1941 lines
56 KiB
Plaintext
1941 lines
56 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-23083092
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
|
// Based on LLVM 3.4svn
|
|
//
|
|
|
|
.version 6.1
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
// .globl _Z6oxMainv
|
|
.global .align 8 .b8 pixelID[8];
|
|
.global .align 8 .b8 resolution[8];
|
|
.global .align 4 .b8 normal[12];
|
|
.global .align 4 .b8 camPos[12];
|
|
.global .align 4 .b8 root[4];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 1 .b8 image[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_Dir[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 lightMeshBuffer[1];
|
|
.global .align 4 .u32 lightMeshBufferSize;
|
|
.global .align 4 .f32 lightInvCutoff;
|
|
.global .align 4 .f32 lightPointSize;
|
|
.global .align 4 .b8 lightColor[12];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 1 .b8 rnd_seeds2[1];
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum19lightMeshBufferSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 16 .b8 __local_depot0[64];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<146>;
|
|
.reg .b16 %rs<56>;
|
|
.reg .f32 %f<1011>;
|
|
.reg .b32 %r<371>;
|
|
.reg .b64 %rd<175>;
|
|
|
|
|
|
mov.u64 %rd174, __local_depot0;
|
|
cvta.local.u64 %SP, %rd174;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
ld.global.v2.u32 {%r105, %r106}, [pixelID];
|
|
cvt.u64.u32 %rd25, %r105;
|
|
cvt.u64.u32 %rd26, %r106;
|
|
mov.u64 %rd29, uvnormal;
|
|
cvta.global.u64 %rd24, %rd29;
|
|
mov.u32 %r103, 2;
|
|
mov.u32 %r104, 4;
|
|
mov.u64 %rd28, 0;
|
|
// inline asm
|
|
call (%rd23), _rt_buffer_get_64, (%rd24, %r103, %r104, %rd25, %rd26, %rd28, %rd28);
|
|
// inline asm
|
|
ld.u32 %r4, [%rd23];
|
|
shr.u32 %r109, %r4, 16;
|
|
cvt.u16.u32 %rs1, %r109;
|
|
and.b16 %rs3, %rs1, 255;
|
|
cvt.u16.u32 %rs4, %r4;
|
|
or.b16 %rs5, %rs4, %rs3;
|
|
setp.eq.s16 %p7, %rs5, 0;
|
|
mov.f32 %f922, 0f00000000;
|
|
mov.f32 %f923, %f922;
|
|
mov.f32 %f924, %f922;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs6, [%rd23+1];
|
|
and.b16 %rs8, %rs4, 255;
|
|
cvt.rn.f32.u16 %f238, %rs8;
|
|
div.rn.f32 %f239, %f238, 0f437F0000;
|
|
fma.rn.f32 %f240, %f239, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f241, %rs6;
|
|
div.rn.f32 %f242, %f241, 0f437F0000;
|
|
fma.rn.f32 %f243, %f242, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f244, %rs3;
|
|
div.rn.f32 %f245, %f244, 0f437F0000;
|
|
fma.rn.f32 %f246, %f245, 0f40000000, 0fBF800000;
|
|
mul.f32 %f247, %f243, %f243;
|
|
fma.rn.f32 %f248, %f240, %f240, %f247;
|
|
fma.rn.f32 %f249, %f246, %f246, %f248;
|
|
sqrt.rn.f32 %f250, %f249;
|
|
rcp.rn.f32 %f251, %f250;
|
|
mul.f32 %f922, %f240, %f251;
|
|
mul.f32 %f923, %f243, %f251;
|
|
mul.f32 %f924, %f246, %f251;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r110, %r111}, [pixelID];
|
|
ld.global.v2.u32 {%r113, %r114}, [tileInfo];
|
|
add.s32 %r5, %r110, %r113;
|
|
add.s32 %r6, %r111, %r114;
|
|
setp.eq.f32 %p8, %f923, 0f00000000;
|
|
setp.eq.f32 %p9, %f922, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f924, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_134;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_134:
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
and.b32 %r317, %r370, 1;
|
|
setp.eq.b32 %p140, %r317, 1;
|
|
@!%p140 bra BB0_136;
|
|
bra.uni BB0_135;
|
|
|
|
BB0_135:
|
|
cvt.u64.u32 %rd127, %r5;
|
|
cvt.u64.u32 %rd128, %r6;
|
|
mov.u64 %rd131, image;
|
|
cvta.global.u64 %rd126, %rd131;
|
|
mov.u64 %rd130, 0;
|
|
// inline asm
|
|
call (%rd125), _rt_buffer_get_64, (%rd126, %r103, %r104, %rd127, %rd128, %rd130, %rd130);
|
|
// inline asm
|
|
mov.u16 %rs37, 0;
|
|
st.v4.u8 [%rd125], {%rs37, %rs37, %rs37, %rs37};
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
|
|
BB0_136:
|
|
and.b32 %r320, %r370, 8;
|
|
setp.eq.s32 %p141, %r320, 0;
|
|
@%p141 bra BB0_138;
|
|
|
|
cvt.u64.u32 %rd134, %r5;
|
|
cvt.u64.u32 %rd135, %r6;
|
|
mov.u64 %rd138, image_Mask;
|
|
cvta.global.u64 %rd133, %rd138;
|
|
mov.u64 %rd137, 0;
|
|
// inline asm
|
|
call (%rd132), _rt_buffer_get_64, (%rd133, %r103, %r103, %rd134, %rd135, %rd137, %rd137);
|
|
// inline asm
|
|
mov.f32 %f905, 0f00000000;
|
|
cvt.rzi.u32.f32 %r323, %f905;
|
|
cvt.u16.u32 %rs38, %r323;
|
|
mov.u16 %rs39, 0;
|
|
st.v2.u8 [%rd132], {%rs38, %rs39};
|
|
ld.global.u32 %r370, [imageEnabled];
|
|
|
|
BB0_138:
|
|
and.b32 %r324, %r370, 4;
|
|
setp.eq.s32 %p142, %r324, 0;
|
|
@%p142 bra BB0_142;
|
|
|
|
ld.global.u32 %r325, [additive];
|
|
setp.eq.s32 %p143, %r325, 0;
|
|
cvt.u64.u32 %rd21, %r5;
|
|
cvt.u64.u32 %rd22, %r6;
|
|
@%p143 bra BB0_141;
|
|
|
|
mov.u64 %rd151, image_HDR;
|
|
cvta.global.u64 %rd140, %rd151;
|
|
mov.u32 %r329, 8;
|
|
mov.u64 %rd150, 0;
|
|
// inline asm
|
|
call (%rd139), _rt_buffer_get_64, (%rd140, %r103, %r329, %rd21, %rd22, %rd150, %rd150);
|
|
// inline asm
|
|
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd139];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f906, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f907, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f908, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd145), _rt_buffer_get_64, (%rd140, %r103, %r329, %rd21, %rd22, %rd150, %rd150);
|
|
// inline asm
|
|
add.f32 %f909, %f906, 0f00000000;
|
|
add.f32 %f910, %f907, 0f00000000;
|
|
add.f32 %f911, %f908, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f911;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f910;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f909;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs50, 0;
|
|
st.v4.u16 [%rd145], {%rs43, %rs44, %rs45, %rs50};
|
|
bra.uni BB0_142;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
|
cvt.u64.u32 %rd32, %r123;
|
|
cvt.u64.u32 %rd33, %r124;
|
|
mov.u64 %rd42, uvpos;
|
|
cvta.global.u64 %rd31, %rd42;
|
|
mov.u32 %r119, 12;
|
|
// inline asm
|
|
call (%rd30), _rt_buffer_get_64, (%rd31, %r103, %r119, %rd32, %rd33, %rd28, %rd28);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd30+8];
|
|
ld.f32 %f8, [%rd30+4];
|
|
ld.f32 %f7, [%rd30];
|
|
mul.f32 %f259, %f7, 0f3456BF95;
|
|
mul.f32 %f260, %f8, 0f3456BF95;
|
|
mul.f32 %f261, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f922;
|
|
div.rn.f32 %f262, %f259, %f10;
|
|
abs.f32 %f263, %f923;
|
|
div.rn.f32 %f264, %f260, %f263;
|
|
abs.f32 %f11, %f924;
|
|
div.rn.f32 %f265, %f261, %f11;
|
|
abs.f32 %f266, %f262;
|
|
abs.f32 %f267, %f264;
|
|
abs.f32 %f268, %f265;
|
|
mov.f32 %f269, 0f38D1B717;
|
|
max.f32 %f270, %f266, %f269;
|
|
max.f32 %f271, %f267, %f269;
|
|
max.f32 %f272, %f268, %f269;
|
|
fma.rn.f32 %f12, %f922, %f270, %f7;
|
|
fma.rn.f32 %f13, %f923, %f271, %f8;
|
|
fma.rn.f32 %f14, %f924, %f272, %f9;
|
|
ld.global.v2.u32 {%r127, %r128}, [pixelID];
|
|
cvt.u64.u32 %rd38, %r127;
|
|
cvt.u64.u32 %rd39, %r128;
|
|
mov.u64 %rd43, rnd_seeds;
|
|
cvta.global.u64 %rd37, %rd43;
|
|
// inline asm
|
|
call (%rd36), _rt_buffer_get_64, (%rd37, %r103, %r104, %rd38, %rd39, %rd28, %rd28);
|
|
// inline asm
|
|
ld.u32 %r346, [%rd36];
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f19, 0f00000000;
|
|
mov.u32 %r9, 0;
|
|
setp.lt.s32 %p14, %r2, 1;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f15, [lightPointSize];
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mul.f32 %f17, %f13, 0f3456BF95;
|
|
mul.f32 %f18, %f14, 0f3456BF95;
|
|
mov.f32 %f19, 0f00000000;
|
|
mov.u32 %r339, 0;
|
|
abs.f32 %f403, %f17;
|
|
abs.f32 %f404, %f16;
|
|
max.f32 %f405, %f404, %f403;
|
|
abs.f32 %f406, %f18;
|
|
max.f32 %f407, %f405, %f406;
|
|
mov.u32 %r9, %r339;
|
|
mov.f32 %f20, %f19;
|
|
mov.f32 %f21, %f19;
|
|
mov.f32 %f22, %f19;
|
|
mov.f32 %f953, %f19;
|
|
mov.f32 %f954, %f19;
|
|
mov.f32 %f955, %f19;
|
|
|
|
BB0_6:
|
|
shl.b32 %r10, %r339, 1;
|
|
cvt.s64.s32 %rd46, %r10;
|
|
mov.u64 %rd50, lightMeshBuffer;
|
|
cvta.global.u64 %rd45, %rd50;
|
|
mov.u32 %r133, 1;
|
|
mov.u64 %rd49, 0;
|
|
// inline asm
|
|
call (%rd44), _rt_buffer_get_64, (%rd45, %r133, %r119, %rd46, %rd49, %rd49, %rd49);
|
|
// inline asm
|
|
ld.f32 %f280, [%rd44];
|
|
sub.f32 %f281, %f280, %f7;
|
|
ld.f32 %f282, [%rd44+4];
|
|
sub.f32 %f283, %f282, %f8;
|
|
ld.f32 %f284, [%rd44+8];
|
|
sub.f32 %f285, %f284, %f9;
|
|
mul.f32 %f286, %f283, %f283;
|
|
fma.rn.f32 %f287, %f281, %f281, %f286;
|
|
fma.rn.f32 %f288, %f285, %f285, %f287;
|
|
sqrt.rn.f32 %f26, %f288;
|
|
rcp.rn.f32 %f289, %f26;
|
|
mul.f32 %f27, %f281, %f289;
|
|
mul.f32 %f28, %f283, %f289;
|
|
mul.f32 %f29, %f285, %f289;
|
|
mul.f32 %f290, %f923, %f28;
|
|
fma.rn.f32 %f291, %f922, %f27, %f290;
|
|
fma.rn.f32 %f30, %f924, %f29, %f291;
|
|
setp.leu.f32 %p15, %f30, 0f00000000;
|
|
@%p15 bra BB0_23;
|
|
|
|
mov.u64 %rd169, 0;
|
|
setp.eq.s32 %p16, %r3, 0;
|
|
mul.f32 %f292, %f26, %f26;
|
|
div.rn.f32 %f31, %f15, %f292;
|
|
ld.global.v2.u32 {%r137, %r138}, [pixelID];
|
|
cvt.u64.u32 %rd57, %r138;
|
|
cvt.u64.u32 %rd58, %r137;
|
|
and.b64 %rd53, %rd58, 255;
|
|
and.b64 %rd54, %rd57, 255;
|
|
mov.u64 %rd59, rnd_seeds2;
|
|
cvta.global.u64 %rd52, %rd59;
|
|
mov.u32 %r136, 3;
|
|
// inline asm
|
|
call (%rd51), _rt_buffer_get_64, (%rd52, %r103, %r136, %rd53, %rd54, %rd169, %rd169);
|
|
// inline asm
|
|
@%p16 bra BB0_9;
|
|
|
|
ld.u8 %rs10, [%rd51];
|
|
cvt.rn.f32.u16 %f293, %rs10;
|
|
div.rn.f32 %f294, %f293, 0fC37F0000;
|
|
fma.rn.f32 %f295, %f294, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p18, %f31, %f295;
|
|
mov.pred %p145, -1;
|
|
@%p18 bra BB0_25;
|
|
|
|
BB0_9:
|
|
ld.global.f32 %f298, [lightInvCutoff];
|
|
mul.f32 %f32, %f26, %f298;
|
|
mov.f32 %f302, 0f40800000;
|
|
abs.f32 %f34, %f32;
|
|
setp.lt.f32 %p19, %f34, 0f00800000;
|
|
mul.f32 %f304, %f34, 0f4B800000;
|
|
selp.f32 %f305, 0fC3170000, 0fC2FE0000, %p19;
|
|
selp.f32 %f306, %f304, %f34, %p19;
|
|
mov.b32 %r141, %f306;
|
|
and.b32 %r142, %r141, 8388607;
|
|
or.b32 %r143, %r142, 1065353216;
|
|
mov.b32 %f307, %r143;
|
|
shr.u32 %r144, %r141, 23;
|
|
cvt.rn.f32.u32 %f308, %r144;
|
|
add.f32 %f309, %f305, %f308;
|
|
setp.gt.f32 %p20, %f307, 0f3FB504F3;
|
|
mul.f32 %f310, %f307, 0f3F000000;
|
|
add.f32 %f311, %f309, 0f3F800000;
|
|
selp.f32 %f312, %f310, %f307, %p20;
|
|
selp.f32 %f313, %f311, %f309, %p20;
|
|
add.f32 %f314, %f312, 0fBF800000;
|
|
add.f32 %f297, %f312, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f296,%f297;
|
|
// inline asm
|
|
add.f32 %f315, %f314, %f314;
|
|
mul.f32 %f316, %f296, %f315;
|
|
mul.f32 %f317, %f316, %f316;
|
|
mov.f32 %f318, 0f3C4CAF63;
|
|
mov.f32 %f319, 0f3B18F0FE;
|
|
fma.rn.f32 %f320, %f319, %f317, %f318;
|
|
mov.f32 %f321, 0f3DAAAABD;
|
|
fma.rn.f32 %f322, %f320, %f317, %f321;
|
|
mul.rn.f32 %f323, %f322, %f317;
|
|
mul.rn.f32 %f324, %f323, %f316;
|
|
sub.f32 %f325, %f314, %f316;
|
|
neg.f32 %f326, %f316;
|
|
add.f32 %f327, %f325, %f325;
|
|
fma.rn.f32 %f328, %f326, %f314, %f327;
|
|
mul.rn.f32 %f329, %f296, %f328;
|
|
add.f32 %f330, %f324, %f316;
|
|
sub.f32 %f331, %f316, %f330;
|
|
add.f32 %f332, %f324, %f331;
|
|
add.f32 %f333, %f329, %f332;
|
|
add.f32 %f334, %f330, %f333;
|
|
sub.f32 %f335, %f330, %f334;
|
|
add.f32 %f336, %f333, %f335;
|
|
mov.f32 %f337, 0f3F317200;
|
|
mul.rn.f32 %f338, %f313, %f337;
|
|
mov.f32 %f339, 0f35BFBE8E;
|
|
mul.rn.f32 %f340, %f313, %f339;
|
|
add.f32 %f341, %f338, %f334;
|
|
sub.f32 %f342, %f338, %f341;
|
|
add.f32 %f343, %f334, %f342;
|
|
add.f32 %f344, %f336, %f343;
|
|
add.f32 %f345, %f340, %f344;
|
|
add.f32 %f346, %f341, %f345;
|
|
sub.f32 %f347, %f341, %f346;
|
|
add.f32 %f348, %f345, %f347;
|
|
mul.rn.f32 %f349, %f302, %f346;
|
|
neg.f32 %f350, %f349;
|
|
fma.rn.f32 %f351, %f302, %f346, %f350;
|
|
fma.rn.f32 %f352, %f302, %f348, %f351;
|
|
mov.f32 %f353, 0f00000000;
|
|
fma.rn.f32 %f354, %f353, %f346, %f352;
|
|
add.rn.f32 %f355, %f349, %f354;
|
|
neg.f32 %f356, %f355;
|
|
add.rn.f32 %f357, %f349, %f356;
|
|
add.rn.f32 %f358, %f357, %f354;
|
|
mov.b32 %r145, %f355;
|
|
setp.eq.s32 %p21, %r145, 1118925336;
|
|
add.s32 %r146, %r145, -1;
|
|
mov.b32 %f359, %r146;
|
|
add.f32 %f360, %f358, 0f37000000;
|
|
selp.f32 %f361, %f359, %f355, %p21;
|
|
selp.f32 %f35, %f360, %f358, %p21;
|
|
mul.f32 %f362, %f361, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f363, %f362;
|
|
mov.f32 %f364, 0fBF317200;
|
|
fma.rn.f32 %f365, %f363, %f364, %f361;
|
|
mov.f32 %f366, 0fB5BFBE8E;
|
|
fma.rn.f32 %f367, %f363, %f366, %f365;
|
|
mul.f32 %f368, %f367, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f369, %f368;
|
|
add.f32 %f370, %f363, 0f00000000;
|
|
ex2.approx.f32 %f371, %f370;
|
|
mul.f32 %f372, %f369, %f371;
|
|
setp.lt.f32 %p22, %f361, 0fC2D20000;
|
|
selp.f32 %f373, 0f00000000, %f372, %p22;
|
|
setp.gt.f32 %p23, %f361, 0f42D20000;
|
|
selp.f32 %f932, 0f7F800000, %f373, %p23;
|
|
setp.eq.f32 %p24, %f932, 0f7F800000;
|
|
@%p24 bra BB0_11;
|
|
|
|
fma.rn.f32 %f932, %f932, %f35, %f932;
|
|
|
|
BB0_11:
|
|
mov.f32 %f918, 0f40000000;
|
|
cvt.rzi.f32.f32 %f917, %f918;
|
|
add.f32 %f916, %f917, %f917;
|
|
mov.f32 %f915, 0f40800000;
|
|
sub.f32 %f914, %f915, %f916;
|
|
abs.f32 %f913, %f914;
|
|
setp.lt.f32 %p25, %f32, 0f00000000;
|
|
setp.eq.f32 %p26, %f913, 0f3F800000;
|
|
and.pred %p1, %p25, %p26;
|
|
mov.b32 %r147, %f932;
|
|
xor.b32 %r148, %r147, -2147483648;
|
|
mov.b32 %f374, %r148;
|
|
selp.f32 %f934, %f374, %f932, %p1;
|
|
setp.eq.f32 %p27, %f32, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
bra.uni BB0_12;
|
|
|
|
BB0_14:
|
|
add.f32 %f377, %f32, %f32;
|
|
selp.f32 %f934, %f377, 0f00000000, %p26;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_12:
|
|
setp.geu.f32 %p28, %f32, 0f00000000;
|
|
@%p28 bra BB0_15;
|
|
|
|
mov.f32 %f921, 0f40800000;
|
|
cvt.rzi.f32.f32 %f376, %f921;
|
|
setp.neu.f32 %p29, %f376, 0f40800000;
|
|
selp.f32 %f934, 0f7FFFFFFF, %f934, %p29;
|
|
|
|
BB0_15:
|
|
add.f32 %f378, %f34, 0f40800000;
|
|
mov.b32 %r149, %f378;
|
|
setp.lt.s32 %p31, %r149, 2139095040;
|
|
@%p31 bra BB0_20;
|
|
|
|
setp.gtu.f32 %p32, %f34, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_19:
|
|
add.f32 %f934, %f32, 0f40800000;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.neu.f32 %p33, %f34, 0f7F800000;
|
|
@%p33 bra BB0_20;
|
|
|
|
selp.f32 %f934, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_20:
|
|
mov.u32 %r335, 1;
|
|
mov.u64 %rd168, lightMeshBuffer;
|
|
cvta.global.u64 %rd167, %rd168;
|
|
shl.b32 %r334, %r339, 1;
|
|
mov.u64 %rd166, 0;
|
|
mov.f32 %f379, 0f3F800000;
|
|
sub.f32 %f380, %f379, %f934;
|
|
setp.eq.f32 %p34, %f32, 0f3F800000;
|
|
selp.f32 %f381, 0f00000000, %f380, %p34;
|
|
cvt.sat.f32.f32 %f382, %f381;
|
|
mul.f32 %f383, %f31, %f382;
|
|
add.s32 %r152, %r334, 1;
|
|
cvt.s64.s32 %rd62, %r152;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd167, %r335, %r119, %rd62, %rd166, %rd166, %rd166);
|
|
// inline asm
|
|
ld.f32 %f384, [%rd60];
|
|
mul.f32 %f385, %f27, %f384;
|
|
ld.f32 %f386, [%rd60+4];
|
|
mul.f32 %f387, %f28, %f386;
|
|
neg.f32 %f388, %f387;
|
|
sub.f32 %f389, %f388, %f385;
|
|
ld.f32 %f390, [%rd60+8];
|
|
mul.f32 %f391, %f29, %f390;
|
|
sub.f32 %f392, %f389, %f391;
|
|
cvt.sat.f32.f32 %f393, %f392;
|
|
mul.f32 %f46, %f383, %f393;
|
|
cvt.sat.f32.f32 %f47, %f30;
|
|
mul.f32 %f394, %f46, %f47;
|
|
setp.leu.f32 %p35, %f394, 0f3727C5AC;
|
|
@%p35 bra BB0_22;
|
|
|
|
mov.u32 %r338, 1;
|
|
add.u64 %rd67, %SP, 28;
|
|
cvta.to.local.u64 %rd68, %rd67;
|
|
max.f32 %f401, %f407, %f269;
|
|
sub.f32 %f402, %f26, %f401;
|
|
mov.u32 %r156, 1065353216;
|
|
st.local.u32 [%rd68+8], %r156;
|
|
st.local.u32 [%rd68+4], %r156;
|
|
st.local.u32 [%rd68], %r156;
|
|
ld.global.u32 %r153, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r153, %f12, %f13, %f14, %f27, %f28, %f29, %r338, %f401, %f402, %rd67, %r119);
|
|
// inline asm
|
|
ld.local.f32 %f409, [%rd68];
|
|
mul.f32 %f410, %f46, 0f3EA2F983;
|
|
mul.f32 %f411, %f410, %f409;
|
|
ld.local.f32 %f412, [%rd68+4];
|
|
mul.f32 %f413, %f410, %f412;
|
|
ld.local.f32 %f414, [%rd68+8];
|
|
mul.f32 %f415, %f410, %f414;
|
|
fma.rn.f32 %f22, %f47, %f411, %f22;
|
|
fma.rn.f32 %f21, %f47, %f413, %f21;
|
|
fma.rn.f32 %f20, %f47, %f415, %f20;
|
|
add.f32 %f19, %f19, %f409;
|
|
fma.rn.f32 %f955, %f27, %f409, %f955;
|
|
fma.rn.f32 %f954, %f28, %f412, %f954;
|
|
fma.rn.f32 %f953, %f29, %f414, %f953;
|
|
|
|
BB0_22:
|
|
add.s32 %r9, %r9, 1;
|
|
|
|
BB0_23:
|
|
add.s32 %r339, %r339, 1;
|
|
setp.lt.s32 %p37, %r339, %r2;
|
|
@%p37 bra BB0_6;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_4:
|
|
mov.f32 %f20, %f19;
|
|
mov.f32 %f21, %f19;
|
|
mov.f32 %f22, %f19;
|
|
mov.f32 %f953, %f19;
|
|
mov.f32 %f954, %f19;
|
|
mov.f32 %f955, %f19;
|
|
|
|
BB0_24:
|
|
mov.pred %p145, %p13;
|
|
|
|
BB0_25:
|
|
cvt.rn.f32.s32 %f416, %r9;
|
|
mov.f32 %f417, 0f3F800000;
|
|
max.f32 %f418, %f416, %f417;
|
|
rcp.rn.f32 %f419, %f418;
|
|
mul.f32 %f995, %f22, %f419;
|
|
mul.f32 %f996, %f21, %f419;
|
|
mul.f32 %f997, %f20, %f419;
|
|
div.rn.f32 %f998, %f19, %f418;
|
|
@!%p145 bra BB0_80;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_26:
|
|
abs.f32 %f920, %f924;
|
|
abs.f32 %f919, %f922;
|
|
setp.gt.f32 %p38, %f919, %f920;
|
|
neg.f32 %f427, %f923;
|
|
selp.f32 %f428, %f427, 0f00000000, %p38;
|
|
neg.f32 %f429, %f924;
|
|
selp.f32 %f430, %f922, %f429, %p38;
|
|
selp.f32 %f431, 0f00000000, %f923, %p38;
|
|
mul.f32 %f432, %f430, %f430;
|
|
fma.rn.f32 %f433, %f428, %f428, %f432;
|
|
fma.rn.f32 %f434, %f431, %f431, %f433;
|
|
sqrt.rn.f32 %f435, %f434;
|
|
rcp.rn.f32 %f436, %f435;
|
|
mul.f32 %f80, %f428, %f436;
|
|
mul.f32 %f81, %f430, %f436;
|
|
mul.f32 %f82, %f431, %f436;
|
|
mov.f32 %f963, 0f00000000;
|
|
setp.lt.s32 %p39, %r3, 1;
|
|
mov.f32 %f964, %f963;
|
|
mov.f32 %f965, %f963;
|
|
mov.f32 %f966, %f963;
|
|
mov.f32 %f953, %f963;
|
|
mov.f32 %f954, %f963;
|
|
mov.f32 %f955, %f963;
|
|
@%p39 bra BB0_79;
|
|
|
|
cvt.rn.f32.s32 %f444, %r3;
|
|
rcp.rn.f32 %f83, %f444;
|
|
add.u64 %rd69, %SP, 0;
|
|
cvta.to.local.u64 %rd3, %rd69;
|
|
mul.f32 %f84, %f12, 0f3456BF95;
|
|
mul.f32 %f85, %f13, 0f3456BF95;
|
|
mul.f32 %f86, %f14, 0f3456BF95;
|
|
add.u64 %rd70, %SP, 48;
|
|
cvta.to.local.u64 %rd4, %rd70;
|
|
mul.f32 %f445, %f922, %f81;
|
|
mul.f32 %f446, %f923, %f80;
|
|
sub.f32 %f87, %f446, %f445;
|
|
mul.f32 %f447, %f924, %f80;
|
|
mul.f32 %f448, %f922, %f82;
|
|
sub.f32 %f88, %f448, %f447;
|
|
mul.f32 %f449, %f923, %f82;
|
|
mul.f32 %f450, %f924, %f81;
|
|
sub.f32 %f89, %f450, %f449;
|
|
mov.f32 %f963, 0f00000000;
|
|
mov.u32 %r157, 0;
|
|
abs.f32 %f451, %f85;
|
|
abs.f32 %f452, %f84;
|
|
max.f32 %f453, %f452, %f451;
|
|
abs.f32 %f454, %f86;
|
|
max.f32 %f455, %f453, %f454;
|
|
mov.u32 %r343, %r157;
|
|
mov.f32 %f964, %f963;
|
|
mov.f32 %f965, %f963;
|
|
mov.f32 %f966, %f963;
|
|
mov.f32 %f953, %f963;
|
|
mov.f32 %f954, %f963;
|
|
mov.f32 %f955, %f963;
|
|
|
|
BB0_28:
|
|
cvt.rn.f32.s32 %f97, %r343;
|
|
max.f32 %f98, %f455, %f269;
|
|
mov.u32 %r345, %r157;
|
|
|
|
BB0_29:
|
|
mad.lo.s32 %r159, %r346, 1664525, 1013904223;
|
|
and.b32 %r160, %r159, 16777215;
|
|
cvt.rn.f32.u32 %f457, %r160;
|
|
fma.rn.f32 %f458, %f457, 0f33800000, %f97;
|
|
mul.f32 %f106, %f83, %f458;
|
|
mad.lo.s32 %r346, %r159, 1664525, 1013904223;
|
|
and.b32 %r161, %r346, 16777215;
|
|
cvt.rn.f32.u32 %f459, %r161;
|
|
cvt.rn.f32.s32 %f460, %r345;
|
|
fma.rn.f32 %f461, %f459, 0f33800000, %f460;
|
|
mul.f32 %f462, %f83, %f461;
|
|
mul.f32 %f463, %f106, %f106;
|
|
sub.f32 %f465, %f417, %f463;
|
|
mov.f32 %f982, 0f00000000;
|
|
max.f32 %f467, %f982, %f465;
|
|
sqrt.rn.f32 %f107, %f467;
|
|
mul.f32 %f976, %f462, 0f40C90FDB;
|
|
abs.f32 %f109, %f976;
|
|
setp.neu.f32 %p40, %f109, 0f7F800000;
|
|
mov.f32 %f970, %f976;
|
|
@%p40 bra BB0_31;
|
|
|
|
mov.f32 %f468, 0f00000000;
|
|
mul.rn.f32 %f970, %f976, %f468;
|
|
|
|
BB0_31:
|
|
mul.f32 %f469, %f970, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r356, %f469;
|
|
cvt.rn.f32.s32 %f470, %r356;
|
|
neg.f32 %f471, %f470;
|
|
mov.f32 %f472, 0f3FC90FDA;
|
|
fma.rn.f32 %f473, %f471, %f472, %f970;
|
|
mov.f32 %f474, 0f33A22168;
|
|
fma.rn.f32 %f475, %f471, %f474, %f473;
|
|
mov.f32 %f476, 0f27C234C5;
|
|
fma.rn.f32 %f971, %f471, %f476, %f475;
|
|
abs.f32 %f477, %f970;
|
|
setp.leu.f32 %p41, %f477, 0f47CE4780;
|
|
@%p41 bra BB0_42;
|
|
|
|
mov.b32 %r21, %f970;
|
|
shr.u32 %r22, %r21, 23;
|
|
shl.b32 %r164, %r21, 8;
|
|
or.b32 %r23, %r164, -2147483648;
|
|
mov.u32 %r347, 0;
|
|
mov.u64 %rd171, 0;
|
|
mov.u64 %rd170, %rd3;
|
|
mov.u32 %r348, %r347;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd72, %rd171, 2;
|
|
mov.u64 %rd73, __cudart_i2opi_f;
|
|
add.s64 %rd74, %rd73, %rd72;
|
|
ld.const.u32 %r167, [%rd74];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r165, %r167, %r23, %r348;
|
|
madc.hi.u32 %r348, %r167, %r23, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd170], %r165;
|
|
add.s32 %r347, %r347, 1;
|
|
cvt.s64.s32 %rd171, %r347;
|
|
mul.wide.s32 %rd75, %r347, 4;
|
|
add.s64 %rd170, %rd3, %rd75;
|
|
setp.ne.s32 %p42, %r347, 6;
|
|
@%p42 bra BB0_33;
|
|
|
|
and.b32 %r170, %r22, 255;
|
|
add.s32 %r171, %r170, -128;
|
|
shr.u32 %r172, %r171, 5;
|
|
and.b32 %r28, %r21, -2147483648;
|
|
st.local.u32 [%rd3+24], %r348;
|
|
mov.u32 %r173, 6;
|
|
sub.s32 %r174, %r173, %r172;
|
|
mul.wide.s32 %rd76, %r174, 4;
|
|
add.s64 %rd10, %rd3, %rd76;
|
|
ld.local.u32 %r349, [%rd10];
|
|
ld.local.u32 %r350, [%rd10+-4];
|
|
and.b32 %r31, %r22, 31;
|
|
setp.eq.s32 %p43, %r31, 0;
|
|
@%p43 bra BB0_36;
|
|
|
|
mov.u32 %r175, 32;
|
|
sub.s32 %r176, %r175, %r31;
|
|
shr.u32 %r177, %r350, %r176;
|
|
shl.b32 %r178, %r349, %r31;
|
|
add.s32 %r349, %r177, %r178;
|
|
ld.local.u32 %r179, [%rd10+-8];
|
|
shr.u32 %r180, %r179, %r176;
|
|
shl.b32 %r181, %r350, %r31;
|
|
add.s32 %r350, %r180, %r181;
|
|
|
|
BB0_36:
|
|
shr.u32 %r182, %r350, 30;
|
|
shl.b32 %r183, %r349, 2;
|
|
add.s32 %r351, %r182, %r183;
|
|
shl.b32 %r37, %r350, 2;
|
|
shr.u32 %r184, %r351, 31;
|
|
shr.u32 %r185, %r349, 30;
|
|
add.s32 %r38, %r184, %r185;
|
|
setp.eq.s32 %p44, %r184, 0;
|
|
@%p44 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r352, %r28;
|
|
mov.u32 %r353, %r37;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
not.b32 %r186, %r351;
|
|
neg.s32 %r353, %r37;
|
|
setp.eq.s32 %p45, %r37, 0;
|
|
selp.u32 %r187, 1, 0, %p45;
|
|
add.s32 %r351, %r187, %r186;
|
|
xor.b32 %r352, %r28, -2147483648;
|
|
|
|
BB0_39:
|
|
clz.b32 %r355, %r351;
|
|
setp.eq.s32 %p46, %r355, 0;
|
|
shl.b32 %r188, %r351, %r355;
|
|
mov.u32 %r189, 32;
|
|
sub.s32 %r190, %r189, %r355;
|
|
shr.u32 %r191, %r353, %r190;
|
|
add.s32 %r192, %r191, %r188;
|
|
selp.b32 %r46, %r351, %r192, %p46;
|
|
mov.u32 %r193, -921707870;
|
|
mul.hi.u32 %r354, %r46, %r193;
|
|
setp.eq.s32 %p47, %r28, 0;
|
|
neg.s32 %r194, %r38;
|
|
selp.b32 %r356, %r38, %r194, %p47;
|
|
setp.lt.s32 %p48, %r354, 1;
|
|
@%p48 bra BB0_41;
|
|
|
|
mul.lo.s32 %r195, %r46, -921707870;
|
|
shr.u32 %r196, %r195, 31;
|
|
shl.b32 %r197, %r354, 1;
|
|
add.s32 %r354, %r196, %r197;
|
|
add.s32 %r355, %r355, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r198, 126;
|
|
sub.s32 %r199, %r198, %r355;
|
|
shl.b32 %r200, %r199, 23;
|
|
add.s32 %r201, %r354, 1;
|
|
shr.u32 %r202, %r201, 7;
|
|
add.s32 %r203, %r202, 1;
|
|
shr.u32 %r204, %r203, 1;
|
|
add.s32 %r205, %r204, %r200;
|
|
or.b32 %r206, %r205, %r352;
|
|
mov.b32 %f971, %r206;
|
|
|
|
BB0_42:
|
|
mul.rn.f32 %f115, %f971, %f971;
|
|
add.s32 %r54, %r356, 1;
|
|
and.b32 %r55, %r54, 1;
|
|
setp.eq.s32 %p49, %r55, 0;
|
|
@%p49 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f480, 0f3C08839E;
|
|
mov.f32 %f481, 0fB94CA1F9;
|
|
fma.rn.f32 %f972, %f481, %f115, %f480;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f478, 0fBAB6061A;
|
|
mov.f32 %f479, 0f37CCF5CE;
|
|
fma.rn.f32 %f972, %f479, %f115, %f478;
|
|
|
|
BB0_45:
|
|
@%p49 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mov.f32 %f485, 0fBE2AAAA3;
|
|
fma.rn.f32 %f486, %f972, %f115, %f485;
|
|
mov.f32 %f487, 0f00000000;
|
|
fma.rn.f32 %f973, %f486, %f115, %f487;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mov.f32 %f482, 0f3D2AAAA5;
|
|
fma.rn.f32 %f483, %f972, %f115, %f482;
|
|
mov.f32 %f484, 0fBF000000;
|
|
fma.rn.f32 %f973, %f483, %f115, %f484;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f974, %f973, %f971, %f971;
|
|
@%p49 bra BB0_50;
|
|
|
|
fma.rn.f32 %f974, %f973, %f115, %f417;
|
|
|
|
BB0_50:
|
|
and.b32 %r207, %r54, 2;
|
|
setp.eq.s32 %p52, %r207, 0;
|
|
@%p52 bra BB0_52;
|
|
|
|
mov.f32 %f490, 0fBF800000;
|
|
fma.rn.f32 %f974, %f974, %f490, %f982;
|
|
|
|
BB0_52:
|
|
@%p40 bra BB0_54;
|
|
|
|
mov.f32 %f491, 0f00000000;
|
|
mul.rn.f32 %f976, %f976, %f491;
|
|
|
|
BB0_54:
|
|
mul.f32 %f492, %f976, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r366, %f492;
|
|
cvt.rn.f32.s32 %f493, %r366;
|
|
neg.f32 %f494, %f493;
|
|
fma.rn.f32 %f496, %f494, %f472, %f976;
|
|
fma.rn.f32 %f498, %f494, %f474, %f496;
|
|
fma.rn.f32 %f977, %f494, %f476, %f498;
|
|
abs.f32 %f500, %f976;
|
|
setp.leu.f32 %p54, %f500, 0f47CE4780;
|
|
@%p54 bra BB0_65;
|
|
|
|
mov.b32 %r57, %f976;
|
|
shr.u32 %r58, %r57, 23;
|
|
shl.b32 %r210, %r57, 8;
|
|
or.b32 %r59, %r210, -2147483648;
|
|
mov.u32 %r357, 0;
|
|
mov.u64 %rd172, %rd3;
|
|
mov.u64 %rd173, %rd28;
|
|
mov.u32 %r358, %r357;
|
|
|
|
BB0_56:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd78, %rd173, 2;
|
|
mov.u64 %rd79, __cudart_i2opi_f;
|
|
add.s64 %rd80, %rd79, %rd78;
|
|
ld.const.u32 %r213, [%rd80];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r211, %r213, %r59, %r358;
|
|
madc.hi.u32 %r358, %r213, %r59, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd172], %r211;
|
|
add.s32 %r357, %r357, 1;
|
|
cvt.s64.s32 %rd173, %r357;
|
|
mul.wide.s32 %rd81, %r357, 4;
|
|
add.s64 %rd172, %rd3, %rd81;
|
|
setp.ne.s32 %p55, %r357, 6;
|
|
@%p55 bra BB0_56;
|
|
|
|
and.b32 %r216, %r58, 255;
|
|
add.s32 %r217, %r216, -128;
|
|
shr.u32 %r218, %r217, 5;
|
|
and.b32 %r64, %r57, -2147483648;
|
|
st.local.u32 [%rd3+24], %r358;
|
|
mov.u32 %r219, 6;
|
|
sub.s32 %r220, %r219, %r218;
|
|
mul.wide.s32 %rd82, %r220, 4;
|
|
add.s64 %rd16, %rd3, %rd82;
|
|
ld.local.u32 %r359, [%rd16];
|
|
ld.local.u32 %r360, [%rd16+-4];
|
|
and.b32 %r67, %r58, 31;
|
|
setp.eq.s32 %p56, %r67, 0;
|
|
@%p56 bra BB0_59;
|
|
|
|
mov.u32 %r221, 32;
|
|
sub.s32 %r222, %r221, %r67;
|
|
shr.u32 %r223, %r360, %r222;
|
|
shl.b32 %r224, %r359, %r67;
|
|
add.s32 %r359, %r223, %r224;
|
|
ld.local.u32 %r225, [%rd16+-8];
|
|
shr.u32 %r226, %r225, %r222;
|
|
shl.b32 %r227, %r360, %r67;
|
|
add.s32 %r360, %r226, %r227;
|
|
|
|
BB0_59:
|
|
shr.u32 %r228, %r360, 30;
|
|
shl.b32 %r229, %r359, 2;
|
|
add.s32 %r361, %r228, %r229;
|
|
shl.b32 %r73, %r360, 2;
|
|
shr.u32 %r230, %r361, 31;
|
|
shr.u32 %r231, %r359, 30;
|
|
add.s32 %r74, %r230, %r231;
|
|
setp.eq.s32 %p57, %r230, 0;
|
|
@%p57 bra BB0_60;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_60:
|
|
mov.u32 %r362, %r64;
|
|
mov.u32 %r363, %r73;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_61:
|
|
not.b32 %r232, %r361;
|
|
neg.s32 %r363, %r73;
|
|
setp.eq.s32 %p58, %r73, 0;
|
|
selp.u32 %r233, 1, 0, %p58;
|
|
add.s32 %r361, %r233, %r232;
|
|
xor.b32 %r362, %r64, -2147483648;
|
|
|
|
BB0_62:
|
|
clz.b32 %r365, %r361;
|
|
setp.eq.s32 %p59, %r365, 0;
|
|
shl.b32 %r234, %r361, %r365;
|
|
mov.u32 %r235, 32;
|
|
sub.s32 %r236, %r235, %r365;
|
|
shr.u32 %r237, %r363, %r236;
|
|
add.s32 %r238, %r237, %r234;
|
|
selp.b32 %r82, %r361, %r238, %p59;
|
|
mov.u32 %r239, -921707870;
|
|
mul.hi.u32 %r364, %r82, %r239;
|
|
setp.eq.s32 %p60, %r64, 0;
|
|
neg.s32 %r240, %r74;
|
|
selp.b32 %r366, %r74, %r240, %p60;
|
|
setp.lt.s32 %p61, %r364, 1;
|
|
@%p61 bra BB0_64;
|
|
|
|
mul.lo.s32 %r241, %r82, -921707870;
|
|
shr.u32 %r242, %r241, 31;
|
|
shl.b32 %r243, %r364, 1;
|
|
add.s32 %r364, %r242, %r243;
|
|
add.s32 %r365, %r365, 1;
|
|
|
|
BB0_64:
|
|
mov.u32 %r244, 126;
|
|
sub.s32 %r245, %r244, %r365;
|
|
shl.b32 %r246, %r245, 23;
|
|
add.s32 %r247, %r364, 1;
|
|
shr.u32 %r248, %r247, 7;
|
|
add.s32 %r249, %r248, 1;
|
|
shr.u32 %r250, %r249, 1;
|
|
add.s32 %r251, %r250, %r246;
|
|
or.b32 %r252, %r251, %r362;
|
|
mov.b32 %f977, %r252;
|
|
|
|
BB0_65:
|
|
mul.rn.f32 %f132, %f977, %f977;
|
|
and.b32 %r90, %r366, 1;
|
|
setp.eq.s32 %p62, %r90, 0;
|
|
@%p62 bra BB0_67;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_67:
|
|
mov.f32 %f503, 0f3C08839E;
|
|
mov.f32 %f504, 0fB94CA1F9;
|
|
fma.rn.f32 %f978, %f504, %f132, %f503;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_66:
|
|
mov.f32 %f501, 0fBAB6061A;
|
|
mov.f32 %f502, 0f37CCF5CE;
|
|
fma.rn.f32 %f978, %f502, %f132, %f501;
|
|
|
|
BB0_68:
|
|
@%p62 bra BB0_70;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_70:
|
|
mov.f32 %f508, 0fBE2AAAA3;
|
|
fma.rn.f32 %f509, %f978, %f132, %f508;
|
|
mov.f32 %f510, 0f00000000;
|
|
fma.rn.f32 %f979, %f509, %f132, %f510;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_69:
|
|
mov.f32 %f505, 0f3D2AAAA5;
|
|
fma.rn.f32 %f506, %f978, %f132, %f505;
|
|
mov.f32 %f507, 0fBF000000;
|
|
fma.rn.f32 %f979, %f506, %f132, %f507;
|
|
|
|
BB0_71:
|
|
fma.rn.f32 %f980, %f979, %f977, %f977;
|
|
@%p62 bra BB0_73;
|
|
|
|
fma.rn.f32 %f980, %f979, %f132, %f417;
|
|
|
|
BB0_73:
|
|
and.b32 %r253, %r366, 2;
|
|
setp.eq.s32 %p65, %r253, 0;
|
|
@%p65 bra BB0_75;
|
|
|
|
mov.f32 %f513, 0fBF800000;
|
|
fma.rn.f32 %f980, %f980, %f513, %f982;
|
|
|
|
BB0_75:
|
|
mul.f32 %f525, %f107, %f974;
|
|
mul.f32 %f526, %f107, %f980;
|
|
mul.f32 %f527, %f80, %f526;
|
|
mul.f32 %f528, %f81, %f526;
|
|
mul.f32 %f529, %f82, %f526;
|
|
fma.rn.f32 %f530, %f89, %f525, %f527;
|
|
fma.rn.f32 %f531, %f88, %f525, %f528;
|
|
fma.rn.f32 %f532, %f87, %f525, %f529;
|
|
fma.rn.f32 %f144, %f922, %f106, %f530;
|
|
fma.rn.f32 %f145, %f923, %f106, %f531;
|
|
fma.rn.f32 %f146, %f924, %f106, %f532;
|
|
mov.f32 %f533, 0f7F7FFFFF;
|
|
st.local.v4.f32 [%rd4], {%f417, %f417, %f417, %f533};
|
|
ld.global.u32 %r254, [root];
|
|
mov.u32 %r255, 0;
|
|
mov.f32 %f521, 0f6C4ECB8F;
|
|
mov.u32 %r256, 16;
|
|
// inline asm
|
|
call _rt_trace_64, (%r254, %f12, %f13, %f14, %f144, %f145, %f146, %r255, %f98, %f521, %rd70, %r256);
|
|
// inline asm
|
|
ld.local.f32 %f147, [%rd4+12];
|
|
setp.ge.f32 %p66, %f147, 0f00000000;
|
|
mov.f32 %f983, %f982;
|
|
mov.f32 %f984, %f982;
|
|
@%p66 bra BB0_77;
|
|
|
|
ld.local.v4.f32 {%f984, %f983, %f982, %f538}, [%rd4];
|
|
|
|
BB0_77:
|
|
setp.ltu.f32 %p67, %f147, 0f00000000;
|
|
selp.f32 %f539, 0f3F800000, 0f00000000, %p67;
|
|
fma.rn.f32 %f955, %f144, %f539, %f955;
|
|
fma.rn.f32 %f954, %f145, %f539, %f954;
|
|
fma.rn.f32 %f953, %f146, %f539, %f953;
|
|
add.f32 %f963, %f963, %f539;
|
|
mul.f32 %f540, %f923, %f145;
|
|
fma.rn.f32 %f541, %f922, %f144, %f540;
|
|
fma.rn.f32 %f542, %f924, %f146, %f541;
|
|
cvt.sat.f32.f32 %f543, %f542;
|
|
fma.rn.f32 %f966, %f984, %f543, %f966;
|
|
fma.rn.f32 %f965, %f983, %f543, %f965;
|
|
fma.rn.f32 %f964, %f982, %f543, %f964;
|
|
add.s32 %r345, %r345, 1;
|
|
setp.lt.s32 %p68, %r345, %r3;
|
|
@%p68 bra BB0_29;
|
|
|
|
add.s32 %r343, %r343, 1;
|
|
setp.lt.s32 %p69, %r343, %r3;
|
|
@%p69 bra BB0_28;
|
|
|
|
BB0_79:
|
|
mul.lo.s32 %r257, %r3, %r3;
|
|
cvt.rn.f32.s32 %f544, %r257;
|
|
rcp.rn.f32 %f545, %f544;
|
|
mul.f32 %f546, %f966, %f545;
|
|
mul.f32 %f547, %f965, %f545;
|
|
mul.f32 %f548, %f964, %f545;
|
|
div.rn.f32 %f998, %f963, %f544;
|
|
fma.rn.f32 %f995, %f966, %f545, %f546;
|
|
fma.rn.f32 %f996, %f965, %f545, %f547;
|
|
fma.rn.f32 %f997, %f964, %f545, %f548;
|
|
|
|
BB0_80:
|
|
setp.gt.s32 %p70, %r1, -1;
|
|
selp.f32 %f179, %f998, %f995, %p70;
|
|
ld.global.u32 %r368, [imageEnabled];
|
|
and.b32 %r258, %r368, 8;
|
|
setp.eq.s32 %p71, %r258, 0;
|
|
@%p71 bra BB0_93;
|
|
|
|
cvt.u64.u32 %rd86, %r5;
|
|
cvt.u64.u32 %rd87, %r6;
|
|
mov.u64 %rd90, image_Mask;
|
|
cvta.global.u64 %rd85, %rd90;
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd85, %r103, %r103, %rd86, %rd87, %rd28, %rd28);
|
|
// inline asm
|
|
mov.f32 %f551, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f552, %f551;
|
|
fma.rn.f32 %f553, %f552, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f180, %f553;
|
|
abs.f32 %f181, %f179;
|
|
setp.lt.f32 %p72, %f181, 0f00800000;
|
|
mul.f32 %f554, %f181, 0f4B800000;
|
|
selp.f32 %f555, 0fC3170000, 0fC2FE0000, %p72;
|
|
selp.f32 %f556, %f554, %f181, %p72;
|
|
mov.b32 %r261, %f556;
|
|
and.b32 %r262, %r261, 8388607;
|
|
or.b32 %r263, %r262, 1065353216;
|
|
mov.b32 %f557, %r263;
|
|
shr.u32 %r264, %r261, 23;
|
|
cvt.rn.f32.u32 %f558, %r264;
|
|
add.f32 %f559, %f555, %f558;
|
|
setp.gt.f32 %p73, %f557, 0f3FB504F3;
|
|
mul.f32 %f560, %f557, 0f3F000000;
|
|
add.f32 %f561, %f559, 0f3F800000;
|
|
selp.f32 %f562, %f560, %f557, %p73;
|
|
selp.f32 %f563, %f561, %f559, %p73;
|
|
add.f32 %f564, %f562, 0fBF800000;
|
|
add.f32 %f550, %f562, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f549,%f550;
|
|
// inline asm
|
|
add.f32 %f565, %f564, %f564;
|
|
mul.f32 %f566, %f549, %f565;
|
|
mul.f32 %f567, %f566, %f566;
|
|
mov.f32 %f568, 0f3C4CAF63;
|
|
mov.f32 %f569, 0f3B18F0FE;
|
|
fma.rn.f32 %f570, %f569, %f567, %f568;
|
|
mov.f32 %f571, 0f3DAAAABD;
|
|
fma.rn.f32 %f572, %f570, %f567, %f571;
|
|
mul.rn.f32 %f573, %f572, %f567;
|
|
mul.rn.f32 %f574, %f573, %f566;
|
|
sub.f32 %f575, %f564, %f566;
|
|
neg.f32 %f576, %f566;
|
|
add.f32 %f577, %f575, %f575;
|
|
fma.rn.f32 %f578, %f576, %f564, %f577;
|
|
mul.rn.f32 %f579, %f549, %f578;
|
|
add.f32 %f580, %f574, %f566;
|
|
sub.f32 %f581, %f566, %f580;
|
|
add.f32 %f582, %f574, %f581;
|
|
add.f32 %f583, %f579, %f582;
|
|
add.f32 %f584, %f580, %f583;
|
|
sub.f32 %f585, %f580, %f584;
|
|
add.f32 %f586, %f583, %f585;
|
|
mov.f32 %f587, 0f3F317200;
|
|
mul.rn.f32 %f588, %f563, %f587;
|
|
mov.f32 %f589, 0f35BFBE8E;
|
|
mul.rn.f32 %f590, %f563, %f589;
|
|
add.f32 %f591, %f588, %f584;
|
|
sub.f32 %f592, %f588, %f591;
|
|
add.f32 %f593, %f584, %f592;
|
|
add.f32 %f594, %f586, %f593;
|
|
add.f32 %f595, %f590, %f594;
|
|
add.f32 %f596, %f591, %f595;
|
|
sub.f32 %f597, %f591, %f596;
|
|
add.f32 %f598, %f595, %f597;
|
|
mov.f32 %f599, 0f3EE8BA2E;
|
|
mul.rn.f32 %f600, %f599, %f596;
|
|
neg.f32 %f601, %f600;
|
|
fma.rn.f32 %f602, %f599, %f596, %f601;
|
|
fma.rn.f32 %f603, %f599, %f598, %f602;
|
|
mov.f32 %f604, 0f00000000;
|
|
fma.rn.f32 %f605, %f604, %f596, %f603;
|
|
add.rn.f32 %f606, %f600, %f605;
|
|
neg.f32 %f607, %f606;
|
|
add.rn.f32 %f608, %f600, %f607;
|
|
add.rn.f32 %f609, %f608, %f605;
|
|
mov.b32 %r265, %f606;
|
|
setp.eq.s32 %p74, %r265, 1118925336;
|
|
add.s32 %r266, %r265, -1;
|
|
mov.b32 %f610, %r266;
|
|
add.f32 %f611, %f609, 0f37000000;
|
|
selp.f32 %f612, %f610, %f606, %p74;
|
|
selp.f32 %f182, %f611, %f609, %p74;
|
|
mul.f32 %f613, %f612, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f614, %f613;
|
|
mov.f32 %f615, 0fBF317200;
|
|
fma.rn.f32 %f616, %f614, %f615, %f612;
|
|
mov.f32 %f617, 0fB5BFBE8E;
|
|
fma.rn.f32 %f618, %f614, %f617, %f616;
|
|
mul.f32 %f619, %f618, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f620, %f619;
|
|
add.f32 %f621, %f614, 0f00000000;
|
|
ex2.approx.f32 %f622, %f621;
|
|
mul.f32 %f623, %f620, %f622;
|
|
setp.lt.f32 %p75, %f612, 0fC2D20000;
|
|
selp.f32 %f624, 0f00000000, %f623, %p75;
|
|
setp.gt.f32 %p76, %f612, 0f42D20000;
|
|
selp.f32 %f999, 0f7F800000, %f624, %p76;
|
|
setp.eq.f32 %p77, %f999, 0f7F800000;
|
|
@%p77 bra BB0_83;
|
|
|
|
fma.rn.f32 %f999, %f999, %f182, %f999;
|
|
|
|
BB0_83:
|
|
setp.lt.f32 %p78, %f179, 0f00000000;
|
|
setp.eq.f32 %p79, %f180, 0f3F800000;
|
|
and.pred %p3, %p78, %p79;
|
|
mov.b32 %r267, %f999;
|
|
xor.b32 %r268, %r267, -2147483648;
|
|
mov.b32 %f625, %r268;
|
|
selp.f32 %f1001, %f625, %f999, %p3;
|
|
setp.eq.f32 %p80, %f179, 0f00000000;
|
|
@%p80 bra BB0_86;
|
|
bra.uni BB0_84;
|
|
|
|
BB0_86:
|
|
add.f32 %f628, %f179, %f179;
|
|
selp.f32 %f1001, %f628, 0f00000000, %p79;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_141:
|
|
mov.u64 %rd158, image_HDR;
|
|
cvta.global.u64 %rd153, %rd158;
|
|
mov.u32 %r331, 8;
|
|
mov.u64 %rd157, 0;
|
|
// inline asm
|
|
call (%rd152), _rt_buffer_get_64, (%rd153, %r103, %r331, %rd21, %rd22, %rd157, %rd157);
|
|
// inline asm
|
|
mov.f32 %f912, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f912;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs52, 0;
|
|
st.v4.u16 [%rd152], {%rs51, %rs51, %rs51, %rs52};
|
|
|
|
BB0_142:
|
|
ld.global.u8 %rs53, [imageEnabled];
|
|
and.b16 %rs54, %rs53, 64;
|
|
setp.eq.s16 %p144, %rs54, 0;
|
|
@%p144 bra BB0_144;
|
|
|
|
cvt.u64.u32 %rd161, %r5;
|
|
cvt.u64.u32 %rd162, %r6;
|
|
mov.u64 %rd165, image_Dir;
|
|
cvta.global.u64 %rd160, %rd165;
|
|
mov.u64 %rd164, 0;
|
|
// inline asm
|
|
call (%rd159), _rt_buffer_get_64, (%rd160, %r103, %r104, %rd161, %rd162, %rd164, %rd164);
|
|
// inline asm
|
|
mov.u16 %rs55, 0;
|
|
st.v4.u8 [%rd159], {%rs55, %rs55, %rs55, %rs55};
|
|
bra.uni BB0_144;
|
|
|
|
BB0_84:
|
|
setp.geu.f32 %p81, %f179, 0f00000000;
|
|
@%p81 bra BB0_87;
|
|
|
|
cvt.rzi.f32.f32 %f627, %f599;
|
|
setp.neu.f32 %p82, %f627, 0f3EE8BA2E;
|
|
selp.f32 %f1001, 0f7FFFFFFF, %f1001, %p82;
|
|
|
|
BB0_87:
|
|
add.f32 %f629, %f181, 0f3EE8BA2E;
|
|
mov.b32 %r269, %f629;
|
|
setp.lt.s32 %p84, %r269, 2139095040;
|
|
@%p84 bra BB0_92;
|
|
|
|
setp.gtu.f32 %p85, %f181, 0f7F800000;
|
|
@%p85 bra BB0_91;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_91:
|
|
add.f32 %f1001, %f179, 0f3EE8BA2E;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_89:
|
|
setp.neu.f32 %p86, %f181, 0f7F800000;
|
|
@%p86 bra BB0_92;
|
|
|
|
selp.f32 %f1001, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_92:
|
|
mul.f32 %f630, %f1001, 0f437F0000;
|
|
setp.eq.f32 %p87, %f179, 0f3F800000;
|
|
selp.f32 %f631, 0f437F0000, %f630, %p87;
|
|
cvt.rzi.u32.f32 %r270, %f631;
|
|
cvt.u16.u32 %rs11, %r270;
|
|
mov.u16 %rs12, 255;
|
|
st.v2.u8 [%rd84], {%rs11, %rs12};
|
|
ld.global.u32 %r368, [imageEnabled];
|
|
|
|
BB0_93:
|
|
ld.global.f32 %f632, [lightColor];
|
|
mul.f32 %f193, %f995, %f632;
|
|
ld.global.f32 %f633, [lightColor+4];
|
|
mul.f32 %f194, %f996, %f633;
|
|
ld.global.f32 %f634, [lightColor+8];
|
|
mul.f32 %f195, %f997, %f634;
|
|
and.b32 %r271, %r368, 1;
|
|
setp.eq.b32 %p88, %r271, 1;
|
|
@!%p88 bra BB0_128;
|
|
bra.uni BB0_94;
|
|
|
|
BB0_94:
|
|
mov.f32 %f637, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f638, %f637;
|
|
fma.rn.f32 %f639, %f638, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f196, %f639;
|
|
abs.f32 %f197, %f193;
|
|
setp.lt.f32 %p89, %f197, 0f00800000;
|
|
mul.f32 %f640, %f197, 0f4B800000;
|
|
selp.f32 %f641, 0fC3170000, 0fC2FE0000, %p89;
|
|
selp.f32 %f642, %f640, %f197, %p89;
|
|
mov.b32 %r272, %f642;
|
|
and.b32 %r273, %r272, 8388607;
|
|
or.b32 %r274, %r273, 1065353216;
|
|
mov.b32 %f643, %r274;
|
|
shr.u32 %r275, %r272, 23;
|
|
cvt.rn.f32.u32 %f644, %r275;
|
|
add.f32 %f645, %f641, %f644;
|
|
setp.gt.f32 %p90, %f643, 0f3FB504F3;
|
|
mul.f32 %f646, %f643, 0f3F000000;
|
|
add.f32 %f647, %f645, 0f3F800000;
|
|
selp.f32 %f648, %f646, %f643, %p90;
|
|
selp.f32 %f649, %f647, %f645, %p90;
|
|
add.f32 %f650, %f648, 0fBF800000;
|
|
add.f32 %f636, %f648, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f635,%f636;
|
|
// inline asm
|
|
add.f32 %f651, %f650, %f650;
|
|
mul.f32 %f652, %f635, %f651;
|
|
mul.f32 %f653, %f652, %f652;
|
|
mov.f32 %f654, 0f3C4CAF63;
|
|
mov.f32 %f655, 0f3B18F0FE;
|
|
fma.rn.f32 %f656, %f655, %f653, %f654;
|
|
mov.f32 %f657, 0f3DAAAABD;
|
|
fma.rn.f32 %f658, %f656, %f653, %f657;
|
|
mul.rn.f32 %f659, %f658, %f653;
|
|
mul.rn.f32 %f660, %f659, %f652;
|
|
sub.f32 %f661, %f650, %f652;
|
|
neg.f32 %f662, %f652;
|
|
add.f32 %f663, %f661, %f661;
|
|
fma.rn.f32 %f664, %f662, %f650, %f663;
|
|
mul.rn.f32 %f665, %f635, %f664;
|
|
add.f32 %f666, %f660, %f652;
|
|
sub.f32 %f667, %f652, %f666;
|
|
add.f32 %f668, %f660, %f667;
|
|
add.f32 %f669, %f665, %f668;
|
|
add.f32 %f670, %f666, %f669;
|
|
sub.f32 %f671, %f666, %f670;
|
|
add.f32 %f672, %f669, %f671;
|
|
mov.f32 %f673, 0f3F317200;
|
|
mul.rn.f32 %f674, %f649, %f673;
|
|
mov.f32 %f675, 0f35BFBE8E;
|
|
mul.rn.f32 %f676, %f649, %f675;
|
|
add.f32 %f677, %f674, %f670;
|
|
sub.f32 %f678, %f674, %f677;
|
|
add.f32 %f679, %f670, %f678;
|
|
add.f32 %f680, %f672, %f679;
|
|
add.f32 %f681, %f676, %f680;
|
|
add.f32 %f682, %f677, %f681;
|
|
sub.f32 %f683, %f677, %f682;
|
|
add.f32 %f684, %f681, %f683;
|
|
mov.f32 %f685, 0f3EE66666;
|
|
mul.rn.f32 %f686, %f685, %f682;
|
|
neg.f32 %f687, %f686;
|
|
fma.rn.f32 %f688, %f685, %f682, %f687;
|
|
fma.rn.f32 %f689, %f685, %f684, %f688;
|
|
mov.f32 %f690, 0f00000000;
|
|
fma.rn.f32 %f691, %f690, %f682, %f689;
|
|
add.rn.f32 %f692, %f686, %f691;
|
|
neg.f32 %f693, %f692;
|
|
add.rn.f32 %f694, %f686, %f693;
|
|
add.rn.f32 %f695, %f694, %f691;
|
|
mov.b32 %r276, %f692;
|
|
setp.eq.s32 %p91, %r276, 1118925336;
|
|
add.s32 %r277, %r276, -1;
|
|
mov.b32 %f696, %r277;
|
|
add.f32 %f697, %f695, 0f37000000;
|
|
selp.f32 %f698, %f696, %f692, %p91;
|
|
selp.f32 %f198, %f697, %f695, %p91;
|
|
mul.f32 %f699, %f698, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f700, %f699;
|
|
mov.f32 %f701, 0fBF317200;
|
|
fma.rn.f32 %f702, %f700, %f701, %f698;
|
|
mov.f32 %f703, 0fB5BFBE8E;
|
|
fma.rn.f32 %f704, %f700, %f703, %f702;
|
|
mul.f32 %f705, %f704, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f706, %f705;
|
|
add.f32 %f707, %f700, 0f00000000;
|
|
ex2.approx.f32 %f708, %f707;
|
|
mul.f32 %f709, %f706, %f708;
|
|
setp.lt.f32 %p92, %f698, 0fC2D20000;
|
|
selp.f32 %f710, 0f00000000, %f709, %p92;
|
|
setp.gt.f32 %p93, %f698, 0f42D20000;
|
|
selp.f32 %f1002, 0f7F800000, %f710, %p93;
|
|
setp.eq.f32 %p94, %f1002, 0f7F800000;
|
|
@%p94 bra BB0_96;
|
|
|
|
fma.rn.f32 %f1002, %f1002, %f198, %f1002;
|
|
|
|
BB0_96:
|
|
setp.lt.f32 %p95, %f193, 0f00000000;
|
|
setp.eq.f32 %p96, %f196, 0f3F800000;
|
|
and.pred %p4, %p95, %p96;
|
|
mov.b32 %r278, %f1002;
|
|
xor.b32 %r279, %r278, -2147483648;
|
|
mov.b32 %f711, %r279;
|
|
selp.f32 %f1004, %f711, %f1002, %p4;
|
|
setp.eq.f32 %p97, %f193, 0f00000000;
|
|
@%p97 bra BB0_99;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_99:
|
|
add.f32 %f714, %f193, %f193;
|
|
selp.f32 %f1004, %f714, 0f00000000, %p96;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_97:
|
|
setp.geu.f32 %p98, %f193, 0f00000000;
|
|
@%p98 bra BB0_100;
|
|
|
|
cvt.rzi.f32.f32 %f713, %f685;
|
|
setp.neu.f32 %p99, %f713, 0f3EE66666;
|
|
selp.f32 %f1004, 0f7FFFFFFF, %f1004, %p99;
|
|
|
|
BB0_100:
|
|
add.f32 %f715, %f197, 0f3EE66666;
|
|
mov.b32 %r280, %f715;
|
|
setp.lt.s32 %p101, %r280, 2139095040;
|
|
@%p101 bra BB0_105;
|
|
|
|
setp.gtu.f32 %p102, %f197, 0f7F800000;
|
|
@%p102 bra BB0_104;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_104:
|
|
add.f32 %f1004, %f193, 0f3EE66666;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_102:
|
|
setp.neu.f32 %p103, %f197, 0f7F800000;
|
|
@%p103 bra BB0_105;
|
|
|
|
selp.f32 %f1004, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_105:
|
|
setp.eq.f32 %p104, %f193, 0f3F800000;
|
|
selp.f32 %f209, 0f3F800000, %f1004, %p104;
|
|
abs.f32 %f210, %f194;
|
|
setp.lt.f32 %p105, %f210, 0f00800000;
|
|
mul.f32 %f718, %f210, 0f4B800000;
|
|
selp.f32 %f719, 0fC3170000, 0fC2FE0000, %p105;
|
|
selp.f32 %f720, %f718, %f210, %p105;
|
|
mov.b32 %r281, %f720;
|
|
and.b32 %r282, %r281, 8388607;
|
|
or.b32 %r283, %r282, 1065353216;
|
|
mov.b32 %f721, %r283;
|
|
shr.u32 %r284, %r281, 23;
|
|
cvt.rn.f32.u32 %f722, %r284;
|
|
add.f32 %f723, %f719, %f722;
|
|
setp.gt.f32 %p106, %f721, 0f3FB504F3;
|
|
mul.f32 %f724, %f721, 0f3F000000;
|
|
add.f32 %f725, %f723, 0f3F800000;
|
|
selp.f32 %f726, %f724, %f721, %p106;
|
|
selp.f32 %f727, %f725, %f723, %p106;
|
|
add.f32 %f728, %f726, 0fBF800000;
|
|
add.f32 %f717, %f726, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f716,%f717;
|
|
// inline asm
|
|
add.f32 %f729, %f728, %f728;
|
|
mul.f32 %f730, %f716, %f729;
|
|
mul.f32 %f731, %f730, %f730;
|
|
fma.rn.f32 %f734, %f655, %f731, %f654;
|
|
fma.rn.f32 %f736, %f734, %f731, %f657;
|
|
mul.rn.f32 %f737, %f736, %f731;
|
|
mul.rn.f32 %f738, %f737, %f730;
|
|
sub.f32 %f739, %f728, %f730;
|
|
neg.f32 %f740, %f730;
|
|
add.f32 %f741, %f739, %f739;
|
|
fma.rn.f32 %f742, %f740, %f728, %f741;
|
|
mul.rn.f32 %f743, %f716, %f742;
|
|
add.f32 %f744, %f738, %f730;
|
|
sub.f32 %f745, %f730, %f744;
|
|
add.f32 %f746, %f738, %f745;
|
|
add.f32 %f747, %f743, %f746;
|
|
add.f32 %f748, %f744, %f747;
|
|
sub.f32 %f749, %f744, %f748;
|
|
add.f32 %f750, %f747, %f749;
|
|
mul.rn.f32 %f752, %f727, %f673;
|
|
mul.rn.f32 %f754, %f727, %f675;
|
|
add.f32 %f755, %f752, %f748;
|
|
sub.f32 %f756, %f752, %f755;
|
|
add.f32 %f757, %f748, %f756;
|
|
add.f32 %f758, %f750, %f757;
|
|
add.f32 %f759, %f754, %f758;
|
|
add.f32 %f760, %f755, %f759;
|
|
sub.f32 %f761, %f755, %f760;
|
|
add.f32 %f762, %f759, %f761;
|
|
mul.rn.f32 %f764, %f685, %f760;
|
|
neg.f32 %f765, %f764;
|
|
fma.rn.f32 %f766, %f685, %f760, %f765;
|
|
fma.rn.f32 %f767, %f685, %f762, %f766;
|
|
fma.rn.f32 %f769, %f690, %f760, %f767;
|
|
add.rn.f32 %f770, %f764, %f769;
|
|
neg.f32 %f771, %f770;
|
|
add.rn.f32 %f772, %f764, %f771;
|
|
add.rn.f32 %f773, %f772, %f769;
|
|
mov.b32 %r285, %f770;
|
|
setp.eq.s32 %p107, %r285, 1118925336;
|
|
add.s32 %r286, %r285, -1;
|
|
mov.b32 %f774, %r286;
|
|
add.f32 %f775, %f773, 0f37000000;
|
|
selp.f32 %f776, %f774, %f770, %p107;
|
|
selp.f32 %f211, %f775, %f773, %p107;
|
|
mul.f32 %f777, %f776, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f778, %f777;
|
|
fma.rn.f32 %f780, %f778, %f701, %f776;
|
|
fma.rn.f32 %f782, %f778, %f703, %f780;
|
|
mul.f32 %f783, %f782, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f784, %f783;
|
|
add.f32 %f785, %f778, 0f00000000;
|
|
ex2.approx.f32 %f786, %f785;
|
|
mul.f32 %f787, %f784, %f786;
|
|
setp.lt.f32 %p108, %f776, 0fC2D20000;
|
|
selp.f32 %f788, 0f00000000, %f787, %p108;
|
|
setp.gt.f32 %p109, %f776, 0f42D20000;
|
|
selp.f32 %f1005, 0f7F800000, %f788, %p109;
|
|
setp.eq.f32 %p110, %f1005, 0f7F800000;
|
|
@%p110 bra BB0_107;
|
|
|
|
fma.rn.f32 %f1005, %f1005, %f211, %f1005;
|
|
|
|
BB0_107:
|
|
setp.lt.f32 %p111, %f194, 0f00000000;
|
|
and.pred %p5, %p111, %p96;
|
|
mov.b32 %r287, %f1005;
|
|
xor.b32 %r288, %r287, -2147483648;
|
|
mov.b32 %f789, %r288;
|
|
selp.f32 %f1007, %f789, %f1005, %p5;
|
|
setp.eq.f32 %p113, %f194, 0f00000000;
|
|
@%p113 bra BB0_110;
|
|
bra.uni BB0_108;
|
|
|
|
BB0_110:
|
|
add.f32 %f792, %f194, %f194;
|
|
selp.f32 %f1007, %f792, 0f00000000, %p96;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_108:
|
|
setp.geu.f32 %p114, %f194, 0f00000000;
|
|
@%p114 bra BB0_111;
|
|
|
|
cvt.rzi.f32.f32 %f791, %f685;
|
|
setp.neu.f32 %p115, %f791, 0f3EE66666;
|
|
selp.f32 %f1007, 0f7FFFFFFF, %f1007, %p115;
|
|
|
|
BB0_111:
|
|
add.f32 %f793, %f210, 0f3EE66666;
|
|
mov.b32 %r289, %f793;
|
|
setp.lt.s32 %p117, %r289, 2139095040;
|
|
@%p117 bra BB0_116;
|
|
|
|
setp.gtu.f32 %p118, %f210, 0f7F800000;
|
|
@%p118 bra BB0_115;
|
|
bra.uni BB0_113;
|
|
|
|
BB0_115:
|
|
add.f32 %f1007, %f194, 0f3EE66666;
|
|
bra.uni BB0_116;
|
|
|
|
BB0_113:
|
|
setp.neu.f32 %p119, %f210, 0f7F800000;
|
|
@%p119 bra BB0_116;
|
|
|
|
selp.f32 %f1007, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_116:
|
|
setp.eq.f32 %p120, %f194, 0f3F800000;
|
|
selp.f32 %f222, 0f3F800000, %f1007, %p120;
|
|
abs.f32 %f223, %f195;
|
|
setp.lt.f32 %p121, %f223, 0f00800000;
|
|
mul.f32 %f796, %f223, 0f4B800000;
|
|
selp.f32 %f797, 0fC3170000, 0fC2FE0000, %p121;
|
|
selp.f32 %f798, %f796, %f223, %p121;
|
|
mov.b32 %r290, %f798;
|
|
and.b32 %r291, %r290, 8388607;
|
|
or.b32 %r292, %r291, 1065353216;
|
|
mov.b32 %f799, %r292;
|
|
shr.u32 %r293, %r290, 23;
|
|
cvt.rn.f32.u32 %f800, %r293;
|
|
add.f32 %f801, %f797, %f800;
|
|
setp.gt.f32 %p122, %f799, 0f3FB504F3;
|
|
mul.f32 %f802, %f799, 0f3F000000;
|
|
add.f32 %f803, %f801, 0f3F800000;
|
|
selp.f32 %f804, %f802, %f799, %p122;
|
|
selp.f32 %f805, %f803, %f801, %p122;
|
|
add.f32 %f806, %f804, 0fBF800000;
|
|
add.f32 %f795, %f804, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f794,%f795;
|
|
// inline asm
|
|
add.f32 %f807, %f806, %f806;
|
|
mul.f32 %f808, %f794, %f807;
|
|
mul.f32 %f809, %f808, %f808;
|
|
fma.rn.f32 %f812, %f655, %f809, %f654;
|
|
fma.rn.f32 %f814, %f812, %f809, %f657;
|
|
mul.rn.f32 %f815, %f814, %f809;
|
|
mul.rn.f32 %f816, %f815, %f808;
|
|
sub.f32 %f817, %f806, %f808;
|
|
neg.f32 %f818, %f808;
|
|
add.f32 %f819, %f817, %f817;
|
|
fma.rn.f32 %f820, %f818, %f806, %f819;
|
|
mul.rn.f32 %f821, %f794, %f820;
|
|
add.f32 %f822, %f816, %f808;
|
|
sub.f32 %f823, %f808, %f822;
|
|
add.f32 %f824, %f816, %f823;
|
|
add.f32 %f825, %f821, %f824;
|
|
add.f32 %f826, %f822, %f825;
|
|
sub.f32 %f827, %f822, %f826;
|
|
add.f32 %f828, %f825, %f827;
|
|
mul.rn.f32 %f830, %f805, %f673;
|
|
mul.rn.f32 %f832, %f805, %f675;
|
|
add.f32 %f833, %f830, %f826;
|
|
sub.f32 %f834, %f830, %f833;
|
|
add.f32 %f835, %f826, %f834;
|
|
add.f32 %f836, %f828, %f835;
|
|
add.f32 %f837, %f832, %f836;
|
|
add.f32 %f838, %f833, %f837;
|
|
sub.f32 %f839, %f833, %f838;
|
|
add.f32 %f840, %f837, %f839;
|
|
mul.rn.f32 %f842, %f685, %f838;
|
|
neg.f32 %f843, %f842;
|
|
fma.rn.f32 %f844, %f685, %f838, %f843;
|
|
fma.rn.f32 %f845, %f685, %f840, %f844;
|
|
fma.rn.f32 %f847, %f690, %f838, %f845;
|
|
add.rn.f32 %f848, %f842, %f847;
|
|
neg.f32 %f849, %f848;
|
|
add.rn.f32 %f850, %f842, %f849;
|
|
add.rn.f32 %f851, %f850, %f847;
|
|
mov.b32 %r294, %f848;
|
|
setp.eq.s32 %p123, %r294, 1118925336;
|
|
add.s32 %r295, %r294, -1;
|
|
mov.b32 %f852, %r295;
|
|
add.f32 %f853, %f851, 0f37000000;
|
|
selp.f32 %f854, %f852, %f848, %p123;
|
|
selp.f32 %f224, %f853, %f851, %p123;
|
|
mul.f32 %f855, %f854, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f856, %f855;
|
|
fma.rn.f32 %f858, %f856, %f701, %f854;
|
|
fma.rn.f32 %f860, %f856, %f703, %f858;
|
|
mul.f32 %f861, %f860, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f862, %f861;
|
|
add.f32 %f863, %f856, 0f00000000;
|
|
ex2.approx.f32 %f864, %f863;
|
|
mul.f32 %f865, %f862, %f864;
|
|
setp.lt.f32 %p124, %f854, 0fC2D20000;
|
|
selp.f32 %f866, 0f00000000, %f865, %p124;
|
|
setp.gt.f32 %p125, %f854, 0f42D20000;
|
|
selp.f32 %f1008, 0f7F800000, %f866, %p125;
|
|
setp.eq.f32 %p126, %f1008, 0f7F800000;
|
|
@%p126 bra BB0_118;
|
|
|
|
fma.rn.f32 %f1008, %f1008, %f224, %f1008;
|
|
|
|
BB0_118:
|
|
setp.lt.f32 %p127, %f195, 0f00000000;
|
|
and.pred %p6, %p127, %p96;
|
|
mov.b32 %r296, %f1008;
|
|
xor.b32 %r297, %r296, -2147483648;
|
|
mov.b32 %f867, %r297;
|
|
selp.f32 %f1010, %f867, %f1008, %p6;
|
|
setp.eq.f32 %p129, %f195, 0f00000000;
|
|
@%p129 bra BB0_121;
|
|
bra.uni BB0_119;
|
|
|
|
BB0_121:
|
|
add.f32 %f870, %f195, %f195;
|
|
selp.f32 %f1010, %f870, 0f00000000, %p96;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_119:
|
|
setp.geu.f32 %p130, %f195, 0f00000000;
|
|
@%p130 bra BB0_122;
|
|
|
|
cvt.rzi.f32.f32 %f869, %f685;
|
|
setp.neu.f32 %p131, %f869, 0f3EE66666;
|
|
selp.f32 %f1010, 0f7FFFFFFF, %f1010, %p131;
|
|
|
|
BB0_122:
|
|
add.f32 %f871, %f223, 0f3EE66666;
|
|
mov.b32 %r298, %f871;
|
|
setp.lt.s32 %p133, %r298, 2139095040;
|
|
@%p133 bra BB0_127;
|
|
|
|
setp.gtu.f32 %p134, %f223, 0f7F800000;
|
|
@%p134 bra BB0_126;
|
|
bra.uni BB0_124;
|
|
|
|
BB0_126:
|
|
add.f32 %f1010, %f195, 0f3EE66666;
|
|
bra.uni BB0_127;
|
|
|
|
BB0_124:
|
|
setp.neu.f32 %p135, %f223, 0f7F800000;
|
|
@%p135 bra BB0_127;
|
|
|
|
selp.f32 %f1010, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_127:
|
|
mov.u32 %r336, 4;
|
|
setp.eq.f32 %p136, %f195, 0f3F800000;
|
|
selp.f32 %f872, 0f3F800000, %f1010, %p136;
|
|
cvt.u64.u32 %rd94, %r6;
|
|
cvt.u64.u32 %rd93, %r5;
|
|
mov.u64 %rd97, image;
|
|
cvta.global.u64 %rd92, %rd97;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd92, %r103, %r336, %rd93, %rd94, %rd28, %rd28);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f873, %f872;
|
|
mul.f32 %f874, %f873, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r301, %f874;
|
|
cvt.sat.f32.f32 %f875, %f222;
|
|
mul.f32 %f876, %f875, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r302, %f876;
|
|
cvt.sat.f32.f32 %f877, %f209;
|
|
mul.f32 %f878, %f877, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r303, %f878;
|
|
cvt.u16.u32 %rs13, %r301;
|
|
cvt.u16.u32 %rs14, %r303;
|
|
cvt.u16.u32 %rs15, %r302;
|
|
mov.u16 %rs16, 255;
|
|
st.v4.u8 [%rd91], {%rs13, %rs15, %rs14, %rs16};
|
|
ld.global.u32 %r368, [imageEnabled];
|
|
|
|
BB0_128:
|
|
and.b32 %r304, %r368, 4;
|
|
setp.eq.s32 %p137, %r304, 0;
|
|
@%p137 bra BB0_132;
|
|
|
|
ld.global.u32 %r305, [additive];
|
|
setp.eq.s32 %p138, %r305, 0;
|
|
cvt.u64.u32 %rd19, %r5;
|
|
cvt.u64.u32 %rd20, %r6;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs17, %f417;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_131;
|
|
|
|
mov.u64 %rd110, image_HDR;
|
|
cvta.global.u64 %rd99, %rd110;
|
|
mov.u32 %r309, 8;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r103, %r309, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs24, %rs25, %rs26, %rs27}, [%rd98];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f880, %rs24;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f881, %rs25;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f882, %rs26;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd99, %r103, %r309, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f883, %f193, %f880;
|
|
add.f32 %f884, %f194, %f881;
|
|
add.f32 %f885, %f195, %f882;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f885;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs22, %f884;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f883;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd104], {%rs21, %rs22, %rs23, %rs17};
|
|
bra.uni BB0_132;
|
|
|
|
BB0_131:
|
|
mov.u64 %rd117, image_HDR;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r311, 8;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r103, %r311, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f195;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f194;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f193;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd111], {%rs28, %rs29, %rs30, %rs17};
|
|
|
|
BB0_132:
|
|
ld.global.u8 %rs31, [imageEnabled];
|
|
and.b16 %rs32, %rs31, 64;
|
|
setp.eq.s16 %p139, %rs32, 0;
|
|
@%p139 bra BB0_144;
|
|
|
|
mov.u32 %r337, 4;
|
|
mul.f32 %f889, %f954, %f954;
|
|
fma.rn.f32 %f890, %f955, %f955, %f889;
|
|
fma.rn.f32 %f891, %f953, %f953, %f890;
|
|
sqrt.rn.f32 %f892, %f891;
|
|
rcp.rn.f32 %f893, %f892;
|
|
mul.f32 %f894, %f955, %f893;
|
|
mul.f32 %f895, %f954, %f893;
|
|
mul.f32 %f896, %f953, %f893;
|
|
cvt.u64.u32 %rd121, %r6;
|
|
cvt.u64.u32 %rd120, %r5;
|
|
mov.u64 %rd124, image_Dir;
|
|
cvta.global.u64 %rd119, %rd124;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r103, %r337, %rd120, %rd121, %rd28, %rd28);
|
|
// inline asm
|
|
fma.rn.f32 %f897, %f894, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f898, %f897, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r314, %f898;
|
|
fma.rn.f32 %f899, %f895, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f900, %f899, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r315, %f900;
|
|
fma.rn.f32 %f901, %f896, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f902, %f901, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r316, %f902;
|
|
cvt.u16.u32 %rs33, %r316;
|
|
cvt.u16.u32 %rs34, %r315;
|
|
cvt.u16.u32 %rs35, %r314;
|
|
mov.u16 %rs36, 255;
|
|
st.v4.u8 [%rd118], {%rs35, %rs34, %rs33, %rs36};
|
|
|
|
BB0_144:
|
|
ret;
|
|
}
|
|
|
|
|