r585, 4;
|
|
add.s64 %rd67, %rd274, %rd275;
|
|
ld.global.u8 %rs46, [%rd67+1];
|
|
and.b16 %rs47, %rs46, 6;
|
|
setp.ne.s16 %p141, %rs47, 0;
|
|
setp.ne.s32 %p142, %r571, 0;
|
|
and.pred %p143, %p141, %p142;
|
|
@%p143 bra $L__BB9_105;
|
|
bra.uni $L__BB9_102;
|
|
|
|
$L__BB9_105:
|
|
cvt.u32.u64 %r391, %rd6;
|
|
ld.const.u32 %r595, [kernel_params+4408];
|
|
div.u32 %r392, %r391, %r595;
|
|
ld.const.u32 %r596, [kernel_params+24];
|
|
mad.lo.s32 %r393, %r392, %r596, %r78;
|
|
ld.const.u64 %rd78, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd299, %rd78;
|
|
add.s64 %rd300, %rd299, 8;
|
|
atom.global.add.u32 %r394, [%rd300], -1;
|
|
add.s64 %rd301, %rd299, 40;
|
|
atom.global.add.u32 %r395, [%rd301], 1;
|
|
ld.const.u64 %rd80, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd302, %rd80;
|
|
add.s64 %rd304, %rd302, %rd244;
|
|
mov.u16 %rs50, 10;
|
|
st.global.u16 [%rd304], %rs50;
|
|
ld.const.u64 %rd79, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd305, %rd79;
|
|
add.s64 %rd307, %rd305, %rd186;
|
|
st.global.u32 [%rd307], %r393;
|
|
ld.const.u64 %rd308, [kernel_params+4360];
|
|
cvta.to.global.u64 %rd309, %rd308;
|
|
mul.wide.s32 %rd310, %r393, 4;
|
|
add.s64 %rd311, %rd309, %rd310;
|
|
atom.global.add.u32 %r396, [%rd311], 1;
|
|
bra.uni $L__BB9_106;
|
|
|
|
$L__BB9_102:
|
|
cvt.u32.u64 %r383, %rd6;
|
|
and.b32 %r384, %r79, 1073741824;
|
|
setp.eq.s32 %p144, %r384, 0;
|
|
ld.const.u32 %r595, [kernel_params+4408];
|
|
div.u32 %r385, %r383, %r595;
|
|
ld.const.u32 %r596, [kernel_params+24];
|
|
mad.lo.s32 %r91, %r385, %r596, %r78;
|
|
ld.const.u64 %rd78, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd69, %rd78;
|
|
add.s64 %rd276, %rd69, 8;
|
|
atom.global.add.u32 %r386, [%rd276], -1;
|
|
cvt.s64.s32 %rd70, %r91;
|
|
@%p144 bra $L__BB9_104;
|
|
bra.uni $L__BB9_103;
|
|
|
|
$L__BB9_104:
|
|
add.s64 %rd288, %rd69, 32;
|
|
atom.global.add.u32 %r389, [%rd288], 1;
|
|
ld.const.u64 %rd80, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd289, %rd80;
|
|
add.s64 %rd291, %rd289, %rd244;
|
|
mov.u16 %rs49, 8;
|
|
st.global.u16 [%rd291], %rs49;
|
|
ld.const.u64 %rd79, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd292, %rd79;
|
|
add.s64 %rd294, %rd292, %rd186;
|
|
st.global.u32 [%rd294], %r91;
|
|
ld.const.u64 %rd295, [kernel_params+4344];
|
|
cvta.to.global.u64 %rd296, %rd295;
|
|
shl.b64 %rd297, %rd70, 2;
|
|
add.s64 %rd298, %rd296, %rd297;
|
|
atom.global.add.u32 %r390, [%rd298], 1;
|
|
bra.uni $L__BB9_106;
|
|
|
|
$L__BB9_103:
|
|
add.s64 %rd277, %rd69, 36;
|
|
atom.global.add.u32 %r387, [%rd277], 1;
|
|
ld.const.u64 %rd80, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd278, %rd80;
|
|
add.s64 %rd280, %rd278, %rd244;
|
|
mov.u16 %rs48, 9;
|
|
st.global.u16 [%rd280], %rs48;
|
|
ld.const.u64 %rd79, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd281, %rd79;
|
|
add.s64 %rd283, %rd281, %rd186;
|
|
st.global.u32 [%rd283], %r91;
|
|
ld.const.u64 %rd284, [kernel_params+4352];
|
|
cvta.to.global.u64 %rd285, %rd284;
|
|
shl.b64 %rd286, %rd70, 2;
|
|
add.s64 %rd287, %rd285, %rd286;
|
|
atom.global.add.u32 %r388, [%rd287], 1;
|
|
|
|
$L__BB9_106:
|
|
ld.const.u32 %r397, [kernel_params+1636];
|
|
setp.eq.s32 %p145, %r397, 0;
|
|
ld.global.u32 %r96, [%rd67];
|
|
and.b32 %r398, %r96, 129;
|
|
setp.ne.s32 %p146, %r398, 128;
|
|
or.pred %p147, %p146, %p145;
|
|
mov.u16 %rs87, 0;
|
|
@%p147 bra $L__BB9_109;
|
|
|
|
ld.global.u32 %r97, [%rd28];
|
|
and.b32 %r399, %r97, 16384;
|
|
setp.eq.s32 %p148, %r399, 0;
|
|
@%p148 bra $L__BB9_109;
|
|
|
|
and.b32 %r400, %r97, 536870912;
|
|
setp.eq.s32 %p149, %r400, 0;
|
|
selp.u16 %rs87, 1, 0, %p149;
|
|
|
|
$L__BB9_109:
|
|
setp.eq.s16 %p150, %rs87, 0;
|
|
@%p150 bra $L__BB9_283;
|
|
|
|
ld.const.u64 %rd739, [kernel_params+8];
|
|
cvta.to.global.u64 %rd312, %rd739;
|
|
ld.const.u64 %rd313, [kernel_params+2464];
|
|
cvta.to.global.u64 %rd81, %rd313;
|
|
add.s64 %rd82, %rd81, %rd186;
|
|
ld.global.u32 %rd315, [%rd82];
|
|
ld.const.s32 %rd316, [kernel_params+1280];
|
|
mul.lo.s64 %rd317, %rd316, %rd315;
|
|
ld.const.s32 %rd318, [kernel_params+1396];
|
|
add.s64 %rd319, %rd317, %rd318;
|
|
shl.b64 %rd320, %rd319, 2;
|
|
add.s64 %rd321, %rd312, %rd320;
|
|
atom.global.add.f32 %f253, [%rd321], 0f3F800000;
|
|
ld.const.u64 %rd83, [kernel_params+2600];
|
|
cvta.to.global.u64 %rd84, %rd83;
|
|
mul.lo.s64 %rd322, %rd6, 12;
|
|
add.s64 %rd85, %rd84, %rd322;
|
|
ld.const.s32 %rd323, [kernel_params+1400];
|
|
add.s64 %rd324, %rd323, %rd317;
|
|
shl.b64 %rd325, %rd324, 2;
|
|
add.s64 %rd326, %rd312, %rd325;
|
|
add.s64 %rd327, %rd326, 12;
|
|
ld.global.f32 %f254, [%rd85+4];
|
|
ld.global.f32 %f255, [%rd85];
|
|
add.ftz.f32 %f256, %f255, %f254;
|
|
ld.global.f32 %f257, [%rd85+8];
|
|
add.ftz.f32 %f258, %f257, %f256;
|
|
mul.ftz.f32 %f259, %f258, 0f3EAAAAAB;
|
|
atom.global.add.f32 %f260, [%rd327], %f259;
|
|
ld.global.u32 %r401, [%rd28];
|
|
or.b32 %r402, %r401, 536870912;
|
|
st.global.u32 [%rd28], %r402;
|
|
ld.const.u64 %rd328, [kernel_params+4400];
|
|
cvta.to.global.u64 %rd329, %rd328;
|
|
atom.global.add.u32 %r98, [%rd329], 1;
|
|
setp.eq.s64 %p151, %rd313, 0;
|
|
@%p151 bra $L__BB9_112;
|
|
|
|
ld.global.u32 %r403, [%rd82];
|
|
mul.wide.s32 %rd330, %r98, 4;
|
|
add.s64 %rd331, %rd81, %rd330;
|
|
st.global.u32 [%rd331], %r403;
|
|
|
|
$L__BB9_112:
|
|
ld.const.u64 %rd707, [kernel_params+2472];
|
|
setp.eq.s64 %p152, %rd707, 0;
|
|
@%p152 bra $L__BB9_114;
|
|
|
|
ld.const.u64 %rd752, [kernel_params+2472];
|
|
cvta.to.global.u64 %rd751, %rd752;
|
|
add.s64 %rd750, %rd751, %rd186;
|
|
ld.const.u64 %rd743, [kernel_params+2472];
|
|
cvta.to.global.u64 %rd742, %rd743;
|
|
ld.global.u32 %r404, [%rd750];
|
|
mul.wide.s32 %rd332, %r98, 4;
|
|
add.s64 %rd333, %rd742, %rd332;
|
|
st.global.u32 [%rd333], %r404;
|
|
|
|
$L__BB9_114:
|
|
ld.const.u64 %rd86, [kernel_params+2480];
|
|
setp.eq.s64 %p153, %rd86, 0;
|
|
@%p153 bra $L__BB9_116;
|
|
|
|
cvta.to.global.u64 %rd334, %rd86;
|
|
add.s64 %rd336, %rd334, %rd244;
|
|
ld.global.u16 %rs53, [%rd336];
|
|
mul.wide.s32 %rd337, %r98, 2;
|
|
add.s64 %rd338, %rd334, %rd337;
|
|
st.global.u16 [%rd338], %rs53;
|
|
|
|
$L__BB9_116:
|
|
ld.const.u64 %rd87, [kernel_params+2488];
|
|
setp.eq.s64 %p154, %rd87, 0;
|
|
@%p154 bra $L__BB9_118;
|
|
|
|
cvta.to.global.u64 %rd339, %rd87;
|
|
add.s64 %rd341, %rd339, %rd244;
|
|
ld.global.u16 %rs54, [%rd341];
|
|
mul.wide.s32 %rd342, %r98, 2;
|
|
add.s64 %rd343, %rd339, %rd342;
|
|
st.global.u16 [%rd343], %rs54;
|
|
|
|
$L__BB9_118:
|
|
ld.const.u64 %rd88, [kernel_params+2496];
|
|
setp.eq.s64 %p155, %rd88, 0;
|
|
@%p155 bra $L__BB9_120;
|
|
|
|
cvta.to.global.u64 %rd344, %rd88;
|
|
add.s64 %rd346, %rd344, %rd244;
|
|
ld.global.u16 %rs55, [%rd346];
|
|
mul.wide.s32 %rd347, %r98, 2;
|
|
add.s64 %rd348, %rd344, %rd347;
|
|
st.global.u16 [%rd348], %rs55;
|
|
|
|
$L__BB9_120:
|
|
ld.const.u64 %rd89, [kernel_params+2504];
|
|
setp.eq.s64 %p156, %rd89, 0;
|
|
@%p156 bra $L__BB9_122;
|
|
|
|
cvta.to.global.u64 %rd349, %rd89;
|
|
add.s64 %rd351, %rd349, %rd244;
|
|
ld.global.u16 %rs56, [%rd351];
|
|
mul.wide.s32 %rd352, %r98, 2;
|
|
add.s64 %rd353, %rd349, %rd352;
|
|
st.global.u16 [%rd353], %rs56;
|
|
|
|
$L__BB9_122:
|
|
ld.const.u64 %rd90, [kernel_params+2512];
|
|
setp.eq.s64 %p157, %rd90, 0;
|
|
@%p157 bra $L__BB9_124;
|
|
|
|
cvta.to.global.u64 %rd354, %rd90;
|
|
add.s64 %rd356, %rd354, %rd244;
|
|
ld.global.u16 %rs57, [%rd356];
|
|
mul.wide.s32 %rd357, %r98, 2;
|
|
add.s64 %rd358, %rd354, %rd357;
|
|
st.global.u16 [%rd358], %rs57;
|
|
|
|
$L__BB9_124:
|
|
ld.const.u64 %rd91, [kernel_params+2520];
|
|
setp.eq.s64 %p158, %rd91, 0;
|
|
@%p158 bra $L__BB9_126;
|
|
|
|
cvta.to.global.u64 %rd359, %rd91;
|
|
add.s64 %rd361, %rd359, %rd244;
|
|
ld.global.u16 %rs58, [%rd361];
|
|
mul.wide.s32 %rd362, %r98, 2;
|
|
add.s64 %rd363, %rd359, %rd362;
|
|
st.global.u16 [%rd363], %rs58;
|
|
|
|
$L__BB9_126:
|
|
ld.const.u64 %rd92, [kernel_params+2528];
|
|
setp.eq.s64 %p159, %rd92, 0;
|
|
@%p159 bra $L__BB9_128;
|
|
|
|
cvta.to.global.u64 %rd364, %rd92;
|
|
add.s64 %rd366, %rd364, %rd244;
|
|
ld.global.u16 %rs59, [%rd366];
|
|
mul.wide.s32 %rd367, %r98, 2;
|
|
add.s64 %rd368, %rd364, %rd367;
|
|
st.global.u16 [%rd368], %rs59;
|
|
|
|
$L__BB9_128:
|
|
setp.eq.s64 %p160, %rd80, 0;
|
|
@%p160 bra $L__BB9_130;
|
|
|
|
cvta.to.global.u64 %rd369, %rd80;
|
|
add.s64 %rd371, %rd369, %rd244;
|
|
ld.global.u16 %rs60, [%rd371];
|
|
mul.wide.s32 %rd372, %r98, 2;
|
|
add.s64 %rd373, %rd369, %rd372;
|
|
st.global.u16 [%rd373], %rs60;
|
|
|
|
$L__BB9_130:
|
|
setp.eq.s64 %p161, %rd241, 0;
|
|
@%p161 bra $L__BB9_132;
|
|
|
|
cvta.to.global.u64 %rd709, %rd241;
|
|
add.s64 %rd708, %rd709, %rd186;
|
|
ld.global.u32 %r405, [%rd708];
|
|
mul.wide.s32 %rd374, %r98, 4;
|
|
add.s64 %rd375, %rd709, %rd374;
|
|
st.global.u32 [%rd375], %r405;
|
|
|
|
$L__BB9_132:
|
|
setp.eq.s64 %p162, %rd242, 0;
|
|
@%p162 bra $L__BB9_134;
|
|
|
|
cvta.to.global.u64 %rd745, %rd242;
|
|
add.s64 %rd744, %rd745, %rd244;
|
|
cvta.to.global.u64 %rd710, %rd242;
|
|
ld.global.u16 %rs61, [%rd744];
|
|
mul.wide.s32 %rd376, %r98, 2;
|
|
add.s64 %rd377, %rd710, %rd376;
|
|
st.global.u16 [%rd377], %rs61;
|
|
|
|
$L__BB9_134:
|
|
ld.const.u64 %rd741, [kernel_params+2560];
|
|
cvta.to.global.u64 %rd740, %rd741;
|
|
ld.const.u64 %rd711, [kernel_params+2560];
|
|
setp.eq.s64 %p163, %rd711, 0;
|
|
cvt.s64.s32 %rd93, %r98;
|
|
mul.wide.s32 %rd378, %r98, 4;
|
|
add.s64 %rd94, %rd740, %rd378;
|
|
@%p163 bra $L__BB9_136;
|
|
|
|
ld.global.u32 %r406, [%rd28];
|
|
st.global.u32 [%rd94], %r406;
|
|
|
|
$L__BB9_136:
|
|
ld.const.u64 %rd95, [kernel_params+2568];
|
|
setp.eq.s64 %p164, %rd95, 0;
|
|
@%p164 bra $L__BB9_138;
|
|
|
|
cvta.to.global.u64 %rd379, %rd95;
|
|
add.s64 %rd380, %rd379, %rd6;
|
|
ld.global.u8 %rs62, [%rd380];
|
|
add.s64 %rd382, %rd379, %rd93;
|
|
st.global.u8 [%rd382], %rs62;
|
|
|
|
$L__BB9_138:
|
|
ld.const.u64 %rd96, [kernel_params+2576];
|
|
setp.eq.s64 %p165, %rd96, 0;
|
|
@%p165 bra $L__BB9_140;
|
|
|
|
cvta.to.global.u64 %rd383, %rd96;
|
|
add.s64 %rd385, %rd383, %rd186;
|
|
ld.global.f32 %f261, [%rd385];
|
|
add.s64 %rd387, %rd383, %rd378;
|
|
st.global.f32 [%rd387], %f261;
|
|
|
|
$L__BB9_140:
|
|
ld.const.u64 %rd97, [kernel_params+2584];
|
|
setp.eq.s64 %p166, %rd97, 0;
|
|
@%p166 bra $L__BB9_142;
|
|
|
|
cvta.to.global.u64 %rd388, %rd97;
|
|
add.s64 %rd390, %rd388, %rd186;
|
|
ld.global.f32 %f262, [%rd390];
|
|
add.s64 %rd392, %rd388, %rd378;
|
|
st.global.f32 [%rd392], %f262;
|
|
|
|
$L__BB9_142:
|
|
ld.const.u64 %rd712, [kernel_params+2592];
|
|
setp.eq.s64 %p167, %rd712, 0;
|
|
@%p167 bra $L__BB9_144;
|
|
|
|
ld.const.u64 %rd755, [kernel_params+2592];
|
|
cvta.to.global.u64 %rd754, %rd755;
|
|
add.s64 %rd753, %rd754, %rd186;
|
|
ld.const.u64 %rd714, [kernel_params+2592];
|
|
cvta.to.global.u64 %rd713, %rd714;
|
|
ld.global.f32 %f263, [%rd753];
|
|
add.s64 %rd394, %rd713, %rd378;
|
|
st.global.f32 [%rd394], %f263;
|
|
|
|
$L__BB9_144:
|
|
ld.const.u64 %rd715, [kernel_params+2600];
|
|
setp.eq.s64 %p168, %rd715, 0;
|
|
@%p168 bra $L__BB9_146;
|
|
|
|
ld.const.u64 %rd757, [kernel_params+2600];
|
|
cvta.to.global.u64 %rd756, %rd757;
|
|
ld.global.f32 %f264, [%rd85];
|
|
ld.global.f32 %f265, [%rd85+4];
|
|
ld.global.f32 %f266, [%rd85+8];
|
|
mul.wide.s32 %rd395, %r98, 12;
|
|
add.s64 %rd396, %rd756, %rd395;
|
|
st.global.f32 [%rd396], %f264;
|
|
st.global.f32 [%rd396+4], %f265;
|
|
st.global.f32 [%rd396+8], %f266;
|
|
|
|
$L__BB9_146:
|
|
ld.const.u64 %rd98, [kernel_params+2608];
|
|
setp.eq.s64 %p169, %rd98, 0;
|
|
@%p169 bra $L__BB9_148;
|
|
|
|
cvta.to.global.u64 %rd397, %rd98;
|
|
add.s64 %rd399, %rd397, %rd186;
|
|
ld.global.f32 %f267, [%rd399];
|
|
add.s64 %rd401, %rd397, %rd378;
|
|
st.global.f32 [%rd401], %f267;
|
|
|
|
$L__BB9_148:
|
|
ld.const.u64 %rd99, [kernel_params+2616];
|
|
setp.eq.s64 %p170, %rd99, 0;
|
|
@%p170 bra $L__BB9_150;
|
|
|
|
cvta.to.global.u64 %rd402, %rd99;
|
|
add.s64 %rd404, %rd402, %rd322;
|
|
ld.global.f32 %f268, [%rd404];
|
|
ld.global.f32 %f269, [%rd404+4];
|
|
ld.global.f32 %f270, [%rd404+8];
|
|
mul.wide.s32 %rd405, %r98, 12;
|
|
add.s64 %rd406, %rd402, %rd405;
|
|
st.global.f32 [%rd406], %f268;
|
|
st.global.f32 [%rd406+4], %f269;
|
|
st.global.f32 [%rd406+8], %f270;
|
|
|
|
$L__BB9_150:
|
|
ld.const.u64 %rd100, [kernel_params+2624];
|
|
setp.eq.s64 %p171, %rd100, 0;
|
|
@%p171 bra $L__BB9_152;
|
|
|
|
cvta.to.global.u64 %rd407, %rd100;
|
|
add.s64 %rd409, %rd407, %rd322;
|
|
ld.global.f32 %f271, [%rd409];
|
|
ld.global.f32 %f272, [%rd409+4];
|
|
ld.global.f32 %f273, [%rd409+8];
|
|
mul.wide.s32 %rd410, %r98, 12;
|
|
add.s64 %rd411, %rd407, %rd410;
|
|
st.global.f32 [%rd411], %f271;
|
|
st.global.f32 [%rd411+4], %f272;
|
|
st.global.f32 [%rd411+8], %f273;
|
|
|
|
$L__BB9_152:
|
|
ld.const.u64 %rd101, [kernel_params+2632];
|
|
setp.eq.s64 %p172, %rd101, 0;
|
|
@%p172 bra $L__BB9_154;
|
|
|
|
cvta.to.global.u64 %rd412, %rd101;
|
|
add.s64 %rd414, %rd412, %rd322;
|
|
ld.global.f32 %f274, [%rd414];
|
|
ld.global.f32 %f275, [%rd414+4];
|
|
ld.global.f32 %f276, [%rd414+8];
|
|
mul.wide.s32 %rd415, %r98, 12;
|
|
add.s64 %rd416, %rd412, %rd415;
|
|
st.global.f32 [%rd416], %f274;
|
|
st.global.f32 [%rd416+4], %f275;
|
|
st.global.f32 [%rd416+8], %f276;
|
|
|
|
$L__BB9_154:
|
|
setp.eq.s64 %p173, %rd79, 0;
|
|
@%p173 bra $L__BB9_156;
|
|
|
|
cvta.to.global.u64 %rd417, %rd79;
|
|
add.s64 %rd419, %rd417, %rd186;
|
|
ld.global.u32 %r407, [%rd419];
|
|
add.s64 %rd421, %rd417, %rd378;
|
|
st.global.u32 [%rd421], %r407;
|
|
|
|
$L__BB9_156:
|
|
ld.const.u64 %rd716, [kernel_params+2648];
|
|
setp.eq.s64 %p174, %rd716, 0;
|
|
@%p174 bra $L__BB9_158;
|
|
|
|
ld.const.u64 %rd797, [kernel_params+2648];
|
|
mul.wide.s32 %rd796, %r3, 12;
|
|
cvta.to.global.u64 %rd795, %rd797;
|
|
add.s64 %rd794, %rd795, %rd796;
|
|
ld.const.u64 %rd718, [kernel_params+2648];
|
|
cvta.to.global.u64 %rd717, %rd718;
|
|
ld.global.f32 %f277, [%rd794];
|
|
ld.global.f32 %f278, [%rd794+4];
|
|
ld.global.f32 %f279, [%rd794+8];
|
|
mul.wide.s32 %rd422, %r98, 12;
|
|
add.s64 %rd423, %rd717, %rd422;
|
|
st.global.f32 [%rd423], %f277;
|
|
st.global.f32 [%rd423+4], %f278;
|
|
st.global.f32 [%rd423+8], %f279;
|
|
|
|
$L__BB9_158:
|
|
setp.eq.s64 %p175, %rd142, 0;
|
|
@%p175 bra $L__BB9_160;
|
|
|
|
mul.wide.s32 %rd785, %r3, 12;
|
|
cvta.to.global.u64 %rd784, %rd142;
|
|
add.s64 %rd783, %rd784, %rd785;
|
|
cvta.to.global.u64 %rd719, %rd142;
|
|
ld.global.f32 %f280, [%rd783];
|
|
ld.global.f32 %f281, [%rd783+4];
|
|
ld.global.f32 %f282, [%rd783+8];
|
|
mul.wide.s32 %rd424, %r98, 12;
|
|
add.s64 %rd425, %rd719, %rd424;
|
|
st.global.f32 [%rd425], %f280;
|
|
st.global.f32 [%rd425+4], %f281;
|
|
st.global.f32 [%rd425+8], %f282;
|
|
|
|
$L__BB9_160:
|
|
setp.eq.s64 %p176, %rd143, 0;
|
|
@%p176 bra $L__BB9_162;
|
|
|
|
mul.wide.s32 %rd760, %r3, 4;
|
|
cvta.to.global.u64 %rd759, %rd143;
|
|
add.s64 %rd758, %rd759, %rd760;
|
|
cvta.to.global.u64 %rd720, %rd143;
|
|
ld.global.f32 %f283, [%rd758];
|
|
add.s64 %rd427, %rd720, %rd378;
|
|
st.global.f32 [%rd427], %f283;
|
|
|
|
$L__BB9_162:
|
|
setp.eq.s64 %p177, %rd145, 0;
|
|
@%p177 bra $L__BB9_164;
|
|
|
|
mul.wide.s32 %rd763, %r3, 4;
|
|
cvta.to.global.u64 %rd762, %rd145;
|
|
add.s64 %rd761, %rd762, %rd763;
|
|
cvta.to.global.u64 %rd721, %rd145;
|
|
ld.global.f32 %f284, [%rd761];
|
|
add.s64 %rd429, %rd721, %rd378;
|
|
st.global.f32 [%rd429], %f284;
|
|
|
|
$L__BB9_164:
|
|
setp.eq.s64 %p178, %rd146, 0;
|
|
@%p178 bra $L__BB9_166;
|
|
|
|
mul.wide.s32 %rd766, %r3, 4;
|
|
cvta.to.global.u64 %rd765, %rd146;
|
|
add.s64 %rd764, %rd765, %rd766;
|
|
cvta.to.global.u64 %rd722, %rd146;
|
|
ld.global.f32 %f285, [%rd764];
|
|
add.s64 %rd431, %rd722, %rd378;
|
|
st.global.f32 [%rd431], %f285;
|
|
|
|
$L__BB9_166:
|
|
setp.eq.s64 %p179, %rd147, 0;
|
|
@%p179 bra $L__BB9_168;
|
|
|
|
mul.wide.s32 %rd769, %r3, 4;
|
|
cvta.to.global.u64 %rd768, %rd147;
|
|
add.s64 %rd767, %rd768, %rd769;
|
|
cvta.to.global.u64 %rd723, %rd147;
|
|
ld.global.f32 %f286, [%rd767];
|
|
add.s64 %rd433, %rd723, %rd378;
|
|
st.global.f32 [%rd433], %f286;
|
|
|
|
$L__BB9_168:
|
|
setp.eq.s64 %p180, %rd148, 0;
|
|
@%p180 bra $L__BB9_170;
|
|
|
|
mul.wide.s32 %rd772, %r3, 4;
|
|
cvta.to.global.u64 %rd771, %rd148;
|
|
add.s64 %rd770, %rd771, %rd772;
|
|
cvta.to.global.u64 %rd724, %rd148;
|
|
ld.global.f32 %f287, [%rd770];
|
|
add.s64 %rd435, %rd724, %rd378;
|
|
st.global.f32 [%rd435], %f287;
|
|
|
|
$L__BB9_170:
|
|
setp.eq.s64 %p181, %rd184, 0;
|
|
@%p181 bra $L__BB9_172;
|
|
|
|
cvta.to.global.u64 %rd726, %rd184;
|
|
add.s64 %rd725, %rd726, %rd186;
|
|
ld.global.f32 %f288, [%rd725];
|
|
add.s64 %rd437, %rd726, %rd378;
|
|
st.global.f32 [%rd437], %f288;
|
|
|
|
$L__BB9_172:
|
|
setp.eq.s64 %p182, %rd185, 0;
|
|
@%p182 bra $L__BB9_174;
|
|
|
|
cvta.to.global.u64 %rd728, %rd185;
|
|
add.s64 %rd727, %rd728, %rd186;
|
|
ld.global.f32 %f289, [%rd727];
|
|
add.s64 %rd439, %rd728, %rd378;
|
|
st.global.f32 [%rd439], %f289;
|
|
|
|
$L__BB9_174:
|
|
ld.const.u64 %rd729, [kernel_params+2720];
|
|
setp.eq.s64 %p183, %rd729, 0;
|
|
@%p183 bra $L__BB9_176;
|
|
|
|
ld.const.u64 %rd775, [kernel_params+2720];
|
|
cvta.to.global.u64 %rd774, %rd775;
|
|
add.s64 %rd773, %rd774, %rd186;
|
|
ld.const.u64 %rd731, [kernel_params+2720];
|
|
cvta.to.global.u64 %rd730, %rd731;
|
|
ld.global.f32 %f290, [%rd773];
|
|
add.s64 %rd441, %rd730, %rd378;
|
|
st.global.f32 [%rd441], %f290;
|
|
|
|
$L__BB9_176:
|
|
ld.const.u64 %rd732, [kernel_params+2728];
|
|
setp.eq.s64 %p184, %rd732, 0;
|
|
@%p184 bra $L__BB9_178;
|
|
|
|
ld.const.u64 %rd734, [kernel_params+2728];
|
|
cvta.to.global.u64 %rd733, %rd734;
|
|
ld.global.u32 %r408, [%rd31];
|
|
add.s64 %rd443, %rd733, %rd378;
|
|
st.global.u32 [%rd443], %r408;
|
|
|
|
$L__BB9_178:
|
|
ld.const.u64 %rd735, [kernel_params+2736];
|
|
setp.eq.s64 %p185, %rd735, 0;
|
|
@%p185 bra $L__BB9_180;
|
|
|
|
ld.const.u64 %rd737, [kernel_params+2736];
|
|
cvta.to.global.u64 %rd736, %rd737;
|
|
ld.global.u32 %r409, [%rd34];
|
|
add.s64 %rd445, %rd736, %rd378;
|
|
st.global.u32 [%rd445], %r409;
|
|
|
|
$L__BB9_180:
|
|
ld.const.u64 %rd738, [kernel_params+2744];
|
|
setp.eq.s64 %p186, %rd738, 0;
|
|
@%p186 bra $L__BB9_182;
|
|
|
|
ld.const.u64 %rd778, [kernel_params+2744];
|
|
cvta.to.global.u64 %rd777, %rd778;
|
|
add.s64 %rd776, %rd777, %rd186;
|
|
ld.const.u64 %rd747, [kernel_params+2744];
|
|
cvta.to.global.u64 %rd746, %rd747;
|
|
ld.global.u32 %r410, [%rd776];
|
|
add.s64 %rd447, %rd746, %rd378;
|
|
st.global.u32 [%rd447], %r410;
|
|
|
|
$L__BB9_182:
|
|
ld.const.u64 %rd102, [kernel_params+2752];
|
|
setp.eq.s64 %p187, %rd102, 0;
|
|
@%p187 bra $L__BB9_184;
|
|
|
|
cvta.to.global.u64 %rd448, %rd102;
|
|
add.s64 %rd450, %rd448, %rd322;
|
|
ld.global.f32 %f291, [%rd450];
|
|
ld.global.f32 %f292, [%rd450+4];
|
|
ld.global.f32 %f293, [%rd450+8];
|
|
mul.wide.s32 %rd451, %r98, 12;
|
|
add.s64 %rd452, %rd448, %rd451;
|
|
st.global.f32 [%rd452], %f291;
|
|
st.global.f32 [%rd452+4], %f292;
|
|
st.global.f32 [%rd452+8], %f293;
|
|
|
|
$L__BB9_184:
|
|
ld.const.u64 %rd103, [kernel_params+2760];
|
|
setp.eq.s64 %p188, %rd103, 0;
|
|
@%p188 bra $L__BB9_186;
|
|
|
|
cvta.to.global.u64 %rd453, %rd103;
|
|
add.s64 %rd455, %rd453, %rd322;
|
|
ld.global.f32 %f294, [%rd455];
|
|
ld.global.f32 %f295, [%rd455+4];
|
|
ld.global.f32 %f296, [%rd455+8];
|
|
mul.wide.s32 %rd456, %r98, 12;
|
|
add.s64 %rd457, %rd453, %rd456;
|
|
st.global.f32 [%rd457], %f294;
|
|
st.global.f32 [%rd457+4], %f295;
|
|
st.global.f32 [%rd457+8], %f296;
|
|
|
|
$L__BB9_186:
|
|
ld.const.u64 %rd104, [kernel_params+2768];
|
|
setp.eq.s64 %p189, %rd104, 0;
|
|
@%p189 bra $L__BB9_188;
|
|
|
|
cvta.to.global.u64 %rd458, %rd104;
|
|
add.s64 %rd460, %rd458, %rd186;
|
|
ld.global.f32 %f297, [%rd460];
|
|
shl.b64 %rd461, %rd93, 2;
|
|
add.s64 %rd462, %rd458, %rd461;
|
|
st.global.f32 [%rd462], %f297;
|
|
|
|
$L__BB9_188:
|
|
ld.const.u64 %rd105, [kernel_params+2776];
|
|
setp.eq.s64 %p190, %rd105, 0;
|
|
@%p190 bra $L__BB9_190;
|
|
|
|
cvta.to.global.u64 %rd463, %rd105;
|
|
add.s64 %rd465, %rd463, %rd322;
|
|
ld.global.f32 %f298, [%rd465];
|
|
ld.global.f32 %f299, [%rd465+4];
|
|
ld.global.f32 %f300, [%rd465+8];
|
|
mul.lo.s64 %rd466, %rd93, 12;
|
|
add.s64 %rd467, %rd463, %rd466;
|
|
st.global.f32 [%rd467], %f298;
|
|
st.global.f32 [%rd467+4], %f299;
|
|
st.global.f32 [%rd467+8], %f300;
|
|
|
|
$L__BB9_190:
|
|
ld.const.u32 %r412, [kernel_params+28];
|
|
max.u32 %r99, %r412, 1;
|
|
add.s32 %r413, %r99, -1;
|
|
and.b32 %r600, %r99, 3;
|
|
setp.lt.u32 %p191, %r413, 3;
|
|
mov.u32 %r599, 0;
|
|
@%p191 bra $L__BB9_209;
|
|
|
|
sub.s32 %r598, %r99, %r600;
|
|
mov.u32 %r599, 0;
|
|
mov.u64 %rd469, kernel_params;
|
|
|
|
$L__BB9_192:
|
|
cvt.s64.s32 %rd106, %r599;
|
|
mul.wide.s32 %rd468, %r599, 16;
|
|
add.s64 %rd470, %rd469, %rd468;
|
|
add.s64 %rd107, %rd470, 2784;
|
|
ld.const.u64 %rd108, [%rd470+2784];
|
|
setp.eq.s64 %p192, %rd108, 0;
|
|
@%p192 bra $L__BB9_194;
|
|
|
|
cvta.to.global.u64 %rd471, %rd108;
|
|
add.s64 %rd473, %rd471, %rd186;
|
|
ld.global.u32 %r415, [%rd473];
|
|
shl.b64 %rd474, %rd93, 2;
|
|
add.s64 %rd475, %rd471, %rd474;
|
|
st.global.u32 [%rd475], %r415;
|
|
|
|
$L__BB9_194:
|
|
ld.const.u64 %rd109, [%rd107+8];
|
|
setp.eq.s64 %p193, %rd109, 0;
|
|
@%p193 bra $L__BB9_196;
|
|
|
|
cvta.to.global.u64 %rd476, %rd109;
|
|
add.s64 %rd478, %rd476, %rd186;
|
|
ld.global.u32 %r416, [%rd478];
|
|
shl.b64 %rd479, %rd93, 2;
|
|
add.s64 %rd480, %rd476, %rd479;
|
|
st.global.u32 [%rd480], %r416;
|
|
|
|
$L__BB9_196:
|
|
ld.const.u64 %rd110, [%rd107+16];
|
|
setp.eq.s64 %p194, %rd110, 0;
|
|
@%p194 bra $L__BB9_198;
|
|
|
|
cvta.to.global.u64 %rd481, %rd110;
|
|
add.s64 %rd483, %rd481, %rd186;
|
|
ld.global.u32 %r417, [%rd483];
|
|
shl.b64 %rd484, %rd93, 2;
|
|
add.s64 %rd485, %rd481, %rd484;
|
|
st.global.u32 [%rd485], %r417;
|
|
|
|
$L__BB9_198:
|
|
ld.const.u64 %rd111, [%rd107+24];
|
|
setp.eq.s64 %p195, %rd111, 0;
|
|
@%p195 bra $L__BB9_200;
|
|
|
|
cvta.to.global.u64 %rd486, %rd111;
|
|
add.s64 %rd488, %rd486, %rd186;
|
|
ld.global.u32 %r418, [%rd488];
|
|
shl.b64 %rd489, %rd93, 2;
|
|
add.s64 %rd490, %rd486, %rd489;
|
|
st.global.u32 [%rd490], %r418;
|
|
|
|
$L__BB9_200:
|
|
ld.const.u64 %rd112, [%rd107+32];
|
|
setp.eq.s64 %p196, %rd112, 0;
|
|
@%p196 bra $L__BB9_202;
|
|
|
|
cvta.to.global.u64 %rd491, %rd112;
|
|
add.s64 %rd493, %rd491, %rd186;
|
|
ld.global.u32 %r419, [%rd493];
|
|
shl.b64 %rd494, %rd93, 2;
|
|
add.s64 %rd495, %rd491, %rd494;
|
|
st.global.u32 [%rd495], %r419;
|
|
|
|
$L__BB9_202:
|
|
ld.const.u64 %rd113, [%rd107+40];
|
|
setp.eq.s64 %p197, %rd113, 0;
|
|
@%p197 bra $L__BB9_204;
|
|
|
|
cvta.to.global.u64 %rd496, %rd113;
|
|
add.s64 %rd498, %rd496, %rd186;
|
|
ld.global.u32 %r420, [%rd498];
|
|
shl.b64 %rd499, %rd93, 2;
|
|
add.s64 %rd500, %rd496, %rd499;
|
|
st.global.u32 [%rd500], %r420;
|
|
|
|
$L__BB9_204:
|
|
ld.const.u64 %rd114, [%rd107+48];
|
|
setp.eq.s64 %p198, %rd114, 0;
|
|
@%p198 bra $L__BB9_206;
|
|
|
|
cvta.to.global.u64 %rd501, %rd114;
|
|
add.s64 %rd503, %rd501, %rd186;
|
|
ld.global.u32 %r421, [%rd503];
|
|
shl.b64 %rd504, %rd93, 2;
|
|
add.s64 %rd505, %rd501, %rd504;
|
|
st.global.u32 [%rd505], %r421;
|
|
|
|
$L__BB9_206:
|
|
ld.const.u64 %rd115, [%rd107+56];
|
|
setp.eq.s64 %p199, %rd115, 0;
|
|
@%p199 bra $L__BB9_208;
|
|
|
|
cvta.to.global.u64 %rd506, %rd115;
|
|
add.s64 %rd508, %rd506, %rd186;
|
|
ld.global.u32 %r422, [%rd508];
|
|
shl.b64 %rd509, %rd93, 2;
|
|
add.s64 %rd510, %rd506, %rd509;
|
|
st.global.u32 [%rd510], %r422;
|
|
|
|
$L__BB9_208:
|
|
cvt.u32.u64 %r423, %rd106;
|
|
add.s32 %r599, %r423, 4;
|
|
add.s32 %r598, %r598, -4;
|
|
setp.ne.s32 %p200, %r598, 0;
|
|
@%p200 bra $L__BB9_192;
|
|
|
|
$L__BB9_209:
|
|
setp.eq.s32 %p201, %r600, 0;
|
|
@%p201 bra $L__BB9_216;
|
|
|
|
mul.wide.s32 %rd511, %r599, 16;
|
|
mov.u64 %rd512, kernel_params;
|
|
add.s64 %rd513, %rd512, %rd511;
|
|
add.s64 %rd803, %rd513, 2792;
|
|
|
|
$L__BB9_211:
|
|
.pragma "nounroll";
|
|
ld.const.u64 %rd119, [%rd803+-8];
|
|
setp.eq.s64 %p202, %rd119, 0;
|
|
@%p202 bra $L__BB9_213;
|
|
|
|
cvta.to.global.u64 %rd514, %rd119;
|
|
add.s64 %rd516, %rd514, %rd186;
|
|
ld.global.u32 %r424, [%rd516];
|
|
shl.b64 %rd517, %rd93, 2;
|
|
add.s64 %rd518, %rd514, %rd517;
|
|
st.global.u32 [%rd518], %r424;
|
|
|
|
$L__BB9_213:
|
|
ld.const.u64 %rd120, [%rd803];
|
|
setp.eq.s64 %p203, %rd120, 0;
|
|
@%p203 bra $L__BB9_215;
|
|
|
|
cvta.to.global.u64 %rd519, %rd120;
|
|
add.s64 %rd521, %rd519, %rd186;
|
|
ld.global.u32 %r425, [%rd521];
|
|
shl.b64 %rd522, %rd93, 2;
|
|
add.s64 %rd523, %rd519, %rd522;
|
|
st.global.u32 [%rd523], %r425;
|
|
|
|
$L__BB9_215:
|
|
add.s32 %r600, %r600, -1;
|
|
add.s64 %rd803, %rd803, 16;
|
|
setp.ne.s32 %p204, %r600, 0;
|
|
@%p204 bra $L__BB9_211;
|
|
|
|
$L__BB9_216:
|
|
ld.const.u64 %rd122, [kernel_params+3296];
|
|
setp.eq.s64 %p205, %rd122, 0;
|
|
@%p205 bra $L__BB9_218;
|
|
|
|
cvta.to.global.u64 %rd524, %rd122;
|
|
shl.b64 %rd525, %rd6, 3;
|
|
add.s64 %rd526, %rd524, %rd525;
|
|
ld.global.u64 %rd527, [%rd526];
|
|
shl.b64 %rd528, %rd93, 3;
|
|
add.s64 %rd529, %rd524, %rd528;
|
|
st.global.u64 [%rd529], %rd527;
|
|
|
|
$L__BB9_218:
|
|
ld.const.u64 %rd123, [kernel_params+3304];
|
|
setp.eq.s64 %p206, %rd123, 0;
|
|
@%p206 bra $L__BB9_220;
|
|
|
|
cvta.to.global.u64 %rd530, %rd123;
|
|
add.s64 %rd531, %rd530, %rd6;
|
|
ld.global.u8 %rs63, [%rd531];
|
|
add.s64 %rd532, %rd530, %rd93;
|
|
st.global.u8 [%rd532], %rs63;
|
|
|
|
$L__BB9_220:
|
|
ld.const.u64 %rd124, [kernel_params+3312];
|
|
setp.eq.s64 %p207, %rd124, 0;
|
|
@%p207 bra $L__BB9_222;
|
|
|
|
cvta.to.global.u64 %rd533, %rd124;
|
|
add.s64 %rd535, %rd533, %rd186;
|
|
ld.global.f32 %f301, [%rd535];
|
|
shl.b64 %rd536, %rd93, 2;
|
|
add.s64 %rd537, %rd533, %rd536;
|
|
st.global.f32 [%rd537], %f301;
|
|
|
|
$L__BB9_222:
|
|
ld.const.u64 %rd125, [kernel_params+3320];
|
|
setp.eq.s64 %p208, %rd125, 0;
|
|
@%p208 bra $L__BB9_224;
|
|
|
|
cvta.to.global.u64 %rd538, %rd125;
|
|
add.s64 %rd540, %rd538, %rd186;
|
|
ld.global.f32 %f302, [%rd540];
|
|
shl.b64 %rd541, %rd93, 2;
|
|
add.s64 %rd542, %rd538, %rd541;
|
|
st.global.f32 [%rd542], %f302;
|
|
|
|
$L__BB9_224:
|
|
ld.const.u64 %rd126, [kernel_params+3328];
|
|
setp.eq.s64 %p209, %rd126, 0;
|
|
@%p209 bra $L__BB9_226;
|
|
|
|
cvta.to.global.u64 %rd543, %rd126;
|
|
add.s64 %rd545, %rd543, %rd186;
|
|
ld.global.f32 %f303, [%rd545];
|
|
shl.b64 %rd546, %rd93, 2;
|
|
add.s64 %rd547, %rd543, %rd546;
|
|
st.global.f32 [%rd547], %f303;
|
|
|
|
$L__BB9_226:
|
|
ld.const.u64 %rd127, [kernel_params+3336];
|
|
setp.eq.s64 %p210, %rd127, 0;
|
|
@%p210 bra $L__BB9_228;
|
|
|
|
cvta.to.global.u64 %rd548, %rd127;
|
|
add.s64 %rd549, %rd548, %rd6;
|
|
ld.global.u8 %rs64, [%rd549];
|
|
add.s64 %rd550, %rd548, %rd93;
|
|
st.global.u8 [%rd550], %rs64;
|
|
|
|
$L__BB9_228:
|
|
ld.const.u64 %rd128, [kernel_params+3344];
|
|
setp.eq.s64 %p211, %rd128, 0;
|
|
@%p211 bra $L__BB9_230;
|
|
|
|
cvta.to.global.u64 %rd551, %rd128;
|
|
add.s64 %rd553, %rd551, %rd186;
|
|
ld.global.f32 %f304, [%rd553];
|
|
shl.b64 %rd554, %rd93, 2;
|
|
add.s64 %rd555, %rd551, %rd554;
|
|
st.global.f32 [%rd555], %f304;
|
|
|
|
$L__BB9_230:
|
|
ld.const.u64 %rd129, [kernel_params+3352];
|
|
setp.eq.s64 %p212, %rd129, 0;
|
|
@%p212 bra $L__BB9_232;
|
|
|
|
cvta.to.global.u64 %rd556, %rd129;
|
|
add.s64 %rd558, %rd556, %rd186;
|
|
ld.global.f32 %f305, [%rd558];
|
|
shl.b64 %rd559, %rd93, 2;
|
|
add.s64 %rd560, %rd556, %rd559;
|
|
st.global.f32 [%rd560], %f305;
|
|
|
|
$L__BB9_232:
|
|
ld.global.u32 %r109, [%rd94];
|
|
or.b32 %r426, %r109, 1073741824;
|
|
st.global.u32 [%rd94], %r426;
|
|
ld.const.u32 %r427, [kernel_params+1372];
|
|
setp.ne.s32 %p213, %r427, -1;
|
|
ld.const.u32 %r428, [kernel_params+1052];
|
|
setp.eq.s32 %p214, %r428, 0;
|
|
and.pred %p215, %p213, %p214;
|
|
@%p215 bra $L__BB9_247;
|
|
bra.uni $L__BB9_233;
|
|
|
|
$L__BB9_247:
|
|
or.b32 %r444, %r109, -1073741824;
|
|
st.global.u32 [%rd94], %r444;
|
|
cvta.to.global.u64 %rd619, %rd78;
|
|
add.s64 %rd620, %rd619, 24;
|
|
atom.global.add.u32 %r445, [%rd620], 1;
|
|
cvta.to.global.u64 %rd621, %rd80;
|
|
shl.b64 %rd622, %rd93, 1;
|
|
add.s64 %rd623, %rd621, %rd622;
|
|
mov.u16 %rs71, 6;
|
|
st.global.u16 [%rd623], %rs71;
|
|
bra.uni $L__BB9_283;
|
|
|
|
$L__BB9_233:
|
|
@%p112 bra $L__BB9_236;
|
|
|
|
ld.const.u64 %rd561, [kernel_params+2792];
|
|
cvta.to.global.u64 %rd562, %rd561;
|
|
shl.b64 %rd563, %rd93, 2;
|
|
add.s64 %rd564, %rd562, %rd563;
|
|
ld.global.u32 %r429, [%rd564];
|
|
setp.eq.s32 %p217, %r429, -1;
|
|
@%p217 bra $L__BB9_236;
|
|
|
|
cvta.to.global.u64 %rd565, %rd78;
|
|
add.s64 %rd566, %rd565, 20;
|
|
atom.global.add.u32 %r430, [%rd566], 1;
|
|
cvta.to.global.u64 %rd567, %rd80;
|
|
shl.b64 %rd568, %rd93, 1;
|
|
add.s64 %rd569, %rd567, %rd568;
|
|
mov.u16 %rs65, 5;
|
|
st.global.u16 [%rd569], %rs65;
|
|
bra.uni $L__BB9_283;
|
|
|
|
$L__BB9_236:
|
|
and.b32 %r573, %r69, 1;
|
|
setp.eq.s32 %p261, %r573, 0;
|
|
@%p261 bra $L__BB9_238;
|
|
bra.uni $L__BB9_237;
|
|
|
|
$L__BB9_238:
|
|
and.b32 %r431, %r69, 8;
|
|
setp.eq.s32 %p219, %r431, 0;
|
|
@%p219 bra $L__BB9_240;
|
|
bra.uni $L__BB9_239;
|
|
|
|
$L__BB9_240:
|
|
and.b32 %r433, %r69, 6;
|
|
setp.eq.s32 %p220, %r433, 0;
|
|
mov.u32 %r601, 0;
|
|
@%p220 bra $L__BB9_242;
|
|
|
|
ld.const.u64 %rd578, [kernel_params+2280];
|
|
cvta.to.global.u64 %rd579, %rd578;
|
|
mul.wide.s32 %rd580, %r71, 16;
|
|
add.s64 %rd581, %rd579, %rd580;
|
|
ld.global.u32 %r601, [%rd581];
|
|
bra.uni $L__BB9_242;
|
|
|
|
$L__BB9_237:
|
|
ld.const.u64 %rd570, [kernel_params+2232];
|
|
cvta.to.global.u64 %rd571, %rd570;
|
|
mul.wide.s32 %rd572, %r71, 4;
|
|
add.s64 %rd573, %rd571, %rd572;
|
|
ld.global.u32 %r601, [%rd573];
|
|
bra.uni $L__BB9_242;
|
|
|
|
$L__BB9_239:
|
|
ld.const.u64 %rd574, [kernel_params+2320];
|
|
cvta.to.global.u64 %rd575, %rd574;
|
|
mul.wide.s32 %rd576, %r71, 4;
|
|
add.s64 %rd577, %rd575, %rd576;
|
|
ld.global.u32 %r601, [%rd577];
|
|
|
|
$L__BB9_242:
|
|
ld.const.u32 %r574, [kernel_params+1608];
|
|
setp.ne.s32 %p262, %r574, 0;
|
|
ld.const.u64 %rd749, [kernel_params+2424];
|
|
cvta.to.global.u64 %rd748, %rd749;
|
|
and.b32 %r434, %r601, 4194303;
|
|
shl.b32 %r435, %r601, 5;
|
|
cvt.u64.u32 %rd582, %r435;
|
|
and.b64 %rd583, %rd582, 134217696;
|
|
add.s64 %rd584, %rd748, %rd583;
|
|
add.s64 %rd130, %rd584, 16;
|
|
and.b32 %r436, %r96, 1536;
|
|
setp.ne.s32 %p221, %r436, 0;
|
|
and.pred %p223, %p221, %p262;
|
|
div.u32 %r437, %r98, %r595;
|
|
mad.lo.s32 %r114, %r437, %r596, %r434;
|
|
cvt.s64.s32 %rd131, %r114;
|
|
@%p223 bra $L__BB9_246;
|
|
bra.uni $L__BB9_243;
|
|
|
|
$L__BB9_246:
|
|
cvta.to.global.u64 %rd607, %rd78;
|
|
add.s64 %rd608, %rd607, 40;
|
|
atom.global.add.u32 %r442, [%rd608], 1;
|
|
cvta.to.global.u64 %rd609, %rd80;
|
|
shl.b64 %rd610, %rd93, 1;
|
|
add.s64 %rd611, %rd609, %rd610;
|
|
mov.u16 %rs70, 10;
|
|
st.global.u16 [%rd611], %rs70;
|
|
cvta.to.global.u64 %rd612, %rd79;
|
|
shl.b64 %rd613, %rd93, 2;
|
|
add.s64 %rd614, %rd612, %rd613;
|
|
st.global.u32 [%rd614], %r114;
|
|
ld.const.u64 %rd615, [kernel_params+4360];
|
|
cvta.to.global.u64 %rd616, %rd615;
|
|
shl.b64 %rd617, %rd131, 2;
|
|
add.s64 %rd618, %rd616, %rd617;
|
|
atom.global.add.u32 %r443, [%rd618], 1;
|
|
bra.uni $L__BB9_283;
|
|
|
|
$L__BB9_243:
|
|
ld.global.u8 %rs66, [%rd130+3];
|
|
and.b16 %rs67, %rs66, 64;
|
|
setp.eq.s16 %p224, %rs67, 0;
|
|
cvta.to.global.u64 %rd132, %rd78;
|
|
@%p224 bra $L__BB9_245;
|
|
bra.uni $L__BB9_244;
|
|
|
|
$L__BB9_245:
|
|
add.s64 %rd596, %rd132, 32;
|
|
atom.global.add.u32 %r440, [%rd596], 1;
|
|
cvta.to.global.u64 %rd597, %rd80;
|
|
shl.b64 %rd598, %rd93, 1;
|
|
add.s64 %rd599, %rd597, %rd598;
|
|
mov.u16 %rs69, 8;
|
|
st.global.u16 [%rd599], %rs69;
|
|
cvta.to.global.u64 %rd600, %rd79;
|
|
shl.b64 %rd601, %rd93, 2;
|
|
add.s64 %rd602, %rd600, %rd601;
|
|
st.global.u32 [%rd602], %r114;
|
|
ld.const.u64 %rd603, [kernel_params+4344];
|
|
cvta.to.global.u64 %rd604, %rd603;
|
|
shl.b64 %rd605, %rd131, 2;
|
|
add.s64 %rd606, %rd604, %rd605;
|
|
atom.global.add.u32 %r441, [%rd606], 1;
|
|
bra.uni $L__BB9_283;
|
|
|
|
$L__BB9_244:
|
|
add.s64 %rd585, %rd132, 36;
|
|
atom.global.add.u32 %r438, [%rd585], 1;
|
|
cvta.to.global.u64 %rd586, %rd80;
|
|
shl.b64 %rd587, %rd93, 1;
|
|
add.s64 %rd588, %rd586, %rd587;
|
|
mov.u16 %rs68, 9;
|
|
st.global.u16 [%rd588], %rs68;
|
|
cvta.to.global.u64 %rd589, %rd79;
|
|
shl.b64 %rd590, %rd93, 2;
|
|
add.s64 %rd591, %rd589, %rd590;
|
|
st.global.u32 [%rd591], %r114;
|
|
ld.const.u64 %rd592, [kernel_params+4352];
|
|
cvta.to.global.u64 %rd593, %rd592;
|
|
shl.b64 %rd594, %rd131, 2;
|
|
add.s64 %rd595, %rd593, %rd594;
|
|
atom.global.add.u32 %r439, [%rd595], 1;
|
|
bra.uni $L__BB9_283;
|
|
|
|
}
|
|
// .globl __raygen__kernel_optix_integrator_intersect_shadow
|
|
.visible .entry __raygen__kernel_optix_integrator_intersect_shadow()
|
|
{
|
|
.local .align 4 .b8 __local_depot10[60];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<59>;
|
|
.reg .b16 %rs<17>;
|
|
.reg .f32 %f<53>;
|
|
.reg .b32 %r<273>;
|
|
.reg .b64 %rd<142>;
|
|
|
|
|
|
mov.u64 %SPL, __local_depot10;
|
|
cvta.local.u64 %SP, %SPL;
|
|
// begin inline asm
|
|
call (%r270), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
ld.const.u64 %rd1, [kernel_params];
|
|
setp.eq.s64 %p6, %rd1, 0;
|
|
@%p6 bra $L__BB10_2;
|
|
|
|
cvta.to.global.u64 %rd11, %rd1;
|
|
mul.wide.s32 %rd12, %r270, 4;
|
|
add.s64 %rd13, %rd11, %rd12;
|
|
ld.global.u32 %r270, [%rd13];
|
|
|
|
$L__BB10_2:
|
|
ld.const.u64 %rd14, [kernel_params+3512];
|
|
cvta.to.global.u64 %rd15, %rd14;
|
|
cvt.s64.s32 %rd2, %r270;
|
|
mul.wide.s32 %rd16, %r270, 12;
|
|
add.s64 %rd17, %rd15, %rd16;
|
|
ld.global.f32 %f1, [%rd17];
|
|
ld.global.f32 %f2, [%rd17+4];
|
|
ld.global.f32 %f3, [%rd17+8];
|
|
add.u64 %rd18, %SP, 0;
|
|
add.u64 %rd19, %SPL, 0;
|
|
st.local.f32 [%rd19], %f1;
|
|
st.local.f32 [%rd19+4], %f2;
|
|
st.local.f32 [%rd19+8], %f3;
|
|
ld.const.v2.u64 {%rd20, %rd21}, [kernel_params+3520];
|
|
mov.u32 %r50, 0;
|
|
cvta.to.global.u64 %rd24, %rd20;
|
|
add.s64 %rd25, %rd24, %rd16;
|
|
ld.global.f32 %f4, [%rd25];
|
|
ld.global.f32 %f5, [%rd25+4];
|
|
ld.global.f32 %f6, [%rd25+8];
|
|
st.local.f32 [%rd19+12], %f4;
|
|
st.local.f32 [%rd19+16], %f5;
|
|
st.local.f32 [%rd19+20], %f6;
|
|
cvta.to.global.u64 %rd26, %rd21;
|
|
mul.wide.s32 %rd27, %r270, 4;
|
|
add.s64 %rd28, %rd26, %rd27;
|
|
ld.global.f32 %f7, [%rd28];
|
|
st.local.f32 [%rd19+24], %f7;
|
|
ld.const.v2.u64 {%rd29, %rd30}, [kernel_params+3536];
|
|
cvta.to.global.u64 %rd33, %rd29;
|
|
add.s64 %rd34, %rd33, %rd27;
|
|
ld.global.f32 %f8, [%rd34];
|
|
st.local.f32 [%rd19+28], %f8;
|
|
cvta.to.global.u64 %rd35, %rd30;
|
|
add.s64 %rd36, %rd35, %rd27;
|
|
ld.global.f32 %f9, [%rd36];
|
|
st.local.f32 [%rd19+32], %f9;
|
|
ld.const.u64 %rd37, [kernel_params+3552];
|
|
cvta.to.global.u64 %rd38, %rd37;
|
|
add.s64 %rd39, %rd38, %rd27;
|
|
ld.global.f32 %f12, [%rd39];
|
|
st.local.f32 [%rd19+52], %f12;
|
|
st.local.u32 [%rd19+56], %r50;
|
|
ld.const.u64 %rd40, [kernel_params+3600];
|
|
cvta.to.global.u64 %rd41, %rd40;
|
|
add.s64 %rd42, %rd41, %rd27;
|
|
ld.global.u32 %r51, [%rd42];
|
|
st.local.u32 [%rd19+40], %r51;
|
|
ld.const.u64 %rd43, [kernel_params+3592];
|
|
cvta.to.global.u64 %rd44, %rd43;
|
|
add.s64 %rd45, %rd44, %rd27;
|
|
ld.global.u32 %r52, [%rd45];
|
|
st.local.u32 [%rd19+36], %r52;
|
|
ld.const.u64 %rd46, [kernel_params+3648];
|
|
cvta.to.global.u64 %rd47, %rd46;
|
|
add.s64 %rd48, %rd47, %rd27;
|
|
ld.global.u32 %r53, [%rd48];
|
|
st.local.u32 [%rd19+48], %r53;
|
|
ld.const.u64 %rd49, [kernel_params+3640];
|
|
cvta.to.global.u64 %rd50, %rd49;
|
|
add.s64 %rd51, %rd50, %rd27;
|
|
ld.global.u32 %r54, [%rd51];
|
|
st.local.u32 [%rd19+44], %r54;
|
|
ld.const.u64 %rd52, [kernel_params+3440];
|
|
cvta.to.global.u64 %rd53, %rd52;
|
|
add.s64 %rd54, %rd53, %rd27;
|
|
ld.global.u8 %rs4, [%rd54+3];
|
|
and.b16 %rs5, %rs4, 64;
|
|
setp.eq.s16 %p7, %rs5, 0;
|
|
selp.b32 %r4, 768, 50331648, %p7;
|
|
ld.const.u32 %r55, [kernel_params+1576];
|
|
setp.eq.s32 %p8, %r55, 0;
|
|
cvt.u32.u64 %r5, %rd18;
|
|
shr.u64 %rd55, %rd18, 32;
|
|
cvt.u32.u64 %r6, %rd55;
|
|
mov.b32 %r7, %f1;
|
|
@%p8 bra $L__BB10_21;
|
|
|
|
setp.nan.ftz.f32 %p10, %f1, %f1;
|
|
ld.const.u64 %rd56, [kernel_params+3400];
|
|
cvta.to.global.u64 %rd57, %rd56;
|
|
shl.b64 %rd58, %rd2, 1;
|
|
add.s64 %rd59, %rd57, %rd58;
|
|
ld.global.u16 %r56, [%rd59];
|
|
not.b32 %r57, %r56;
|
|
ld.const.u32 %r58, [kernel_params+1572];
|
|
add.s32 %r59, %r58, %r57;
|
|
max.s32 %r8, %r59, 0;
|
|
mov.pred %p56, -1;
|
|
@%p10 bra $L__BB10_8;
|
|
|
|
or.b32 %r60, %r7, -2147483648;
|
|
setp.ne.s32 %p12, %r60, -2147483648;
|
|
add.ftz.f32 %f13, %f1, %f1;
|
|
setp.eq.ftz.f32 %p13, %f13, %f1;
|
|
and.pred %p14, %p12, %p13;
|
|
shl.b32 %r61, %r7, 1;
|
|
setp.gt.u32 %p15, %r61, -16777216;
|
|
or.pred %p16, %p15, %p14;
|
|
@%p16 bra $L__BB10_8;
|
|
|
|
setp.nan.ftz.f32 %p18, %f4, %f4;
|
|
@%p18 bra $L__BB10_8;
|
|
|
|
mov.b32 %r62, %f4;
|
|
or.b32 %r63, %r62, -2147483648;
|
|
setp.ne.s32 %p20, %r63, -2147483648;
|
|
add.ftz.f32 %f14, %f4, %f4;
|
|
setp.eq.ftz.f32 %p21, %f14, %f4;
|
|
and.pred %p22, %p20, %p21;
|
|
shl.b32 %r64, %r62, 1;
|
|
setp.gt.u32 %p23, %r64, -16777216;
|
|
or.pred %p24, %p23, %p22;
|
|
@%p24 bra $L__BB10_8;
|
|
|
|
mul.ftz.f32 %f15, %f5, %f5;
|
|
fma.rn.ftz.f32 %f16, %f4, %f4, %f15;
|
|
fma.rn.ftz.f32 %f17, %f6, %f6, %f16;
|
|
setp.eq.ftz.f32 %p56, %f17, 0f00000000;
|
|
|
|
$L__BB10_8:
|
|
ld.const.u64 %rd61, [kernel_params+2096];
|
|
selp.b64 %rd60, 0, %rd61, %p56;
|
|
cvt.u32.u64 %r104, %rd2;
|
|
mov.u32 %r98, 255;
|
|
mov.u32 %r99, 2;
|
|
mov.u32 %r100, 1;
|
|
mov.u32 %r103, 8;
|
|
mov.u32 %r105, 1065353216;
|
|
// begin inline asm
|
|
call(%r65,%r66,%r67,%r68,%r69,%r70,%r71,%r72,%r73,%r74,%r75,%r76,%r77,%r78,%r79,%r80,%r81,%r82,%r83,%r84,%r85,%r86,%r87,%r88,%r89,%r90,%r91,%r92,%r93,%r94,%r95,%r96),_optix_trace_typed_32,(%r50,%rd60,%f1,%f2,%f3,%f4,%f5,%f6,%f7,%f8,%f9,%r98,%r99,%r100,%r50,%r50,%r103,%r104,%r105,%r50,%r8,%r4,%r50,%r5,%r6,%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159);
|
|
// end inline asm
|
|
shr.u32 %r41, %r67, 16;
|
|
mov.b32 %f10, %r66;
|
|
setp.eq.ftz.f32 %p25, %f10, 0f3F800000;
|
|
@%p25 bra $L__BB10_10;
|
|
|
|
ld.const.u64 %rd62, [kernel_params+3448];
|
|
cvta.to.global.u64 %rd63, %rd62;
|
|
mul.lo.s64 %rd64, %rd2, 12;
|
|
add.s64 %rd65, %rd63, %rd64;
|
|
ld.global.f32 %f27, [%rd65];
|
|
mul.ftz.f32 %f28, %f27, %f10;
|
|
ld.global.f32 %f29, [%rd65+4];
|
|
mul.ftz.f32 %f30, %f29, %f10;
|
|
ld.global.f32 %f31, [%rd65+8];
|
|
mul.ftz.f32 %f32, %f31, %f10;
|
|
st.global.f32 [%rd65], %f28;
|
|
st.global.f32 [%rd65+4], %f30;
|
|
st.global.f32 [%rd65+8], %f32;
|
|
|
|
$L__BB10_10:
|
|
setp.ne.s32 %p26, %r70, 0;
|
|
setp.gt.u32 %p27, %r41, %r8;
|
|
or.pred %p28, %p26, %p27;
|
|
@%p28 bra $L__BB10_20;
|
|
bra.uni $L__BB10_11;
|
|
|
|
$L__BB10_20:
|
|
ld.const.u64 %rd119, [kernel_params+3480];
|
|
cvta.to.global.u64 %rd120, %rd119;
|
|
add.s64 %rd122, %rd120, %rd58;
|
|
mov.u16 %rs10, 0;
|
|
st.global.u16 [%rd122], %rs10;
|
|
mov.pred %p58, -1;
|
|
bra.uni $L__BB10_28;
|
|
|
|
$L__BB10_21:
|
|
setp.nan.ftz.f32 %p37, %f1, %f1;
|
|
mov.pred %p58, -1;
|
|
mov.pred %p57, %p58;
|
|
@%p37 bra $L__BB10_26;
|
|
|
|
or.b32 %r168, %r7, -2147483648;
|
|
setp.ne.s32 %p39, %r168, -2147483648;
|
|
add.ftz.f32 %f39, %f1, %f1;
|
|
setp.eq.ftz.f32 %p40, %f39, %f1;
|
|
and.pred %p41, %p39, %p40;
|
|
shl.b32 %r169, %r7, 1;
|
|
setp.gt.u32 %p42, %r169, -16777216;
|
|
or.pred %p43, %p42, %p41;
|
|
mov.pred %p57, %p58;
|
|
@%p43 bra $L__BB10_26;
|
|
|
|
setp.nan.ftz.f32 %p45, %f4, %f4;
|
|
mov.pred %p57, %p58;
|
|
@%p45 bra $L__BB10_26;
|
|
|
|
mov.b32 %r170, %f4;
|
|
or.b32 %r171, %r170, -2147483648;
|
|
setp.ne.s32 %p47, %r171, -2147483648;
|
|
add.ftz.f32 %f40, %f4, %f4;
|
|
setp.eq.ftz.f32 %p48, %f40, %f4;
|
|
and.pred %p49, %p47, %p48;
|
|
shl.b32 %r172, %r170, 1;
|
|
setp.gt.u32 %p50, %r172, -16777216;
|
|
or.pred %p51, %p50, %p49;
|
|
mov.pred %p57, %p58;
|
|
@%p51 bra $L__BB10_26;
|
|
|
|
mul.ftz.f32 %f41, %f5, %f5;
|
|
fma.rn.ftz.f32 %f42, %f4, %f4, %f41;
|
|
fma.rn.ftz.f32 %f43, %f6, %f6, %f42;
|
|
setp.eq.ftz.f32 %p57, %f43, 0f00000000;
|
|
|
|
$L__BB10_26:
|
|
and.b32 %r216, %r4, 16777472;
|
|
ld.const.u64 %rd124, [kernel_params+2096];
|
|
selp.b64 %rd123, 0, %rd124, %p57;
|
|
setp.eq.s32 %p53, %r216, 0;
|
|
mov.u32 %r217, 0;
|
|
selp.b32 %r206, 0, 255, %p53;
|
|
mov.u32 %r207, 2;
|
|
mov.u32 %r211, 8;
|
|
// begin inline asm
|
|
call(%r173,%r174,%r175,%r176,%r177,%r178,%r179,%r180,%r181,%r182,%r183,%r184,%r185,%r186,%r187,%r188,%r189,%r190,%r191,%r192,%r193,%r194,%r195,%r196,%r197,%r198,%r199,%r200,%r201,%r202,%r203,%r204),_optix_trace_typed_32,(%r217,%rd123,%f1,%f2,%f3,%f4,%f5,%f6,%f7,%f8,%f9,%r206,%r207,%r217,%r217,%r217,%r211,%r217,%r217,%r217,%r217,%r216,%r217,%r5,%r6,%r244,%r245,%r246,%r247,%r248,%r249,%r250,%r251,%r252,%r253,%r254,%r255,%r256,%r257,%r258,%r259,%r260,%r261,%r262,%r263,%r264,%r265,%r266,%r267);
|
|
// end inline asm
|
|
setp.ne.s32 %p54, %r178, 0;
|
|
@%p54 bra $L__BB10_28;
|
|
|
|
ld.const.u64 %rd125, [kernel_params+3480];
|
|
cvta.to.global.u64 %rd126, %rd125;
|
|
shl.b64 %rd127, %rd2, 1;
|
|
add.s64 %rd128, %rd126, %rd127;
|
|
mov.u16 %rs11, 0;
|
|
st.global.u16 [%rd128], %rs11;
|
|
mov.pred %p58, 0;
|
|
bra.uni $L__BB10_28;
|
|
|
|
$L__BB10_11:
|
|
min.u32 %r160, %r8, 4;
|
|
min.u32 %r271, %r41, %r160;
|
|
setp.eq.s32 %p29, %r271, 0;
|
|
@%p29 bra $L__BB10_19;
|
|
|
|
ld.const.u64 %rd3, [kernel_params+3568];
|
|
mov.u16 %rs6, 0;
|
|
mov.u64 %rd66, kernel_params;
|
|
shl.b64 %rd68, %rd2, 2;
|
|
mov.u16 %rs8, 1;
|
|
|
|
$L__BB10_13:
|
|
add.s32 %r271, %r271, -1;
|
|
setp.eq.s32 %p30, %r271, 0;
|
|
mov.u16 %rs15, %rs6;
|
|
@%p30 bra $L__BB10_18;
|
|
|
|
mov.u64 %rd140, %rd66;
|
|
mov.u64 %rd141, %rd3;
|
|
mov.u32 %r272, %r50;
|
|
mov.u16 %rs15, %rs6;
|
|
|
|
$L__BB10_15:
|
|
cvta.to.global.u64 %rd67, %rd141;
|
|
add.s64 %rd69, %rd67, %rd68;
|
|
add.s32 %r272, %r272, 1;
|
|
ld.const.u64 %rd141, [%rd140+3616];
|
|
cvta.to.global.u64 %rd70, %rd141;
|
|
add.s64 %rd8, %rd70, %rd68;
|
|
ld.global.f32 %f11, [%rd8];
|
|
ld.global.f32 %f33, [%rd69];
|
|
setp.leu.ftz.f32 %p31, %f33, %f11;
|
|
@%p31 bra $L__BB10_17;
|
|
|
|
ld.const.v2.u64 {%rd71, %rd72}, [%rd140+3584];
|
|
cvta.to.global.u64 %rd75, %rd72;
|
|
add.s64 %rd77, %rd75, %rd68;
|
|
ld.global.u32 %r162, [%rd77];
|
|
ld.const.v2.u64 {%rd78, %rd79}, [%rd140+3600];
|
|
cvta.to.global.u64 %rd82, %rd78;
|
|
add.s64 %rd83, %rd82, %rd68;
|
|
ld.global.u32 %r163, [%rd83];
|
|
cvta.to.global.u64 %rd84, %rd79;
|
|
add.s64 %rd85, %rd84, %rd68;
|
|
ld.global.u32 %r164, [%rd85];
|
|
ld.const.v2.u64 {%rd86, %rd87}, [%rd140+3568];
|
|
cvta.to.global.u64 %rd90, %rd87;
|
|
add.s64 %rd91, %rd90, %rd68;
|
|
ld.global.f32 %f34, [%rd91];
|
|
cvta.to.global.u64 %rd92, %rd71;
|
|
add.s64 %rd93, %rd92, %rd68;
|
|
ld.global.f32 %f35, [%rd93];
|
|
cvta.to.global.u64 %rd94, %rd86;
|
|
add.s64 %rd95, %rd94, %rd68;
|
|
ld.global.f32 %f36, [%rd95];
|
|
ld.const.v2.u64 {%rd96, %rd97}, [%rd140+3632];
|
|
cvta.to.global.u64 %rd100, %rd97;
|
|
add.s64 %rd101, %rd100, %rd68;
|
|
ld.global.u32 %r165, [%rd101];
|
|
ld.const.v2.u64 {%rd102, %rd103}, [%rd140+3648];
|
|
cvta.to.global.u64 %rd106, %rd102;
|
|
add.s64 %rd107, %rd106, %rd68;
|
|
ld.global.u32 %r166, [%rd107];
|
|
cvta.to.global.u64 %rd108, %rd103;
|
|
add.s64 %rd109, %rd108, %rd68;
|
|
ld.global.u32 %r167, [%rd109];
|
|
ld.const.u64 %rd110, [%rd140+3624];
|
|
cvta.to.global.u64 %rd111, %rd110;
|
|
add.s64 %rd112, %rd111, %rd68;
|
|
ld.global.f32 %f37, [%rd112];
|
|
cvta.to.global.u64 %rd113, %rd96;
|
|
add.s64 %rd114, %rd113, %rd68;
|
|
ld.global.f32 %f38, [%rd114];
|
|
st.global.f32 [%rd95], %f11;
|
|
st.global.f32 [%rd91], %f37;
|
|
st.global.f32 [%rd93], %f38;
|
|
st.global.u32 [%rd83], %r166;
|
|
st.global.u32 [%rd77], %r165;
|
|
st.global.u32 [%rd85], %r167;
|
|
st.global.f32 [%rd8], %f36;
|
|
st.global.f32 [%rd112], %f34;
|
|
st.global.f32 [%rd114], %f35;
|
|
st.global.u32 [%rd107], %r163;
|
|
st.global.u32 [%rd101], %r162;
|
|
st.global.u32 [%rd109], %r164;
|
|
mov.u16 %rs15, %rs8;
|
|
|
|
$L__BB10_17:
|
|
add.s64 %rd140, %rd140, 48;
|
|
setp.lt.u32 %p32, %r272, %r271;
|
|
@%p32 bra $L__BB10_15;
|
|
|
|
$L__BB10_18:
|
|
and.b16 %rs9, %rs15, 255;
|
|
setp.ne.s16 %p33, %rs9, 0;
|
|
@%p33 bra $L__BB10_13;
|
|
|
|
$L__BB10_19:
|
|
ld.const.u64 %rd115, [kernel_params+3480];
|
|
cvta.to.global.u64 %rd116, %rd115;
|
|
add.s64 %rd118, %rd116, %rd58;
|
|
st.global.u16 [%rd118], %r41;
|
|
mov.pred %p58, 0;
|
|
|
|
$L__BB10_28:
|
|
ld.const.u64 %rd129, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd10, %rd129;
|
|
add.s64 %rd130, %rd10, 12;
|
|
atom.global.add.u32 %r268, [%rd130], -1;
|
|
@%p58 bra $L__BB10_30;
|
|
bra.uni $L__BB10_29;
|
|
|
|
$L__BB10_30:
|
|
ld.const.u64 %rd136, [kernel_params+3432];
|
|
cvta.to.global.u64 %rd137, %rd136;
|
|
shl.b64 %rd138, %rd2, 1;
|
|
add.s64 %rd139, %rd137, %rd138;
|
|
mov.u16 %rs13, 0;
|
|
st.global.u16 [%rd139], %rs13;
|
|
bra.uni $L__BB10_31;
|
|
|
|
$L__BB10_29:
|
|
add.s64 %rd131, %rd10, 48;
|
|
atom.global.add.u32 %r269, [%rd131], 1;
|
|
ld.const.u64 %rd132, [kernel_params+3432];
|
|
cvta.to.global.u64 %rd133, %rd132;
|
|
shl.b64 %rd134, %rd2, 1;
|
|
add.s64 %rd135, %rd133, %rd134;
|
|
mov.u16 %rs12, 12;
|
|
st.global.u16 [%rd135], %rs12;
|
|
|
|
$L__BB10_31:
|
|
ret;
|
|
|
|
}
|
|
// .globl __raygen__kernel_optix_integrator_intersect_subsurface
|
|
.visible .entry __raygen__kernel_optix_integrator_intersect_subsurface()
|
|
{
|
|
.local .align 16 .b8 __local_depot11[576];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<376>;
|
|
.reg .b16 %rs<114>;
|
|
.reg .f32 %f<4005>;
|
|
.reg .b32 %r<1529>;
|
|
.reg .b64 %rd<718>;
|
|
|
|
|
|
mov.u64 %SPL, __local_depot11;
|
|
cvta.local.u64 %SP, %SPL;
|
|
add.u64 %rd145, %SP, 0;
|
|
add.u64 %rd1, %SPL, 0;
|
|
add.u64 %rd146, %SP, 424;
|
|
add.u64 %rd2, %SPL, 424;
|
|
// begin inline asm
|
|
call (%r1488), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
ld.const.u64 %rd3, [kernel_params];
|
|
setp.eq.s64 %p13, %rd3, 0;
|
|
@%p13 bra $L__BB11_2;
|
|
|
|
cvta.to.global.u64 %rd147, %rd3;
|
|
mul.wide.s32 %rd148, %r1488, 4;
|
|
add.s64 %rd149, %rd147, %rd148;
|
|
ld.global.u32 %r1488, [%rd149];
|
|
|
|
$L__BB11_2:
|
|
ld.const.v2.u64 {%rd150, %rd151}, [kernel_params+2544];
|
|
mov.u32 %r216, 0;
|
|
cvta.to.global.u64 %rd154, %rd150;
|
|
cvt.s64.s32 %rd4, %r1488;
|
|
mul.wide.s32 %rd155, %r1488, 4;
|
|
add.s64 %rd156, %rd154, %rd155;
|
|
ld.global.u32 %r4, [%rd156];
|
|
cvta.to.global.u64 %rd157, %rd151;
|
|
mul.wide.s32 %rd158, %r1488, 2;
|
|
add.s64 %rd5, %rd157, %rd158;
|
|
ld.global.u16 %rs1, [%rd5];
|
|
cvt.u32.u16 %r1495, %rs1;
|
|
ld.const.u64 %rd159, [kernel_params+2472];
|
|
cvta.to.global.u64 %rd160, %rd159;
|
|
add.s64 %rd161, %rd160, %rd155;
|
|
ld.global.u32 %r6, [%rd161];
|
|
add.u64 %rd162, %SP, 364;
|
|
add.u64 %rd6, %SPL, 364;
|
|
st.local.u32 [%rd6], %r216;
|
|
st.local.u32 [%rd6+4], %r216;
|
|
st.local.u32 [%rd6+8], %r216;
|
|
st.local.u32 [%rd6+12], %r216;
|
|
st.local.u32 [%rd6+16], %r216;
|
|
st.local.u32 [%rd6+20], %r216;
|
|
st.local.u32 [%rd6+24], %r216;
|
|
st.local.u32 [%rd6+28], %r216;
|
|
st.local.u32 [%rd6+32], %r216;
|
|
st.local.u32 [%rd6+36], %r216;
|
|
st.local.u32 [%rd6+40], %r216;
|
|
st.local.u32 [%rd6+44], %r216;
|
|
st.local.u32 [%rd6+48], %r216;
|
|
st.local.u32 [%rd6+52], %r216;
|
|
st.local.u32 [%rd6+56], %r216;
|
|
cvta.to.local.u64 %rd697, %rd146;
|
|
add.u64 %rd164, %SP, 304;
|
|
add.u64 %rd8, %SPL, 304;
|
|
mov.u32 %r1489, %r216;
|
|
|
|
$L__BB11_3:
|
|
st.local.u32 [%rd697], %r216;
|
|
add.s64 %rd697, %rd697, 4;
|
|
add.s32 %r1489, %r1489, 1;
|
|
setp.lt.u32 %p14, %r1489, 37;
|
|
@%p14 bra $L__BB11_3;
|
|
|
|
ld.const.u64 %rd166, [kernel_params+2560];
|
|
cvta.to.global.u64 %rd167, %rd166;
|
|
shl.b64 %rd168, %rd4, 2;
|
|
add.s64 %rd12, %rd167, %rd168;
|
|
ld.global.u32 %r9, [%rd12];
|
|
and.b32 %r218, %r9, 1048576;
|
|
setp.eq.s32 %p15, %r218, 0;
|
|
brev.b32 %r219, %r6;
|
|
mul.lo.s32 %r220, %r219, 1025551850;
|
|
xor.b32 %r10, %r220, %r219;
|
|
@%p15 bra $L__BB11_116;
|
|
bra.uni $L__BB11_5;
|
|
|
|
$L__BB11_116:
|
|
ld.const.u32 %r63, [kernel_params+1612];
|
|
setp.eq.s32 %p104, %r63, 0;
|
|
@%p104 bra $L__BB11_119;
|
|
|
|
ld.const.v2.u32 {%r888, %r889}, [kernel_params+1616];
|
|
mov.b32 %f294, %r889;
|
|
setp.geu.ftz.f32 %p105, %f294, 0f3F800000;
|
|
setp.lt.ftz.f32 %p106, %f294, 0f3F800000;
|
|
ld.const.u32 %r892, [kernel_params+1592];
|
|
selp.b32 %r893, %r892, %r4, %p106;
|
|
and.b32 %r894, %r1495, 255;
|
|
xor.b32 %r895, %r893, %r894;
|
|
mul.lo.s32 %r896, %r895, -512718531;
|
|
shr.u32 %r897, %r893, 16;
|
|
xor.b32 %r898, %r896, %r897;
|
|
bfe.u32 %r899, %r898, 4, 4;
|
|
shr.u32 %r900, %r893, 8;
|
|
xor.b32 %r901, %r898, %r900;
|
|
xor.b32 %r902, %r901, %r899;
|
|
mul.lo.s32 %r903, %r902, 153742143;
|
|
shr.u32 %r904, %r893, 23;
|
|
xor.b32 %r905, %r903, %r904;
|
|
bfe.u32 %r906, %r905, 1, 7;
|
|
xor.b32 %r907, %r906, %r905;
|
|
shr.u32 %r908, %r893, 27;
|
|
or.b32 %r909, %r908, 1;
|
|
mul.lo.s32 %r910, %r909, %r907;
|
|
mul.lo.s32 %r911, %r910, -1074439621;
|
|
bfe.u32 %r912, %r911, 2, 6;
|
|
xor.b32 %r913, %r912, %r911;
|
|
mul.lo.s32 %r914, %r913, -1638916925;
|
|
bfe.u32 %r915, %r914, 2, 6;
|
|
xor.b32 %r916, %r915, %r914;
|
|
mul.lo.s32 %r917, %r916, -933190689;
|
|
and.b32 %r918, %r917, 255;
|
|
bfe.u32 %r919, %r917, 5, 3;
|
|
xor.b32 %r920, %r919, %r918;
|
|
add.s32 %r921, %r888, -1;
|
|
xor.b32 %r64, %r1495, 61;
|
|
xor.b32 %r922, %r893, %r64;
|
|
mul.lo.s32 %r923, %r922, 9;
|
|
shr.u32 %r924, %r923, 4;
|
|
xor.b32 %r925, %r924, %r923;
|
|
mul.lo.s32 %r926, %r925, 668265261;
|
|
add.s32 %r927, %r926, %r10;
|
|
shr.u32 %r928, %r926, 16;
|
|
or.b32 %r929, %r928, 1;
|
|
mul.lo.s32 %r930, %r929, %r927;
|
|
mul.lo.s32 %r931, %r930, 89287766;
|
|
xor.b32 %r932, %r931, %r930;
|
|
mul.lo.s32 %r933, %r932, 1403136100;
|
|
xor.b32 %r934, %r933, %r932;
|
|
brev.b32 %r935, %r934;
|
|
neg.s32 %r936, %r888;
|
|
and.b32 %r937, %r6, %r936;
|
|
and.b32 %r938, %r935, %r921;
|
|
or.b32 %r939, %r938, %r937;
|
|
mad.lo.s32 %r940, %r920, %r888, %r939;
|
|
shl.b32 %r941, %r888, 8;
|
|
rem.u32 %r942, %r940, %r941;
|
|
ld.const.u64 %rd235, [kernel_params+2440];
|
|
cvta.to.global.u64 %rd236, %rd235;
|
|
shl.b32 %r943, %r942, 1;
|
|
mul.wide.u32 %rd237, %r943, 4;
|
|
add.s64 %rd238, %rd236, %rd237;
|
|
ld.global.f32 %f3795, [%rd238];
|
|
ld.global.f32 %f306, [%rd238+4];
|
|
@%p105 bra $L__BB11_123;
|
|
|
|
xor.b32 %r944, %r64, %r4;
|
|
mul.lo.s32 %r945, %r944, 9;
|
|
shr.u32 %r946, %r945, 4;
|
|
xor.b32 %r947, %r946, %r945;
|
|
mul.lo.s32 %r948, %r947, 668265261;
|
|
cvt.rn.f32.u32 %f1685, %r948;
|
|
mul.ftz.f32 %f1686, %f1685, 0f2F7FFFFE;
|
|
xor.b32 %r949, %r4, %r1495;
|
|
xor.b32 %r950, %r949, -905047700;
|
|
mul.lo.s32 %r951, %r950, 9;
|
|
shr.u32 %r952, %r951, 4;
|
|
xor.b32 %r953, %r952, %r951;
|
|
mul.lo.s32 %r954, %r953, 668265261;
|
|
cvt.rn.f32.u32 %f1687, %r954;
|
|
mul.ftz.f32 %f1688, %f1687, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f1689, %f1686, %f294, %f3795;
|
|
fma.rn.ftz.f32 %f1690, %f1688, %f294, %f306;
|
|
cvt.rmi.ftz.f32.f32 %f1691, %f1689;
|
|
sub.ftz.f32 %f3795, %f1689, %f1691;
|
|
cvt.rmi.ftz.f32.f32 %f1692, %f1690;
|
|
sub.ftz.f32 %f306, %f1690, %f1692;
|
|
bra.uni $L__BB11_123;
|
|
|
|
$L__BB11_5:
|
|
ld.const.u32 %r11, [kernel_params+1612];
|
|
setp.eq.s32 %p16, %r11, 0;
|
|
@%p16 bra $L__BB11_8;
|
|
|
|
ld.const.v2.u32 {%r221, %r222}, [kernel_params+1616];
|
|
mov.b32 %f1, %r222;
|
|
setp.geu.ftz.f32 %p17, %f1, 0f3F800000;
|
|
setp.lt.ftz.f32 %p18, %f1, 0f3F800000;
|
|
ld.const.u32 %r225, [kernel_params+1592];
|
|
selp.b32 %r226, %r225, %r4, %p18;
|
|
and.b32 %r227, %r1495, 255;
|
|
xor.b32 %r228, %r226, %r227;
|
|
mul.lo.s32 %r229, %r228, -512718531;
|
|
shr.u32 %r230, %r226, 16;
|
|
xor.b32 %r231, %r229, %r230;
|
|
bfe.u32 %r232, %r231, 4, 4;
|
|
shr.u32 %r233, %r226, 8;
|
|
xor.b32 %r234, %r231, %r233;
|
|
xor.b32 %r235, %r234, %r232;
|
|
mul.lo.s32 %r236, %r235, 153742143;
|
|
shr.u32 %r237, %r226, 23;
|
|
xor.b32 %r238, %r236, %r237;
|
|
bfe.u32 %r239, %r238, 1, 7;
|
|
xor.b32 %r240, %r239, %r238;
|
|
shr.u32 %r241, %r226, 27;
|
|
or.b32 %r242, %r241, 1;
|
|
mul.lo.s32 %r243, %r242, %r240;
|
|
mul.lo.s32 %r244, %r243, -1074439621;
|
|
bfe.u32 %r245, %r244, 2, 6;
|
|
xor.b32 %r246, %r245, %r244;
|
|
mul.lo.s32 %r247, %r246, -1638916925;
|
|
bfe.u32 %r248, %r247, 2, 6;
|
|
xor.b32 %r249, %r248, %r247;
|
|
mul.lo.s32 %r250, %r249, -933190689;
|
|
and.b32 %r251, %r250, 255;
|
|
bfe.u32 %r252, %r250, 5, 3;
|
|
xor.b32 %r253, %r252, %r251;
|
|
add.s32 %r254, %r221, -1;
|
|
xor.b32 %r12, %r1495, 61;
|
|
xor.b32 %r255, %r226, %r12;
|
|
mul.lo.s32 %r256, %r255, 9;
|
|
shr.u32 %r257, %r256, 4;
|
|
xor.b32 %r258, %r257, %r256;
|
|
mul.lo.s32 %r259, %r258, 668265261;
|
|
add.s32 %r260, %r259, %r10;
|
|
shr.u32 %r261, %r259, 16;
|
|
or.b32 %r262, %r261, 1;
|
|
mul.lo.s32 %r263, %r262, %r260;
|
|
mul.lo.s32 %r264, %r263, 89287766;
|
|
xor.b32 %r265, %r264, %r263;
|
|
mul.lo.s32 %r266, %r265, 1403136100;
|
|
xor.b32 %r267, %r266, %r265;
|
|
brev.b32 %r268, %r267;
|
|
neg.s32 %r269, %r221;
|
|
and.b32 %r270, %r6, %r269;
|
|
and.b32 %r271, %r268, %r254;
|
|
or.b32 %r272, %r271, %r270;
|
|
mad.lo.s32 %r273, %r253, %r221, %r272;
|
|
shl.b32 %r274, %r221, 8;
|
|
rem.u32 %r275, %r273, %r274;
|
|
ld.const.u64 %rd169, [kernel_params+2440];
|
|
cvta.to.global.u64 %rd170, %rd169;
|
|
shl.b32 %r276, %r275, 1;
|
|
mul.wide.u32 %rd171, %r276, 4;
|
|
add.s64 %rd172, %rd170, %rd171;
|
|
ld.global.f32 %f3717, [%rd172];
|
|
ld.global.f32 %f3718, [%rd172+4];
|
|
@%p17 bra $L__BB11_12;
|
|
|
|
xor.b32 %r277, %r12, %r4;
|
|
mul.lo.s32 %r278, %r277, 9;
|
|
shr.u32 %r279, %r278, 4;
|
|
xor.b32 %r280, %r279, %r278;
|
|
mul.lo.s32 %r281, %r280, 668265261;
|
|
cvt.rn.f32.u32 %f1094, %r281;
|
|
mul.ftz.f32 %f1095, %f1094, 0f2F7FFFFE;
|
|
xor.b32 %r282, %r4, %r1495;
|
|
xor.b32 %r283, %r282, -905047700;
|
|
mul.lo.s32 %r284, %r283, 9;
|
|
shr.u32 %r285, %r284, 4;
|
|
xor.b32 %r286, %r285, %r284;
|
|
mul.lo.s32 %r287, %r286, 668265261;
|
|
cvt.rn.f32.u32 %f1096, %r287;
|
|
mul.ftz.f32 %f1097, %f1096, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f1098, %f1095, %f1, %f3717;
|
|
fma.rn.ftz.f32 %f1099, %f1097, %f1, %f3718;
|
|
cvt.rmi.ftz.f32.f32 %f1100, %f1098;
|
|
sub.ftz.f32 %f3717, %f1098, %f1100;
|
|
cvt.rmi.ftz.f32.f32 %f1101, %f1099;
|
|
sub.ftz.f32 %f3718, %f1099, %f1101;
|
|
bra.uni $L__BB11_12;
|
|
|
|
$L__BB11_119:
|
|
mul.lo.s32 %r956, %r1495, 569420461;
|
|
shr.u32 %r957, %r956, 15;
|
|
xor.b32 %r958, %r957, %r956;
|
|
mul.lo.s32 %r959, %r958, -749064809;
|
|
shr.u32 %r960, %r959, 15;
|
|
xor.b32 %r961, %r4, %r959;
|
|
xor.b32 %r962, %r961, %r960;
|
|
xor.b32 %r65, %r962, -419546133;
|
|
xor.b32 %r963, %r962, 508809841;
|
|
add.s32 %r964, %r963, %r10;
|
|
shr.u32 %r965, %r963, 16;
|
|
or.b32 %r966, %r965, 1;
|
|
mul.lo.s32 %r967, %r966, %r964;
|
|
mul.lo.s32 %r968, %r967, 89287766;
|
|
xor.b32 %r969, %r968, %r967;
|
|
mul.lo.s32 %r970, %r969, 1403136100;
|
|
xor.b32 %r66, %r970, %r969;
|
|
setp.eq.s32 %p107, %r66, 0;
|
|
mov.u32 %r1503, 0;
|
|
@%p107 bra $L__BB11_122;
|
|
|
|
mov.u32 %r1500, 0;
|
|
mov.u64 %rd240, sobol_burley_table;
|
|
mov.u32 %r1501, %r66;
|
|
mov.u32 %r1503, %r1500;
|
|
|
|
$L__BB11_121:
|
|
clz.b32 %r973, %r1501;
|
|
add.s32 %r974, %r973, %r1500;
|
|
mul.wide.u32 %rd239, %r974, 4;
|
|
add.s64 %rd241, %rd240, %rd239;
|
|
ld.const.u32 %r975, [%rd241+128];
|
|
xor.b32 %r1503, %r975, %r1503;
|
|
add.s32 %r1500, %r974, 1;
|
|
shl.b32 %r976, %r1501, %r973;
|
|
shl.b32 %r1501, %r976, 1;
|
|
setp.ne.s32 %p108, %r1501, 0;
|
|
@%p108 bra $L__BB11_121;
|
|
|
|
$L__BB11_122:
|
|
xor.b32 %r977, %r65, -525684874;
|
|
brev.b32 %r978, %r66;
|
|
shr.u32 %r979, %r977, 16;
|
|
or.b32 %r980, %r979, 1;
|
|
mul.lo.s32 %r981, %r978, 1025551850;
|
|
xor.b32 %r982, %r981, %r978;
|
|
add.s32 %r983, %r982, %r977;
|
|
mul.lo.s32 %r984, %r983, %r980;
|
|
mul.lo.s32 %r985, %r984, 89287766;
|
|
xor.b32 %r986, %r985, %r984;
|
|
xor.b32 %r987, %r65, -1802089138;
|
|
mul.lo.s32 %r988, %r1503, 1025551850;
|
|
xor.b32 %r989, %r988, %r1503;
|
|
add.s32 %r990, %r989, %r987;
|
|
shr.u32 %r991, %r987, 16;
|
|
or.b32 %r992, %r991, 1;
|
|
mul.lo.s32 %r993, %r990, %r992;
|
|
mul.lo.s32 %r994, %r993, 89287766;
|
|
xor.b32 %r995, %r994, %r993;
|
|
mul.lo.s32 %r996, %r995, 1403136100;
|
|
xor.b32 %r997, %r996, %r995;
|
|
brev.b32 %r998, %r997;
|
|
cvt.rn.f32.u32 %f1693, %r998;
|
|
mul.lo.s32 %r999, %r986, 1403136100;
|
|
xor.b32 %r1000, %r999, %r986;
|
|
brev.b32 %r1001, %r1000;
|
|
cvt.rn.f32.u32 %f1694, %r1001;
|
|
mul.ftz.f32 %f306, %f1693, 0f2F7FFFFE;
|
|
mul.ftz.f32 %f3795, %f1694, 0f2F7FFFFE;
|
|
|
|
$L__BB11_123:
|
|
ld.const.u64 %rd707, [kernel_params+2648];
|
|
cvta.to.global.u64 %rd242, %rd707;
|
|
mul.lo.s64 %rd243, %rd4, 12;
|
|
add.s64 %rd244, %rd242, %rd243;
|
|
ld.global.f32 %f307, [%rd244];
|
|
ld.global.f32 %f308, [%rd244+4];
|
|
ld.global.f32 %f309, [%rd244+8];
|
|
ld.const.u64 %rd704, [kernel_params+2688];
|
|
cvta.to.global.u64 %rd245, %rd704;
|
|
add.s64 %rd247, %rd245, %rd168;
|
|
ld.global.f32 %f310, [%rd247];
|
|
ld.const.u64 %rd705, [kernel_params+2680];
|
|
cvta.to.global.u64 %rd248, %rd705;
|
|
add.s64 %rd249, %rd248, %rd168;
|
|
ld.global.f32 %f311, [%rd249];
|
|
ld.const.u64 %rd250, [kernel_params+2776];
|
|
cvta.to.global.u64 %rd251, %rd250;
|
|
add.s64 %rd252, %rd251, %rd243;
|
|
ld.const.u64 %rd706, [kernel_params+2736];
|
|
cvta.to.global.u64 %rd253, %rd706;
|
|
add.s64 %rd254, %rd253, %rd168;
|
|
ld.global.u32 %r74, [%rd254];
|
|
ld.const.u64 %rd255, [kernel_params+2760];
|
|
cvta.to.global.u64 %rd256, %rd255;
|
|
add.s64 %rd257, %rd256, %rd243;
|
|
ld.global.f32 %f312, [%rd257];
|
|
ld.global.f32 %f313, [%rd257+4];
|
|
ld.global.f32 %f314, [%rd257+8];
|
|
ld.global.f32 %f315, [%rd252+4];
|
|
ld.global.f32 %f316, [%rd252];
|
|
setp.neu.ftz.f32 %p109, %f316, %f315;
|
|
ld.global.f32 %f317, [%rd252+8];
|
|
setp.neu.ftz.f32 %p110, %f316, %f317;
|
|
or.pred %p111, %p109, %p110;
|
|
sub.ftz.f32 %f1695, %f317, %f315;
|
|
neg.ftz.f32 %f1696, %f315;
|
|
neg.ftz.f32 %f1697, %f317;
|
|
selp.f32 %f1698, %f1697, %f317, %p111;
|
|
add.ftz.f32 %f1699, %f316, %f1698;
|
|
selp.f32 %f1700, %f315, %f1696, %p111;
|
|
sub.ftz.f32 %f1701, %f1700, %f316;
|
|
mul.ftz.f32 %f1702, %f1699, %f1699;
|
|
fma.rn.ftz.f32 %f1703, %f1695, %f1695, %f1702;
|
|
fma.rn.ftz.f32 %f1704, %f1701, %f1701, %f1703;
|
|
sqrt.approx.ftz.f32 %f1705, %f1704;
|
|
div.approx.ftz.f32 %f318, %f1695, %f1705;
|
|
div.approx.ftz.f32 %f319, %f1699, %f1705;
|
|
div.approx.ftz.f32 %f320, %f1701, %f1705;
|
|
mul.ftz.f32 %f1706, %f315, %f320;
|
|
mul.ftz.f32 %f1707, %f317, %f319;
|
|
sub.ftz.f32 %f3803, %f1706, %f1707;
|
|
mul.ftz.f32 %f1708, %f317, %f318;
|
|
mul.ftz.f32 %f1709, %f316, %f320;
|
|
sub.ftz.f32 %f3804, %f1708, %f1709;
|
|
mul.ftz.f32 %f1710, %f316, %f319;
|
|
mul.ftz.f32 %f1711, %f315, %f318;
|
|
sub.ftz.f32 %f3805, %f1710, %f1711;
|
|
setp.lt.ftz.f32 %p112, %f306, 0f3F000000;
|
|
@%p112 bra $L__BB11_127;
|
|
bra.uni $L__BB11_124;
|
|
|
|
$L__BB11_127:
|
|
add.ftz.f32 %f3806, %f306, %f306;
|
|
mov.f32 %f3808, 0f3E800000;
|
|
mov.f32 %f3807, 0f3F000000;
|
|
mov.f32 %f3797, %f3803;
|
|
mov.f32 %f3798, %f3804;
|
|
mov.f32 %f3799, %f3805;
|
|
mov.f32 %f3800, %f318;
|
|
mov.f32 %f3801, %f319;
|
|
mov.f32 %f3802, %f320;
|
|
mov.f32 %f3803, %f316;
|
|
mov.f32 %f3804, %f315;
|
|
mov.f32 %f3805, %f317;
|
|
mov.f32 %f3809, %f3808;
|
|
bra.uni $L__BB11_128;
|
|
|
|
$L__BB11_124:
|
|
setp.lt.ftz.f32 %p113, %f306, 0f3F400000;
|
|
@%p113 bra $L__BB11_126;
|
|
bra.uni $L__BB11_125;
|
|
|
|
$L__BB11_126:
|
|
add.ftz.f32 %f1719, %f306, 0fBF000000;
|
|
mov.f32 %f3808, 0f3F000000;
|
|
mul.ftz.f32 %f3806, %f1719, 0f40800000;
|
|
mov.f32 %f3807, 0f3E800000;
|
|
mov.f32 %f3797, %f3803;
|
|
mov.f32 %f3798, %f3804;
|
|
mov.f32 %f3799, %f3805;
|
|
mov.f32 %f3800, %f316;
|
|
mov.f32 %f3801, %f315;
|
|
mov.f32 %f3802, %f317;
|
|
mov.f32 %f3803, %f318;
|
|
mov.f32 %f3804, %f319;
|
|
mov.f32 %f3805, %f320;
|
|
mov.f32 %f3809, %f3807;
|
|
bra.uni $L__BB11_128;
|
|
|
|
$L__BB11_8:
|
|
mul.lo.s32 %r289, %r1495, 569420461;
|
|
shr.u32 %r290, %r289, 15;
|
|
xor.b32 %r291, %r290, %r289;
|
|
mul.lo.s32 %r292, %r291, -749064809;
|
|
shr.u32 %r293, %r292, 15;
|
|
xor.b32 %r294, %r4, %r292;
|
|
xor.b32 %r295, %r294, %r293;
|
|
xor.b32 %r13, %r295, -419546133;
|
|
xor.b32 %r296, %r295, 508809841;
|
|
add.s32 %r297, %r296, %r10;
|
|
shr.u32 %r298, %r296, 16;
|
|
or.b32 %r299, %r298, 1;
|
|
mul.lo.s32 %r300, %r299, %r297;
|
|
mul.lo.s32 %r301, %r300, 89287766;
|
|
xor.b32 %r302, %r301, %r300;
|
|
mul.lo.s32 %r303, %r302, 1403136100;
|
|
xor.b32 %r14, %r303, %r302;
|
|
setp.eq.s32 %p19, %r14, 0;
|
|
mov.u32 %r1493, 0;
|
|
@%p19 bra $L__BB11_11;
|
|
|
|
mov.u32 %r1490, 0;
|
|
mov.u64 %rd174, sobol_burley_table;
|
|
mov.u32 %r1491, %r14;
|
|
mov.u32 %r1493, %r1490;
|
|
|
|
$L__BB11_10:
|
|
clz.b32 %r306, %r1491;
|
|
add.s32 %r307, %r306, %r1490;
|
|
mul.wide.u32 %rd173, %r307, 4;
|
|
add.s64 %rd175, %rd174, %rd173;
|
|
ld.const.u32 %r308, [%rd175+128];
|
|
xor.b32 %r1493, %r308, %r1493;
|
|
add.s32 %r1490, %r307, 1;
|
|
shl.b32 %r309, %r1491, %r306;
|
|
shl.b32 %r1491, %r309, 1;
|
|
setp.ne.s32 %p20, %r1491, 0;
|
|
@%p20 bra $L__BB11_10;
|
|
|
|
$L__BB11_11:
|
|
xor.b32 %r310, %r13, -525684874;
|
|
brev.b32 %r311, %r14;
|
|
shr.u32 %r312, %r310, 16;
|
|
or.b32 %r313, %r312, 1;
|
|
mul.lo.s32 %r314, %r311, 1025551850;
|
|
xor.b32 %r315, %r314, %r311;
|
|
add.s32 %r316, %r315, %r310;
|
|
mul.lo.s32 %r317, %r316, %r313;
|
|
mul.lo.s32 %r318, %r317, 89287766;
|
|
xor.b32 %r319, %r318, %r317;
|
|
xor.b32 %r320, %r13, -1802089138;
|
|
mul.lo.s32 %r321, %r1493, 1025551850;
|
|
xor.b32 %r322, %r321, %r1493;
|
|
add.s32 %r323, %r322, %r320;
|
|
shr.u32 %r324, %r320, 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 %f1102, %r331;
|
|
mul.lo.s32 %r332, %r319, 1403136100;
|
|
xor.b32 %r333, %r332, %r319;
|
|
brev.b32 %r334, %r333;
|
|
cvt.rn.f32.u32 %f1103, %r334;
|
|
mul.ftz.f32 %f3718, %f1102, 0f2F7FFFFE;
|
|
mul.ftz.f32 %f3717, %f1103, 0f2F7FFFFE;
|
|
|
|
$L__BB11_12:
|
|
ld.const.u64 %rd14, [kernel_params+2648];
|
|
cvta.to.global.u64 %rd176, %rd14;
|
|
mul.lo.s64 %rd177, %rd4, 12;
|
|
add.s64 %rd178, %rd176, %rd177;
|
|
ld.global.f32 %f14, [%rd178];
|
|
ld.global.f32 %f15, [%rd178+4];
|
|
ld.global.f32 %f16, [%rd178+8];
|
|
ld.const.u64 %rd179, [kernel_params+2656];
|
|
cvta.to.global.u64 %rd180, %rd179;
|
|
add.s64 %rd181, %rd180, %rd177;
|
|
ld.const.u64 %rd15, [kernel_params+2688];
|
|
ld.const.u64 %rd16, [kernel_params+2680];
|
|
ld.const.u64 %rd182, [kernel_params+2776];
|
|
cvta.to.global.u64 %rd183, %rd182;
|
|
add.s64 %rd184, %rd183, %rd177;
|
|
ld.const.u64 %rd17, [kernel_params+2736];
|
|
cvta.to.global.u64 %rd185, %rd17;
|
|
add.s64 %rd187, %rd185, %rd168;
|
|
ld.global.u32 %r22, [%rd187];
|
|
ld.global.f32 %f17, [%rd181];
|
|
neg.ftz.f32 %f18, %f17;
|
|
ld.global.f32 %f19, [%rd181+4];
|
|
neg.ftz.f32 %f20, %f19;
|
|
ld.global.f32 %f21, [%rd181+8];
|
|
neg.ftz.f32 %f22, %f21;
|
|
mul.ftz.f32 %f1104, %f3717, 0f40C90FDB;
|
|
sqrt.approx.ftz.f32 %f1105, %f3718;
|
|
cos.approx.ftz.f32 %f1106, %f1104;
|
|
mul.ftz.f32 %f1107, %f1105, %f1106;
|
|
sin.approx.ftz.f32 %f1108, %f1104;
|
|
mul.ftz.f32 %f1109, %f1105, %f1108;
|
|
mul.ftz.f32 %f1110, %f1107, %f1107;
|
|
mov.f32 %f1111, 0f3F800000;
|
|
sub.ftz.f32 %f1112, %f1111, %f1110;
|
|
mul.ftz.f32 %f1113, %f1109, %f1109;
|
|
sub.ftz.f32 %f1114, %f1112, %f1113;
|
|
mov.f32 %f3732, 0f00000000;
|
|
max.ftz.f32 %f1116, %f1114, %f3732;
|
|
sqrt.approx.ftz.f32 %f1117, %f1116;
|
|
setp.neu.ftz.f32 %p22, %f17, %f19;
|
|
setp.neu.ftz.f32 %p23, %f17, %f21;
|
|
or.pred %p1, %p22, %p23;
|
|
sub.ftz.f32 %f1118, %f19, %f21;
|
|
selp.f32 %f1119, %f20, %f19, %p1;
|
|
add.ftz.f32 %f1120, %f17, %f1119;
|
|
selp.f32 %f1121, %f21, %f22, %p1;
|
|
sub.ftz.f32 %f1122, %f1121, %f17;
|
|
mul.ftz.f32 %f1123, %f1122, %f1122;
|
|
fma.rn.ftz.f32 %f1124, %f1118, %f1118, %f1123;
|
|
fma.rn.ftz.f32 %f1125, %f1120, %f1120, %f1124;
|
|
sqrt.approx.ftz.f32 %f1126, %f1125;
|
|
div.approx.ftz.f32 %f1127, %f1118, %f1126;
|
|
div.approx.ftz.f32 %f1128, %f1122, %f1126;
|
|
div.approx.ftz.f32 %f1129, %f1120, %f1126;
|
|
mul.ftz.f32 %f1130, %f1129, %f19;
|
|
mul.ftz.f32 %f1131, %f1128, %f21;
|
|
sub.ftz.f32 %f1132, %f1131, %f1130;
|
|
mul.ftz.f32 %f1133, %f1127, %f21;
|
|
mul.ftz.f32 %f1134, %f1129, %f17;
|
|
sub.ftz.f32 %f1135, %f1134, %f1133;
|
|
mul.ftz.f32 %f1136, %f1128, %f17;
|
|
mul.ftz.f32 %f1137, %f1127, %f19;
|
|
sub.ftz.f32 %f1138, %f1137, %f1136;
|
|
mul.ftz.f32 %f1139, %f1109, %f1132;
|
|
mul.ftz.f32 %f1140, %f1109, %f1135;
|
|
mul.ftz.f32 %f1141, %f1109, %f1138;
|
|
fma.rn.ftz.f32 %f1142, %f1107, %f1127, %f1139;
|
|
fma.rn.ftz.f32 %f1143, %f1107, %f1128, %f1140;
|
|
fma.rn.ftz.f32 %f1144, %f1107, %f1129, %f1141;
|
|
mul.ftz.f32 %f1145, %f1117, %f17;
|
|
mul.ftz.f32 %f1146, %f1117, %f19;
|
|
mul.ftz.f32 %f1147, %f1117, %f21;
|
|
sub.ftz.f32 %f3767, %f1142, %f1145;
|
|
sub.ftz.f32 %f3766, %f1143, %f1146;
|
|
sub.ftz.f32 %f3765, %f1144, %f1147;
|
|
ld.global.f32 %f1148, [%rd184];
|
|
mul.ftz.f32 %f1149, %f1148, %f3767;
|
|
ld.global.f32 %f1150, [%rd184+4];
|
|
mul.ftz.f32 %f1151, %f1150, %f3766;
|
|
neg.ftz.f32 %f1152, %f1151;
|
|
sub.ftz.f32 %f1153, %f1152, %f1149;
|
|
ld.global.f32 %f1154, [%rd184+8];
|
|
mul.ftz.f32 %f1155, %f1154, %f3765;
|
|
sub.ftz.f32 %f1156, %f1153, %f1155;
|
|
setp.le.ftz.f32 %p24, %f1156, 0f00000000;
|
|
mov.pred %p375, -1;
|
|
@%p24 bra $L__BB11_369;
|
|
|
|
ld.const.u64 %rd188, [kernel_params+2728];
|
|
cvta.to.global.u64 %rd189, %rd188;
|
|
add.s64 %rd191, %rd189, %rd168;
|
|
ld.global.u32 %r335, [%rd191];
|
|
cvta.to.global.u64 %rd192, %rd16;
|
|
add.s64 %rd193, %rd192, %rd168;
|
|
ld.global.f32 %f1157, [%rd193];
|
|
cvta.to.global.u64 %rd194, %rd15;
|
|
add.s64 %rd195, %rd194, %rd168;
|
|
ld.global.f32 %f1158, [%rd195];
|
|
st.local.f32 [%rd6], %f14;
|
|
st.local.f32 [%rd6+4], %f15;
|
|
st.local.f32 [%rd6+8], %f16;
|
|
st.local.f32 [%rd6+12], %f3767;
|
|
st.local.f32 [%rd6+16], %f3766;
|
|
st.local.f32 [%rd6+20], %f3765;
|
|
mov.u32 %r1494, 0;
|
|
st.local.u32 [%rd6+24], %r1494;
|
|
mov.u32 %r337, 2139095039;
|
|
st.local.u32 [%rd6+28], %r337;
|
|
st.local.f32 [%rd6+32], %f1157;
|
|
st.local.f32 [%rd6+52], %f1158;
|
|
st.local.u32 [%rd6+56], %r1494;
|
|
st.local.u32 [%rd6+40], %r22;
|
|
st.local.u32 [%rd6+36], %r335;
|
|
mov.u32 %r338, -1;
|
|
st.local.u32 [%rd6+48], %r338;
|
|
st.local.u32 [%rd6+44], %r338;
|
|
ld.const.v2.u64 {%rd196, %rd197}, [kernel_params+2752];
|
|
cvta.to.global.u64 %rd200, %rd196;
|
|
add.s64 %rd202, %rd200, %rd177;
|
|
ld.global.f32 %f26, [%rd202+4];
|
|
ld.global.f32 %f27, [%rd202+8];
|
|
cvta.to.global.u64 %rd203, %rd197;
|
|
add.s64 %rd204, %rd203, %rd177;
|
|
ld.global.f32 %f28, [%rd204];
|
|
ld.global.f32 %f29, [%rd204+4];
|
|
ld.global.f32 %f30, [%rd204+8];
|
|
ld.const.u64 %rd205, [kernel_params+2768];
|
|
cvta.to.global.u64 %rd206, %rd205;
|
|
add.s64 %rd207, %rd206, %rd168;
|
|
ld.const.u64 %rd208, [kernel_params+2600];
|
|
cvta.to.global.u64 %rd209, %rd208;
|
|
add.s64 %rd18, %rd209, %rd177;
|
|
ld.global.f32 %f31, [%rd18];
|
|
ld.global.f32 %f32, [%rd18+4];
|
|
ld.global.f32 %f33, [%rd18+8];
|
|
ld.global.f32 %f34, [%rd207];
|
|
mul.ftz.f32 %f1159, %f34, %f34;
|
|
mul.ftz.f32 %f1160, %f34, %f1159;
|
|
mul.ftz.f32 %f1161, %f34, %f1160;
|
|
mul.ftz.f32 %f1162, %f34, %f1161;
|
|
mul.ftz.f32 %f1163, %f34, %f1162;
|
|
mul.ftz.f32 %f1164, %f34, %f1163;
|
|
fma.rn.ftz.f32 %f1165, %f34, 0fBFA46AD8, 0f3FE9BC16;
|
|
fma.rn.ftz.f32 %f1166, %f1159, 0fBFE64726, %f1165;
|
|
fma.rn.ftz.f32 %f1167, %f1160, 0f41131A59, %f1166;
|
|
fma.rn.ftz.f32 %f1168, %f1161, 0fC1B6928D, %f1167;
|
|
fma.rn.ftz.f32 %f1169, %f1162, 0f4200180D, %f1168;
|
|
fma.rn.ftz.f32 %f1170, %f1163, 0fC1BD0308, %f1169;
|
|
fma.rn.ftz.f32 %f35, %f1164, 0f40E6BDCF, %f1170;
|
|
mul.ftz.f32 %f1171, %f1159, 0fC349D8D6;
|
|
fma.rn.ftz.f32 %f1172, %f34, 0f41F9317A, %f1171;
|
|
fma.rn.ftz.f32 %f1173, %f1160, 0f445264DD, %f1172;
|
|
fma.rn.ftz.f32 %f1174, %f1161, 0fC4FC42F9, %f1173;
|
|
fma.rn.ftz.f32 %f1175, %f1162, 0f452ABB73, %f1174;
|
|
fma.rn.ftz.f32 %f1176, %f1163, 0fC4F1ED41, %f1175;
|
|
fma.rn.ftz.f32 %f1177, %f1164, 0f440BC094, %f1176;
|
|
mul.ftz.f32 %f1178, %f1177, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1179, %f1178;
|
|
fma.rn.ftz.f32 %f36, %f1179, 0f3E02699A, 0f409F8609;
|
|
fma.rn.ftz.f32 %f1180, %f34, 0fBECA16A7, 0f3F8C65F1;
|
|
fma.rn.ftz.f32 %f1181, %f1159, 0f3F86BAFB, %f1180;
|
|
fma.rn.ftz.f32 %f1182, %f1160, 0fC10D6F27, %f1181;
|
|
fma.rn.ftz.f32 %f1183, %f1161, 0f41E6EA22, %f1182;
|
|
fma.rn.ftz.f32 %f1184, %f1162, 0fC23B856B, %f1183;
|
|
fma.rn.ftz.f32 %f1185, %f1163, 0f421A2940, %f1184;
|
|
fma.rn.ftz.f32 %f37, %f1164, 0fC14B7D5B, %f1185;
|
|
fma.rn.ftz.f32 %f1186, %f34, 0f3EB86522, 0f3EFE1C5F;
|
|
fma.rn.ftz.f32 %f1187, %f1159, 0fC009B06D, %f1186;
|
|
fma.rn.ftz.f32 %f1188, %f1160, 0f418F1E16, %f1187;
|
|
fma.rn.ftz.f32 %f1189, %f1161, 0fC25D3190, %f1188;
|
|
fma.rn.ftz.f32 %f1190, %f1162, 0f42A421C8, %f1189;
|
|
fma.rn.ftz.f32 %f1191, %f1163, 0fC26A0ADB, %f1190;
|
|
fma.rn.ftz.f32 %f38, %f1164, 0f417D90B6, %f1191;
|
|
mul.ftz.f32 %f1192, %f1159, 0fC41496D5;
|
|
fma.rn.ftz.f32 %f1193, %f34, 0f42997698, %f1192;
|
|
fma.rn.ftz.f32 %f1194, %f1160, 0f45190E22, %f1193;
|
|
fma.rn.ftz.f32 %f1195, %f1161, 0fC5AE457B, %f1194;
|
|
fma.rn.ftz.f32 %f1196, %f1162, 0f45DE64D0, %f1195;
|
|
fma.rn.ftz.f32 %f1197, %f1163, 0fC594DC5C, %f1196;
|
|
fma.rn.ftz.f32 %f1198, %f1164, 0f44A2F105, %f1197;
|
|
mul.ftz.f32 %f1199, %f1198, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1200, %f1199;
|
|
fma.rn.ftz.f32 %f39, %f1200, 0f3B4B8EB2, 0f40876BC0;
|
|
fma.rn.ftz.f32 %f1201, %f34, 0fC0212958, 0f4019FC65;
|
|
fma.rn.ftz.f32 %f1202, %f1159, 0f4112F58D, %f1201;
|
|
fma.rn.ftz.f32 %f1203, %f1160, 0fC29E7037, %f1202;
|
|
fma.rn.ftz.f32 %f1204, %f1161, 0f43818A9B, %f1203;
|
|
fma.rn.ftz.f32 %f1205, %f1162, 0fC3C9CE91, %f1204;
|
|
fma.rn.ftz.f32 %f1206, %f1163, 0f43976DB6, %f1205;
|
|
fma.rn.ftz.f32 %f40, %f1164, 0fC2AEDFC5, %f1206;
|
|
sub.ftz.f32 %f41, %f1111, %f34;
|
|
ld.global.f32 %f42, [%rd202];
|
|
mul.ftz.f32 %f43, %f42, %f36;
|
|
abs.ftz.f32 %f44, %f43;
|
|
setp.leu.ftz.f32 %p25, %f44, 0f3F800000;
|
|
mov.f32 %f3719, %f44;
|
|
@%p25 bra $L__BB11_15;
|
|
|
|
rcp.approx.ftz.f32 %f3719, %f44;
|
|
|
|
$L__BB11_15:
|
|
lg2.approx.ftz.f32 %f1208, %f42;
|
|
mul.ftz.f32 %f1209, %f1208, 0f3E800000;
|
|
ex2.approx.ftz.f32 %f47, %f1209;
|
|
mul.rn.ftz.f32 %f1210, %f3719, %f3719;
|
|
mov.f32 %f1211, 0fC0B59883;
|
|
mov.f32 %f1212, 0fBF52C7EA;
|
|
fma.rn.ftz.f32 %f1213, %f1210, %f1212, %f1211;
|
|
mov.f32 %f1214, 0fC0D21907;
|
|
fma.rn.ftz.f32 %f1215, %f1213, %f1210, %f1214;
|
|
mul.ftz.f32 %f1216, %f1210, %f1215;
|
|
mul.ftz.f32 %f1217, %f3719, %f1216;
|
|
add.ftz.f32 %f1218, %f1210, 0f41355DC0;
|
|
mov.f32 %f1219, 0f41E6BD60;
|
|
fma.rn.ftz.f32 %f1220, %f1218, %f1210, %f1219;
|
|
mov.f32 %f1221, 0f419D92C8;
|
|
fma.rn.ftz.f32 %f1222, %f1220, %f1210, %f1221;
|
|
rcp.approx.ftz.f32 %f1223, %f1222;
|
|
fma.rn.ftz.f32 %f1224, %f1217, %f1223, %f3719;
|
|
mov.f32 %f1225, 0f3FC90FDB;
|
|
sub.ftz.f32 %f1226, %f1225, %f1224;
|
|
setp.gt.ftz.f32 %p26, %f44, 0f3F800000;
|
|
selp.f32 %f1227, %f1226, %f1224, %p26;
|
|
mov.b32 %r339, %f1227;
|
|
mov.b32 %r340, %f43;
|
|
and.b32 %r341, %r340, -2147483648;
|
|
or.b32 %r342, %r341, %r339;
|
|
mov.b32 %f1228, %r342;
|
|
setp.le.ftz.f32 %p27, %f44, 0f7F800000;
|
|
selp.f32 %f48, %f1228, %f1227, %p27;
|
|
mul.ftz.f32 %f49, %f42, %f39;
|
|
abs.ftz.f32 %f50, %f49;
|
|
setp.leu.ftz.f32 %p28, %f50, 0f3F800000;
|
|
mov.f32 %f3720, %f50;
|
|
@%p28 bra $L__BB11_17;
|
|
|
|
rcp.approx.ftz.f32 %f3720, %f50;
|
|
|
|
$L__BB11_17:
|
|
mul.rn.ftz.f32 %f1229, %f3720, %f3720;
|
|
fma.rn.ftz.f32 %f1232, %f1229, %f1212, %f1211;
|
|
fma.rn.ftz.f32 %f1234, %f1232, %f1229, %f1214;
|
|
mul.ftz.f32 %f1235, %f1229, %f1234;
|
|
mul.ftz.f32 %f1236, %f3720, %f1235;
|
|
add.ftz.f32 %f1237, %f1229, 0f41355DC0;
|
|
fma.rn.ftz.f32 %f1239, %f1237, %f1229, %f1219;
|
|
fma.rn.ftz.f32 %f1241, %f1239, %f1229, %f1221;
|
|
rcp.approx.ftz.f32 %f1242, %f1241;
|
|
fma.rn.ftz.f32 %f1244, %f1236, %f1242, %f3720;
|
|
sub.ftz.f32 %f1246, %f1225, %f1244;
|
|
setp.gt.ftz.f32 %p29, %f50, 0f3F800000;
|
|
selp.f32 %f1247, %f1246, %f1244, %p29;
|
|
mov.b32 %r343, %f1247;
|
|
mov.b32 %r344, %f49;
|
|
and.b32 %r345, %r344, -2147483648;
|
|
or.b32 %r346, %r345, %r343;
|
|
mov.b32 %f1248, %r346;
|
|
setp.le.ftz.f32 %p30, %f50, 0f7F800000;
|
|
selp.f32 %f1249, %f1248, %f1247, %p30;
|
|
lg2.approx.ftz.f32 %f1250, %f1249;
|
|
mul.ftz.f32 %f1251, %f40, %f1250;
|
|
ex2.approx.ftz.f32 %f1252, %f1251;
|
|
mul.ftz.f32 %f1253, %f38, %f47;
|
|
mul.ftz.f32 %f1254, %f1253, %f1252;
|
|
sub.ftz.f32 %f1255, %f1111, %f47;
|
|
mul.ftz.f32 %f1256, %f35, %f1255;
|
|
lg2.approx.ftz.f32 %f1257, %f48;
|
|
mul.ftz.f32 %f1258, %f37, %f1257;
|
|
ex2.approx.ftz.f32 %f1259, %f1258;
|
|
fma.rn.ftz.f32 %f1260, %f1256, %f1259, %f1254;
|
|
max.ftz.f32 %f1262, %f1260, %f3732;
|
|
mov.f32 %f1263, 0f3F7FFFEF;
|
|
min.ftz.f32 %f3729, %f1262, %f1263;
|
|
mov.f32 %f1264, 0f24E69595;
|
|
max.ftz.f32 %f1265, %f28, %f1264;
|
|
rcp.approx.ftz.f32 %f1266, %f1265;
|
|
div.approx.ftz.f32 %f54, %f1266, %f41;
|
|
mul.ftz.f32 %f55, %f26, %f36;
|
|
abs.ftz.f32 %f56, %f55;
|
|
setp.leu.ftz.f32 %p31, %f56, 0f3F800000;
|
|
mov.f32 %f3721, %f56;
|
|
@%p31 bra $L__BB11_19;
|
|
|
|
rcp.approx.ftz.f32 %f3721, %f56;
|
|
|
|
$L__BB11_19:
|
|
lg2.approx.ftz.f32 %f1267, %f26;
|
|
mul.ftz.f32 %f1268, %f1267, 0f3E800000;
|
|
ex2.approx.ftz.f32 %f59, %f1268;
|
|
mul.rn.ftz.f32 %f1269, %f3721, %f3721;
|
|
fma.rn.ftz.f32 %f1272, %f1269, %f1212, %f1211;
|
|
fma.rn.ftz.f32 %f1274, %f1272, %f1269, %f1214;
|
|
mul.ftz.f32 %f1275, %f1269, %f1274;
|
|
mul.ftz.f32 %f1276, %f3721, %f1275;
|
|
add.ftz.f32 %f1277, %f1269, 0f41355DC0;
|
|
fma.rn.ftz.f32 %f1279, %f1277, %f1269, %f1219;
|
|
fma.rn.ftz.f32 %f1281, %f1279, %f1269, %f1221;
|
|
rcp.approx.ftz.f32 %f1282, %f1281;
|
|
fma.rn.ftz.f32 %f1283, %f1276, %f1282, %f3721;
|
|
sub.ftz.f32 %f1285, %f1225, %f1283;
|
|
setp.gt.ftz.f32 %p32, %f56, 0f3F800000;
|
|
selp.f32 %f1286, %f1285, %f1283, %p32;
|
|
mov.b32 %r347, %f1286;
|
|
mov.b32 %r348, %f55;
|
|
and.b32 %r349, %r348, -2147483648;
|
|
or.b32 %r350, %r349, %r347;
|
|
mov.b32 %f1287, %r350;
|
|
setp.le.ftz.f32 %p33, %f56, 0f7F800000;
|
|
selp.f32 %f60, %f1287, %f1286, %p33;
|
|
mul.ftz.f32 %f61, %f26, %f39;
|
|
abs.ftz.f32 %f62, %f61;
|
|
setp.leu.ftz.f32 %p34, %f62, 0f3F800000;
|
|
mov.f32 %f3722, %f62;
|
|
@%p34 bra $L__BB11_21;
|
|
|
|
rcp.approx.ftz.f32 %f3722, %f62;
|
|
|
|
$L__BB11_21:
|
|
mul.rn.ftz.f32 %f1288, %f3722, %f3722;
|
|
fma.rn.ftz.f32 %f1291, %f1288, %f1212, %f1211;
|
|
fma.rn.ftz.f32 %f1293, %f1291, %f1288, %f1214;
|
|
mul.ftz.f32 %f1294, %f1288, %f1293;
|
|
mul.ftz.f32 %f1295, %f3722, %f1294;
|
|
add.ftz.f32 %f1296, %f1288, 0f41355DC0;
|
|
fma.rn.ftz.f32 %f1298, %f1296, %f1288, %f1219;
|
|
fma.rn.ftz.f32 %f1300, %f1298, %f1288, %f1221;
|
|
rcp.approx.ftz.f32 %f1301, %f1300;
|
|
fma.rn.ftz.f32 %f1303, %f1295, %f1301, %f3722;
|
|
sub.ftz.f32 %f1305, %f1225, %f1303;
|
|
setp.gt.ftz.f32 %p35, %f62, 0f3F800000;
|
|
selp.f32 %f1306, %f1305, %f1303, %p35;
|
|
mov.b32 %r351, %f1306;
|
|
mov.b32 %r352, %f61;
|
|
and.b32 %r353, %r352, -2147483648;
|
|
or.b32 %r354, %r353, %r351;
|
|
mov.b32 %f1307, %r354;
|
|
setp.le.ftz.f32 %p36, %f62, 0f7F800000;
|
|
selp.f32 %f1308, %f1307, %f1306, %p36;
|
|
lg2.approx.ftz.f32 %f1309, %f1308;
|
|
mul.ftz.f32 %f1310, %f40, %f1309;
|
|
ex2.approx.ftz.f32 %f1311, %f1310;
|
|
mul.ftz.f32 %f1312, %f38, %f59;
|
|
mul.ftz.f32 %f1313, %f1312, %f1311;
|
|
sub.ftz.f32 %f1314, %f1111, %f59;
|
|
mul.ftz.f32 %f1315, %f35, %f1314;
|
|
lg2.approx.ftz.f32 %f1316, %f60;
|
|
mul.ftz.f32 %f1317, %f37, %f1316;
|
|
ex2.approx.ftz.f32 %f1318, %f1317;
|
|
fma.rn.ftz.f32 %f1319, %f1315, %f1318, %f1313;
|
|
max.ftz.f32 %f1321, %f1319, %f3732;
|
|
min.ftz.f32 %f3731, %f1321, %f1263;
|
|
max.ftz.f32 %f1324, %f29, %f1264;
|
|
rcp.approx.ftz.f32 %f1325, %f1324;
|
|
div.approx.ftz.f32 %f66, %f1325, %f41;
|
|
mul.ftz.f32 %f67, %f27, %f36;
|
|
abs.ftz.f32 %f68, %f67;
|
|
setp.leu.ftz.f32 %p37, %f68, 0f3F800000;
|
|
mov.f32 %f3723, %f68;
|
|
@%p37 bra $L__BB11_23;
|
|
|
|
rcp.approx.ftz.f32 %f3723, %f68;
|
|
|
|
$L__BB11_23:
|
|
lg2.approx.ftz.f32 %f1326, %f27;
|
|
mul.ftz.f32 %f1327, %f1326, 0f3E800000;
|
|
ex2.approx.ftz.f32 %f1328, %f1327;
|
|
mul.rn.ftz.f32 %f1329, %f3723, %f3723;
|
|
fma.rn.ftz.f32 %f1332, %f1329, %f1212, %f1211;
|
|
fma.rn.ftz.f32 %f1334, %f1332, %f1329, %f1214;
|
|
mul.ftz.f32 %f1335, %f1329, %f1334;
|
|
mul.ftz.f32 %f1336, %f3723, %f1335;
|
|
add.ftz.f32 %f1337, %f1329, 0f41355DC0;
|
|
fma.rn.ftz.f32 %f1339, %f1337, %f1329, %f1219;
|
|
fma.rn.ftz.f32 %f1341, %f1339, %f1329, %f1221;
|
|
rcp.approx.ftz.f32 %f1342, %f1341;
|
|
fma.rn.ftz.f32 %f1344, %f1336, %f1342, %f3723;
|
|
sub.ftz.f32 %f1346, %f1225, %f1344;
|
|
setp.gt.ftz.f32 %p38, %f68, 0f3F800000;
|
|
selp.f32 %f1347, %f1346, %f1344, %p38;
|
|
mov.b32 %r355, %f1347;
|
|
mov.b32 %r356, %f67;
|
|
and.b32 %r357, %r356, -2147483648;
|
|
or.b32 %r358, %r357, %r355;
|
|
mov.b32 %f1348, %r358;
|
|
setp.le.ftz.f32 %p39, %f68, 0f7F800000;
|
|
selp.f32 %f1349, %f1348, %f1347, %p39;
|
|
lg2.approx.ftz.f32 %f1350, %f1349;
|
|
mul.ftz.f32 %f1351, %f37, %f1350;
|
|
ex2.approx.ftz.f32 %f1352, %f1351;
|
|
sub.ftz.f32 %f1353, %f1111, %f1328;
|
|
mul.ftz.f32 %f1354, %f35, %f1353;
|
|
mul.ftz.f32 %f71, %f1354, %f1352;
|
|
mul.ftz.f32 %f72, %f38, %f1328;
|
|
mul.ftz.f32 %f73, %f27, %f39;
|
|
abs.ftz.f32 %f74, %f73;
|
|
setp.leu.ftz.f32 %p40, %f74, 0f3F800000;
|
|
mov.f32 %f3724, %f74;
|
|
@%p40 bra $L__BB11_25;
|
|
|
|
rcp.approx.ftz.f32 %f3724, %f74;
|
|
|
|
$L__BB11_25:
|
|
mul.rn.ftz.f32 %f1356, %f3724, %f3724;
|
|
fma.rn.ftz.f32 %f1359, %f1356, %f1212, %f1211;
|
|
fma.rn.ftz.f32 %f1361, %f1359, %f1356, %f1214;
|
|
mul.ftz.f32 %f1362, %f1356, %f1361;
|
|
mul.ftz.f32 %f1363, %f3724, %f1362;
|
|
add.ftz.f32 %f1364, %f1356, 0f41355DC0;
|
|
fma.rn.ftz.f32 %f1366, %f1364, %f1356, %f1219;
|
|
fma.rn.ftz.f32 %f1368, %f1366, %f1356, %f1221;
|
|
rcp.approx.ftz.f32 %f1369, %f1368;
|
|
fma.rn.ftz.f32 %f1370, %f1363, %f1369, %f3724;
|
|
sub.ftz.f32 %f1372, %f1225, %f1370;
|
|
setp.gt.ftz.f32 %p41, %f74, 0f3F800000;
|
|
selp.f32 %f1373, %f1372, %f1370, %p41;
|
|
mov.b32 %r359, %f1373;
|
|
mov.b32 %r360, %f73;
|
|
and.b32 %r361, %r360, -2147483648;
|
|
or.b32 %r362, %r361, %r359;
|
|
mov.b32 %f1374, %r362;
|
|
setp.le.ftz.f32 %p42, %f74, 0f7F800000;
|
|
selp.f32 %f1375, %f1374, %f1373, %p42;
|
|
lg2.approx.ftz.f32 %f1376, %f1375;
|
|
mul.ftz.f32 %f1377, %f40, %f1376;
|
|
ex2.approx.ftz.f32 %f1378, %f1377;
|
|
fma.rn.ftz.f32 %f1379, %f72, %f1378, %f71;
|
|
max.ftz.f32 %f1380, %f1379, %f3732;
|
|
min.ftz.f32 %f3733, %f1380, %f1263;
|
|
max.ftz.f32 %f1383, %f30, %f1264;
|
|
rcp.approx.ftz.f32 %f1384, %f1383;
|
|
div.approx.ftz.f32 %f78, %f1384, %f41;
|
|
setp.eq.ftz.f32 %p43, %f42, 0f00000000;
|
|
mov.f32 %f3728, %f3732;
|
|
@%p43 bra $L__BB11_27;
|
|
|
|
div.approx.ftz.f32 %f3728, %f31, %f42;
|
|
|
|
$L__BB11_27:
|
|
setp.eq.ftz.f32 %p44, %f26, 0f00000000;
|
|
mov.f32 %f3730, %f3732;
|
|
@%p44 bra $L__BB11_29;
|
|
|
|
div.approx.ftz.f32 %f3730, %f32, %f26;
|
|
|
|
$L__BB11_29:
|
|
setp.eq.ftz.f32 %p45, %f27, 0f00000000;
|
|
@%p45 bra $L__BB11_31;
|
|
|
|
div.approx.ftz.f32 %f3732, %f33, %f27;
|
|
|
|
$L__BB11_31:
|
|
setp.geu.ftz.f32 %p46, %f3729, 0f3E4CCCCD;
|
|
@%p46 bra $L__BB11_33;
|
|
|
|
mov.f32 %f1387, 0f3E4CCCCD;
|
|
div.approx.ftz.f32 %f1388, %f3729, %f1387;
|
|
mul.ftz.f32 %f3728, %f1388, %f3728;
|
|
mov.f32 %f3729, %f1387;
|
|
|
|
$L__BB11_33:
|
|
setp.geu.ftz.f32 %p47, %f3731, 0f3E4CCCCD;
|
|
@%p47 bra $L__BB11_35;
|
|
|
|
mov.f32 %f1389, 0f3E4CCCCD;
|
|
div.approx.ftz.f32 %f1390, %f3731, %f1389;
|
|
mul.ftz.f32 %f3730, %f1390, %f3730;
|
|
mov.f32 %f3731, %f1389;
|
|
|
|
$L__BB11_35:
|
|
setp.geu.ftz.f32 %p48, %f3733, 0f3E4CCCCD;
|
|
@%p48 bra $L__BB11_37;
|
|
|
|
mov.f32 %f1391, 0f3E4CCCCD;
|
|
div.approx.ftz.f32 %f1392, %f3733, %f1391;
|
|
mul.ftz.f32 %f3732, %f1392, %f3732;
|
|
mov.f32 %f3733, %f1391;
|
|
|
|
$L__BB11_37:
|
|
mul.ftz.f32 %f94, %f54, %f3729;
|
|
mul.ftz.f32 %f95, %f66, %f3731;
|
|
mul.ftz.f32 %f96, %f78, %f3733;
|
|
max.ftz.f32 %f1393, %f3729, %f3731;
|
|
max.ftz.f32 %f1394, %f1393, %f3733;
|
|
fma.rn.ftz.f32 %f1395, %f1394, 0fBCB0CB44, 0f401C5921;
|
|
mov.f32 %f1396, 0f3F14218E;
|
|
div.approx.ftz.f32 %f1397, %f1396, %f1394;
|
|
add.ftz.f32 %f1398, %f1397, %f1395;
|
|
lg2.approx.ftz.f32 %f1399, %f1394;
|
|
mul.ftz.f32 %f1400, %f1399, %f1398;
|
|
ex2.approx.ftz.f32 %f1401, %f1400;
|
|
sub.ftz.f32 %f1403, %f1111, %f1401;
|
|
rsqrt.approx.ftz.f32 %f97, %f1403;
|
|
setp.eq.ftz.f32 %p50, %f97, 0f3F800000;
|
|
@%p50 bra $L__BB11_369;
|
|
|
|
add.ftz.f32 %f98, %f97, 0f3F800000;
|
|
add.ftz.f32 %f1406, %f97, 0fBF800000;
|
|
div.approx.ftz.f32 %f1407, %f98, %f1406;
|
|
lg2.approx.ftz.f32 %f1408, %f1407;
|
|
mul.ftz.f32 %f99, %f1408, 0f3F317218;
|
|
add.s32 %r364, %r4, %r1495;
|
|
xor.b32 %r365, %r364, 701873903;
|
|
shr.u32 %r366, %r365, 16;
|
|
xor.b32 %r367, %r366, %r365;
|
|
mul.lo.s32 %r368, %r367, 569420461;
|
|
shr.u32 %r369, %r368, 15;
|
|
xor.b32 %r370, %r369, %r368;
|
|
mul.lo.s32 %r371, %r370, -749064809;
|
|
shr.u32 %r372, %r371, 15;
|
|
xor.b32 %r24, %r372, %r371;
|
|
xor.b32 %r23, %r24, -419546133;
|
|
abs.ftz.f32 %f1409, %f34;
|
|
lg2.approx.ftz.f32 %f1410, %f1409;
|
|
mul.ftz.f32 %f1411, %f1410, 0f3E000000;
|
|
ex2.approx.ftz.f32 %f1412, %f1411;
|
|
mov.f32 %f1413, 0f3F000000;
|
|
max.ftz.f32 %f1414, %f1413, %f1412;
|
|
sub.ftz.f32 %f100, %f1111, %f1414;
|
|
sub.ftz.f32 %f1415, %f54, %f94;
|
|
mul.ftz.f32 %f101, %f41, %f94;
|
|
add.ftz.f32 %f102, %f101, %f1415;
|
|
sub.ftz.f32 %f1416, %f66, %f95;
|
|
mul.ftz.f32 %f103, %f41, %f95;
|
|
add.ftz.f32 %f104, %f103, %f1416;
|
|
sub.ftz.f32 %f1417, %f78, %f96;
|
|
mul.ftz.f32 %f105, %f41, %f96;
|
|
add.ftz.f32 %f106, %f105, %f1417;
|
|
ld.const.v2.u32 {%r373, %r374}, [kernel_params+1616];
|
|
mov.b32 %f107, %r374;
|
|
setp.lt.ftz.f32 %p51, %f107, 0f3F800000;
|
|
ld.const.u32 %r376, [kernel_params+1592];
|
|
selp.b32 %r26, %r376, %r23, %p51;
|
|
shr.u32 %r27, %r26, 16;
|
|
shr.u32 %r28, %r26, 8;
|
|
shr.u32 %r29, %r26, 23;
|
|
shr.u32 %r377, %r26, 27;
|
|
or.b32 %r378, %r377, 1;
|
|
mul.lo.s32 %r30, %r378, -1074439621;
|
|
add.s32 %r31, %r373, -1;
|
|
neg.s32 %r379, %r373;
|
|
and.b32 %r32, %r6, %r379;
|
|
shl.b32 %r33, %r373, 8;
|
|
ld.const.u64 %rd210, [kernel_params+2440];
|
|
cvta.to.global.u64 %rd19, %rd210;
|
|
cvta.to.local.u64 %rd20, %rd145;
|
|
cvt.u32.u64 %r34, %rd146;
|
|
shr.u64 %rd213, %rd146, 32;
|
|
cvt.u32.u64 %r35, %rd213;
|
|
cvt.u32.u64 %r36, %rd162;
|
|
shr.u64 %rd215, %rd162, 32;
|
|
cvt.u32.u64 %r37, %rd215;
|
|
ld.const.u64 %rd703, [kernel_params+2096];
|
|
selp.f32 %f1418, %f22, %f21, %p1;
|
|
add.ftz.f32 %f108, %f17, %f1418;
|
|
selp.f32 %f1419, %f19, %f20, %p1;
|
|
sub.ftz.f32 %f109, %f1419, %f17;
|
|
sub.ftz.f32 %f110, %f21, %f19;
|
|
mul.ftz.f32 %f1420, %f108, %f108;
|
|
fma.rn.ftz.f32 %f1421, %f110, %f110, %f1420;
|
|
fma.rn.ftz.f32 %f111, %f109, %f109, %f1421;
|
|
xor.b32 %r38, %r24, 753937031;
|
|
mov.u16 %rs41, 0;
|
|
mov.f32 %f3777, 0f00000000;
|
|
mov.u16 %rs50, 1;
|
|
sqrt.approx.ftz.f32 %f1558, %f111;
|
|
mov.f32 %f3737, %f16;
|
|
mov.f32 %f3738, %f15;
|
|
mov.f32 %f3739, %f14;
|
|
mov.u16 %rs94, %rs41;
|
|
|
|
$L__BB11_39:
|
|
setp.lt.u32 %p2, %r1494, 10;
|
|
selp.f32 %f122, %f78, %f106, %p2;
|
|
selp.f32 %f123, %f66, %f104, %p2;
|
|
selp.f32 %f124, %f54, %f102, %p2;
|
|
selp.f32 %f125, %f34, 0f00000000, %p2;
|
|
selp.f32 %f126, %f100, 0f3F400000, %p2;
|
|
@%p16 bra $L__BB11_42;
|
|
|
|
setp.geu.ftz.f32 %p53, %f107, 0f3F800000;
|
|
add.s32 %r380, %r1495, 17;
|
|
and.b32 %r381, %r380, 255;
|
|
xor.b32 %r382, %r26, %r381;
|
|
mul.lo.s32 %r383, %r382, -512718531;
|
|
xor.b32 %r384, %r383, %r27;
|
|
bfe.u32 %r385, %r384, 4, 4;
|
|
xor.b32 %r386, %r384, %r28;
|
|
xor.b32 %r387, %r386, %r385;
|
|
mul.lo.s32 %r388, %r387, 153742143;
|
|
xor.b32 %r389, %r388, %r29;
|
|
bfe.u32 %r390, %r389, 1, 7;
|
|
xor.b32 %r391, %r390, %r389;
|
|
mul.lo.s32 %r392, %r30, %r391;
|
|
bfe.u32 %r393, %r392, 2, 6;
|
|
xor.b32 %r394, %r393, %r392;
|
|
mul.lo.s32 %r395, %r394, -1638916925;
|
|
bfe.u32 %r396, %r395, 2, 6;
|
|
xor.b32 %r397, %r396, %r395;
|
|
mul.lo.s32 %r398, %r397, -933190689;
|
|
and.b32 %r399, %r398, 255;
|
|
bfe.u32 %r400, %r398, 5, 3;
|
|
xor.b32 %r401, %r400, %r399;
|
|
xor.b32 %r41, %r380, 61;
|
|
xor.b32 %r402, %r26, %r41;
|
|
mul.lo.s32 %r403, %r402, 9;
|
|
shr.u32 %r404, %r403, 4;
|
|
xor.b32 %r405, %r404, %r403;
|
|
mul.lo.s32 %r406, %r405, 668265261;
|
|
add.s32 %r407, %r406, %r10;
|
|
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;
|
|
brev.b32 %r415, %r414;
|
|
and.b32 %r416, %r415, %r31;
|
|
or.b32 %r417, %r416, %r32;
|
|
mad.lo.s32 %r418, %r401, %r373, %r417;
|
|
rem.u32 %r419, %r418, %r33;
|
|
shl.b32 %r420, %r419, 1;
|
|
mul.wide.u32 %rd216, %r420, 4;
|
|
add.s64 %rd217, %rd19, %rd216;
|
|
ld.global.f32 %f3744, [%rd217];
|
|
@%p53 bra $L__BB11_43;
|
|
|
|
xor.b32 %r421, %r41, %r23;
|
|
mul.lo.s32 %r422, %r421, 9;
|
|
shr.u32 %r423, %r422, 4;
|
|
xor.b32 %r424, %r423, %r422;
|
|
mul.lo.s32 %r425, %r424, 668265261;
|
|
cvt.rn.f32.u32 %f1422, %r425;
|
|
mul.ftz.f32 %f1423, %f1422, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f1424, %f1423, %f107, %f3744;
|
|
cvt.rmi.ftz.f32.f32 %f1425, %f1424;
|
|
sub.ftz.f32 %f3744, %f1424, %f1425;
|
|
bra.uni $L__BB11_43;
|
|
|
|
$L__BB11_42:
|
|
add.s32 %r426, %r1495, 17;
|
|
shr.u32 %r427, %r426, 16;
|
|
xor.b32 %r428, %r427, %r426;
|
|
mul.lo.s32 %r429, %r428, 569420461;
|
|
shr.u32 %r430, %r429, 15;
|
|
xor.b32 %r431, %r430, %r429;
|
|
mul.lo.s32 %r432, %r431, -749064809;
|
|
shr.u32 %r433, %r432, 15;
|
|
xor.b32 %r434, %r24, %r432;
|
|
xor.b32 %r435, %r434, %r433;
|
|
xor.b32 %r436, %r435, -1074177026;
|
|
add.s32 %r437, %r436, %r10;
|
|
shr.u32 %r438, %r436, 16;
|
|
or.b32 %r439, %r438, 1;
|
|
mul.lo.s32 %r440, %r439, %r437;
|
|
mul.lo.s32 %r441, %r440, 89287766;
|
|
xor.b32 %r442, %r441, %r440;
|
|
mul.lo.s32 %r443, %r442, 1403136100;
|
|
xor.b32 %r444, %r443, %r442;
|
|
xor.b32 %r445, %r435, 1667004349;
|
|
brev.b32 %r446, %r444;
|
|
mul.lo.s32 %r447, %r446, 1025551850;
|
|
xor.b32 %r448, %r447, %r446;
|
|
add.s32 %r449, %r448, %r445;
|
|
shr.u32 %r450, %r445, 16;
|
|
or.b32 %r451, %r450, 1;
|
|
mul.lo.s32 %r452, %r449, %r451;
|
|
mul.lo.s32 %r453, %r452, 89287766;
|
|
xor.b32 %r454, %r453, %r452;
|
|
mul.lo.s32 %r455, %r454, 1403136100;
|
|
xor.b32 %r456, %r455, %r454;
|
|
brev.b32 %r457, %r456;
|
|
cvt.rn.f32.u32 %f1426, %r457;
|
|
mul.ftz.f32 %f3744, %f1426, 0f2F7FFFFE;
|
|
|
|
$L__BB11_43:
|
|
mul.ftz.f32 %f1430, %f3729, %f3728;
|
|
abs.ftz.f32 %f131, %f1430;
|
|
mul.ftz.f32 %f1431, %f3731, %f3730;
|
|
abs.ftz.f32 %f132, %f1431;
|
|
mul.ftz.f32 %f1432, %f3733, %f3732;
|
|
abs.ftz.f32 %f133, %f1432;
|
|
add.ftz.f32 %f1433, %f131, %f132;
|
|
add.ftz.f32 %f134, %f133, %f1433;
|
|
setp.leu.ftz.f32 %p54, %f134, 0f00000000;
|
|
mov.f32 %f3745, 0f3EAAAAAB;
|
|
mov.f32 %f3746, %f3745;
|
|
mov.f32 %f3747, %f3745;
|
|
@%p54 bra $L__BB11_45;
|
|
|
|
div.approx.ftz.f32 %f3745, %f131, %f134;
|
|
div.approx.ftz.f32 %f3746, %f132, %f134;
|
|
div.approx.ftz.f32 %f3747, %f133, %f134;
|
|
|
|
$L__BB11_45:
|
|
add.ftz.f32 %f141, %f3745, 0f00000000;
|
|
setp.gt.ftz.f32 %p55, %f141, %f3744;
|
|
mov.u64 %rd698, 0;
|
|
@%p55 bra $L__BB11_48;
|
|
|
|
add.ftz.f32 %f1434, %f141, %f3746;
|
|
setp.gt.ftz.f32 %p56, %f1434, %f3744;
|
|
mov.u64 %rd698, 1;
|
|
@%p56 bra $L__BB11_48;
|
|
|
|
mov.u64 %rd698, 2;
|
|
|
|
$L__BB11_48:
|
|
st.local.f32 [%rd20], %f124;
|
|
st.local.f32 [%rd20+4], %f123;
|
|
st.local.f32 [%rd20+8], %f122;
|
|
shl.b64 %rd221, %rd698, 2;
|
|
add.s64 %rd222, %rd20, %rd221;
|
|
ld.local.f32 %f3768, [%rd222];
|
|
add.s32 %r42, %r1495, 18;
|
|
@%p16 bra $L__BB11_51;
|
|
|
|
setp.geu.ftz.f32 %p58, %f107, 0f3F800000;
|
|
and.b32 %r458, %r42, 255;
|
|
xor.b32 %r459, %r26, %r458;
|
|
mul.lo.s32 %r460, %r459, -512718531;
|
|
xor.b32 %r461, %r460, %r27;
|
|
bfe.u32 %r462, %r461, 4, 4;
|
|
xor.b32 %r463, %r461, %r28;
|
|
xor.b32 %r464, %r463, %r462;
|
|
mul.lo.s32 %r465, %r464, 153742143;
|
|
xor.b32 %r466, %r465, %r29;
|
|
bfe.u32 %r467, %r466, 1, 7;
|
|
xor.b32 %r468, %r467, %r466;
|
|
mul.lo.s32 %r469, %r30, %r468;
|
|
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;
|
|
xor.b32 %r43, %r42, 61;
|
|
xor.b32 %r479, %r26, %r43;
|
|
mul.lo.s32 %r480, %r479, 9;
|
|
shr.u32 %r481, %r480, 4;
|
|
xor.b32 %r482, %r481, %r480;
|
|
mul.lo.s32 %r483, %r482, 668265261;
|
|
add.s32 %r484, %r483, %r10;
|
|
shr.u32 %r485, %r483, 16;
|
|
or.b32 %r486, %r485, 1;
|
|
mul.lo.s32 %r487, %r486, %r484;
|
|
mul.lo.s32 %r488, %r487, 89287766;
|
|
xor.b32 %r489, %r488, %r487;
|
|
mul.lo.s32 %r490, %r489, 1403136100;
|
|
xor.b32 %r491, %r490, %r489;
|
|
brev.b32 %r492, %r491;
|
|
and.b32 %r493, %r492, %r31;
|
|
or.b32 %r494, %r493, %r32;
|
|
mad.lo.s32 %r495, %r478, %r373, %r494;
|
|
rem.u32 %r496, %r495, %r33;
|
|
shl.b32 %r497, %r496, 1;
|
|
mul.wide.u32 %rd223, %r497, 4;
|
|
add.s64 %rd224, %rd19, %rd223;
|
|
ld.global.f32 %f3748, [%rd224];
|
|
@%p58 bra $L__BB11_52;
|
|
|
|
xor.b32 %r498, %r43, %r23;
|
|
mul.lo.s32 %r499, %r498, 9;
|
|
shr.u32 %r500, %r499, 4;
|
|
xor.b32 %r501, %r500, %r499;
|
|
mul.lo.s32 %r502, %r501, 668265261;
|
|
cvt.rn.f32.u32 %f1435, %r502;
|
|
mul.ftz.f32 %f1436, %f1435, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f1437, %f1436, %f107, %f3748;
|
|
cvt.rmi.ftz.f32.f32 %f1438, %f1437;
|
|
sub.ftz.f32 %f3748, %f1437, %f1438;
|
|
bra.uni $L__BB11_52;
|
|
|
|
$L__BB11_51:
|
|
shr.u32 %r503, %r42, 16;
|
|
xor.b32 %r504, %r503, %r42;
|
|
mul.lo.s32 %r505, %r504, 569420461;
|
|
shr.u32 %r506, %r505, 15;
|
|
xor.b32 %r507, %r506, %r505;
|
|
mul.lo.s32 %r508, %r507, -749064809;
|
|
shr.u32 %r509, %r508, 15;
|
|
xor.b32 %r510, %r24, %r508;
|
|
xor.b32 %r511, %r510, %r509;
|
|
xor.b32 %r512, %r511, -1074177026;
|
|
add.s32 %r513, %r512, %r10;
|
|
shr.u32 %r514, %r512, 16;
|
|
or.b32 %r515, %r514, 1;
|
|
mul.lo.s32 %r516, %r515, %r513;
|
|
mul.lo.s32 %r517, %r516, 89287766;
|
|
xor.b32 %r518, %r517, %r516;
|
|
mul.lo.s32 %r519, %r518, 1403136100;
|
|
xor.b32 %r520, %r519, %r518;
|
|
xor.b32 %r521, %r511, 1667004349;
|
|
brev.b32 %r522, %r520;
|
|
mul.lo.s32 %r523, %r522, 1025551850;
|
|
xor.b32 %r524, %r523, %r522;
|
|
add.s32 %r525, %r524, %r521;
|
|
shr.u32 %r526, %r521, 16;
|
|
or.b32 %r527, %r526, 1;
|
|
mul.lo.s32 %r528, %r525, %r527;
|
|
mul.lo.s32 %r529, %r528, 89287766;
|
|
xor.b32 %r530, %r529, %r528;
|
|
mul.lo.s32 %r531, %r530, 1403136100;
|
|
xor.b32 %r532, %r531, %r530;
|
|
brev.b32 %r533, %r532;
|
|
cvt.rn.f32.u32 %f1439, %r533;
|
|
mul.ftz.f32 %f3748, %f1439, 0f2F7FFFFE;
|
|
|
|
$L__BB11_52:
|
|
setp.eq.s32 %p59, %r1494, 0;
|
|
mov.f32 %f3769, 0f00000000;
|
|
mov.f32 %f3770, %f3769;
|
|
mov.f32 %f3771, %f1111;
|
|
mov.f32 %f3772, %f3769;
|
|
mov.f32 %f3773, %f1111;
|
|
@%p59 bra $L__BB11_81;
|
|
|
|
add.s32 %r44, %r1495, 19;
|
|
@%p16 bra $L__BB11_56;
|
|
|
|
setp.geu.ftz.f32 %p61, %f107, 0f3F800000;
|
|
and.b32 %r534, %r44, 255;
|
|
xor.b32 %r535, %r26, %r534;
|
|
mul.lo.s32 %r536, %r535, -512718531;
|
|
xor.b32 %r537, %r536, %r27;
|
|
bfe.u32 %r538, %r537, 4, 4;
|
|
xor.b32 %r539, %r537, %r28;
|
|
xor.b32 %r540, %r539, %r538;
|
|
mul.lo.s32 %r541, %r540, 153742143;
|
|
xor.b32 %r542, %r541, %r29;
|
|
bfe.u32 %r543, %r542, 1, 7;
|
|
xor.b32 %r544, %r543, %r542;
|
|
mul.lo.s32 %r545, %r30, %r544;
|
|
bfe.u32 %r546, %r545, 2, 6;
|
|
xor.b32 %r547, %r546, %r545;
|
|
mul.lo.s32 %r548, %r547, -1638916925;
|
|
bfe.u32 %r549, %r548, 2, 6;
|
|
xor.b32 %r550, %r549, %r548;
|
|
mul.lo.s32 %r551, %r550, -933190689;
|
|
and.b32 %r552, %r551, 255;
|
|
bfe.u32 %r553, %r551, 5, 3;
|
|
xor.b32 %r554, %r553, %r552;
|
|
xor.b32 %r45, %r44, 61;
|
|
xor.b32 %r555, %r26, %r45;
|
|
mul.lo.s32 %r556, %r555, 9;
|
|
shr.u32 %r557, %r556, 4;
|
|
xor.b32 %r558, %r557, %r556;
|
|
mul.lo.s32 %r559, %r558, 668265261;
|
|
add.s32 %r560, %r559, %r10;
|
|
shr.u32 %r561, %r559, 16;
|
|
or.b32 %r562, %r561, 1;
|
|
mul.lo.s32 %r563, %r562, %r560;
|
|
mul.lo.s32 %r564, %r563, 89287766;
|
|
xor.b32 %r565, %r564, %r563;
|
|
mul.lo.s32 %r566, %r565, 1403136100;
|
|
xor.b32 %r567, %r566, %r565;
|
|
brev.b32 %r568, %r567;
|
|
and.b32 %r569, %r568, %r31;
|
|
or.b32 %r570, %r569, %r32;
|
|
mad.lo.s32 %r571, %r554, %r373, %r570;
|
|
rem.u32 %r572, %r571, %r33;
|
|
shl.b32 %r573, %r572, 1;
|
|
mul.wide.u32 %rd225, %r573, 4;
|
|
add.s64 %rd226, %rd19, %rd225;
|
|
ld.global.f32 %f3749, [%rd226];
|
|
@%p61 bra $L__BB11_57;
|
|
|
|
xor.b32 %r574, %r45, %r23;
|
|
mul.lo.s32 %r575, %r574, 9;
|
|
shr.u32 %r576, %r575, 4;
|
|
xor.b32 %r577, %r576, %r575;
|
|
mul.lo.s32 %r578, %r577, 668265261;
|
|
cvt.rn.f32.u32 %f1445, %r578;
|
|
mul.ftz.f32 %f1446, %f1445, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f1447, %f1446, %f107, %f3749;
|
|
cvt.rmi.ftz.f32.f32 %f1448, %f1447;
|
|
sub.ftz.f32 %f3749, %f1447, %f1448;
|
|
bra.uni $L__BB11_57;
|
|
|
|
$L__BB11_56:
|
|
shr.u32 %r579, %r44, 16;
|
|
xor.b32 %r580, %r579, %r44;
|
|
mul.lo.s32 %r581, %r580, 569420461;
|
|
shr.u32 %r582, %r581, 15;
|
|
xor.b32 %r583, %r582, %r581;
|
|
mul.lo.s32 %r584, %r583, -749064809;
|
|
shr.u32 %r585, %r584, 15;
|
|
xor.b32 %r586, %r24, %r584;
|
|
xor.b32 %r587, %r586, %r585;
|
|
xor.b32 %r588, %r587, -1074177026;
|
|
add.s32 %r589, %r588, %r10;
|
|
shr.u32 %r590, %r588, 16;
|
|
or.b32 %r591, %r590, 1;
|
|
mul.lo.s32 %r592, %r591, %r589;
|
|
mul.lo.s32 %r593, %r592, 89287766;
|
|
xor.b32 %r594, %r593, %r592;
|
|
mul.lo.s32 %r595, %r594, 1403136100;
|
|
xor.b32 %r596, %r595, %r594;
|
|
xor.b32 %r597, %r587, 1667004349;
|
|
brev.b32 %r598, %r596;
|
|
mul.lo.s32 %r599, %r598, 1025551850;
|
|
xor.b32 %r600, %r599, %r598;
|
|
add.s32 %r601, %r600, %r597;
|
|
shr.u32 %r602, %r597, 16;
|
|
or.b32 %r603, %r602, 1;
|
|
mul.lo.s32 %r604, %r601, %r603;
|
|
mul.lo.s32 %r605, %r604, 89287766;
|
|
xor.b32 %r606, %r605, %r604;
|
|
mul.lo.s32 %r607, %r606, 1403136100;
|
|
xor.b32 %r608, %r607, %r606;
|
|
brev.b32 %r609, %r608;
|
|
cvt.rn.f32.u32 %f1449, %r609;
|
|
mul.ftz.f32 %f3749, %f1449, 0f2F7FFFFE;
|
|
|
|
$L__BB11_57:
|
|
and.b16 %rs43, %rs94, 255;
|
|
setp.eq.s16 %p62, %rs43, 0;
|
|
mov.u16 %rs89, 0;
|
|
mov.f32 %f3769, 0f00000000;
|
|
@%p62 bra $L__BB11_63;
|
|
|
|
sub.ftz.f32 %f1451, %f3739, %f14;
|
|
sub.ftz.f32 %f1452, %f3738, %f15;
|
|
mul.ftz.f32 %f1453, %f1452, %f20;
|
|
fma.rn.ftz.f32 %f1454, %f1451, %f18, %f1453;
|
|
sub.ftz.f32 %f1455, %f3737, %f16;
|
|
fma.rn.ftz.f32 %f1456, %f1455, %f22, %f1454;
|
|
mov.f32 %f1457, 0f00000000;
|
|
max.ftz.f32 %f1458, %f1456, %f1457;
|
|
min.ftz.f32 %f1459, %f1458, %f3777;
|
|
add.ftz.f32 %f1460, %f1459, %f1459;
|
|
sub.ftz.f32 %f1461, %f3777, %f1460;
|
|
div.approx.ftz.f32 %f1462, %f1461, %f97;
|
|
mul.ftz.f32 %f1463, %f1462, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1464, %f1463;
|
|
add.ftz.f32 %f1465, %f1464, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f3769, %f1465;
|
|
add.s32 %r46, %r1495, 20;
|
|
@%p16 bra $L__BB11_61;
|
|
|
|
setp.geu.ftz.f32 %p64, %f107, 0f3F800000;
|
|
and.b32 %r610, %r46, 255;
|
|
xor.b32 %r611, %r26, %r610;
|
|
mul.lo.s32 %r612, %r611, -512718531;
|
|
xor.b32 %r613, %r612, %r27;
|
|
bfe.u32 %r614, %r613, 4, 4;
|
|
xor.b32 %r615, %r613, %r28;
|
|
xor.b32 %r616, %r615, %r614;
|
|
mul.lo.s32 %r617, %r616, 153742143;
|
|
xor.b32 %r618, %r617, %r29;
|
|
bfe.u32 %r619, %r618, 1, 7;
|
|
xor.b32 %r620, %r619, %r618;
|
|
mul.lo.s32 %r621, %r30, %r620;
|
|
bfe.u32 %r622, %r621, 2, 6;
|
|
xor.b32 %r623, %r622, %r621;
|
|
mul.lo.s32 %r624, %r623, -1638916925;
|
|
bfe.u32 %r625, %r624, 2, 6;
|
|
xor.b32 %r626, %r625, %r624;
|
|
mul.lo.s32 %r627, %r626, -933190689;
|
|
and.b32 %r628, %r627, 255;
|
|
bfe.u32 %r629, %r627, 5, 3;
|
|
xor.b32 %r630, %r629, %r628;
|
|
xor.b32 %r47, %r46, 61;
|
|
xor.b32 %r631, %r26, %r47;
|
|
mul.lo.s32 %r632, %r631, 9;
|
|
shr.u32 %r633, %r632, 4;
|
|
xor.b32 %r634, %r633, %r632;
|
|
mul.lo.s32 %r635, %r634, 668265261;
|
|
add.s32 %r636, %r635, %r10;
|
|
shr.u32 %r637, %r635, 16;
|
|
or.b32 %r638, %r637, 1;
|
|
mul.lo.s32 %r639, %r638, %r636;
|
|
mul.lo.s32 %r640, %r639, 89287766;
|
|
xor.b32 %r641, %r640, %r639;
|
|
mul.lo.s32 %r642, %r641, 1403136100;
|
|
xor.b32 %r643, %r642, %r641;
|
|
brev.b32 %r644, %r643;
|
|
and.b32 %r645, %r644, %r31;
|
|
or.b32 %r646, %r645, %r32;
|
|
mad.lo.s32 %r647, %r630, %r373, %r646;
|
|
rem.u32 %r648, %r647, %r33;
|
|
shl.b32 %r649, %r648, 1;
|
|
mul.wide.u32 %rd227, %r649, 4;
|
|
add.s64 %rd228, %rd19, %rd227;
|
|
ld.global.f32 %f3750, [%rd228];
|
|
@%p64 bra $L__BB11_62;
|
|
|
|
xor.b32 %r650, %r47, %r23;
|
|
mul.lo.s32 %r651, %r650, 9;
|
|
shr.u32 %r652, %r651, 4;
|
|
xor.b32 %r653, %r652, %r651;
|
|
mul.lo.s32 %r654, %r653, 668265261;
|
|
cvt.rn.f32.u32 %f1466, %r654;
|
|
mul.ftz.f32 %f1467, %f1466, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f1468, %f1467, %f107, %f3750;
|
|
cvt.rmi.ftz.f32.f32 %f1469, %f1468;
|
|
sub.ftz.f32 %f3750, %f1468, %f1469;
|
|
bra.uni $L__BB11_62;
|
|
|
|
$L__BB11_61:
|
|
shr.u32 %r655, %r46, 16;
|
|
xor.b32 %r656, %r655, %r46;
|
|
mul.lo.s32 %r657, %r656, 569420461;
|
|
shr.u32 %r658, %r657, 15;
|
|
xor.b32 %r659, %r658, %r657;
|
|
mul.lo.s32 %r660, %r659, -749064809;
|
|
shr.u32 %r661, %r660, 15;
|
|
xor.b32 %r662, %r24, %r660;
|
|
xor.b32 %r663, %r662, %r661;
|
|
xor.b32 %r664, %r663, -1074177026;
|
|
add.s32 %r665, %r664, %r10;
|
|
shr.u32 %r666, %r664, 16;
|
|
or.b32 %r667, %r666, 1;
|
|
mul.lo.s32 %r668, %r667, %r665;
|
|
mul.lo.s32 %r669, %r668, 89287766;
|
|
xor.b32 %r670, %r669, %r668;
|
|
mul.lo.s32 %r671, %r670, 1403136100;
|
|
xor.b32 %r672, %r671, %r670;
|
|
xor.b32 %r673, %r663, 1667004349;
|
|
brev.b32 %r674, %r672;
|
|
mul.lo.s32 %r675, %r674, 1025551850;
|
|
xor.b32 %r676, %r675, %r674;
|
|
add.s32 %r677, %r676, %r673;
|
|
shr.u32 %r678, %r673, 16;
|
|
or.b32 %r679, %r678, 1;
|
|
mul.lo.s32 %r680, %r677, %r679;
|
|
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;
|
|
cvt.rn.f32.u32 %f1470, %r685;
|
|
mul.ftz.f32 %f3750, %f1470, 0f2F7FFFFE;
|
|
|
|
$L__BB11_62:
|
|
setp.lt.ftz.f32 %p65, %f3750, %f3769;
|
|
selp.u16 %rs89, 1, 0, %p65;
|
|
|
|
$L__BB11_63:
|
|
@%p16 bra $L__BB11_66;
|
|
|
|
setp.geu.ftz.f32 %p67, %f107, 0f3F800000;
|
|
add.s32 %r686, %r1495, 16;
|
|
and.b32 %r687, %r686, 255;
|
|
xor.b32 %r688, %r26, %r687;
|
|
mul.lo.s32 %r689, %r688, -512718531;
|
|
xor.b32 %r690, %r689, %r27;
|
|
bfe.u32 %r691, %r690, 4, 4;
|
|
xor.b32 %r692, %r690, %r28;
|
|
xor.b32 %r693, %r692, %r691;
|
|
mul.lo.s32 %r694, %r693, 153742143;
|
|
xor.b32 %r695, %r694, %r29;
|
|
bfe.u32 %r696, %r695, 1, 7;
|
|
xor.b32 %r697, %r696, %r695;
|
|
mul.lo.s32 %r698, %r30, %r697;
|
|
bfe.u32 %r699, %r698, 2, 6;
|
|
xor.b32 %r700, %r699, %r698;
|
|
mul.lo.s32 %r701, %r700, -1638916925;
|
|
bfe.u32 %r702, %r701, 2, 6;
|
|
xor.b32 %r703, %r702, %r701;
|
|
mul.lo.s32 %r704, %r703, -933190689;
|
|
and.b32 %r705, %r704, 255;
|
|
bfe.u32 %r706, %r704, 5, 3;
|
|
xor.b32 %r707, %r706, %r705;
|
|
xor.b32 %r48, %r686, 61;
|
|
xor.b32 %r708, %r26, %r48;
|
|
mul.lo.s32 %r709, %r708, 9;
|
|
shr.u32 %r710, %r709, 4;
|
|
xor.b32 %r711, %r710, %r709;
|
|
mul.lo.s32 %r712, %r711, 668265261;
|
|
add.s32 %r713, %r712, %r10;
|
|
shr.u32 %r714, %r712, 16;
|
|
or.b32 %r715, %r714, 1;
|
|
mul.lo.s32 %r716, %r715, %r713;
|
|
mul.lo.s32 %r717, %r716, 89287766;
|
|
xor.b32 %r718, %r717, %r716;
|
|
mul.lo.s32 %r719, %r718, 1403136100;
|
|
xor.b32 %r720, %r719, %r718;
|
|
brev.b32 %r721, %r720;
|
|
and.b32 %r722, %r721, %r31;
|
|
or.b32 %r723, %r722, %r32;
|
|
mad.lo.s32 %r724, %r707, %r373, %r723;
|
|
rem.u32 %r725, %r724, %r33;
|
|
shl.b32 %r726, %r725, 1;
|
|
mul.wide.u32 %rd229, %r726, 4;
|
|
add.s64 %rd230, %rd19, %rd229;
|
|
ld.global.f32 %f167, [%rd230];
|
|
ld.global.f32 %f3755, [%rd230+4];
|
|
@%p67 bra $L__BB11_70;
|
|
|
|
xor.b32 %r727, %r48, %r23;
|
|
mul.lo.s32 %r728, %r727, 9;
|
|
shr.u32 %r729, %r728, 4;
|
|
xor.b32 %r730, %r729, %r728;
|
|
mul.lo.s32 %r731, %r730, 668265261;
|
|
cvt.rn.f32.u32 %f1471, %r731;
|
|
mul.ftz.f32 %f1472, %f1471, 0f2F7FFFFE;
|
|
xor.b32 %r733, %r38, %r686;
|
|
mul.lo.s32 %r734, %r733, 9;
|
|
shr.u32 %r735, %r734, 4;
|
|
xor.b32 %r736, %r735, %r734;
|
|
mul.lo.s32 %r737, %r736, 668265261;
|
|
cvt.rn.f32.u32 %f1473, %r737;
|
|
mul.ftz.f32 %f1474, %f1473, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f1475, %f1472, %f107, %f167;
|
|
fma.rn.ftz.f32 %f1476, %f1474, %f107, %f3755;
|
|
cvt.rmi.ftz.f32.f32 %f1477, %f1475;
|
|
sub.ftz.f32 %f167, %f1475, %f1477;
|
|
cvt.rmi.ftz.f32.f32 %f1478, %f1476;
|
|
sub.ftz.f32 %f3755, %f1476, %f1478;
|
|
bra.uni $L__BB11_70;
|
|
|
|
$L__BB11_66:
|
|
add.s32 %r739, %r1495, 16;
|
|
shr.u32 %r740, %r739, 16;
|
|
xor.b32 %r741, %r740, %r739;
|
|
mul.lo.s32 %r742, %r741, 569420461;
|
|
shr.u32 %r743, %r742, 15;
|
|
xor.b32 %r744, %r743, %r742;
|
|
mul.lo.s32 %r745, %r744, -749064809;
|
|
shr.u32 %r746, %r745, 15;
|
|
xor.b32 %r747, %r24, %r745;
|
|
xor.b32 %r49, %r747, %r746;
|
|
xor.b32 %r748, %r49, -122820198;
|
|
add.s32 %r749, %r748, %r10;
|
|
shr.u32 %r750, %r748, 16;
|
|
or.b32 %r751, %r750, 1;
|
|
mul.lo.s32 %r752, %r751, %r749;
|
|
mul.lo.s32 %r753, %r752, 89287766;
|
|
xor.b32 %r754, %r753, %r752;
|
|
mul.lo.s32 %r755, %r754, 1403136100;
|
|
xor.b32 %r50, %r755, %r754;
|
|
setp.eq.s32 %p68, %r50, 0;
|
|
mov.u32 %r1499, 0;
|
|
@%p68 bra $L__BB11_69;
|
|
|
|
mov.u32 %r1496, 0;
|
|
mov.u32 %r1497, %r50;
|
|
mov.u32 %r1499, %r1496;
|
|
|
|
$L__BB11_68:
|
|
clz.b32 %r758, %r1497;
|
|
add.s32 %r759, %r758, %r1496;
|
|
mul.wide.u32 %rd231, %r759, 4;
|
|
mov.u64 %rd232, sobol_burley_table;
|
|
add.s64 %rd233, %rd232, %rd231;
|
|
ld.const.u32 %r760, [%rd233+128];
|
|
xor.b32 %r1499, %r760, %r1499;
|
|
add.s32 %r1496, %r759, 1;
|
|
shl.b32 %r761, %r1497, %r758;
|
|
shl.b32 %r1497, %r761, 1;
|
|
setp.ne.s32 %p69, %r1497, 0;
|
|
@%p69 bra $L__BB11_68;
|
|
|
|
$L__BB11_69:
|
|
xor.b32 %r762, %r49, -525684874;
|
|
brev.b32 %r763, %r50;
|
|
shr.u32 %r764, %r762, 16;
|
|
or.b32 %r765, %r764, 1;
|
|
mul.lo.s32 %r766, %r763, 1025551850;
|
|
xor.b32 %r767, %r766, %r763;
|
|
add.s32 %r768, %r767, %r762;
|
|
mul.lo.s32 %r769, %r768, %r765;
|
|
mul.lo.s32 %r770, %r769, 89287766;
|
|
xor.b32 %r771, %r770, %r769;
|
|
xor.b32 %r772, %r49, -1802089138;
|
|
mul.lo.s32 %r773, %r1499, 1025551850;
|
|
xor.b32 %r774, %r773, %r1499;
|
|
add.s32 %r775, %r774, %r772;
|
|
shr.u32 %r776, %r772, 16;
|
|
or.b32 %r777, %r776, 1;
|
|
mul.lo.s32 %r778, %r775, %r777;
|
|
mul.lo.s32 %r779, %r778, 89287766;
|
|
xor.b32 %r780, %r779, %r778;
|
|
mul.lo.s32 %r781, %r780, 1403136100;
|
|
xor.b32 %r782, %r781, %r780;
|
|
brev.b32 %r783, %r782;
|
|
cvt.rn.f32.u32 %f1479, %r783;
|
|
mul.lo.s32 %r784, %r771, 1403136100;
|
|
xor.b32 %r785, %r784, %r771;
|
|
brev.b32 %r786, %r785;
|
|
cvt.rn.f32.u32 %f1480, %r786;
|
|
mul.ftz.f32 %f3755, %f1479, 0f2F7FFFFE;
|
|
mul.ftz.f32 %f167, %f1480, 0f2F7FFFFE;
|
|
|
|
$L__BB11_70:
|
|
setp.lt.ftz.f32 %p70, %f3749, %f126;
|
|
@%p70 bra $L__BB11_77;
|
|
bra.uni $L__BB11_71;
|
|
|
|
$L__BB11_77:
|
|
mul.ftz.f32 %f1542, %f99, %f167;
|
|
mul.ftz.f32 %f1543, %f1542, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f1544, %f1543;
|
|
mul.ftz.f32 %f1545, %f98, %f1544;
|
|
sub.ftz.f32 %f1546, %f97, %f1545;
|
|
neg.ftz.f32 %f1547, %f1546;
|
|
setp.eq.s16 %p76, %rs89, 0;
|
|
selp.f32 %f3764, %f1546, %f1547, %p76;
|
|
mul.ftz.f32 %f1548, %f3764, %f3764;
|
|
mov.f32 %f1549, 0f3F800000;
|
|
sub.ftz.f32 %f1550, %f1549, %f1548;
|
|
mov.f32 %f3759, 0f00000000;
|
|
max.ftz.f32 %f1551, %f1550, %f3759;
|
|
sqrt.approx.ftz.f32 %f1552, %f1551;
|
|
mul.ftz.f32 %f1553, %f3755, 0f40C90FDB;
|
|
cos.approx.ftz.f32 %f1554, %f1553;
|
|
mul.ftz.f32 %f1555, %f1552, %f1554;
|
|
sin.approx.ftz.f32 %f1556, %f1553;
|
|
mul.ftz.f32 %f1557, %f1552, %f1556;
|
|
div.approx.ftz.f32 %f1559, %f110, %f1558;
|
|
div.approx.ftz.f32 %f1560, %f108, %f1558;
|
|
div.approx.ftz.f32 %f1561, %f109, %f1558;
|
|
mul.ftz.f32 %f1562, %f19, %f1561;
|
|
mul.ftz.f32 %f1563, %f21, %f1560;
|
|
sub.ftz.f32 %f1564, %f1562, %f1563;
|
|
mul.ftz.f32 %f1565, %f21, %f1559;
|
|
mul.ftz.f32 %f1566, %f17, %f1561;
|
|
sub.ftz.f32 %f1567, %f1565, %f1566;
|
|
mul.ftz.f32 %f1568, %f17, %f1560;
|
|
mul.ftz.f32 %f1569, %f19, %f1559;
|
|
sub.ftz.f32 %f1570, %f1568, %f1569;
|
|
mul.ftz.f32 %f1571, %f1557, %f1564;
|
|
mul.ftz.f32 %f1572, %f1557, %f1567;
|
|
mul.ftz.f32 %f1573, %f1557, %f1570;
|
|
fma.rn.ftz.f32 %f1574, %f1555, %f1559, %f1571;
|
|
fma.rn.ftz.f32 %f1575, %f1555, %f1560, %f1572;
|
|
fma.rn.ftz.f32 %f1576, %f1555, %f1561, %f1573;
|
|
fma.rn.ftz.f32 %f3760, %f17, %f3764, %f1574;
|
|
fma.rn.ftz.f32 %f3761, %f19, %f3764, %f1575;
|
|
fma.rn.ftz.f32 %f3762, %f21, %f3764, %f1576;
|
|
mul.ftz.f32 %f1577, %f3766, %f3761;
|
|
fma.rn.ftz.f32 %f1578, %f3767, %f3760, %f1577;
|
|
fma.rn.ftz.f32 %f1579, %f3765, %f3762, %f1578;
|
|
mul.ftz.f32 %f188, %f125, %f125;
|
|
add.ftz.f32 %f1580, %f188, 0f3F800000;
|
|
mul.ftz.f32 %f1581, %f125, 0fC0000000;
|
|
fma.rn.ftz.f32 %f189, %f1581, %f1579, %f1580;
|
|
setp.le.ftz.f32 %p77, %f189, 0f00000000;
|
|
@%p77 bra $L__BB11_79;
|
|
|
|
lg2.approx.ftz.f32 %f1582, %f189;
|
|
mul.ftz.f32 %f1583, %f1582, 0f3FC00000;
|
|
ex2.approx.ftz.f32 %f3759, %f1583;
|
|
|
|
$L__BB11_79:
|
|
sub.ftz.f32 %f1585, %f1549, %f188;
|
|
div.approx.ftz.f32 %f1586, %f1585, %f3759;
|
|
mul.ftz.f32 %f3763, %f1586, 0f3DA2F983;
|
|
bra.uni $L__BB11_80;
|
|
|
|
$L__BB11_71:
|
|
abs.ftz.f32 %f1481, %f125;
|
|
setp.lt.ftz.f32 %p71, %f1481, 0f3A83126F;
|
|
@%p71 bra $L__BB11_75;
|
|
bra.uni $L__BB11_72;
|
|
|
|
$L__BB11_75:
|
|
add.ftz.f32 %f1497, %f167, %f167;
|
|
mov.f32 %f1498, 0f3F800000;
|
|
sub.ftz.f32 %f3758, %f1498, %f1497;
|
|
mov.f32 %f3763, 0f3DA2F983;
|
|
bra.uni $L__BB11_76;
|
|
|
|
$L__BB11_72:
|
|
mul.ftz.f32 %f1483, %f125, %f125;
|
|
mov.f32 %f1484, 0f3F800000;
|
|
sub.ftz.f32 %f171, %f1484, %f1483;
|
|
add.ftz.f32 %f1485, %f125, %f125;
|
|
sub.ftz.f32 %f1486, %f1484, %f125;
|
|
fma.rn.ftz.f32 %f1487, %f1485, %f167, %f1486;
|
|
div.approx.ftz.f32 %f1488, %f171, %f1487;
|
|
add.ftz.f32 %f1489, %f1483, 0f3F800000;
|
|
mul.ftz.f32 %f1490, %f1488, %f1488;
|
|
sub.ftz.f32 %f1491, %f1489, %f1490;
|
|
div.approx.ftz.f32 %f3758, %f1491, %f1485;
|
|
mul.ftz.f32 %f1492, %f1485, %f3758;
|
|
sub.ftz.f32 %f173, %f1489, %f1492;
|
|
setp.le.ftz.f32 %p72, %f173, 0f00000000;
|
|
mov.f32 %f3756, 0f00000000;
|
|
@%p72 bra $L__BB11_74;
|
|
|
|
lg2.approx.ftz.f32 %f1493, %f173;
|
|
mul.ftz.f32 %f1494, %f1493, 0f3FC00000;
|
|
ex2.approx.ftz.f32 %f3756, %f1494;
|
|
|
|
$L__BB11_74:
|
|
div.approx.ftz.f32 %f1495, %f171, %f3756;
|
|
mul.ftz.f32 %f3763, %f1495, 0f3DA2F983;
|
|
|
|
$L__BB11_76:
|
|
mul.ftz.f32 %f1499, %f3758, %f3758;
|
|
mov.f32 %f1500, 0f3F800000;
|
|
sub.ftz.f32 %f1501, %f1500, %f1499;
|
|
mov.f32 %f1502, 0f00000000;
|
|
max.ftz.f32 %f1503, %f1501, %f1502;
|
|
sqrt.approx.ftz.f32 %f1504, %f1503;
|
|
mul.ftz.f32 %f1505, %f3755, 0f40C90FDB;
|
|
cos.approx.ftz.f32 %f1506, %f1505;
|
|
mul.ftz.f32 %f1507, %f1504, %f1506;
|
|
sin.approx.ftz.f32 %f1508, %f1505;
|
|
mul.ftz.f32 %f1509, %f1504, %f1508;
|
|
setp.neu.ftz.f32 %p73, %f3767, %f3765;
|
|
setp.neu.ftz.f32 %p74, %f3767, %f3766;
|
|
or.pred %p75, %p74, %p73;
|
|
neg.ftz.f32 %f1510, %f3765;
|
|
selp.f32 %f1511, %f1510, %f3765, %p75;
|
|
add.ftz.f32 %f1512, %f3767, %f1511;
|
|
neg.ftz.f32 %f1513, %f3766;
|
|
selp.f32 %f1514, %f3766, %f1513, %p75;
|
|
sub.ftz.f32 %f1515, %f1514, %f3767;
|
|
sub.ftz.f32 %f1516, %f3765, %f3766;
|
|
mul.ftz.f32 %f1517, %f1512, %f1512;
|
|
fma.rn.ftz.f32 %f1518, %f1516, %f1516, %f1517;
|
|
fma.rn.ftz.f32 %f1519, %f1515, %f1515, %f1518;
|
|
sqrt.approx.ftz.f32 %f1520, %f1519;
|
|
div.approx.ftz.f32 %f1521, %f1516, %f1520;
|
|
div.approx.ftz.f32 %f1522, %f1512, %f1520;
|
|
div.approx.ftz.f32 %f1523, %f1515, %f1520;
|
|
mul.ftz.f32 %f1524, %f3766, %f1523;
|
|
mul.ftz.f32 %f1525, %f3765, %f1522;
|
|
sub.ftz.f32 %f1526, %f1524, %f1525;
|
|
mul.ftz.f32 %f1527, %f3765, %f1521;
|
|
mul.ftz.f32 %f1528, %f3767, %f1523;
|
|
sub.ftz.f32 %f1529, %f1527, %f1528;
|
|
mul.ftz.f32 %f1530, %f3767, %f1522;
|
|
mul.ftz.f32 %f1531, %f3766, %f1521;
|
|
sub.ftz.f32 %f1532, %f1530, %f1531;
|
|
mul.ftz.f32 %f1533, %f1509, %f1526;
|
|
mul.ftz.f32 %f1534, %f1509, %f1529;
|
|
mul.ftz.f32 %f1535, %f1509, %f1532;
|
|
fma.rn.ftz.f32 %f1536, %f1507, %f1521, %f1533;
|
|
fma.rn.ftz.f32 %f1537, %f1507, %f1522, %f1534;
|
|
fma.rn.ftz.f32 %f1538, %f1507, %f1523, %f1535;
|
|
fma.rn.ftz.f32 %f3760, %f3767, %f3758, %f1536;
|
|
fma.rn.ftz.f32 %f3761, %f3766, %f3758, %f1537;
|
|
fma.rn.ftz.f32 %f3762, %f3765, %f3758, %f1538;
|
|
mul.ftz.f32 %f1539, %f19, %f3761;
|
|
fma.rn.ftz.f32 %f1540, %f17, %f3760, %f1539;
|
|
fma.rn.ftz.f32 %f3764, %f21, %f3762, %f1540;
|
|
|
|
$L__BB11_80:
|
|
st.local.f32 [%rd6+12], %f3760;
|
|
st.local.f32 [%rd6+16], %f3761;
|
|
st.local.f32 [%rd6+20], %f3762;
|
|
sub.ftz.f32 %f1587, %f97, %f3764;
|
|
mul.ftz.f32 %f1588, %f99, %f1587;
|
|
rcp.approx.ftz.f32 %f1589, %f1588;
|
|
mov.f32 %f1590, 0f3F800000;
|
|
mul.ftz.f32 %f1591, %f1589, 0f3E22F983;
|
|
div.approx.ftz.f32 %f3770, %f1591, %f3763;
|
|
add.ftz.f32 %f1592, %f97, %f3764;
|
|
mul.ftz.f32 %f1593, %f99, %f1592;
|
|
rcp.approx.ftz.f32 %f1594, %f1593;
|
|
mul.ftz.f32 %f1595, %f1594, 0f3E22F983;
|
|
div.approx.ftz.f32 %f3772, %f1595, %f3763;
|
|
div.approx.ftz.f32 %f1596, %f3764, %f97;
|
|
sub.ftz.f32 %f3771, %f1590, %f1596;
|
|
add.ftz.f32 %f3773, %f1596, 0f3F800000;
|
|
setp.eq.s16 %p78, %rs89, 0;
|
|
selp.f32 %f1597, %f3771, %f3773, %p78;
|
|
mul.ftz.f32 %f1598, %f3768, %f1597;
|
|
selp.f32 %f3768, %f1598, %f3768, %p70;
|
|
mov.f32 %f3765, %f3762;
|
|
mov.f32 %f3766, %f3761;
|
|
mov.f32 %f3767, %f3760;
|
|
|
|
$L__BB11_81:
|
|
sub.ftz.f32 %f1600, %f1111, %f3748;
|
|
lg2.approx.ftz.f32 %f1601, %f1600;
|
|
mul.ftz.f32 %f1602, %f1601, 0fBF317218;
|
|
div.approx.ftz.f32 %f212, %f1602, %f3768;
|
|
@%p59 bra $L__BB11_83;
|
|
|
|
st.local.f32 [%rd6+28], %f212;
|
|
st.local.u32 [%rd6+40], %r338;
|
|
st.local.u32 [%rd6+36], %r338;
|
|
mov.f32 %f3774, %f212;
|
|
bra.uni $L__BB11_84;
|
|
|
|
$L__BB11_83:
|
|
min.ftz.f32 %f1603, %f124, %f123;
|
|
min.ftz.f32 %f1604, %f1603, %f122;
|
|
mov.f32 %f1605, 0f41200000;
|
|
div.approx.ftz.f32 %f1606, %f1605, %f1604;
|
|
max.ftz.f32 %f3774, %f212, %f1606;
|
|
st.local.f32 [%rd6+28], %f3774;
|
|
|
|
$L__BB11_84:
|
|
mov.u32 %r788, 0;
|
|
st.local.u32 [%rd2], %r788;
|
|
mov.b32 %r58, %f3739;
|
|
setp.nan.ftz.f32 %p81, %f3739, %f3739;
|
|
mov.u16 %rs90, %rs41;
|
|
@%p81 bra $L__BB11_87;
|
|
|
|
or.b32 %r789, %r58, -2147483648;
|
|
setp.ne.s32 %p82, %r789, -2147483648;
|
|
add.ftz.f32 %f1607, %f3739, %f3739;
|
|
setp.eq.ftz.f32 %p83, %f1607, %f3739;
|
|
and.pred %p84, %p83, %p82;
|
|
mov.u16 %rs90, %rs41;
|
|
@%p84 bra $L__BB11_87;
|
|
|
|
shl.b32 %r790, %r58, 1;
|
|
setp.lt.u32 %p85, %r790, -16777215;
|
|
selp.u16 %rs90, 1, 0, %p85;
|
|
|
|
$L__BB11_87:
|
|
setp.eq.s16 %p86, %rs90, 0;
|
|
mov.u16 %rs92, %rs41;
|
|
@%p86 bra $L__BB11_93;
|
|
|
|
mov.b32 %r59, %f3767;
|
|
setp.nan.ftz.f32 %p87, %f3767, %f3767;
|
|
mov.u16 %rs92, 0;
|
|
mov.u16 %rs91, %rs92;
|
|
@%p87 bra $L__BB11_91;
|
|
|
|
or.b32 %r791, %r59, -2147483648;
|
|
setp.ne.s32 %p88, %r791, -2147483648;
|
|
add.ftz.f32 %f1608, %f3767, %f3767;
|
|
setp.eq.ftz.f32 %p89, %f1608, %f3767;
|
|
and.pred %p90, %p89, %p88;
|
|
mov.u16 %rs91, %rs92;
|
|
@%p90 bra $L__BB11_91;
|
|
|
|
shl.b32 %r792, %r59, 1;
|
|
setp.lt.u32 %p91, %r792, -16777215;
|
|
selp.u16 %rs91, 1, 0, %p91;
|
|
|
|
$L__BB11_91:
|
|
setp.eq.s16 %p92, %rs91, 0;
|
|
@%p92 bra $L__BB11_93;
|
|
|
|
mul.ftz.f32 %f1609, %f3766, %f3766;
|
|
fma.rn.ftz.f32 %f1610, %f3767, %f3767, %f1609;
|
|
fma.rn.ftz.f32 %f1611, %f3765, %f3765, %f1610;
|
|
setp.neu.ftz.f32 %p93, %f1611, 0f00000000;
|
|
selp.u16 %rs92, 1, 0, %p93;
|
|
|
|
$L__BB11_93:
|
|
setp.eq.s16 %p94, %rs92, 0;
|
|
selp.b64 %rd234, 0, %rd703, %p94;
|
|
ld.local.f32 %f1618, [%rd6+24];
|
|
ld.local.f32 %f1620, [%rd6+32];
|
|
mov.u32 %r826, 255;
|
|
mov.u32 %r828, 2;
|
|
mov.u32 %r831, 8;
|
|
mov.u32 %r837, 1;
|
|
// begin inline asm
|
|
call(%r793,%r794,%r795,%r796,%r797,%r798,%r799,%r800,%r801,%r802,%r803,%r804,%r805,%r806,%r807,%r808,%r809,%r810,%r811,%r812,%r813,%r814,%r815,%r816,%r817,%r818,%r819,%r820,%r821,%r822,%r823,%r824),_optix_trace_typed_32,(%r788,%rd234,%f3739,%f3738,%f3737,%f3767,%f3766,%f3765,%f1618,%f3774,%f1620,%r826,%r828,%r828,%r788,%r788,%r831,%r788,%r788,%r34,%r35,%r22,%r837,%r36,%r37,%r864,%r865,%r866,%r867,%r868,%r869,%r870,%r871,%r872,%r873,%r874,%r875,%r876,%r877,%r878,%r879,%r880,%r881,%r882,%r883,%r884,%r885,%r886,%r887);
|
|
// end inline asm
|
|
ld.local.u32 %r60, [%rd2];
|
|
setp.gt.s32 %p3, %r60, 0;
|
|
@%p3 bra $L__BB11_95;
|
|
bra.uni $L__BB11_94;
|
|
|
|
$L__BB11_95:
|
|
ld.local.f32 %f3775, [%rd2+4];
|
|
st.local.f32 [%rd6+28], %f3775;
|
|
bra.uni $L__BB11_96;
|
|
|
|
$L__BB11_94:
|
|
ld.local.f32 %f3775, [%rd6+28];
|
|
|
|
$L__BB11_96:
|
|
setp.gt.s32 %p372, %r60, 0;
|
|
setp.ne.s32 %p95, %r1494, 0;
|
|
@%p95 bra $L__BB11_100;
|
|
|
|
setp.lt.s32 %p96, %r60, 1;
|
|
@%p96 bra $L__BB11_99;
|
|
|
|
ld.local.f32 %f1621, [%rd6+12];
|
|
ld.local.f32 %f1622, [%rd6+16];
|
|
ld.local.f32 %f1623, [%rd6+20];
|
|
ld.local.f32 %f1624, [%rd6];
|
|
fma.rn.ftz.f32 %f1625, %f3775, %f1621, %f1624;
|
|
ld.local.f32 %f1626, [%rd6+4];
|
|
fma.rn.ftz.f32 %f1627, %f3775, %f1622, %f1626;
|
|
ld.local.f32 %f1628, [%rd6+8];
|
|
fma.rn.ftz.f32 %f1629, %f3775, %f1623, %f1628;
|
|
sub.ftz.f32 %f1630, %f1625, %f14;
|
|
sub.ftz.f32 %f1631, %f1627, %f15;
|
|
sub.ftz.f32 %f1632, %f1629, %f16;
|
|
mul.ftz.f32 %f1633, %f1631, %f20;
|
|
fma.rn.ftz.f32 %f1634, %f1630, %f18, %f1633;
|
|
fma.rn.ftz.f32 %f3777, %f1632, %f22, %f1634;
|
|
mov.u16 %rs94, %rs50;
|
|
|
|
$L__BB11_99:
|
|
setp.lt.ftz.f32 %p372, %f3775, %f212;
|
|
|
|
$L__BB11_100:
|
|
ld.local.f32 %f3767, [%rd6+12];
|
|
selp.f32 %f222, %f3775, %f212, %p372;
|
|
ld.local.f32 %f3766, [%rd6+16];
|
|
ld.local.f32 %f3765, [%rd6+20];
|
|
ld.local.f32 %f1635, [%rd6];
|
|
fma.rn.ftz.f32 %f3739, %f3767, %f222, %f1635;
|
|
ld.local.f32 %f1636, [%rd6+4];
|
|
fma.rn.ftz.f32 %f3738, %f222, %f3766, %f1636;
|
|
ld.local.f32 %f1637, [%rd6+8];
|
|
fma.rn.ftz.f32 %f3737, %f222, %f3765, %f1637;
|
|
st.local.f32 [%rd6], %f3739;
|
|
st.local.f32 [%rd6+4], %f3738;
|
|
st.local.f32 [%rd6+8], %f3737;
|
|
mul.ftz.f32 %f1638, %f124, %f222;
|
|
mul.ftz.f32 %f1639, %f123, %f222;
|
|
mul.ftz.f32 %f1640, %f122, %f222;
|
|
mul.ftz.f32 %f1641, %f1638, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f228, %f1641;
|
|
mul.ftz.f32 %f1642, %f1639, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f229, %f1642;
|
|
mul.ftz.f32 %f1643, %f1640, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f230, %f1643;
|
|
@%p372 bra $L__BB11_102;
|
|
bra.uni $L__BB11_101;
|
|
|
|
$L__BB11_102:
|
|
mov.f32 %f3792, %f228;
|
|
mov.f32 %f3791, %f229;
|
|
mov.f32 %f3790, %f230;
|
|
bra.uni $L__BB11_103;
|
|
|
|
$L__BB11_101:
|
|
mul.ftz.f32 %f3790, %f122, %f230;
|
|
mul.ftz.f32 %f3791, %f123, %f229;
|
|
mul.ftz.f32 %f3792, %f124, %f228;
|
|
|
|
$L__BB11_103:
|
|
@%p59 bra $L__BB11_112;
|
|
|
|
mul.ftz.f32 %f243, %f124, %f3771;
|
|
mul.ftz.f32 %f1644, %f243, %f222;
|
|
mul.ftz.f32 %f244, %f123, %f3771;
|
|
mul.ftz.f32 %f1645, %f244, %f222;
|
|
mul.ftz.f32 %f245, %f122, %f3771;
|
|
mul.ftz.f32 %f1646, %f245, %f222;
|
|
mul.ftz.f32 %f1647, %f1644, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f3781, %f1647;
|
|
mul.ftz.f32 %f1648, %f1645, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f3782, %f1648;
|
|
mul.ftz.f32 %f1649, %f1646, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f3783, %f1649;
|
|
@%p372 bra $L__BB11_106;
|
|
|
|
mul.ftz.f32 %f3783, %f245, %f3783;
|
|
mul.ftz.f32 %f3782, %f244, %f3782;
|
|
mul.ftz.f32 %f3781, %f243, %f3781;
|
|
|
|
$L__BB11_106:
|
|
and.b16 %rs51, %rs94, 255;
|
|
setp.eq.s16 %p98, %rs51, 0;
|
|
@%p98 bra $L__BB11_110;
|
|
|
|
mul.ftz.f32 %f261, %f124, %f3773;
|
|
mul.ftz.f32 %f1650, %f261, %f222;
|
|
mul.ftz.f32 %f262, %f123, %f3773;
|
|
mul.ftz.f32 %f1651, %f262, %f222;
|
|
mul.ftz.f32 %f263, %f122, %f3773;
|
|
mul.ftz.f32 %f1652, %f263, %f222;
|
|
mul.ftz.f32 %f1653, %f1650, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f3784, %f1653;
|
|
mul.ftz.f32 %f1654, %f1651, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f3785, %f1654;
|
|
mul.ftz.f32 %f1655, %f1652, 0fBFB8AA3B;
|
|
ex2.approx.ftz.f32 %f3786, %f1655;
|
|
@%p372 bra $L__BB11_109;
|
|
|
|
mul.ftz.f32 %f3786, %f263, %f3786;
|
|
mul.ftz.f32 %f3785, %f262, %f3785;
|
|
mul.ftz.f32 %f3784, %f261, %f3784;
|
|
|
|
$L__BB11_109:
|
|
mul.ftz.f32 %f1656, %f3772, %f3784;
|
|
mul.ftz.f32 %f1657, %f3772, %f3785;
|
|
mul.ftz.f32 %f1658, %f3772, %f3786;
|
|
mul.ftz.f32 %f1659, %f3770, %f3781;
|
|
sub.ftz.f32 %f1660, %f1656, %f1659;
|
|
mul.ftz.f32 %f1661, %f3770, %f3782;
|
|
sub.ftz.f32 %f1662, %f1657, %f1661;
|
|
mul.ftz.f32 %f1663, %f3770, %f3783;
|
|
sub.ftz.f32 %f1664, %f1658, %f1663;
|
|
fma.rn.ftz.f32 %f3789, %f3769, %f1660, %f1659;
|
|
fma.rn.ftz.f32 %f3788, %f3769, %f1662, %f1661;
|
|
fma.rn.ftz.f32 %f3787, %f3769, %f1664, %f1663;
|
|
bra.uni $L__BB11_111;
|
|
|
|
$L__BB11_110:
|
|
mul.ftz.f32 %f3789, %f3770, %f3781;
|
|
mul.ftz.f32 %f3788, %f3770, %f3782;
|
|
mul.ftz.f32 %f3787, %f3770, %f3783;
|
|
|
|
$L__BB11_111:
|
|
sub.ftz.f32 %f1665, %f3789, %f3792;
|
|
sub.ftz.f32 %f1666, %f3788, %f3791;
|
|
sub.ftz.f32 %f1667, %f3787, %f3790;
|
|
fma.rn.ftz.f32 %f3792, %f126, %f1665, %f3792;
|
|
fma.rn.ftz.f32 %f3791, %f126, %f1666, %f3791;
|
|
fma.rn.ftz.f32 %f3790, %f126, %f1667, %f3790;
|
|
|
|
$L__BB11_112:
|
|
setp.lt.u32 %p370, %r1494, 10;
|
|
selp.f32 %f1668, %f96, %f105, %p370;
|
|
mul.ftz.f32 %f1669, %f1668, %f230;
|
|
selp.f32 %f1670, %f230, %f1669, %p372;
|
|
selp.f32 %f1671, %f95, %f103, %p370;
|
|
mul.ftz.f32 %f1672, %f1671, %f229;
|
|
selp.f32 %f1673, %f229, %f1672, %p372;
|
|
selp.f32 %f1674, %f94, %f101, %p370;
|
|
mul.ftz.f32 %f1675, %f1674, %f228;
|
|
selp.f32 %f1676, %f228, %f1675, %p372;
|
|
mul.ftz.f32 %f1677, %f3791, %f3746;
|
|
fma.rn.ftz.f32 %f1678, %f3792, %f3745, %f1677;
|
|
fma.rn.ftz.f32 %f1679, %f3790, %f3747, %f1678;
|
|
div.approx.ftz.f32 %f1680, %f1676, %f1679;
|
|
div.approx.ftz.f32 %f1681, %f1673, %f1679;
|
|
div.approx.ftz.f32 %f1682, %f1670, %f1679;
|
|
mul.ftz.f32 %f3728, %f1680, %f3728;
|
|
mul.ftz.f32 %f3730, %f1681, %f3730;
|
|
mul.ftz.f32 %f3732, %f1682, %f3732;
|
|
@%p372 bra $L__BB11_114;
|
|
|
|
max.ftz.f32 %f1683, %f3728, %f3730;
|
|
max.ftz.f32 %f1684, %f1683, %f3732;
|
|
setp.geu.ftz.f32 %p99, %f1684, 0f358637BD;
|
|
add.s32 %r1494, %r1494, 1;
|
|
setp.lt.u32 %p100, %r1494, 256;
|
|
and.pred %p101, %p99, %p100;
|
|
add.s32 %r1495, %r1495, 16;
|
|
@%p101 bra $L__BB11_39;
|
|
|
|
$L__BB11_114:
|
|
mov.pred %p375, -1;
|
|
not.pred %p103, %p372;
|
|
@%p103 bra $L__BB11_369;
|
|
|
|
ld.const.u64 %rd707, [kernel_params+2648];
|
|
ld.const.u64 %rd706, [kernel_params+2736];
|
|
ld.const.u64 %rd704, [kernel_params+2688];
|
|
ld.const.u64 %rd705, [kernel_params+2680];
|
|
st.global.f32 [%rd18], %f3728;
|
|
st.global.f32 [%rd18+4], %f3730;
|
|
st.global.f32 [%rd18+8], %f3732;
|
|
bra.uni $L__BB11_224;
|
|
|
|
$L__BB11_125:
|
|
add.ftz.f32 %f1715, %f306, 0fBF400000;
|
|
mul.ftz.f32 %f3806, %f1715, 0f40800000;
|
|
mov.f32 %f3809, 0f3F000000;
|
|
mov.f32 %f3807, 0f3E800000;
|
|
mov.f32 %f3797, %f316;
|
|
mov.f32 %f3798, %f315;
|
|
mov.f32 %f3799, %f317;
|
|
mov.f32 %f3800, %f318;
|
|
mov.f32 %f3801, %f319;
|
|
mov.f32 %f3802, %f320;
|
|
mov.f32 %f3808, %f3807;
|
|
|
|
$L__BB11_128:
|
|
setp.leu.ftz.f32 %p114, %f312, 0f00000000;
|
|
setp.gt.ftz.f32 %p115, %f312, 0f00000000;
|
|
mov.f32 %f3810, 0f00000000;
|
|
selp.f32 %f1724, 0f3F800000, 0f00000000, %p115;
|
|
add.ftz.f32 %f1725, %f1724, 0f3F800000;
|
|
setp.gt.ftz.f32 %p116, %f313, 0f00000000;
|
|
selp.f32 %f1726, %f1725, %f1724, %p116;
|
|
add.ftz.f32 %f1727, %f1726, 0f3F800000;
|
|
setp.gt.ftz.f32 %p117, %f314, 0f00000000;
|
|
selp.f32 %f341, %f1727, %f1726, %p117;
|
|
mul.ftz.f32 %f3815, %f3795, %f341;
|
|
@%p114 bra $L__BB11_130;
|
|
|
|
setp.lt.ftz.f32 %p118, %f3815, 0f3F800000;
|
|
mov.f32 %f3810, 0f3F800000;
|
|
mov.f32 %f3812, 0f00000000;
|
|
mov.f32 %f3813, %f312;
|
|
@%p118 bra $L__BB11_133;
|
|
|
|
$L__BB11_130:
|
|
setp.leu.ftz.f32 %p119, %f313, 0f00000000;
|
|
mov.f32 %f3811, %f3810;
|
|
@%p119 bra $L__BB11_132;
|
|
|
|
add.ftz.f32 %f3811, %f3810, 0f3F800000;
|
|
setp.lt.ftz.f32 %p120, %f3815, %f3811;
|
|
mov.f32 %f3812, %f3810;
|
|
mov.f32 %f3813, %f313;
|
|
@%p120 bra $L__BB11_133;
|
|
|
|
$L__BB11_132:
|
|
add.ftz.f32 %f1731, %f3811, 0f3F800000;
|
|
setp.geu.ftz.f32 %p121, %f3815, %f1731;
|
|
setp.leu.ftz.f32 %p122, %f314, 0f00000000;
|
|
mov.f32 %f3814, 0f00000000;
|
|
or.pred %p123, %p122, %p121;
|
|
mov.f32 %f3812, %f3811;
|
|
mov.f32 %f3813, %f314;
|
|
@%p123 bra $L__BB11_134;
|
|
|
|
$L__BB11_133:
|
|
sub.ftz.f32 %f3815, %f3815, %f3812;
|
|
mov.f32 %f3814, %f3813;
|
|
|
|
$L__BB11_134:
|
|
mul.ftz.f32 %f351, %f3815, 0f3F7F12B2;
|
|
setp.gtu.ftz.f32 %p124, %f351, 0f3F666666;
|
|
mov.f32 %f3817, 0f41700000;
|
|
@%p124 bra $L__BB11_136;
|
|
|
|
mul.ftz.f32 %f1733, %f351, %f351;
|
|
mul.ftz.f32 %f1734, %f1733, 0f4019999A;
|
|
mul.ftz.f32 %f1735, %f1734, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1736, %f1735;
|
|
add.ftz.f32 %f3817, %f1736, 0fBF800000;
|
|
|
|
$L__BB11_136:
|
|
neg.ftz.f32 %f1737, %f3817;
|
|
mov.f32 %f1738, 0f40400000;
|
|
div.approx.ftz.f32 %f1739, %f1737, %f1738;
|
|
mul.ftz.f32 %f1740, %f1739, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1741, %f1740;
|
|
mul.ftz.f32 %f1742, %f1741, %f1741;
|
|
mul.ftz.f32 %f1743, %f1741, %f1742;
|
|
mul.ftz.f32 %f1744, %f1743, 0f3E800000;
|
|
mov.f32 %f1745, 0f3F800000;
|
|
sub.ftz.f32 %f1746, %f1745, %f1744;
|
|
fma.rn.ftz.f32 %f1747, %f1741, 0fBF400000, %f1746;
|
|
sub.ftz.f32 %f354, %f1747, %f351;
|
|
fma.rn.ftz.f32 %f355, %f1741, 0f3E800000, %f1744;
|
|
abs.ftz.f32 %f1748, %f354;
|
|
setp.lt.ftz.f32 %p125, %f1748, 0f358637BD;
|
|
setp.eq.ftz.f32 %p126, %f355, 0f00000000;
|
|
or.pred %p127, %p125, %p126;
|
|
@%p127 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1749, %f354, %f355;
|
|
sub.ftz.f32 %f1750, %f3817, %f1749;
|
|
setp.lt.ftz.f32 %p128, %f1750, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1750, %p128;
|
|
neg.ftz.f32 %f1751, %f3817;
|
|
div.approx.ftz.f32 %f1753, %f1751, %f1738;
|
|
mul.ftz.f32 %f1754, %f1753, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1755, %f1754;
|
|
mul.ftz.f32 %f1756, %f1755, %f1755;
|
|
mul.ftz.f32 %f1757, %f1755, %f1756;
|
|
mul.ftz.f32 %f1758, %f1757, 0f3E800000;
|
|
sub.ftz.f32 %f1760, %f1745, %f1758;
|
|
fma.rn.ftz.f32 %f1761, %f1755, 0fBF400000, %f1760;
|
|
sub.ftz.f32 %f357, %f1761, %f351;
|
|
fma.rn.ftz.f32 %f358, %f1755, 0f3E800000, %f1758;
|
|
abs.ftz.f32 %f1762, %f357;
|
|
setp.lt.ftz.f32 %p129, %f1762, 0f358637BD;
|
|
setp.eq.ftz.f32 %p130, %f358, 0f00000000;
|
|
or.pred %p131, %p129, %p130;
|
|
@%p131 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1763, %f357, %f358;
|
|
sub.ftz.f32 %f1764, %f3817, %f1763;
|
|
setp.lt.ftz.f32 %p132, %f1764, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1764, %p132;
|
|
neg.ftz.f32 %f1765, %f3817;
|
|
div.approx.ftz.f32 %f1767, %f1765, %f1738;
|
|
mul.ftz.f32 %f1768, %f1767, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1769, %f1768;
|
|
mul.ftz.f32 %f1770, %f1769, %f1769;
|
|
mul.ftz.f32 %f1771, %f1769, %f1770;
|
|
mul.ftz.f32 %f1772, %f1771, 0f3E800000;
|
|
sub.ftz.f32 %f1774, %f1745, %f1772;
|
|
fma.rn.ftz.f32 %f1775, %f1769, 0fBF400000, %f1774;
|
|
sub.ftz.f32 %f360, %f1775, %f351;
|
|
fma.rn.ftz.f32 %f361, %f1769, 0f3E800000, %f1772;
|
|
abs.ftz.f32 %f1776, %f360;
|
|
setp.lt.ftz.f32 %p133, %f1776, 0f358637BD;
|
|
setp.eq.ftz.f32 %p134, %f361, 0f00000000;
|
|
or.pred %p135, %p133, %p134;
|
|
@%p135 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1777, %f360, %f361;
|
|
sub.ftz.f32 %f1778, %f3817, %f1777;
|
|
setp.lt.ftz.f32 %p136, %f1778, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1778, %p136;
|
|
neg.ftz.f32 %f1779, %f3817;
|
|
div.approx.ftz.f32 %f1781, %f1779, %f1738;
|
|
mul.ftz.f32 %f1782, %f1781, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1783, %f1782;
|
|
mul.ftz.f32 %f1784, %f1783, %f1783;
|
|
mul.ftz.f32 %f1785, %f1783, %f1784;
|
|
mul.ftz.f32 %f1786, %f1785, 0f3E800000;
|
|
sub.ftz.f32 %f1788, %f1745, %f1786;
|
|
fma.rn.ftz.f32 %f1789, %f1783, 0fBF400000, %f1788;
|
|
sub.ftz.f32 %f363, %f1789, %f351;
|
|
fma.rn.ftz.f32 %f364, %f1783, 0f3E800000, %f1786;
|
|
abs.ftz.f32 %f1790, %f363;
|
|
setp.lt.ftz.f32 %p137, %f1790, 0f358637BD;
|
|
setp.eq.ftz.f32 %p138, %f364, 0f00000000;
|
|
or.pred %p139, %p137, %p138;
|
|
@%p139 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1791, %f363, %f364;
|
|
sub.ftz.f32 %f1792, %f3817, %f1791;
|
|
setp.lt.ftz.f32 %p140, %f1792, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1792, %p140;
|
|
neg.ftz.f32 %f1793, %f3817;
|
|
div.approx.ftz.f32 %f1795, %f1793, %f1738;
|
|
mul.ftz.f32 %f1796, %f1795, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1797, %f1796;
|
|
mul.ftz.f32 %f1798, %f1797, %f1797;
|
|
mul.ftz.f32 %f1799, %f1797, %f1798;
|
|
mul.ftz.f32 %f1800, %f1799, 0f3E800000;
|
|
sub.ftz.f32 %f1802, %f1745, %f1800;
|
|
fma.rn.ftz.f32 %f1803, %f1797, 0fBF400000, %f1802;
|
|
sub.ftz.f32 %f366, %f1803, %f351;
|
|
fma.rn.ftz.f32 %f367, %f1797, 0f3E800000, %f1800;
|
|
abs.ftz.f32 %f1804, %f366;
|
|
setp.lt.ftz.f32 %p141, %f1804, 0f358637BD;
|
|
setp.eq.ftz.f32 %p142, %f367, 0f00000000;
|
|
or.pred %p143, %p141, %p142;
|
|
@%p143 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1805, %f366, %f367;
|
|
sub.ftz.f32 %f1806, %f3817, %f1805;
|
|
setp.lt.ftz.f32 %p144, %f1806, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1806, %p144;
|
|
neg.ftz.f32 %f1807, %f3817;
|
|
div.approx.ftz.f32 %f1809, %f1807, %f1738;
|
|
mul.ftz.f32 %f1810, %f1809, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1811, %f1810;
|
|
mul.ftz.f32 %f1812, %f1811, %f1811;
|
|
mul.ftz.f32 %f1813, %f1811, %f1812;
|
|
mul.ftz.f32 %f1814, %f1813, 0f3E800000;
|
|
sub.ftz.f32 %f1816, %f1745, %f1814;
|
|
fma.rn.ftz.f32 %f1817, %f1811, 0fBF400000, %f1816;
|
|
sub.ftz.f32 %f369, %f1817, %f351;
|
|
fma.rn.ftz.f32 %f370, %f1811, 0f3E800000, %f1814;
|
|
abs.ftz.f32 %f1818, %f369;
|
|
setp.lt.ftz.f32 %p145, %f1818, 0f358637BD;
|
|
setp.eq.ftz.f32 %p146, %f370, 0f00000000;
|
|
or.pred %p147, %p145, %p146;
|
|
@%p147 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1819, %f369, %f370;
|
|
sub.ftz.f32 %f1820, %f3817, %f1819;
|
|
setp.lt.ftz.f32 %p148, %f1820, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1820, %p148;
|
|
neg.ftz.f32 %f1821, %f3817;
|
|
div.approx.ftz.f32 %f1823, %f1821, %f1738;
|
|
mul.ftz.f32 %f1824, %f1823, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1825, %f1824;
|
|
mul.ftz.f32 %f1826, %f1825, %f1825;
|
|
mul.ftz.f32 %f1827, %f1825, %f1826;
|
|
mul.ftz.f32 %f1828, %f1827, 0f3E800000;
|
|
sub.ftz.f32 %f1830, %f1745, %f1828;
|
|
fma.rn.ftz.f32 %f1831, %f1825, 0fBF400000, %f1830;
|
|
sub.ftz.f32 %f372, %f1831, %f351;
|
|
fma.rn.ftz.f32 %f373, %f1825, 0f3E800000, %f1828;
|
|
abs.ftz.f32 %f1832, %f372;
|
|
setp.lt.ftz.f32 %p149, %f1832, 0f358637BD;
|
|
setp.eq.ftz.f32 %p150, %f373, 0f00000000;
|
|
or.pred %p151, %p149, %p150;
|
|
@%p151 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1833, %f372, %f373;
|
|
sub.ftz.f32 %f1834, %f3817, %f1833;
|
|
setp.lt.ftz.f32 %p152, %f1834, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1834, %p152;
|
|
neg.ftz.f32 %f1835, %f3817;
|
|
div.approx.ftz.f32 %f1837, %f1835, %f1738;
|
|
mul.ftz.f32 %f1838, %f1837, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1839, %f1838;
|
|
mul.ftz.f32 %f1840, %f1839, %f1839;
|
|
mul.ftz.f32 %f1841, %f1839, %f1840;
|
|
mul.ftz.f32 %f1842, %f1841, 0f3E800000;
|
|
sub.ftz.f32 %f1844, %f1745, %f1842;
|
|
fma.rn.ftz.f32 %f1845, %f1839, 0fBF400000, %f1844;
|
|
sub.ftz.f32 %f375, %f1845, %f351;
|
|
fma.rn.ftz.f32 %f376, %f1839, 0f3E800000, %f1842;
|
|
abs.ftz.f32 %f1846, %f375;
|
|
setp.lt.ftz.f32 %p153, %f1846, 0f358637BD;
|
|
setp.eq.ftz.f32 %p154, %f376, 0f00000000;
|
|
or.pred %p155, %p153, %p154;
|
|
@%p155 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1847, %f375, %f376;
|
|
sub.ftz.f32 %f1848, %f3817, %f1847;
|
|
setp.lt.ftz.f32 %p156, %f1848, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1848, %p156;
|
|
neg.ftz.f32 %f1849, %f3817;
|
|
div.approx.ftz.f32 %f1851, %f1849, %f1738;
|
|
mul.ftz.f32 %f1852, %f1851, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1853, %f1852;
|
|
mul.ftz.f32 %f1854, %f1853, %f1853;
|
|
mul.ftz.f32 %f1855, %f1853, %f1854;
|
|
mul.ftz.f32 %f1856, %f1855, 0f3E800000;
|
|
sub.ftz.f32 %f1858, %f1745, %f1856;
|
|
fma.rn.ftz.f32 %f1859, %f1853, 0fBF400000, %f1858;
|
|
sub.ftz.f32 %f378, %f1859, %f351;
|
|
fma.rn.ftz.f32 %f379, %f1853, 0f3E800000, %f1856;
|
|
abs.ftz.f32 %f1860, %f378;
|
|
setp.lt.ftz.f32 %p157, %f1860, 0f358637BD;
|
|
setp.eq.ftz.f32 %p158, %f379, 0f00000000;
|
|
or.pred %p159, %p157, %p158;
|
|
@%p159 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1861, %f378, %f379;
|
|
sub.ftz.f32 %f1862, %f3817, %f1861;
|
|
setp.lt.ftz.f32 %p160, %f1862, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1862, %p160;
|
|
neg.ftz.f32 %f1863, %f3817;
|
|
div.approx.ftz.f32 %f1865, %f1863, %f1738;
|
|
mul.ftz.f32 %f1866, %f1865, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1867, %f1866;
|
|
mul.ftz.f32 %f1868, %f1867, %f1867;
|
|
mul.ftz.f32 %f1869, %f1867, %f1868;
|
|
mul.ftz.f32 %f1870, %f1869, 0f3E800000;
|
|
sub.ftz.f32 %f1872, %f1745, %f1870;
|
|
fma.rn.ftz.f32 %f1873, %f1867, 0fBF400000, %f1872;
|
|
sub.ftz.f32 %f381, %f1873, %f351;
|
|
fma.rn.ftz.f32 %f382, %f1867, 0f3E800000, %f1870;
|
|
abs.ftz.f32 %f1874, %f381;
|
|
setp.lt.ftz.f32 %p161, %f1874, 0f358637BD;
|
|
setp.eq.ftz.f32 %p162, %f382, 0f00000000;
|
|
or.pred %p163, %p161, %p162;
|
|
@%p163 bra $L__BB11_147;
|
|
|
|
div.approx.ftz.f32 %f1875, %f381, %f382;
|
|
sub.ftz.f32 %f1876, %f3817, %f1875;
|
|
setp.lt.ftz.f32 %p164, %f1876, 0f00000000;
|
|
selp.f32 %f3817, 0f00000000, %f1876, %p164;
|
|
|
|
$L__BB11_147:
|
|
mul.ftz.f32 %f1877, %f3806, 0f40C90FDB;
|
|
mul.ftz.f32 %f1878, %f3814, 0f41800000;
|
|
mul.ftz.f32 %f385, %f3814, %f3817;
|
|
mul.ftz.f32 %f1879, %f385, %f385;
|
|
mul.ftz.f32 %f1880, %f1878, %f1878;
|
|
sub.ftz.f32 %f1881, %f1880, %f1879;
|
|
mov.f32 %f3855, 0f00000000;
|
|
max.ftz.f32 %f1883, %f1881, %f3855;
|
|
sqrt.approx.ftz.f32 %f1884, %f1883;
|
|
cos.approx.ftz.f32 %f1885, %f1877;
|
|
mul.ftz.f32 %f1886, %f385, %f1885;
|
|
sin.approx.ftz.f32 %f1887, %f1877;
|
|
mul.ftz.f32 %f1888, %f385, %f1887;
|
|
mul.ftz.f32 %f1889, %f3797, %f1888;
|
|
mul.ftz.f32 %f1890, %f3798, %f1888;
|
|
mul.ftz.f32 %f1891, %f3799, %f1888;
|
|
fma.rn.ftz.f32 %f1892, %f3800, %f1886, %f1889;
|
|
fma.rn.ftz.f32 %f1893, %f3801, %f1886, %f1890;
|
|
fma.rn.ftz.f32 %f1894, %f3802, %f1886, %f1891;
|
|
fma.rn.ftz.f32 %f1895, %f3803, %f1884, %f307;
|
|
fma.rn.ftz.f32 %f1896, %f3804, %f1884, %f308;
|
|
fma.rn.ftz.f32 %f1897, %f3805, %f1884, %f309;
|
|
add.ftz.f32 %f386, %f1895, %f1892;
|
|
add.ftz.f32 %f387, %f1896, %f1893;
|
|
add.ftz.f32 %f388, %f1897, %f1894;
|
|
st.local.f32 [%rd6], %f386;
|
|
st.local.f32 [%rd6+4], %f387;
|
|
st.local.f32 [%rd6+8], %f388;
|
|
neg.ftz.f32 %f389, %f3803;
|
|
st.local.f32 [%rd6+12], %f389;
|
|
neg.ftz.f32 %f390, %f3804;
|
|
st.local.f32 [%rd6+16], %f390;
|
|
neg.ftz.f32 %f391, %f3805;
|
|
st.local.f32 [%rd6+20], %f391;
|
|
mov.u32 %r1504, 0;
|
|
st.local.u32 [%rd6+24], %r1504;
|
|
add.ftz.f32 %f392, %f1884, %f1884;
|
|
st.local.f32 [%rd6+28], %f392;
|
|
st.local.f32 [%rd6+52], %f310;
|
|
st.local.u32 [%rd6+56], %r1504;
|
|
st.local.f32 [%rd6+32], %f311;
|
|
mov.u32 %r1003, -1;
|
|
st.local.u32 [%rd6+40], %r1003;
|
|
st.local.u32 [%rd6+36], %r1003;
|
|
st.local.u32 [%rd6+48], %r1003;
|
|
st.local.u32 [%rd6+44], %r1003;
|
|
add.s32 %r1004, %r4, %r1495;
|
|
mad.lo.s32 %r1005, %r6, 1757159915, %r1004;
|
|
mad.lo.s32 %r1006, %r1005, 1103515245, 12345;
|
|
st.local.u32 [%rd8], %r1006;
|
|
st.local.u32 [%rd2], %r1504;
|
|
mov.b32 %r75, %f386;
|
|
setp.nan.ftz.f32 %p165, %f386, %f386;
|
|
mov.u16 %rs95, 0;
|
|
@%p165 bra $L__BB11_150;
|
|
|
|
or.b32 %r1007, %r75, -2147483648;
|
|
setp.ne.s32 %p166, %r1007, -2147483648;
|
|
add.ftz.f32 %f1898, %f386, %f386;
|
|
setp.eq.ftz.f32 %p167, %f1898, %f386;
|
|
and.pred %p168, %p167, %p166;
|
|
@%p168 bra $L__BB11_150;
|
|
|
|
shl.b32 %r1008, %r75, 1;
|
|
setp.lt.u32 %p169, %r1008, -16777215;
|
|
selp.u16 %rs95, 1, 0, %p169;
|
|
|
|
$L__BB11_150:
|
|
setp.eq.s16 %p171, %rs95, 0;
|
|
mov.pred %p375, -1;
|
|
mov.pred %p373, %p375;
|
|
@%p171 bra $L__BB11_156;
|
|
|
|
mov.b32 %r76, %f389;
|
|
setp.nan.ftz.f32 %p172, %f3803, %f3803;
|
|
mov.u16 %rs96, 0;
|
|
@%p172 bra $L__BB11_154;
|
|
|
|
or.b32 %r1009, %r76, -2147483648;
|
|
setp.ne.s32 %p173, %r1009, -2147483648;
|
|
add.ftz.f32 %f1899, %f3803, %f3803;
|
|
setp.eq.ftz.f32 %p174, %f1899, %f3803;
|
|
and.pred %p175, %p174, %p173;
|
|
@%p175 bra $L__BB11_154;
|
|
|
|
shl.b32 %r1010, %r76, 1;
|
|
setp.lt.u32 %p176, %r1010, -16777215;
|
|
selp.u16 %rs96, 1, 0, %p176;
|
|
|
|
$L__BB11_154:
|
|
setp.eq.s16 %p178, %rs96, 0;
|
|
mov.pred %p373, %p375;
|
|
@%p178 bra $L__BB11_156;
|
|
|
|
mul.ftz.f32 %f1900, %f3804, %f3804;
|
|
fma.rn.ftz.f32 %f1901, %f3803, %f3803, %f1900;
|
|
fma.rn.ftz.f32 %f1902, %f3805, %f3805, %f1901;
|
|
setp.eq.ftz.f32 %p373, %f1902, 0f00000000;
|
|
|
|
$L__BB11_156:
|
|
ld.const.u64 %rd703, [kernel_params+2096];
|
|
selp.b64 %rd258, 0, %rd703, %p373;
|
|
cvt.u32.u64 %r1050, %rd164;
|
|
shr.u64 %rd260, %rd164, 32;
|
|
cvt.u32.u64 %r1051, %rd260;
|
|
cvt.u32.u64 %r1052, %rd146;
|
|
shr.u64 %rd262, %rd146, 32;
|
|
cvt.u32.u64 %r1053, %rd262;
|
|
cvt.u32.u64 %r1056, %rd162;
|
|
shr.u64 %rd264, %rd162, 32;
|
|
cvt.u32.u64 %r1057, %rd264;
|
|
mov.u32 %r1044, 255;
|
|
mov.u32 %r1046, 2;
|
|
mov.u32 %r1049, 8;
|
|
mov.u32 %r1055, 4;
|
|
// begin inline asm
|
|
call(%r1011,%r1012,%r1013,%r1014,%r1015,%r1016,%r1017,%r1018,%r1019,%r1020,%r1021,%r1022,%r1023,%r1024,%r1025,%r1026,%r1027,%r1028,%r1029,%r1030,%r1031,%r1032,%r1033,%r1034,%r1035,%r1036,%r1037,%r1038,%r1039,%r1040,%r1041,%r1042),_optix_trace_typed_32,(%r1504,%rd258,%f386,%f387,%f388,%f389,%f390,%f391,%f3855,%f392,%f311,%r1044,%r1046,%r1046,%r1504,%r1504,%r1049,%r1050,%r1051,%r1052,%r1053,%r74,%r1055,%r1056,%r1057,%r1082,%r1083,%r1084,%r1085,%r1086,%r1087,%r1088,%r1089,%r1090,%r1091,%r1092,%r1093,%r1094,%r1095,%r1096,%r1097,%r1098,%r1099,%r1100,%r1101,%r1102,%r1103,%r1104,%r1105);
|
|
// end inline asm
|
|
ld.local.u32 %r77, [%rd2];
|
|
min.s32 %r78, %r77, 4;
|
|
setp.eq.s32 %p180, %r78, 0;
|
|
mov.pred %p374, %p375;
|
|
@%p180 bra $L__BB11_223;
|
|
|
|
not.b32 %r1107, %r77;
|
|
max.s32 %r1108, %r1107, -5;
|
|
mov.u32 %r1109, -2;
|
|
sub.s32 %r79, %r1109, %r1108;
|
|
mov.u16 %rs56, 0;
|
|
mov.u16 %rs64, 1;
|
|
|
|
$L__BB11_158:
|
|
sub.s32 %r81, %r79, %r1504;
|
|
setp.eq.s32 %p181, %r81, 0;
|
|
mov.u16 %rs98, %rs56;
|
|
@%p181 bra $L__BB11_179;
|
|
|
|
and.b32 %r82, %r81, 3;
|
|
add.s32 %r1111, %r81, -1;
|
|
setp.lt.u32 %p182, %r1111, 3;
|
|
mov.u32 %r1507, 0;
|
|
mov.u16 %rs98, %rs56;
|
|
@%p182 bra $L__BB11_170;
|
|
|
|
sub.s32 %r1506, %r81, %r82;
|
|
mov.u16 %rs98, 0;
|
|
mov.u32 %r1507, 0;
|
|
|
|
$L__BB11_161:
|
|
cvt.s64.s32 %rd28, %r1507;
|
|
mul.wide.s32 %rd265, %r1507, 24;
|
|
add.s64 %rd266, %rd2, %rd265;
|
|
add.s64 %rd29, %rd266, 4;
|
|
ld.local.f32 %f3818, [%rd266+28];
|
|
ld.local.f32 %f394, [%rd266+4];
|
|
setp.leu.ftz.f32 %p183, %f394, %f3818;
|
|
mul.wide.s32 %rd267, %r1507, 12;
|
|
add.s64 %rd268, %rd2, %rd267;
|
|
add.s64 %rd30, %rd268, 100;
|
|
@%p183 bra $L__BB11_163;
|
|
|
|
ld.local.f32 %f1912, [%rd29+4];
|
|
ld.local.f32 %f1913, [%rd29+8];
|
|
ld.local.u32 %r1113, [%rd29+12];
|
|
ld.local.u32 %r1114, [%rd29+16];
|
|
ld.local.u32 %r1115, [%rd29+20];
|
|
ld.local.f32 %f1914, [%rd30];
|
|
ld.local.f32 %f1915, [%rd30+4];
|
|
ld.local.f32 %f1916, [%rd30+8];
|
|
ld.local.f32 %f1917, [%rd29+28];
|
|
ld.local.f32 %f1918, [%rd29+32];
|
|
ld.local.u32 %r1116, [%rd29+36];
|
|
ld.local.u32 %r1117, [%rd29+40];
|
|
ld.local.u32 %r1118, [%rd29+44];
|
|
st.local.f32 [%rd29], %f3818;
|
|
st.local.f32 [%rd29+4], %f1917;
|
|
st.local.f32 [%rd29+8], %f1918;
|
|
st.local.u32 [%rd29+12], %r1116;
|
|
st.local.u32 [%rd29+16], %r1117;
|
|
st.local.u32 [%rd29+20], %r1118;
|
|
ld.local.f32 %f1919, [%rd30+12];
|
|
ld.local.f32 %f1920, [%rd30+16];
|
|
ld.local.f32 %f1921, [%rd30+20];
|
|
st.local.f32 [%rd30], %f1919;
|
|
st.local.f32 [%rd30+4], %f1920;
|
|
st.local.f32 [%rd30+8], %f1921;
|
|
st.local.f32 [%rd29+24], %f394;
|
|
st.local.f32 [%rd29+28], %f1912;
|
|
st.local.f32 [%rd29+32], %f1913;
|
|
st.local.u32 [%rd29+36], %r1113;
|
|
st.local.u32 [%rd29+40], %r1114;
|
|
st.local.u32 [%rd29+44], %r1115;
|
|
st.local.f32 [%rd30+12], %f1914;
|
|
st.local.f32 [%rd30+16], %f1915;
|
|
st.local.f32 [%rd30+20], %f1916;
|
|
ld.local.f32 %f3818, [%rd29+24];
|
|
mov.u16 %rs98, 1;
|
|
|
|
$L__BB11_163:
|
|
ld.local.f32 %f3819, [%rd29+48];
|
|
setp.leu.ftz.f32 %p184, %f3818, %f3819;
|
|
@%p184 bra $L__BB11_165;
|
|
|
|
ld.local.f32 %f1922, [%rd29+28];
|
|
ld.local.f32 %f1923, [%rd29+32];
|
|
ld.local.u32 %r1119, [%rd29+36];
|
|
ld.local.u32 %r1120, [%rd29+40];
|
|
ld.local.u32 %r1121, [%rd29+44];
|
|
ld.local.f32 %f1924, [%rd30+12];
|
|
ld.local.f32 %f1925, [%rd30+16];
|
|
ld.local.f32 %f1926, [%rd30+20];
|
|
ld.local.f32 %f1927, [%rd29+52];
|
|
ld.local.f32 %f1928, [%rd29+56];
|
|
ld.local.u32 %r1122, [%rd29+60];
|
|
ld.local.u32 %r1123, [%rd29+64];
|
|
ld.local.u32 %r1124, [%rd29+68];
|
|
st.local.f32 [%rd29+24], %f3819;
|
|
st.local.f32 [%rd29+28], %f1927;
|
|
st.local.f32 [%rd29+32], %f1928;
|
|
st.local.u32 [%rd29+36], %r1122;
|
|
st.local.u32 [%rd29+40], %r1123;
|
|
st.local.u32 [%rd29+44], %r1124;
|
|
ld.local.f32 %f1929, [%rd30+24];
|
|
ld.local.f32 %f1930, [%rd30+28];
|
|
ld.local.f32 %f1931, [%rd30+32];
|
|
st.local.f32 [%rd30+12], %f1929;
|
|
st.local.f32 [%rd30+16], %f1930;
|
|
st.local.f32 [%rd30+20], %f1931;
|
|
st.local.f32 [%rd29+48], %f3818;
|
|
st.local.f32 [%rd29+52], %f1922;
|
|
st.local.f32 [%rd29+56], %f1923;
|
|
st.local.u32 [%rd29+60], %r1119;
|
|
st.local.u32 [%rd29+64], %r1120;
|
|
st.local.u32 [%rd29+68], %r1121;
|
|
st.local.f32 [%rd30+24], %f1924;
|
|
st.local.f32 [%rd30+28], %f1925;
|
|
st.local.f32 [%rd30+32], %f1926;
|
|
ld.local.f32 %f3819, [%rd29+48];
|
|
mov.u16 %rs98, 1;
|
|
|
|
$L__BB11_165:
|
|
ld.local.f32 %f3820, [%rd29+72];
|
|
setp.leu.ftz.f32 %p185, %f3819, %f3820;
|
|
@%p185 bra $L__BB11_167;
|
|
|
|
ld.local.f32 %f1932, [%rd29+52];
|
|
ld.local.f32 %f1933, [%rd29+56];
|
|
ld.local.u32 %r1125, [%rd29+60];
|
|
ld.local.u32 %r1126, [%rd29+64];
|
|
ld.local.u32 %r1127, [%rd29+68];
|
|
ld.local.f32 %f1934, [%rd30+24];
|
|
ld.local.f32 %f1935, [%rd30+28];
|
|
ld.local.f32 %f1936, [%rd30+32];
|
|
ld.local.f32 %f1937, [%rd29+76];
|
|
ld.local.f32 %f1938, [%rd29+80];
|
|
ld.local.u32 %r1128, [%rd29+84];
|
|
ld.local.u32 %r1129, [%rd29+88];
|
|
ld.local.u32 %r1130, [%rd29+92];
|
|
st.local.f32 [%rd29+48], %f3820;
|
|
st.local.f32 [%rd29+52], %f1937;
|
|
st.local.f32 [%rd29+56], %f1938;
|
|
st.local.u32 [%rd29+60], %r1128;
|
|
st.local.u32 [%rd29+64], %r1129;
|
|
st.local.u32 [%rd29+68], %r1130;
|
|
ld.local.f32 %f1939, [%rd30+36];
|
|
ld.local.f32 %f1940, [%rd30+40];
|
|
ld.local.f32 %f1941, [%rd30+44];
|
|
st.local.f32 [%rd30+24], %f1939;
|
|
st.local.f32 [%rd30+28], %f1940;
|
|
st.local.f32 [%rd30+32], %f1941;
|
|
st.local.f32 [%rd29+72], %f3819;
|
|
st.local.f32 [%rd29+76], %f1932;
|
|
st.local.f32 [%rd29+80], %f1933;
|
|
st.local.u32 [%rd29+84], %r1125;
|
|
st.local.u32 [%rd29+88], %r1126;
|
|
st.local.u32 [%rd29+92], %r1127;
|
|
st.local.f32 [%rd30+36], %f1934;
|
|
st.local.f32 [%rd30+40], %f1935;
|
|
st.local.f32 [%rd30+44], %f1936;
|
|
ld.local.f32 %f3820, [%rd29+72];
|
|
mov.u16 %rs98, 1;
|
|
|
|
$L__BB11_167:
|
|
cvt.u32.u64 %r1131, %rd28;
|
|
add.s32 %r1507, %r1131, 4;
|
|
ld.local.f32 %f403, [%rd29+96];
|
|
setp.leu.ftz.f32 %p186, %f3820, %f403;
|
|
@%p186 bra $L__BB11_169;
|
|
|
|
ld.local.f32 %f1942, [%rd29+76];
|
|
ld.local.f32 %f1943, [%rd29+80];
|
|
ld.local.u32 %r1132, [%rd29+84];
|
|
ld.local.u32 %r1133, [%rd29+88];
|
|
ld.local.u32 %r1134, [%rd29+92];
|
|
ld.local.f32 %f1944, [%rd30+36];
|
|
ld.local.f32 %f1945, [%rd30+40];
|
|
ld.local.f32 %f1946, [%rd30+44];
|
|
ld.local.f32 %f1947, [%rd29+100];
|
|
ld.local.f32 %f1948, [%rd29+104];
|
|
ld.local.u32 %r1135, [%rd29+108];
|
|
ld.local.u32 %r1136, [%rd29+112];
|
|
ld.local.u32 %r1137, [%rd29+116];
|
|
st.local.f32 [%rd29+72], %f403;
|
|
st.local.f32 [%rd29+76], %f1947;
|
|
st.local.f32 [%rd29+80], %f1948;
|
|
st.local.u32 [%rd29+84], %r1135;
|
|
st.local.u32 [%rd29+88], %r1136;
|
|
st.local.u32 [%rd29+92], %r1137;
|
|
ld.local.f32 %f1949, [%rd30+48];
|
|
ld.local.f32 %f1950, [%rd30+52];
|
|
ld.local.f32 %f1951, [%rd30+56];
|
|
st.local.f32 [%rd30+36], %f1949;
|
|
st.local.f32 [%rd30+40], %f1950;
|
|
st.local.f32 [%rd30+44], %f1951;
|
|
st.local.f32 [%rd29+96], %f3820;
|
|
st.local.f32 [%rd29+100], %f1942;
|
|
st.local.f32 [%rd29+104], %f1943;
|
|
st.local.u32 [%rd29+108], %r1132;
|
|
st.local.u32 [%rd29+112], %r1133;
|
|
st.local.u32 [%rd29+116], %r1134;
|
|
st.local.f32 [%rd30+48], %f1944;
|
|
st.local.f32 [%rd30+52], %f1945;
|
|
st.local.f32 [%rd30+56], %f1946;
|
|
mov.u16 %rs98, 1;
|
|
|
|
$L__BB11_169:
|
|
add.s32 %r1506, %r1506, -4;
|
|
setp.ne.s32 %p187, %r1506, 0;
|
|
@%p187 bra $L__BB11_161;
|
|
|
|
$L__BB11_170:
|
|
setp.eq.s32 %p188, %r82, 0;
|
|
@%p188 bra $L__BB11_179;
|
|
|
|
mul.wide.s32 %rd269, %r1507, 24;
|
|
add.s64 %rd270, %rd2, %rd269;
|
|
add.s64 %rd31, %rd270, 4;
|
|
ld.local.f32 %f404, [%rd270+28];
|
|
ld.local.f32 %f405, [%rd270+4];
|
|
setp.leu.ftz.f32 %p189, %f405, %f404;
|
|
mul.wide.s32 %rd271, %r1507, 12;
|
|
add.s64 %rd272, %rd2, %rd271;
|
|
add.s64 %rd32, %rd272, 100;
|
|
@%p189 bra $L__BB11_173;
|
|
|
|
ld.local.f32 %f1952, [%rd31+4];
|
|
ld.local.f32 %f1953, [%rd31+8];
|
|
ld.local.u32 %r1138, [%rd31+12];
|
|
ld.local.u32 %r1139, [%rd31+16];
|
|
ld.local.u32 %r1140, [%rd31+20];
|
|
ld.local.f32 %f1954, [%rd32];
|
|
ld.local.f32 %f1955, [%rd32+4];
|
|
ld.local.f32 %f1956, [%rd32+8];
|
|
ld.local.f32 %f1957, [%rd31+28];
|
|
ld.local.f32 %f1958, [%rd31+32];
|
|
ld.local.u32 %r1141, [%rd31+36];
|
|
ld.local.u32 %r1142, [%rd31+40];
|
|
ld.local.u32 %r1143, [%rd31+44];
|
|
st.local.f32 [%rd31], %f404;
|
|
st.local.f32 [%rd31+4], %f1957;
|
|
st.local.f32 [%rd31+8], %f1958;
|
|
st.local.u32 [%rd31+12], %r1141;
|
|
st.local.u32 [%rd31+16], %r1142;
|
|
st.local.u32 [%rd31+20], %r1143;
|
|
ld.local.f32 %f1959, [%rd32+12];
|
|
ld.local.f32 %f1960, [%rd32+16];
|
|
ld.local.f32 %f1961, [%rd32+20];
|
|
st.local.f32 [%rd32], %f1959;
|
|
st.local.f32 [%rd32+4], %f1960;
|
|
st.local.f32 [%rd32+8], %f1961;
|
|
st.local.f32 [%rd31+24], %f405;
|
|
st.local.f32 [%rd31+28], %f1952;
|
|
st.local.f32 [%rd31+32], %f1953;
|
|
st.local.u32 [%rd31+36], %r1138;
|
|
st.local.u32 [%rd31+40], %r1139;
|
|
st.local.u32 [%rd31+44], %r1140;
|
|
st.local.f32 [%rd32+12], %f1954;
|
|
st.local.f32 [%rd32+16], %f1955;
|
|
st.local.f32 [%rd32+20], %f1956;
|
|
mov.u16 %rs98, %rs64;
|
|
|
|
$L__BB11_173:
|
|
setp.eq.s32 %p190, %r82, 1;
|
|
@%p190 bra $L__BB11_179;
|
|
|
|
ld.local.f32 %f406, [%rd31+48];
|
|
ld.local.f32 %f407, [%rd31+24];
|
|
setp.leu.ftz.f32 %p191, %f407, %f406;
|
|
@%p191 bra $L__BB11_176;
|
|
|
|
ld.local.f32 %f1962, [%rd31+28];
|
|
ld.local.f32 %f1963, [%rd31+32];
|
|
ld.local.u32 %r1144, [%rd31+36];
|
|
ld.local.u32 %r1145, [%rd31+40];
|
|
ld.local.u32 %r1146, [%rd31+44];
|
|
ld.local.f32 %f1964, [%rd32+12];
|
|
ld.local.f32 %f1965, [%rd32+16];
|
|
ld.local.f32 %f1966, [%rd32+20];
|
|
ld.local.f32 %f1967, [%rd31+52];
|
|
ld.local.f32 %f1968, [%rd31+56];
|
|
ld.local.u32 %r1147, [%rd31+60];
|
|
ld.local.u32 %r1148, [%rd31+64];
|
|
ld.local.u32 %r1149, [%rd31+68];
|
|
st.local.f32 [%rd31+24], %f406;
|
|
st.local.f32 [%rd31+28], %f1967;
|
|
st.local.f32 [%rd31+32], %f1968;
|
|
st.local.u32 [%rd31+36], %r1147;
|
|
st.local.u32 [%rd31+40], %r1148;
|
|
st.local.u32 [%rd31+44], %r1149;
|
|
ld.local.f32 %f1969, [%rd32+24];
|
|
ld.local.f32 %f1970, [%rd32+28];
|
|
ld.local.f32 %f1971, [%rd32+32];
|
|
st.local.f32 [%rd32+12], %f1969;
|
|
st.local.f32 [%rd32+16], %f1970;
|
|
st.local.f32 [%rd32+20], %f1971;
|
|
st.local.f32 [%rd31+48], %f407;
|
|
st.local.f32 [%rd31+52], %f1962;
|
|
st.local.f32 [%rd31+56], %f1963;
|
|
st.local.u32 [%rd31+60], %r1144;
|
|
st.local.u32 [%rd31+64], %r1145;
|
|
st.local.u32 [%rd31+68], %r1146;
|
|
st.local.f32 [%rd32+24], %f1964;
|
|
st.local.f32 [%rd32+28], %f1965;
|
|
st.local.f32 [%rd32+32], %f1966;
|
|
mov.u16 %rs98, %rs64;
|
|
|
|
$L__BB11_176:
|
|
setp.eq.s32 %p192, %r82, 2;
|
|
@%p192 bra $L__BB11_179;
|
|
|
|
ld.local.f32 %f408, [%rd31+72];
|
|
ld.local.f32 %f409, [%rd31+48];
|
|
setp.leu.ftz.f32 %p193, %f409, %f408;
|
|
@%p193 bra $L__BB11_179;
|
|
|
|
ld.local.f32 %f1972, [%rd31+52];
|
|
ld.local.f32 %f1973, [%rd31+56];
|
|
ld.local.u32 %r1150, [%rd31+60];
|
|
ld.local.u32 %r1151, [%rd31+64];
|
|
ld.local.u32 %r1152, [%rd31+68];
|
|
ld.local.f32 %f1974, [%rd32+24];
|
|
ld.local.f32 %f1975, [%rd32+28];
|
|
ld.local.f32 %f1976, [%rd32+32];
|
|
ld.local.f32 %f1977, [%rd31+76];
|
|
ld.local.f32 %f1978, [%rd31+80];
|
|
ld.local.u32 %r1153, [%rd31+84];
|
|
ld.local.u32 %r1154, [%rd31+88];
|
|
ld.local.u32 %r1155, [%rd31+92];
|
|
st.local.f32 [%rd31+48], %f408;
|
|
st.local.f32 [%rd31+52], %f1977;
|
|
st.local.f32 [%rd31+56], %f1978;
|
|
st.local.u32 [%rd31+60], %r1153;
|
|
st.local.u32 [%rd31+64], %r1154;
|
|
st.local.u32 [%rd31+68], %r1155;
|
|
ld.local.f32 %f1979, [%rd32+36];
|
|
ld.local.f32 %f1980, [%rd32+40];
|
|
ld.local.f32 %f1981, [%rd32+44];
|
|
st.local.f32 [%rd32+24], %f1979;
|
|
st.local.f32 [%rd32+28], %f1980;
|
|
st.local.f32 [%rd32+32], %f1981;
|
|
st.local.f32 [%rd31+72], %f409;
|
|
st.local.f32 [%rd31+76], %f1972;
|
|
st.local.f32 [%rd31+80], %f1973;
|
|
st.local.u32 [%rd31+84], %r1150;
|
|
st.local.u32 [%rd31+88], %r1151;
|
|
st.local.u32 [%rd31+92], %r1152;
|
|
st.local.f32 [%rd32+36], %f1974;
|
|
st.local.f32 [%rd32+40], %f1975;
|
|
st.local.f32 [%rd32+44], %f1976;
|
|
mov.u16 %rs98, 1;
|
|
|
|
$L__BB11_179:
|
|
add.s32 %r1504, %r1504, 1;
|
|
and.b16 %rs67, %rs98, 255;
|
|
setp.ne.s16 %p194, %rs67, 0;
|
|
@%p194 bra $L__BB11_158;
|
|
|
|
setp.lt.s32 %p195, %r77, 1;
|
|
@%p195 bra $L__BB11_209;
|
|
|
|
ld.const.v2.u64 {%rd273, %rd274}, [kernel_params+2192];
|
|
mov.u32 %r1508, 0;
|
|
cvta.to.global.u64 %rd33, %rd274;
|
|
and.b32 %r90, %r9, 8388608;
|
|
ld.const.u64 %rd277, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd34, %rd277;
|
|
ld.local.f32 %f410, [%rd6+12];
|
|
ld.local.f32 %f411, [%rd6+16];
|
|
ld.local.f32 %f412, [%rd6+20];
|
|
ld.local.f32 %f413, [%rd6];
|
|
ld.local.f32 %f414, [%rd6+4];
|
|
ld.local.f32 %f415, [%rd6+8];
|
|
cvta.to.global.u64 %rd35, %rd273;
|
|
ld.local.u32 %r91, [%rd2];
|
|
cvt.rn.f32.s32 %f416, %r91;
|
|
mul.ftz.f32 %f417, %f312, 0f40400000;
|
|
mul.ftz.f32 %f418, %f312, 0f40800000;
|
|
mul.ftz.f32 %f419, %f313, 0f40400000;
|
|
mul.ftz.f32 %f420, %f313, 0f40800000;
|
|
mul.ftz.f32 %f421, %f314, 0f40400000;
|
|
mul.ftz.f32 %f422, %f314, 0f40800000;
|
|
neg.ftz.f32 %f423, %f385;
|
|
mul.ftz.f32 %f424, %f312, 0f41800000;
|
|
setp.gtu.ftz.f32 %p196, %f424, %f385;
|
|
setp.neu.ftz.f32 %p197, %f385, 0f00000000;
|
|
mov.f32 %f3855, 0f00000000;
|
|
and.pred %p8, %p197, %p196;
|
|
mul.ftz.f32 %f425, %f313, 0f41800000;
|
|
setp.gtu.ftz.f32 %p198, %f425, %f385;
|
|
and.pred %p9, %p197, %p198;
|
|
mul.ftz.f32 %f426, %f314, 0f41800000;
|
|
setp.gtu.ftz.f32 %p199, %f426, %f385;
|
|
and.pred %p10, %p197, %p199;
|
|
not.pred %p219, %p8;
|
|
not.pred %p220, %p9;
|
|
not.pred %p221, %p10;
|
|
mov.u64 %rd699, %rd2;
|
|
mov.u64 %rd700, %rd1;
|
|
mov.u64 %rd701, %rd2;
|
|
|
|
$L__BB11_182:
|
|
ld.local.u32 %r1157, [%rd701+20];
|
|
cvt.s64.s32 %rd39, %r1157;
|
|
mul.wide.s32 %rd278, %r1157, 4;
|
|
add.s64 %rd279, %rd33, %rd278;
|
|
ld.local.f32 %f1984, [%rd699+100];
|
|
neg.ftz.f32 %f1985, %f1984;
|
|
ld.local.f32 %f1986, [%rd699+104];
|
|
neg.ftz.f32 %f1987, %f1986;
|
|
ld.local.f32 %f1988, [%rd699+108];
|
|
neg.ftz.f32 %f1989, %f1988;
|
|
setp.eq.s32 %p200, %r90, 0;
|
|
selp.f32 %f1990, %f1984, %f1985, %p200;
|
|
selp.f32 %f1991, %f1986, %f1987, %p200;
|
|
selp.f32 %f1992, %f1988, %f1989, %p200;
|
|
ld.global.u32 %r93, [%rd279];
|
|
and.b32 %r1158, %r93, 8;
|
|
setp.eq.s32 %p201, %r1158, 0;
|
|
neg.ftz.f32 %f1993, %f1990;
|
|
neg.ftz.f32 %f1994, %f1991;
|
|
neg.ftz.f32 %f1995, %f1992;
|
|
selp.f32 %f3842, %f1990, %f1993, %p201;
|
|
selp.f32 %f3843, %f1991, %f1994, %p201;
|
|
selp.f32 %f3844, %f1992, %f1995, %p201;
|
|
and.b32 %r1159, %r93, 4;
|
|
setp.ne.s32 %p202, %r1159, 0;
|
|
@%p202 bra $L__BB11_192;
|
|
|
|
and.b32 %r1160, %r93, 2;
|
|
setp.eq.s32 %p203, %r1160, 0;
|
|
mul.lo.s64 %rd280, %rd39, 208;
|
|
add.s64 %rd281, %rd34, %rd280;
|
|
add.s64 %rd40, %rd281, 48;
|
|
@%p203 bra $L__BB11_190;
|
|
bra.uni $L__BB11_184;
|
|
|
|
$L__BB11_190:
|
|
ld.global.v4.f32 {%f3833, %f3834, %f3835, %f2206}, [%rd40];
|
|
ld.global.v4.f32 {%f3836, %f3837, %f3838, %f2210}, [%rd40+16];
|
|
ld.global.v4.f32 {%f3839, %f3840, %f3841, %f2214}, [%rd40+32];
|
|
bra.uni $L__BB11_191;
|
|
|
|
$L__BB11_184:
|
|
ld.global.u32 %rd282, [%rd40+120];
|
|
ld.global.u32 %r1161, [%rd40+104];
|
|
shl.b32 %r1162, %r1161, 1;
|
|
cvt.rn.f32.s32 %f1996, %r1162;
|
|
mul.ftz.f32 %f1997, %f311, %f1996;
|
|
cvt.rzi.ftz.s32.f32 %r1163, %f1997;
|
|
add.s32 %r1164, %r1162, -1;
|
|
min.s32 %r1165, %r1163, %r1164;
|
|
cvt.rn.f32.s32 %f1998, %r1165;
|
|
sub.ftz.f32 %f431, %f1997, %f1998;
|
|
cvt.s64.s32 %rd283, %r1165;
|
|
add.s64 %rd284, %rd283, %rd282;
|
|
shl.b64 %rd285, %rd284, 6;
|
|
add.s64 %rd41, %rd35, %rd285;
|
|
ld.global.v4.f32 {%f1999, %f2000, %f2001, %f2002}, [%rd41];
|
|
ld.global.v4.f32 {%f2003, %f2004, %f2005, %f2006}, [%rd41+64];
|
|
mul.ftz.f32 %f2007, %f2000, %f2004;
|
|
fma.rn.ftz.f32 %f2008, %f1999, %f2003, %f2007;
|
|
mul.ftz.f32 %f2009, %f2002, %f2006;
|
|
fma.rn.ftz.f32 %f2010, %f2001, %f2005, %f2009;
|
|
add.ftz.f32 %f440, %f2008, %f2010;
|
|
setp.gt.ftz.f32 %p204, %f440, 0f3F7FDF3B;
|
|
sub.ftz.f32 %f441, %f1745, %f431;
|
|
@%p204 bra $L__BB11_186;
|
|
bra.uni $L__BB11_185;
|
|
|
|
$L__BB11_186:
|
|
mul.ftz.f32 %f2065, %f1999, %f441;
|
|
fma.rn.ftz.f32 %f2066, %f2003, %f431, %f2065;
|
|
mul.ftz.f32 %f2067, %f2000, %f441;
|
|
fma.rn.ftz.f32 %f2068, %f2004, %f431, %f2067;
|
|
mul.ftz.f32 %f2069, %f2001, %f441;
|
|
fma.rn.ftz.f32 %f2070, %f2005, %f431, %f2069;
|
|
mul.ftz.f32 %f2071, %f2002, %f441;
|
|
fma.rn.ftz.f32 %f2072, %f2006, %f431, %f2071;
|
|
mul.ftz.f32 %f2073, %f2068, %f2068;
|
|
fma.rn.ftz.f32 %f2074, %f2066, %f2066, %f2073;
|
|
mul.ftz.f32 %f2075, %f2072, %f2072;
|
|
fma.rn.ftz.f32 %f2076, %f2070, %f2070, %f2075;
|
|
add.ftz.f32 %f2077, %f2074, %f2076;
|
|
rsqrt.approx.ftz.f32 %f2078, %f2077;
|
|
mul.ftz.f32 %f3825, %f2072, %f2078;
|
|
mul.ftz.f32 %f3824, %f2070, %f2078;
|
|
mul.ftz.f32 %f3823, %f2068, %f2078;
|
|
mul.ftz.f32 %f3822, %f2066, %f2078;
|
|
bra.uni $L__BB11_187;
|
|
|
|
$L__BB11_185:
|
|
mov.f32 %f2012, 0fBF800000;
|
|
max.ftz.f32 %f2013, %f440, %f2012;
|
|
min.ftz.f32 %f2015, %f2013, %f1745;
|
|
abs.ftz.f32 %f2016, %f2015;
|
|
sub.ftz.f32 %f2017, %f1745, %f2016;
|
|
mul.ftz.f32 %f2018, %f2017, 0f3F000000;
|
|
sqrt.approx.ftz.f32 %f2019, %f2018;
|
|
setp.gt.ftz.f32 %p205, %f2016, 0f3F11EB85;
|
|
selp.f32 %f2020, %f2019, %f2016, %p205;
|
|
mul.ftz.f32 %f2021, %f2020, %f2020;
|
|
mov.f32 %f2022, 0f3C94D2E9;
|
|
mov.f32 %f2023, 0f3D53F941;
|
|
fma.rn.ftz.f32 %f2024, %f2023, %f2021, %f2022;
|
|
mov.f32 %f2025, 0f3D3F841F;
|
|
fma.rn.ftz.f32 %f2026, %f2024, %f2021, %f2025;
|
|
mov.f32 %f2027, 0f3D994929;
|
|
fma.rn.ftz.f32 %f2028, %f2026, %f2021, %f2027;
|
|
mov.f32 %f2029, 0f3E2AAB94;
|
|
fma.rn.ftz.f32 %f2030, %f2028, %f2021, %f2029;
|
|
mul.ftz.f32 %f2031, %f2021, %f2030;
|
|
fma.rn.ftz.f32 %f2032, %f2031, %f2020, %f2020;
|
|
add.ftz.f32 %f2033, %f2032, %f2032;
|
|
mov.f32 %f2034, 0f3FC90FDB;
|
|
sub.ftz.f32 %f2035, %f2034, %f2032;
|
|
selp.f32 %f2036, %f2033, %f2035, %p205;
|
|
setp.lt.ftz.f32 %p206, %f2015, 0f00000000;
|
|
mov.f32 %f2037, 0f40490FDB;
|
|
sub.ftz.f32 %f2038, %f2037, %f2036;
|
|
selp.f32 %f2039, %f2038, %f2036, %p206;
|
|
mul.ftz.f32 %f2040, %f1999, %f440;
|
|
sub.ftz.f32 %f2041, %f2003, %f2040;
|
|
mul.ftz.f32 %f2042, %f2000, %f440;
|
|
sub.ftz.f32 %f2043, %f2004, %f2042;
|
|
mul.ftz.f32 %f2044, %f2001, %f440;
|
|
sub.ftz.f32 %f2045, %f2005, %f2044;
|
|
mul.ftz.f32 %f2046, %f2002, %f440;
|
|
sub.ftz.f32 %f2047, %f2006, %f2046;
|
|
mul.ftz.f32 %f2048, %f2043, %f2043;
|
|
fma.rn.ftz.f32 %f2049, %f2041, %f2041, %f2048;
|
|
mul.ftz.f32 %f2050, %f2047, %f2047;
|
|
fma.rn.ftz.f32 %f2051, %f2045, %f2045, %f2050;
|
|
add.ftz.f32 %f2052, %f2049, %f2051;
|
|
rsqrt.approx.ftz.f32 %f2053, %f2052;
|
|
mul.ftz.f32 %f2054, %f2041, %f2053;
|
|
mul.ftz.f32 %f2055, %f2043, %f2053;
|
|
mul.ftz.f32 %f2056, %f2045, %f2053;
|
|
mul.ftz.f32 %f2057, %f2047, %f2053;
|
|
mul.ftz.f32 %f2058, %f431, %f2039;
|
|
cos.approx.ftz.f32 %f2059, %f2058;
|
|
sin.approx.ftz.f32 %f2060, %f2058;
|
|
mul.ftz.f32 %f2061, %f2054, %f2060;
|
|
mul.ftz.f32 %f2062, %f2055, %f2060;
|
|
mul.ftz.f32 %f2063, %f2056, %f2060;
|
|
mul.ftz.f32 %f2064, %f2057, %f2060;
|
|
fma.rn.ftz.f32 %f3825, %f2002, %f2059, %f2064;
|
|
fma.rn.ftz.f32 %f3824, %f2001, %f2059, %f2063;
|
|
fma.rn.ftz.f32 %f3823, %f2000, %f2059, %f2062;
|
|
fma.rn.ftz.f32 %f3822, %f1999, %f2059, %f2061;
|
|
|
|
$L__BB11_187:
|
|
ld.global.f32 %f2079, [%rd41+28];
|
|
ld.global.f32 %f2080, [%rd41+92];
|
|
mul.ftz.f32 %f2081, %f431, %f2080;
|
|
fma.rn.ftz.f32 %f2082, %f441, %f2079, %f2081;
|
|
ld.global.v4.f32 {%f2083, %f2084, %f2085, %f2086}, [%rd41+32];
|
|
ld.global.v4.f32 {%f2091, %f2092, %f2093, %f2094}, [%rd41+96];
|
|
mul.ftz.f32 %f2099, %f431, %f2091;
|
|
mul.ftz.f32 %f2100, %f431, %f2092;
|
|
mul.ftz.f32 %f2101, %f431, %f2093;
|
|
mul.ftz.f32 %f2102, %f431, %f2094;
|
|
fma.rn.ftz.f32 %f2103, %f441, %f2083, %f2099;
|
|
fma.rn.ftz.f32 %f2104, %f441, %f2084, %f2100;
|
|
fma.rn.ftz.f32 %f2105, %f441, %f2085, %f2101;
|
|
fma.rn.ftz.f32 %f2106, %f441, %f2086, %f2102;
|
|
ld.global.v4.f32 {%f2107, %f2108, %f2109, %f2110}, [%rd41+48];
|
|
ld.global.v4.f32 {%f2115, %f2116, %f2117, %f2118}, [%rd41+112];
|
|
mul.ftz.f32 %f2123, %f431, %f2115;
|
|
mul.ftz.f32 %f2124, %f431, %f2116;
|
|
mul.ftz.f32 %f2125, %f431, %f2117;
|
|
mul.ftz.f32 %f2126, %f431, %f2118;
|
|
fma.rn.ftz.f32 %f2127, %f441, %f2107, %f2123;
|
|
fma.rn.ftz.f32 %f2128, %f441, %f2108, %f2124;
|
|
fma.rn.ftz.f32 %f2129, %f441, %f2109, %f2125;
|
|
fma.rn.ftz.f32 %f2130, %f441, %f2110, %f2126;
|
|
mul.ftz.f32 %f2131, %f3825, 0f3FB504F3;
|
|
mul.ftz.f32 %f2132, %f3822, 0f3FB504F3;
|
|
mul.ftz.f32 %f2133, %f3823, 0f3FB504F3;
|
|
mul.ftz.f32 %f2134, %f3824, 0f3FB504F3;
|
|
mul.ftz.f32 %f2135, %f2131, %f2132;
|
|
mul.ftz.f32 %f2136, %f2131, %f2133;
|
|
mul.ftz.f32 %f2137, %f2131, %f2134;
|
|
mul.ftz.f32 %f2138, %f2132, %f2132;
|
|
mul.ftz.f32 %f2139, %f2132, %f2133;
|
|
mul.ftz.f32 %f2140, %f2132, %f2134;
|
|
mul.ftz.f32 %f2141, %f2133, %f2133;
|
|
mul.ftz.f32 %f2142, %f2133, %f2134;
|
|
mul.ftz.f32 %f2143, %f2134, %f2134;
|
|
sub.ftz.f32 %f2145, %f1745, %f2141;
|
|
sub.ftz.f32 %f2146, %f2145, %f2143;
|
|
sub.ftz.f32 %f2147, %f2139, %f2137;
|
|
add.ftz.f32 %f2148, %f2136, %f2140;
|
|
add.ftz.f32 %f2149, %f2137, %f2139;
|
|
sub.ftz.f32 %f2150, %f1745, %f2138;
|
|
sub.ftz.f32 %f2151, %f2150, %f2143;
|
|
sub.ftz.f32 %f2152, %f2142, %f2135;
|
|
sub.ftz.f32 %f2153, %f2140, %f2136;
|
|
add.ftz.f32 %f2154, %f2135, %f2142;
|
|
sub.ftz.f32 %f2155, %f2150, %f2141;
|
|
mul.ftz.f32 %f2156, %f2147, %f2105;
|
|
fma.rn.ftz.f32 %f2157, %f2082, %f2146, %f2156;
|
|
fma.rn.ftz.f32 %f3831, %f2148, %f2128, %f2157;
|
|
mul.ftz.f32 %f2158, %f2147, %f2106;
|
|
fma.rn.ftz.f32 %f2159, %f2146, %f2103, %f2158;
|
|
fma.rn.ftz.f32 %f455, %f2148, %f2129, %f2159;
|
|
mul.ftz.f32 %f2160, %f2147, %f2127;
|
|
fma.rn.ftz.f32 %f2161, %f2146, %f2104, %f2160;
|
|
fma.rn.ftz.f32 %f456, %f2148, %f2130, %f2161;
|
|
mul.ftz.f32 %f2162, %f2151, %f2105;
|
|
fma.rn.ftz.f32 %f2163, %f2149, %f2082, %f2162;
|
|
fma.rn.ftz.f32 %f457, %f2152, %f2128, %f2163;
|
|
mul.ftz.f32 %f2164, %f2151, %f2106;
|
|
fma.rn.ftz.f32 %f2165, %f2149, %f2103, %f2164;
|
|
fma.rn.ftz.f32 %f3830, %f2152, %f2129, %f2165;
|
|
mul.ftz.f32 %f2166, %f2151, %f2127;
|
|
fma.rn.ftz.f32 %f2167, %f2149, %f2104, %f2166;
|
|
fma.rn.ftz.f32 %f459, %f2152, %f2130, %f2167;
|
|
mul.ftz.f32 %f2168, %f2154, %f2105;
|
|
fma.rn.ftz.f32 %f2169, %f2153, %f2082, %f2168;
|
|
fma.rn.ftz.f32 %f460, %f2155, %f2128, %f2169;
|
|
mul.ftz.f32 %f2170, %f2154, %f2106;
|
|
fma.rn.ftz.f32 %f2171, %f2153, %f2103, %f2170;
|
|
fma.rn.ftz.f32 %f461, %f2155, %f2129, %f2171;
|
|
mul.ftz.f32 %f2172, %f2154, %f2127;
|
|
fma.rn.ftz.f32 %f2173, %f2153, %f2104, %f2172;
|
|
fma.rn.ftz.f32 %f3829, %f2155, %f2130, %f2173;
|
|
mul.ftz.f32 %f2174, %f3830, %f3829;
|
|
mul.ftz.f32 %f463, %f461, %f459;
|
|
sub.ftz.f32 %f3828, %f2174, %f463;
|
|
mul.ftz.f32 %f465, %f461, %f456;
|
|
mul.ftz.f32 %f2175, %f455, %f3829;
|
|
sub.ftz.f32 %f3827, %f465, %f2175;
|
|
mul.ftz.f32 %f467, %f455, %f459;
|
|
mul.ftz.f32 %f2176, %f3830, %f456;
|
|
sub.ftz.f32 %f3826, %f467, %f2176;
|
|
mul.ftz.f32 %f2177, %f457, %f3827;
|
|
fma.rn.ftz.f32 %f2178, %f3831, %f3828, %f2177;
|
|
fma.rn.ftz.f32 %f3832, %f460, %f3826, %f2178;
|
|
setp.neu.ftz.f32 %p207, %f3832, 0f00000000;
|
|
@%p207 bra $L__BB11_189;
|
|
|
|
add.ftz.f32 %f3831, %f3831, 0f322BCC77;
|
|
add.ftz.f32 %f3829, %f3829, 0f322BCC77;
|
|
add.ftz.f32 %f3830, %f3830, 0f322BCC77;
|
|
mul.ftz.f32 %f2179, %f3830, %f3829;
|
|
sub.ftz.f32 %f3828, %f2179, %f463;
|
|
mul.ftz.f32 %f2180, %f455, %f3829;
|
|
sub.ftz.f32 %f3827, %f465, %f2180;
|
|
mul.ftz.f32 %f2181, %f3830, %f456;
|
|
sub.ftz.f32 %f3826, %f467, %f2181;
|
|
mul.ftz.f32 %f2182, %f457, %f3827;
|
|
fma.rn.ftz.f32 %f2183, %f3831, %f3828, %f2182;
|
|
fma.rn.ftz.f32 %f2184, %f460, %f3826, %f2183;
|
|
setp.eq.ftz.f32 %p208, %f2184, 0f00000000;
|
|
selp.f32 %f3832, 0f7F7FFFFF, %f2184, %p208;
|
|
|
|
$L__BB11_189:
|
|
div.approx.ftz.f32 %f3833, %f3828, %f3832;
|
|
div.approx.ftz.f32 %f3834, %f3827, %f3832;
|
|
div.approx.ftz.f32 %f3835, %f3826, %f3832;
|
|
mul.ftz.f32 %f2185, %f457, %f3829;
|
|
mul.ftz.f32 %f2186, %f460, %f459;
|
|
sub.ftz.f32 %f2187, %f2186, %f2185;
|
|
mul.ftz.f32 %f2188, %f460, %f456;
|
|
mul.ftz.f32 %f2189, %f3829, %f3831;
|
|
sub.ftz.f32 %f2190, %f2189, %f2188;
|
|
mul.ftz.f32 %f2191, %f459, %f3831;
|
|
mul.ftz.f32 %f2192, %f457, %f456;
|
|
sub.ftz.f32 %f2193, %f2192, %f2191;
|
|
div.approx.ftz.f32 %f3836, %f2187, %f3832;
|
|
div.approx.ftz.f32 %f3837, %f2190, %f3832;
|
|
div.approx.ftz.f32 %f3838, %f2193, %f3832;
|
|
mul.ftz.f32 %f2194, %f460, %f3830;
|
|
mul.ftz.f32 %f2195, %f457, %f461;
|
|
sub.ftz.f32 %f2196, %f2195, %f2194;
|
|
mul.ftz.f32 %f2197, %f461, %f3831;
|
|
mul.ftz.f32 %f2198, %f460, %f455;
|
|
sub.ftz.f32 %f2199, %f2198, %f2197;
|
|
mul.ftz.f32 %f2200, %f457, %f455;
|
|
mul.ftz.f32 %f2201, %f3830, %f3831;
|
|
sub.ftz.f32 %f2202, %f2201, %f2200;
|
|
div.approx.ftz.f32 %f3839, %f2196, %f3832;
|
|
div.approx.ftz.f32 %f3840, %f2199, %f3832;
|
|
div.approx.ftz.f32 %f3841, %f2202, %f3832;
|
|
|
|
$L__BB11_191:
|
|
mul.ftz.f32 %f2215, %f3843, %f3836;
|
|
fma.rn.ftz.f32 %f2216, %f3842, %f3833, %f2215;
|
|
fma.rn.ftz.f32 %f2217, %f3844, %f3839, %f2216;
|
|
mul.ftz.f32 %f2218, %f3843, %f3837;
|
|
fma.rn.ftz.f32 %f2219, %f3842, %f3834, %f2218;
|
|
fma.rn.ftz.f32 %f2220, %f3844, %f3840, %f2219;
|
|
mul.ftz.f32 %f2221, %f3843, %f3838;
|
|
fma.rn.ftz.f32 %f2222, %f3842, %f3835, %f2221;
|
|
fma.rn.ftz.f32 %f2223, %f3844, %f3841, %f2222;
|
|
mul.ftz.f32 %f2224, %f2220, %f2220;
|
|
fma.rn.ftz.f32 %f2225, %f2217, %f2217, %f2224;
|
|
fma.rn.ftz.f32 %f2226, %f2223, %f2223, %f2225;
|
|
sqrt.approx.ftz.f32 %f2227, %f2226;
|
|
div.approx.ftz.f32 %f3842, %f2217, %f2227;
|
|
div.approx.ftz.f32 %f3843, %f2220, %f2227;
|
|
div.approx.ftz.f32 %f3844, %f2223, %f2227;
|
|
|
|
$L__BB11_192:
|
|
ld.local.f32 %f2228, [%rd701+4];
|
|
fma.rn.ftz.f32 %f517, %f2228, %f410, %f413;
|
|
fma.rn.ftz.f32 %f518, %f2228, %f411, %f414;
|
|
fma.rn.ftz.f32 %f519, %f2228, %f412, %f415;
|
|
mul.ftz.f32 %f2229, %f3804, %f3843;
|
|
fma.rn.ftz.f32 %f2230, %f3803, %f3842, %f2229;
|
|
fma.rn.ftz.f32 %f2231, %f3805, %f3844, %f2230;
|
|
abs.ftz.f32 %f2232, %f2231;
|
|
mul.ftz.f32 %f2233, %f3807, %f2232;
|
|
mul.ftz.f32 %f2234, %f3801, %f3843;
|
|
fma.rn.ftz.f32 %f2235, %f3800, %f3842, %f2234;
|
|
fma.rn.ftz.f32 %f2236, %f3802, %f3844, %f2235;
|
|
abs.ftz.f32 %f2237, %f2236;
|
|
mul.ftz.f32 %f2238, %f3808, %f2237;
|
|
mul.ftz.f32 %f2239, %f3798, %f3843;
|
|
fma.rn.ftz.f32 %f2240, %f3797, %f3842, %f2239;
|
|
fma.rn.ftz.f32 %f2241, %f3799, %f3844, %f2240;
|
|
abs.ftz.f32 %f2242, %f2241;
|
|
mul.ftz.f32 %f2243, %f3809, %f2242;
|
|
mul.ftz.f32 %f2244, %f2238, %f2238;
|
|
fma.rn.ftz.f32 %f2245, %f2233, %f2233, %f2244;
|
|
fma.rn.ftz.f32 %f2246, %f2243, %f2243, %f2245;
|
|
div.approx.ftz.f32 %f3845, %f2233, %f2246;
|
|
setp.lt.s32 %p209, %r91, 5;
|
|
@%p209 bra $L__BB11_194;
|
|
|
|
mov.f32 %f2247, 0f40800000;
|
|
div.approx.ftz.f32 %f2248, %f416, %f2247;
|
|
mul.ftz.f32 %f3845, %f3845, %f2248;
|
|
|
|
$L__BB11_194:
|
|
sub.ftz.f32 %f2250, %f517, %f307;
|
|
sub.ftz.f32 %f2251, %f518, %f308;
|
|
mul.ftz.f32 %f2252, %f2251, %f2251;
|
|
fma.rn.ftz.f32 %f2253, %f2250, %f2250, %f2252;
|
|
sub.ftz.f32 %f2254, %f519, %f309;
|
|
fma.rn.ftz.f32 %f2255, %f2254, %f2254, %f2253;
|
|
sqrt.approx.ftz.f32 %f523, %f2255;
|
|
setp.eq.ftz.f32 %p210, %f523, 0f00000000;
|
|
mov.f32 %f3852, 0f00000000;
|
|
neg.ftz.f32 %f524, %f523;
|
|
setp.le.ftz.f32 %p211, %f424, %f523;
|
|
or.pred %p212, %p210, %p211;
|
|
mov.f32 %f3846, %f3852;
|
|
@%p212 bra $L__BB11_196;
|
|
|
|
div.approx.ftz.f32 %f2256, %f524, %f417;
|
|
mul.ftz.f32 %f2257, %f2256, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f2258, %f2257;
|
|
mul.ftz.f32 %f2259, %f2258, %f2258;
|
|
fma.rn.ftz.f32 %f2260, %f2258, %f2259, %f2258;
|
|
div.approx.ftz.f32 %f2261, %f2260, %f418;
|
|
mul.ftz.f32 %f3846, %f2261, 0f3F807715;
|
|
|
|
$L__BB11_196:
|
|
setp.le.ftz.f32 %p213, %f425, %f523;
|
|
or.pred %p215, %p210, %p213;
|
|
mov.f32 %f3847, %f3852;
|
|
@%p215 bra $L__BB11_198;
|
|
|
|
div.approx.ftz.f32 %f2263, %f524, %f419;
|
|
mul.ftz.f32 %f2264, %f2263, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f2265, %f2264;
|
|
mul.ftz.f32 %f2266, %f2265, %f2265;
|
|
fma.rn.ftz.f32 %f2267, %f2265, %f2266, %f2265;
|
|
div.approx.ftz.f32 %f2268, %f2267, %f420;
|
|
mul.ftz.f32 %f3847, %f2268, 0f3F807715;
|
|
|
|
$L__BB11_198:
|
|
setp.le.ftz.f32 %p216, %f426, %f523;
|
|
or.pred %p218, %p210, %p216;
|
|
mov.f32 %f3848, %f3852;
|
|
@%p218 bra $L__BB11_200;
|
|
|
|
div.approx.ftz.f32 %f2270, %f524, %f421;
|
|
mul.ftz.f32 %f2271, %f2270, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f2272, %f2271;
|
|
mul.ftz.f32 %f2273, %f2272, %f2272;
|
|
fma.rn.ftz.f32 %f2274, %f2272, %f2273, %f2272;
|
|
div.approx.ftz.f32 %f2275, %f2274, %f422;
|
|
mul.ftz.f32 %f3848, %f2275, 0f3F807715;
|
|
|
|
$L__BB11_200:
|
|
mov.f32 %f3849, %f3852;
|
|
@%p219 bra $L__BB11_202;
|
|
|
|
div.approx.ftz.f32 %f2277, %f423, %f417;
|
|
mul.ftz.f32 %f2278, %f2277, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f2279, %f2278;
|
|
mul.ftz.f32 %f2280, %f2279, %f2279;
|
|
fma.rn.ftz.f32 %f2281, %f2279, %f2280, %f2279;
|
|
div.approx.ftz.f32 %f2282, %f2281, %f418;
|
|
mul.ftz.f32 %f3849, %f2282, 0f3F807715;
|
|
|
|
$L__BB11_202:
|
|
mov.f32 %f3850, %f3852;
|
|
@%p220 bra $L__BB11_204;
|
|
|
|
div.approx.ftz.f32 %f2284, %f423, %f419;
|
|
mul.ftz.f32 %f2285, %f2284, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f2286, %f2285;
|
|
mul.ftz.f32 %f2287, %f2286, %f2286;
|
|
fma.rn.ftz.f32 %f2288, %f2286, %f2287, %f2286;
|
|
div.approx.ftz.f32 %f2289, %f2288, %f420;
|
|
mul.ftz.f32 %f3850, %f2289, 0f3F807715;
|
|
|
|
$L__BB11_204:
|
|
mov.f32 %f3851, %f3852;
|
|
@%p221 bra $L__BB11_206;
|
|
|
|
div.approx.ftz.f32 %f2291, %f423, %f421;
|
|
mul.ftz.f32 %f2292, %f2291, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f2293, %f2292;
|
|
mul.ftz.f32 %f2294, %f2293, %f2293;
|
|
fma.rn.ftz.f32 %f2295, %f2293, %f2294, %f2293;
|
|
div.approx.ftz.f32 %f2296, %f2295, %f422;
|
|
mul.ftz.f32 %f3851, %f2296, 0f3F807715;
|
|
|
|
$L__BB11_206:
|
|
add.ftz.f32 %f2300, %f3849, %f3850;
|
|
add.ftz.f32 %f2301, %f2300, %f3851;
|
|
div.approx.ftz.f32 %f537, %f2301, %f341;
|
|
setp.leu.ftz.f32 %p222, %f537, 0f00000000;
|
|
mov.f32 %f3853, %f3852;
|
|
mov.f32 %f3854, %f3852;
|
|
@%p222 bra $L__BB11_208;
|
|
|
|
div.approx.ftz.f32 %f3854, %f3848, %f537;
|
|
div.approx.ftz.f32 %f3853, %f3847, %f537;
|
|
div.approx.ftz.f32 %f3852, %f3846, %f537;
|
|
|
|
$L__BB11_208:
|
|
mul.ftz.f32 %f2302, %f3845, %f3852;
|
|
mul.ftz.f32 %f2303, %f3845, %f3853;
|
|
mul.ftz.f32 %f2304, %f3845, %f3854;
|
|
st.local.f32 [%rd699+100], %f3842;
|
|
st.local.f32 [%rd699+104], %f3843;
|
|
st.local.f32 [%rd699+108], %f3844;
|
|
st.local.f32 [%rd700], %f2302;
|
|
st.local.f32 [%rd700+4], %f2303;
|
|
st.local.f32 [%rd700+8], %f2304;
|
|
abs.ftz.f32 %f2305, %f2302;
|
|
abs.ftz.f32 %f2306, %f2303;
|
|
abs.ftz.f32 %f2307, %f2304;
|
|
add.ftz.f32 %f2308, %f2305, %f2306;
|
|
add.ftz.f32 %f2309, %f2307, %f2308;
|
|
fma.rn.ftz.f32 %f3855, %f2309, 0f3EAAAAAB, %f3855;
|
|
add.s64 %rd701, %rd701, 24;
|
|
add.s64 %rd700, %rd700, 12;
|
|
add.s64 %rd699, %rd699, 12;
|
|
add.s32 %r1508, %r1508, 1;
|
|
setp.lt.s32 %p223, %r1508, %r78;
|
|
@%p223 bra $L__BB11_182;
|
|
|
|
$L__BB11_209:
|
|
setp.eq.ftz.f32 %p225, %f3855, 0f00000000;
|
|
mov.pred %p374, %p375;
|
|
@%p225 bra $L__BB11_223;
|
|
bra.uni $L__BB11_210;
|
|
|
|
$L__BB11_223:
|
|
@%p374 bra $L__BB11_369;
|
|
|
|
$L__BB11_224:
|
|
ld.const.u32 %r1263, [kernel_params+1624];
|
|
setp.eq.s32 %p244, %r1263, 0;
|
|
@%p244 bra $L__BB11_358;
|
|
|
|
ld.const.u64 %rd304, [kernel_params+2200];
|
|
cvta.to.global.u64 %rd52, %rd304;
|
|
ld.local.u32 %r1264, [%rd2+20];
|
|
mul.wide.s32 %rd305, %r1264, 4;
|
|
add.s64 %rd306, %rd52, %rd305;
|
|
ld.global.u8 %rs70, [%rd306];
|
|
and.b16 %rs71, %rs70, 32;
|
|
setp.eq.s16 %p245, %rs71, 0;
|
|
@%p245 bra $L__BB11_358;
|
|
|
|
ld.const.u64 %rd695, [kernel_params+2560];
|
|
cvta.to.global.u64 %rd694, %rd695;
|
|
add.s64 %rd693, %rd694, %rd168;
|
|
add.u64 %rd691, %SPL, 304;
|
|
cvta.to.global.u64 %rd307, %rd707;
|
|
mul.lo.s64 %rd308, %rd4, 12;
|
|
add.s64 %rd309, %rd307, %rd308;
|
|
ld.global.f32 %f3862, [%rd309];
|
|
ld.global.f32 %f2377, [%rd309+4];
|
|
ld.global.f32 %f2378, [%rd309+8];
|
|
ld.local.f32 %f2379, [%rd6];
|
|
ld.local.f32 %f2380, [%rd6+4];
|
|
ld.local.f32 %f2381, [%rd6+8];
|
|
add.s64 %rd53, %rd691, 32;
|
|
mov.u32 %r1509, 0;
|
|
st.local.u32 [%rd691+32], %r1509;
|
|
st.local.u32 [%rd691+52], %r1509;
|
|
st.local.u32 [%rd691+56], %r1509;
|
|
st.local.f32 [%rd691], %f3862;
|
|
st.local.f32 [%rd691+4], %f2377;
|
|
st.local.f32 [%rd691+8], %f2378;
|
|
sub.ftz.f32 %f2382, %f2379, %f3862;
|
|
sub.ftz.f32 %f2383, %f2380, %f2377;
|
|
sub.ftz.f32 %f2384, %f2381, %f2378;
|
|
mul.ftz.f32 %f2385, %f2383, %f2383;
|
|
fma.rn.ftz.f32 %f2386, %f2382, %f2382, %f2385;
|
|
fma.rn.ftz.f32 %f2387, %f2384, %f2384, %f2386;
|
|
sqrt.approx.ftz.f32 %f2388, %f2387;
|
|
st.local.f32 [%rd691+28], %f2388;
|
|
rsqrt.approx.ftz.f32 %f2389, %f2387;
|
|
mul.ftz.f32 %f2390, %f2382, %f2389;
|
|
mul.ftz.f32 %f2391, %f2383, %f2389;
|
|
mul.ftz.f32 %f2392, %f2384, %f2389;
|
|
st.local.f32 [%rd691+12], %f2390;
|
|
st.local.f32 [%rd691+16], %f2391;
|
|
st.local.f32 [%rd691+20], %f2392;
|
|
st.local.u32 [%rd691+24], %r1509;
|
|
cvta.to.global.u64 %rd310, %rd706;
|
|
add.s64 %rd312, %rd310, %rd168;
|
|
ld.global.u32 %r1266, [%rd312];
|
|
st.local.u32 [%rd691+40], %r1266;
|
|
ld.const.u64 %rd313, [kernel_params+2728];
|
|
cvta.to.global.u64 %rd314, %rd313;
|
|
add.s64 %rd315, %rd314, %rd168;
|
|
ld.global.u32 %r1267, [%rd315];
|
|
st.local.u32 [%rd691+36], %r1267;
|
|
mov.u32 %r1268, -1;
|
|
st.local.u32 [%rd691+48], %r1268;
|
|
st.local.u32 [%rd691+44], %r1268;
|
|
ld.global.u8 %rs72, [%rd693+3];
|
|
and.b16 %rs73, %rs72, 64;
|
|
setp.eq.s16 %p246, %rs73, 0;
|
|
selp.b32 %r97, 2047, 134152192, %p246;
|
|
ld.const.u32 %r98, [kernel_params+28];
|
|
shl.b32 %r99, %r98, 1;
|
|
setp.eq.s32 %p247, %r99, 0;
|
|
@%p247 bra $L__BB11_358;
|
|
|
|
add.u64 %rd696, %SP, 304;
|
|
cvt.u32.u64 %r100, %rd696;
|
|
shr.u64 %rd317, %rd696, 32;
|
|
cvt.u32.u64 %r101, %rd317;
|
|
and.b32 %r1270, %r97, 255;
|
|
setp.eq.s32 %p248, %r1270, 0;
|
|
selp.b32 %r102, 255, %r1270, %p248;
|
|
add.s64 %rd319, %rd145, 72;
|
|
add.s64 %rd320, %rd145, 88;
|
|
cvta.to.local.u64 %rd54, %rd320;
|
|
add.s64 %rd55, %rd145, 64;
|
|
add.s64 %rd321, %rd145, 76;
|
|
add.s64 %rd322, %rd145, 56;
|
|
cvta.to.local.u64 %rd56, %rd322;
|
|
add.s64 %rd323, %rd145, 60;
|
|
cvta.to.local.u64 %rd57, %rd323;
|
|
add.s64 %rd58, %rd145, 80;
|
|
add.s64 %rd324, %rd145, 52;
|
|
cvta.to.local.u64 %rd59, %rd324;
|
|
add.s64 %rd325, %rd145, 36;
|
|
cvta.to.local.u64 %rd60, %rd325;
|
|
add.s64 %rd61, %rd145, 40;
|
|
cvta.to.local.u64 %rd62, %rd61;
|
|
add.s64 %rd63, %rd145, 44;
|
|
cvta.to.local.u64 %rd64, %rd63;
|
|
add.s64 %rd65, %rd145, 144;
|
|
ld.const.u64 %rd326, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd66, %rd326;
|
|
ld.const.u64 %rd327, [kernel_params+2192];
|
|
cvta.to.global.u64 %rd67, %rd327;
|
|
add.s64 %rd68, %rd145, 24;
|
|
ld.const.u64 %rd328, [kernel_params+2424];
|
|
cvta.to.global.u64 %rd69, %rd328;
|
|
add.s64 %rd329, %rd145, 48;
|
|
cvta.to.local.u64 %rd70, %rd329;
|
|
add.s64 %rd330, %rd145, 32;
|
|
cvta.to.local.u64 %rd71, %rd330;
|
|
add.s64 %rd331, %rd145, 116;
|
|
add.s64 %rd72, %rd145, 120;
|
|
add.s64 %rd332, %rd145, 124;
|
|
add.s64 %rd73, %rd145, 128;
|
|
add.s64 %rd74, %rd145, 176;
|
|
add.s64 %rd75, %rd145, 160;
|
|
add.s64 %rd76, %rd145, 192;
|
|
ld.const.v2.u64 {%rd333, %rd334}, [kernel_params+2320];
|
|
cvta.to.global.u64 %rd77, %rd333;
|
|
or.b64 %rd337, %rd145, 8;
|
|
cvta.to.local.u64 %rd78, %rd337;
|
|
or.b64 %rd338, %rd145, 12;
|
|
cvta.to.local.u64 %rd79, %rd338;
|
|
add.s64 %rd80, %rd145, 16;
|
|
add.s64 %rd81, %rd145, 20;
|
|
cvta.to.local.u64 %rd82, %rd331;
|
|
cvta.to.local.u64 %rd83, %rd72;
|
|
cvta.to.local.u64 %rd84, %rd332;
|
|
add.s64 %rd339, %rd145, 136;
|
|
cvta.to.local.u64 %rd85, %rd339;
|
|
add.s64 %rd340, %rd145, 92;
|
|
cvta.to.local.u64 %rd86, %rd340;
|
|
add.s64 %rd341, %rd145, 96;
|
|
cvta.to.local.u64 %rd87, %rd341;
|
|
ld.const.v2.u64 {%rd342, %rd343}, [kernel_params+2272];
|
|
cvta.to.global.u64 %rd88, %rd343;
|
|
add.s64 %rd89, %rd145, 224;
|
|
add.s64 %rd90, %rd145, 208;
|
|
ld.const.v2.u64 {%rd346, %rd347}, [kernel_params+2240];
|
|
cvta.to.global.u64 %rd91, %rd347;
|
|
cvta.to.global.u64 %rd92, %rd342;
|
|
ld.const.u64 %rd350, [kernel_params+2312];
|
|
cvta.to.global.u64 %rd93, %rd350;
|
|
cvta.to.global.u64 %rd94, %rd334;
|
|
ld.const.u64 %rd351, [kernel_params+2288];
|
|
cvta.to.global.u64 %rd95, %rd351;
|
|
ld.const.u64 %rd352, [kernel_params+2232];
|
|
cvta.to.global.u64 %rd96, %rd352;
|
|
add.s64 %rd353, %rd145, 100;
|
|
cvta.to.local.u64 %rd97, %rd353;
|
|
add.s64 %rd98, %rd145, 104;
|
|
add.s64 %rd354, %rd145, 112;
|
|
cvta.to.local.u64 %rd99, %rd354;
|
|
cvta.to.local.u64 %rd100, %rd80;
|
|
cvta.to.local.u64 %rd101, %rd81;
|
|
cvta.to.local.u64 %rd102, %rd321;
|
|
ld.const.u64 %rd355, [kernel_params+2360];
|
|
cvta.to.global.u64 %rd103, %rd355;
|
|
ld.const.u64 %rd356, [kernel_params+2792];
|
|
cvta.to.global.u64 %rd357, %rd356;
|
|
add.s64 %rd104, %rd357, %rd168;
|
|
cvta.to.local.u64 %rd105, %rd319;
|
|
cvta.to.global.u64 %rd106, %rd346;
|
|
add.s32 %r103, %r98, -1;
|
|
mov.f32 %f3861, 0f00000000;
|
|
mov.u16 %rs74, 0;
|
|
cvta.to.local.u64 %rd360, %rd55;
|
|
cvta.to.local.u64 %rd363, %rd58;
|
|
cvta.to.local.u64 %rd495, %rd98;
|
|
mov.pred %p337, 0;
|
|
bra.uni $L__BB11_228;
|
|
|
|
$L__BB11_357:
|
|
ld.local.f32 %f3862, [%rd53+-32];
|
|
|
|
$L__BB11_228:
|
|
mov.b32 %r105, %f3862;
|
|
setp.nan.ftz.f32 %p249, %f3862, %f3862;
|
|
mov.u16 %rs107, %rs74;
|
|
@%p249 bra $L__BB11_231;
|
|
|
|
or.b32 %r1271, %r105, -2147483648;
|
|
setp.ne.s32 %p250, %r1271, -2147483648;
|
|
add.ftz.f32 %f2394, %f3862, %f3862;
|
|
setp.eq.ftz.f32 %p251, %f2394, %f3862;
|
|
and.pred %p252, %p251, %p250;
|
|
mov.u16 %rs107, %rs74;
|
|
@%p252 bra $L__BB11_231;
|
|
|
|
shl.b32 %r1272, %r105, 1;
|
|
setp.lt.u32 %p253, %r1272, -16777215;
|
|
selp.u16 %rs107, 1, 0, %p253;
|
|
|
|
$L__BB11_231:
|
|
ld.local.f32 %f578, [%rd53+-20];
|
|
setp.eq.s16 %p254, %rs107, 0;
|
|
mov.u16 %rs109, %rs74;
|
|
@%p254 bra $L__BB11_237;
|
|
|
|
mov.b32 %r106, %f578;
|
|
setp.nan.ftz.f32 %p255, %f578, %f578;
|
|
mov.u16 %rs109, 0;
|
|
mov.u16 %rs108, %rs109;
|
|
@%p255 bra $L__BB11_235;
|
|
|
|
or.b32 %r1273, %r106, -2147483648;
|
|
setp.ne.s32 %p256, %r1273, -2147483648;
|
|
add.ftz.f32 %f2395, %f578, %f578;
|
|
setp.eq.ftz.f32 %p257, %f2395, %f578;
|
|
and.pred %p258, %p257, %p256;
|
|
mov.u16 %rs108, %rs109;
|
|
@%p258 bra $L__BB11_235;
|
|
|
|
shl.b32 %r1274, %r106, 1;
|
|
setp.lt.u32 %p259, %r1274, -16777215;
|
|
selp.u16 %rs108, 1, 0, %p259;
|
|
|
|
$L__BB11_235:
|
|
setp.eq.s16 %p260, %rs108, 0;
|
|
@%p260 bra $L__BB11_237;
|
|
|
|
ld.local.f32 %f2396, [%rd53+-16];
|
|
mul.ftz.f32 %f2397, %f2396, %f2396;
|
|
fma.rn.ftz.f32 %f2398, %f578, %f578, %f2397;
|
|
ld.local.f32 %f2399, [%rd53+-12];
|
|
fma.rn.ftz.f32 %f2400, %f2399, %f2399, %f2398;
|
|
setp.neu.ftz.f32 %p261, %f2400, 0f00000000;
|
|
selp.u16 %rs109, 1, 0, %p261;
|
|
|
|
$L__BB11_237:
|
|
setp.eq.s16 %p262, %rs109, 0;
|
|
selp.b64 %rd359, 0, %rd703, %p262;
|
|
ld.local.f32 %f2402, [%rd53+-28];
|
|
ld.local.f32 %f2403, [%rd53+-24];
|
|
ld.local.f32 %f2405, [%rd53+-16];
|
|
ld.local.f32 %f2406, [%rd53+-12];
|
|
ld.local.f32 %f2408, [%rd53+-4];
|
|
ld.local.f32 %f2409, [%rd53];
|
|
mov.u32 %r1309, 2;
|
|
mov.u32 %r1310, 3;
|
|
mov.u32 %r1313, 8;
|
|
mov.u32 %r1319, 0;
|
|
// begin inline asm
|
|
call(%r1275,%r1276,%r1277,%r1278,%r1279,%r1280,%r1281,%r1282,%r1283,%r1284,%r1285,%r1286,%r1287,%r1288,%r1289,%r1290,%r1291,%r1292,%r1293,%r1294,%r1295,%r1296,%r1297,%r1298,%r1299,%r1300,%r1301,%r1302,%r1303,%r1304,%r1305,%r1306),_optix_trace_typed_32,(%r1319,%rd359,%f3862,%f2402,%f2403,%f578,%f2405,%f2406,%f3861,%f2408,%f2409,%r102,%r1309,%r1310,%r1319,%r1319,%r1313,%r1319,%r1319,%r1319,%r1319,%r97,%r1319,%r100,%r101,%r1346,%r1347,%r1348,%r1349,%r1350,%r1351,%r1352,%r1353,%r1354,%r1355,%r1356,%r1357,%r1358,%r1359,%r1360,%r1361,%r1362,%r1363,%r1364,%r1365,%r1366,%r1367,%r1368,%r1369);
|
|
// end inline asm
|
|
setp.eq.s32 %p263, %r1280, 0;
|
|
@%p263 bra $L__BB11_358;
|
|
|
|
mov.u32 %r1481, 0;
|
|
mov.b32 %f3697, %r1275;
|
|
st.local.f32 [%rd54], %f3697;
|
|
st.local.v4.u32 [%rd360], {%r1280, %r1276, %r1277, %r1279};
|
|
mul.wide.s32 %rd361, %r1279, 4;
|
|
add.s64 %rd362, %rd52, %rd361;
|
|
ld.global.u32 %r1520, [%rd362];
|
|
st.local.u32 [%rd56], %r1520;
|
|
st.local.u32 [%rd57], %r1278;
|
|
st.local.u32 [%rd59], %r1481;
|
|
ld.local.f32 %f582, [%rd53];
|
|
mov.b32 %r1371, %f582;
|
|
st.local.v2.u32 [%rd363], {%r1268, %r1371};
|
|
and.b32 %r146, %r1520, 2;
|
|
setp.eq.s32 %p264, %r146, 0;
|
|
mul.wide.s32 %rd364, %r1279, 208;
|
|
add.s64 %rd107, %rd66, %rd364;
|
|
@%p264 bra $L__BB11_245;
|
|
|
|
ld.global.u32 %rd365, [%rd107+168];
|
|
ld.global.u32 %r1373, [%rd107+152];
|
|
shl.b32 %r1374, %r1373, 1;
|
|
cvt.rn.f32.s32 %f2410, %r1374;
|
|
mul.ftz.f32 %f2411, %f582, %f2410;
|
|
cvt.rzi.ftz.s32.f32 %r1375, %f2411;
|
|
add.s32 %r1376, %r1374, -1;
|
|
min.s32 %r1377, %r1375, %r1376;
|
|
cvt.rn.f32.s32 %f2412, %r1377;
|
|
sub.ftz.f32 %f583, %f2411, %f2412;
|
|
cvt.s64.s32 %rd366, %r1377;
|
|
add.s64 %rd367, %rd366, %rd365;
|
|
shl.b64 %rd368, %rd367, 6;
|
|
add.s64 %rd108, %rd67, %rd368;
|
|
ld.global.v4.f32 {%f2413, %f2414, %f2415, %f2416}, [%rd108];
|
|
ld.global.v4.f32 {%f2417, %f2418, %f2419, %f2420}, [%rd108+64];
|
|
mul.ftz.f32 %f2421, %f2414, %f2418;
|
|
fma.rn.ftz.f32 %f2422, %f2413, %f2417, %f2421;
|
|
mul.ftz.f32 %f2423, %f2416, %f2420;
|
|
fma.rn.ftz.f32 %f2424, %f2415, %f2419, %f2423;
|
|
add.ftz.f32 %f592, %f2422, %f2424;
|
|
setp.gt.ftz.f32 %p265, %f592, 0f3F7FDF3B;
|
|
@%p265 bra $L__BB11_241;
|
|
bra.uni $L__BB11_240;
|
|
|
|
$L__BB11_241:
|
|
mov.f32 %f3714, 0f3F800000;
|
|
sub.ftz.f32 %f3713, %f3714, %f583;
|
|
mul.ftz.f32 %f2479, %f2413, %f3713;
|
|
fma.rn.ftz.f32 %f2480, %f2417, %f583, %f2479;
|
|
mul.ftz.f32 %f2481, %f2414, %f3713;
|
|
fma.rn.ftz.f32 %f2482, %f2418, %f583, %f2481;
|
|
mul.ftz.f32 %f2483, %f2415, %f3713;
|
|
fma.rn.ftz.f32 %f2484, %f2419, %f583, %f2483;
|
|
mul.ftz.f32 %f2485, %f2416, %f3713;
|
|
fma.rn.ftz.f32 %f2486, %f2420, %f583, %f2485;
|
|
mul.ftz.f32 %f2487, %f2482, %f2482;
|
|
fma.rn.ftz.f32 %f2488, %f2480, %f2480, %f2487;
|
|
mul.ftz.f32 %f2489, %f2486, %f2486;
|
|
fma.rn.ftz.f32 %f2490, %f2484, %f2484, %f2489;
|
|
add.ftz.f32 %f2491, %f2488, %f2490;
|
|
rsqrt.approx.ftz.f32 %f2492, %f2491;
|
|
mul.ftz.f32 %f3866, %f2486, %f2492;
|
|
mul.ftz.f32 %f3865, %f2484, %f2492;
|
|
mul.ftz.f32 %f3864, %f2482, %f2492;
|
|
mul.ftz.f32 %f3863, %f2480, %f2492;
|
|
bra.uni $L__BB11_242;
|
|
|
|
$L__BB11_240:
|
|
mov.f32 %f3710, 0f3F800000;
|
|
mov.f32 %f2426, 0fBF800000;
|
|
max.ftz.f32 %f2427, %f592, %f2426;
|
|
min.ftz.f32 %f2429, %f2427, %f3710;
|
|
abs.ftz.f32 %f2430, %f2429;
|
|
sub.ftz.f32 %f2431, %f3710, %f2430;
|
|
mul.ftz.f32 %f2432, %f2431, 0f3F000000;
|
|
sqrt.approx.ftz.f32 %f2433, %f2432;
|
|
setp.gt.ftz.f32 %p266, %f2430, 0f3F11EB85;
|
|
selp.f32 %f2434, %f2433, %f2430, %p266;
|
|
mul.ftz.f32 %f2435, %f2434, %f2434;
|
|
mov.f32 %f2436, 0f3C94D2E9;
|
|
mov.f32 %f2437, 0f3D53F941;
|
|
fma.rn.ftz.f32 %f2438, %f2437, %f2435, %f2436;
|
|
mov.f32 %f2439, 0f3D3F841F;
|
|
fma.rn.ftz.f32 %f2440, %f2438, %f2435, %f2439;
|
|
mov.f32 %f2441, 0f3D994929;
|
|
fma.rn.ftz.f32 %f2442, %f2440, %f2435, %f2441;
|
|
mov.f32 %f2443, 0f3E2AAB94;
|
|
fma.rn.ftz.f32 %f2444, %f2442, %f2435, %f2443;
|
|
mul.ftz.f32 %f2445, %f2435, %f2444;
|
|
fma.rn.ftz.f32 %f2446, %f2445, %f2434, %f2434;
|
|
add.ftz.f32 %f2447, %f2446, %f2446;
|
|
mov.f32 %f2448, 0f3FC90FDB;
|
|
sub.ftz.f32 %f2449, %f2448, %f2446;
|
|
selp.f32 %f2450, %f2447, %f2449, %p266;
|
|
setp.lt.ftz.f32 %p267, %f2429, 0f00000000;
|
|
mov.f32 %f2451, 0f40490FDB;
|
|
sub.ftz.f32 %f2452, %f2451, %f2450;
|
|
selp.f32 %f2453, %f2452, %f2450, %p267;
|
|
mul.ftz.f32 %f2454, %f2413, %f592;
|
|
sub.ftz.f32 %f2455, %f2417, %f2454;
|
|
mul.ftz.f32 %f2456, %f2414, %f592;
|
|
sub.ftz.f32 %f2457, %f2418, %f2456;
|
|
mul.ftz.f32 %f2458, %f2415, %f592;
|
|
sub.ftz.f32 %f2459, %f2419, %f2458;
|
|
mul.ftz.f32 %f2460, %f2416, %f592;
|
|
sub.ftz.f32 %f2461, %f2420, %f2460;
|
|
mul.ftz.f32 %f2462, %f2457, %f2457;
|
|
fma.rn.ftz.f32 %f2463, %f2455, %f2455, %f2462;
|
|
mul.ftz.f32 %f2464, %f2461, %f2461;
|
|
fma.rn.ftz.f32 %f2465, %f2459, %f2459, %f2464;
|
|
add.ftz.f32 %f2466, %f2463, %f2465;
|
|
rsqrt.approx.ftz.f32 %f2467, %f2466;
|
|
mul.ftz.f32 %f2468, %f2455, %f2467;
|
|
mul.ftz.f32 %f2469, %f2457, %f2467;
|
|
mul.ftz.f32 %f2470, %f2459, %f2467;
|
|
mul.ftz.f32 %f2471, %f2461, %f2467;
|
|
mul.ftz.f32 %f2472, %f583, %f2453;
|
|
cos.approx.ftz.f32 %f2473, %f2472;
|
|
sin.approx.ftz.f32 %f2474, %f2472;
|
|
mul.ftz.f32 %f2475, %f2468, %f2474;
|
|
mul.ftz.f32 %f2476, %f2469, %f2474;
|
|
mul.ftz.f32 %f2477, %f2470, %f2474;
|
|
mul.ftz.f32 %f2478, %f2471, %f2474;
|
|
fma.rn.ftz.f32 %f3866, %f2416, %f2473, %f2478;
|
|
fma.rn.ftz.f32 %f3865, %f2415, %f2473, %f2477;
|
|
fma.rn.ftz.f32 %f3864, %f2414, %f2473, %f2476;
|
|
fma.rn.ftz.f32 %f3863, %f2413, %f2473, %f2475;
|
|
|
|
$L__BB11_242:
|
|
mov.f32 %f3712, 0f3F800000;
|
|
sub.ftz.f32 %f3711, %f3712, %f583;
|
|
ld.global.v4.f32 {%f2493, %f2494, %f2495, %f2496}, [%rd108+16];
|
|
ld.global.v4.f32 {%f2501, %f2502, %f2503, %f2504}, [%rd108+80];
|
|
mul.ftz.f32 %f2509, %f583, %f2501;
|
|
mul.ftz.f32 %f2510, %f583, %f2502;
|
|
mul.ftz.f32 %f2511, %f583, %f2503;
|
|
mul.ftz.f32 %f2512, %f583, %f2504;
|
|
fma.rn.ftz.f32 %f606, %f3711, %f2493, %f2509;
|
|
fma.rn.ftz.f32 %f607, %f3711, %f2494, %f2510;
|
|
fma.rn.ftz.f32 %f608, %f3711, %f2495, %f2511;
|
|
fma.rn.ftz.f32 %f2513, %f3711, %f2496, %f2512;
|
|
ld.global.v4.f32 {%f2514, %f2515, %f2516, %f2517}, [%rd108+32];
|
|
ld.global.v4.f32 {%f2522, %f2523, %f2524, %f2525}, [%rd108+96];
|
|
mul.ftz.f32 %f2530, %f583, %f2522;
|
|
mul.ftz.f32 %f2531, %f583, %f2523;
|
|
mul.ftz.f32 %f2532, %f583, %f2524;
|
|
mul.ftz.f32 %f2533, %f583, %f2525;
|
|
fma.rn.ftz.f32 %f2534, %f3711, %f2514, %f2530;
|
|
fma.rn.ftz.f32 %f2535, %f3711, %f2515, %f2531;
|
|
fma.rn.ftz.f32 %f2536, %f3711, %f2516, %f2532;
|
|
fma.rn.ftz.f32 %f2537, %f3711, %f2517, %f2533;
|
|
ld.global.v4.f32 {%f2538, %f2539, %f2540, %f2541}, [%rd108+48];
|
|
ld.global.v4.f32 {%f2546, %f2547, %f2548, %f2549}, [%rd108+112];
|
|
mul.ftz.f32 %f2554, %f583, %f2546;
|
|
mul.ftz.f32 %f2555, %f583, %f2547;
|
|
mul.ftz.f32 %f2556, %f583, %f2548;
|
|
mul.ftz.f32 %f2557, %f583, %f2549;
|
|
fma.rn.ftz.f32 %f2558, %f3711, %f2538, %f2554;
|
|
fma.rn.ftz.f32 %f2559, %f3711, %f2539, %f2555;
|
|
fma.rn.ftz.f32 %f2560, %f3711, %f2540, %f2556;
|
|
fma.rn.ftz.f32 %f2561, %f3711, %f2541, %f2557;
|
|
mul.ftz.f32 %f2562, %f3866, 0f3FB504F3;
|
|
mul.ftz.f32 %f2563, %f3863, 0f3FB504F3;
|
|
mul.ftz.f32 %f2564, %f3864, 0f3FB504F3;
|
|
mul.ftz.f32 %f2565, %f3865, 0f3FB504F3;
|
|
mul.ftz.f32 %f2566, %f2562, %f2563;
|
|
mul.ftz.f32 %f2567, %f2562, %f2564;
|
|
mul.ftz.f32 %f2568, %f2562, %f2565;
|
|
mul.ftz.f32 %f2569, %f2563, %f2563;
|
|
mul.ftz.f32 %f2570, %f2563, %f2564;
|
|
mul.ftz.f32 %f2571, %f2563, %f2565;
|
|
mul.ftz.f32 %f2572, %f2564, %f2564;
|
|
mul.ftz.f32 %f2573, %f2564, %f2565;
|
|
mul.ftz.f32 %f2574, %f2565, %f2565;
|
|
sub.ftz.f32 %f2576, %f3712, %f2572;
|
|
sub.ftz.f32 %f2577, %f2576, %f2574;
|
|
sub.ftz.f32 %f2578, %f2570, %f2568;
|
|
add.ftz.f32 %f2579, %f2567, %f2571;
|
|
add.ftz.f32 %f2580, %f2568, %f2570;
|
|
sub.ftz.f32 %f2581, %f3712, %f2569;
|
|
sub.ftz.f32 %f2582, %f2581, %f2574;
|
|
sub.ftz.f32 %f2583, %f2573, %f2566;
|
|
sub.ftz.f32 %f2584, %f2571, %f2567;
|
|
add.ftz.f32 %f2585, %f2566, %f2573;
|
|
sub.ftz.f32 %f2586, %f2581, %f2572;
|
|
mul.ftz.f32 %f2587, %f2578, %f2536;
|
|
fma.rn.ftz.f32 %f2588, %f2577, %f2513, %f2587;
|
|
fma.rn.ftz.f32 %f3872, %f2579, %f2559, %f2588;
|
|
mul.ftz.f32 %f2589, %f2578, %f2537;
|
|
fma.rn.ftz.f32 %f2590, %f2577, %f2534, %f2589;
|
|
fma.rn.ftz.f32 %f610, %f2579, %f2560, %f2590;
|
|
mul.ftz.f32 %f2591, %f2578, %f2558;
|
|
fma.rn.ftz.f32 %f2592, %f2577, %f2535, %f2591;
|
|
fma.rn.ftz.f32 %f611, %f2579, %f2561, %f2592;
|
|
mul.ftz.f32 %f2593, %f2582, %f2536;
|
|
fma.rn.ftz.f32 %f2594, %f2580, %f2513, %f2593;
|
|
fma.rn.ftz.f32 %f612, %f2583, %f2559, %f2594;
|
|
mul.ftz.f32 %f2595, %f2582, %f2537;
|
|
fma.rn.ftz.f32 %f2596, %f2580, %f2534, %f2595;
|
|
fma.rn.ftz.f32 %f3871, %f2583, %f2560, %f2596;
|
|
mul.ftz.f32 %f2597, %f2582, %f2558;
|
|
fma.rn.ftz.f32 %f2598, %f2580, %f2535, %f2597;
|
|
fma.rn.ftz.f32 %f614, %f2583, %f2561, %f2598;
|
|
mul.ftz.f32 %f2599, %f2585, %f2536;
|
|
fma.rn.ftz.f32 %f2600, %f2584, %f2513, %f2599;
|
|
fma.rn.ftz.f32 %f615, %f2586, %f2559, %f2600;
|
|
mul.ftz.f32 %f2601, %f2585, %f2537;
|
|
fma.rn.ftz.f32 %f2602, %f2584, %f2534, %f2601;
|
|
fma.rn.ftz.f32 %f616, %f2586, %f2560, %f2602;
|
|
mul.ftz.f32 %f2603, %f2585, %f2558;
|
|
fma.rn.ftz.f32 %f2604, %f2584, %f2535, %f2603;
|
|
fma.rn.ftz.f32 %f3870, %f2586, %f2561, %f2604;
|
|
cvta.to.local.u64 %rd369, %rd65;
|
|
st.local.v4.f32 [%rd369], {%f3872, %f610, %f611, %f606};
|
|
cvta.to.local.u64 %rd370, %rd75;
|
|
st.local.v4.f32 [%rd370], {%f612, %f3871, %f614, %f607};
|
|
cvta.to.local.u64 %rd371, %rd74;
|
|
st.local.v4.f32 [%rd371], {%f615, %f616, %f3870, %f608};
|
|
mul.ftz.f32 %f2605, %f3871, %f3870;
|
|
mul.ftz.f32 %f618, %f616, %f614;
|
|
sub.ftz.f32 %f3869, %f2605, %f618;
|
|
mul.ftz.f32 %f620, %f616, %f611;
|
|
mul.ftz.f32 %f2606, %f610, %f3870;
|
|
sub.ftz.f32 %f3868, %f620, %f2606;
|
|
mul.ftz.f32 %f622, %f610, %f614;
|
|
mul.ftz.f32 %f2607, %f3871, %f611;
|
|
sub.ftz.f32 %f3867, %f622, %f2607;
|
|
mul.ftz.f32 %f2608, %f612, %f3868;
|
|
fma.rn.ftz.f32 %f2609, %f3872, %f3869, %f2608;
|
|
fma.rn.ftz.f32 %f3873, %f615, %f3867, %f2609;
|
|
setp.neu.ftz.f32 %p268, %f3873, 0f00000000;
|
|
@%p268 bra $L__BB11_244;
|
|
|
|
add.ftz.f32 %f3872, %f3872, 0f322BCC77;
|
|
add.ftz.f32 %f3870, %f3870, 0f322BCC77;
|
|
add.ftz.f32 %f3871, %f3871, 0f322BCC77;
|
|
mul.ftz.f32 %f2610, %f3871, %f3870;
|
|
sub.ftz.f32 %f3869, %f2610, %f618;
|
|
mul.ftz.f32 %f2611, %f610, %f3870;
|
|
sub.ftz.f32 %f3868, %f620, %f2611;
|
|
mul.ftz.f32 %f2612, %f3871, %f611;
|
|
sub.ftz.f32 %f3867, %f622, %f2612;
|
|
mul.ftz.f32 %f2613, %f612, %f3868;
|
|
fma.rn.ftz.f32 %f2614, %f3872, %f3869, %f2613;
|
|
fma.rn.ftz.f32 %f2615, %f615, %f3867, %f2614;
|
|
setp.eq.ftz.f32 %p269, %f2615, 0f00000000;
|
|
selp.f32 %f3873, 0f7F7FFFFF, %f2615, %p269;
|
|
|
|
$L__BB11_244:
|
|
mul.ftz.f32 %f2616, %f612, %f3870;
|
|
mul.ftz.f32 %f2617, %f615, %f614;
|
|
sub.ftz.f32 %f2618, %f2617, %f2616;
|
|
mul.ftz.f32 %f2619, %f615, %f611;
|
|
mul.ftz.f32 %f2620, %f3870, %f3872;
|
|
sub.ftz.f32 %f2621, %f2620, %f2619;
|
|
mul.ftz.f32 %f2622, %f614, %f3872;
|
|
mul.ftz.f32 %f2623, %f612, %f611;
|
|
sub.ftz.f32 %f2624, %f2623, %f2622;
|
|
mul.ftz.f32 %f2625, %f615, %f3871;
|
|
mul.ftz.f32 %f2626, %f612, %f616;
|
|
sub.ftz.f32 %f2627, %f2626, %f2625;
|
|
mul.ftz.f32 %f2628, %f616, %f3872;
|
|
mul.ftz.f32 %f2629, %f615, %f610;
|
|
sub.ftz.f32 %f2630, %f2629, %f2628;
|
|
mul.ftz.f32 %f2631, %f612, %f610;
|
|
mul.ftz.f32 %f2632, %f3871, %f3872;
|
|
sub.ftz.f32 %f2633, %f2632, %f2631;
|
|
div.approx.ftz.f32 %f2634, %f3868, %f3873;
|
|
mul.ftz.f32 %f2635, %f607, %f2634;
|
|
div.approx.ftz.f32 %f2636, %f3869, %f3873;
|
|
fma.rn.ftz.f32 %f2637, %f606, %f2636, %f2635;
|
|
div.approx.ftz.f32 %f2638, %f3867, %f3873;
|
|
fma.rn.ftz.f32 %f2639, %f608, %f2638, %f2637;
|
|
div.approx.ftz.f32 %f2640, %f2621, %f3873;
|
|
mul.ftz.f32 %f2641, %f607, %f2640;
|
|
div.approx.ftz.f32 %f2642, %f2618, %f3873;
|
|
fma.rn.ftz.f32 %f2643, %f606, %f2642, %f2641;
|
|
div.approx.ftz.f32 %f2644, %f2624, %f3873;
|
|
fma.rn.ftz.f32 %f2645, %f608, %f2644, %f2643;
|
|
div.approx.ftz.f32 %f2646, %f2630, %f3873;
|
|
mul.ftz.f32 %f2647, %f607, %f2646;
|
|
div.approx.ftz.f32 %f2648, %f2627, %f3873;
|
|
fma.rn.ftz.f32 %f2649, %f606, %f2648, %f2647;
|
|
div.approx.ftz.f32 %f2650, %f2633, %f3873;
|
|
fma.rn.ftz.f32 %f2651, %f608, %f2650, %f2649;
|
|
neg.ftz.f32 %f2652, %f2639;
|
|
cvta.to.local.u64 %rd372, %rd76;
|
|
st.local.v4.f32 [%rd372], {%f2636, %f2634, %f2638, %f2652};
|
|
neg.ftz.f32 %f2653, %f2645;
|
|
cvta.to.local.u64 %rd373, %rd90;
|
|
st.local.v4.f32 [%rd373], {%f2642, %f2640, %f2644, %f2653};
|
|
neg.ftz.f32 %f2654, %f2651;
|
|
cvta.to.local.u64 %rd374, %rd89;
|
|
st.local.v4.f32 [%rd374], {%f2648, %f2646, %f2650, %f2654};
|
|
|
|
$L__BB11_245:
|
|
ld.local.f32 %f639, [%rd53+-20];
|
|
neg.ftz.f32 %f640, %f639;
|
|
ld.local.f32 %f641, [%rd53+-16];
|
|
neg.ftz.f32 %f642, %f641;
|
|
ld.local.f32 %f3935, [%rd53+-12];
|
|
neg.ftz.f32 %f2655, %f3935;
|
|
st.local.f32 [%rd60], %f640;
|
|
st.local.f32 [%rd62], %f642;
|
|
st.local.f32 [%rd64], %f2655;
|
|
and.b32 %r1378, %r1280, 6;
|
|
setp.eq.s32 %p270, %r1378, 0;
|
|
@%p270 bra $L__BB11_285;
|
|
|
|
mov.b32 %f3889, %r1275;
|
|
ld.local.f32 %f3893, [%rd53+-32];
|
|
ld.local.f32 %f3894, [%rd53+-28];
|
|
ld.local.f32 %f3895, [%rd53+-24];
|
|
and.b32 %r147, %r1520, 4;
|
|
setp.ne.s32 %p271, %r147, 0;
|
|
mov.f32 %f3890, %f639;
|
|
mov.f32 %f3891, %f641;
|
|
mov.f32 %f3892, %f3935;
|
|
@%p271 bra $L__BB11_253;
|
|
|
|
@%p264 bra $L__BB11_249;
|
|
bra.uni $L__BB11_248;
|
|
|
|
$L__BB11_249:
|
|
ld.global.v4.f32 {%f3874, %f3875, %f3876, %f3877}, [%rd107+48];
|
|
ld.global.v4.f32 {%f3878, %f3879, %f3880, %f3881}, [%rd107+64];
|
|
ld.global.v4.f32 {%f3882, %f3883, %f3884, %f3885}, [%rd107+80];
|
|
bra.uni $L__BB11_250;
|
|
|
|
$L__BB11_285:
|
|
and.b32 %r1405, %r1280, 8;
|
|
setp.eq.s32 %p295, %r1405, 0;
|
|
@%p295 bra $L__BB11_306;
|
|
bra.uni $L__BB11_286;
|
|
|
|
$L__BB11_306:
|
|
setp.eq.s32 %p308, %r1280, 1;
|
|
@%p308 bra $L__BB11_308;
|
|
bra.uni $L__BB11_307;
|
|
|
|
$L__BB11_308:
|
|
mul.wide.s32 %rd434, %r1278, 16;
|
|
add.s64 %rd435, %rd91, %rd434;
|
|
ld.global.u32 %r1422, [%rd435+12];
|
|
mul.wide.u32 %rd436, %r1422, 12;
|
|
add.s64 %rd437, %rd92, %rd436;
|
|
add.s32 %r1423, %r1422, 1;
|
|
mul.wide.u32 %rd438, %r1423, 12;
|
|
add.s64 %rd439, %rd92, %rd438;
|
|
add.s32 %r1424, %r1422, 2;
|
|
mul.wide.u32 %rd440, %r1424, 12;
|
|
add.s64 %rd441, %rd92, %rd440;
|
|
and.b32 %r1425, %r1520, 8;
|
|
setp.eq.s32 %p309, %r1425, 0;
|
|
ld.global.f32 %f3212, [%rd441];
|
|
ld.global.f32 %f924, [%rd437];
|
|
sub.ftz.f32 %f925, %f3212, %f924;
|
|
ld.global.f32 %f3213, [%rd441+4];
|
|
ld.global.f32 %f926, [%rd437+4];
|
|
sub.ftz.f32 %f927, %f3213, %f926;
|
|
ld.global.f32 %f3214, [%rd441+8];
|
|
ld.global.f32 %f928, [%rd437+8];
|
|
sub.ftz.f32 %f929, %f3214, %f928;
|
|
ld.global.f32 %f3215, [%rd439];
|
|
sub.ftz.f32 %f930, %f3215, %f924;
|
|
ld.global.f32 %f3216, [%rd439+4];
|
|
sub.ftz.f32 %f931, %f3216, %f926;
|
|
ld.global.f32 %f3217, [%rd439+8];
|
|
sub.ftz.f32 %f932, %f3217, %f928;
|
|
mul.ftz.f32 %f933, %f932, %f927;
|
|
mul.ftz.f32 %f934, %f931, %f929;
|
|
mul.ftz.f32 %f935, %f930, %f929;
|
|
mul.ftz.f32 %f936, %f932, %f925;
|
|
mul.ftz.f32 %f937, %f931, %f925;
|
|
mul.ftz.f32 %f938, %f930, %f927;
|
|
@%p309 bra $L__BB11_310;
|
|
bra.uni $L__BB11_309;
|
|
|
|
$L__BB11_310:
|
|
sub.ftz.f32 %f3225, %f934, %f933;
|
|
sub.ftz.f32 %f3226, %f936, %f935;
|
|
mul.ftz.f32 %f3227, %f3226, %f3226;
|
|
fma.rn.ftz.f32 %f3228, %f3225, %f3225, %f3227;
|
|
sub.ftz.f32 %f3229, %f938, %f937;
|
|
fma.rn.ftz.f32 %f3230, %f3229, %f3229, %f3228;
|
|
sqrt.approx.ftz.f32 %f3231, %f3230;
|
|
div.approx.ftz.f32 %f3960, %f3229, %f3231;
|
|
div.approx.ftz.f32 %f3959, %f3226, %f3231;
|
|
div.approx.ftz.f32 %f3958, %f3225, %f3231;
|
|
bra.uni $L__BB11_311;
|
|
|
|
$L__BB11_286:
|
|
mov.b32 %f3701, %r1275;
|
|
cvt.s64.s32 %rd116, %r1278;
|
|
mul.wide.s32 %rd414, %r1278, 4;
|
|
add.s64 %rd415, %rd77, %rd414;
|
|
ld.global.u32 %r1406, [%rd415];
|
|
st.local.u32 [%rd70], %r1406;
|
|
ld.local.f32 %f3110, [%rd53+-32];
|
|
fma.rn.ftz.f32 %f887, %f639, %f3701, %f3110;
|
|
ld.local.f32 %f3111, [%rd53+-28];
|
|
fma.rn.ftz.f32 %f888, %f641, %f3701, %f3111;
|
|
ld.local.f32 %f3112, [%rd53+-24];
|
|
fma.rn.ftz.f32 %f889, %f3935, %f3701, %f3112;
|
|
cvta.to.local.u64 %rd417, %rd145;
|
|
st.local.v2.f32 [%rd417], {%f887, %f888};
|
|
st.local.f32 [%rd78], %f889;
|
|
and.b32 %r1407, %r1280, 64;
|
|
setp.eq.s32 %p296, %r1407, 0;
|
|
@%p296 bra $L__BB11_300;
|
|
bra.uni $L__BB11_287;
|
|
|
|
$L__BB11_300:
|
|
shl.b64 %rd426, %rd116, 4;
|
|
add.s64 %rd427, %rd93, %rd426;
|
|
ld.global.v4.f32 {%f3955, %f3956, %f3957, %f3137}, [%rd427];
|
|
bra.uni $L__BB11_301;
|
|
|
|
$L__BB11_248:
|
|
cvta.to.local.u64 %rd375, %rd76;
|
|
ld.local.v4.f32 {%f3874, %f3875, %f3876, %f3877}, [%rd375];
|
|
cvta.to.local.u64 %rd376, %rd90;
|
|
ld.local.v4.f32 {%f3878, %f3879, %f3880, %f3881}, [%rd376];
|
|
cvta.to.local.u64 %rd377, %rd89;
|
|
ld.local.v4.f32 {%f3882, %f3883, %f3884, %f3885}, [%rd377];
|
|
|
|
$L__BB11_250:
|
|
mov.b32 %f3700, %r1275;
|
|
mul.ftz.f32 %f2680, %f3894, %f3875;
|
|
fma.rn.ftz.f32 %f2681, %f3893, %f3874, %f2680;
|
|
fma.rn.ftz.f32 %f2682, %f3895, %f3876, %f2681;
|
|
add.ftz.f32 %f683, %f3877, %f2682;
|
|
mul.ftz.f32 %f2683, %f3894, %f3879;
|
|
fma.rn.ftz.f32 %f2684, %f3893, %f3878, %f2683;
|
|
fma.rn.ftz.f32 %f2685, %f3895, %f3880, %f2684;
|
|
add.ftz.f32 %f684, %f3881, %f2685;
|
|
mul.ftz.f32 %f2686, %f3894, %f3883;
|
|
fma.rn.ftz.f32 %f2687, %f3893, %f3882, %f2686;
|
|
fma.rn.ftz.f32 %f2688, %f3895, %f3884, %f2687;
|
|
add.ftz.f32 %f3895, %f3885, %f2688;
|
|
mul.ftz.f32 %f2689, %f639, %f3700;
|
|
mul.ftz.f32 %f2690, %f641, %f3700;
|
|
mul.ftz.f32 %f2691, %f2690, %f3875;
|
|
fma.rn.ftz.f32 %f2692, %f2689, %f3874, %f2691;
|
|
mul.ftz.f32 %f2693, %f3935, %f3700;
|
|
fma.rn.ftz.f32 %f3890, %f2693, %f3876, %f2692;
|
|
mul.ftz.f32 %f2694, %f2690, %f3879;
|
|
fma.rn.ftz.f32 %f2695, %f2689, %f3878, %f2694;
|
|
fma.rn.ftz.f32 %f3891, %f2693, %f3880, %f2695;
|
|
mul.ftz.f32 %f2696, %f2690, %f3883;
|
|
fma.rn.ftz.f32 %f2697, %f2689, %f3882, %f2696;
|
|
fma.rn.ftz.f32 %f3892, %f2693, %f3884, %f2697;
|
|
mul.ftz.f32 %f2698, %f3891, %f3891;
|
|
fma.rn.ftz.f32 %f2699, %f3890, %f3890, %f2698;
|
|
fma.rn.ftz.f32 %f2700, %f3892, %f3892, %f2699;
|
|
sqrt.approx.ftz.f32 %f3889, %f2700;
|
|
setp.neu.ftz.f32 %p273, %f3889, 0f00000000;
|
|
@%p273 bra $L__BB11_251;
|
|
bra.uni $L__BB11_252;
|
|
|
|
$L__BB11_251:
|
|
div.approx.ftz.f32 %f3892, %f3892, %f3889;
|
|
div.approx.ftz.f32 %f3891, %f3891, %f3889;
|
|
div.approx.ftz.f32 %f3890, %f3890, %f3889;
|
|
|
|
$L__BB11_252:
|
|
mov.f32 %f3893, %f683;
|
|
mov.f32 %f3894, %f684;
|
|
|
|
$L__BB11_253:
|
|
mul.wide.s32 %rd378, %r1278, 16;
|
|
add.s64 %rd379, %rd88, %rd378;
|
|
shr.s32 %r1379, %r1280, 7;
|
|
ld.global.u32 %r1380, [%rd379+4];
|
|
add.s32 %r148, %r1379, %r1380;
|
|
add.s32 %r1381, %r148, -1;
|
|
max.s32 %r149, %r1381, %r1380;
|
|
add.s32 %r1382, %r148, 2;
|
|
ld.global.u32 %r1383, [%rd379+8];
|
|
add.s32 %r1384, %r1380, %r1383;
|
|
add.s32 %r1385, %r1384, -1;
|
|
min.s32 %r150, %r1382, %r1385;
|
|
and.b32 %r1386, %r1280, 64;
|
|
setp.eq.s32 %p274, %r1386, 0;
|
|
mul.wide.s32 %rd380, %r149, 16;
|
|
add.s64 %rd110, %rd95, %rd380;
|
|
mul.wide.s32 %rd381, %r148, 16;
|
|
add.s64 %rd111, %rd95, %rd381;
|
|
mul.wide.s32 %rd382, %r150, 16;
|
|
add.s64 %rd112, %rd95, %rd382;
|
|
@%p274 bra $L__BB11_267;
|
|
bra.uni $L__BB11_254;
|
|
|
|
$L__BB11_267:
|
|
ld.global.v4.f32 {%f3920, %f3921, %f3922, %f2785}, [%rd110];
|
|
ld.global.v4.f32 {%f3923, %f3924, %f3925, %f2789}, [%rd111];
|
|
ld.global.v4.f32 {%f3926, %f3927, %f3928, %f2793}, [%rd111+16];
|
|
ld.global.v4.f32 {%f3929, %f3930, %f3931, %f2797}, [%rd112];
|
|
bra.uni $L__BB11_268;
|
|
|
|
$L__BB11_254:
|
|
ld.global.u32 %r151, [%rd107+148];
|
|
ld.global.u32 %r152, [%rd107+152];
|
|
shl.b32 %r1387, %r152, 1;
|
|
cvt.rn.f32.s32 %f2701, %r1387;
|
|
mul.ftz.f32 %f709, %f582, %f2701;
|
|
cvt.rzi.ftz.s32.f32 %r1388, %f709;
|
|
add.s32 %r1389, %r1387, -1;
|
|
min.s32 %r153, %r1388, %r1389;
|
|
ld.global.u32 %r1511, [%rd107+164];
|
|
mul.wide.u32 %rd383, %r1511, 16;
|
|
add.s64 %rd384, %rd94, %rd383;
|
|
ld.global.u16 %rs110, [%rd384+12];
|
|
ld.global.u32 %r1510, [%rd384+8];
|
|
ld.global.u64 %rd708, [%rd384];
|
|
setp.eq.s64 %p275, %rd708, 11;
|
|
@%p275 bra $L__BB11_259;
|
|
|
|
$L__BB11_255:
|
|
setp.eq.s64 %p276, %rd708, 0;
|
|
@%p276 bra $L__BB11_257;
|
|
|
|
add.s32 %r1511, %r1511, 2;
|
|
bra.uni $L__BB11_258;
|
|
|
|
$L__BB11_257:
|
|
setp.eq.s16 %p277, %rs110, 0;
|
|
mov.u32 %r1511, %r1510;
|
|
mov.u32 %r1514, %r1268;
|
|
@%p277 bra $L__BB11_260;
|
|
|
|
$L__BB11_258:
|
|
mul.wide.u32 %rd385, %r1511, 16;
|
|
add.s64 %rd386, %rd94, %rd385;
|
|
ld.global.u16 %rs110, [%rd386+12];
|
|
ld.global.u32 %r1510, [%rd386+8];
|
|
ld.global.u64 %rd708, [%rd386];
|
|
setp.ne.s64 %p278, %rd708, 11;
|
|
@%p278 bra $L__BB11_255;
|
|
|
|
$L__BB11_259:
|
|
setp.eq.s16 %p279, %rs110, 0;
|
|
selp.b32 %r1514, -1, %r1510, %p279;
|
|
|
|
$L__BB11_260:
|
|
setp.eq.s32 %p280, %r153, %r152;
|
|
@%p280 bra $L__BB11_262;
|
|
bra.uni $L__BB11_261;
|
|
|
|
$L__BB11_262:
|
|
ld.global.v4.f32 {%f3896, %f3897, %f3898, %f2721}, [%rd110];
|
|
ld.global.v4.f32 {%f3899, %f3900, %f3901, %f2725}, [%rd111];
|
|
ld.global.v4.f32 {%f3902, %f3903, %f3904, %f2729}, [%rd111+16];
|
|
ld.global.v4.f32 {%f3905, %f3906, %f3907, %f2733}, [%rd112];
|
|
bra.uni $L__BB11_263;
|
|
|
|
$L__BB11_261:
|
|
shr.s32 %r1484, %r1280, 7;
|
|
add.s32 %r1483, %r1484, %r1380;
|
|
setp.gt.s32 %p281, %r153, %r152;
|
|
selp.b32 %r1391, -1, 0, %p281;
|
|
add.s32 %r1392, %r153, %r1391;
|
|
mad.lo.s32 %r1393, %r1392, %r151, %r1514;
|
|
add.s32 %r1394, %r1393, %r149;
|
|
mul.wide.s32 %rd387, %r1394, 16;
|
|
add.s64 %rd388, %rd103, %rd387;
|
|
ld.global.v4.f32 {%f3896, %f3897, %f3898, %f2705}, [%rd388];
|
|
add.s32 %r1395, %r1393, %r1483;
|
|
mul.wide.s32 %rd389, %r1395, 16;
|
|
add.s64 %rd390, %rd103, %rd389;
|
|
ld.global.v4.f32 {%f3899, %f3900, %f3901, %f2709}, [%rd390];
|
|
ld.global.v4.f32 {%f3902, %f3903, %f3904, %f2713}, [%rd390+16];
|
|
add.s32 %r1396, %r1393, %r150;
|
|
mul.wide.s32 %rd391, %r1396, 16;
|
|
add.s64 %rd392, %rd103, %rd391;
|
|
ld.global.v4.f32 {%f3905, %f3906, %f3907, %f2717}, [%rd392];
|
|
|
|
$L__BB11_263:
|
|
add.s32 %r164, %r153, 1;
|
|
setp.eq.s32 %p282, %r164, %r152;
|
|
@%p282 bra $L__BB11_265;
|
|
bra.uni $L__BB11_264;
|
|
|
|
$L__BB11_265:
|
|
ld.global.v4.f32 {%f3908, %f3909, %f3910, %f2753}, [%rd110];
|
|
ld.global.v4.f32 {%f3911, %f3912, %f3913, %f2757}, [%rd111];
|
|
ld.global.v4.f32 {%f3914, %f3915, %f3916, %f2761}, [%rd111+16];
|
|
ld.global.v4.f32 {%f3917, %f3918, %f3919, %f2765}, [%rd112];
|
|
bra.uni $L__BB11_266;
|
|
|
|
$L__BB11_264:
|
|
shr.s32 %r1486, %r1280, 7;
|
|
add.s32 %r1485, %r1486, %r1380;
|
|
setp.ge.s32 %p283, %r153, %r152;
|
|
selp.b32 %r1397, -1, 0, %p283;
|
|
add.s32 %r1398, %r164, %r1397;
|
|
mad.lo.s32 %r1399, %r1398, %r151, %r1514;
|
|
add.s32 %r1400, %r1399, %r149;
|
|
mul.wide.s32 %rd393, %r1400, 16;
|
|
add.s64 %rd394, %rd103, %rd393;
|
|
ld.global.v4.f32 {%f3908, %f3909, %f3910, %f2737}, [%rd394];
|
|
add.s32 %r1401, %r1399, %r1485;
|
|
mul.wide.s32 %rd395, %r1401, 16;
|
|
add.s64 %rd396, %rd103, %rd395;
|
|
ld.global.v4.f32 {%f3911, %f3912, %f3913, %f2741}, [%rd396];
|
|
ld.global.v4.f32 {%f3914, %f3915, %f3916, %f2745}, [%rd396+16];
|
|
add.s32 %r1402, %r1399, %r150;
|
|
mul.wide.s32 %rd397, %r1402, 16;
|
|
add.s64 %rd398, %rd103, %rd397;
|
|
ld.global.v4.f32 {%f3917, %f3918, %f3919, %f2749}, [%rd398];
|
|
|
|
$L__BB11_266:
|
|
shl.b32 %r1487, %r152, 1;
|
|
cvt.rn.f32.s32 %f3709, %r1487;
|
|
mul.ftz.f32 %f3708, %f582, %f3709;
|
|
cvt.rn.f32.s32 %f2766, %r153;
|
|
sub.ftz.f32 %f2767, %f3708, %f2766;
|
|
mov.f32 %f2768, 0f3F800000;
|
|
sub.ftz.f32 %f2769, %f2768, %f2767;
|
|
mul.ftz.f32 %f2770, %f2767, %f3908;
|
|
fma.rn.ftz.f32 %f3920, %f2769, %f3896, %f2770;
|
|
mul.ftz.f32 %f2771, %f2767, %f3909;
|
|
fma.rn.ftz.f32 %f3921, %f2769, %f3897, %f2771;
|
|
mul.ftz.f32 %f2772, %f2767, %f3910;
|
|
fma.rn.ftz.f32 %f3922, %f2769, %f3898, %f2772;
|
|
mul.ftz.f32 %f2773, %f2767, %f3911;
|
|
fma.rn.ftz.f32 %f3923, %f2769, %f3899, %f2773;
|
|
mul.ftz.f32 %f2774, %f2767, %f3912;
|
|
fma.rn.ftz.f32 %f3924, %f2769, %f3900, %f2774;
|
|
mul.ftz.f32 %f2775, %f2767, %f3913;
|
|
fma.rn.ftz.f32 %f3925, %f2769, %f3901, %f2775;
|
|
mul.ftz.f32 %f2776, %f2767, %f3914;
|
|
fma.rn.ftz.f32 %f3926, %f2769, %f3902, %f2776;
|
|
mul.ftz.f32 %f2777, %f2767, %f3915;
|
|
fma.rn.ftz.f32 %f3927, %f2769, %f3903, %f2777;
|
|
mul.ftz.f32 %f2778, %f2767, %f3916;
|
|
fma.rn.ftz.f32 %f3928, %f2769, %f3904, %f2778;
|
|
mul.ftz.f32 %f2779, %f2767, %f3917;
|
|
fma.rn.ftz.f32 %f3929, %f2769, %f3905, %f2779;
|
|
mul.ftz.f32 %f2780, %f2767, %f3918;
|
|
fma.rn.ftz.f32 %f3930, %f2769, %f3906, %f2780;
|
|
mul.ftz.f32 %f2781, %f2767, %f3919;
|
|
fma.rn.ftz.f32 %f3931, %f2769, %f3907, %f2781;
|
|
|
|
$L__BB11_268:
|
|
mov.b32 %f3699, %r1276;
|
|
fma.rn.ftz.f32 %f3949, %f3889, %f3890, %f3893;
|
|
fma.rn.ftz.f32 %f3950, %f3889, %f3891, %f3894;
|
|
fma.rn.ftz.f32 %f3951, %f3889, %f3892, %f3895;
|
|
mov.f32 %f2798, 0f3F800000;
|
|
sub.ftz.f32 %f821, %f2798, %f3699;
|
|
mul.ftz.f32 %f822, %f821, %f821;
|
|
add.ftz.f32 %f2799, %f821, %f821;
|
|
mul.ftz.f32 %f2800, %f2799, %f3699;
|
|
sub.ftz.f32 %f2801, %f2800, %f822;
|
|
mul.ftz.f32 %f2802, %f3699, 0f40400000;
|
|
add.ftz.f32 %f823, %f2802, 0fC0A00000;
|
|
add.ftz.f32 %f2803, %f3699, %f3699;
|
|
mul.ftz.f32 %f2804, %f2803, %f823;
|
|
fma.rn.ftz.f32 %f2805, %f2802, %f3699, %f2804;
|
|
add.ftz.f32 %f2806, %f2802, 0f40000000;
|
|
mul.ftz.f32 %f2807, %f2799, %f2806;
|
|
mul.ftz.f32 %f824, %f821, 0f40400000;
|
|
mul.ftz.f32 %f2808, %f821, %f824;
|
|
sub.ftz.f32 %f2809, %f2807, %f2808;
|
|
mul.ftz.f32 %f2810, %f821, 0fC0000000;
|
|
mul.ftz.f32 %f825, %f3699, %f3699;
|
|
fma.rn.ftz.f32 %f2811, %f2810, %f3699, %f825;
|
|
mul.ftz.f32 %f2812, %f3923, %f2805;
|
|
mul.ftz.f32 %f2813, %f3924, %f2805;
|
|
mul.ftz.f32 %f2814, %f3925, %f2805;
|
|
fma.rn.ftz.f32 %f2815, %f3920, %f2801, %f2812;
|
|
fma.rn.ftz.f32 %f2816, %f3921, %f2801, %f2813;
|
|
fma.rn.ftz.f32 %f2817, %f3922, %f2801, %f2814;
|
|
fma.rn.ftz.f32 %f2818, %f3926, %f2809, %f2815;
|
|
fma.rn.ftz.f32 %f2819, %f3927, %f2809, %f2816;
|
|
fma.rn.ftz.f32 %f2820, %f3928, %f2809, %f2817;
|
|
fma.rn.ftz.f32 %f2821, %f3929, %f2811, %f2818;
|
|
fma.rn.ftz.f32 %f2822, %f3930, %f2811, %f2819;
|
|
fma.rn.ftz.f32 %f2823, %f3931, %f2811, %f2820;
|
|
mul.ftz.f32 %f3944, %f2821, 0f3F000000;
|
|
mul.ftz.f32 %f3946, %f2822, 0f3F000000;
|
|
mul.ftz.f32 %f3945, %f2823, 0f3F000000;
|
|
and.b32 %r165, %r1280, 4;
|
|
setp.eq.s32 %p284, %r165, 0;
|
|
@%p284 bra $L__BB11_270;
|
|
bra.uni $L__BB11_269;
|
|
|
|
$L__BB11_270:
|
|
mov.b32 %f3704, %r1276;
|
|
mul.ftz.f32 %f2882, %f821, %f3704;
|
|
mul.ftz.f32 %f2883, %f821, %f2882;
|
|
fma.rn.ftz.f32 %f2884, %f825, %f823, 0f40000000;
|
|
add.ftz.f32 %f2885, %f824, 0fC0A00000;
|
|
fma.rn.ftz.f32 %f2886, %f822, %f2885, 0f40000000;
|
|
mul.ftz.f32 %f2887, %f2882, %f3704;
|
|
mul.ftz.f32 %f2888, %f3920, %f2883;
|
|
mul.ftz.f32 %f2889, %f3921, %f2883;
|
|
mul.ftz.f32 %f2890, %f3922, %f2883;
|
|
mul.ftz.f32 %f2891, %f3923, %f2884;
|
|
mul.ftz.f32 %f2892, %f3924, %f2884;
|
|
mul.ftz.f32 %f2893, %f3925, %f2884;
|
|
sub.ftz.f32 %f2894, %f2891, %f2888;
|
|
sub.ftz.f32 %f2895, %f2892, %f2889;
|
|
sub.ftz.f32 %f2896, %f2893, %f2890;
|
|
fma.rn.ftz.f32 %f2897, %f3926, %f2886, %f2894;
|
|
fma.rn.ftz.f32 %f2898, %f3927, %f2886, %f2895;
|
|
fma.rn.ftz.f32 %f2899, %f3928, %f2886, %f2896;
|
|
mul.ftz.f32 %f2900, %f3929, %f2887;
|
|
mul.ftz.f32 %f2901, %f3930, %f2887;
|
|
mul.ftz.f32 %f2902, %f3931, %f2887;
|
|
sub.ftz.f32 %f2903, %f2897, %f2900;
|
|
sub.ftz.f32 %f2904, %f2898, %f2901;
|
|
sub.ftz.f32 %f2905, %f2899, %f2902;
|
|
mul.ftz.f32 %f832, %f2903, 0f3F000000;
|
|
mul.ftz.f32 %f833, %f2904, 0f3F000000;
|
|
mul.ftz.f32 %f834, %f2905, 0f3F000000;
|
|
setp.eq.ftz.f32 %p285, %f3949, %f832;
|
|
setp.eq.ftz.f32 %p286, %f3950, %f833;
|
|
and.pred %p287, %p285, %p286;
|
|
setp.eq.ftz.f32 %p288, %f3951, %f834;
|
|
and.pred %p289, %p288, %p287;
|
|
@%p289 bra $L__BB11_272;
|
|
bra.uni $L__BB11_271;
|
|
|
|
$L__BB11_272:
|
|
mov.f32 %f3948, %f639;
|
|
mov.f32 %f3947, %f641;
|
|
bra.uni $L__BB11_273;
|
|
|
|
$L__BB11_269:
|
|
mov.b32 %f3705, %r1277;
|
|
mul.ftz.f32 %f2824, %f3946, %f3946;
|
|
fma.rn.ftz.f32 %f2825, %f3944, %f3944, %f2824;
|
|
fma.rn.ftz.f32 %f2826, %f3945, %f3945, %f2825;
|
|
sqrt.approx.ftz.f32 %f2827, %f2826;
|
|
div.approx.ftz.f32 %f2828, %f3944, %f2827;
|
|
div.approx.ftz.f32 %f2829, %f3946, %f2827;
|
|
div.approx.ftz.f32 %f2830, %f3945, %f2827;
|
|
mul.ftz.f32 %f2831, %f2829, %f3892;
|
|
mul.ftz.f32 %f2832, %f2830, %f3891;
|
|
sub.ftz.f32 %f2833, %f2832, %f2831;
|
|
mul.ftz.f32 %f2834, %f2830, %f3890;
|
|
mul.ftz.f32 %f2835, %f2828, %f3892;
|
|
sub.ftz.f32 %f2836, %f2835, %f2834;
|
|
mul.ftz.f32 %f2837, %f2828, %f3891;
|
|
mul.ftz.f32 %f2838, %f2829, %f3890;
|
|
sub.ftz.f32 %f2839, %f2838, %f2837;
|
|
mul.ftz.f32 %f2840, %f2836, %f2836;
|
|
fma.rn.ftz.f32 %f2841, %f2833, %f2833, %f2840;
|
|
fma.rn.ftz.f32 %f2842, %f2839, %f2839, %f2841;
|
|
sqrt.approx.ftz.f32 %f2843, %f2842;
|
|
div.approx.ftz.f32 %f2844, %f2833, %f2843;
|
|
div.approx.ftz.f32 %f2845, %f2836, %f2843;
|
|
div.approx.ftz.f32 %f2846, %f2839, %f2843;
|
|
mul.ftz.f32 %f2847, %f3705, %f3705;
|
|
sub.ftz.f32 %f2849, %f2798, %f2847;
|
|
mov.f32 %f2850, 0f00000000;
|
|
max.ftz.f32 %f2851, %f2849, %f2850;
|
|
sqrt.approx.ftz.f32 %f2852, %f2851;
|
|
mul.ftz.f32 %f2853, %f2844, %f3705;
|
|
mul.ftz.f32 %f2854, %f2845, %f3705;
|
|
mul.ftz.f32 %f2855, %f2846, %f3705;
|
|
mul.ftz.f32 %f2856, %f2829, %f2846;
|
|
mul.ftz.f32 %f2857, %f2830, %f2845;
|
|
sub.ftz.f32 %f2858, %f2856, %f2857;
|
|
mul.ftz.f32 %f2859, %f2830, %f2844;
|
|
mul.ftz.f32 %f2860, %f2828, %f2846;
|
|
sub.ftz.f32 %f2861, %f2859, %f2860;
|
|
mul.ftz.f32 %f2862, %f2828, %f2845;
|
|
mul.ftz.f32 %f2863, %f2829, %f2844;
|
|
sub.ftz.f32 %f2864, %f2862, %f2863;
|
|
mul.ftz.f32 %f2865, %f2861, %f2861;
|
|
fma.rn.ftz.f32 %f2866, %f2858, %f2858, %f2865;
|
|
fma.rn.ftz.f32 %f2867, %f2864, %f2864, %f2866;
|
|
sqrt.approx.ftz.f32 %f2868, %f2867;
|
|
div.approx.ftz.f32 %f2869, %f2858, %f2868;
|
|
div.approx.ftz.f32 %f2870, %f2861, %f2868;
|
|
div.approx.ftz.f32 %f2871, %f2864, %f2868;
|
|
mul.ftz.f32 %f2872, %f2852, %f2869;
|
|
mul.ftz.f32 %f2873, %f2852, %f2870;
|
|
mul.ftz.f32 %f2874, %f2852, %f2871;
|
|
sub.ftz.f32 %f2875, %f2853, %f2872;
|
|
sub.ftz.f32 %f2876, %f2854, %f2873;
|
|
sub.ftz.f32 %f2877, %f2855, %f2874;
|
|
mul.ftz.f32 %f2878, %f2876, %f2876;
|
|
fma.rn.ftz.f32 %f2879, %f2875, %f2875, %f2878;
|
|
fma.rn.ftz.f32 %f2880, %f2877, %f2877, %f2879;
|
|
sqrt.approx.ftz.f32 %f2881, %f2880;
|
|
div.approx.ftz.f32 %f3948, %f2875, %f2881;
|
|
div.approx.ftz.f32 %f3947, %f2876, %f2881;
|
|
div.approx.ftz.f32 %f3935, %f2877, %f2881;
|
|
st.local.f32 [%rd79], %f3948;
|
|
st.local.f32 [%rd100], %f3947;
|
|
st.local.f32 [%rd101], %f3935;
|
|
bra.uni $L__BB11_274;
|
|
|
|
$L__BB11_271:
|
|
sub.ftz.f32 %f2906, %f3949, %f832;
|
|
sub.ftz.f32 %f2907, %f3950, %f833;
|
|
mul.ftz.f32 %f2908, %f2907, %f2907;
|
|
fma.rn.ftz.f32 %f2909, %f2906, %f2906, %f2908;
|
|
sub.ftz.f32 %f2910, %f3951, %f834;
|
|
fma.rn.ftz.f32 %f2911, %f2910, %f2910, %f2909;
|
|
sqrt.approx.ftz.f32 %f2912, %f2911;
|
|
div.approx.ftz.f32 %f3935, %f2910, %f2912;
|
|
div.approx.ftz.f32 %f3947, %f2907, %f2912;
|
|
div.approx.ftz.f32 %f3948, %f2906, %f2912;
|
|
|
|
$L__BB11_273:
|
|
mov.u32 %r1482, 0;
|
|
st.local.f32 [%rd79], %f3948;
|
|
st.local.f32 [%rd100], %f3947;
|
|
st.local.f32 [%rd101], %f3935;
|
|
st.local.u32 [%rd105], %r1482;
|
|
|
|
$L__BB11_274:
|
|
and.b32 %r1479, %r1520, 4;
|
|
setp.ne.s32 %p368, %r1479, 0;
|
|
st.local.f32 [%rd82], %f3944;
|
|
st.local.f32 [%rd83], %f3946;
|
|
st.local.f32 [%rd84], %f3945;
|
|
@%p368 bra $L__BB11_284;
|
|
|
|
@%p264 bra $L__BB11_277;
|
|
bra.uni $L__BB11_276;
|
|
|
|
$L__BB11_277:
|
|
ld.global.v4.f32 {%f2946, %f2947, %f2948, %f2949}, [%rd107];
|
|
ld.global.v4.f32 {%f2954, %f2955, %f2956, %f2957}, [%rd107+16];
|
|
ld.global.v4.f32 {%f2962, %f2963, %f2964, %f2965}, [%rd107+32];
|
|
mul.ftz.f32 %f2970, %f3950, %f2947;
|
|
fma.rn.ftz.f32 %f2971, %f3949, %f2946, %f2970;
|
|
fma.rn.ftz.f32 %f2972, %f3951, %f2948, %f2971;
|
|
add.ftz.f32 %f3938, %f2949, %f2972;
|
|
mul.ftz.f32 %f2973, %f3950, %f2955;
|
|
fma.rn.ftz.f32 %f2974, %f3949, %f2954, %f2973;
|
|
fma.rn.ftz.f32 %f2975, %f3951, %f2956, %f2974;
|
|
add.ftz.f32 %f3939, %f2957, %f2975;
|
|
mul.ftz.f32 %f2976, %f3950, %f2963;
|
|
fma.rn.ftz.f32 %f2977, %f3949, %f2962, %f2976;
|
|
fma.rn.ftz.f32 %f2978, %f3951, %f2964, %f2977;
|
|
add.ftz.f32 %f3951, %f2965, %f2978;
|
|
bra.uni $L__BB11_278;
|
|
|
|
$L__BB11_276:
|
|
cvta.to.local.u64 %rd399, %rd65;
|
|
ld.local.v4.f32 {%f2913, %f2914, %f2915, %f2916}, [%rd399];
|
|
mul.ftz.f32 %f2921, %f3950, %f2914;
|
|
fma.rn.ftz.f32 %f2922, %f3949, %f2913, %f2921;
|
|
fma.rn.ftz.f32 %f2923, %f3951, %f2915, %f2922;
|
|
add.ftz.f32 %f3938, %f2916, %f2923;
|
|
cvta.to.local.u64 %rd400, %rd75;
|
|
ld.local.v4.f32 {%f2924, %f2925, %f2926, %f2927}, [%rd400];
|
|
mul.ftz.f32 %f2932, %f3950, %f2925;
|
|
fma.rn.ftz.f32 %f2933, %f3949, %f2924, %f2932;
|
|
fma.rn.ftz.f32 %f2934, %f3951, %f2926, %f2933;
|
|
add.ftz.f32 %f3939, %f2927, %f2934;
|
|
cvta.to.local.u64 %rd401, %rd74;
|
|
ld.local.v4.f32 {%f2935, %f2936, %f2937, %f2938}, [%rd401];
|
|
mul.ftz.f32 %f2943, %f3950, %f2936;
|
|
fma.rn.ftz.f32 %f2944, %f3949, %f2935, %f2943;
|
|
fma.rn.ftz.f32 %f2945, %f3951, %f2937, %f2944;
|
|
add.ftz.f32 %f3951, %f2938, %f2945;
|
|
|
|
$L__BB11_278:
|
|
@%p264 bra $L__BB11_280;
|
|
bra.uni $L__BB11_279;
|
|
|
|
$L__BB11_280:
|
|
ld.global.v4.f32 {%f3013, %f3014, %f3015, %f3016}, [%rd107+48];
|
|
ld.global.v4.f32 {%f3020, %f3021, %f3022, %f3023}, [%rd107+64];
|
|
ld.global.v4.f32 {%f3027, %f3028, %f3029, %f3030}, [%rd107+80];
|
|
mul.ftz.f32 %f3034, %f3020, %f3947;
|
|
fma.rn.ftz.f32 %f3035, %f3013, %f3948, %f3034;
|
|
fma.rn.ftz.f32 %f3036, %f3027, %f3935, %f3035;
|
|
mul.ftz.f32 %f3037, %f3021, %f3947;
|
|
fma.rn.ftz.f32 %f3038, %f3014, %f3948, %f3037;
|
|
fma.rn.ftz.f32 %f3039, %f3028, %f3935, %f3038;
|
|
mul.ftz.f32 %f3040, %f3022, %f3947;
|
|
fma.rn.ftz.f32 %f3041, %f3015, %f3948, %f3040;
|
|
fma.rn.ftz.f32 %f3042, %f3029, %f3935, %f3041;
|
|
mul.ftz.f32 %f3043, %f3039, %f3039;
|
|
fma.rn.ftz.f32 %f3044, %f3036, %f3036, %f3043;
|
|
fma.rn.ftz.f32 %f3045, %f3042, %f3042, %f3044;
|
|
sqrt.approx.ftz.f32 %f3046, %f3045;
|
|
div.approx.ftz.f32 %f3948, %f3036, %f3046;
|
|
div.approx.ftz.f32 %f3947, %f3039, %f3046;
|
|
div.approx.ftz.f32 %f3941, %f3042, %f3046;
|
|
bra.uni $L__BB11_281;
|
|
|
|
$L__BB11_279:
|
|
cvta.to.local.u64 %rd402, %rd76;
|
|
ld.local.v4.f32 {%f2979, %f2980, %f2981, %f2982}, [%rd402];
|
|
cvta.to.local.u64 %rd403, %rd90;
|
|
ld.local.v4.f32 {%f2986, %f2987, %f2988, %f2989}, [%rd403];
|
|
cvta.to.local.u64 %rd404, %rd89;
|
|
ld.local.v4.f32 {%f2993, %f2994, %f2995, %f2996}, [%rd404];
|
|
mul.ftz.f32 %f3000, %f3947, %f2986;
|
|
fma.rn.ftz.f32 %f3001, %f3948, %f2979, %f3000;
|
|
fma.rn.ftz.f32 %f3002, %f3935, %f2993, %f3001;
|
|
mul.ftz.f32 %f3003, %f3947, %f2987;
|
|
fma.rn.ftz.f32 %f3004, %f3948, %f2980, %f3003;
|
|
fma.rn.ftz.f32 %f3005, %f3935, %f2994, %f3004;
|
|
mul.ftz.f32 %f3006, %f3947, %f2988;
|
|
fma.rn.ftz.f32 %f3007, %f3948, %f2981, %f3006;
|
|
fma.rn.ftz.f32 %f3008, %f3935, %f2995, %f3007;
|
|
mul.ftz.f32 %f3009, %f3005, %f3005;
|
|
fma.rn.ftz.f32 %f3010, %f3002, %f3002, %f3009;
|
|
fma.rn.ftz.f32 %f3011, %f3008, %f3008, %f3010;
|
|
sqrt.approx.ftz.f32 %f3012, %f3011;
|
|
div.approx.ftz.f32 %f3948, %f3002, %f3012;
|
|
div.approx.ftz.f32 %f3947, %f3005, %f3012;
|
|
div.approx.ftz.f32 %f3941, %f3008, %f3012;
|
|
|
|
$L__BB11_281:
|
|
st.local.f32 [%rd79], %f3948;
|
|
st.local.f32 [%rd100], %f3947;
|
|
st.local.f32 [%rd101], %f3941;
|
|
@%p264 bra $L__BB11_283;
|
|
bra.uni $L__BB11_282;
|
|
|
|
$L__BB11_283:
|
|
ld.global.v4.f32 {%f3074, %f3075, %f3076, %f3077}, [%rd107];
|
|
ld.global.v4.f32 {%f3081, %f3082, %f3083, %f3084}, [%rd107+16];
|
|
ld.global.v4.f32 {%f3088, %f3089, %f3090, %f3091}, [%rd107+32];
|
|
mul.ftz.f32 %f3095, %f3075, %f3946;
|
|
fma.rn.ftz.f32 %f3096, %f3944, %f3074, %f3095;
|
|
fma.rn.ftz.f32 %f871, %f3076, %f3945, %f3096;
|
|
mul.ftz.f32 %f3097, %f3082, %f3946;
|
|
fma.rn.ftz.f32 %f3098, %f3944, %f3081, %f3097;
|
|
fma.rn.ftz.f32 %f872, %f3083, %f3945, %f3098;
|
|
mul.ftz.f32 %f3099, %f3089, %f3946;
|
|
fma.rn.ftz.f32 %f3100, %f3944, %f3088, %f3099;
|
|
fma.rn.ftz.f32 %f3945, %f3090, %f3945, %f3100;
|
|
st.local.f32 [%rd82], %f871;
|
|
st.local.f32 [%rd83], %f872;
|
|
st.local.f32 [%rd84], %f3945;
|
|
mov.f32 %f3944, %f871;
|
|
mov.f32 %f3946, %f872;
|
|
mov.f32 %f3949, %f3938;
|
|
mov.f32 %f3950, %f3939;
|
|
bra.uni $L__BB11_284;
|
|
|
|
$L__BB11_282:
|
|
cvta.to.local.u64 %rd405, %rd65;
|
|
ld.local.v4.f32 {%f3047, %f3048, %f3049, %f3050}, [%rd405];
|
|
mul.ftz.f32 %f3054, %f3946, %f3048;
|
|
fma.rn.ftz.f32 %f3055, %f3944, %f3047, %f3054;
|
|
fma.rn.ftz.f32 %f868, %f3945, %f3049, %f3055;
|
|
cvta.to.local.u64 %rd406, %rd75;
|
|
ld.local.v4.f32 {%f3056, %f3057, %f3058, %f3059}, [%rd406];
|
|
mul.ftz.f32 %f3063, %f3946, %f3057;
|
|
fma.rn.ftz.f32 %f3064, %f3944, %f3056, %f3063;
|
|
fma.rn.ftz.f32 %f869, %f3945, %f3058, %f3064;
|
|
cvta.to.local.u64 %rd407, %rd74;
|
|
ld.local.v4.f32 {%f3065, %f3066, %f3067, %f3068}, [%rd407];
|
|
mul.ftz.f32 %f3072, %f3946, %f3066;
|
|
fma.rn.ftz.f32 %f3073, %f3944, %f3065, %f3072;
|
|
fma.rn.ftz.f32 %f3945, %f3945, %f3067, %f3073;
|
|
st.local.f32 [%rd82], %f868;
|
|
st.local.f32 [%rd83], %f869;
|
|
st.local.f32 [%rd84], %f3945;
|
|
mov.f32 %f3944, %f868;
|
|
mov.f32 %f3946, %f869;
|
|
mov.f32 %f3949, %f3938;
|
|
mov.f32 %f3950, %f3939;
|
|
|
|
$L__BB11_284:
|
|
mul.wide.s32 %rd678, %r1278, 16;
|
|
add.s64 %rd677, %rd88, %rd678;
|
|
add.s64 %rd676, %rd677, 4;
|
|
neg.ftz.f32 %f3695, %f641;
|
|
neg.ftz.f32 %f3694, %f639;
|
|
cvta.to.local.u64 %rd409, %rd145;
|
|
st.local.v2.f32 [%rd409], {%f3949, %f3950};
|
|
st.local.f32 [%rd78], %f3951;
|
|
selp.f32 %f3988, %f3948, %f3694, %p284;
|
|
selp.f32 %f3989, %f3947, %f3695, %p284;
|
|
selp.b64 %rd410, %rd81, %rd63, %p284;
|
|
cvta.to.local.u64 %rd411, %rd410;
|
|
ld.local.f32 %f3990, [%rd411];
|
|
cvta.to.local.u64 %rd412, %rd68;
|
|
st.local.v2.f32 [%rd412], {%f3988, %f3989};
|
|
st.local.f32 [%rd71], %f3990;
|
|
mul.ftz.f32 %f3101, %f3946, %f3990;
|
|
mul.ftz.f32 %f3102, %f3945, %f3989;
|
|
mul.ftz.f32 %f3103, %f3945, %f3988;
|
|
mul.ftz.f32 %f3104, %f3944, %f3990;
|
|
mul.ftz.f32 %f3105, %f3944, %f3989;
|
|
mul.ftz.f32 %f3106, %f3946, %f3988;
|
|
sub.ftz.f32 %f3107, %f3105, %f3106;
|
|
sub.ftz.f32 %f3108, %f3103, %f3104;
|
|
sub.ftz.f32 %f3109, %f3101, %f3102;
|
|
cvta.to.local.u64 %rd413, %rd73;
|
|
st.local.v2.f32 [%rd413], {%f3109, %f3108};
|
|
st.local.f32 [%rd85], %f3107;
|
|
ld.global.u32 %r1404, [%rd676+-4];
|
|
st.local.u32 [%rd70], %r1404;
|
|
bra.uni $L__BB11_335;
|
|
|
|
$L__BB11_307:
|
|
{ // callseq 0, 0
|
|
.reg .b32 temp_param_reg;
|
|
.param .b64 param0;
|
|
st.param.b64 [param0+0], %rd145;
|
|
call.uni
|
|
_Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib,
|
|
(
|
|
param0
|
|
);
|
|
} // callseq 0
|
|
ld.local.u32 %r1520, [%rd56];
|
|
bra.uni $L__BB11_321;
|
|
|
|
$L__BB11_287:
|
|
ld.global.u32 %r166, [%rd107+148];
|
|
ld.global.u32 %r167, [%rd107+152];
|
|
shl.b32 %r1408, %r167, 1;
|
|
cvt.rn.f32.s32 %f3113, %r1408;
|
|
mul.ftz.f32 %f890, %f582, %f3113;
|
|
cvt.rzi.ftz.s32.f32 %r1409, %f890;
|
|
add.s32 %r1410, %r1408, -1;
|
|
min.s32 %r168, %r1409, %r1410;
|
|
ld.global.u32 %r1516, [%rd107+164];
|
|
mul.wide.u32 %rd418, %r1516, 16;
|
|
add.s64 %rd419, %rd94, %rd418;
|
|
ld.global.u16 %rs112, [%rd419+12];
|
|
ld.global.u32 %r1515, [%rd419+8];
|
|
ld.global.u64 %rd709, [%rd419];
|
|
setp.eq.s64 %p297, %rd709, 11;
|
|
@%p297 bra $L__BB11_292;
|
|
|
|
$L__BB11_288:
|
|
setp.eq.s64 %p298, %rd709, 0;
|
|
@%p298 bra $L__BB11_290;
|
|
|
|
add.s32 %r1516, %r1516, 2;
|
|
bra.uni $L__BB11_291;
|
|
|
|
$L__BB11_290:
|
|
setp.eq.s16 %p299, %rs112, 0;
|
|
mov.u32 %r1519, -1;
|
|
mov.u32 %r1516, %r1515;
|
|
@%p299 bra $L__BB11_293;
|
|
|
|
$L__BB11_291:
|
|
mul.wide.u32 %rd420, %r1516, 16;
|
|
add.s64 %rd421, %rd94, %rd420;
|
|
ld.global.u16 %rs112, [%rd421+12];
|
|
ld.global.u32 %r1515, [%rd421+8];
|
|
ld.global.u64 %rd709, [%rd421];
|
|
setp.ne.s64 %p300, %rd709, 11;
|
|
@%p300 bra $L__BB11_288;
|
|
|
|
$L__BB11_292:
|
|
setp.eq.s16 %p301, %rs112, 0;
|
|
selp.b32 %r1519, -1, %r1515, %p301;
|
|
|
|
$L__BB11_293:
|
|
setp.eq.s32 %p302, %r168, %r167;
|
|
@%p302 bra $L__BB11_295;
|
|
bra.uni $L__BB11_294;
|
|
|
|
$L__BB11_295:
|
|
shl.b64 %rd423, %rd116, 4;
|
|
add.s64 %rd710, %rd93, %rd423;
|
|
bra.uni $L__BB11_296;
|
|
|
|
$L__BB11_309:
|
|
sub.ftz.f32 %f3218, %f933, %f934;
|
|
sub.ftz.f32 %f3219, %f935, %f936;
|
|
mul.ftz.f32 %f3220, %f3219, %f3219;
|
|
fma.rn.ftz.f32 %f3221, %f3218, %f3218, %f3220;
|
|
sub.ftz.f32 %f3222, %f937, %f938;
|
|
fma.rn.ftz.f32 %f3223, %f3222, %f3222, %f3221;
|
|
sqrt.approx.ftz.f32 %f3224, %f3223;
|
|
div.approx.ftz.f32 %f3960, %f3222, %f3224;
|
|
div.approx.ftz.f32 %f3959, %f3219, %f3224;
|
|
div.approx.ftz.f32 %f3958, %f3218, %f3224;
|
|
|
|
$L__BB11_311:
|
|
cvt.s64.s32 %rd683, %r1278;
|
|
mov.b32 %f3706, %r1277;
|
|
mov.b32 %f3702, %r1276;
|
|
shl.b64 %rd442, %rd683, 2;
|
|
add.s64 %rd443, %rd96, %rd442;
|
|
ld.global.u32 %r181, [%rd443];
|
|
st.local.u32 [%rd70], %r181;
|
|
fma.rn.ftz.f32 %f3232, %f930, %f3702, %f924;
|
|
fma.rn.ftz.f32 %f3233, %f931, %f3702, %f926;
|
|
fma.rn.ftz.f32 %f3234, %f932, %f3702, %f928;
|
|
fma.rn.ftz.f32 %f3973, %f925, %f3706, %f3232;
|
|
fma.rn.ftz.f32 %f3974, %f927, %f3706, %f3233;
|
|
fma.rn.ftz.f32 %f3975, %f929, %f3706, %f3234;
|
|
and.b32 %r1426, %r1520, 4;
|
|
setp.ne.s32 %p310, %r1426, 0;
|
|
@%p310 bra $L__BB11_316;
|
|
|
|
@%p264 bra $L__BB11_314;
|
|
bra.uni $L__BB11_313;
|
|
|
|
$L__BB11_314:
|
|
ld.global.v4.f32 {%f3961, %f3962, %f3963, %f3964}, [%rd107];
|
|
ld.global.v4.f32 {%f3965, %f3966, %f3967, %f3968}, [%rd107+16];
|
|
ld.global.v4.f32 {%f3969, %f3970, %f3971, %f3972}, [%rd107+32];
|
|
bra.uni $L__BB11_315;
|
|
|
|
$L__BB11_294:
|
|
setp.gt.s32 %p303, %r168, %r167;
|
|
selp.b32 %r1412, -1, 0, %p303;
|
|
add.s32 %r1413, %r168, %r1412;
|
|
mad.lo.s32 %r1414, %r1413, %r166, %r1278;
|
|
add.s32 %r1415, %r1414, %r1519;
|
|
mul.wide.s32 %rd422, %r1415, 16;
|
|
add.s64 %rd710, %rd103, %rd422;
|
|
|
|
$L__BB11_296:
|
|
ld.global.v4.f32 {%f3114, %f3115, %f3116, %f3117}, [%rd710];
|
|
add.s32 %r179, %r168, 1;
|
|
setp.eq.s32 %p304, %r179, %r167;
|
|
@%p304 bra $L__BB11_298;
|
|
bra.uni $L__BB11_297;
|
|
|
|
$L__BB11_298:
|
|
shl.b64 %rd425, %rd116, 4;
|
|
add.s64 %rd711, %rd93, %rd425;
|
|
bra.uni $L__BB11_299;
|
|
|
|
$L__BB11_297:
|
|
setp.ge.s32 %p305, %r168, %r167;
|
|
selp.b32 %r1416, -1, 0, %p305;
|
|
add.s32 %r1417, %r179, %r1416;
|
|
mad.lo.s32 %r1418, %r1417, %r166, %r1278;
|
|
add.s32 %r1419, %r1418, %r1519;
|
|
mul.wide.s32 %rd424, %r1419, 16;
|
|
add.s64 %rd711, %rd103, %rd424;
|
|
|
|
$L__BB11_299:
|
|
cvt.rn.f32.s32 %f3118, %r168;
|
|
sub.ftz.f32 %f3119, %f890, %f3118;
|
|
ld.global.v4.f32 {%f3120, %f3121, %f3122, %f3123}, [%rd711];
|
|
mov.f32 %f3128, 0f3F800000;
|
|
sub.ftz.f32 %f3129, %f3128, %f3119;
|
|
mul.ftz.f32 %f3130, %f3119, %f3120;
|
|
mul.ftz.f32 %f3131, %f3119, %f3121;
|
|
mul.ftz.f32 %f3132, %f3119, %f3122;
|
|
fma.rn.ftz.f32 %f3957, %f3129, %f3116, %f3132;
|
|
fma.rn.ftz.f32 %f3956, %f3129, %f3115, %f3131;
|
|
fma.rn.ftz.f32 %f3955, %f3129, %f3114, %f3130;
|
|
|
|
$L__BB11_301:
|
|
and.b32 %r1420, %r1520, 4;
|
|
setp.ne.s32 %p306, %r1420, 0;
|
|
@%p306 bra $L__BB11_305;
|
|
|
|
@%p264 bra $L__BB11_304;
|
|
bra.uni $L__BB11_303;
|
|
|
|
$L__BB11_304:
|
|
ld.global.v4.f32 {%f3171, %f3172, %f3173, %f3174}, [%rd107];
|
|
ld.global.v4.f32 {%f3179, %f3180, %f3181, %f3182}, [%rd107+16];
|
|
ld.global.v4.f32 {%f3187, %f3188, %f3189, %f3190}, [%rd107+32];
|
|
mul.ftz.f32 %f3195, %f3956, %f3172;
|
|
fma.rn.ftz.f32 %f3196, %f3955, %f3171, %f3195;
|
|
fma.rn.ftz.f32 %f3197, %f3957, %f3173, %f3196;
|
|
add.ftz.f32 %f913, %f3174, %f3197;
|
|
mul.ftz.f32 %f3198, %f3956, %f3180;
|
|
fma.rn.ftz.f32 %f3199, %f3955, %f3179, %f3198;
|
|
fma.rn.ftz.f32 %f3200, %f3957, %f3181, %f3199;
|
|
add.ftz.f32 %f914, %f3182, %f3200;
|
|
mul.ftz.f32 %f3201, %f3956, %f3188;
|
|
fma.rn.ftz.f32 %f3202, %f3955, %f3187, %f3201;
|
|
fma.rn.ftz.f32 %f3203, %f3957, %f3189, %f3202;
|
|
add.ftz.f32 %f3957, %f3190, %f3203;
|
|
mov.f32 %f3955, %f913;
|
|
mov.f32 %f3956, %f914;
|
|
bra.uni $L__BB11_305;
|
|
|
|
$L__BB11_303:
|
|
cvta.to.local.u64 %rd428, %rd65;
|
|
ld.local.v4.f32 {%f3138, %f3139, %f3140, %f3141}, [%rd428];
|
|
mul.ftz.f32 %f3146, %f3956, %f3139;
|
|
fma.rn.ftz.f32 %f3147, %f3955, %f3138, %f3146;
|
|
fma.rn.ftz.f32 %f3148, %f3957, %f3140, %f3147;
|
|
add.ftz.f32 %f910, %f3141, %f3148;
|
|
cvta.to.local.u64 %rd429, %rd75;
|
|
ld.local.v4.f32 {%f3149, %f3150, %f3151, %f3152}, [%rd429];
|
|
mul.ftz.f32 %f3157, %f3956, %f3150;
|
|
fma.rn.ftz.f32 %f3158, %f3955, %f3149, %f3157;
|
|
fma.rn.ftz.f32 %f3159, %f3957, %f3151, %f3158;
|
|
add.ftz.f32 %f911, %f3152, %f3159;
|
|
cvta.to.local.u64 %rd430, %rd74;
|
|
ld.local.v4.f32 {%f3160, %f3161, %f3162, %f3163}, [%rd430];
|
|
mul.ftz.f32 %f3168, %f3956, %f3161;
|
|
fma.rn.ftz.f32 %f3169, %f3955, %f3160, %f3168;
|
|
fma.rn.ftz.f32 %f3170, %f3957, %f3162, %f3169;
|
|
add.ftz.f32 %f3957, %f3163, %f3170;
|
|
mov.f32 %f3955, %f910;
|
|
mov.f32 %f3956, %f911;
|
|
|
|
$L__BB11_305:
|
|
sub.ftz.f32 %f3204, %f887, %f3955;
|
|
sub.ftz.f32 %f3205, %f888, %f3956;
|
|
mul.ftz.f32 %f3206, %f3205, %f3205;
|
|
fma.rn.ftz.f32 %f3207, %f3204, %f3204, %f3206;
|
|
sub.ftz.f32 %f3208, %f889, %f3957;
|
|
fma.rn.ftz.f32 %f3209, %f3208, %f3208, %f3207;
|
|
sqrt.approx.ftz.f32 %f3210, %f3209;
|
|
div.approx.ftz.f32 %f3988, %f3204, %f3210;
|
|
div.approx.ftz.f32 %f3989, %f3205, %f3210;
|
|
div.approx.ftz.f32 %f3990, %f3208, %f3210;
|
|
st.local.f32 [%rd71], %f3990;
|
|
st.local.f32 [%rd79], %f3988;
|
|
st.local.v4.f32 [%rd100], {%f3989, %f3990, %f3988, %f3989};
|
|
mov.u32 %r1421, 0;
|
|
st.local.u32 [%rd82], %r1421;
|
|
st.local.u32 [%rd83], %r1421;
|
|
st.local.u32 [%rd84], %r1421;
|
|
cvta.to.local.u64 %rd432, %rd73;
|
|
mov.f32 %f3211, 0f00000000;
|
|
st.local.v2.f32 [%rd432], {%f3211, %f3211};
|
|
st.local.u32 [%rd85], %r1421;
|
|
bra.uni $L__BB11_335;
|
|
|
|
$L__BB11_313:
|
|
cvta.to.local.u64 %rd444, %rd65;
|
|
ld.local.v4.f32 {%f3961, %f3962, %f3963, %f3964}, [%rd444];
|
|
cvta.to.local.u64 %rd445, %rd75;
|
|
ld.local.v4.f32 {%f3965, %f3966, %f3967, %f3968}, [%rd445];
|
|
cvta.to.local.u64 %rd446, %rd74;
|
|
ld.local.v4.f32 {%f3969, %f3970, %f3971, %f3972}, [%rd446];
|
|
|
|
$L__BB11_315:
|
|
mul.ftz.f32 %f3259, %f3974, %f3962;
|
|
fma.rn.ftz.f32 %f3260, %f3973, %f3961, %f3259;
|
|
fma.rn.ftz.f32 %f3261, %f3975, %f3963, %f3260;
|
|
mul.ftz.f32 %f3262, %f3974, %f3966;
|
|
fma.rn.ftz.f32 %f3263, %f3973, %f3965, %f3262;
|
|
fma.rn.ftz.f32 %f3264, %f3975, %f3967, %f3263;
|
|
mul.ftz.f32 %f3265, %f3974, %f3970;
|
|
fma.rn.ftz.f32 %f3266, %f3973, %f3969, %f3265;
|
|
fma.rn.ftz.f32 %f3267, %f3975, %f3971, %f3266;
|
|
add.ftz.f32 %f3973, %f3964, %f3261;
|
|
add.ftz.f32 %f3974, %f3968, %f3264;
|
|
add.ftz.f32 %f3975, %f3972, %f3267;
|
|
|
|
$L__BB11_316:
|
|
st.local.f32 [%rd71], %f3960;
|
|
cvta.to.local.u64 %rd448, %rd145;
|
|
st.local.v4.f32 [%rd448], {%f3973, %f3974, %f3975, %f3958};
|
|
st.local.v4.f32 [%rd100], {%f3959, %f3960, %f3958, %f3959};
|
|
setp.gt.s32 %p312, %r181, -1;
|
|
@%p312 bra $L__BB11_320;
|
|
|
|
mul.wide.s32 %rd686, %r1278, 16;
|
|
add.s64 %rd685, %rd91, %rd686;
|
|
add.s64 %rd684, %rd685, 12;
|
|
mov.b32 %f3707, %r1277;
|
|
mov.b32 %f3703, %r1276;
|
|
ld.global.v4.u32 {%r1427, %r1428, %r1429, %r1430}, [%rd684+-12];
|
|
mul.wide.u32 %rd450, %r1427, 12;
|
|
add.s64 %rd451, %rd106, %rd450;
|
|
mul.wide.u32 %rd452, %r1428, 12;
|
|
add.s64 %rd453, %rd106, %rd452;
|
|
mul.wide.u32 %rd454, %r1429, 12;
|
|
add.s64 %rd455, %rd106, %rd454;
|
|
mov.f32 %f3268, 0f3F800000;
|
|
sub.ftz.f32 %f3269, %f3268, %f3703;
|
|
sub.ftz.f32 %f3270, %f3269, %f3707;
|
|
ld.global.f32 %f3271, [%rd451];
|
|
ld.global.f32 %f3272, [%rd451+4];
|
|
ld.global.f32 %f3273, [%rd451+8];
|
|
ld.global.f32 %f3274, [%rd453];
|
|
mul.ftz.f32 %f3275, %f3274, %f3703;
|
|
ld.global.f32 %f3276, [%rd453+4];
|
|
mul.ftz.f32 %f3277, %f3276, %f3703;
|
|
ld.global.f32 %f3278, [%rd453+8];
|
|
mul.ftz.f32 %f3279, %f3278, %f3703;
|
|
fma.rn.ftz.f32 %f3280, %f3270, %f3271, %f3275;
|
|
fma.rn.ftz.f32 %f3281, %f3270, %f3272, %f3277;
|
|
fma.rn.ftz.f32 %f3282, %f3270, %f3273, %f3279;
|
|
ld.global.f32 %f3283, [%rd455];
|
|
ld.global.f32 %f3284, [%rd455+4];
|
|
ld.global.f32 %f3285, [%rd455+8];
|
|
fma.rn.ftz.f32 %f3976, %f3283, %f3707, %f3280;
|
|
fma.rn.ftz.f32 %f3977, %f3284, %f3707, %f3281;
|
|
fma.rn.ftz.f32 %f3978, %f3285, %f3707, %f3282;
|
|
mul.ftz.f32 %f3286, %f3977, %f3977;
|
|
fma.rn.ftz.f32 %f3287, %f3976, %f3976, %f3286;
|
|
fma.rn.ftz.f32 %f3288, %f3978, %f3978, %f3287;
|
|
sqrt.approx.ftz.f32 %f999, %f3288;
|
|
setp.neu.ftz.f32 %p313, %f999, 0f00000000;
|
|
@%p313 bra $L__BB11_318;
|
|
bra.uni $L__BB11_319;
|
|
|
|
$L__BB11_318:
|
|
rcp.approx.ftz.f32 %f3289, %f999;
|
|
mul.ftz.f32 %f3978, %f3978, %f3289;
|
|
mul.ftz.f32 %f3977, %f3977, %f3289;
|
|
mul.ftz.f32 %f3976, %f3976, %f3289;
|
|
|
|
$L__BB11_319:
|
|
setp.eq.ftz.f32 %p314, %f3976, 0f00000000;
|
|
setp.eq.ftz.f32 %p315, %f3977, 0f00000000;
|
|
and.pred %p316, %p314, %p315;
|
|
setp.eq.ftz.f32 %p317, %f3978, 0f00000000;
|
|
and.pred %p318, %p317, %p316;
|
|
selp.f32 %f3290, %f3960, %f3978, %p318;
|
|
selp.f32 %f3291, %f3959, %f3977, %p318;
|
|
selp.f32 %f3292, %f3958, %f3976, %p318;
|
|
st.local.f32 [%rd79], %f3292;
|
|
st.local.f32 [%rd100], %f3291;
|
|
st.local.f32 [%rd101], %f3290;
|
|
|
|
$L__BB11_320:
|
|
st.local.f32 [%rd82], %f930;
|
|
st.local.f32 [%rd83], %f931;
|
|
st.local.f32 [%rd84], %f932;
|
|
cvta.to.local.u64 %rd456, %rd73;
|
|
st.local.v2.f32 [%rd456], {%f925, %f927};
|
|
st.local.f32 [%rd85], %f929;
|
|
|
|
$L__BB11_321:
|
|
and.b32 %r1434, %r1520, 4;
|
|
setp.eq.s32 %p319, %r1434, 0;
|
|
@%p319 bra $L__BB11_323;
|
|
bra.uni $L__BB11_322;
|
|
|
|
$L__BB11_323:
|
|
and.b32 %r183, %r1520, 2;
|
|
setp.eq.s32 %p320, %r183, 0;
|
|
@%p320 bra $L__BB11_325;
|
|
bra.uni $L__BB11_324;
|
|
|
|
$L__BB11_325:
|
|
ld.local.u32 %r1435, [%rd102];
|
|
mul.wide.s32 %rd462, %r1435, 208;
|
|
add.s64 %rd463, %rd66, %rd462;
|
|
ld.global.v4.f32 {%f3334, %f3335, %f3336, %f3337}, [%rd463+48];
|
|
ld.global.v4.f32 {%f3341, %f3342, %f3343, %f3344}, [%rd463+64];
|
|
ld.global.v4.f32 {%f3348, %f3349, %f3350, %f3351}, [%rd463+80];
|
|
ld.local.v2.f32 {%f3355, %f3356}, [%rd100];
|
|
ld.local.f32 %f3359, [%rd79];
|
|
mul.ftz.f32 %f3360, %f3341, %f3355;
|
|
fma.rn.ftz.f32 %f3361, %f3334, %f3359, %f3360;
|
|
fma.rn.ftz.f32 %f3362, %f3348, %f3356, %f3361;
|
|
mul.ftz.f32 %f3363, %f3342, %f3355;
|
|
fma.rn.ftz.f32 %f3364, %f3335, %f3359, %f3363;
|
|
fma.rn.ftz.f32 %f3365, %f3349, %f3356, %f3364;
|
|
mul.ftz.f32 %f3366, %f3343, %f3355;
|
|
fma.rn.ftz.f32 %f3367, %f3336, %f3359, %f3366;
|
|
fma.rn.ftz.f32 %f3368, %f3350, %f3356, %f3367;
|
|
mul.ftz.f32 %f3369, %f3365, %f3365;
|
|
fma.rn.ftz.f32 %f3370, %f3362, %f3362, %f3369;
|
|
fma.rn.ftz.f32 %f3371, %f3368, %f3368, %f3370;
|
|
sqrt.approx.ftz.f32 %f3372, %f3371;
|
|
div.approx.ftz.f32 %f3981, %f3362, %f3372;
|
|
div.approx.ftz.f32 %f3980, %f3365, %f3372;
|
|
div.approx.ftz.f32 %f3979, %f3368, %f3372;
|
|
bra.uni $L__BB11_326;
|
|
|
|
$L__BB11_322:
|
|
cvta.to.local.u64 %rd457, %rd68;
|
|
ld.local.v2.f32 {%f3988, %f3989}, [%rd457];
|
|
ld.local.f32 %f3990, [%rd71];
|
|
bra.uni $L__BB11_335;
|
|
|
|
$L__BB11_324:
|
|
ld.local.v2.f32 {%f3295, %f3296}, [%rd100];
|
|
cvta.to.local.u64 %rd459, %rd76;
|
|
ld.local.v4.f32 {%f3299, %f3300, %f3301, %f3302}, [%rd459];
|
|
cvta.to.local.u64 %rd460, %rd90;
|
|
ld.local.v4.f32 {%f3306, %f3307, %f3308, %f3309}, [%rd460];
|
|
cvta.to.local.u64 %rd461, %rd89;
|
|
ld.local.v4.f32 {%f3313, %f3314, %f3315, %f3316}, [%rd461];
|
|
ld.local.f32 %f3320, [%rd79];
|
|
mul.ftz.f32 %f3321, %f3295, %f3306;
|
|
fma.rn.ftz.f32 %f3322, %f3320, %f3299, %f3321;
|
|
fma.rn.ftz.f32 %f3323, %f3296, %f3313, %f3322;
|
|
mul.ftz.f32 %f3324, %f3295, %f3307;
|
|
fma.rn.ftz.f32 %f3325, %f3320, %f3300, %f3324;
|
|
fma.rn.ftz.f32 %f3326, %f3296, %f3314, %f3325;
|
|
mul.ftz.f32 %f3327, %f3295, %f3308;
|
|
fma.rn.ftz.f32 %f3328, %f3320, %f3301, %f3327;
|
|
fma.rn.ftz.f32 %f3329, %f3296, %f3315, %f3328;
|
|
mul.ftz.f32 %f3330, %f3326, %f3326;
|
|
fma.rn.ftz.f32 %f3331, %f3323, %f3323, %f3330;
|
|
fma.rn.ftz.f32 %f3332, %f3329, %f3329, %f3331;
|
|
sqrt.approx.ftz.f32 %f3333, %f3332;
|
|
div.approx.ftz.f32 %f3981, %f3323, %f3333;
|
|
div.approx.ftz.f32 %f3980, %f3326, %f3333;
|
|
div.approx.ftz.f32 %f3979, %f3329, %f3333;
|
|
|
|
$L__BB11_326:
|
|
st.local.f32 [%rd79], %f3981;
|
|
st.local.f32 [%rd100], %f3980;
|
|
st.local.f32 [%rd101], %f3979;
|
|
@%p320 bra $L__BB11_328;
|
|
bra.uni $L__BB11_327;
|
|
|
|
$L__BB11_328:
|
|
ld.local.u32 %r1436, [%rd102];
|
|
mul.wide.s32 %rd469, %r1436, 208;
|
|
add.s64 %rd470, %rd66, %rd469;
|
|
ld.global.v4.f32 {%f3412, %f3413, %f3414, %f3415}, [%rd470+48];
|
|
ld.global.v4.f32 {%f3419, %f3420, %f3421, %f3422}, [%rd470+64];
|
|
ld.global.v4.f32 {%f3426, %f3427, %f3428, %f3429}, [%rd470+80];
|
|
cvta.to.local.u64 %rd471, %rd68;
|
|
ld.local.v2.f32 {%f3433, %f3434}, [%rd471];
|
|
mul.ftz.f32 %f3437, %f3419, %f3434;
|
|
fma.rn.ftz.f32 %f3438, %f3412, %f3433, %f3437;
|
|
ld.local.f32 %f3439, [%rd71];
|
|
fma.rn.ftz.f32 %f3440, %f3426, %f3439, %f3438;
|
|
mul.ftz.f32 %f3441, %f3420, %f3434;
|
|
fma.rn.ftz.f32 %f3442, %f3413, %f3433, %f3441;
|
|
fma.rn.ftz.f32 %f3443, %f3427, %f3439, %f3442;
|
|
mul.ftz.f32 %f3444, %f3421, %f3434;
|
|
fma.rn.ftz.f32 %f3445, %f3414, %f3433, %f3444;
|
|
fma.rn.ftz.f32 %f3446, %f3428, %f3439, %f3445;
|
|
mul.ftz.f32 %f3447, %f3443, %f3443;
|
|
fma.rn.ftz.f32 %f3448, %f3440, %f3440, %f3447;
|
|
fma.rn.ftz.f32 %f3449, %f3446, %f3446, %f3448;
|
|
sqrt.approx.ftz.f32 %f3450, %f3449;
|
|
div.approx.ftz.f32 %f3988, %f3440, %f3450;
|
|
div.approx.ftz.f32 %f3989, %f3443, %f3450;
|
|
div.approx.ftz.f32 %f3990, %f3446, %f3450;
|
|
bra.uni $L__BB11_329;
|
|
|
|
$L__BB11_327:
|
|
cvta.to.local.u64 %rd465, %rd68;
|
|
ld.local.v2.f32 {%f3373, %f3374}, [%rd465];
|
|
cvta.to.local.u64 %rd466, %rd76;
|
|
ld.local.v4.f32 {%f3377, %f3378, %f3379, %f3380}, [%rd466];
|
|
cvta.to.local.u64 %rd467, %rd90;
|
|
ld.local.v4.f32 {%f3384, %f3385, %f3386, %f3387}, [%rd467];
|
|
cvta.to.local.u64 %rd468, %rd89;
|
|
ld.local.v4.f32 {%f3391, %f3392, %f3393, %f3394}, [%rd468];
|
|
mul.ftz.f32 %f3398, %f3374, %f3384;
|
|
fma.rn.ftz.f32 %f3399, %f3373, %f3377, %f3398;
|
|
ld.local.f32 %f3400, [%rd71];
|
|
fma.rn.ftz.f32 %f3401, %f3400, %f3391, %f3399;
|
|
mul.ftz.f32 %f3402, %f3374, %f3385;
|
|
fma.rn.ftz.f32 %f3403, %f3373, %f3378, %f3402;
|
|
fma.rn.ftz.f32 %f3404, %f3400, %f3392, %f3403;
|
|
mul.ftz.f32 %f3405, %f3374, %f3386;
|
|
fma.rn.ftz.f32 %f3406, %f3373, %f3379, %f3405;
|
|
fma.rn.ftz.f32 %f3407, %f3400, %f3393, %f3406;
|
|
mul.ftz.f32 %f3408, %f3404, %f3404;
|
|
fma.rn.ftz.f32 %f3409, %f3401, %f3401, %f3408;
|
|
fma.rn.ftz.f32 %f3410, %f3407, %f3407, %f3409;
|
|
sqrt.approx.ftz.f32 %f3411, %f3410;
|
|
div.approx.ftz.f32 %f3988, %f3401, %f3411;
|
|
div.approx.ftz.f32 %f3989, %f3404, %f3411;
|
|
div.approx.ftz.f32 %f3990, %f3407, %f3411;
|
|
|
|
$L__BB11_329:
|
|
cvta.to.local.u64 %rd472, %rd68;
|
|
st.local.v2.f32 [%rd472], {%f3988, %f3989};
|
|
st.local.f32 [%rd71], %f3990;
|
|
@%p320 bra $L__BB11_331;
|
|
bra.uni $L__BB11_330;
|
|
|
|
$L__BB11_331:
|
|
ld.local.u32 %r1437, [%rd102];
|
|
mul.wide.s32 %rd477, %r1437, 208;
|
|
add.s64 %rd478, %rd66, %rd477;
|
|
ld.global.v4.f32 {%f3483, %f3484, %f3485, %f3486}, [%rd478];
|
|
ld.global.v4.f32 {%f3490, %f3491, %f3492, %f3493}, [%rd478+16];
|
|
ld.global.v4.f32 {%f3497, %f3498, %f3499, %f3500}, [%rd478+32];
|
|
ld.local.v2.f32 {%f3504, %f3505}, [%rd83];
|
|
ld.local.f32 %f3508, [%rd82];
|
|
mul.ftz.f32 %f3509, %f3484, %f3504;
|
|
fma.rn.ftz.f32 %f3510, %f3508, %f3483, %f3509;
|
|
fma.rn.ftz.f32 %f3987, %f3485, %f3505, %f3510;
|
|
mul.ftz.f32 %f3511, %f3491, %f3504;
|
|
fma.rn.ftz.f32 %f3512, %f3508, %f3490, %f3511;
|
|
fma.rn.ftz.f32 %f3986, %f3492, %f3505, %f3512;
|
|
mul.ftz.f32 %f3513, %f3498, %f3504;
|
|
fma.rn.ftz.f32 %f3514, %f3508, %f3497, %f3513;
|
|
fma.rn.ftz.f32 %f3985, %f3499, %f3505, %f3514;
|
|
bra.uni $L__BB11_332;
|
|
|
|
$L__BB11_330:
|
|
ld.local.v2.f32 {%f3451, %f3452}, [%rd83];
|
|
cvta.to.local.u64 %rd474, %rd65;
|
|
ld.local.v4.f32 {%f3455, %f3456, %f3457, %f3458}, [%rd474];
|
|
ld.local.f32 %f3462, [%rd82];
|
|
mul.ftz.f32 %f3463, %f3451, %f3456;
|
|
fma.rn.ftz.f32 %f3464, %f3462, %f3455, %f3463;
|
|
fma.rn.ftz.f32 %f3987, %f3452, %f3457, %f3464;
|
|
cvta.to.local.u64 %rd475, %rd75;
|
|
ld.local.v4.f32 {%f3465, %f3466, %f3467, %f3468}, [%rd475];
|
|
mul.ftz.f32 %f3472, %f3451, %f3466;
|
|
fma.rn.ftz.f32 %f3473, %f3462, %f3465, %f3472;
|
|
fma.rn.ftz.f32 %f3986, %f3452, %f3467, %f3473;
|
|
cvta.to.local.u64 %rd476, %rd74;
|
|
ld.local.v4.f32 {%f3474, %f3475, %f3476, %f3477}, [%rd476];
|
|
mul.ftz.f32 %f3481, %f3451, %f3475;
|
|
fma.rn.ftz.f32 %f3482, %f3462, %f3474, %f3481;
|
|
fma.rn.ftz.f32 %f3985, %f3452, %f3476, %f3482;
|
|
|
|
$L__BB11_332:
|
|
st.local.f32 [%rd82], %f3987;
|
|
st.local.f32 [%rd83], %f3986;
|
|
st.local.f32 [%rd84], %f3985;
|
|
@%p320 bra $L__BB11_334;
|
|
bra.uni $L__BB11_333;
|
|
|
|
$L__BB11_334:
|
|
ld.local.u32 %r1438, [%rd102];
|
|
mul.wide.s32 %rd484, %r1438, 208;
|
|
add.s64 %rd485, %rd66, %rd484;
|
|
ld.global.v4.f32 {%f3552, %f3553, %f3554, %f3555}, [%rd485];
|
|
ld.global.v4.f32 {%f3559, %f3560, %f3561, %f3562}, [%rd485+16];
|
|
ld.global.v4.f32 {%f3566, %f3567, %f3568, %f3569}, [%rd485+32];
|
|
cvta.to.local.u64 %rd486, %rd73;
|
|
ld.local.v4.f32 {%f3573, %f3574, %f3575, %f3576}, [%rd486];
|
|
mul.ftz.f32 %f3580, %f3553, %f3574;
|
|
fma.rn.ftz.f32 %f3581, %f3573, %f3552, %f3580;
|
|
mul.ftz.f32 %f3582, %f3560, %f3574;
|
|
fma.rn.ftz.f32 %f3583, %f3573, %f3559, %f3582;
|
|
mul.ftz.f32 %f3584, %f3567, %f3574;
|
|
fma.rn.ftz.f32 %f3585, %f3573, %f3566, %f3584;
|
|
fma.rn.ftz.f32 %f3586, %f3568, %f3575, %f3585;
|
|
fma.rn.ftz.f32 %f3587, %f3561, %f3575, %f3583;
|
|
fma.rn.ftz.f32 %f3588, %f3554, %f3575, %f3581;
|
|
st.local.v2.f32 [%rd486], {%f3588, %f3587};
|
|
st.local.f32 [%rd85], %f3586;
|
|
bra.uni $L__BB11_335;
|
|
|
|
$L__BB11_333:
|
|
cvta.to.local.u64 %rd480, %rd73;
|
|
ld.local.v4.f32 {%f3515, %f3516, %f3517, %f3518}, [%rd480];
|
|
cvta.to.local.u64 %rd481, %rd65;
|
|
ld.local.v4.f32 {%f3522, %f3523, %f3524, %f3525}, [%rd481];
|
|
mul.ftz.f32 %f3529, %f3516, %f3523;
|
|
fma.rn.ftz.f32 %f3530, %f3515, %f3522, %f3529;
|
|
cvta.to.local.u64 %rd482, %rd75;
|
|
ld.local.v4.f32 {%f3531, %f3532, %f3533, %f3534}, [%rd482];
|
|
mul.ftz.f32 %f3538, %f3516, %f3532;
|
|
fma.rn.ftz.f32 %f3539, %f3515, %f3531, %f3538;
|
|
cvta.to.local.u64 %rd483, %rd74;
|
|
ld.local.v4.f32 {%f3540, %f3541, %f3542, %f3543}, [%rd483];
|
|
mul.ftz.f32 %f3547, %f3516, %f3541;
|
|
fma.rn.ftz.f32 %f3548, %f3515, %f3540, %f3547;
|
|
fma.rn.ftz.f32 %f3549, %f3517, %f3542, %f3548;
|
|
fma.rn.ftz.f32 %f3550, %f3517, %f3533, %f3539;
|
|
fma.rn.ftz.f32 %f3551, %f3517, %f3524, %f3530;
|
|
st.local.v2.f32 [%rd480], {%f3551, %f3550};
|
|
st.local.f32 [%rd85], %f3549;
|
|
|
|
$L__BB11_335:
|
|
ld.local.u32 %r184, [%rd70];
|
|
shl.b32 %r1439, %r184, 5;
|
|
cvt.u64.u32 %rd487, %r1439;
|
|
and.b64 %rd488, %rd487, 134217696;
|
|
add.s64 %rd489, %rd69, %rd488;
|
|
ld.global.u32 %r1521, [%rd489+16];
|
|
st.local.u32 [%rd59], %r1521;
|
|
ld.local.f32 %f3589, [%rd60];
|
|
ld.local.v2.f32 {%f3590, %f3591}, [%rd62];
|
|
mul.ftz.f32 %f3594, %f3989, %f3590;
|
|
fma.rn.ftz.f32 %f3595, %f3988, %f3589, %f3594;
|
|
fma.rn.ftz.f32 %f3596, %f3990, %f3591, %f3595;
|
|
setp.lt.ftz.f32 %p324, %f3596, 0f00000000;
|
|
@%p324 bra $L__BB11_337;
|
|
bra.uni $L__BB11_336;
|
|
|
|
$L__BB11_337:
|
|
or.b32 %r1521, %r1521, 1;
|
|
st.local.u32 [%rd59], %r1521;
|
|
neg.ftz.f32 %f3990, %f3990;
|
|
st.local.f32 [%rd71], %f3990;
|
|
ld.local.f32 %f3597, [%rd79];
|
|
neg.ftz.f32 %f3598, %f3597;
|
|
ld.local.v2.f32 {%f3599, %f3600}, [%rd100];
|
|
st.local.f32 [%rd79], %f3598;
|
|
neg.ftz.f32 %f3603, %f3600;
|
|
neg.ftz.f32 %f3604, %f3599;
|
|
neg.ftz.f32 %f3989, %f3989;
|
|
neg.ftz.f32 %f3988, %f3988;
|
|
st.local.v4.f32 [%rd100], {%f3604, %f3603, %f3988, %f3989};
|
|
ld.local.f32 %f3605, [%rd82];
|
|
neg.ftz.f32 %f4002, %f3605;
|
|
ld.local.v2.f32 {%f3606, %f3607}, [%rd83];
|
|
neg.ftz.f32 %f4003, %f3606;
|
|
neg.ftz.f32 %f3610, %f3607;
|
|
st.local.f32 [%rd82], %f4002;
|
|
st.local.f32 [%rd83], %f4003;
|
|
st.local.f32 [%rd84], %f3610;
|
|
cvta.to.local.u64 %rd493, %rd73;
|
|
ld.local.v4.f32 {%f3611, %f3612, %f3613, %f3614}, [%rd493];
|
|
neg.ftz.f32 %f3618, %f3613;
|
|
neg.ftz.f32 %f3619, %f3612;
|
|
neg.ftz.f32 %f3620, %f3611;
|
|
st.local.v2.f32 [%rd493], {%f3620, %f3619};
|
|
st.local.f32 [%rd85], %f3618;
|
|
bra.uni $L__BB11_338;
|
|
|
|
$L__BB11_336:
|
|
ld.local.f32 %f4002, [%rd82];
|
|
ld.local.f32 %f4003, [%rd83];
|
|
|
|
$L__BB11_338:
|
|
ld.local.f32 %f3621, [%rd54];
|
|
ld.local.f32 %f3622, [%rd53+24];
|
|
ld.local.f32 %f3623, [%rd53+20];
|
|
fma.rn.ftz.f32 %f3624, %f3622, %f3621, %f3623;
|
|
st.local.f32 [%rd86], %f3624;
|
|
st.local.f32 [%rd87], %f3622;
|
|
cvta.to.local.u64 %rd494, %rd73;
|
|
ld.local.v4.f32 {%f4000, %f4001, %f3627, %f3628}, [%rd494];
|
|
setp.neu.ftz.f32 %p325, %f3988, %f3989;
|
|
setp.neu.ftz.f32 %p326, %f3988, %f3990;
|
|
or.pred %p327, %p325, %p326;
|
|
sub.ftz.f32 %f3629, %f3990, %f3989;
|
|
neg.ftz.f32 %f3630, %f3989;
|
|
neg.ftz.f32 %f3631, %f3990;
|
|
selp.f32 %f3632, %f3631, %f3990, %p327;
|
|
add.ftz.f32 %f3633, %f3988, %f3632;
|
|
selp.f32 %f3634, %f3989, %f3630, %p327;
|
|
sub.ftz.f32 %f3635, %f3634, %f3988;
|
|
mul.ftz.f32 %f3636, %f3633, %f3633;
|
|
fma.rn.ftz.f32 %f3637, %f3629, %f3629, %f3636;
|
|
fma.rn.ftz.f32 %f3638, %f3635, %f3635, %f3637;
|
|
sqrt.approx.ftz.f32 %f3639, %f3638;
|
|
div.approx.ftz.f32 %f3640, %f3629, %f3639;
|
|
div.approx.ftz.f32 %f3641, %f3633, %f3639;
|
|
div.approx.ftz.f32 %f3642, %f3635, %f3639;
|
|
mul.ftz.f32 %f3643, %f3989, %f3642;
|
|
mul.ftz.f32 %f3644, %f3990, %f3641;
|
|
sub.ftz.f32 %f3645, %f3643, %f3644;
|
|
mul.ftz.f32 %f3646, %f3990, %f3640;
|
|
mul.ftz.f32 %f3647, %f3988, %f3642;
|
|
sub.ftz.f32 %f3648, %f3646, %f3647;
|
|
mul.ftz.f32 %f3649, %f3988, %f3641;
|
|
mul.ftz.f32 %f3650, %f3989, %f3640;
|
|
sub.ftz.f32 %f3651, %f3649, %f3650;
|
|
mul.ftz.f32 %f3996, %f3624, %f3640;
|
|
mul.ftz.f32 %f3997, %f3624, %f3641;
|
|
mul.ftz.f32 %f1068, %f3624, %f3642;
|
|
mul.ftz.f32 %f3998, %f3624, %f3645;
|
|
mul.ftz.f32 %f3999, %f3624, %f3648;
|
|
mul.ftz.f32 %f1071, %f3624, %f3651;
|
|
abs.ftz.f32 %f1072, %f3988;
|
|
abs.ftz.f32 %f1073, %f3989;
|
|
abs.ftz.f32 %f1074, %f3990;
|
|
setp.geu.ftz.f32 %p328, %f1074, %f1072;
|
|
setp.geu.ftz.f32 %p329, %f1074, %f1073;
|
|
and.pred %p330, %p328, %p329;
|
|
@%p330 bra $L__BB11_340;
|
|
|
|
ld.local.f32 %f1076, [%rd84];
|
|
setp.lt.ftz.f32 %p331, %f1073, %f1074;
|
|
setp.lt.ftz.f32 %p332, %f1073, %f1072;
|
|
or.pred %p333, %p332, %p331;
|
|
selp.f32 %f3996, %f3997, %f3996, %p333;
|
|
selp.f32 %f3998, %f3999, %f3998, %p333;
|
|
selp.f32 %f4000, %f4001, %f4000, %p333;
|
|
selp.f32 %f4002, %f4003, %f4002, %p333;
|
|
mov.f32 %f3997, %f1068;
|
|
mov.f32 %f3999, %f1071;
|
|
mov.f32 %f4001, %f3627;
|
|
mov.f32 %f4003, %f1076;
|
|
|
|
$L__BB11_340:
|
|
mul.ftz.f32 %f3652, %f4000, %f4003;
|
|
mul.ftz.f32 %f3653, %f4001, %f4002;
|
|
sub.ftz.f32 %f4004, %f3653, %f3652;
|
|
setp.eq.ftz.f32 %p334, %f4004, 0f00000000;
|
|
@%p334 bra $L__BB11_342;
|
|
|
|
rcp.approx.ftz.f32 %f4004, %f4004;
|
|
|
|
$L__BB11_342:
|
|
mul.ftz.f32 %f3654, %f3997, %f4000;
|
|
mul.ftz.f32 %f3655, %f3996, %f4001;
|
|
sub.ftz.f32 %f3656, %f3655, %f3654;
|
|
mul.ftz.f32 %f3657, %f3656, %f4004;
|
|
st.local.f32 [%rd97], %f3657;
|
|
mul.ftz.f32 %f3658, %f3996, %f4003;
|
|
mul.ftz.f32 %f3659, %f3997, %f4002;
|
|
sub.ftz.f32 %f3660, %f3659, %f3658;
|
|
mul.ftz.f32 %f3661, %f3999, %f4000;
|
|
mul.ftz.f32 %f3662, %f3998, %f4001;
|
|
sub.ftz.f32 %f3663, %f3662, %f3661;
|
|
mul.ftz.f32 %f3664, %f3660, %f4004;
|
|
mul.ftz.f32 %f3665, %f3663, %f4004;
|
|
st.local.v2.f32 [%rd495], {%f3665, %f3664};
|
|
mul.ftz.f32 %f3666, %f3998, %f4003;
|
|
mul.ftz.f32 %f3667, %f3999, %f4002;
|
|
sub.ftz.f32 %f3668, %f3667, %f3666;
|
|
mul.ftz.f32 %f3669, %f3668, %f4004;
|
|
st.local.f32 [%rd99], %f3669;
|
|
and.b32 %r1440, %r1521, 262144;
|
|
setp.eq.s32 %p335, %r1440, 0;
|
|
@%p335 bra $L__BB11_356;
|
|
bra.uni $L__BB11_343;
|
|
|
|
$L__BB11_356:
|
|
mov.b32 %f3696, %r1275;
|
|
setp.eq.ftz.f32 %p352, %f3696, 0f00000000;
|
|
add.s32 %r1452, %r1275, 1;
|
|
mov.b32 %f3670, %r1452;
|
|
selp.f32 %f3861, 0f00000001, %f3670, %p352;
|
|
st.local.f32 [%rd53+-8], %f3861;
|
|
st.local.u32 [%rd53+8], %r1279;
|
|
st.local.u32 [%rd53+4], %r1278;
|
|
add.s32 %r1509, %r1509, 1;
|
|
setp.ge.u32 %p353, %r1509, %r99;
|
|
@%p353 bra $L__BB11_358;
|
|
bra.uni $L__BB11_357;
|
|
|
|
$L__BB11_343:
|
|
and.b32 %r1441, %r1521, 1;
|
|
setp.eq.b32 %p336, %r1441, 1;
|
|
xor.pred %p338, %p336, %p337;
|
|
not.pred %p339, %p338;
|
|
ld.global.u32 %r1522, [%rd104];
|
|
@%p339 bra $L__BB11_350;
|
|
bra.uni $L__BB11_344;
|
|
|
|
$L__BB11_350:
|
|
mov.u64 %rd530, kernel_params;
|
|
add.s64 %rd717, %rd530, 2792;
|
|
add.s64 %rd716, %rd530, 2784;
|
|
setp.eq.s32 %p346, %r1522, -1;
|
|
mov.u32 %r1527, 0;
|
|
@%p346 bra $L__BB11_354;
|
|
|
|
add.s64 %rd716, %rd530, 2784;
|
|
ld.local.u32 %r197, [%rd102];
|
|
mov.u32 %r1527, 0;
|
|
|
|
$L__BB11_352:
|
|
ld.const.u64 %rd533, [%rd716];
|
|
cvta.to.global.u64 %rd534, %rd533;
|
|
add.s64 %rd536, %rd534, %rd168;
|
|
ld.global.u32 %r1448, [%rd536];
|
|
setp.eq.s32 %p347, %r1448, %r197;
|
|
setp.eq.s32 %p348, %r1522, %r184;
|
|
and.pred %p349, %p347, %p348;
|
|
@%p349 bra $L__BB11_356;
|
|
|
|
add.s32 %r1527, %r1527, 1;
|
|
mul.wide.s32 %rd537, %r1527, 16;
|
|
add.s64 %rd539, %rd530, %rd537;
|
|
add.s64 %rd716, %rd539, 2784;
|
|
add.s64 %rd717, %rd539, 2792;
|
|
ld.const.u64 %rd540, [%rd539+2792];
|
|
cvta.to.global.u64 %rd541, %rd540;
|
|
add.s64 %rd543, %rd541, %rd168;
|
|
ld.global.u32 %r1522, [%rd543];
|
|
setp.ne.s32 %p350, %r1522, -1;
|
|
@%p350 bra $L__BB11_352;
|
|
|
|
$L__BB11_354:
|
|
setp.ge.u32 %p351, %r1527, %r103;
|
|
@%p351 bra $L__BB11_356;
|
|
|
|
ld.local.u32 %r1449, [%rd102];
|
|
ld.const.u64 %rd544, [%rd716];
|
|
cvta.to.global.u64 %rd545, %rd544;
|
|
add.s64 %rd547, %rd545, %rd168;
|
|
st.global.u32 [%rd547], %r1449;
|
|
ld.const.u64 %rd548, [%rd717];
|
|
cvta.to.global.u64 %rd549, %rd548;
|
|
add.s64 %rd550, %rd549, %rd168;
|
|
st.global.u32 [%rd550], %r184;
|
|
add.s32 %r1450, %r1527, 1;
|
|
mul.wide.s32 %rd551, %r1450, 16;
|
|
add.s64 %rd553, %rd530, %rd551;
|
|
ld.const.v2.u64 {%rd554, %rd555}, [%rd553+2784];
|
|
cvta.to.global.u64 %rd558, %rd554;
|
|
add.s64 %rd559, %rd558, %rd168;
|
|
mov.u32 %r1451, -1;
|
|
st.global.u32 [%rd559], %r1451;
|
|
cvta.to.global.u64 %rd560, %rd555;
|
|
add.s64 %rd561, %rd560, %rd168;
|
|
st.global.u32 [%rd561], %r1451;
|
|
bra.uni $L__BB11_356;
|
|
|
|
$L__BB11_344:
|
|
setp.eq.s32 %p340, %r1522, -1;
|
|
@%p340 bra $L__BB11_356;
|
|
|
|
mov.u32 %r1523, 0;
|
|
mov.u64 %rd497, kernel_params;
|
|
add.s64 %rd712, %rd497, 2784;
|
|
ld.local.u32 %r189, [%rd102];
|
|
|
|
$L__BB11_346:
|
|
ld.const.u64 %rd498, [%rd712];
|
|
cvta.to.global.u64 %rd499, %rd498;
|
|
add.s64 %rd501, %rd499, %rd168;
|
|
ld.global.u32 %r1443, [%rd501];
|
|
setp.eq.s32 %p341, %r1443, %r189;
|
|
setp.eq.s32 %p342, %r1522, %r184;
|
|
and.pred %p343, %p341, %p342;
|
|
add.s32 %r192, %r1523, 1;
|
|
mul.wide.s32 %rd502, %r192, 16;
|
|
add.s64 %rd504, %rd497, %rd502;
|
|
add.s64 %rd129, %rd504, 2792;
|
|
@%p343 bra $L__BB11_348;
|
|
|
|
add.s64 %rd712, %rd504, 2784;
|
|
ld.const.u64 %rd508, [%rd129];
|
|
cvta.to.global.u64 %rd509, %rd508;
|
|
add.s64 %rd511, %rd509, %rd168;
|
|
ld.global.u32 %r1522, [%rd511];
|
|
setp.eq.s32 %p344, %r1522, -1;
|
|
mov.u32 %r1523, %r192;
|
|
@%p344 bra $L__BB11_356;
|
|
bra.uni $L__BB11_346;
|
|
|
|
$L__BB11_348:
|
|
ld.const.v2.u64 {%rd714, %rd713}, [%rd129+-24];
|
|
|
|
$L__BB11_349:
|
|
add.s32 %r1523, %r1523, 1;
|
|
mul.wide.s32 %rd514, %r1523, 16;
|
|
add.s64 %rd516, %rd497, %rd514;
|
|
ld.const.v2.u64 {%rd517, %rd518}, [%rd516+2784];
|
|
cvta.to.global.u64 %rd519, %rd517;
|
|
add.s64 %rd521, %rd519, %rd168;
|
|
ld.global.u32 %r1444, [%rd521];
|
|
cvta.to.global.u64 %rd522, %rd518;
|
|
add.s64 %rd523, %rd522, %rd168;
|
|
ld.global.u32 %r1445, [%rd523];
|
|
cvta.to.global.u64 %rd524, %rd714;
|
|
add.s64 %rd525, %rd524, %rd168;
|
|
st.global.u32 [%rd525], %r1444;
|
|
cvta.to.global.u64 %rd526, %rd713;
|
|
add.s64 %rd527, %rd526, %rd168;
|
|
st.global.u32 [%rd527], %r1445;
|
|
setp.eq.s32 %p345, %r1445, -1;
|
|
mov.u64 %rd713, %rd518;
|
|
mov.u64 %rd714, %rd517;
|
|
@%p345 bra $L__BB11_356;
|
|
bra.uni $L__BB11_349;
|
|
|
|
$L__BB11_358:
|
|
add.u64 %rd681, %SPL, 424;
|
|
add.u64 %rd679, %SPL, 364;
|
|
ld.local.f32 %f3671, [%rd679+12];
|
|
ld.local.f32 %f3672, [%rd679+28];
|
|
mul.ftz.f32 %f3673, %f3672, %f3671;
|
|
ld.local.f32 %f3674, [%rd679+16];
|
|
mul.ftz.f32 %f3675, %f3672, %f3674;
|
|
ld.local.f32 %f3676, [%rd679+20];
|
|
mul.ftz.f32 %f3677, %f3672, %f3676;
|
|
ld.local.f32 %f3678, [%rd679];
|
|
fma.rn.ftz.f32 %f3679, %f3673, 0f40000000, %f3678;
|
|
ld.local.f32 %f3680, [%rd679+4];
|
|
fma.rn.ftz.f32 %f3681, %f3675, 0f40000000, %f3680;
|
|
ld.local.f32 %f3682, [%rd679+8];
|
|
fma.rn.ftz.f32 %f3683, %f3677, 0f40000000, %f3682;
|
|
st.local.f32 [%rd679], %f3679;
|
|
st.local.f32 [%rd679+4], %f3681;
|
|
st.local.f32 [%rd679+8], %f3683;
|
|
neg.ftz.f32 %f3684, %f3671;
|
|
neg.ftz.f32 %f3685, %f3674;
|
|
neg.ftz.f32 %f3686, %f3676;
|
|
st.local.f32 [%rd679+12], %f3684;
|
|
st.local.f32 [%rd679+16], %f3685;
|
|
st.local.f32 [%rd679+20], %f3686;
|
|
ld.local.f32 %f3687, [%rd681+4];
|
|
ld.const.v2.u64 {%rd562, %rd563}, [kernel_params+2704];
|
|
cvta.to.global.u64 %rd566, %rd562;
|
|
add.s64 %rd568, %rd566, %rd168;
|
|
st.global.f32 [%rd568], %f3687;
|
|
ld.local.f32 %f3688, [%rd681+8];
|
|
cvta.to.global.u64 %rd569, %rd563;
|
|
add.s64 %rd570, %rd569, %rd168;
|
|
st.global.f32 [%rd570], %f3688;
|
|
ld.local.f32 %f3689, [%rd681+12];
|
|
ld.const.v2.u64 {%rd571, %rd572}, [kernel_params+2720];
|
|
cvta.to.global.u64 %rd575, %rd571;
|
|
add.s64 %rd576, %rd575, %rd168;
|
|
st.global.f32 [%rd576], %f3689;
|
|
ld.local.u32 %r204, [%rd681+20];
|
|
cvta.to.global.u64 %rd577, %rd706;
|
|
add.s64 %rd578, %rd577, %rd168;
|
|
st.global.u32 [%rd578], %r204;
|
|
ld.local.u32 %r205, [%rd681+16];
|
|
cvt.s64.s32 %rd142, %r205;
|
|
cvta.to.global.u64 %rd579, %rd572;
|
|
add.s64 %rd580, %rd579, %rd168;
|
|
st.global.u32 [%rd580], %r205;
|
|
ld.local.u32 %r206, [%rd681+24];
|
|
ld.const.u64 %rd581, [kernel_params+2744];
|
|
cvta.to.global.u64 %rd582, %rd581;
|
|
add.s64 %rd583, %rd582, %rd168;
|
|
st.global.u32 [%rd583], %r206;
|
|
cvta.to.global.u64 %rd584, %rd707;
|
|
mul.lo.s64 %rd585, %rd4, 12;
|
|
add.s64 %rd586, %rd584, %rd585;
|
|
st.global.f32 [%rd586], %f3679;
|
|
st.global.f32 [%rd586+4], %f3681;
|
|
st.global.f32 [%rd586+8], %f3683;
|
|
ld.const.v2.u64 {%rd587, %rd588}, [kernel_params+2656];
|
|
cvta.to.global.u64 %rd591, %rd587;
|
|
add.s64 %rd592, %rd591, %rd585;
|
|
st.global.f32 [%rd592], %f3684;
|
|
st.global.f32 [%rd592+4], %f3685;
|
|
st.global.f32 [%rd592+8], %f3686;
|
|
ld.local.f32 %f3690, [%rd679+24];
|
|
cvta.to.global.u64 %rd593, %rd588;
|
|
add.s64 %rd594, %rd593, %rd168;
|
|
st.global.f32 [%rd594], %f3690;
|
|
ld.const.u64 %rd595, [kernel_params+2672];
|
|
cvta.to.global.u64 %rd596, %rd595;
|
|
add.s64 %rd597, %rd596, %rd168;
|
|
st.global.f32 [%rd597], %f3672;
|
|
ld.local.f32 %f3691, [%rd679+32];
|
|
cvta.to.global.u64 %rd598, %rd705;
|
|
add.s64 %rd599, %rd598, %rd168;
|
|
st.global.f32 [%rd599], %f3691;
|
|
ld.local.f32 %f3692, [%rd679+52];
|
|
cvta.to.global.u64 %rd600, %rd704;
|
|
add.s64 %rd601, %rd600, %rd168;
|
|
st.global.f32 [%rd601], %f3692;
|
|
ld.local.f32 %f3693, [%rd679+56];
|
|
ld.const.u64 %rd602, [kernel_params+2696];
|
|
cvta.to.global.u64 %rd603, %rd602;
|
|
add.s64 %rd604, %rd603, %rd168;
|
|
st.global.f32 [%rd604], %f3693;
|
|
ld.global.u16 %rs80, [%rd5];
|
|
add.s16 %rs81, %rs80, 16;
|
|
st.global.u16 [%rd5], %rs81;
|
|
and.b32 %r1453, %r206, 1;
|
|
setp.eq.b32 %p354, %r1453, 1;
|
|
mov.pred %p375, 0;
|
|
xor.pred %p356, %p354, %p375;
|
|
not.pred %p357, %p356;
|
|
@%p357 bra $L__BB11_360;
|
|
bra.uni $L__BB11_359;
|
|
|
|
$L__BB11_360:
|
|
and.b32 %r1454, %r206, 8;
|
|
setp.eq.s32 %p358, %r1454, 0;
|
|
@%p358 bra $L__BB11_362;
|
|
bra.uni $L__BB11_361;
|
|
|
|
$L__BB11_362:
|
|
and.b32 %r1457, %r206, 6;
|
|
setp.eq.s32 %p359, %r1457, 0;
|
|
mov.u32 %r1528, 0;
|
|
@%p359 bra $L__BB11_364;
|
|
|
|
ld.const.u64 %rd613, [kernel_params+2280];
|
|
cvta.to.global.u64 %rd614, %rd613;
|
|
mul.wide.s32 %rd615, %r205, 16;
|
|
add.s64 %rd616, %rd614, %rd615;
|
|
ld.global.u32 %r1528, [%rd616];
|
|
bra.uni $L__BB11_364;
|
|
|
|
$L__BB11_359:
|
|
ld.const.u64 %rd605, [kernel_params+2232];
|
|
cvta.to.global.u64 %rd606, %rd605;
|
|
shl.b64 %rd607, %rd142, 2;
|
|
add.s64 %rd608, %rd606, %rd607;
|
|
ld.global.u32 %r1528, [%rd608];
|
|
bra.uni $L__BB11_364;
|
|
|
|
$L__BB11_210:
|
|
add.s32 %r95, %r1495, 1;
|
|
@%p104 bra $L__BB11_213;
|
|
|
|
ld.const.v2.u32 {%r1166, %r1167}, [kernel_params+1616];
|
|
mov.b32 %f546, %r1167;
|
|
setp.geu.ftz.f32 %p227, %f546, 0f3F800000;
|
|
setp.lt.ftz.f32 %p228, %f546, 0f3F800000;
|
|
ld.const.u32 %r1170, [kernel_params+1592];
|
|
selp.b32 %r1171, %r1170, %r4, %p228;
|
|
add.s16 %rs68, %rs1, 1;
|
|
and.b16 %rs69, %rs68, 255;
|
|
cvt.u32.u16 %r1172, %rs69;
|
|
xor.b32 %r1173, %r1171, %r1172;
|
|
mul.lo.s32 %r1174, %r1173, -512718531;
|
|
shr.u32 %r1175, %r1171, 16;
|
|
xor.b32 %r1176, %r1174, %r1175;
|
|
bfe.u32 %r1177, %r1176, 4, 4;
|
|
shr.u32 %r1178, %r1171, 8;
|
|
xor.b32 %r1179, %r1176, %r1178;
|
|
xor.b32 %r1180, %r1179, %r1177;
|
|
mul.lo.s32 %r1181, %r1180, 153742143;
|
|
shr.u32 %r1182, %r1171, 23;
|
|
xor.b32 %r1183, %r1181, %r1182;
|
|
bfe.u32 %r1184, %r1183, 1, 7;
|
|
xor.b32 %r1185, %r1184, %r1183;
|
|
shr.u32 %r1186, %r1171, 27;
|
|
or.b32 %r1187, %r1186, 1;
|
|
mul.lo.s32 %r1188, %r1187, %r1185;
|
|
mul.lo.s32 %r1189, %r1188, -1074439621;
|
|
bfe.u32 %r1190, %r1189, 2, 6;
|
|
xor.b32 %r1191, %r1190, %r1189;
|
|
mul.lo.s32 %r1192, %r1191, -1638916925;
|
|
bfe.u32 %r1193, %r1192, 2, 6;
|
|
xor.b32 %r1194, %r1193, %r1192;
|
|
mul.lo.s32 %r1195, %r1194, -933190689;
|
|
and.b32 %r1196, %r1195, 255;
|
|
bfe.u32 %r1197, %r1195, 5, 3;
|
|
xor.b32 %r1198, %r1197, %r1196;
|
|
add.s32 %r1199, %r1166, -1;
|
|
xor.b32 %r96, %r95, 61;
|
|
xor.b32 %r1200, %r1171, %r96;
|
|
mul.lo.s32 %r1201, %r1200, 9;
|
|
shr.u32 %r1202, %r1201, 4;
|
|
xor.b32 %r1203, %r1202, %r1201;
|
|
mul.lo.s32 %r1204, %r1203, 668265261;
|
|
add.s32 %r1205, %r1204, %r10;
|
|
shr.u32 %r1206, %r1204, 16;
|
|
or.b32 %r1207, %r1206, 1;
|
|
mul.lo.s32 %r1208, %r1207, %r1205;
|
|
mul.lo.s32 %r1209, %r1208, 89287766;
|
|
xor.b32 %r1210, %r1209, %r1208;
|
|
mul.lo.s32 %r1211, %r1210, 1403136100;
|
|
xor.b32 %r1212, %r1211, %r1210;
|
|
brev.b32 %r1213, %r1212;
|
|
neg.s32 %r1214, %r1166;
|
|
and.b32 %r1215, %r6, %r1214;
|
|
and.b32 %r1216, %r1213, %r1199;
|
|
or.b32 %r1217, %r1216, %r1215;
|
|
mad.lo.s32 %r1218, %r1198, %r1166, %r1217;
|
|
shl.b32 %r1219, %r1166, 8;
|
|
rem.u32 %r1220, %r1218, %r1219;
|
|
ld.const.u64 %rd286, [kernel_params+2440];
|
|
cvta.to.global.u64 %rd287, %rd286;
|
|
shl.b32 %r1221, %r1220, 1;
|
|
mul.wide.u32 %rd288, %r1221, 4;
|
|
add.s64 %rd289, %rd287, %rd288;
|
|
ld.global.f32 %f3856, [%rd289];
|
|
@%p227 bra $L__BB11_214;
|
|
|
|
xor.b32 %r1222, %r96, %r4;
|
|
mul.lo.s32 %r1223, %r1222, 9;
|
|
shr.u32 %r1224, %r1223, 4;
|
|
xor.b32 %r1225, %r1224, %r1223;
|
|
mul.lo.s32 %r1226, %r1225, 668265261;
|
|
cvt.rn.f32.u32 %f2310, %r1226;
|
|
mul.ftz.f32 %f2311, %f2310, 0f2F7FFFFE;
|
|
fma.rn.ftz.f32 %f2312, %f2311, %f546, %f3856;
|
|
cvt.rmi.ftz.f32.f32 %f2313, %f2312;
|
|
sub.ftz.f32 %f3856, %f2312, %f2313;
|
|
bra.uni $L__BB11_214;
|
|
|
|
$L__BB11_361:
|
|
cvt.u32.u64 %r1455, %rd142;
|
|
ld.const.u64 %rd609, [kernel_params+2320];
|
|
cvta.to.global.u64 %rd610, %rd609;
|
|
mul.wide.s32 %rd611, %r1455, 4;
|
|
add.s64 %rd612, %rd610, %rd611;
|
|
ld.global.u32 %r1528, [%rd612];
|
|
|
|
$L__BB11_364:
|
|
and.b32 %r211, %r1528, 4194303;
|
|
ld.const.u64 %rd617, [kernel_params+2200];
|
|
cvta.to.global.u64 %rd618, %rd617;
|
|
mul.wide.s32 %rd619, %r204, 4;
|
|
add.s64 %rd620, %rd618, %rd619;
|
|
ld.const.u32 %r1458, [kernel_params+1608];
|
|
setp.ne.s32 %p360, %r1458, 0;
|
|
ld.global.u8 %rs82, [%rd620+1];
|
|
and.b16 %rs83, %rs82, 6;
|
|
setp.ne.s16 %p361, %rs83, 0;
|
|
and.pred %p362, %p361, %p360;
|
|
@%p362 bra $L__BB11_368;
|
|
bra.uni $L__BB11_365;
|
|
|
|
$L__BB11_368:
|
|
cvt.u32.u64 %r1470, %rd4;
|
|
ld.const.u32 %r1471, [kernel_params+4408];
|
|
div.u32 %r1472, %r1470, %r1471;
|
|
ld.const.u32 %r1473, [kernel_params+24];
|
|
mad.lo.s32 %r1474, %r1472, %r1473, %r211;
|
|
ld.const.u64 %rd653, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd654, %rd653;
|
|
add.s64 %rd655, %rd654, 16;
|
|
atom.global.add.u32 %r1475, [%rd655], -1;
|
|
add.s64 %rd656, %rd654, 40;
|
|
atom.global.add.u32 %r1476, [%rd656], 1;
|
|
ld.const.u64 %rd657, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd658, %rd657;
|
|
shl.b64 %rd659, %rd4, 1;
|
|
add.s64 %rd660, %rd658, %rd659;
|
|
mov.u16 %rs86, 10;
|
|
st.global.u16 [%rd660], %rs86;
|
|
ld.const.u64 %rd661, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd662, %rd661;
|
|
add.s64 %rd664, %rd662, %rd168;
|
|
st.global.u32 [%rd664], %r1474;
|
|
ld.const.u64 %rd665, [kernel_params+4360];
|
|
cvta.to.global.u64 %rd666, %rd665;
|
|
mul.wide.s32 %rd667, %r1474, 4;
|
|
add.s64 %rd668, %rd666, %rd667;
|
|
atom.global.add.u32 %r1477, [%rd668], 1;
|
|
bra.uni $L__BB11_369;
|
|
|
|
$L__BB11_365:
|
|
ld.const.u64 %rd621, [kernel_params+2424];
|
|
cvt.u32.u64 %r1459, %rd4;
|
|
cvta.to.global.u64 %rd622, %rd621;
|
|
mul.wide.u32 %rd623, %r211, 32;
|
|
add.s64 %rd624, %rd622, %rd623;
|
|
ld.global.u32 %r1460, [%rd624+16];
|
|
and.b32 %r1461, %r1460, 1073741824;
|
|
setp.eq.s32 %p363, %r1461, 0;
|
|
ld.const.u32 %r1462, [kernel_params+4408];
|
|
div.u32 %r1463, %r1459, %r1462;
|
|
ld.const.u32 %r1464, [kernel_params+24];
|
|
mad.lo.s32 %r212, %r1463, %r1464, %r211;
|
|
ld.const.u64 %rd625, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd143, %rd625;
|
|
add.s64 %rd626, %rd143, 16;
|
|
atom.global.add.u32 %r1465, [%rd626], -1;
|
|
cvt.s64.s32 %rd144, %r212;
|
|
@%p363 bra $L__BB11_367;
|
|
bra.uni $L__BB11_366;
|
|
|
|
$L__BB11_367:
|
|
add.s64 %rd640, %rd143, 32;
|
|
atom.global.add.u32 %r1468, [%rd640], 1;
|
|
ld.const.u64 %rd641, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd642, %rd641;
|
|
shl.b64 %rd643, %rd4, 1;
|
|
add.s64 %rd644, %rd642, %rd643;
|
|
mov.u16 %rs85, 8;
|
|
st.global.u16 [%rd644], %rs85;
|
|
ld.const.u64 %rd645, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd646, %rd645;
|
|
add.s64 %rd648, %rd646, %rd168;
|
|
st.global.u32 [%rd648], %r212;
|
|
ld.const.u64 %rd649, [kernel_params+4344];
|
|
cvta.to.global.u64 %rd650, %rd649;
|
|
shl.b64 %rd651, %rd144, 2;
|
|
add.s64 %rd652, %rd650, %rd651;
|
|
atom.global.add.u32 %r1469, [%rd652], 1;
|
|
bra.uni $L__BB11_369;
|
|
|
|
$L__BB11_366:
|
|
add.s64 %rd627, %rd143, 36;
|
|
atom.global.add.u32 %r1466, [%rd627], 1;
|
|
ld.const.u64 %rd628, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd629, %rd628;
|
|
shl.b64 %rd630, %rd4, 1;
|
|
add.s64 %rd631, %rd629, %rd630;
|
|
mov.u16 %rs84, 9;
|
|
st.global.u16 [%rd631], %rs84;
|
|
ld.const.u64 %rd632, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd633, %rd632;
|
|
add.s64 %rd635, %rd633, %rd168;
|
|
st.global.u32 [%rd635], %r212;
|
|
ld.const.u64 %rd636, [kernel_params+4352];
|
|
cvta.to.global.u64 %rd637, %rd636;
|
|
shl.b64 %rd638, %rd144, 2;
|
|
add.s64 %rd639, %rd637, %rd638;
|
|
atom.global.add.u32 %r1467, [%rd639], 1;
|
|
|
|
$L__BB11_369:
|
|
not.pred %p367, %p375;
|
|
@%p367 bra $L__BB11_371;
|
|
|
|
ld.const.u64 %rd669, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd670, %rd669;
|
|
add.s64 %rd671, %rd670, 16;
|
|
atom.global.add.u32 %r1478, [%rd671], -1;
|
|
ld.const.u64 %rd672, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd673, %rd672;
|
|
shl.b64 %rd674, %rd4, 1;
|
|
add.s64 %rd675, %rd673, %rd674;
|
|
mov.u16 %rs87, 0;
|
|
st.global.u16 [%rd675], %rs87;
|
|
|
|
$L__BB11_371:
|
|
ret;
|
|
|
|
$L__BB11_213:
|
|
shr.u32 %r1227, %r95, 16;
|
|
xor.b32 %r1228, %r1227, %r95;
|
|
mul.lo.s32 %r1229, %r1228, 569420461;
|
|
shr.u32 %r1230, %r1229, 15;
|
|
xor.b32 %r1231, %r1230, %r1229;
|
|
mul.lo.s32 %r1232, %r1231, -749064809;
|
|
shr.u32 %r1233, %r1232, 15;
|
|
xor.b32 %r1234, %r4, %r1232;
|
|
xor.b32 %r1235, %r1234, %r1233;
|
|
xor.b32 %r1236, %r1235, 1493655573;
|
|
add.s32 %r1237, %r1236, %r10;
|
|
shr.u32 %r1238, %r1236, 16;
|
|
or.b32 %r1239, %r1238, 1;
|
|
mul.lo.s32 %r1240, %r1239, %r1237;
|
|
mul.lo.s32 %r1241, %r1240, 89287766;
|
|
xor.b32 %r1242, %r1241, %r1240;
|
|
mul.lo.s32 %r1243, %r1242, 1403136100;
|
|
xor.b32 %r1244, %r1243, %r1242;
|
|
xor.b32 %r1245, %r1235, -2052961194;
|
|
brev.b32 %r1246, %r1244;
|
|
mul.lo.s32 %r1247, %r1246, 1025551850;
|
|
xor.b32 %r1248, %r1247, %r1246;
|
|
add.s32 %r1249, %r1248, %r1245;
|
|
shr.u32 %r1250, %r1245, 16;
|
|
or.b32 %r1251, %r1250, 1;
|
|
mul.lo.s32 %r1252, %r1249, %r1251;
|
|
mul.lo.s32 %r1253, %r1252, 89287766;
|
|
xor.b32 %r1254, %r1253, %r1252;
|
|
mul.lo.s32 %r1255, %r1254, 1403136100;
|
|
xor.b32 %r1256, %r1255, %r1254;
|
|
brev.b32 %r1257, %r1256;
|
|
cvt.rn.f32.u32 %f2314, %r1257;
|
|
mul.ftz.f32 %f3856, %f2314, 0f2F7FFFFE;
|
|
|
|
$L__BB11_214:
|
|
mul.ftz.f32 %f551, %f3855, %f3856;
|
|
mov.pred %p374, %p375;
|
|
@%p195 bra $L__BB11_223;
|
|
|
|
cvta.to.local.u64 %rd292, %rd145;
|
|
ld.local.v4.f32 {%f3857, %f3858, %f3859, %f2318}, [%rd292];
|
|
mov.u64 %rd702, 0;
|
|
abs.ftz.f32 %f2319, %f3857;
|
|
abs.ftz.f32 %f2320, %f3858;
|
|
abs.ftz.f32 %f2321, %f3859;
|
|
add.ftz.f32 %f2322, %f2319, %f2320;
|
|
add.ftz.f32 %f2323, %f2321, %f2322;
|
|
mul.ftz.f32 %f3860, %f2323, 0f3EAAAAAB;
|
|
add.ftz.f32 %f556, %f3860, 0f00000000;
|
|
setp.lt.ftz.f32 %p231, %f551, %f556;
|
|
@%p231 bra $L__BB11_222;
|
|
|
|
setp.lt.s32 %p233, %r78, 2;
|
|
mov.pred %p374, %p375;
|
|
@%p233 bra $L__BB11_223;
|
|
|
|
add.s64 %rd45, %rd1, 12;
|
|
ld.local.v2.f32 {%f3858, %f3859}, [%rd1+16];
|
|
mov.u64 %rd702, 1;
|
|
ld.local.f32 %f3857, [%rd1+12];
|
|
abs.ftz.f32 %f2326, %f3857;
|
|
abs.ftz.f32 %f2327, %f3858;
|
|
abs.ftz.f32 %f2328, %f3859;
|
|
add.ftz.f32 %f2329, %f2326, %f2327;
|
|
add.ftz.f32 %f2330, %f2328, %f2329;
|
|
mul.ftz.f32 %f3860, %f2330, 0f3EAAAAAB;
|
|
add.ftz.f32 %f561, %f556, %f3860;
|
|
setp.lt.ftz.f32 %p234, %f551, %f561;
|
|
@%p234 bra $L__BB11_222;
|
|
|
|
setp.lt.s32 %p236, %r78, 3;
|
|
mov.pred %p374, %p375;
|
|
@%p236 bra $L__BB11_223;
|
|
|
|
ld.local.v2.f32 {%f3857, %f3858}, [%rd45+12];
|
|
abs.ftz.f32 %f2333, %f3857;
|
|
abs.ftz.f32 %f2334, %f3858;
|
|
ld.local.f32 %f3859, [%rd45+20];
|
|
abs.ftz.f32 %f2335, %f3859;
|
|
add.ftz.f32 %f2336, %f2333, %f2334;
|
|
add.ftz.f32 %f2337, %f2335, %f2336;
|
|
mul.ftz.f32 %f3860, %f2337, 0f3EAAAAAB;
|
|
add.ftz.f32 %f566, %f561, %f3860;
|
|
setp.lt.ftz.f32 %p237, %f551, %f566;
|
|
mov.u64 %rd702, 2;
|
|
@%p237 bra $L__BB11_222;
|
|
|
|
setp.lt.s32 %p239, %r78, 4;
|
|
mov.pred %p374, %p375;
|
|
@%p239 bra $L__BB11_223;
|
|
|
|
ld.local.v2.f32 {%f3858, %f3859}, [%rd45+28];
|
|
ld.local.f32 %f3857, [%rd45+24];
|
|
abs.ftz.f32 %f2340, %f3857;
|
|
abs.ftz.f32 %f2341, %f3858;
|
|
abs.ftz.f32 %f2342, %f3859;
|
|
add.ftz.f32 %f2343, %f2340, %f2341;
|
|
add.ftz.f32 %f2344, %f2342, %f2343;
|
|
mul.ftz.f32 %f3860, %f2344, 0f3EAAAAAB;
|
|
add.ftz.f32 %f2345, %f566, %f3860;
|
|
setp.geu.ftz.f32 %p241, %f551, %f2345;
|
|
mov.u64 %rd702, 3;
|
|
mov.pred %p374, %p375;
|
|
@%p241 bra $L__BB11_223;
|
|
|
|
$L__BB11_222:
|
|
mul.ftz.f32 %f2346, %f3855, %f3857;
|
|
div.approx.ftz.f32 %f2347, %f2346, %f3860;
|
|
mul.ftz.f32 %f2348, %f3855, %f3858;
|
|
div.approx.ftz.f32 %f2349, %f2348, %f3860;
|
|
mul.ftz.f32 %f2350, %f3855, %f3859;
|
|
div.approx.ftz.f32 %f2351, %f2350, %f3860;
|
|
ld.const.u64 %rd296, [kernel_params+2600];
|
|
cvta.to.global.u64 %rd297, %rd296;
|
|
add.s64 %rd299, %rd297, %rd243;
|
|
ld.global.f32 %f2352, [%rd299];
|
|
mul.ftz.f32 %f2353, %f2347, %f2352;
|
|
ld.global.f32 %f2354, [%rd299+4];
|
|
mul.ftz.f32 %f2355, %f2349, %f2354;
|
|
ld.global.f32 %f2356, [%rd299+8];
|
|
mul.ftz.f32 %f2357, %f2351, %f2356;
|
|
st.global.f32 [%rd299], %f2353;
|
|
st.global.f32 [%rd299+4], %f2355;
|
|
st.global.f32 [%rd299+8], %f2357;
|
|
mul.lo.s64 %rd300, %rd702, 24;
|
|
add.s64 %rd301, %rd2, %rd300;
|
|
ld.local.f32 %f2358, [%rd301+4];
|
|
ld.local.f32 %f2359, [%rd301+8];
|
|
ld.local.f32 %f2360, [%rd301+12];
|
|
ld.local.u32 %r1258, [%rd301+16];
|
|
ld.local.u32 %r1259, [%rd301+20];
|
|
ld.local.u32 %r1260, [%rd301+24];
|
|
st.local.f32 [%rd2+4], %f2358;
|
|
st.local.f32 [%rd2+8], %f2359;
|
|
st.local.f32 [%rd2+12], %f2360;
|
|
st.local.u32 [%rd2+16], %r1258;
|
|
st.local.u32 [%rd2+20], %r1259;
|
|
st.local.u32 [%rd2+24], %r1260;
|
|
mul.lo.s64 %rd302, %rd702, 12;
|
|
add.s64 %rd303, %rd2, %rd302;
|
|
ld.local.f32 %f2361, [%rd303+100];
|
|
ld.local.f32 %f2362, [%rd303+104];
|
|
ld.local.f32 %f2363, [%rd303+108];
|
|
st.local.f32 [%rd2+100], %f2361;
|
|
st.local.f32 [%rd2+104], %f2362;
|
|
st.local.f32 [%rd2+108], %f2363;
|
|
ld.local.f32 %f2364, [%rd6+12];
|
|
ld.local.f32 %f2365, [%rd301+4];
|
|
ld.local.f32 %f2366, [%rd6+16];
|
|
ld.local.f32 %f2367, [%rd6+20];
|
|
ld.local.f32 %f2368, [%rd6];
|
|
fma.rn.ftz.f32 %f2369, %f2365, %f2364, %f2368;
|
|
ld.local.f32 %f2370, [%rd6+4];
|
|
fma.rn.ftz.f32 %f2371, %f2365, %f2366, %f2370;
|
|
ld.local.f32 %f2372, [%rd6+8];
|
|
fma.rn.ftz.f32 %f2373, %f2365, %f2367, %f2372;
|
|
st.local.f32 [%rd6], %f2369;
|
|
st.local.f32 [%rd6+4], %f2371;
|
|
st.local.f32 [%rd6+8], %f2373;
|
|
ld.local.f32 %f2374, [%rd303+100];
|
|
ld.local.f32 %f2375, [%rd303+104];
|
|
ld.local.f32 %f2376, [%rd303+108];
|
|
st.local.f32 [%rd6+12], %f2374;
|
|
st.local.f32 [%rd6+16], %f2375;
|
|
st.local.f32 [%rd6+20], %f2376;
|
|
mov.u32 %r1261, 0;
|
|
st.local.u32 [%rd6+24], %r1261;
|
|
mov.u32 %r1262, 1065353216;
|
|
st.local.u32 [%rd6+28], %r1262;
|
|
mov.pred %p374, 0;
|
|
bra.uni $L__BB11_223;
|
|
|
|
}
|
|
// .globl __raygen__kernel_optix_integrator_intersect_volume_stack
|
|
.visible .entry __raygen__kernel_optix_integrator_intersect_volume_stack()
|
|
{
|
|
.local .align 16 .b8 __local_depot12[496];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<135>;
|
|
.reg .b16 %rs<49>;
|
|
.reg .f32 %f<1968>;
|
|
.reg .b32 %r<377>;
|
|
.reg .b64 %rd<441>;
|
|
|
|
|
|
mov.u64 %SPL, __local_depot12;
|
|
cvta.local.u64 %SP, %SPL;
|
|
add.u64 %rd1, %SPL, 368;
|
|
// begin inline asm
|
|
call (%r354), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
ld.const.u64 %rd2, [kernel_params];
|
|
setp.eq.s64 %p3, %rd2, 0;
|
|
@%p3 bra $L__BB12_2;
|
|
|
|
cvta.to.global.u64 %rd87, %rd2;
|
|
mul.wide.s32 %rd88, %r354, 4;
|
|
add.s64 %rd89, %rd87, %rd88;
|
|
ld.global.u32 %r354, [%rd89];
|
|
|
|
$L__BB12_2:
|
|
ld.const.u64 %rd90, [kernel_params+2648];
|
|
cvta.to.global.u64 %rd91, %rd90;
|
|
cvt.s64.s32 %rd3, %r354;
|
|
mul.wide.s32 %rd92, %r354, 12;
|
|
add.s64 %rd93, %rd91, %rd92;
|
|
ld.global.f32 %f1825, [%rd93];
|
|
ld.global.f32 %f520, [%rd93+4];
|
|
ld.global.f32 %f521, [%rd93+8];
|
|
add.u64 %rd94, %SP, 304;
|
|
add.u64 %rd4, %SPL, 304;
|
|
st.local.f32 [%rd4], %f1825;
|
|
st.local.f32 [%rd4+4], %f520;
|
|
st.local.f32 [%rd4+8], %f521;
|
|
ld.const.u64 %rd95, [kernel_params+2680];
|
|
cvta.to.global.u64 %rd96, %rd95;
|
|
mul.wide.s32 %rd97, %r354, 4;
|
|
add.s64 %rd98, %rd96, %rd97;
|
|
ld.global.f32 %f522, [%rd98];
|
|
st.local.f32 [%rd4+32], %f522;
|
|
ld.const.v2.u64 {%rd99, %rd100}, [kernel_params+2688];
|
|
mov.u32 %r375, 0;
|
|
cvta.to.global.u64 %rd103, %rd99;
|
|
add.s64 %rd104, %rd103, %rd97;
|
|
ld.global.f32 %f523, [%rd104];
|
|
st.local.f32 [%rd4+52], %f523;
|
|
cvta.to.global.u64 %rd105, %rd100;
|
|
add.s64 %rd106, %rd105, %rd97;
|
|
ld.global.f32 %f524, [%rd106];
|
|
st.local.f32 [%rd4+56], %f524;
|
|
st.local.u32 [%rd4+12], %r375;
|
|
st.local.u32 [%rd4+16], %r375;
|
|
mov.u32 %r124, 1065353216;
|
|
st.local.u32 [%rd4+20], %r124;
|
|
st.local.u32 [%rd4+24], %r375;
|
|
mov.u32 %r125, 2139095039;
|
|
st.local.u32 [%rd4+28], %r125;
|
|
mov.u32 %r126, -1;
|
|
st.local.u32 [%rd4+40], %r126;
|
|
st.local.u32 [%rd4+36], %r126;
|
|
st.local.u32 [%rd4+48], %r126;
|
|
st.local.u32 [%rd4+44], %r126;
|
|
ld.const.u64 %rd107, [kernel_params+2560];
|
|
cvta.to.global.u64 %rd108, %rd107;
|
|
add.s64 %rd5, %rd108, %rd97;
|
|
ld.global.u32 %r127, [%rd5];
|
|
and.b32 %r4, %r127, 1073741824;
|
|
shr.u32 %r128, %r127, 30;
|
|
and.b32 %r129, %r128, 1;
|
|
setp.eq.b32 %p4, %r129, 1;
|
|
setp.eq.s32 %p5, %r4, 0;
|
|
selp.b32 %r5, 1, 65536, %p5;
|
|
ld.const.u32 %r6, [kernel_params+1044];
|
|
setp.eq.s32 %p6, %r6, -1;
|
|
or.pred %p7, %p6, %p4;
|
|
@%p7 bra $L__BB12_4;
|
|
|
|
ld.const.v2.u64 {%rd109, %rd110}, [kernel_params+2784];
|
|
mov.u32 %r375, 1;
|
|
cvta.to.global.u64 %rd113, %rd109;
|
|
shl.b64 %rd114, %rd3, 2;
|
|
add.s64 %rd115, %rd113, %rd114;
|
|
mov.u32 %r131, -1;
|
|
st.global.u32 [%rd115], %r131;
|
|
cvta.to.global.u64 %rd116, %rd110;
|
|
add.s64 %rd117, %rd116, %rd114;
|
|
st.global.u32 [%rd117], %r6;
|
|
|
|
$L__BB12_4:
|
|
ld.const.u32 %r132, [kernel_params+28];
|
|
add.s32 %r8, %r132, -1;
|
|
setp.ge.u32 %p8, %r375, %r8;
|
|
shl.b32 %r9, %r132, 1;
|
|
setp.eq.s32 %p9, %r9, 0;
|
|
or.pred %p10, %p9, %p8;
|
|
@%p10 bra $L__BB12_133;
|
|
|
|
mov.u32 %r373, 0;
|
|
shr.u64 %rd119, %rd94, 32;
|
|
cvt.u32.u64 %r11, %rd119;
|
|
and.b32 %r135, %r5, 1;
|
|
setp.eq.s32 %p11, %r135, 0;
|
|
setp.ne.s32 %p12, %r4, 0;
|
|
and.pred %p13, %p11, %p12;
|
|
selp.b32 %r12, 255, %r135, %p13;
|
|
add.u64 %rd120, %SP, 0;
|
|
add.s64 %rd121, %rd120, 72;
|
|
add.s64 %rd122, %rd120, 88;
|
|
cvta.to.local.u64 %rd7, %rd122;
|
|
add.s64 %rd8, %rd120, 64;
|
|
add.s64 %rd123, %rd120, 76;
|
|
ld.const.v2.u64 {%rd124, %rd125}, [kernel_params+2192];
|
|
cvta.to.global.u64 %rd9, %rd125;
|
|
add.s64 %rd128, %rd120, 56;
|
|
cvta.to.local.u64 %rd10, %rd128;
|
|
add.s64 %rd129, %rd120, 60;
|
|
cvta.to.local.u64 %rd11, %rd129;
|
|
add.s64 %rd12, %rd120, 80;
|
|
add.s64 %rd130, %rd120, 52;
|
|
cvta.to.local.u64 %rd13, %rd130;
|
|
add.s64 %rd131, %rd120, 36;
|
|
cvta.to.local.u64 %rd14, %rd131;
|
|
add.s64 %rd15, %rd120, 40;
|
|
cvta.to.local.u64 %rd16, %rd15;
|
|
add.s64 %rd17, %rd120, 44;
|
|
cvta.to.local.u64 %rd18, %rd17;
|
|
add.s64 %rd19, %rd120, 144;
|
|
ld.const.u64 %rd132, [kernel_params+2176];
|
|
cvta.to.global.u64 %rd20, %rd132;
|
|
cvta.to.global.u64 %rd21, %rd124;
|
|
add.s64 %rd22, %rd120, 24;
|
|
ld.const.u64 %rd133, [kernel_params+2424];
|
|
cvta.to.global.u64 %rd23, %rd133;
|
|
add.s64 %rd134, %rd120, 48;
|
|
cvta.to.local.u64 %rd24, %rd134;
|
|
add.s64 %rd135, %rd120, 32;
|
|
cvta.to.local.u64 %rd25, %rd135;
|
|
add.s64 %rd136, %rd120, 116;
|
|
add.s64 %rd26, %rd120, 120;
|
|
add.s64 %rd137, %rd120, 124;
|
|
add.s64 %rd27, %rd120, 128;
|
|
add.s64 %rd28, %rd120, 176;
|
|
add.s64 %rd29, %rd120, 160;
|
|
add.s64 %rd30, %rd120, 192;
|
|
ld.const.v2.u64 {%rd138, %rd139}, [kernel_params+2320];
|
|
cvta.to.global.u64 %rd31, %rd138;
|
|
or.b64 %rd142, %rd120, 8;
|
|
cvta.to.local.u64 %rd32, %rd142;
|
|
or.b64 %rd143, %rd120, 12;
|
|
cvta.to.local.u64 %rd33, %rd143;
|
|
add.s64 %rd34, %rd120, 16;
|
|
add.s64 %rd35, %rd120, 20;
|
|
cvta.to.local.u64 %rd36, %rd136;
|
|
cvta.to.local.u64 %rd37, %rd26;
|
|
cvta.to.local.u64 %rd38, %rd137;
|
|
add.s64 %rd144, %rd120, 136;
|
|
cvta.to.local.u64 %rd39, %rd144;
|
|
add.s64 %rd145, %rd120, 92;
|
|
cvta.to.local.u64 %rd40, %rd145;
|
|
add.s64 %rd146, %rd120, 96;
|
|
cvta.to.local.u64 %rd41, %rd146;
|
|
ld.const.v2.u64 {%rd147, %rd148}, [kernel_params+2272];
|
|
add.s64 %rd43, %rd120, 224;
|
|
add.s64 %rd44, %rd120, 208;
|
|
ld.const.v2.u64 {%rd151, %rd152}, [kernel_params+2240];
|
|
cvta.to.global.u64 %rd46, %rd147;
|
|
ld.const.u64 %rd155, [kernel_params+2312];
|
|
cvta.to.global.u64 %rd47, %rd155;
|
|
cvta.to.global.u64 %rd48, %rd139;
|
|
ld.const.u64 %rd156, [kernel_params+2288];
|
|
cvta.to.global.u64 %rd49, %rd156;
|
|
ld.const.u64 %rd157, [kernel_params+2232];
|
|
cvta.to.global.u64 %rd50, %rd157;
|
|
add.s64 %rd158, %rd120, 100;
|
|
cvta.to.local.u64 %rd51, %rd158;
|
|
add.s64 %rd52, %rd120, 104;
|
|
add.s64 %rd159, %rd120, 112;
|
|
cvta.to.local.u64 %rd53, %rd159;
|
|
cvta.to.local.u64 %rd54, %rd34;
|
|
cvta.to.local.u64 %rd55, %rd35;
|
|
cvta.to.local.u64 %rd56, %rd123;
|
|
ld.const.u64 %rd160, [kernel_params+2360];
|
|
cvta.to.global.u64 %rd57, %rd160;
|
|
cvta.to.local.u64 %rd58, %rd121;
|
|
cvta.to.global.u64 %rd59, %rd151;
|
|
add.u64 %rd60, %SPL, 0;
|
|
mov.f32 %f1824, 0f00000000;
|
|
mov.u16 %rs19, 0;
|
|
cvta.to.local.u64 %rd162, %rd8;
|
|
cvta.to.local.u64 %rd165, %rd12;
|
|
cvta.to.local.u64 %rd297, %rd52;
|
|
add.s64 %rd82, %rd60, 76;
|
|
mov.u32 %r358, %r373;
|
|
bra.uni $L__BB12_6;
|
|
|
|
$L__BB12_132:
|
|
ld.local.f32 %f1825, [%rd4];
|
|
|
|
$L__BB12_6:
|
|
mov.b32 %r16, %f1825;
|
|
setp.nan.ftz.f32 %p14, %f1825, %f1825;
|
|
mov.u16 %rs39, %rs19;
|
|
@%p14 bra $L__BB12_9;
|
|
|
|
or.b32 %r136, %r16, -2147483648;
|
|
setp.ne.s32 %p15, %r136, -2147483648;
|
|
add.ftz.f32 %f526, %f1825, %f1825;
|
|
setp.eq.ftz.f32 %p16, %f526, %f1825;
|
|
and.pred %p17, %p16, %p15;
|
|
mov.u16 %rs39, %rs19;
|
|
@%p17 bra $L__BB12_9;
|
|
|
|
shl.b32 %r137, %r16, 1;
|
|
setp.lt.u32 %p18, %r137, -16777215;
|
|
selp.u16 %rs39, 1, 0, %p18;
|
|
|
|
$L__BB12_9:
|
|
ld.local.f32 %f4, [%rd4+12];
|
|
setp.eq.s16 %p19, %rs39, 0;
|
|
mov.u16 %rs41, %rs19;
|
|
@%p19 bra $L__BB12_15;
|
|
|
|
mov.b32 %r17, %f4;
|
|
setp.nan.ftz.f32 %p20, %f4, %f4;
|
|
mov.u16 %rs41, 0;
|
|
mov.u16 %rs40, %rs41;
|
|
@%p20 bra $L__BB12_13;
|
|
|
|
or.b32 %r138, %r17, -2147483648;
|
|
setp.ne.s32 %p21, %r138, -2147483648;
|
|
add.ftz.f32 %f527, %f4, %f4;
|
|
setp.eq.ftz.f32 %p22, %f527, %f4;
|
|
and.pred %p23, %p22, %p21;
|
|
mov.u16 %rs40, %rs41;
|
|
@%p23 bra $L__BB12_13;
|
|
|
|
shl.b32 %r139, %r17, 1;
|
|
setp.lt.u32 %p24, %r139, -16777215;
|
|
selp.u16 %rs40, 1, 0, %p24;
|
|
|
|
$L__BB12_13:
|
|
setp.eq.s16 %p25, %rs40, 0;
|
|
@%p25 bra $L__BB12_15;
|
|
|
|
ld.local.f32 %f528, [%rd4+16];
|
|
mul.ftz.f32 %f529, %f528, %f528;
|
|
fma.rn.ftz.f32 %f530, %f4, %f4, %f529;
|
|
ld.local.f32 %f531, [%rd4+20];
|
|
fma.rn.ftz.f32 %f532, %f531, %f531, %f530;
|
|
setp.neu.ftz.f32 %p26, %f532, 0f00000000;
|
|
selp.u16 %rs41, 1, 0, %p26;
|
|
|
|
$L__BB12_15:
|
|
add.u64 %rd429, %SP, 304;
|
|
cvt.u32.u64 %r345, %rd429;
|
|
ld.const.u64 %rd428, [kernel_params+2096];
|
|
setp.eq.s16 %p27, %rs41, 0;
|
|
selp.b64 %rd161, 0, %rd428, %p27;
|
|
ld.local.f32 %f534, [%rd4+4];
|
|
ld.local.f32 %f535, [%rd4+8];
|
|
ld.local.f32 %f537, [%rd4+16];
|
|
ld.local.f32 %f538, [%rd4+20];
|
|
ld.local.f32 %f540, [%rd4+28];
|
|
ld.local.f32 %f541, [%rd4+32];
|
|
mov.u32 %r174, 2;
|
|
mov.u32 %r175, 3;
|
|
mov.u32 %r178, 8;
|
|
mov.u32 %r184, 0;
|
|
// begin inline asm
|
|
call(%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167,%r168,%r169,%r170,%r171),_optix_trace_typed_32,(%r184,%rd161,%f1825,%f534,%f535,%f4,%f537,%f538,%f1824,%f540,%f541,%r12,%r174,%r175,%r184,%r184,%r178,%r184,%r184,%r184,%r184,%r5,%r184,%r345,%r11,%r211,%r212,%r213,%r214,%r215,%r216,%r217,%r218,%r219,%r220,%r221,%r222,%r223,%r224,%r225,%r226,%r227,%r228,%r229,%r230,%r231,%r232,%r233,%r234);
|
|
// end inline asm
|
|
setp.eq.s32 %p28, %r145, 0;
|
|
@%p28 bra $L__BB12_133;
|
|
|
|
mov.u32 %r346, 0;
|
|
mov.b32 %f1806, %r140;
|
|
st.local.f32 [%rd7], %f1806;
|
|
st.local.v4.u32 [%rd162], {%r145, %r141, %r142, %r144};
|
|
mul.wide.s32 %rd163, %r144, 4;
|
|
add.s64 %rd164, %rd9, %rd163;
|
|
ld.global.u32 %r369, [%rd164];
|
|
st.local.u32 [%rd10], %r369;
|
|
st.local.u32 [%rd11], %r143;
|
|
st.local.u32 [%rd13], %r346;
|
|
ld.local.f32 %f8, [%rd4+32];
|
|
mov.b32 %r236, %f8;
|
|
st.local.v2.u32 [%rd165], {%r126, %r236};
|
|
and.b32 %r57, %r369, 2;
|
|
setp.eq.s32 %p29, %r57, 0;
|
|
mul.wide.s32 %rd166, %r144, 208;
|
|
add.s64 %rd61, %rd20, %rd166;
|
|
@%p29 bra $L__BB12_23;
|
|
|
|
ld.global.u32 %rd167, [%rd61+168];
|
|
ld.global.u32 %r238, [%rd61+152];
|
|
shl.b32 %r239, %r238, 1;
|
|
cvt.rn.f32.s32 %f542, %r239;
|
|
mul.ftz.f32 %f543, %f8, %f542;
|
|
cvt.rzi.ftz.s32.f32 %r240, %f543;
|
|
add.s32 %r241, %r239, -1;
|
|
min.s32 %r242, %r240, %r241;
|
|
cvt.rn.f32.s32 %f544, %r242;
|
|
sub.ftz.f32 %f9, %f543, %f544;
|
|
cvt.s64.s32 %rd168, %r242;
|
|
add.s64 %rd169, %rd168, %rd167;
|
|
shl.b64 %rd170, %rd169, 6;
|
|
add.s64 %rd62, %rd21, %rd170;
|
|
ld.global.v4.f32 {%f545, %f546, %f547, %f548}, [%rd62];
|
|
ld.global.v4.f32 {%f549, %f550, %f551, %f552}, [%rd62+64];
|
|
mul.ftz.f32 %f553, %f546, %f550;
|
|
fma.rn.ftz.f32 %f554, %f545, %f549, %f553;
|
|
mul.ftz.f32 %f555, %f548, %f552;
|
|
fma.rn.ftz.f32 %f556, %f547, %f551, %f555;
|
|
add.ftz.f32 %f18, %f554, %f556;
|
|
setp.gt.ftz.f32 %p30, %f18, 0f3F7FDF3B;
|
|
@%p30 bra $L__BB12_19;
|
|
bra.uni $L__BB12_18;
|
|
|
|
$L__BB12_19:
|
|
mov.f32 %f1823, 0f3F800000;
|
|
sub.ftz.f32 %f1822, %f1823, %f9;
|
|
mul.ftz.f32 %f611, %f545, %f1822;
|
|
fma.rn.ftz.f32 %f612, %f549, %f9, %f611;
|
|
mul.ftz.f32 %f613, %f546, %f1822;
|
|
fma.rn.ftz.f32 %f614, %f550, %f9, %f613;
|
|
mul.ftz.f32 %f615, %f547, %f1822;
|
|
fma.rn.ftz.f32 %f616, %f551, %f9, %f615;
|
|
mul.ftz.f32 %f617, %f548, %f1822;
|
|
fma.rn.ftz.f32 %f618, %f552, %f9, %f617;
|
|
mul.ftz.f32 %f619, %f614, %f614;
|
|
fma.rn.ftz.f32 %f620, %f612, %f612, %f619;
|
|
mul.ftz.f32 %f621, %f618, %f618;
|
|
fma.rn.ftz.f32 %f622, %f616, %f616, %f621;
|
|
add.ftz.f32 %f623, %f620, %f622;
|
|
rsqrt.approx.ftz.f32 %f624, %f623;
|
|
mul.ftz.f32 %f1829, %f618, %f624;
|
|
mul.ftz.f32 %f1828, %f616, %f624;
|
|
mul.ftz.f32 %f1827, %f614, %f624;
|
|
mul.ftz.f32 %f1826, %f612, %f624;
|
|
bra.uni $L__BB12_20;
|
|
|
|
$L__BB12_18:
|
|
mov.f32 %f1819, 0f3F800000;
|
|
mov.f32 %f558, 0fBF800000;
|
|
max.ftz.f32 %f559, %f18, %f558;
|
|
min.ftz.f32 %f561, %f559, %f1819;
|
|
abs.ftz.f32 %f562, %f561;
|
|
sub.ftz.f32 %f563, %f1819, %f562;
|
|
mul.ftz.f32 %f564, %f563, 0f3F000000;
|
|
sqrt.approx.ftz.f32 %f565, %f564;
|
|
setp.gt.ftz.f32 %p31, %f562, 0f3F11EB85;
|
|
selp.f32 %f566, %f565, %f562, %p31;
|
|
mul.ftz.f32 %f567, %f566, %f566;
|
|
mov.f32 %f568, 0f3C94D2E9;
|
|
mov.f32 %f569, 0f3D53F941;
|
|
fma.rn.ftz.f32 %f570, %f569, %f567, %f568;
|
|
mov.f32 %f571, 0f3D3F841F;
|
|
fma.rn.ftz.f32 %f572, %f570, %f567, %f571;
|
|
mov.f32 %f573, 0f3D994929;
|
|
fma.rn.ftz.f32 %f574, %f572, %f567, %f573;
|
|
mov.f32 %f575, 0f3E2AAB94;
|
|
fma.rn.ftz.f32 %f576, %f574, %f567, %f575;
|
|
mul.ftz.f32 %f577, %f567, %f576;
|
|
fma.rn.ftz.f32 %f578, %f577, %f566, %f566;
|
|
add.ftz.f32 %f579, %f578, %f578;
|
|
mov.f32 %f580, 0f3FC90FDB;
|
|
sub.ftz.f32 %f581, %f580, %f578;
|
|
selp.f32 %f582, %f579, %f581, %p31;
|
|
setp.lt.ftz.f32 %p32, %f561, 0f00000000;
|
|
mov.f32 %f583, 0f40490FDB;
|
|
sub.ftz.f32 %f584, %f583, %f582;
|
|
selp.f32 %f585, %f584, %f582, %p32;
|
|
mul.ftz.f32 %f586, %f545, %f18;
|
|
sub.ftz.f32 %f587, %f549, %f586;
|
|
mul.ftz.f32 %f588, %f546, %f18;
|
|
sub.ftz.f32 %f589, %f550, %f588;
|
|
mul.ftz.f32 %f590, %f547, %f18;
|
|
sub.ftz.f32 %f591, %f551, %f590;
|
|
mul.ftz.f32 %f592, %f548, %f18;
|
|
sub.ftz.f32 %f593, %f552, %f592;
|
|
mul.ftz.f32 %f594, %f589, %f589;
|
|
fma.rn.ftz.f32 %f595, %f587, %f587, %f594;
|
|
mul.ftz.f32 %f596, %f593, %f593;
|
|
fma.rn.ftz.f32 %f597, %f591, %f591, %f596;
|
|
add.ftz.f32 %f598, %f595, %f597;
|
|
rsqrt.approx.ftz.f32 %f599, %f598;
|
|
mul.ftz.f32 %f600, %f587, %f599;
|
|
mul.ftz.f32 %f601, %f589, %f599;
|
|
mul.ftz.f32 %f602, %f591, %f599;
|
|
mul.ftz.f32 %f603, %f593, %f599;
|
|
mul.ftz.f32 %f604, %f9, %f585;
|
|
cos.approx.ftz.f32 %f605, %f604;
|
|
sin.approx.ftz.f32 %f606, %f604;
|
|
mul.ftz.f32 %f607, %f600, %f606;
|
|
mul.ftz.f32 %f608, %f601, %f606;
|
|
mul.ftz.f32 %f609, %f602, %f606;
|
|
mul.ftz.f32 %f610, %f603, %f606;
|
|
fma.rn.ftz.f32 %f1829, %f548, %f605, %f610;
|
|
fma.rn.ftz.f32 %f1828, %f547, %f605, %f609;
|
|
fma.rn.ftz.f32 %f1827, %f546, %f605, %f608;
|
|
fma.rn.ftz.f32 %f1826, %f545, %f605, %f607;
|
|
|
|
$L__BB12_20:
|
|
mov.f32 %f1821, 0f3F800000;
|
|
sub.ftz.f32 %f1820, %f1821, %f9;
|
|
ld.global.v4.f32 {%f625, %f626, %f627, %f628}, [%rd62+16];
|
|
ld.global.v4.f32 {%f633, %f634, %f635, %f636}, [%rd62+80];
|
|
mul.ftz.f32 %f641, %f9, %f633;
|
|
mul.ftz.f32 %f642, %f9, %f634;
|
|
mul.ftz.f32 %f643, %f9, %f635;
|
|
mul.ftz.f32 %f644, %f9, %f636;
|
|
fma.rn.ftz.f32 %f32, %f1820, %f625, %f641;
|
|
fma.rn.ftz.f32 %f33, %f1820, %f626, %f642;
|
|
fma.rn.ftz.f32 %f34, %f1820, %f627, %f643;
|
|
fma.rn.ftz.f32 %f645, %f1820, %f628, %f644;
|
|
ld.global.v4.f32 {%f646, %f647, %f648, %f649}, [%rd62+32];
|
|
ld.global.v4.f32 {%f654, %f655, %f656, %f657}, [%rd62+96];
|
|
mul.ftz.f32 %f662, %f9, %f654;
|
|
mul.ftz.f32 %f663, %f9, %f655;
|
|
mul.ftz.f32 %f664, %f9, %f656;
|
|
mul.ftz.f32 %f665, %f9, %f657;
|
|
fma.rn.ftz.f32 %f666, %f1820, %f646, %f662;
|
|
fma.rn.ftz.f32 %f667, %f1820, %f647, %f663;
|
|
fma.rn.ftz.f32 %f668, %f1820, %f648, %f664;
|
|
fma.rn.ftz.f32 %f669, %f1820, %f649, %f665;
|
|
ld.global.v4.f32 {%f670, %f671, %f672, %f673}, [%rd62+48];
|
|
ld.global.v4.f32 {%f678, %f679, %f680, %f681}, [%rd62+112];
|
|
mul.ftz.f32 %f686, %f9, %f678;
|
|
mul.ftz.f32 %f687, %f9, %f679;
|
|
mul.ftz.f32 %f688, %f9, %f680;
|
|
mul.ftz.f32 %f689, %f9, %f681;
|
|
fma.rn.ftz.f32 %f690, %f1820, %f670, %f686;
|
|
fma.rn.ftz.f32 %f691, %f1820, %f671, %f687;
|
|
fma.rn.ftz.f32 %f692, %f1820, %f672, %f688;
|
|
fma.rn.ftz.f32 %f693, %f1820, %f673, %f689;
|
|
mul.ftz.f32 %f694, %f1829, 0f3FB504F3;
|
|
mul.ftz.f32 %f695, %f1826, 0f3FB504F3;
|
|
mul.ftz.f32 %f696, %f1827, 0f3FB504F3;
|
|
mul.ftz.f32 %f697, %f1828, 0f3FB504F3;
|
|
mul.ftz.f32 %f698, %f694, %f695;
|
|
mul.ftz.f32 %f699, %f694, %f696;
|
|
mul.ftz.f32 %f700, %f694, %f697;
|
|
mul.ftz.f32 %f701, %f695, %f695;
|
|
mul.ftz.f32 %f702, %f695, %f696;
|
|
mul.ftz.f32 %f703, %f695, %f697;
|
|
mul.ftz.f32 %f704, %f696, %f696;
|
|
mul.ftz.f32 %f705, %f696, %f697;
|
|
mul.ftz.f32 %f706, %f697, %f697;
|
|
sub.ftz.f32 %f708, %f1821, %f704;
|
|
sub.ftz.f32 %f709, %f708, %f706;
|
|
sub.ftz.f32 %f710, %f702, %f700;
|
|
add.ftz.f32 %f711, %f699, %f703;
|
|
add.ftz.f32 %f712, %f700, %f702;
|
|
sub.ftz.f32 %f713, %f1821, %f701;
|
|
sub.ftz.f32 %f714, %f713, %f706;
|
|
sub.ftz.f32 %f715, %f705, %f698;
|
|
sub.ftz.f32 %f716, %f703, %f699;
|
|
add.ftz.f32 %f717, %f698, %f705;
|
|
sub.ftz.f32 %f718, %f713, %f704;
|
|
mul.ftz.f32 %f719, %f710, %f668;
|
|
fma.rn.ftz.f32 %f720, %f709, %f645, %f719;
|
|
fma.rn.ftz.f32 %f1835, %f711, %f691, %f720;
|
|
mul.ftz.f32 %f721, %f710, %f669;
|
|
fma.rn.ftz.f32 %f722, %f709, %f666, %f721;
|
|
fma.rn.ftz.f32 %f36, %f711, %f692, %f722;
|
|
mul.ftz.f32 %f723, %f710, %f690;
|
|
fma.rn.ftz.f32 %f724, %f709, %f667, %f723;
|
|
fma.rn.ftz.f32 %f37, %f711, %f693, %f724;
|
|
mul.ftz.f32 %f725, %f714, %f668;
|
|
fma.rn.ftz.f32 %f726, %f712, %f645, %f725;
|
|
fma.rn.ftz.f32 %f38, %f715, %f691, %f726;
|
|
mul.ftz.f32 %f727, %f714, %f669;
|
|
fma.rn.ftz.f32 %f728, %f712, %f666, %f727;
|
|
fma.rn.ftz.f32 %f1834, %f715, %f692, %f728;
|
|
mul.ftz.f32 %f729, %f714, %f690;
|
|
fma.rn.ftz.f32 %f730, %f712, %f667, %f729;
|
|
fma.rn.ftz.f32 %f40, %f715, %f693, %f730;
|
|
mul.ftz.f32 %f731, %f717, %f668;
|
|
fma.rn.ftz.f32 %f732, %f716, %f645, %f731;
|
|
fma.rn.ftz.f32 %f41, %f718, %f691, %f732;
|
|
mul.ftz.f32 %f733, %f717, %f669;
|
|
fma.rn.ftz.f32 %f734, %f716, %f666, %f733;
|
|
fma.rn.ftz.f32 %f42, %f718, %f692, %f734;
|
|
mul.ftz.f32 %f735, %f717, %f690;
|
|
fma.rn.ftz.f32 %f736, %f716, %f667, %f735;
|
|
fma.rn.ftz.f32 %f1833, %f718, %f693, %f736;
|
|
cvta.to.local.u64 %rd171, %rd19;
|
|
st.local.v4.f32 [%rd171], {%f1835, %f36, %f37, %f32};
|
|
cvta.to.local.u64 %rd172, %rd29;
|
|
st.local.v4.f32 [%rd172], {%f38, %f1834, %f40, %f33};
|
|
cvta.to.local.u64 %rd173, %rd28;
|
|
st.local.v4.f32 [%rd173], {%f41, %f42, %f1833, %f34};
|
|
mul.ftz.f32 %f737, %f1834, %f1833;
|
|
mul.ftz.f32 %f44, %f42, %f40;
|
|
sub.ftz.f32 %f1832, %f737, %f44;
|
|
mul.ftz.f32 %f46, %f42, %f37;
|
|
mul.ftz.f32 %f738, %f36, %f1833;
|
|
sub.ftz.f32 %f1831, %f46, %f738;
|
|
mul.ftz.f32 %f48, %f36, %f40;
|
|
mul.ftz.f32 %f739, %f1834, %f37;
|
|
sub.ftz.f32 %f1830, %f48, %f739;
|
|
mul.ftz.f32 %f740, %f38, %f1831;
|
|
fma.rn.ftz.f32 %f741, %f1835, %f1832, %f740;
|
|
fma.rn.ftz.f32 %f1836, %f41, %f1830, %f741;
|
|
setp.neu.ftz.f32 %p33, %f1836, 0f00000000;
|
|
@%p33 bra $L__BB12_22;
|
|
|
|
add.ftz.f32 %f1835, %f1835, 0f322BCC77;
|
|
add.ftz.f32 %f1833, %f1833, 0f322BCC77;
|
|
add.ftz.f32 %f1834, %f1834, 0f322BCC77;
|
|
mul.ftz.f32 %f742, %f1834, %f1833;
|
|
sub.ftz.f32 %f1832, %f742, %f44;
|
|
mul.ftz.f32 %f743, %f36, %f1833;
|
|
sub.ftz.f32 %f1831, %f46, %f743;
|
|
mul.ftz.f32 %f744, %f1834, %f37;
|
|
sub.ftz.f32 %f1830, %f48, %f744;
|
|
mul.ftz.f32 %f745, %f38, %f1831;
|
|
fma.rn.ftz.f32 %f746, %f1835, %f1832, %f745;
|
|
fma.rn.ftz.f32 %f747, %f41, %f1830, %f746;
|
|
setp.eq.ftz.f32 %p34, %f747, 0f00000000;
|
|
selp.f32 %f1836, 0f7F7FFFFF, %f747, %p34;
|
|
|
|
$L__BB12_22:
|
|
mul.ftz.f32 %f748, %f38, %f1833;
|
|
mul.ftz.f32 %f749, %f41, %f40;
|
|
sub.ftz.f32 %f750, %f749, %f748;
|
|
mul.ftz.f32 %f751, %f41, %f37;
|
|
mul.ftz.f32 %f752, %f1833, %f1835;
|
|
sub.ftz.f32 %f753, %f752, %f751;
|
|
mul.ftz.f32 %f754, %f40, %f1835;
|
|
mul.ftz.f32 %f755, %f38, %f37;
|
|
sub.ftz.f32 %f756, %f755, %f754;
|
|
mul.ftz.f32 %f757, %f41, %f1834;
|
|
mul.ftz.f32 %f758, %f38, %f42;
|
|
sub.ftz.f32 %f759, %f758, %f757;
|
|
mul.ftz.f32 %f760, %f42, %f1835;
|
|
mul.ftz.f32 %f761, %f41, %f36;
|
|
sub.ftz.f32 %f762, %f761, %f760;
|
|
mul.ftz.f32 %f763, %f38, %f36;
|
|
mul.ftz.f32 %f764, %f1834, %f1835;
|
|
sub.ftz.f32 %f765, %f764, %f763;
|
|
div.approx.ftz.f32 %f766, %f1831, %f1836;
|
|
mul.ftz.f32 %f767, %f33, %f766;
|
|
div.approx.ftz.f32 %f768, %f1832, %f1836;
|
|
fma.rn.ftz.f32 %f769, %f32, %f768, %f767;
|
|
div.approx.ftz.f32 %f770, %f1830, %f1836;
|
|
fma.rn.ftz.f32 %f771, %f34, %f770, %f769;
|
|
div.approx.ftz.f32 %f772, %f753, %f1836;
|
|
mul.ftz.f32 %f773, %f33, %f772;
|
|
div.approx.ftz.f32 %f774, %f750, %f1836;
|
|
fma.rn.ftz.f32 %f775, %f32, %f774, %f773;
|
|
div.approx.ftz.f32 %f776, %f756, %f1836;
|
|
fma.rn.ftz.f32 %f777, %f34, %f776, %f775;
|
|
div.approx.ftz.f32 %f778, %f762, %f1836;
|
|
mul.ftz.f32 %f779, %f33, %f778;
|
|
div.approx.ftz.f32 %f780, %f759, %f1836;
|
|
fma.rn.ftz.f32 %f781, %f32, %f780, %f779;
|
|
div.approx.ftz.f32 %f782, %f765, %f1836;
|
|
fma.rn.ftz.f32 %f783, %f34, %f782, %f781;
|
|
neg.ftz.f32 %f784, %f771;
|
|
cvta.to.local.u64 %rd174, %rd30;
|
|
st.local.v4.f32 [%rd174], {%f768, %f766, %f770, %f784};
|
|
neg.ftz.f32 %f785, %f777;
|
|
cvta.to.local.u64 %rd175, %rd44;
|
|
st.local.v4.f32 [%rd175], {%f774, %f772, %f776, %f785};
|
|
neg.ftz.f32 %f786, %f783;
|
|
cvta.to.local.u64 %rd176, %rd43;
|
|
st.local.v4.f32 [%rd176], {%f780, %f778, %f782, %f786};
|
|
|
|
$L__BB12_23:
|
|
ld.local.f32 %f65, [%rd4+12];
|
|
neg.ftz.f32 %f66, %f65;
|
|
ld.local.f32 %f67, [%rd4+16];
|
|
neg.ftz.f32 %f68, %f67;
|
|
ld.local.f32 %f1898, [%rd4+20];
|
|
neg.ftz.f32 %f787, %f1898;
|
|
st.local.f32 [%rd14], %f66;
|
|
st.local.f32 [%rd16], %f68;
|
|
st.local.f32 [%rd18], %f787;
|
|
and.b32 %r243, %r145, 6;
|
|
setp.eq.s32 %p35, %r243, 0;
|
|
@%p35 bra $L__BB12_63;
|
|
|
|
mov.b32 %f1852, %r140;
|
|
ld.local.f32 %f1856, [%rd4];
|
|
ld.local.f32 %f1857, [%rd4+4];
|
|
ld.local.f32 %f1858, [%rd4+8];
|
|
and.b32 %r58, %r369, 4;
|
|
setp.ne.s32 %p36, %r58, 0;
|
|
mov.f32 %f1853, %f65;
|
|
mov.f32 %f1854, %f67;
|
|
mov.f32 %f1855, %f1898;
|
|
@%p36 bra $L__BB12_31;
|
|
|
|
@%p29 bra $L__BB12_27;
|
|
bra.uni $L__BB12_26;
|
|
|
|
$L__BB12_27:
|
|
ld.global.v4.f32 {%f1837, %f1838, %f1839, %f1840}, [%rd61+48];
|
|
ld.global.v4.f32 {%f1841, %f1842, %f1843, %f1844}, [%rd61+64];
|
|
ld.global.v4.f32 {%f1845, %f1846, %f1847, %f1848}, [%rd61+80];
|
|
bra.uni $L__BB12_28;
|
|
|
|
$L__BB12_63:
|
|
and.b32 %r270, %r145, 8;
|
|
setp.eq.s32 %p60, %r270, 0;
|
|
@%p60 bra $L__BB12_84;
|
|
bra.uni $L__BB12_64;
|
|
|
|
$L__BB12_84:
|
|
setp.eq.s32 %p73, %r145, 1;
|
|
@%p73 bra $L__BB12_86;
|
|
bra.uni $L__BB12_85;
|
|
|
|
$L__BB12_86:
|
|
cvta.to.global.u64 %rd433, %rd152;
|
|
mul.wide.s32 %rd236, %r143, 16;
|
|
add.s64 %rd237, %rd433, %rd236;
|
|
ld.global.u32 %r287, [%rd237+12];
|
|
mul.wide.u32 %rd238, %r287, 12;
|
|
add.s64 %rd239, %rd46, %rd238;
|
|
add.s32 %r288, %r287, 1;
|
|
mul.wide.u32 %rd240, %r288, 12;
|
|
add.s64 %rd241, %rd46, %rd240;
|
|
add.s32 %r289, %r287, 2;
|
|
mul.wide.u32 %rd242, %r289, 12;
|
|
add.s64 %rd243, %rd46, %rd242;
|
|
and.b32 %r290, %r369, 8;
|
|
setp.eq.s32 %p74, %r290, 0;
|
|
ld.global.f32 %f1344, [%rd243];
|
|
ld.global.f32 %f350, [%rd239];
|
|
sub.ftz.f32 %f351, %f1344, %f350;
|
|
ld.global.f32 %f1345, [%rd243+4];
|
|
ld.global.f32 %f352, [%rd239+4];
|
|
sub.ftz.f32 %f353, %f1345, %f352;
|
|
ld.global.f32 %f1346, [%rd243+8];
|
|
ld.global.f32 %f354, [%rd239+8];
|
|
sub.ftz.f32 %f355, %f1346, %f354;
|
|
ld.global.f32 %f1347, [%rd241];
|
|
sub.ftz.f32 %f356, %f1347, %f350;
|
|
ld.global.f32 %f1348, [%rd241+4];
|
|
sub.ftz.f32 %f357, %f1348, %f352;
|
|
ld.global.f32 %f1349, [%rd241+8];
|
|
sub.ftz.f32 %f358, %f1349, %f354;
|
|
mul.ftz.f32 %f359, %f358, %f353;
|
|
mul.ftz.f32 %f360, %f357, %f355;
|
|
mul.ftz.f32 %f361, %f356, %f355;
|
|
mul.ftz.f32 %f362, %f358, %f351;
|
|
mul.ftz.f32 %f363, %f357, %f351;
|
|
mul.ftz.f32 %f364, %f356, %f353;
|
|
@%p74 bra $L__BB12_88;
|
|
bra.uni $L__BB12_87;
|
|
|
|
$L__BB12_88:
|
|
sub.ftz.f32 %f1357, %f360, %f359;
|
|
sub.ftz.f32 %f1358, %f362, %f361;
|
|
mul.ftz.f32 %f1359, %f1358, %f1358;
|
|
fma.rn.ftz.f32 %f1360, %f1357, %f1357, %f1359;
|
|
sub.ftz.f32 %f1361, %f364, %f363;
|
|
fma.rn.ftz.f32 %f1362, %f1361, %f1361, %f1360;
|
|
sqrt.approx.ftz.f32 %f1363, %f1362;
|
|
div.approx.ftz.f32 %f1923, %f1361, %f1363;
|
|
div.approx.ftz.f32 %f1922, %f1358, %f1363;
|
|
div.approx.ftz.f32 %f1921, %f1357, %f1363;
|
|
bra.uni $L__BB12_89;
|
|
|
|
$L__BB12_64:
|
|
mov.b32 %f1810, %r140;
|
|
cvt.s64.s32 %rd70, %r143;
|
|
mul.wide.s32 %rd216, %r143, 4;
|
|
add.s64 %rd217, %rd31, %rd216;
|
|
ld.global.u32 %r271, [%rd217];
|
|
st.local.u32 [%rd24], %r271;
|
|
ld.local.f32 %f1242, [%rd4];
|
|
fma.rn.ftz.f32 %f313, %f65, %f1810, %f1242;
|
|
ld.local.f32 %f1243, [%rd4+4];
|
|
fma.rn.ftz.f32 %f314, %f67, %f1810, %f1243;
|
|
ld.local.f32 %f1244, [%rd4+8];
|
|
fma.rn.ftz.f32 %f315, %f1898, %f1810, %f1244;
|
|
st.local.v2.f32 [%rd60], {%f313, %f314};
|
|
st.local.f32 [%rd32], %f315;
|
|
and.b32 %r272, %r145, 64;
|
|
setp.eq.s32 %p61, %r272, 0;
|
|
@%p61 bra $L__BB12_78;
|
|
bra.uni $L__BB12_65;
|
|
|
|
$L__BB12_78:
|
|
shl.b64 %rd228, %rd70, 4;
|
|
add.s64 %rd229, %rd47, %rd228;
|
|
ld.global.v4.f32 {%f1918, %f1919, %f1920, %f1269}, [%rd229];
|
|
bra.uni $L__BB12_79;
|
|
|
|
$L__BB12_26:
|
|
cvta.to.local.u64 %rd177, %rd30;
|
|
ld.local.v4.f32 {%f1837, %f1838, %f1839, %f1840}, [%rd177];
|
|
cvta.to.local.u64 %rd178, %rd44;
|
|
ld.local.v4.f32 {%f1841, %f1842, %f1843, %f1844}, [%rd178];
|
|
cvta.to.local.u64 %rd179, %rd43;
|
|
ld.local.v4.f32 {%f1845, %f1846, %f1847, %f1848}, [%rd179];
|
|
|
|
$L__BB12_28:
|
|
mov.b32 %f1809, %r140;
|
|
mul.ftz.f32 %f812, %f1857, %f1838;
|
|
fma.rn.ftz.f32 %f813, %f1856, %f1837, %f812;
|
|
fma.rn.ftz.f32 %f814, %f1858, %f1839, %f813;
|
|
add.ftz.f32 %f109, %f1840, %f814;
|
|
mul.ftz.f32 %f815, %f1857, %f1842;
|
|
fma.rn.ftz.f32 %f816, %f1856, %f1841, %f815;
|
|
fma.rn.ftz.f32 %f817, %f1858, %f1843, %f816;
|
|
add.ftz.f32 %f110, %f1844, %f817;
|
|
mul.ftz.f32 %f818, %f1857, %f1846;
|
|
fma.rn.ftz.f32 %f819, %f1856, %f1845, %f818;
|
|
fma.rn.ftz.f32 %f820, %f1858, %f1847, %f819;
|
|
add.ftz.f32 %f1858, %f1848, %f820;
|
|
mul.ftz.f32 %f821, %f65, %f1809;
|
|
mul.ftz.f32 %f822, %f67, %f1809;
|
|
mul.ftz.f32 %f823, %f822, %f1838;
|
|
fma.rn.ftz.f32 %f824, %f821, %f1837, %f823;
|
|
mul.ftz.f32 %f825, %f1898, %f1809;
|
|
fma.rn.ftz.f32 %f1853, %f825, %f1839, %f824;
|
|
mul.ftz.f32 %f826, %f822, %f1842;
|
|
fma.rn.ftz.f32 %f827, %f821, %f1841, %f826;
|
|
fma.rn.ftz.f32 %f1854, %f825, %f1843, %f827;
|
|
mul.ftz.f32 %f828, %f822, %f1846;
|
|
fma.rn.ftz.f32 %f829, %f821, %f1845, %f828;
|
|
fma.rn.ftz.f32 %f1855, %f825, %f1847, %f829;
|
|
mul.ftz.f32 %f830, %f1854, %f1854;
|
|
fma.rn.ftz.f32 %f831, %f1853, %f1853, %f830;
|
|
fma.rn.ftz.f32 %f832, %f1855, %f1855, %f831;
|
|
sqrt.approx.ftz.f32 %f1852, %f832;
|
|
setp.neu.ftz.f32 %p38, %f1852, 0f00000000;
|
|
@%p38 bra $L__BB12_29;
|
|
bra.uni $L__BB12_30;
|
|
|
|
$L__BB12_29:
|
|
div.approx.ftz.f32 %f1855, %f1855, %f1852;
|
|
div.approx.ftz.f32 %f1854, %f1854, %f1852;
|
|
div.approx.ftz.f32 %f1853, %f1853, %f1852;
|
|
|
|
$L__BB12_30:
|
|
mov.f32 %f1856, %f109;
|
|
mov.f32 %f1857, %f110;
|
|
|
|
$L__BB12_31:
|
|
cvta.to.global.u64 %rd430, %rd148;
|
|
mul.wide.s32 %rd180, %r143, 16;
|
|
add.s64 %rd181, %rd430, %rd180;
|
|
shr.s32 %r244, %r145, 7;
|
|
ld.global.u32 %r245, [%rd181+4];
|
|
add.s32 %r59, %r244, %r245;
|
|
add.s32 %r246, %r59, -1;
|
|
max.s32 %r60, %r246, %r245;
|
|
add.s32 %r247, %r59, 2;
|
|
ld.global.u32 %r248, [%rd181+8];
|
|
add.s32 %r249, %r245, %r248;
|
|
add.s32 %r250, %r249, -1;
|
|
min.s32 %r61, %r247, %r250;
|
|
and.b32 %r251, %r145, 64;
|
|
setp.eq.s32 %p39, %r251, 0;
|
|
mul.wide.s32 %rd182, %r60, 16;
|
|
add.s64 %rd64, %rd49, %rd182;
|
|
mul.wide.s32 %rd183, %r59, 16;
|
|
add.s64 %rd65, %rd49, %rd183;
|
|
mul.wide.s32 %rd184, %r61, 16;
|
|
add.s64 %rd66, %rd49, %rd184;
|
|
@%p39 bra $L__BB12_45;
|
|
bra.uni $L__BB12_32;
|
|
|
|
$L__BB12_45:
|
|
ld.global.v4.f32 {%f1883, %f1884, %f1885, %f917}, [%rd64];
|
|
ld.global.v4.f32 {%f1886, %f1887, %f1888, %f921}, [%rd65];
|
|
ld.global.v4.f32 {%f1889, %f1890, %f1891, %f925}, [%rd65+16];
|
|
ld.global.v4.f32 {%f1892, %f1893, %f1894, %f929}, [%rd66];
|
|
bra.uni $L__BB12_46;
|
|
|
|
$L__BB12_32:
|
|
ld.global.u32 %r62, [%rd61+148];
|
|
ld.global.u32 %r63, [%rd61+152];
|
|
shl.b32 %r252, %r63, 1;
|
|
cvt.rn.f32.s32 %f833, %r252;
|
|
mul.ftz.f32 %f135, %f8, %f833;
|
|
cvt.rzi.ftz.s32.f32 %r253, %f135;
|
|
add.s32 %r254, %r252, -1;
|
|
min.s32 %r64, %r253, %r254;
|
|
ld.global.u32 %r360, [%rd61+164];
|
|
mul.wide.u32 %rd185, %r360, 16;
|
|
add.s64 %rd186, %rd48, %rd185;
|
|
ld.global.u16 %rs42, [%rd186+12];
|
|
ld.global.u32 %r359, [%rd186+8];
|
|
ld.global.u64 %rd437, [%rd186];
|
|
setp.eq.s64 %p40, %rd437, 11;
|
|
@%p40 bra $L__BB12_37;
|
|
|
|
$L__BB12_33:
|
|
setp.eq.s64 %p41, %rd437, 0;
|
|
@%p41 bra $L__BB12_35;
|
|
|
|
add.s32 %r360, %r360, 2;
|
|
bra.uni $L__BB12_36;
|
|
|
|
$L__BB12_35:
|
|
setp.eq.s16 %p42, %rs42, 0;
|
|
mov.u32 %r360, %r359;
|
|
mov.u32 %r363, %r126;
|
|
@%p42 bra $L__BB12_38;
|
|
|
|
$L__BB12_36:
|
|
mul.wide.u32 %rd187, %r360, 16;
|
|
add.s64 %rd188, %rd48, %rd187;
|
|
ld.global.u16 %rs42, [%rd188+12];
|
|
ld.global.u32 %r359, [%rd188+8];
|
|
ld.global.u64 %rd437, [%rd188];
|
|
setp.ne.s64 %p43, %rd437, 11;
|
|
@%p43 bra $L__BB12_33;
|
|
|
|
$L__BB12_37:
|
|
setp.eq.s16 %p44, %rs42, 0;
|
|
selp.b32 %r363, -1, %r359, %p44;
|
|
|
|
$L__BB12_38:
|
|
setp.eq.s32 %p45, %r64, %r63;
|
|
@%p45 bra $L__BB12_40;
|
|
bra.uni $L__BB12_39;
|
|
|
|
$L__BB12_40:
|
|
ld.global.v4.f32 {%f1859, %f1860, %f1861, %f853}, [%rd64];
|
|
ld.global.v4.f32 {%f1862, %f1863, %f1864, %f857}, [%rd65];
|
|
ld.global.v4.f32 {%f1865, %f1866, %f1867, %f861}, [%rd65+16];
|
|
ld.global.v4.f32 {%f1868, %f1869, %f1870, %f865}, [%rd66];
|
|
bra.uni $L__BB12_41;
|
|
|
|
$L__BB12_39:
|
|
shr.s32 %r350, %r145, 7;
|
|
add.s32 %r349, %r350, %r245;
|
|
setp.gt.s32 %p46, %r64, %r63;
|
|
selp.b32 %r256, -1, 0, %p46;
|
|
add.s32 %r257, %r64, %r256;
|
|
mad.lo.s32 %r258, %r257, %r62, %r363;
|
|
add.s32 %r259, %r258, %r60;
|
|
mul.wide.s32 %rd189, %r259, 16;
|
|
add.s64 %rd190, %rd57, %rd189;
|
|
ld.global.v4.f32 {%f1859, %f1860, %f1861, %f837}, [%rd190];
|
|
add.s32 %r260, %r258, %r349;
|
|
mul.wide.s32 %rd191, %r260, 16;
|
|
add.s64 %rd192, %rd57, %rd191;
|
|
ld.global.v4.f32 {%f1862, %f1863, %f1864, %f841}, [%rd192];
|
|
ld.global.v4.f32 {%f1865, %f1866, %f1867, %f845}, [%rd192+16];
|
|
add.s32 %r261, %r258, %r61;
|
|
mul.wide.s32 %rd193, %r261, 16;
|
|
add.s64 %rd194, %rd57, %rd193;
|
|
ld.global.v4.f32 {%f1868, %f1869, %f1870, %f849}, [%rd194];
|
|
|
|
$L__BB12_41:
|
|
add.s32 %r75, %r64, 1;
|
|
setp.eq.s32 %p47, %r75, %r63;
|
|
@%p47 bra $L__BB12_43;
|
|
bra.uni $L__BB12_42;
|
|
|
|
$L__BB12_43:
|
|
ld.global.v4.f32 {%f1871, %f1872, %f1873, %f885}, [%rd64];
|
|
ld.global.v4.f32 {%f1874, %f1875, %f1876, %f889}, [%rd65];
|
|
ld.global.v4.f32 {%f1877, %f1878, %f1879, %f893}, [%rd65+16];
|
|
ld.global.v4.f32 {%f1880, %f1881, %f1882, %f897}, [%rd66];
|
|
bra.uni $L__BB12_44;
|
|
|
|
$L__BB12_42:
|
|
shr.s32 %r352, %r145, 7;
|
|
add.s32 %r351, %r352, %r245;
|
|
setp.ge.s32 %p48, %r64, %r63;
|
|
selp.b32 %r262, -1, 0, %p48;
|
|
add.s32 %r263, %r75, %r262;
|
|
mad.lo.s32 %r264, %r263, %r62, %r363;
|
|
add.s32 %r265, %r264, %r60;
|
|
mul.wide.s32 %rd195, %r265, 16;
|
|
add.s64 %rd196, %rd57, %rd195;
|
|
ld.global.v4.f32 {%f1871, %f1872, %f1873, %f869}, [%rd196];
|
|
add.s32 %r266, %r264, %r351;
|
|
mul.wide.s32 %rd197, %r266, 16;
|
|
add.s64 %rd198, %rd57, %rd197;
|
|
ld.global.v4.f32 {%f1874, %f1875, %f1876, %f873}, [%rd198];
|
|
ld.global.v4.f32 {%f1877, %f1878, %f1879, %f877}, [%rd198+16];
|
|
add.s32 %r267, %r264, %r61;
|
|
mul.wide.s32 %rd199, %r267, 16;
|
|
add.s64 %rd200, %rd57, %rd199;
|
|
ld.global.v4.f32 {%f1880, %f1881, %f1882, %f881}, [%rd200];
|
|
|
|
$L__BB12_44:
|
|
shl.b32 %r353, %r63, 1;
|
|
cvt.rn.f32.s32 %f1818, %r353;
|
|
mul.ftz.f32 %f1817, %f8, %f1818;
|
|
cvt.rn.f32.s32 %f898, %r64;
|
|
sub.ftz.f32 %f899, %f1817, %f898;
|
|
mov.f32 %f900, 0f3F800000;
|
|
sub.ftz.f32 %f901, %f900, %f899;
|
|
mul.ftz.f32 %f902, %f899, %f1871;
|
|
fma.rn.ftz.f32 %f1883, %f901, %f1859, %f902;
|
|
mul.ftz.f32 %f903, %f899, %f1872;
|
|
fma.rn.ftz.f32 %f1884, %f901, %f1860, %f903;
|
|
mul.ftz.f32 %f904, %f899, %f1873;
|
|
fma.rn.ftz.f32 %f1885, %f901, %f1861, %f904;
|
|
mul.ftz.f32 %f905, %f899, %f1874;
|
|
fma.rn.ftz.f32 %f1886, %f901, %f1862, %f905;
|
|
mul.ftz.f32 %f906, %f899, %f1875;
|
|
fma.rn.ftz.f32 %f1887, %f901, %f1863, %f906;
|
|
mul.ftz.f32 %f907, %f899, %f1876;
|
|
fma.rn.ftz.f32 %f1888, %f901, %f1864, %f907;
|
|
mul.ftz.f32 %f908, %f899, %f1877;
|
|
fma.rn.ftz.f32 %f1889, %f901, %f1865, %f908;
|
|
mul.ftz.f32 %f909, %f899, %f1878;
|
|
fma.rn.ftz.f32 %f1890, %f901, %f1866, %f909;
|
|
mul.ftz.f32 %f910, %f899, %f1879;
|
|
fma.rn.ftz.f32 %f1891, %f901, %f1867, %f910;
|
|
mul.ftz.f32 %f911, %f899, %f1880;
|
|
fma.rn.ftz.f32 %f1892, %f901, %f1868, %f911;
|
|
mul.ftz.f32 %f912, %f899, %f1881;
|
|
fma.rn.ftz.f32 %f1893, %f901, %f1869, %f912;
|
|
mul.ftz.f32 %f913, %f899, %f1882;
|
|
fma.rn.ftz.f32 %f1894, %f901, %f1870, %f913;
|
|
|
|
$L__BB12_46:
|
|
mov.b32 %f1808, %r141;
|
|
fma.rn.ftz.f32 %f1912, %f1852, %f1853, %f1856;
|
|
fma.rn.ftz.f32 %f1913, %f1852, %f1854, %f1857;
|
|
fma.rn.ftz.f32 %f1914, %f1852, %f1855, %f1858;
|
|
mov.f32 %f930, 0f3F800000;
|
|
sub.ftz.f32 %f247, %f930, %f1808;
|
|
mul.ftz.f32 %f248, %f247, %f247;
|
|
add.ftz.f32 %f931, %f247, %f247;
|
|
mul.ftz.f32 %f932, %f931, %f1808;
|
|
sub.ftz.f32 %f933, %f932, %f248;
|
|
mul.ftz.f32 %f934, %f1808, 0f40400000;
|
|
add.ftz.f32 %f249, %f934, 0fC0A00000;
|
|
add.ftz.f32 %f935, %f1808, %f1808;
|
|
mul.ftz.f32 %f936, %f935, %f249;
|
|
fma.rn.ftz.f32 %f937, %f934, %f1808, %f936;
|
|
add.ftz.f32 %f938, %f934, 0f40000000;
|
|
mul.ftz.f32 %f939, %f931, %f938;
|
|
mul.ftz.f32 %f250, %f247, 0f40400000;
|
|
mul.ftz.f32 %f940, %f247, %f250;
|
|
sub.ftz.f32 %f941, %f939, %f940;
|
|
mul.ftz.f32 %f942, %f247, 0fC0000000;
|
|
mul.ftz.f32 %f251, %f1808, %f1808;
|
|
fma.rn.ftz.f32 %f943, %f942, %f1808, %f251;
|
|
mul.ftz.f32 %f944, %f1886, %f937;
|
|
mul.ftz.f32 %f945, %f1887, %f937;
|
|
mul.ftz.f32 %f946, %f1888, %f937;
|
|
fma.rn.ftz.f32 %f947, %f1883, %f933, %f944;
|
|
fma.rn.ftz.f32 %f948, %f1884, %f933, %f945;
|
|
fma.rn.ftz.f32 %f949, %f1885, %f933, %f946;
|
|
fma.rn.ftz.f32 %f950, %f1889, %f941, %f947;
|
|
fma.rn.ftz.f32 %f951, %f1890, %f941, %f948;
|
|
fma.rn.ftz.f32 %f952, %f1891, %f941, %f949;
|
|
fma.rn.ftz.f32 %f953, %f1892, %f943, %f950;
|
|
fma.rn.ftz.f32 %f954, %f1893, %f943, %f951;
|
|
fma.rn.ftz.f32 %f955, %f1894, %f943, %f952;
|
|
mul.ftz.f32 %f1907, %f953, 0f3F000000;
|
|
mul.ftz.f32 %f1909, %f954, 0f3F000000;
|
|
mul.ftz.f32 %f1908, %f955, 0f3F000000;
|
|
and.b32 %r76, %r145, 4;
|
|
setp.eq.s32 %p49, %r76, 0;
|
|
@%p49 bra $L__BB12_48;
|
|
bra.uni $L__BB12_47;
|
|
|
|
$L__BB12_48:
|
|
mov.b32 %f1813, %r141;
|
|
mul.ftz.f32 %f1014, %f247, %f1813;
|
|
mul.ftz.f32 %f1015, %f247, %f1014;
|
|
fma.rn.ftz.f32 %f1016, %f251, %f249, 0f40000000;
|
|
add.ftz.f32 %f1017, %f250, 0fC0A00000;
|
|
fma.rn.ftz.f32 %f1018, %f248, %f1017, 0f40000000;
|
|
mul.ftz.f32 %f1019, %f1014, %f1813;
|
|
mul.ftz.f32 %f1020, %f1883, %f1015;
|
|
mul.ftz.f32 %f1021, %f1884, %f1015;
|
|
mul.ftz.f32 %f1022, %f1885, %f1015;
|
|
mul.ftz.f32 %f1023, %f1886, %f1016;
|
|
mul.ftz.f32 %f1024, %f1887, %f1016;
|
|
mul.ftz.f32 %f1025, %f1888, %f1016;
|
|
sub.ftz.f32 %f1026, %f1023, %f1020;
|
|
sub.ftz.f32 %f1027, %f1024, %f1021;
|
|
sub.ftz.f32 %f1028, %f1025, %f1022;
|
|
fma.rn.ftz.f32 %f1029, %f1889, %f1018, %f1026;
|
|
fma.rn.ftz.f32 %f1030, %f1890, %f1018, %f1027;
|
|
fma.rn.ftz.f32 %f1031, %f1891, %f1018, %f1028;
|
|
mul.ftz.f32 %f1032, %f1892, %f1019;
|
|
mul.ftz.f32 %f1033, %f1893, %f1019;
|
|
mul.ftz.f32 %f1034, %f1894, %f1019;
|
|
sub.ftz.f32 %f1035, %f1029, %f1032;
|
|
sub.ftz.f32 %f1036, %f1030, %f1033;
|
|
sub.ftz.f32 %f1037, %f1031, %f1034;
|
|
mul.ftz.f32 %f258, %f1035, 0f3F000000;
|
|
mul.ftz.f32 %f259, %f1036, 0f3F000000;
|
|
mul.ftz.f32 %f260, %f1037, 0f3F000000;
|
|
setp.eq.ftz.f32 %p50, %f1912, %f258;
|
|
setp.eq.ftz.f32 %p51, %f1913, %f259;
|
|
and.pred %p52, %p50, %p51;
|
|
setp.eq.ftz.f32 %p53, %f1914, %f260;
|
|
and.pred %p54, %p53, %p52;
|
|
@%p54 bra $L__BB12_50;
|
|
bra.uni $L__BB12_49;
|
|
|
|
$L__BB12_50:
|
|
mov.f32 %f1911, %f65;
|
|
mov.f32 %f1910, %f67;
|
|
bra.uni $L__BB12_51;
|
|
|
|
$L__BB12_47:
|
|
mov.b32 %f1814, %r142;
|
|
mul.ftz.f32 %f956, %f1909, %f1909;
|
|
fma.rn.ftz.f32 %f957, %f1907, %f1907, %f956;
|
|
fma.rn.ftz.f32 %f958, %f1908, %f1908, %f957;
|
|
sqrt.approx.ftz.f32 %f959, %f958;
|
|
div.approx.ftz.f32 %f960, %f1907, %f959;
|
|
div.approx.ftz.f32 %f961, %f1909, %f959;
|
|
div.approx.ftz.f32 %f962, %f1908, %f959;
|
|
mul.ftz.f32 %f963, %f961, %f1855;
|
|
mul.ftz.f32 %f964, %f962, %f1854;
|
|
sub.ftz.f32 %f965, %f964, %f963;
|
|
mul.ftz.f32 %f966, %f962, %f1853;
|
|
mul.ftz.f32 %f967, %f960, %f1855;
|
|
sub.ftz.f32 %f968, %f967, %f966;
|
|
mul.ftz.f32 %f969, %f960, %f1854;
|
|
mul.ftz.f32 %f970, %f961, %f1853;
|
|
sub.ftz.f32 %f971, %f970, %f969;
|
|
mul.ftz.f32 %f972, %f968, %f968;
|
|
fma.rn.ftz.f32 %f973, %f965, %f965, %f972;
|
|
fma.rn.ftz.f32 %f974, %f971, %f971, %f973;
|
|
sqrt.approx.ftz.f32 %f975, %f974;
|
|
div.approx.ftz.f32 %f976, %f965, %f975;
|
|
div.approx.ftz.f32 %f977, %f968, %f975;
|
|
div.approx.ftz.f32 %f978, %f971, %f975;
|
|
mul.ftz.f32 %f979, %f1814, %f1814;
|
|
sub.ftz.f32 %f981, %f930, %f979;
|
|
mov.f32 %f982, 0f00000000;
|
|
max.ftz.f32 %f983, %f981, %f982;
|
|
sqrt.approx.ftz.f32 %f984, %f983;
|
|
mul.ftz.f32 %f985, %f976, %f1814;
|
|
mul.ftz.f32 %f986, %f977, %f1814;
|
|
mul.ftz.f32 %f987, %f978, %f1814;
|
|
mul.ftz.f32 %f988, %f961, %f978;
|
|
mul.ftz.f32 %f989, %f962, %f977;
|
|
sub.ftz.f32 %f990, %f988, %f989;
|
|
mul.ftz.f32 %f991, %f962, %f976;
|
|
mul.ftz.f32 %f992, %f960, %f978;
|
|
sub.ftz.f32 %f993, %f991, %f992;
|
|
mul.ftz.f32 %f994, %f960, %f977;
|
|
mul.ftz.f32 %f995, %f961, %f976;
|
|
sub.ftz.f32 %f996, %f994, %f995;
|
|
mul.ftz.f32 %f997, %f993, %f993;
|
|
fma.rn.ftz.f32 %f998, %f990, %f990, %f997;
|
|
fma.rn.ftz.f32 %f999, %f996, %f996, %f998;
|
|
sqrt.approx.ftz.f32 %f1000, %f999;
|
|
div.approx.ftz.f32 %f1001, %f990, %f1000;
|
|
div.approx.ftz.f32 %f1002, %f993, %f1000;
|
|
div.approx.ftz.f32 %f1003, %f996, %f1000;
|
|
mul.ftz.f32 %f1004, %f984, %f1001;
|
|
mul.ftz.f32 %f1005, %f984, %f1002;
|
|
mul.ftz.f32 %f1006, %f984, %f1003;
|
|
sub.ftz.f32 %f1007, %f985, %f1004;
|
|
sub.ftz.f32 %f1008, %f986, %f1005;
|
|
sub.ftz.f32 %f1009, %f987, %f1006;
|
|
mul.ftz.f32 %f1010, %f1008, %f1008;
|
|
fma.rn.ftz.f32 %f1011, %f1007, %f1007, %f1010;
|
|
fma.rn.ftz.f32 %f1012, %f1009, %f1009, %f1011;
|
|
sqrt.approx.ftz.f32 %f1013, %f1012;
|
|
div.approx.ftz.f32 %f1911, %f1007, %f1013;
|
|
div.approx.ftz.f32 %f1910, %f1008, %f1013;
|
|
div.approx.ftz.f32 %f1898, %f1009, %f1013;
|
|
st.local.f32 [%rd33], %f1911;
|
|
st.local.f32 [%rd54], %f1910;
|
|
st.local.f32 [%rd55], %f1898;
|
|
bra.uni $L__BB12_52;
|
|
|
|
$L__BB12_49:
|
|
sub.ftz.f32 %f1038, %f1912, %f258;
|
|
sub.ftz.f32 %f1039, %f1913, %f259;
|
|
mul.ftz.f32 %f1040, %f1039, %f1039;
|
|
fma.rn.ftz.f32 %f1041, %f1038, %f1038, %f1040;
|
|
sub.ftz.f32 %f1042, %f1914, %f260;
|
|
fma.rn.ftz.f32 %f1043, %f1042, %f1042, %f1041;
|
|
sqrt.approx.ftz.f32 %f1044, %f1043;
|
|
div.approx.ftz.f32 %f1898, %f1042, %f1044;
|
|
div.approx.ftz.f32 %f1910, %f1039, %f1044;
|
|
div.approx.ftz.f32 %f1911, %f1038, %f1044;
|
|
|
|
$L__BB12_51:
|
|
mov.u32 %r348, 0;
|
|
st.local.f32 [%rd33], %f1911;
|
|
st.local.f32 [%rd54], %f1910;
|
|
st.local.f32 [%rd55], %f1898;
|
|
st.local.u32 [%rd58], %r348;
|
|
|
|
$L__BB12_52:
|
|
and.b32 %r341, %r369, 4;
|
|
setp.ne.s32 %p132, %r341, 0;
|
|
st.local.f32 [%rd36], %f1907;
|
|
st.local.f32 [%rd37], %f1909;
|
|
st.local.f32 [%rd38], %f1908;
|
|
@%p132 bra $L__BB12_62;
|
|
|
|
@%p29 bra $L__BB12_55;
|
|
bra.uni $L__BB12_54;
|
|
|
|
$L__BB12_55:
|
|
ld.global.v4.f32 {%f1078, %f1079, %f1080, %f1081}, [%rd61];
|
|
ld.global.v4.f32 {%f1086, %f1087, %f1088, %f1089}, [%rd61+16];
|
|
ld.global.v4.f32 {%f1094, %f1095, %f1096, %f1097}, [%rd61+32];
|
|
mul.ftz.f32 %f1102, %f1913, %f1079;
|
|
fma.rn.ftz.f32 %f1103, %f1912, %f1078, %f1102;
|
|
fma.rn.ftz.f32 %f1104, %f1914, %f1080, %f1103;
|
|
add.ftz.f32 %f1901, %f1081, %f1104;
|
|
mul.ftz.f32 %f1105, %f1913, %f1087;
|
|
fma.rn.ftz.f32 %f1106, %f1912, %f1086, %f1105;
|
|
fma.rn.ftz.f32 %f1107, %f1914, %f1088, %f1106;
|
|
add.ftz.f32 %f1902, %f1089, %f1107;
|
|
mul.ftz.f32 %f1108, %f1913, %f1095;
|
|
fma.rn.ftz.f32 %f1109, %f1912, %f1094, %f1108;
|
|
fma.rn.ftz.f32 %f1110, %f1914, %f1096, %f1109;
|
|
add.ftz.f32 %f1914, %f1097, %f1110;
|
|
bra.uni $L__BB12_56;
|
|
|
|
$L__BB12_54:
|
|
cvta.to.local.u64 %rd201, %rd19;
|
|
ld.local.v4.f32 {%f1045, %f1046, %f1047, %f1048}, [%rd201];
|
|
mul.ftz.f32 %f1053, %f1913, %f1046;
|
|
fma.rn.ftz.f32 %f1054, %f1912, %f1045, %f1053;
|
|
fma.rn.ftz.f32 %f1055, %f1914, %f1047, %f1054;
|
|
add.ftz.f32 %f1901, %f1048, %f1055;
|
|
cvta.to.local.u64 %rd202, %rd29;
|
|
ld.local.v4.f32 {%f1056, %f1057, %f1058, %f1059}, [%rd202];
|
|
mul.ftz.f32 %f1064, %f1913, %f1057;
|
|
fma.rn.ftz.f32 %f1065, %f1912, %f1056, %f1064;
|
|
fma.rn.ftz.f32 %f1066, %f1914, %f1058, %f1065;
|
|
add.ftz.f32 %f1902, %f1059, %f1066;
|
|
cvta.to.local.u64 %rd203, %rd28;
|
|
ld.local.v4.f32 {%f1067, %f1068, %f1069, %f1070}, [%rd203];
|
|
mul.ftz.f32 %f1075, %f1913, %f1068;
|
|
fma.rn.ftz.f32 %f1076, %f1912, %f1067, %f1075;
|
|
fma.rn.ftz.f32 %f1077, %f1914, %f1069, %f1076;
|
|
add.ftz.f32 %f1914, %f1070, %f1077;
|
|
|
|
$L__BB12_56:
|
|
@%p29 bra $L__BB12_58;
|
|
bra.uni $L__BB12_57;
|
|
|
|
$L__BB12_58:
|
|
ld.global.v4.f32 {%f1145, %f1146, %f1147, %f1148}, [%rd61+48];
|
|
ld.global.v4.f32 {%f1152, %f1153, %f1154, %f1155}, [%rd61+64];
|
|
ld.global.v4.f32 {%f1159, %f1160, %f1161, %f1162}, [%rd61+80];
|
|
mul.ftz.f32 %f1166, %f1152, %f1910;
|
|
fma.rn.ftz.f32 %f1167, %f1145, %f1911, %f1166;
|
|
fma.rn.ftz.f32 %f1168, %f1159, %f1898, %f1167;
|
|
mul.ftz.f32 %f1169, %f1153, %f1910;
|
|
fma.rn.ftz.f32 %f1170, %f1146, %f1911, %f1169;
|
|
fma.rn.ftz.f32 %f1171, %f1160, %f1898, %f1170;
|
|
mul.ftz.f32 %f1172, %f1154, %f1910;
|
|
fma.rn.ftz.f32 %f1173, %f1147, %f1911, %f1172;
|
|
fma.rn.ftz.f32 %f1174, %f1161, %f1898, %f1173;
|
|
mul.ftz.f32 %f1175, %f1171, %f1171;
|
|
fma.rn.ftz.f32 %f1176, %f1168, %f1168, %f1175;
|
|
fma.rn.ftz.f32 %f1177, %f1174, %f1174, %f1176;
|
|
sqrt.approx.ftz.f32 %f1178, %f1177;
|
|
div.approx.ftz.f32 %f1911, %f1168, %f1178;
|
|
div.approx.ftz.f32 %f1910, %f1171, %f1178;
|
|
div.approx.ftz.f32 %f1904, %f1174, %f1178;
|
|
bra.uni $L__BB12_59;
|
|
|
|
$L__BB12_57:
|
|
cvta.to.local.u64 %rd204, %rd30;
|
|
ld.local.v4.f32 {%f1111, %f1112, %f1113, %f1114}, [%rd204];
|
|
cvta.to.local.u64 %rd205, %rd44;
|
|
ld.local.v4.f32 {%f1118, %f1119, %f1120, %f1121}, [%rd205];
|
|
cvta.to.local.u64 %rd206, %rd43;
|
|
ld.local.v4.f32 {%f1125, %f1126, %f1127, %f1128}, [%rd206];
|
|
mul.ftz.f32 %f1132, %f1910, %f1118;
|
|
fma.rn.ftz.f32 %f1133, %f1911, %f1111, %f1132;
|
|
fma.rn.ftz.f32 %f1134, %f1898, %f1125, %f1133;
|
|
mul.ftz.f32 %f1135, %f1910, %f1119;
|
|
fma.rn.ftz.f32 %f1136, %f1911, %f1112, %f1135;
|
|
fma.rn.ftz.f32 %f1137, %f1898, %f1126, %f1136;
|
|
mul.ftz.f32 %f1138, %f1910, %f1120;
|
|
fma.rn.ftz.f32 %f1139, %f1911, %f1113, %f1138;
|
|
fma.rn.ftz.f32 %f1140, %f1898, %f1127, %f1139;
|
|
mul.ftz.f32 %f1141, %f1137, %f1137;
|
|
fma.rn.ftz.f32 %f1142, %f1134, %f1134, %f1141;
|
|
fma.rn.ftz.f32 %f1143, %f1140, %f1140, %f1142;
|
|
sqrt.approx.ftz.f32 %f1144, %f1143;
|
|
div.approx.ftz.f32 %f1911, %f1134, %f1144;
|
|
div.approx.ftz.f32 %f1910, %f1137, %f1144;
|
|
div.approx.ftz.f32 %f1904, %f1140, %f1144;
|
|
|
|
$L__BB12_59:
|
|
st.local.f32 [%rd33], %f1911;
|
|
st.local.f32 [%rd54], %f1910;
|
|
st.local.f32 [%rd55], %f1904;
|
|
@%p29 bra $L__BB12_61;
|
|
bra.uni $L__BB12_60;
|
|
|
|
$L__BB12_61:
|
|
ld.global.v4.f32 {%f1206, %f1207, %f1208, %f1209}, [%rd61];
|
|
ld.global.v4.f32 {%f1213, %f1214, %f1215, %f1216}, [%rd61+16];
|
|
ld.global.v4.f32 {%f1220, %f1221, %f1222, %f1223}, [%rd61+32];
|
|
mul.ftz.f32 %f1227, %f1207, %f1909;
|
|
fma.rn.ftz.f32 %f1228, %f1907, %f1206, %f1227;
|
|
fma.rn.ftz.f32 %f297, %f1208, %f1908, %f1228;
|
|
mul.ftz.f32 %f1229, %f1214, %f1909;
|
|
fma.rn.ftz.f32 %f1230, %f1907, %f1213, %f1229;
|
|
fma.rn.ftz.f32 %f298, %f1215, %f1908, %f1230;
|
|
mul.ftz.f32 %f1231, %f1221, %f1909;
|
|
fma.rn.ftz.f32 %f1232, %f1907, %f1220, %f1231;
|
|
fma.rn.ftz.f32 %f1908, %f1222, %f1908, %f1232;
|
|
st.local.f32 [%rd36], %f297;
|
|
st.local.f32 [%rd37], %f298;
|
|
st.local.f32 [%rd38], %f1908;
|
|
mov.f32 %f1907, %f297;
|
|
mov.f32 %f1909, %f298;
|
|
mov.f32 %f1912, %f1901;
|
|
mov.f32 %f1913, %f1902;
|
|
bra.uni $L__BB12_62;
|
|
|
|
$L__BB12_60:
|
|
cvta.to.local.u64 %rd207, %rd19;
|
|
ld.local.v4.f32 {%f1179, %f1180, %f1181, %f1182}, [%rd207];
|
|
mul.ftz.f32 %f1186, %f1909, %f1180;
|
|
fma.rn.ftz.f32 %f1187, %f1907, %f1179, %f1186;
|
|
fma.rn.ftz.f32 %f294, %f1908, %f1181, %f1187;
|
|
cvta.to.local.u64 %rd208, %rd29;
|
|
ld.local.v4.f32 {%f1188, %f1189, %f1190, %f1191}, [%rd208];
|
|
mul.ftz.f32 %f1195, %f1909, %f1189;
|
|
fma.rn.ftz.f32 %f1196, %f1907, %f1188, %f1195;
|
|
fma.rn.ftz.f32 %f295, %f1908, %f1190, %f1196;
|
|
cvta.to.local.u64 %rd209, %rd28;
|
|
ld.local.v4.f32 {%f1197, %f1198, %f1199, %f1200}, [%rd209];
|
|
mul.ftz.f32 %f1204, %f1909, %f1198;
|
|
fma.rn.ftz.f32 %f1205, %f1907, %f1197, %f1204;
|
|
fma.rn.ftz.f32 %f1908, %f1908, %f1199, %f1205;
|
|
st.local.f32 [%rd36], %f294;
|
|
st.local.f32 [%rd37], %f295;
|
|
st.local.f32 [%rd38], %f1908;
|
|
mov.f32 %f1907, %f294;
|
|
mov.f32 %f1909, %f295;
|
|
mov.f32 %f1912, %f1901;
|
|
mov.f32 %f1913, %f1902;
|
|
|
|
$L__BB12_62:
|
|
add.u64 %rd423, %SP, 0;
|
|
mul.wide.s32 %rd422, %r143, 16;
|
|
cvta.to.global.u64 %rd421, %rd148;
|
|
add.s64 %rd420, %rd421, %rd422;
|
|
add.s64 %rd419, %rd420, 4;
|
|
add.s64 %rd418, %rd423, 44;
|
|
add.s64 %rd417, %rd423, 20;
|
|
neg.ftz.f32 %f1804, %f67;
|
|
neg.ftz.f32 %f1803, %f65;
|
|
st.local.v2.f32 [%rd60], {%f1912, %f1913};
|
|
st.local.f32 [%rd32], %f1914;
|
|
selp.f32 %f1951, %f1911, %f1803, %p49;
|
|
selp.f32 %f1952, %f1910, %f1804, %p49;
|
|
selp.b64 %rd212, %rd417, %rd418, %p49;
|
|
cvta.to.local.u64 %rd213, %rd212;
|
|
ld.local.f32 %f1953, [%rd213];
|
|
cvta.to.local.u64 %rd214, %rd22;
|
|
st.local.v2.f32 [%rd214], {%f1951, %f1952};
|
|
st.local.f32 [%rd25], %f1953;
|
|
mul.ftz.f32 %f1233, %f1909, %f1953;
|
|
mul.ftz.f32 %f1234, %f1908, %f1952;
|
|
mul.ftz.f32 %f1235, %f1908, %f1951;
|
|
mul.ftz.f32 %f1236, %f1907, %f1953;
|
|
mul.ftz.f32 %f1237, %f1907, %f1952;
|
|
mul.ftz.f32 %f1238, %f1909, %f1951;
|
|
sub.ftz.f32 %f1239, %f1237, %f1238;
|
|
sub.ftz.f32 %f1240, %f1235, %f1236;
|
|
sub.ftz.f32 %f1241, %f1233, %f1234;
|
|
cvta.to.local.u64 %rd215, %rd27;
|
|
st.local.v2.f32 [%rd215], {%f1241, %f1240};
|
|
st.local.f32 [%rd39], %f1239;
|
|
ld.global.u32 %r269, [%rd419+-4];
|
|
st.local.u32 [%rd24], %r269;
|
|
bra.uni $L__BB12_113;
|
|
|
|
$L__BB12_85:
|
|
add.u64 %rd431, %SP, 0;
|
|
{ // callseq 1, 0
|
|
.reg .b32 temp_param_reg;
|
|
.param .b64 param0;
|
|
st.param.b64 [param0+0], %rd431;
|
|
call.uni
|
|
_Z28motion_triangle_shader_setupPK16KernelGlobalsGPUP10ShaderData6float3S4_fiib,
|
|
(
|
|
param0
|
|
);
|
|
} // callseq 1
|
|
ld.local.u32 %r369, [%rd10];
|
|
bra.uni $L__BB12_99;
|
|
|
|
$L__BB12_65:
|
|
ld.global.u32 %r77, [%rd61+148];
|
|
ld.global.u32 %r78, [%rd61+152];
|
|
shl.b32 %r273, %r78, 1;
|
|
cvt.rn.f32.s32 %f1245, %r273;
|
|
mul.ftz.f32 %f316, %f8, %f1245;
|
|
cvt.rzi.ftz.s32.f32 %r274, %f316;
|
|
add.s32 %r275, %r273, -1;
|
|
min.s32 %r79, %r274, %r275;
|
|
ld.global.u32 %r365, [%rd61+164];
|
|
mul.wide.u32 %rd220, %r365, 16;
|
|
add.s64 %rd221, %rd48, %rd220;
|
|
ld.global.u16 %rs44, [%rd221+12];
|
|
ld.global.u32 %r364, [%rd221+8];
|
|
ld.global.u64 %rd438, [%rd221];
|
|
setp.eq.s64 %p62, %rd438, 11;
|
|
@%p62 bra $L__BB12_70;
|
|
|
|
$L__BB12_66:
|
|
setp.eq.s64 %p63, %rd438, 0;
|
|
@%p63 bra $L__BB12_68;
|
|
|
|
add.s32 %r365, %r365, 2;
|
|
bra.uni $L__BB12_69;
|
|
|
|
$L__BB12_68:
|
|
setp.eq.s16 %p64, %rs44, 0;
|
|
mov.u32 %r368, -1;
|
|
mov.u32 %r365, %r364;
|
|
@%p64 bra $L__BB12_71;
|
|
|
|
$L__BB12_69:
|
|
mul.wide.u32 %rd222, %r365, 16;
|
|
add.s64 %rd223, %rd48, %rd222;
|
|
ld.global.u16 %rs44, [%rd223+12];
|
|
ld.global.u32 %r364, [%rd223+8];
|
|
ld.global.u64 %rd438, [%rd223];
|
|
setp.ne.s64 %p65, %rd438, 11;
|
|
@%p65 bra $L__BB12_66;
|
|
|
|
$L__BB12_70:
|
|
setp.eq.s16 %p66, %rs44, 0;
|
|
selp.b32 %r368, -1, %r364, %p66;
|
|
|
|
$L__BB12_71:
|
|
setp.eq.s32 %p67, %r79, %r78;
|
|
@%p67 bra $L__BB12_73;
|
|
bra.uni $L__BB12_72;
|
|
|
|
$L__BB12_73:
|
|
shl.b64 %rd225, %rd70, 4;
|
|
add.s64 %rd439, %rd47, %rd225;
|
|
bra.uni $L__BB12_74;
|
|
|
|
$L__BB12_87:
|
|
sub.ftz.f32 %f1350, %f359, %f360;
|
|
sub.ftz.f32 %f1351, %f361, %f362;
|
|
mul.ftz.f32 %f1352, %f1351, %f1351;
|
|
fma.rn.ftz.f32 %f1353, %f1350, %f1350, %f1352;
|
|
sub.ftz.f32 %f1354, %f363, %f364;
|
|
fma.rn.ftz.f32 %f1355, %f1354, %f1354, %f1353;
|
|
sqrt.approx.ftz.f32 %f1356, %f1355;
|
|
div.approx.ftz.f32 %f1923, %f1354, %f1356;
|
|
div.approx.ftz.f32 %f1922, %f1351, %f1356;
|
|
div.approx.ftz.f32 %f1921, %f1350, %f1356;
|
|
|
|
$L__BB12_89:
|
|
cvt.s64.s32 %rd432, %r143;
|
|
mov.b32 %f1815, %r142;
|
|
mov.b32 %f1811, %r141;
|
|
shl.b64 %rd244, %rd432, 2;
|
|
add.s64 %rd245, %rd50, %rd244;
|
|
ld.global.u32 %r92, [%rd245];
|
|
st.local.u32 [%rd24], %r92;
|
|
fma.rn.ftz.f32 %f1364, %f356, %f1811, %f350;
|
|
fma.rn.ftz.f32 %f1365, %f357, %f1811, %f352;
|
|
fma.rn.ftz.f32 %f1366, %f358, %f1811, %f354;
|
|
fma.rn.ftz.f32 %f1936, %f351, %f1815, %f1364;
|
|
fma.rn.ftz.f32 %f1937, %f353, %f1815, %f1365;
|
|
fma.rn.ftz.f32 %f1938, %f355, %f1815, %f1366;
|
|
and.b32 %r291, %r369, 4;
|
|
setp.ne.s32 %p75, %r291, 0;
|
|
@%p75 bra $L__BB12_94;
|
|
|
|
@%p29 bra $L__BB12_92;
|
|
bra.uni $L__BB12_91;
|
|
|
|
$L__BB12_92:
|
|
ld.global.v4.f32 {%f1924, %f1925, %f1926, %f1927}, [%rd61];
|
|
ld.global.v4.f32 {%f1928, %f1929, %f1930, %f1931}, [%rd61+16];
|
|
ld.global.v4.f32 {%f1932, %f1933, %f1934, %f1935}, [%rd61+32];
|
|
bra.uni $L__BB12_93;
|
|
|
|
$L__BB12_72:
|
|
setp.gt.s32 %p68, %r79, %r78;
|
|
selp.b32 %r277, -1, 0, %p68;
|
|
add.s32 %r278, %r79, %r277;
|
|
mad.lo.s32 %r279, %r278, %r77, %r143;
|
|
add.s32 %r280, %r279, %r368;
|
|
mul.wide.s32 %rd224, %r280, 16;
|
|
add.s64 %rd439, %rd57, %rd224;
|
|
|
|
$L__BB12_74:
|
|
ld.global.v4.f32 {%f1246, %f1247, %f1248, %f1249}, [%rd439];
|
|
add.s32 %r90, %r79, 1;
|
|
setp.eq.s32 %p69, %r90, %r78;
|
|
@%p69 bra $L__BB12_76;
|
|
bra.uni $L__BB12_75;
|
|
|
|
$L__BB12_76:
|
|
shl.b64 %rd227, %rd70, 4;
|
|
add.s64 %rd440, %rd47, %rd227;
|
|
bra.uni $L__BB12_77;
|
|
|
|
$L__BB12_75:
|
|
setp.ge.s32 %p70, %r79, %r78;
|
|
selp.b32 %r281, -1, 0, %p70;
|
|
add.s32 %r282, %r90, %r281;
|
|
mad.lo.s32 %r283, %r282, %r77, %r143;
|
|
add.s32 %r284, %r283, %r368;
|
|
mul.wide.s32 %rd226, %r284, 16;
|
|
add.s64 %rd440, %rd57, %rd226;
|
|
|
|
$L__BB12_77:
|
|
cvt.rn.f32.s32 %f1250, %r79;
|
|
sub.ftz.f32 %f1251, %f316, %f1250;
|
|
ld.global.v4.f32 {%f1252, %f1253, %f1254, %f1255}, [%rd440];
|
|
mov.f32 %f1260, 0f3F800000;
|
|
sub.ftz.f32 %f1261, %f1260, %f1251;
|
|
mul.ftz.f32 %f1262, %f1251, %f1252;
|
|
mul.ftz.f32 %f1263, %f1251, %f1253;
|
|
mul.ftz.f32 %f1264, %f1251, %f1254;
|
|
fma.rn.ftz.f32 %f1920, %f1261, %f1248, %f1264;
|
|
fma.rn.ftz.f32 %f1919, %f1261, %f1247, %f1263;
|
|
fma.rn.ftz.f32 %f1918, %f1261, %f1246, %f1262;
|
|
|
|
$L__BB12_79:
|
|
and.b32 %r285, %r369, 4;
|
|
setp.ne.s32 %p71, %r285, 0;
|
|
@%p71 bra $L__BB12_83;
|
|
|
|
@%p29 bra $L__BB12_82;
|
|
bra.uni $L__BB12_81;
|
|
|
|
$L__BB12_82:
|
|
ld.global.v4.f32 {%f1303, %f1304, %f1305, %f1306}, [%rd61];
|
|
ld.global.v4.f32 {%f1311, %f1312, %f1313, %f1314}, [%rd61+16];
|
|
ld.global.v4.f32 {%f1319, %f1320, %f1321, %f1322}, [%rd61+32];
|
|
mul.ftz.f32 %f1327, %f1919, %f1304;
|
|
fma.rn.ftz.f32 %f1328, %f1918, %f1303, %f1327;
|
|
fma.rn.ftz.f32 %f1329, %f1920, %f1305, %f1328;
|
|
add.ftz.f32 %f339, %f1306, %f1329;
|
|
mul.ftz.f32 %f1330, %f1919, %f1312;
|
|
fma.rn.ftz.f32 %f1331, %f1918, %f1311, %f1330;
|
|
fma.rn.ftz.f32 %f1332, %f1920, %f1313, %f1331;
|
|
add.ftz.f32 %f340, %f1314, %f1332;
|
|
mul.ftz.f32 %f1333, %f1919, %f1320;
|
|
fma.rn.ftz.f32 %f1334, %f1918, %f1319, %f1333;
|
|
fma.rn.ftz.f32 %f1335, %f1920, %f1321, %f1334;
|
|
add.ftz.f32 %f1920, %f1322, %f1335;
|
|
mov.f32 %f1918, %f339;
|
|
mov.f32 %f1919, %f340;
|
|
bra.uni $L__BB12_83;
|
|
|
|
$L__BB12_81:
|
|
cvta.to.local.u64 %rd230, %rd19;
|
|
ld.local.v4.f32 {%f1270, %f1271, %f1272, %f1273}, [%rd230];
|
|
mul.ftz.f32 %f1278, %f1919, %f1271;
|
|
fma.rn.ftz.f32 %f1279, %f1918, %f1270, %f1278;
|
|
fma.rn.ftz.f32 %f1280, %f1920, %f1272, %f1279;
|
|
add.ftz.f32 %f336, %f1273, %f1280;
|
|
cvta.to.local.u64 %rd231, %rd29;
|
|
ld.local.v4.f32 {%f1281, %f1282, %f1283, %f1284}, [%rd231];
|
|
mul.ftz.f32 %f1289, %f1919, %f1282;
|
|
fma.rn.ftz.f32 %f1290, %f1918, %f1281, %f1289;
|
|
fma.rn.ftz.f32 %f1291, %f1920, %f1283, %f1290;
|
|
add.ftz.f32 %f337, %f1284, %f1291;
|
|
cvta.to.local.u64 %rd232, %rd28;
|
|
ld.local.v4.f32 {%f1292, %f1293, %f1294, %f1295}, [%rd232];
|
|
mul.ftz.f32 %f1300, %f1919, %f1293;
|
|
fma.rn.ftz.f32 %f1301, %f1918, %f1292, %f1300;
|
|
fma.rn.ftz.f32 %f1302, %f1920, %f1294, %f1301;
|
|
add.ftz.f32 %f1920, %f1295, %f1302;
|
|
mov.f32 %f1918, %f336;
|
|
mov.f32 %f1919, %f337;
|
|
|
|
$L__BB12_83:
|
|
sub.ftz.f32 %f1336, %f313, %f1918;
|
|
sub.ftz.f32 %f1337, %f314, %f1919;
|
|
mul.ftz.f32 %f1338, %f1337, %f1337;
|
|
fma.rn.ftz.f32 %f1339, %f1336, %f1336, %f1338;
|
|
sub.ftz.f32 %f1340, %f315, %f1920;
|
|
fma.rn.ftz.f32 %f1341, %f1340, %f1340, %f1339;
|
|
sqrt.approx.ftz.f32 %f1342, %f1341;
|
|
div.approx.ftz.f32 %f1951, %f1336, %f1342;
|
|
div.approx.ftz.f32 %f1952, %f1337, %f1342;
|
|
div.approx.ftz.f32 %f1953, %f1340, %f1342;
|
|
st.local.f32 [%rd25], %f1953;
|
|
st.local.f32 [%rd33], %f1951;
|
|
st.local.v4.f32 [%rd54], {%f1952, %f1953, %f1951, %f1952};
|
|
mov.u32 %r286, 0;
|
|
st.local.u32 [%rd36], %r286;
|
|
st.local.u32 [%rd37], %r286;
|
|
st.local.u32 [%rd38], %r286;
|
|
cvta.to.local.u64 %rd234, %rd27;
|
|
mov.f32 %f1343, 0f00000000;
|
|
st.local.v2.f32 [%rd234], {%f1343, %f1343};
|
|
st.local.u32 [%rd39], %r286;
|
|
bra.uni $L__BB12_113;
|
|
|
|
$L__BB12_91:
|
|
cvta.to.local.u64 %rd246, %rd19;
|
|
ld.local.v4.f32 {%f1924, %f1925, %f1926, %f1927}, [%rd246];
|
|
cvta.to.local.u64 %rd247, %rd29;
|
|
ld.local.v4.f32 {%f1928, %f1929, %f1930, %f1931}, [%rd247];
|
|
cvta.to.local.u64 %rd248, %rd28;
|
|
ld.local.v4.f32 {%f1932, %f1933, %f1934, %f1935}, [%rd248];
|
|
|
|
$L__BB12_93:
|
|
mul.ftz.f32 %f1391, %f1937, %f1925;
|
|
fma.rn.ftz.f32 %f1392, %f1936, %f1924, %f1391;
|
|
fma.rn.ftz.f32 %f1393, %f1938, %f1926, %f1392;
|
|
mul.ftz.f32 %f1394, %f1937, %f1929;
|
|
fma.rn.ftz.f32 %f1395, %f1936, %f1928, %f1394;
|
|
fma.rn.ftz.f32 %f1396, %f1938, %f1930, %f1395;
|
|
mul.ftz.f32 %f1397, %f1937, %f1933;
|
|
fma.rn.ftz.f32 %f1398, %f1936, %f1932, %f1397;
|
|
fma.rn.ftz.f32 %f1399, %f1938, %f1934, %f1398;
|
|
add.ftz.f32 %f1936, %f1927, %f1393;
|
|
add.ftz.f32 %f1937, %f1931, %f1396;
|
|
add.ftz.f32 %f1938, %f1935, %f1399;
|
|
|
|
$L__BB12_94:
|
|
st.local.f32 [%rd25], %f1923;
|
|
st.local.v4.f32 [%rd60], {%f1936, %f1937, %f1938, %f1921};
|
|
st.local.v4.f32 [%rd54], {%f1922, %f1923, %f1921, %f1922};
|
|
setp.gt.s32 %p77, %r92, -1;
|
|
@%p77 bra $L__BB12_98;
|
|
|
|
mul.wide.s32 %rd436, %r143, 16;
|
|
add.s64 %rd435, %rd433, %rd436;
|
|
add.s64 %rd434, %rd435, 12;
|
|
mov.b32 %f1816, %r142;
|
|
mov.b32 %f1812, %r141;
|
|
ld.global.v4.u32 {%r292, %r293, %r294, %r295}, [%rd434+-12];
|
|
mul.wide.u32 %rd252, %r292, 12;
|
|
add.s64 %rd253, %rd59, %rd252;
|
|
mul.wide.u32 %rd254, %r293, 12;
|
|
add.s64 %rd255, %rd59, %rd254;
|
|
mul.wide.u32 %rd256, %r294, 12;
|
|
add.s64 %rd257, %rd59, %rd256;
|
|
mov.f32 %f1400, 0f3F800000;
|
|
sub.ftz.f32 %f1401, %f1400, %f1812;
|
|
sub.ftz.f32 %f1402, %f1401, %f1816;
|
|
ld.global.f32 %f1403, [%rd253];
|
|
ld.global.f32 %f1404, [%rd253+4];
|
|
ld.global.f32 %f1405, [%rd253+8];
|
|
ld.global.f32 %f1406, [%rd255];
|
|
mul.ftz.f32 %f1407, %f1406, %f1812;
|
|
ld.global.f32 %f1408, [%rd255+4];
|
|
mul.ftz.f32 %f1409, %f1408, %f1812;
|
|
ld.global.f32 %f1410, [%rd255+8];
|
|
mul.ftz.f32 %f1411, %f1410, %f1812;
|
|
fma.rn.ftz.f32 %f1412, %f1402, %f1403, %f1407;
|
|
fma.rn.ftz.f32 %f1413, %f1402, %f1404, %f1409;
|
|
fma.rn.ftz.f32 %f1414, %f1402, %f1405, %f1411;
|
|
ld.global.f32 %f1415, [%rd257];
|
|
ld.global.f32 %f1416, [%rd257+4];
|
|
ld.global.f32 %f1417, [%rd257+8];
|
|
fma.rn.ftz.f32 %f1939, %f1415, %f1816, %f1412;
|
|
fma.rn.ftz.f32 %f1940, %f1416, %f1816, %f1413;
|
|
fma.rn.ftz.f32 %f1941, %f1417, %f1816, %f1414;
|
|
mul.ftz.f32 %f1418, %f1940, %f1940;
|
|
fma.rn.ftz.f32 %f1419, %f1939, %f1939, %f1418;
|
|
fma.rn.ftz.f32 %f1420, %f1941, %f1941, %f1419;
|
|
sqrt.approx.ftz.f32 %f425, %f1420;
|
|
setp.neu.ftz.f32 %p78, %f425, 0f00000000;
|
|
@%p78 bra $L__BB12_96;
|
|
bra.uni $L__BB12_97;
|
|
|
|
$L__BB12_96:
|
|
rcp.approx.ftz.f32 %f1421, %f425;
|
|
mul.ftz.f32 %f1941, %f1941, %f1421;
|
|
mul.ftz.f32 %f1940, %f1940, %f1421;
|
|
mul.ftz.f32 %f1939, %f1939, %f1421;
|
|
|
|
$L__BB12_97:
|
|
setp.eq.ftz.f32 %p79, %f1939, 0f00000000;
|
|
setp.eq.ftz.f32 %p80, %f1940, 0f00000000;
|
|
and.pred %p81, %p79, %p80;
|
|
setp.eq.ftz.f32 %p82, %f1941, 0f00000000;
|
|
and.pred %p83, %p82, %p81;
|
|
selp.f32 %f1422, %f1923, %f1941, %p83;
|
|
selp.f32 %f1423, %f1922, %f1940, %p83;
|
|
selp.f32 %f1424, %f1921, %f1939, %p83;
|
|
st.local.f32 [%rd33], %f1424;
|
|
st.local.f32 [%rd54], %f1423;
|
|
st.local.f32 [%rd55], %f1422;
|
|
|
|
$L__BB12_98:
|
|
st.local.f32 [%rd36], %f356;
|
|
st.local.f32 [%rd37], %f357;
|
|
st.local.f32 [%rd38], %f358;
|
|
cvta.to.local.u64 %rd258, %rd27;
|
|
st.local.v2.f32 [%rd258], {%f351, %f353};
|
|
st.local.f32 [%rd39], %f355;
|
|
|
|
$L__BB12_99:
|
|
and.b32 %r299, %r369, 4;
|
|
setp.eq.s32 %p84, %r299, 0;
|
|
@%p84 bra $L__BB12_101;
|
|
bra.uni $L__BB12_100;
|
|
|
|
$L__BB12_101:
|
|
and.b32 %r94, %r369, 2;
|
|
setp.eq.s32 %p85, %r94, 0;
|
|
@%p85 bra $L__BB12_103;
|
|
bra.uni $L__BB12_102;
|
|
|
|
$L__BB12_103:
|
|
ld.local.u32 %r300, [%rd56];
|
|
mul.wide.s32 %rd264, %r300, 208;
|
|
add.s64 %rd265, %rd20, %rd264;
|
|
ld.global.v4.f32 {%f1466, %f1467, %f1468, %f1469}, [%rd265+48];
|
|
ld.global.v4.f32 {%f1473, %f1474, %f1475, %f1476}, [%rd265+64];
|
|
ld.global.v4.f32 {%f1480, %f1481, %f1482, %f1483}, [%rd265+80];
|
|
ld.local.v2.f32 {%f1487, %f1488}, [%rd54];
|
|
ld.local.f32 %f1491, [%rd33];
|
|
mul.ftz.f32 %f1492, %f1473, %f1487;
|
|
fma.rn.ftz.f32 %f1493, %f1466, %f1491, %f1492;
|
|
fma.rn.ftz.f32 %f1494, %f1480, %f1488, %f1493;
|
|
mul.ftz.f32 %f1495, %f1474, %f1487;
|
|
fma.rn.ftz.f32 %f1496, %f1467, %f1491, %f1495;
|
|
fma.rn.ftz.f32 %f1497, %f1481, %f1488, %f1496;
|
|
mul.ftz.f32 %f1498, %f1475, %f1487;
|
|
fma.rn.ftz.f32 %f1499, %f1468, %f1491, %f1498;
|
|
fma.rn.ftz.f32 %f1500, %f1482, %f1488, %f1499;
|
|
mul.ftz.f32 %f1501, %f1497, %f1497;
|
|
fma.rn.ftz.f32 %f1502, %f1494, %f1494, %f1501;
|
|
fma.rn.ftz.f32 %f1503, %f1500, %f1500, %f1502;
|
|
sqrt.approx.ftz.f32 %f1504, %f1503;
|
|
div.approx.ftz.f32 %f1944, %f1494, %f1504;
|
|
div.approx.ftz.f32 %f1943, %f1497, %f1504;
|
|
div.approx.ftz.f32 %f1942, %f1500, %f1504;
|
|
bra.uni $L__BB12_104;
|
|
|
|
$L__BB12_100:
|
|
cvta.to.local.u64 %rd259, %rd22;
|
|
ld.local.v2.f32 {%f1951, %f1952}, [%rd259];
|
|
ld.local.f32 %f1953, [%rd25];
|
|
bra.uni $L__BB12_113;
|
|
|
|
$L__BB12_102:
|
|
ld.local.v2.f32 {%f1427, %f1428}, [%rd54];
|
|
cvta.to.local.u64 %rd261, %rd30;
|
|
ld.local.v4.f32 {%f1431, %f1432, %f1433, %f1434}, [%rd261];
|
|
cvta.to.local.u64 %rd262, %rd44;
|
|
ld.local.v4.f32 {%f1438, %f1439, %f1440, %f1441}, [%rd262];
|
|
cvta.to.local.u64 %rd263, %rd43;
|
|
ld.local.v4.f32 {%f1445, %f1446, %f1447, %f1448}, [%rd263];
|
|
ld.local.f32 %f1452, [%rd33];
|
|
mul.ftz.f32 %f1453, %f1427, %f1438;
|
|
fma.rn.ftz.f32 %f1454, %f1452, %f1431, %f1453;
|
|
fma.rn.ftz.f32 %f1455, %f1428, %f1445, %f1454;
|
|
mul.ftz.f32 %f1456, %f1427, %f1439;
|
|
fma.rn.ftz.f32 %f1457, %f1452, %f1432, %f1456;
|
|
fma.rn.ftz.f32 %f1458, %f1428, %f1446, %f1457;
|
|
mul.ftz.f32 %f1459, %f1427, %f1440;
|
|
fma.rn.ftz.f32 %f1460, %f1452, %f1433, %f1459;
|
|
fma.rn.ftz.f32 %f1461, %f1428, %f1447, %f1460;
|
|
mul.ftz.f32 %f1462, %f1458, %f1458;
|
|
fma.rn.ftz.f32 %f1463, %f1455, %f1455, %f1462;
|
|
fma.rn.ftz.f32 %f1464, %f1461, %f1461, %f1463;
|
|
sqrt.approx.ftz.f32 %f1465, %f1464;
|
|
div.approx.ftz.f32 %f1944, %f1455, %f1465;
|
|
div.approx.ftz.f32 %f1943, %f1458, %f1465;
|
|
div.approx.ftz.f32 %f1942, %f1461, %f1465;
|
|
|
|
$L__BB12_104:
|
|
st.local.f32 [%rd33], %f1944;
|
|
st.local.f32 [%rd54], %f1943;
|
|
st.local.f32 [%rd55], %f1942;
|
|
@%p85 bra $L__BB12_106;
|
|
bra.uni $L__BB12_105;
|
|
|
|
$L__BB12_106:
|
|
ld.local.u32 %r301, [%rd56];
|
|
mul.wide.s32 %rd271, %r301, 208;
|
|
add.s64 %rd272, %rd20, %rd271;
|
|
ld.global.v4.f32 {%f1544, %f1545, %f1546, %f1547}, [%rd272+48];
|
|
ld.global.v4.f32 {%f1551, %f1552, %f1553, %f1554}, [%rd272+64];
|
|
ld.global.v4.f32 {%f1558, %f1559, %f1560, %f1561}, [%rd272+80];
|
|
cvta.to.local.u64 %rd273, %rd22;
|
|
ld.local.v2.f32 {%f1565, %f1566}, [%rd273];
|
|
mul.ftz.f32 %f1569, %f1551, %f1566;
|
|
fma.rn.ftz.f32 %f1570, %f1544, %f1565, %f1569;
|
|
ld.local.f32 %f1571, [%rd25];
|
|
fma.rn.ftz.f32 %f1572, %f1558, %f1571, %f1570;
|
|
mul.ftz.f32 %f1573, %f1552, %f1566;
|
|
fma.rn.ftz.f32 %f1574, %f1545, %f1565, %f1573;
|
|
fma.rn.ftz.f32 %f1575, %f1559, %f1571, %f1574;
|
|
mul.ftz.f32 %f1576, %f1553, %f1566;
|
|
fma.rn.ftz.f32 %f1577, %f1546, %f1565, %f1576;
|
|
fma.rn.ftz.f32 %f1578, %f1560, %f1571, %f1577;
|
|
mul.ftz.f32 %f1579, %f1575, %f1575;
|
|
fma.rn.ftz.f32 %f1580, %f1572, %f1572, %f1579;
|
|
fma.rn.ftz.f32 %f1581, %f1578, %f1578, %f1580;
|
|
sqrt.approx.ftz.f32 %f1582, %f1581;
|
|
div.approx.ftz.f32 %f1951, %f1572, %f1582;
|
|
div.approx.ftz.f32 %f1952, %f1575, %f1582;
|
|
div.approx.ftz.f32 %f1953, %f1578, %f1582;
|
|
bra.uni $L__BB12_107;
|
|
|
|
$L__BB12_105:
|
|
cvta.to.local.u64 %rd267, %rd22;
|
|
ld.local.v2.f32 {%f1505, %f1506}, [%rd267];
|
|
cvta.to.local.u64 %rd268, %rd30;
|
|
ld.local.v4.f32 {%f1509, %f1510, %f1511, %f1512}, [%rd268];
|
|
cvta.to.local.u64 %rd269, %rd44;
|
|
ld.local.v4.f32 {%f1516, %f1517, %f1518, %f1519}, [%rd269];
|
|
cvta.to.local.u64 %rd270, %rd43;
|
|
ld.local.v4.f32 {%f1523, %f1524, %f1525, %f1526}, [%rd270];
|
|
mul.ftz.f32 %f1530, %f1506, %f1516;
|
|
fma.rn.ftz.f32 %f1531, %f1505, %f1509, %f1530;
|
|
ld.local.f32 %f1532, [%rd25];
|
|
fma.rn.ftz.f32 %f1533, %f1532, %f1523, %f1531;
|
|
mul.ftz.f32 %f1534, %f1506, %f1517;
|
|
fma.rn.ftz.f32 %f1535, %f1505, %f1510, %f1534;
|
|
fma.rn.ftz.f32 %f1536, %f1532, %f1524, %f1535;
|
|
mul.ftz.f32 %f1537, %f1506, %f1518;
|
|
fma.rn.ftz.f32 %f1538, %f1505, %f1511, %f1537;
|
|
fma.rn.ftz.f32 %f1539, %f1532, %f1525, %f1538;
|
|
mul.ftz.f32 %f1540, %f1536, %f1536;
|
|
fma.rn.ftz.f32 %f1541, %f1533, %f1533, %f1540;
|
|
fma.rn.ftz.f32 %f1542, %f1539, %f1539, %f1541;
|
|
sqrt.approx.ftz.f32 %f1543, %f1542;
|
|
div.approx.ftz.f32 %f1951, %f1533, %f1543;
|
|
div.approx.ftz.f32 %f1952, %f1536, %f1543;
|
|
div.approx.ftz.f32 %f1953, %f1539, %f1543;
|
|
|
|
$L__BB12_107:
|
|
cvta.to.local.u64 %rd274, %rd22;
|
|
st.local.v2.f32 [%rd274], {%f1951, %f1952};
|
|
st.local.f32 [%rd25], %f1953;
|
|
@%p85 bra $L__BB12_109;
|
|
bra.uni $L__BB12_108;
|
|
|
|
$L__BB12_109:
|
|
ld.local.u32 %r302, [%rd56];
|
|
mul.wide.s32 %rd279, %r302, 208;
|
|
add.s64 %rd280, %rd20, %rd279;
|
|
ld.global.v4.f32 {%f1615, %f1616, %f1617, %f1618}, [%rd280];
|
|
ld.global.v4.f32 {%f1622, %f1623, %f1624, %f1625}, [%rd280+16];
|
|
ld.global.v4.f32 {%f1629, %f1630, %f1631, %f1632}, [%rd280+32];
|
|
ld.local.v2.f32 {%f1636, %f1637}, [%rd37];
|
|
ld.local.f32 %f1640, [%rd36];
|
|
mul.ftz.f32 %f1641, %f1616, %f1636;
|
|
fma.rn.ftz.f32 %f1642, %f1640, %f1615, %f1641;
|
|
fma.rn.ftz.f32 %f1950, %f1617, %f1637, %f1642;
|
|
mul.ftz.f32 %f1643, %f1623, %f1636;
|
|
fma.rn.ftz.f32 %f1644, %f1640, %f1622, %f1643;
|
|
fma.rn.ftz.f32 %f1949, %f1624, %f1637, %f1644;
|
|
mul.ftz.f32 %f1645, %f1630, %f1636;
|
|
fma.rn.ftz.f32 %f1646, %f1640, %f1629, %f1645;
|
|
fma.rn.ftz.f32 %f1948, %f1631, %f1637, %f1646;
|
|
bra.uni $L__BB12_110;
|
|
|
|
$L__BB12_108:
|
|
ld.local.v2.f32 {%f1583, %f1584}, [%rd37];
|
|
cvta.to.local.u64 %rd276, %rd19;
|
|
ld.local.v4.f32 {%f1587, %f1588, %f1589, %f1590}, [%rd276];
|
|
ld.local.f32 %f1594, [%rd36];
|
|
mul.ftz.f32 %f1595, %f1583, %f1588;
|
|
fma.rn.ftz.f32 %f1596, %f1594, %f1587, %f1595;
|
|
fma.rn.ftz.f32 %f1950, %f1584, %f1589, %f1596;
|
|
cvta.to.local.u64 %rd277, %rd29;
|
|
ld.local.v4.f32 {%f1597, %f1598, %f1599, %f1600}, [%rd277];
|
|
mul.ftz.f32 %f1604, %f1583, %f1598;
|
|
fma.rn.ftz.f32 %f1605, %f1594, %f1597, %f1604;
|
|
fma.rn.ftz.f32 %f1949, %f1584, %f1599, %f1605;
|
|
cvta.to.local.u64 %rd278, %rd28;
|
|
ld.local.v4.f32 {%f1606, %f1607, %f1608, %f1609}, [%rd278];
|
|
mul.ftz.f32 %f1613, %f1583, %f1607;
|
|
fma.rn.ftz.f32 %f1614, %f1594, %f1606, %f1613;
|
|
fma.rn.ftz.f32 %f1948, %f1584, %f1608, %f1614;
|
|
|
|
$L__BB12_110:
|
|
st.local.f32 [%rd36], %f1950;
|
|
st.local.f32 [%rd37], %f1949;
|
|
st.local.f32 [%rd38], %f1948;
|
|
@%p85 bra $L__BB12_112;
|
|
bra.uni $L__BB12_111;
|
|
|
|
$L__BB12_112:
|
|
ld.local.u32 %r303, [%rd56];
|
|
mul.wide.s32 %rd286, %r303, 208;
|
|
add.s64 %rd287, %rd20, %rd286;
|
|
ld.global.v4.f32 {%f1684, %f1685, %f1686, %f1687}, [%rd287];
|
|
ld.global.v4.f32 {%f1691, %f1692, %f1693, %f1694}, [%rd287+16];
|
|
ld.global.v4.f32 {%f1698, %f1699, %f1700, %f1701}, [%rd287+32];
|
|
cvta.to.local.u64 %rd288, %rd27;
|
|
ld.local.v4.f32 {%f1705, %f1706, %f1707, %f1708}, [%rd288];
|
|
mul.ftz.f32 %f1712, %f1685, %f1706;
|
|
fma.rn.ftz.f32 %f1713, %f1705, %f1684, %f1712;
|
|
mul.ftz.f32 %f1714, %f1692, %f1706;
|
|
fma.rn.ftz.f32 %f1715, %f1705, %f1691, %f1714;
|
|
mul.ftz.f32 %f1716, %f1699, %f1706;
|
|
fma.rn.ftz.f32 %f1717, %f1705, %f1698, %f1716;
|
|
fma.rn.ftz.f32 %f1718, %f1700, %f1707, %f1717;
|
|
fma.rn.ftz.f32 %f1719, %f1693, %f1707, %f1715;
|
|
fma.rn.ftz.f32 %f1720, %f1686, %f1707, %f1713;
|
|
st.local.v2.f32 [%rd288], {%f1720, %f1719};
|
|
st.local.f32 [%rd39], %f1718;
|
|
bra.uni $L__BB12_113;
|
|
|
|
$L__BB12_111:
|
|
cvta.to.local.u64 %rd282, %rd27;
|
|
ld.local.v4.f32 {%f1647, %f1648, %f1649, %f1650}, [%rd282];
|
|
cvta.to.local.u64 %rd283, %rd19;
|
|
ld.local.v4.f32 {%f1654, %f1655, %f1656, %f1657}, [%rd283];
|
|
mul.ftz.f32 %f1661, %f1648, %f1655;
|
|
fma.rn.ftz.f32 %f1662, %f1647, %f1654, %f1661;
|
|
cvta.to.local.u64 %rd284, %rd29;
|
|
ld.local.v4.f32 {%f1663, %f1664, %f1665, %f1666}, [%rd284];
|
|
mul.ftz.f32 %f1670, %f1648, %f1664;
|
|
fma.rn.ftz.f32 %f1671, %f1647, %f1663, %f1670;
|
|
cvta.to.local.u64 %rd285, %rd28;
|
|
ld.local.v4.f32 {%f1672, %f1673, %f1674, %f1675}, [%rd285];
|
|
mul.ftz.f32 %f1679, %f1648, %f1673;
|
|
fma.rn.ftz.f32 %f1680, %f1647, %f1672, %f1679;
|
|
fma.rn.ftz.f32 %f1681, %f1649, %f1674, %f1680;
|
|
fma.rn.ftz.f32 %f1682, %f1649, %f1665, %f1671;
|
|
fma.rn.ftz.f32 %f1683, %f1649, %f1656, %f1662;
|
|
st.local.v2.f32 [%rd282], {%f1683, %f1682};
|
|
st.local.f32 [%rd39], %f1681;
|
|
|
|
$L__BB12_113:
|
|
ld.local.u32 %r95, [%rd24];
|
|
shl.b32 %r304, %r95, 5;
|
|
cvt.u64.u32 %rd289, %r304;
|
|
and.b64 %rd290, %rd289, 134217696;
|
|
add.s64 %rd291, %rd23, %rd290;
|
|
ld.global.u32 %r370, [%rd291+16];
|
|
st.local.u32 [%rd13], %r370;
|
|
ld.local.f32 %f1721, [%rd14];
|
|
ld.local.v2.f32 {%f1722, %f1723}, [%rd16];
|
|
mul.ftz.f32 %f1726, %f1952, %f1722;
|
|
fma.rn.ftz.f32 %f1727, %f1951, %f1721, %f1726;
|
|
fma.rn.ftz.f32 %f1728, %f1953, %f1723, %f1727;
|
|
setp.lt.ftz.f32 %p89, %f1728, 0f00000000;
|
|
@%p89 bra $L__BB12_115;
|
|
bra.uni $L__BB12_114;
|
|
|
|
$L__BB12_115:
|
|
or.b32 %r370, %r370, 1;
|
|
st.local.u32 [%rd13], %r370;
|
|
neg.ftz.f32 %f1953, %f1953;
|
|
st.local.f32 [%rd25], %f1953;
|
|
ld.local.f32 %f1729, [%rd33];
|
|
neg.ftz.f32 %f1730, %f1729;
|
|
ld.local.v2.f32 {%f1731, %f1732}, [%rd54];
|
|
st.local.f32 [%rd33], %f1730;
|
|
neg.ftz.f32 %f1735, %f1732;
|
|
neg.ftz.f32 %f1736, %f1731;
|
|
neg.ftz.f32 %f1952, %f1952;
|
|
neg.ftz.f32 %f1951, %f1951;
|
|
st.local.v4.f32 [%rd54], {%f1736, %f1735, %f1951, %f1952};
|
|
ld.local.f32 %f1737, [%rd36];
|
|
neg.ftz.f32 %f1965, %f1737;
|
|
ld.local.v2.f32 {%f1738, %f1739}, [%rd37];
|
|
neg.ftz.f32 %f1966, %f1738;
|
|
neg.ftz.f32 %f1742, %f1739;
|
|
st.local.f32 [%rd36], %f1965;
|
|
st.local.f32 [%rd37], %f1966;
|
|
st.local.f32 [%rd38], %f1742;
|
|
cvta.to.local.u64 %rd295, %rd27;
|
|
ld.local.v4.f32 {%f1743, %f1744, %f1745, %f1746}, [%rd295];
|
|
neg.ftz.f32 %f1750, %f1745;
|
|
neg.ftz.f32 %f1751, %f1744;
|
|
neg.ftz.f32 %f1752, %f1743;
|
|
st.local.v2.f32 [%rd295], {%f1752, %f1751};
|
|
st.local.f32 [%rd39], %f1750;
|
|
bra.uni $L__BB12_116;
|
|
|
|
$L__BB12_114:
|
|
ld.local.f32 %f1965, [%rd36];
|
|
ld.local.f32 %f1966, [%rd37];
|
|
|
|
$L__BB12_116:
|
|
ld.local.f32 %f1753, [%rd7];
|
|
ld.local.f32 %f1754, [%rd4+56];
|
|
ld.local.f32 %f1755, [%rd4+52];
|
|
fma.rn.ftz.f32 %f1756, %f1754, %f1753, %f1755;
|
|
st.local.f32 [%rd40], %f1756;
|
|
st.local.f32 [%rd41], %f1754;
|
|
cvta.to.local.u64 %rd296, %rd27;
|
|
ld.local.v4.f32 {%f1963, %f1964, %f1759, %f1760}, [%rd296];
|
|
setp.neu.ftz.f32 %p90, %f1951, %f1952;
|
|
setp.neu.ftz.f32 %p91, %f1951, %f1953;
|
|
or.pred %p92, %p90, %p91;
|
|
sub.ftz.f32 %f1761, %f1953, %f1952;
|
|
neg.ftz.f32 %f1762, %f1952;
|
|
neg.ftz.f32 %f1763, %f1953;
|
|
selp.f32 %f1764, %f1763, %f1953, %p92;
|
|
add.ftz.f32 %f1765, %f1951, %f1764;
|
|
selp.f32 %f1766, %f1952, %f1762, %p92;
|
|
sub.ftz.f32 %f1767, %f1766, %f1951;
|
|
mul.ftz.f32 %f1768, %f1765, %f1765;
|
|
fma.rn.ftz.f32 %f1769, %f1761, %f1761, %f1768;
|
|
fma.rn.ftz.f32 %f1770, %f1767, %f1767, %f1769;
|
|
sqrt.approx.ftz.f32 %f1771, %f1770;
|
|
div.approx.ftz.f32 %f1772, %f1761, %f1771;
|
|
div.approx.ftz.f32 %f1773, %f1765, %f1771;
|
|
div.approx.ftz.f32 %f1774, %f1767, %f1771;
|
|
mul.ftz.f32 %f1775, %f1952, %f1774;
|
|
mul.ftz.f32 %f1776, %f1953, %f1773;
|
|
sub.ftz.f32 %f1777, %f1775, %f1776;
|
|
mul.ftz.f32 %f1778, %f1953, %f1772;
|
|
mul.ftz.f32 %f1779, %f1951, %f1774;
|
|
sub.ftz.f32 %f1780, %f1778, %f1779;
|
|
mul.ftz.f32 %f1781, %f1951, %f1773;
|
|
mul.ftz.f32 %f1782, %f1952, %f1772;
|
|
sub.ftz.f32 %f1783, %f1781, %f1782;
|
|
mul.ftz.f32 %f1959, %f1756, %f1772;
|
|
mul.ftz.f32 %f1960, %f1756, %f1773;
|
|
mul.ftz.f32 %f494, %f1756, %f1774;
|
|
mul.ftz.f32 %f1961, %f1756, %f1777;
|
|
mul.ftz.f32 %f1962, %f1756, %f1780;
|
|
mul.ftz.f32 %f497, %f1756, %f1783;
|
|
abs.ftz.f32 %f498, %f1951;
|
|
abs.ftz.f32 %f499, %f1952;
|
|
abs.ftz.f32 %f500, %f1953;
|
|
setp.geu.ftz.f32 %p93, %f500, %f498;
|
|
setp.geu.ftz.f32 %p94, %f500, %f499;
|
|
and.pred %p95, %p93, %p94;
|
|
@%p95 bra $L__BB12_118;
|
|
|
|
ld.local.f32 %f502, [%rd38];
|
|
setp.lt.ftz.f32 %p96, %f499, %f500;
|
|
setp.lt.ftz.f32 %p97, %f499, %f498;
|
|
or.pred %p98, %p97, %p96;
|
|
selp.f32 %f1959, %f1960, %f1959, %p98;
|
|
selp.f32 %f1961, %f1962, %f1961, %p98;
|
|
selp.f32 %f1963, %f1964, %f1963, %p98;
|
|
selp.f32 %f1965, %f1966, %f1965, %p98;
|
|
mov.f32 %f1960, %f494;
|
|
mov.f32 %f1962, %f497;
|
|
mov.f32 %f1964, %f1759;
|
|
mov.f32 %f1966, %f502;
|
|
|
|
$L__BB12_118:
|
|
mul.ftz.f32 %f1784, %f1963, %f1966;
|
|
mul.ftz.f32 %f1785, %f1964, %f1965;
|
|
sub.ftz.f32 %f1967, %f1785, %f1784;
|
|
setp.eq.ftz.f32 %p99, %f1967, 0f00000000;
|
|
@%p99 bra $L__BB12_120;
|
|
|
|
rcp.approx.ftz.f32 %f1967, %f1967;
|
|
|
|
$L__BB12_120:
|
|
mov.pred %p133, 0;
|
|
mul.ftz.f32 %f1786, %f1960, %f1963;
|
|
mul.ftz.f32 %f1787, %f1959, %f1964;
|
|
sub.ftz.f32 %f1788, %f1787, %f1786;
|
|
mul.ftz.f32 %f1789, %f1788, %f1967;
|
|
st.local.f32 [%rd51], %f1789;
|
|
mul.ftz.f32 %f1790, %f1959, %f1966;
|
|
mul.ftz.f32 %f1791, %f1960, %f1965;
|
|
sub.ftz.f32 %f1792, %f1791, %f1790;
|
|
mul.ftz.f32 %f1793, %f1962, %f1963;
|
|
mul.ftz.f32 %f1794, %f1961, %f1964;
|
|
sub.ftz.f32 %f1795, %f1794, %f1793;
|
|
mul.ftz.f32 %f1796, %f1792, %f1967;
|
|
mul.ftz.f32 %f1797, %f1795, %f1967;
|
|
st.local.v2.f32 [%rd297], {%f1797, %f1796};
|
|
mul.ftz.f32 %f1798, %f1961, %f1966;
|
|
mul.ftz.f32 %f1799, %f1962, %f1965;
|
|
sub.ftz.f32 %f1800, %f1799, %f1798;
|
|
mul.ftz.f32 %f1801, %f1800, %f1967;
|
|
st.local.f32 [%rd53], %f1801;
|
|
and.b32 %r305, %r370, 1;
|
|
setp.eq.b32 %p100, %r305, 1;
|
|
xor.pred %p102, %p100, %p133;
|
|
not.pred %p103, %p102;
|
|
@%p103 bra $L__BB12_130;
|
|
bra.uni $L__BB12_121;
|
|
|
|
$L__BB12_130:
|
|
ld.local.u32 %r311, [%rd82];
|
|
add.s32 %r106, %r373, 1;
|
|
mul.wide.s32 %rd319, %r373, 4;
|
|
add.s64 %rd320, %rd1, %rd319;
|
|
st.local.u32 [%rd320], %r311;
|
|
mov.u32 %r373, %r106;
|
|
bra.uni $L__BB12_131;
|
|
|
|
$L__BB12_121:
|
|
mov.u16 %rs47, 1;
|
|
setp.lt.s32 %p105, %r373, 1;
|
|
mov.pred %p134, -1;
|
|
@%p105 bra $L__BB12_124;
|
|
|
|
ld.local.u32 %r99, [%rd82];
|
|
mov.u16 %rs47, 1;
|
|
mov.u32 %r371, 0;
|
|
|
|
$L__BB12_123:
|
|
mul.wide.s32 %rd298, %r371, 4;
|
|
add.s64 %rd299, %rd1, %rd298;
|
|
ld.local.u32 %r307, [%rd299];
|
|
setp.eq.s32 %p106, %r307, %r99;
|
|
selp.b16 %rs47, 0, %rs47, %p106;
|
|
and.b16 %rs27, %rs47, 255;
|
|
add.s32 %r371, %r371, 1;
|
|
setp.lt.s32 %p107, %r371, %r373;
|
|
setp.ne.s16 %p134, %rs27, 0;
|
|
and.pred %p108, %p134, %p107;
|
|
@%p108 bra $L__BB12_123;
|
|
|
|
$L__BB12_124:
|
|
setp.lt.s32 %p109, %r375, 1;
|
|
not.pred %p110, %p134;
|
|
or.pred %p111, %p110, %p109;
|
|
mov.u16 %rs48, %rs47;
|
|
@%p111 bra $L__BB12_128;
|
|
|
|
mov.u32 %r372, 0;
|
|
ld.local.u32 %r102, [%rd82];
|
|
|
|
$L__BB12_126:
|
|
mul.wide.s32 %rd300, %r372, 16;
|
|
mov.u64 %rd301, kernel_params;
|
|
add.s64 %rd302, %rd301, %rd300;
|
|
ld.const.u64 %rd303, [%rd302+2784];
|
|
cvta.to.global.u64 %rd304, %rd303;
|
|
shl.b64 %rd305, %rd3, 2;
|
|
add.s64 %rd306, %rd304, %rd305;
|
|
ld.global.u32 %r309, [%rd306];
|
|
setp.eq.s32 %p112, %r309, %r102;
|
|
add.s32 %r372, %r372, 1;
|
|
mov.u16 %rs48, 0;
|
|
@%p112 bra $L__BB12_128;
|
|
|
|
setp.lt.s32 %p113, %r372, %r375;
|
|
mov.u16 %rs48, %rs47;
|
|
@%p113 bra $L__BB12_126;
|
|
|
|
$L__BB12_128:
|
|
and.b16 %rs29, %rs48, 255;
|
|
setp.eq.s16 %p114, %rs29, 0;
|
|
@%p114 bra $L__BB12_131;
|
|
|
|
ld.local.u32 %r310, [%rd82];
|
|
mul.wide.s32 %rd307, %r375, 16;
|
|
mov.u64 %rd308, kernel_params;
|
|
add.s64 %rd309, %rd308, %rd307;
|
|
ld.const.v2.u64 {%rd310, %rd311}, [%rd309+2784];
|
|
cvta.to.global.u64 %rd314, %rd310;
|
|
shl.b64 %rd315, %rd3, 2;
|
|
add.s64 %rd316, %rd314, %rd315;
|
|
st.global.u32 [%rd316], %r310;
|
|
cvta.to.global.u64 %rd317, %rd311;
|
|
add.s64 %rd318, %rd317, %rd315;
|
|
st.global.u32 [%rd318], %r95;
|
|
add.s32 %r375, %r375, 1;
|
|
|
|
$L__BB12_131:
|
|
ld.const.u32 %r344, [kernel_params+28];
|
|
shl.b32 %r343, %r344, 1;
|
|
add.s32 %r342, %r344, -1;
|
|
mov.b32 %f1805, %r140;
|
|
add.s32 %r312, %r140, 1;
|
|
mov.b32 %f1802, %r312;
|
|
setp.eq.ftz.f32 %p115, %f1805, 0f00000000;
|
|
selp.f32 %f1824, 0f00000001, %f1802, %p115;
|
|
st.local.f32 [%rd4+24], %f1824;
|
|
st.local.u32 [%rd4+40], %r144;
|
|
st.local.u32 [%rd4+36], %r143;
|
|
setp.gt.s32 %p116, %r373, 30;
|
|
setp.ge.u32 %p117, %r375, %r342;
|
|
or.pred %p118, %p116, %p117;
|
|
add.s32 %r358, %r358, 1;
|
|
setp.ge.u32 %p119, %r358, %r343;
|
|
or.pred %p120, %p119, %p118;
|
|
@%p120 bra $L__BB12_133;
|
|
bra.uni $L__BB12_132;
|
|
|
|
$L__BB12_133:
|
|
ld.const.u64 %rd427, [kernel_params+2560];
|
|
mul.wide.s32 %rd426, %r354, 4;
|
|
cvta.to.global.u64 %rd425, %rd427;
|
|
add.s64 %rd424, %rd425, %rd426;
|
|
mul.wide.s32 %rd321, %r375, 16;
|
|
mov.u64 %rd322, kernel_params;
|
|
add.s64 %rd323, %rd322, %rd321;
|
|
ld.const.v2.u64 {%rd324, %rd325}, [%rd323+2784];
|
|
cvta.to.global.u64 %rd328, %rd324;
|
|
shl.b64 %rd329, %rd3, 2;
|
|
add.s64 %rd330, %rd328, %rd329;
|
|
st.global.u32 [%rd330], %r126;
|
|
cvta.to.global.u64 %rd331, %rd325;
|
|
add.s64 %rd332, %rd331, %rd329;
|
|
st.global.u32 [%rd332], %r126;
|
|
ld.global.u8 %rs30, [%rd424+3];
|
|
and.b16 %rs31, %rs30, 64;
|
|
setp.eq.s16 %p121, %rs31, 0;
|
|
@%p121 bra $L__BB12_145;
|
|
bra.uni $L__BB12_134;
|
|
|
|
$L__BB12_145:
|
|
ld.const.u64 %rd409, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd410, %rd409;
|
|
add.s64 %rd411, %rd410, 20;
|
|
atom.global.add.u32 %r339, [%rd411], -1;
|
|
add.s64 %rd412, %rd410, 8;
|
|
atom.global.add.u32 %r340, [%rd412], 1;
|
|
ld.const.u64 %rd413, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd414, %rd413;
|
|
shl.b64 %rd415, %rd3, 1;
|
|
add.s64 %rd416, %rd414, %rd415;
|
|
mov.u16 %rs37, 2;
|
|
st.global.u16 [%rd416], %rs37;
|
|
bra.uni $L__BB12_146;
|
|
|
|
$L__BB12_134:
|
|
ld.const.u64 %rd333, [kernel_params+2728];
|
|
cvta.to.global.u64 %rd334, %rd333;
|
|
add.s64 %rd336, %rd334, %rd329;
|
|
ld.global.u32 %r111, [%rd336];
|
|
cvt.s64.s32 %rd83, %r111;
|
|
ld.const.v2.u64 {%rd337, %rd338}, [kernel_params+2736];
|
|
cvta.to.global.u64 %rd341, %rd337;
|
|
add.s64 %rd342, %rd341, %rd329;
|
|
ld.global.u32 %r112, [%rd342];
|
|
cvta.to.global.u64 %rd343, %rd338;
|
|
add.s64 %rd344, %rd343, %rd329;
|
|
ld.global.u32 %r113, [%rd344];
|
|
and.b32 %r314, %r113, 1;
|
|
setp.eq.b32 %p122, %r314, 1;
|
|
mov.pred %p123, 0;
|
|
xor.pred %p124, %p122, %p123;
|
|
not.pred %p125, %p124;
|
|
@%p125 bra $L__BB12_136;
|
|
bra.uni $L__BB12_135;
|
|
|
|
$L__BB12_136:
|
|
and.b32 %r315, %r113, 8;
|
|
setp.eq.s32 %p126, %r315, 0;
|
|
@%p126 bra $L__BB12_138;
|
|
bra.uni $L__BB12_137;
|
|
|
|
$L__BB12_138:
|
|
and.b32 %r317, %r113, 6;
|
|
setp.eq.s32 %p127, %r317, 0;
|
|
mov.u32 %r376, 0;
|
|
@%p127 bra $L__BB12_140;
|
|
|
|
cvt.u32.u64 %r318, %rd83;
|
|
ld.const.u64 %rd353, [kernel_params+2280];
|
|
cvta.to.global.u64 %rd354, %rd353;
|
|
mul.wide.s32 %rd355, %r318, 16;
|
|
add.s64 %rd356, %rd354, %rd355;
|
|
ld.global.u32 %r376, [%rd356];
|
|
bra.uni $L__BB12_140;
|
|
|
|
$L__BB12_135:
|
|
ld.const.u64 %rd345, [kernel_params+2232];
|
|
cvta.to.global.u64 %rd346, %rd345;
|
|
shl.b64 %rd347, %rd83, 2;
|
|
add.s64 %rd348, %rd346, %rd347;
|
|
ld.global.u32 %r376, [%rd348];
|
|
bra.uni $L__BB12_140;
|
|
|
|
$L__BB12_137:
|
|
ld.const.u64 %rd349, [kernel_params+2320];
|
|
cvta.to.global.u64 %rd350, %rd349;
|
|
mul.wide.s32 %rd351, %r111, 4;
|
|
add.s64 %rd352, %rd350, %rd351;
|
|
ld.global.u32 %r376, [%rd352];
|
|
|
|
$L__BB12_140:
|
|
and.b32 %r118, %r376, 4194303;
|
|
ld.const.u64 %rd357, [kernel_params+2200];
|
|
cvta.to.global.u64 %rd358, %rd357;
|
|
mul.wide.s32 %rd359, %r112, 4;
|
|
add.s64 %rd360, %rd358, %rd359;
|
|
ld.const.u32 %r319, [kernel_params+1608];
|
|
setp.ne.s32 %p128, %r319, 0;
|
|
ld.global.u8 %rs32, [%rd360+1];
|
|
and.b16 %rs33, %rs32, 6;
|
|
setp.ne.s16 %p129, %rs33, 0;
|
|
and.pred %p130, %p129, %p128;
|
|
@%p130 bra $L__BB12_144;
|
|
bra.uni $L__BB12_141;
|
|
|
|
$L__BB12_144:
|
|
cvt.u32.u64 %r331, %rd3;
|
|
ld.const.u32 %r332, [kernel_params+4408];
|
|
div.u32 %r333, %r331, %r332;
|
|
ld.const.u32 %r334, [kernel_params+24];
|
|
mad.lo.s32 %r335, %r333, %r334, %r118;
|
|
ld.const.u64 %rd393, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd394, %rd393;
|
|
add.s64 %rd395, %rd394, 20;
|
|
atom.global.add.u32 %r336, [%rd395], -1;
|
|
add.s64 %rd396, %rd394, 40;
|
|
atom.global.add.u32 %r337, [%rd396], 1;
|
|
ld.const.u64 %rd397, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd398, %rd397;
|
|
shl.b64 %rd399, %rd3, 1;
|
|
add.s64 %rd400, %rd398, %rd399;
|
|
mov.u16 %rs36, 10;
|
|
st.global.u16 [%rd400], %rs36;
|
|
ld.const.u64 %rd401, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd402, %rd401;
|
|
add.s64 %rd404, %rd402, %rd329;
|
|
st.global.u32 [%rd404], %r335;
|
|
ld.const.u64 %rd405, [kernel_params+4360];
|
|
cvta.to.global.u64 %rd406, %rd405;
|
|
mul.wide.s32 %rd407, %r335, 4;
|
|
add.s64 %rd408, %rd406, %rd407;
|
|
atom.global.add.u32 %r338, [%rd408], 1;
|
|
bra.uni $L__BB12_146;
|
|
|
|
$L__BB12_141:
|
|
ld.const.u64 %rd361, [kernel_params+2424];
|
|
cvt.u32.u64 %r320, %rd3;
|
|
cvta.to.global.u64 %rd362, %rd361;
|
|
mul.wide.u32 %rd363, %r118, 32;
|
|
add.s64 %rd364, %rd362, %rd363;
|
|
ld.global.u32 %r321, [%rd364+16];
|
|
and.b32 %r322, %r321, 1073741824;
|
|
setp.eq.s32 %p131, %r322, 0;
|
|
ld.const.u32 %r323, [kernel_params+4408];
|
|
div.u32 %r324, %r320, %r323;
|
|
ld.const.u32 %r325, [kernel_params+24];
|
|
mad.lo.s32 %r119, %r324, %r325, %r118;
|
|
ld.const.u64 %rd365, [kernel_params+4272];
|
|
cvta.to.global.u64 %rd84, %rd365;
|
|
add.s64 %rd366, %rd84, 20;
|
|
atom.global.add.u32 %r326, [%rd366], -1;
|
|
cvt.s64.s32 %rd85, %r119;
|
|
@%p131 bra $L__BB12_143;
|
|
bra.uni $L__BB12_142;
|
|
|
|
$L__BB12_143:
|
|
add.s64 %rd380, %rd84, 32;
|
|
atom.global.add.u32 %r329, [%rd380], 1;
|
|
ld.const.u64 %rd381, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd382, %rd381;
|
|
shl.b64 %rd383, %rd3, 1;
|
|
add.s64 %rd384, %rd382, %rd383;
|
|
mov.u16 %rs35, 8;
|
|
st.global.u16 [%rd384], %rs35;
|
|
ld.const.u64 %rd385, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd386, %rd385;
|
|
add.s64 %rd388, %rd386, %rd329;
|
|
st.global.u32 [%rd388], %r119;
|
|
ld.const.u64 %rd389, [kernel_params+4344];
|
|
cvta.to.global.u64 %rd390, %rd389;
|
|
shl.b64 %rd391, %rd85, 2;
|
|
add.s64 %rd392, %rd390, %rd391;
|
|
atom.global.add.u32 %r330, [%rd392], 1;
|
|
bra.uni $L__BB12_146;
|
|
|
|
$L__BB12_142:
|
|
add.s64 %rd367, %rd84, 36;
|
|
atom.global.add.u32 %r327, [%rd367], 1;
|
|
ld.const.u64 %rd368, [kernel_params+2536];
|
|
cvta.to.global.u64 %rd369, %rd368;
|
|
shl.b64 %rd370, %rd3, 1;
|
|
add.s64 %rd371, %rd369, %rd370;
|
|
mov.u16 %rs34, 9;
|
|
st.global.u16 [%rd371], %rs34;
|
|
ld.const.u64 %rd372, [kernel_params+2640];
|
|
cvta.to.global.u64 %rd373, %rd372;
|
|
add.s64 %rd375, %rd373, %rd329;
|
|
st.global.u32 [%rd375], %r119;
|
|
ld.const.u64 %rd376, [kernel_params+4352];
|
|
cvta.to.global.u64 %rd377, %rd376;
|
|
shl.b64 %rd378, %rd85, 2;
|
|
add.s64 %rd379, %rd377, %rd378;
|
|
atom.global.add.u32 %r328, [%rd379], 1;
|
|
|
|
$L__BB12_146:
|
|
ret;
|
|
|
|
}
|
|
|