Files
FORT-BO/Assets/Editor/x64/Bakery/lmSunTranslucent.ptx
2026-08-06 18:56:32 +03:00

1826 lines
52 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 rnd_seeds[1];
.global .align 4 .b8 directDir[12];
.global .align 4 .b8 directColor[12];
.global .align 4 .f32 shadowSpread;
.global .align 4 .u32 samples;
.global .align 4 .u32 ignoreNormal;
.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_typeinfo9directDirE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo11directColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12shadowSpreadE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12ignoreNormalE[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 8 .b8 _ZN21rti_internal_typename9directDirE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename11directColorE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename12shadowSpreadE[6] = {102, 108, 111, 97, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[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_typeenum9directDirE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum11directColorE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12shadowSpreadE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 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_semantic9directDirE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic11directColorE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12shadowSpreadE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[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_annotation9directDirE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation11directColorE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12shadowSpreadE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[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 4 .b8 __local_depot0[40];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<129>;
.reg .b16 %rs<76>;
.reg .f32 %f<762>;
.reg .b32 %r<371>;
.reg .b64 %rd<198>;
mov.u64 %rd197, __local_depot0;
cvta.local.u64 %SP, %rd197;
ld.global.v2.u32 {%r106, %r107}, [pixelID];
cvt.u64.u32 %rd25, %r106;
cvt.u64.u32 %rd26, %r107;
mov.u64 %rd29, uvnormal;
cvta.global.u64 %rd24, %rd29;
mov.u32 %r104, 2;
mov.u32 %r105, 4;
mov.u64 %rd28, 0;
// inline asm
call (%rd23), _rt_buffer_get_64, (%rd24, %r104, %r105, %rd25, %rd26, %rd28, %rd28);
// inline asm
ld.u32 %r1, [%rd23];
shr.u32 %r110, %r1, 16;
cvt.u16.u32 %rs1, %r110;
and.b16 %rs4, %rs1, 255;
cvt.u16.u32 %rs5, %r1;
or.b16 %rs6, %rs5, %rs4;
setp.eq.s16 %p5, %rs6, 0;
mov.f32 %f720, 0f00000000;
mov.f32 %f721, %f720;
mov.f32 %f722, %f720;
@%p5 bra BB0_2;
ld.u8 %rs7, [%rd23+1];
and.b16 %rs9, %rs5, 255;
cvt.rn.f32.u16 %f142, %rs9;
div.rn.f32 %f143, %f142, 0f437F0000;
fma.rn.f32 %f144, %f143, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f145, %rs7;
div.rn.f32 %f146, %f145, 0f437F0000;
fma.rn.f32 %f147, %f146, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f148, %rs4;
div.rn.f32 %f149, %f148, 0f437F0000;
fma.rn.f32 %f150, %f149, 0f40000000, 0fBF800000;
mul.f32 %f151, %f147, %f147;
fma.rn.f32 %f152, %f144, %f144, %f151;
fma.rn.f32 %f153, %f150, %f150, %f152;
sqrt.rn.f32 %f154, %f153;
rcp.rn.f32 %f155, %f154;
mul.f32 %f720, %f144, %f155;
mul.f32 %f721, %f147, %f155;
mul.f32 %f722, %f150, %f155;
BB0_2:
ld.global.v2.u32 {%r111, %r112}, [pixelID];
ld.global.v2.u32 {%r114, %r115}, [tileInfo];
add.s32 %r2, %r111, %r114;
add.s32 %r3, %r112, %r115;
setp.eq.f32 %p6, %f721, 0f00000000;
setp.eq.f32 %p7, %f720, 0f00000000;
and.pred %p8, %p7, %p6;
setp.eq.f32 %p9, %f722, 0f00000000;
and.pred %p10, %p8, %p9;
@%p10 bra BB0_123;
bra.uni BB0_3;
BB0_123:
ld.global.u32 %r370, [imageEnabled];
and.b32 %r314, %r370, 1;
setp.eq.b32 %p124, %r314, 1;
@!%p124 bra BB0_125;
bra.uni BB0_124;
BB0_124:
cvt.u64.u32 %rd150, %r2;
cvt.u64.u32 %rd151, %r3;
mov.u64 %rd154, image;
cvta.global.u64 %rd149, %rd154;
mov.u64 %rd153, 0;
// inline asm
call (%rd148), _rt_buffer_get_64, (%rd149, %r104, %r105, %rd150, %rd151, %rd153, %rd153);
// inline asm
mov.u16 %rs57, 0;
st.v4.u8 [%rd148], {%rs57, %rs57, %rs57, %rs57};
ld.global.u32 %r370, [imageEnabled];
BB0_125:
and.b32 %r317, %r370, 8;
setp.eq.s32 %p125, %r317, 0;
@%p125 bra BB0_127;
cvt.u64.u32 %rd157, %r2;
cvt.u64.u32 %rd158, %r3;
mov.u64 %rd161, image_Mask;
cvta.global.u64 %rd156, %rd161;
mov.u64 %rd160, 0;
// inline asm
call (%rd155), _rt_buffer_get_64, (%rd156, %r104, %r104, %rd157, %rd158, %rd160, %rd160);
// inline asm
mov.f32 %f684, 0f00000000;
cvt.rzi.u32.f32 %r320, %f684;
cvt.u16.u32 %rs58, %r320;
mov.u16 %rs59, 0;
st.v2.u8 [%rd155], {%rs58, %rs59};
ld.global.u32 %r370, [imageEnabled];
BB0_127:
and.b32 %r321, %r370, 4;
setp.eq.s32 %p126, %r321, 0;
@%p126 bra BB0_131;
ld.global.u32 %r322, [additive];
setp.eq.s32 %p127, %r322, 0;
cvt.u64.u32 %rd21, %r2;
cvt.u64.u32 %rd22, %r3;
@%p127 bra BB0_130;
mov.u64 %rd174, image_HDR;
cvta.global.u64 %rd163, %rd174;
mov.u32 %r326, 8;
mov.u64 %rd173, 0;
// inline asm
call (%rd162), _rt_buffer_get_64, (%rd163, %r104, %r326, %rd21, %rd22, %rd173, %rd173);
// inline asm
ld.v4.u16 {%rs66, %rs67, %rs68, %rs69}, [%rd162];
// inline asm
{ cvt.f32.f16 %f685, %rs66;}
// inline asm
// inline asm
{ cvt.f32.f16 %f686, %rs67;}
// inline asm
// inline asm
{ cvt.f32.f16 %f687, %rs68;}
// inline asm
// inline asm
call (%rd168), _rt_buffer_get_64, (%rd163, %r104, %r326, %rd21, %rd22, %rd173, %rd173);
// inline asm
add.f32 %f688, %f685, 0f00000000;
add.f32 %f689, %f686, 0f00000000;
add.f32 %f690, %f687, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs65, %f690;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs64, %f689;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs63, %f688;}
// inline asm
mov.u16 %rs70, 0;
st.v4.u16 [%rd168], {%rs63, %rs64, %rs65, %rs70};
bra.uni BB0_131;
BB0_3:
ld.global.f32 %f9, [directDir+8];
ld.global.f32 %f8, [directDir+4];
ld.global.f32 %f7, [directDir];
mul.f32 %f156, %f720, %f7;
mul.f32 %f157, %f721, %f8;
neg.f32 %f158, %f157;
sub.f32 %f159, %f158, %f156;
mul.f32 %f160, %f722, %f9;
sub.f32 %f10, %f159, %f160;
ld.global.u32 %r119, [ignoreNormal];
setp.eq.s32 %p11, %r119, 0;
setp.le.f32 %p12, %f10, 0f00000000;
and.pred %p13, %p11, %p12;
ld.global.u32 %r368, [imageEnabled];
and.b32 %r120, %r368, 32;
setp.eq.s32 %p14, %r120, 0;
and.pred %p15, %p13, %p14;
@%p15 bra BB0_113;
bra.uni BB0_4;
BB0_113:
and.b32 %r297, %r368, 1;
setp.eq.b32 %p119, %r297, 1;
@!%p119 bra BB0_115;
bra.uni BB0_114;
BB0_114:
cvt.u64.u32 %rd109, %r2;
cvt.u64.u32 %rd110, %r3;
mov.u64 %rd113, image;
cvta.global.u64 %rd108, %rd113;
mov.u64 %rd112, 0;
// inline asm
call (%rd107), _rt_buffer_get_64, (%rd108, %r104, %r105, %rd109, %rd110, %rd112, %rd112);
// inline asm
mov.u16 %rs37, 255;
mov.u16 %rs38, 0;
st.v4.u8 [%rd107], {%rs38, %rs38, %rs38, %rs37};
ld.global.u32 %r368, [imageEnabled];
BB0_115:
and.b32 %r300, %r368, 8;
setp.eq.s32 %p120, %r300, 0;
@%p120 bra BB0_117;
cvt.u64.u32 %rd116, %r2;
cvt.u64.u32 %rd117, %r3;
mov.u64 %rd120, image_Mask;
cvta.global.u64 %rd115, %rd120;
mov.u64 %rd119, 0;
// inline asm
call (%rd114), _rt_buffer_get_64, (%rd115, %r104, %r104, %rd116, %rd117, %rd119, %rd119);
// inline asm
mov.f32 %f673, 0f00000000;
cvt.rzi.u32.f32 %r303, %f673;
cvt.u16.u32 %rs39, %r303;
mov.u16 %rs40, 255;
st.v2.u8 [%rd114], {%rs39, %rs40};
ld.global.u32 %r368, [imageEnabled];
BB0_117:
and.b32 %r304, %r368, 4;
setp.eq.s32 %p121, %r304, 0;
@%p121 bra BB0_121;
ld.global.u32 %r305, [additive];
setp.eq.s32 %p122, %r305, 0;
cvt.u64.u32 %rd19, %r2;
cvt.u64.u32 %rd20, %r3;
mov.f32 %f674, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs41, %f674;}
// inline asm
@%p122 bra BB0_120;
mov.u64 %rd133, image_HDR;
cvta.global.u64 %rd122, %rd133;
mov.u32 %r309, 8;
mov.u64 %rd132, 0;
// inline asm
call (%rd121), _rt_buffer_get_64, (%rd122, %r104, %r309, %rd19, %rd20, %rd132, %rd132);
// inline asm
ld.v4.u16 {%rs48, %rs49, %rs50, %rs51}, [%rd121];
// inline asm
{ cvt.f32.f16 %f675, %rs48;}
// inline asm
// inline asm
{ cvt.f32.f16 %f676, %rs49;}
// inline asm
// inline asm
{ cvt.f32.f16 %f677, %rs50;}
// inline asm
// inline asm
call (%rd127), _rt_buffer_get_64, (%rd122, %r104, %r309, %rd19, %rd20, %rd132, %rd132);
// inline asm
add.f32 %f678, %f675, 0f00000000;
add.f32 %f679, %f676, 0f00000000;
add.f32 %f680, %f677, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs47, %f680;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs46, %f679;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs45, %f678;}
// inline asm
st.v4.u16 [%rd127], {%rs45, %rs46, %rs47, %rs41};
bra.uni BB0_121;
BB0_4:
ld.global.v2.u32 {%r125, %r126}, [pixelID];
cvt.u64.u32 %rd32, %r125;
cvt.u64.u32 %rd33, %r126;
mov.u64 %rd42, uvpos;
cvta.global.u64 %rd31, %rd42;
mov.u32 %r122, 12;
// inline asm
call (%rd30), _rt_buffer_get_64, (%rd31, %r104, %r122, %rd32, %rd33, %rd28, %rd28);
// inline asm
ld.f32 %f164, [%rd30+8];
ld.f32 %f165, [%rd30+4];
ld.f32 %f166, [%rd30];
mul.f32 %f167, %f166, 0f3456BF95;
mul.f32 %f168, %f165, 0f3456BF95;
mul.f32 %f169, %f164, 0f3456BF95;
abs.f32 %f170, %f720;
div.rn.f32 %f171, %f167, %f170;
abs.f32 %f172, %f721;
div.rn.f32 %f173, %f168, %f172;
abs.f32 %f174, %f722;
div.rn.f32 %f175, %f169, %f174;
abs.f32 %f176, %f171;
abs.f32 %f177, %f173;
abs.f32 %f178, %f175;
mov.f32 %f179, 0f38D1B717;
max.f32 %f180, %f176, %f179;
max.f32 %f181, %f177, %f179;
max.f32 %f182, %f178, %f179;
fma.rn.f32 %f11, %f720, %f180, %f166;
fma.rn.f32 %f12, %f721, %f181, %f165;
fma.rn.f32 %f13, %f722, %f182, %f164;
abs.f32 %f183, %f9;
abs.f32 %f184, %f7;
setp.gt.f32 %p16, %f184, %f183;
neg.f32 %f185, %f8;
selp.f32 %f186, %f185, 0f00000000, %p16;
neg.f32 %f187, %f9;
selp.f32 %f188, %f7, %f187, %p16;
selp.f32 %f189, 0f00000000, %f8, %p16;
mul.f32 %f190, %f188, %f188;
fma.rn.f32 %f191, %f186, %f186, %f190;
fma.rn.f32 %f192, %f189, %f189, %f191;
sqrt.rn.f32 %f193, %f192;
rcp.rn.f32 %f194, %f193;
mul.f32 %f14, %f186, %f194;
mul.f32 %f15, %f188, %f194;
mul.f32 %f16, %f189, %f194;
ld.global.v2.u32 {%r129, %r130}, [pixelID];
cvt.u64.u32 %rd38, %r129;
cvt.u64.u32 %rd39, %r130;
mov.u64 %rd43, rnd_seeds;
cvta.global.u64 %rd37, %rd43;
// inline asm
call (%rd36), _rt_buffer_get_64, (%rd37, %r104, %r105, %rd38, %rd39, %rd28, %rd28);
// inline asm
ld.global.u32 %r337, [samples];
mov.f32 %f741, 0f00000000;
setp.lt.s32 %p17, %r337, 1;
@%p17 bra BB0_5;
cvt.rn.f32.s32 %f198, %r337;
rcp.rn.f32 %f17, %f198;
ld.u32 %r363, [%rd36];
mul.f32 %f18, %f11, 0f3456BF95;
mul.f32 %f19, %f12, 0f3456BF95;
mul.f32 %f20, %f13, 0f3456BF95;
mul.f32 %f199, %f7, %f15;
mul.f32 %f200, %f8, %f14;
sub.f32 %f21, %f200, %f199;
mul.f32 %f201, %f9, %f14;
mul.f32 %f202, %f7, %f16;
sub.f32 %f22, %f202, %f201;
mul.f32 %f203, %f8, %f16;
mul.f32 %f204, %f9, %f15;
sub.f32 %f23, %f204, %f203;
mov.f32 %f741, 0f00000000;
mov.u32 %r338, 0;
abs.f32 %f205, %f19;
abs.f32 %f206, %f18;
max.f32 %f207, %f206, %f205;
abs.f32 %f208, %f20;
max.f32 %f209, %f207, %f208;
mov.f32 %f742, %f741;
mov.f32 %f743, %f741;
BB0_7:
setp.lt.s32 %p18, %r337, 1;
@%p18 bra BB0_56;
cvt.rn.f32.s32 %f27, %r338;
max.f32 %f28, %f209, %f179;
mov.u32 %r340, 0;
BB0_9:
mad.lo.s32 %r135, %r363, 1664525, 1013904223;
and.b32 %r136, %r135, 16777215;
cvt.rn.f32.u32 %f211, %r136;
fma.rn.f32 %f212, %f211, 0f33800000, %f27;
mul.f32 %f213, %f17, %f212;
mad.lo.s32 %r363, %r135, 1664525, 1013904223;
and.b32 %r137, %r363, 16777215;
cvt.rn.f32.u32 %f214, %r137;
cvt.rn.f32.s32 %f215, %r340;
fma.rn.f32 %f216, %f214, 0f33800000, %f215;
mul.f32 %f217, %f17, %f216;
sqrt.rn.f32 %f32, %f213;
mul.f32 %f735, %f217, 0f40C90FDB;
abs.f32 %f34, %f735;
setp.neu.f32 %p19, %f34, 0f7F800000;
mov.f32 %f729, %f735;
@%p19 bra BB0_11;
mov.f32 %f218, 0f00000000;
mul.rn.f32 %f729, %f735, %f218;
BB0_11:
mul.f32 %f219, %f729, 0f3F22F983;
cvt.rni.s32.f32 %r351, %f219;
cvt.rn.f32.s32 %f220, %r351;
neg.f32 %f221, %f220;
mov.f32 %f222, 0f3FC90FDA;
fma.rn.f32 %f223, %f221, %f222, %f729;
mov.f32 %f224, 0f33A22168;
fma.rn.f32 %f225, %f221, %f224, %f223;
mov.f32 %f226, 0f27C234C5;
fma.rn.f32 %f730, %f221, %f226, %f225;
abs.f32 %f227, %f729;
setp.leu.f32 %p20, %f227, 0f47CE4780;
@%p20 bra BB0_22;
add.u64 %rd45, %SP, 0;
cvta.to.local.u64 %rd3, %rd45;
mov.b32 %r14, %f729;
shr.u32 %r15, %r14, 23;
shl.b32 %r140, %r14, 8;
or.b32 %r16, %r140, -2147483648;
mov.u32 %r342, 0;
mov.u64 %rd194, 0;
mov.u64 %rd193, %rd3;
mov.u32 %r343, %r342;
BB0_13:
.pragma "nounroll";
shl.b64 %rd46, %rd194, 2;
mov.u64 %rd47, __cudart_i2opi_f;
add.s64 %rd48, %rd47, %rd46;
ld.const.u32 %r143, [%rd48];
// inline asm
{
mad.lo.cc.u32 %r141, %r143, %r16, %r343;
madc.hi.u32 %r343, %r143, %r16, 0;
}
// inline asm
st.local.u32 [%rd193], %r141;
add.s32 %r342, %r342, 1;
cvt.s64.s32 %rd194, %r342;
mul.wide.s32 %rd51, %r342, 4;
add.s64 %rd193, %rd3, %rd51;
setp.ne.s32 %p21, %r342, 6;
@%p21 bra BB0_13;
and.b32 %r146, %r15, 255;
add.s32 %r147, %r146, -128;
shr.u32 %r148, %r147, 5;
and.b32 %r21, %r14, -2147483648;
cvta.to.local.u64 %rd53, %rd45;
st.local.u32 [%rd53+24], %r343;
mov.u32 %r149, 6;
sub.s32 %r150, %r149, %r148;
mul.wide.s32 %rd54, %r150, 4;
add.s64 %rd8, %rd53, %rd54;
ld.local.u32 %r344, [%rd8];
ld.local.u32 %r345, [%rd8+-4];
and.b32 %r24, %r15, 31;
setp.eq.s32 %p22, %r24, 0;
@%p22 bra BB0_16;
mov.u32 %r151, 32;
sub.s32 %r152, %r151, %r24;
shr.u32 %r153, %r345, %r152;
shl.b32 %r154, %r344, %r24;
add.s32 %r344, %r153, %r154;
ld.local.u32 %r155, [%rd8+-8];
shr.u32 %r156, %r155, %r152;
shl.b32 %r157, %r345, %r24;
add.s32 %r345, %r156, %r157;
BB0_16:
shr.u32 %r158, %r345, 30;
shl.b32 %r159, %r344, 2;
add.s32 %r346, %r158, %r159;
shl.b32 %r30, %r345, 2;
shr.u32 %r160, %r346, 31;
shr.u32 %r161, %r344, 30;
add.s32 %r31, %r160, %r161;
setp.eq.s32 %p23, %r160, 0;
@%p23 bra BB0_17;
bra.uni BB0_18;
BB0_17:
mov.u32 %r347, %r21;
mov.u32 %r348, %r30;
bra.uni BB0_19;
BB0_18:
not.b32 %r162, %r346;
neg.s32 %r348, %r30;
setp.eq.s32 %p24, %r30, 0;
selp.u32 %r163, 1, 0, %p24;
add.s32 %r346, %r163, %r162;
xor.b32 %r347, %r21, -2147483648;
BB0_19:
clz.b32 %r350, %r346;
setp.eq.s32 %p25, %r350, 0;
shl.b32 %r164, %r346, %r350;
mov.u32 %r165, 32;
sub.s32 %r166, %r165, %r350;
shr.u32 %r167, %r348, %r166;
add.s32 %r168, %r167, %r164;
selp.b32 %r39, %r346, %r168, %p25;
mov.u32 %r169, -921707870;
mul.hi.u32 %r349, %r39, %r169;
setp.eq.s32 %p26, %r21, 0;
neg.s32 %r170, %r31;
selp.b32 %r351, %r31, %r170, %p26;
setp.lt.s32 %p27, %r349, 1;
@%p27 bra BB0_21;
mul.lo.s32 %r171, %r39, -921707870;
shr.u32 %r172, %r171, 31;
shl.b32 %r173, %r349, 1;
add.s32 %r349, %r172, %r173;
add.s32 %r350, %r350, 1;
BB0_21:
mov.u32 %r174, 126;
sub.s32 %r175, %r174, %r350;
shl.b32 %r176, %r175, 23;
add.s32 %r177, %r349, 1;
shr.u32 %r178, %r177, 7;
add.s32 %r179, %r178, 1;
shr.u32 %r180, %r179, 1;
add.s32 %r181, %r180, %r176;
or.b32 %r182, %r181, %r347;
mov.b32 %f730, %r182;
BB0_22:
mul.rn.f32 %f40, %f730, %f730;
add.s32 %r47, %r351, 1;
and.b32 %r48, %r47, 1;
setp.eq.s32 %p28, %r48, 0;
@%p28 bra BB0_24;
bra.uni BB0_23;
BB0_24:
mov.f32 %f230, 0f3C08839E;
mov.f32 %f231, 0fB94CA1F9;
fma.rn.f32 %f731, %f231, %f40, %f230;
bra.uni BB0_25;
BB0_23:
mov.f32 %f228, 0fBAB6061A;
mov.f32 %f229, 0f37CCF5CE;
fma.rn.f32 %f731, %f229, %f40, %f228;
BB0_25:
@%p28 bra BB0_27;
bra.uni BB0_26;
BB0_27:
mov.f32 %f235, 0fBE2AAAA3;
fma.rn.f32 %f236, %f731, %f40, %f235;
mov.f32 %f237, 0f00000000;
fma.rn.f32 %f732, %f236, %f40, %f237;
bra.uni BB0_28;
BB0_26:
mov.f32 %f232, 0f3D2AAAA5;
fma.rn.f32 %f233, %f731, %f40, %f232;
mov.f32 %f234, 0fBF000000;
fma.rn.f32 %f732, %f233, %f40, %f234;
BB0_28:
fma.rn.f32 %f733, %f732, %f730, %f730;
@%p28 bra BB0_30;
mov.f32 %f238, 0f3F800000;
fma.rn.f32 %f733, %f732, %f40, %f238;
BB0_30:
and.b32 %r183, %r47, 2;
setp.eq.s32 %p31, %r183, 0;
@%p31 bra BB0_32;
mov.f32 %f239, 0f00000000;
mov.f32 %f240, 0fBF800000;
fma.rn.f32 %f733, %f733, %f240, %f239;
BB0_32:
@%p19 bra BB0_34;
mov.f32 %f241, 0f00000000;
mul.rn.f32 %f735, %f735, %f241;
BB0_34:
mul.f32 %f242, %f735, 0f3F22F983;
cvt.rni.s32.f32 %r361, %f242;
cvt.rn.f32.s32 %f243, %r361;
neg.f32 %f244, %f243;
fma.rn.f32 %f246, %f244, %f222, %f735;
fma.rn.f32 %f248, %f244, %f224, %f246;
fma.rn.f32 %f736, %f244, %f226, %f248;
abs.f32 %f250, %f735;
setp.leu.f32 %p33, %f250, 0f47CE4780;
@%p33 bra BB0_45;
add.u64 %rd56, %SP, 0;
cvta.to.local.u64 %rd9, %rd56;
mov.b32 %r50, %f735;
shr.u32 %r51, %r50, 23;
shl.b32 %r186, %r50, 8;
or.b32 %r52, %r186, -2147483648;
mov.u32 %r352, 0;
mov.u64 %rd195, %rd9;
mov.u64 %rd196, %rd28;
mov.u32 %r353, %r352;
BB0_36:
.pragma "nounroll";
shl.b64 %rd57, %rd196, 2;
mov.u64 %rd58, __cudart_i2opi_f;
add.s64 %rd59, %rd58, %rd57;
ld.const.u32 %r189, [%rd59];
// inline asm
{
mad.lo.cc.u32 %r187, %r189, %r52, %r353;
madc.hi.u32 %r353, %r189, %r52, 0;
}
// inline asm
st.local.u32 [%rd195], %r187;
add.s32 %r352, %r352, 1;
cvt.s64.s32 %rd196, %r352;
mul.wide.s32 %rd60, %r352, 4;
add.s64 %rd195, %rd9, %rd60;
setp.ne.s32 %p34, %r352, 6;
@%p34 bra BB0_36;
and.b32 %r192, %r51, 255;
add.s32 %r193, %r192, -128;
shr.u32 %r194, %r193, 5;
and.b32 %r57, %r50, -2147483648;
cvta.to.local.u64 %rd62, %rd56;
st.local.u32 [%rd62+24], %r353;
mov.u32 %r195, 6;
sub.s32 %r196, %r195, %r194;
mul.wide.s32 %rd63, %r196, 4;
add.s64 %rd15, %rd62, %rd63;
ld.local.u32 %r354, [%rd15];
ld.local.u32 %r355, [%rd15+-4];
and.b32 %r60, %r51, 31;
setp.eq.s32 %p35, %r60, 0;
@%p35 bra BB0_39;
mov.u32 %r197, 32;
sub.s32 %r198, %r197, %r60;
shr.u32 %r199, %r355, %r198;
shl.b32 %r200, %r354, %r60;
add.s32 %r354, %r199, %r200;
ld.local.u32 %r201, [%rd15+-8];
shr.u32 %r202, %r201, %r198;
shl.b32 %r203, %r355, %r60;
add.s32 %r355, %r202, %r203;
BB0_39:
shr.u32 %r204, %r355, 30;
shl.b32 %r205, %r354, 2;
add.s32 %r356, %r204, %r205;
shl.b32 %r66, %r355, 2;
shr.u32 %r206, %r356, 31;
shr.u32 %r207, %r354, 30;
add.s32 %r67, %r206, %r207;
setp.eq.s32 %p36, %r206, 0;
@%p36 bra BB0_40;
bra.uni BB0_41;
BB0_40:
mov.u32 %r357, %r57;
mov.u32 %r358, %r66;
bra.uni BB0_42;
BB0_41:
not.b32 %r208, %r356;
neg.s32 %r358, %r66;
setp.eq.s32 %p37, %r66, 0;
selp.u32 %r209, 1, 0, %p37;
add.s32 %r356, %r209, %r208;
xor.b32 %r357, %r57, -2147483648;
BB0_42:
clz.b32 %r360, %r356;
setp.eq.s32 %p38, %r360, 0;
shl.b32 %r210, %r356, %r360;
mov.u32 %r211, 32;
sub.s32 %r212, %r211, %r360;
shr.u32 %r213, %r358, %r212;
add.s32 %r214, %r213, %r210;
selp.b32 %r75, %r356, %r214, %p38;
mov.u32 %r215, -921707870;
mul.hi.u32 %r359, %r75, %r215;
setp.eq.s32 %p39, %r57, 0;
neg.s32 %r216, %r67;
selp.b32 %r361, %r67, %r216, %p39;
setp.lt.s32 %p40, %r359, 1;
@%p40 bra BB0_44;
mul.lo.s32 %r217, %r75, -921707870;
shr.u32 %r218, %r217, 31;
shl.b32 %r219, %r359, 1;
add.s32 %r359, %r218, %r219;
add.s32 %r360, %r360, 1;
BB0_44:
mov.u32 %r220, 126;
sub.s32 %r221, %r220, %r360;
shl.b32 %r222, %r221, 23;
add.s32 %r223, %r359, 1;
shr.u32 %r224, %r223, 7;
add.s32 %r225, %r224, 1;
shr.u32 %r226, %r225, 1;
add.s32 %r227, %r226, %r222;
or.b32 %r228, %r227, %r357;
mov.b32 %f736, %r228;
BB0_45:
mul.rn.f32 %f57, %f736, %f736;
and.b32 %r83, %r361, 1;
setp.eq.s32 %p41, %r83, 0;
@%p41 bra BB0_47;
bra.uni BB0_46;
BB0_47:
mov.f32 %f253, 0f3C08839E;
mov.f32 %f254, 0fB94CA1F9;
fma.rn.f32 %f737, %f254, %f57, %f253;
bra.uni BB0_48;
BB0_46:
mov.f32 %f251, 0fBAB6061A;
mov.f32 %f252, 0f37CCF5CE;
fma.rn.f32 %f737, %f252, %f57, %f251;
BB0_48:
@%p41 bra BB0_50;
bra.uni BB0_49;
BB0_50:
mov.f32 %f258, 0fBE2AAAA3;
fma.rn.f32 %f259, %f737, %f57, %f258;
mov.f32 %f260, 0f00000000;
fma.rn.f32 %f738, %f259, %f57, %f260;
bra.uni BB0_51;
BB0_49:
mov.f32 %f255, 0f3D2AAAA5;
fma.rn.f32 %f256, %f737, %f57, %f255;
mov.f32 %f257, 0fBF000000;
fma.rn.f32 %f738, %f256, %f57, %f257;
BB0_51:
fma.rn.f32 %f739, %f738, %f736, %f736;
@%p41 bra BB0_53;
mov.f32 %f261, 0f3F800000;
fma.rn.f32 %f739, %f738, %f57, %f261;
BB0_53:
and.b32 %r229, %r361, 2;
setp.eq.s32 %p44, %r229, 0;
@%p44 bra BB0_55;
mov.f32 %f262, 0f00000000;
mov.f32 %f263, 0fBF800000;
fma.rn.f32 %f739, %f739, %f263, %f262;
BB0_55:
mul.f32 %f272, %f32, %f733;
add.u64 %rd64, %SP, 28;
cvta.to.local.u64 %rd65, %rd64;
mul.f32 %f273, %f272, %f272;
mov.f32 %f274, 0f3F800000;
sub.f32 %f275, %f274, %f273;
mul.f32 %f276, %f32, %f739;
mul.f32 %f277, %f276, %f276;
sub.f32 %f278, %f275, %f277;
mov.f32 %f279, 0f00000000;
max.f32 %f280, %f279, %f278;
sqrt.rn.f32 %f281, %f280;
mul.f32 %f282, %f14, %f276;
mul.f32 %f283, %f15, %f276;
mul.f32 %f284, %f16, %f276;
fma.rn.f32 %f285, %f23, %f272, %f282;
fma.rn.f32 %f286, %f22, %f272, %f283;
fma.rn.f32 %f287, %f21, %f272, %f284;
fma.rn.f32 %f288, %f7, %f281, %f285;
fma.rn.f32 %f289, %f8, %f281, %f286;
fma.rn.f32 %f290, %f9, %f281, %f287;
add.f32 %f291, %f7, %f288;
add.f32 %f292, %f8, %f289;
add.f32 %f293, %f9, %f290;
ld.global.f32 %f294, [shadowSpread];
mul.f32 %f295, %f294, %f291;
mul.f32 %f296, %f294, %f292;
mul.f32 %f297, %f294, %f293;
sub.f32 %f298, %f295, %f7;
sub.f32 %f299, %f296, %f8;
sub.f32 %f300, %f297, %f9;
mul.f32 %f301, %f299, %f299;
fma.rn.f32 %f302, %f298, %f298, %f301;
fma.rn.f32 %f303, %f300, %f300, %f302;
sqrt.rn.f32 %f304, %f303;
rcp.rn.f32 %f305, %f304;
mul.f32 %f267, %f305, %f298;
mul.f32 %f268, %f305, %f299;
mul.f32 %f269, %f305, %f300;
ld.global.u32 %r233, [imageEnabled];
and.b32 %r234, %r233, 32;
setp.eq.s32 %p45, %r234, 0;
selp.f32 %f306, 0f3F800000, 0f41200000, %p45;
mul.f32 %f270, %f306, %f28;
mov.u32 %r235, 1065353216;
st.local.u32 [%rd65+8], %r235;
st.local.u32 [%rd65+4], %r235;
st.local.u32 [%rd65], %r235;
ld.global.u32 %r230, [root];
mov.u32 %r231, 1;
mov.f32 %f271, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r230, %f11, %f12, %f13, %f267, %f268, %f269, %r231, %f270, %f271, %rd64, %r122);
// inline asm
ld.local.f32 %f307, [%rd65];
add.f32 %f743, %f743, %f307;
ld.local.f32 %f308, [%rd65+4];
add.f32 %f742, %f742, %f308;
ld.local.f32 %f309, [%rd65+8];
add.f32 %f741, %f741, %f309;
ld.global.u32 %r337, [samples];
add.s32 %r340, %r340, 1;
setp.lt.s32 %p46, %r340, %r337;
@%p46 bra BB0_9;
BB0_56:
add.s32 %r338, %r338, 1;
setp.lt.s32 %p47, %r338, %r337;
@%p47 bra BB0_7;
bra.uni BB0_57;
BB0_130:
mov.u64 %rd181, image_HDR;
cvta.global.u64 %rd176, %rd181;
mov.u32 %r328, 8;
mov.u64 %rd180, 0;
// inline asm
call (%rd175), _rt_buffer_get_64, (%rd176, %r104, %r328, %rd21, %rd22, %rd180, %rd180);
// inline asm
mov.f32 %f691, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs71, %f691;}
// inline asm
mov.u16 %rs72, 0;
st.v4.u16 [%rd175], {%rs71, %rs71, %rs71, %rs72};
BB0_131:
ld.global.u8 %rs73, [imageEnabled];
and.b16 %rs74, %rs73, 64;
setp.eq.s16 %p128, %rs74, 0;
@%p128 bra BB0_133;
cvt.u64.u32 %rd184, %r2;
cvt.u64.u32 %rd185, %r3;
mov.u64 %rd188, image_Dir;
cvta.global.u64 %rd183, %rd188;
mov.u64 %rd187, 0;
// inline asm
call (%rd182), _rt_buffer_get_64, (%rd183, %r104, %r105, %rd184, %rd185, %rd187, %rd187);
// inline asm
mov.u16 %rs75, 0;
st.v4.u8 [%rd182], {%rs75, %rs75, %rs75, %rs75};
bra.uni BB0_133;
BB0_5:
mov.f32 %f742, %f741;
mov.f32 %f743, %f741;
BB0_57:
setp.eq.s32 %p48, %r337, 0;
mov.f32 %f747, 0f3F800000;
mov.f32 %f748, %f747;
mov.f32 %f749, %f747;
@%p48 bra BB0_59;
mul.lo.s32 %r236, %r337, %r337;
cvt.rn.f32.s32 %f313, %r236;
rcp.rn.f32 %f314, %f313;
mul.f32 %f747, %f743, %f314;
mul.f32 %f748, %f742, %f314;
mul.f32 %f749, %f741, %f314;
BB0_59:
ld.global.f32 %f315, [directColor];
mul.f32 %f316, %f747, %f315;
ld.global.f32 %f317, [directColor+4];
mul.f32 %f318, %f748, %f317;
ld.global.f32 %f319, [directColor+8];
mul.f32 %f320, %f749, %f319;
cvt.sat.f32.f32 %f321, %f10;
mul.f32 %f322, %f316, %f321;
mul.f32 %f323, %f318, %f321;
mul.f32 %f324, %f320, %f321;
ld.global.u32 %r237, [ignoreNormal];
setp.eq.s32 %p49, %r237, 0;
selp.f32 %f84, %f322, %f316, %p49;
selp.f32 %f85, %f323, %f318, %p49;
selp.f32 %f86, %f324, %f320, %p49;
ld.global.u32 %r366, [imageEnabled];
and.b32 %r238, %r366, 8;
setp.eq.s32 %p50, %r238, 0;
@%p50 bra BB0_72;
cvt.u64.u32 %rd68, %r2;
cvt.u64.u32 %rd69, %r3;
mov.u64 %rd72, image_Mask;
cvta.global.u64 %rd67, %rd72;
// inline asm
call (%rd66), _rt_buffer_get_64, (%rd67, %r104, %r104, %rd68, %rd69, %rd28, %rd28);
// inline asm
mov.f32 %f327, 0f3E68BA2E;
cvt.rzi.f32.f32 %f328, %f327;
fma.rn.f32 %f329, %f328, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f87, %f329;
abs.f32 %f88, %f747;
setp.lt.f32 %p51, %f88, 0f00800000;
mul.f32 %f330, %f88, 0f4B800000;
selp.f32 %f331, 0fC3170000, 0fC2FE0000, %p51;
selp.f32 %f332, %f330, %f88, %p51;
mov.b32 %r241, %f332;
and.b32 %r242, %r241, 8388607;
or.b32 %r243, %r242, 1065353216;
mov.b32 %f333, %r243;
shr.u32 %r244, %r241, 23;
cvt.rn.f32.u32 %f334, %r244;
add.f32 %f335, %f331, %f334;
setp.gt.f32 %p52, %f333, 0f3FB504F3;
mul.f32 %f336, %f333, 0f3F000000;
add.f32 %f337, %f335, 0f3F800000;
selp.f32 %f338, %f336, %f333, %p52;
selp.f32 %f339, %f337, %f335, %p52;
add.f32 %f340, %f338, 0fBF800000;
add.f32 %f326, %f338, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f325,%f326;
// inline asm
add.f32 %f341, %f340, %f340;
mul.f32 %f342, %f325, %f341;
mul.f32 %f343, %f342, %f342;
mov.f32 %f344, 0f3C4CAF63;
mov.f32 %f345, 0f3B18F0FE;
fma.rn.f32 %f346, %f345, %f343, %f344;
mov.f32 %f347, 0f3DAAAABD;
fma.rn.f32 %f348, %f346, %f343, %f347;
mul.rn.f32 %f349, %f348, %f343;
mul.rn.f32 %f350, %f349, %f342;
sub.f32 %f351, %f340, %f342;
neg.f32 %f352, %f342;
add.f32 %f353, %f351, %f351;
fma.rn.f32 %f354, %f352, %f340, %f353;
mul.rn.f32 %f355, %f325, %f354;
add.f32 %f356, %f350, %f342;
sub.f32 %f357, %f342, %f356;
add.f32 %f358, %f350, %f357;
add.f32 %f359, %f355, %f358;
add.f32 %f360, %f356, %f359;
sub.f32 %f361, %f356, %f360;
add.f32 %f362, %f359, %f361;
mov.f32 %f363, 0f3F317200;
mul.rn.f32 %f364, %f339, %f363;
mov.f32 %f365, 0f35BFBE8E;
mul.rn.f32 %f366, %f339, %f365;
add.f32 %f367, %f364, %f360;
sub.f32 %f368, %f364, %f367;
add.f32 %f369, %f360, %f368;
add.f32 %f370, %f362, %f369;
add.f32 %f371, %f366, %f370;
add.f32 %f372, %f367, %f371;
sub.f32 %f373, %f367, %f372;
add.f32 %f374, %f371, %f373;
mov.f32 %f375, 0f3EE8BA2E;
mul.rn.f32 %f376, %f375, %f372;
neg.f32 %f377, %f376;
fma.rn.f32 %f378, %f375, %f372, %f377;
fma.rn.f32 %f379, %f375, %f374, %f378;
mov.f32 %f380, 0f00000000;
fma.rn.f32 %f381, %f380, %f372, %f379;
add.rn.f32 %f382, %f376, %f381;
neg.f32 %f383, %f382;
add.rn.f32 %f384, %f376, %f383;
add.rn.f32 %f385, %f384, %f381;
mov.b32 %r245, %f382;
setp.eq.s32 %p53, %r245, 1118925336;
add.s32 %r246, %r245, -1;
mov.b32 %f386, %r246;
add.f32 %f387, %f385, 0f37000000;
selp.f32 %f388, %f386, %f382, %p53;
selp.f32 %f89, %f387, %f385, %p53;
mul.f32 %f389, %f388, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f390, %f389;
mov.f32 %f391, 0fBF317200;
fma.rn.f32 %f392, %f390, %f391, %f388;
mov.f32 %f393, 0fB5BFBE8E;
fma.rn.f32 %f394, %f390, %f393, %f392;
mul.f32 %f395, %f394, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f396, %f395;
add.f32 %f397, %f390, 0f00000000;
ex2.approx.f32 %f398, %f397;
mul.f32 %f399, %f396, %f398;
setp.lt.f32 %p54, %f388, 0fC2D20000;
selp.f32 %f400, 0f00000000, %f399, %p54;
setp.gt.f32 %p55, %f388, 0f42D20000;
selp.f32 %f750, 0f7F800000, %f400, %p55;
setp.eq.f32 %p56, %f750, 0f7F800000;
@%p56 bra BB0_62;
fma.rn.f32 %f750, %f750, %f89, %f750;
BB0_62:
setp.lt.f32 %p57, %f747, 0f00000000;
setp.eq.f32 %p58, %f87, 0f3F800000;
and.pred %p1, %p57, %p58;
mov.b32 %r247, %f750;
xor.b32 %r248, %r247, -2147483648;
mov.b32 %f401, %r248;
selp.f32 %f752, %f401, %f750, %p1;
setp.eq.f32 %p59, %f747, 0f00000000;
@%p59 bra BB0_65;
bra.uni BB0_63;
BB0_65:
add.f32 %f404, %f747, %f747;
selp.f32 %f752, %f404, 0f00000000, %p58;
bra.uni BB0_66;
BB0_120:
mov.u64 %rd140, image_HDR;
cvta.global.u64 %rd135, %rd140;
mov.u32 %r311, 8;
mov.u64 %rd139, 0;
// inline asm
call (%rd134), _rt_buffer_get_64, (%rd135, %r104, %r311, %rd19, %rd20, %rd139, %rd139);
// inline asm
mov.f32 %f681, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs52, %f681;}
// inline asm
st.v4.u16 [%rd134], {%rs52, %rs52, %rs52, %rs41};
BB0_121:
ld.global.u8 %rs53, [imageEnabled];
and.b16 %rs54, %rs53, 64;
setp.eq.s16 %p123, %rs54, 0;
@%p123 bra BB0_133;
cvt.u64.u32 %rd143, %r2;
cvt.u64.u32 %rd144, %r3;
mov.u64 %rd147, image_Dir;
cvta.global.u64 %rd142, %rd147;
mov.u64 %rd146, 0;
// inline asm
call (%rd141), _rt_buffer_get_64, (%rd142, %r104, %r105, %rd143, %rd144, %rd146, %rd146);
// inline asm
mov.u16 %rs55, 255;
mov.u16 %rs56, 0;
st.v4.u8 [%rd141], {%rs56, %rs56, %rs56, %rs55};
bra.uni BB0_133;
BB0_63:
setp.geu.f32 %p60, %f747, 0f00000000;
@%p60 bra BB0_66;
cvt.rzi.f32.f32 %f403, %f375;
setp.neu.f32 %p61, %f403, 0f3EE8BA2E;
selp.f32 %f752, 0f7FFFFFFF, %f752, %p61;
BB0_66:
add.f32 %f405, %f88, 0f3EE8BA2E;
mov.b32 %r249, %f405;
setp.lt.s32 %p63, %r249, 2139095040;
@%p63 bra BB0_71;
setp.gtu.f32 %p64, %f88, 0f7F800000;
@%p64 bra BB0_70;
bra.uni BB0_68;
BB0_70:
add.f32 %f752, %f747, 0f3EE8BA2E;
bra.uni BB0_71;
BB0_68:
setp.neu.f32 %p65, %f88, 0f7F800000;
@%p65 bra BB0_71;
selp.f32 %f752, 0fFF800000, 0f7F800000, %p1;
BB0_71:
mul.f32 %f406, %f752, 0f437F0000;
setp.eq.f32 %p66, %f747, 0f3F800000;
selp.f32 %f407, 0f437F0000, %f406, %p66;
cvt.rzi.u32.f32 %r250, %f407;
cvt.u16.u32 %rs11, %r250;
mov.u16 %rs12, 255;
st.v2.u8 [%rd66], {%rs11, %rs12};
ld.global.u32 %r366, [imageEnabled];
BB0_72:
and.b32 %r251, %r366, 1;
setp.eq.b32 %p67, %r251, 1;
@!%p67 bra BB0_107;
bra.uni BB0_73;
BB0_73:
mov.f32 %f410, 0f3E666666;
cvt.rzi.f32.f32 %f411, %f410;
fma.rn.f32 %f412, %f411, 0fC0000000, 0f3EE66666;
abs.f32 %f100, %f412;
abs.f32 %f101, %f84;
setp.lt.f32 %p68, %f101, 0f00800000;
mul.f32 %f413, %f101, 0f4B800000;
selp.f32 %f414, 0fC3170000, 0fC2FE0000, %p68;
selp.f32 %f415, %f413, %f101, %p68;
mov.b32 %r252, %f415;
and.b32 %r253, %r252, 8388607;
or.b32 %r254, %r253, 1065353216;
mov.b32 %f416, %r254;
shr.u32 %r255, %r252, 23;
cvt.rn.f32.u32 %f417, %r255;
add.f32 %f418, %f414, %f417;
setp.gt.f32 %p69, %f416, 0f3FB504F3;
mul.f32 %f419, %f416, 0f3F000000;
add.f32 %f420, %f418, 0f3F800000;
selp.f32 %f421, %f419, %f416, %p69;
selp.f32 %f422, %f420, %f418, %p69;
add.f32 %f423, %f421, 0fBF800000;
add.f32 %f409, %f421, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f408,%f409;
// inline asm
add.f32 %f424, %f423, %f423;
mul.f32 %f425, %f408, %f424;
mul.f32 %f426, %f425, %f425;
mov.f32 %f427, 0f3C4CAF63;
mov.f32 %f428, 0f3B18F0FE;
fma.rn.f32 %f429, %f428, %f426, %f427;
mov.f32 %f430, 0f3DAAAABD;
fma.rn.f32 %f431, %f429, %f426, %f430;
mul.rn.f32 %f432, %f431, %f426;
mul.rn.f32 %f433, %f432, %f425;
sub.f32 %f434, %f423, %f425;
neg.f32 %f435, %f425;
add.f32 %f436, %f434, %f434;
fma.rn.f32 %f437, %f435, %f423, %f436;
mul.rn.f32 %f438, %f408, %f437;
add.f32 %f439, %f433, %f425;
sub.f32 %f440, %f425, %f439;
add.f32 %f441, %f433, %f440;
add.f32 %f442, %f438, %f441;
add.f32 %f443, %f439, %f442;
sub.f32 %f444, %f439, %f443;
add.f32 %f445, %f442, %f444;
mov.f32 %f446, 0f3F317200;
mul.rn.f32 %f447, %f422, %f446;
mov.f32 %f448, 0f35BFBE8E;
mul.rn.f32 %f449, %f422, %f448;
add.f32 %f450, %f447, %f443;
sub.f32 %f451, %f447, %f450;
add.f32 %f452, %f443, %f451;
add.f32 %f453, %f445, %f452;
add.f32 %f454, %f449, %f453;
add.f32 %f455, %f450, %f454;
sub.f32 %f456, %f450, %f455;
add.f32 %f457, %f454, %f456;
mov.f32 %f458, 0f3EE66666;
mul.rn.f32 %f459, %f458, %f455;
neg.f32 %f460, %f459;
fma.rn.f32 %f461, %f458, %f455, %f460;
fma.rn.f32 %f462, %f458, %f457, %f461;
mov.f32 %f463, 0f00000000;
fma.rn.f32 %f464, %f463, %f455, %f462;
add.rn.f32 %f465, %f459, %f464;
neg.f32 %f466, %f465;
add.rn.f32 %f467, %f459, %f466;
add.rn.f32 %f468, %f467, %f464;
mov.b32 %r256, %f465;
setp.eq.s32 %p70, %r256, 1118925336;
add.s32 %r257, %r256, -1;
mov.b32 %f469, %r257;
add.f32 %f470, %f468, 0f37000000;
selp.f32 %f471, %f469, %f465, %p70;
selp.f32 %f102, %f470, %f468, %p70;
mul.f32 %f472, %f471, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f473, %f472;
mov.f32 %f474, 0fBF317200;
fma.rn.f32 %f475, %f473, %f474, %f471;
mov.f32 %f476, 0fB5BFBE8E;
fma.rn.f32 %f477, %f473, %f476, %f475;
mul.f32 %f478, %f477, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f479, %f478;
add.f32 %f480, %f473, 0f00000000;
ex2.approx.f32 %f481, %f480;
mul.f32 %f482, %f479, %f481;
setp.lt.f32 %p71, %f471, 0fC2D20000;
selp.f32 %f483, 0f00000000, %f482, %p71;
setp.gt.f32 %p72, %f471, 0f42D20000;
selp.f32 %f753, 0f7F800000, %f483, %p72;
setp.eq.f32 %p73, %f753, 0f7F800000;
@%p73 bra BB0_75;
fma.rn.f32 %f753, %f753, %f102, %f753;
BB0_75:
setp.lt.f32 %p74, %f84, 0f00000000;
setp.eq.f32 %p75, %f100, 0f3F800000;
and.pred %p2, %p74, %p75;
mov.b32 %r258, %f753;
xor.b32 %r259, %r258, -2147483648;
mov.b32 %f484, %r259;
selp.f32 %f755, %f484, %f753, %p2;
setp.eq.f32 %p76, %f84, 0f00000000;
@%p76 bra BB0_78;
bra.uni BB0_76;
BB0_78:
add.f32 %f487, %f84, %f84;
selp.f32 %f755, %f487, 0f00000000, %p75;
bra.uni BB0_79;
BB0_76:
setp.geu.f32 %p77, %f84, 0f00000000;
@%p77 bra BB0_79;
cvt.rzi.f32.f32 %f486, %f458;
setp.neu.f32 %p78, %f486, 0f3EE66666;
selp.f32 %f755, 0f7FFFFFFF, %f755, %p78;
BB0_79:
abs.f32 %f692, %f84;
add.f32 %f488, %f692, 0f3EE66666;
mov.b32 %r260, %f488;
setp.lt.s32 %p80, %r260, 2139095040;
@%p80 bra BB0_84;
abs.f32 %f709, %f84;
setp.gtu.f32 %p81, %f709, 0f7F800000;
@%p81 bra BB0_83;
bra.uni BB0_81;
BB0_83:
add.f32 %f755, %f84, 0f3EE66666;
bra.uni BB0_84;
BB0_81:
abs.f32 %f710, %f84;
setp.neu.f32 %p82, %f710, 0f7F800000;
@%p82 bra BB0_84;
selp.f32 %f755, 0fFF800000, 0f7F800000, %p2;
BB0_84:
mov.f32 %f700, 0fB5BFBE8E;
mov.f32 %f699, 0fBF317200;
mov.f32 %f698, 0f00000000;
mov.f32 %f697, 0f35BFBE8E;
mov.f32 %f696, 0f3F317200;
mov.f32 %f695, 0f3DAAAABD;
mov.f32 %f694, 0f3C4CAF63;
mov.f32 %f693, 0f3B18F0FE;
setp.eq.f32 %p83, %f84, 0f3F800000;
selp.f32 %f113, 0f3F800000, %f755, %p83;
abs.f32 %f114, %f85;
setp.lt.f32 %p84, %f114, 0f00800000;
mul.f32 %f491, %f114, 0f4B800000;
selp.f32 %f492, 0fC3170000, 0fC2FE0000, %p84;
selp.f32 %f493, %f491, %f114, %p84;
mov.b32 %r261, %f493;
and.b32 %r262, %r261, 8388607;
or.b32 %r263, %r262, 1065353216;
mov.b32 %f494, %r263;
shr.u32 %r264, %r261, 23;
cvt.rn.f32.u32 %f495, %r264;
add.f32 %f496, %f492, %f495;
setp.gt.f32 %p85, %f494, 0f3FB504F3;
mul.f32 %f497, %f494, 0f3F000000;
add.f32 %f498, %f496, 0f3F800000;
selp.f32 %f499, %f497, %f494, %p85;
selp.f32 %f500, %f498, %f496, %p85;
add.f32 %f501, %f499, 0fBF800000;
add.f32 %f490, %f499, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f489,%f490;
// inline asm
add.f32 %f502, %f501, %f501;
mul.f32 %f503, %f489, %f502;
mul.f32 %f504, %f503, %f503;
fma.rn.f32 %f507, %f693, %f504, %f694;
fma.rn.f32 %f509, %f507, %f504, %f695;
mul.rn.f32 %f510, %f509, %f504;
mul.rn.f32 %f511, %f510, %f503;
sub.f32 %f512, %f501, %f503;
neg.f32 %f513, %f503;
add.f32 %f514, %f512, %f512;
fma.rn.f32 %f515, %f513, %f501, %f514;
mul.rn.f32 %f516, %f489, %f515;
add.f32 %f517, %f511, %f503;
sub.f32 %f518, %f503, %f517;
add.f32 %f519, %f511, %f518;
add.f32 %f520, %f516, %f519;
add.f32 %f521, %f517, %f520;
sub.f32 %f522, %f517, %f521;
add.f32 %f523, %f520, %f522;
mul.rn.f32 %f525, %f500, %f696;
mul.rn.f32 %f527, %f500, %f697;
add.f32 %f528, %f525, %f521;
sub.f32 %f529, %f525, %f528;
add.f32 %f530, %f521, %f529;
add.f32 %f531, %f523, %f530;
add.f32 %f532, %f527, %f531;
add.f32 %f533, %f528, %f532;
sub.f32 %f534, %f528, %f533;
add.f32 %f535, %f532, %f534;
mul.rn.f32 %f537, %f458, %f533;
neg.f32 %f538, %f537;
fma.rn.f32 %f539, %f458, %f533, %f538;
fma.rn.f32 %f540, %f458, %f535, %f539;
fma.rn.f32 %f542, %f698, %f533, %f540;
add.rn.f32 %f543, %f537, %f542;
neg.f32 %f544, %f543;
add.rn.f32 %f545, %f537, %f544;
add.rn.f32 %f546, %f545, %f542;
mov.b32 %r265, %f543;
setp.eq.s32 %p86, %r265, 1118925336;
add.s32 %r266, %r265, -1;
mov.b32 %f547, %r266;
add.f32 %f548, %f546, 0f37000000;
selp.f32 %f549, %f547, %f543, %p86;
selp.f32 %f115, %f548, %f546, %p86;
mul.f32 %f550, %f549, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f551, %f550;
fma.rn.f32 %f553, %f551, %f699, %f549;
fma.rn.f32 %f555, %f551, %f700, %f553;
mul.f32 %f556, %f555, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f557, %f556;
add.f32 %f558, %f551, 0f00000000;
ex2.approx.f32 %f559, %f558;
mul.f32 %f560, %f557, %f559;
setp.lt.f32 %p87, %f549, 0fC2D20000;
selp.f32 %f561, 0f00000000, %f560, %p87;
setp.gt.f32 %p88, %f549, 0f42D20000;
selp.f32 %f756, 0f7F800000, %f561, %p88;
setp.eq.f32 %p89, %f756, 0f7F800000;
@%p89 bra BB0_86;
fma.rn.f32 %f756, %f756, %f115, %f756;
BB0_86:
setp.lt.f32 %p90, %f85, 0f00000000;
and.pred %p3, %p90, %p75;
mov.b32 %r267, %f756;
xor.b32 %r268, %r267, -2147483648;
mov.b32 %f562, %r268;
selp.f32 %f758, %f562, %f756, %p3;
setp.eq.f32 %p92, %f85, 0f00000000;
@%p92 bra BB0_89;
bra.uni BB0_87;
BB0_89:
add.f32 %f565, %f85, %f85;
selp.f32 %f758, %f565, 0f00000000, %p75;
bra.uni BB0_90;
BB0_87:
setp.geu.f32 %p93, %f85, 0f00000000;
@%p93 bra BB0_90;
mov.f32 %f716, 0f3EE66666;
cvt.rzi.f32.f32 %f564, %f716;
setp.neu.f32 %p94, %f564, 0f3EE66666;
selp.f32 %f758, 0f7FFFFFFF, %f758, %p94;
BB0_90:
abs.f32 %f711, %f85;
add.f32 %f566, %f711, 0f3EE66666;
mov.b32 %r269, %f566;
setp.lt.s32 %p96, %r269, 2139095040;
@%p96 bra BB0_95;
abs.f32 %f714, %f85;
setp.gtu.f32 %p97, %f714, 0f7F800000;
@%p97 bra BB0_94;
bra.uni BB0_92;
BB0_94:
add.f32 %f758, %f85, 0f3EE66666;
bra.uni BB0_95;
BB0_92:
abs.f32 %f715, %f85;
setp.neu.f32 %p98, %f715, 0f7F800000;
@%p98 bra BB0_95;
selp.f32 %f758, 0fFF800000, 0f7F800000, %p3;
BB0_95:
mov.f32 %f712, 0f3EE66666;
mov.f32 %f708, 0fB5BFBE8E;
mov.f32 %f707, 0fBF317200;
mov.f32 %f706, 0f00000000;
mov.f32 %f705, 0f35BFBE8E;
mov.f32 %f704, 0f3F317200;
mov.f32 %f703, 0f3DAAAABD;
mov.f32 %f702, 0f3C4CAF63;
mov.f32 %f701, 0f3B18F0FE;
setp.eq.f32 %p99, %f85, 0f3F800000;
selp.f32 %f126, 0f3F800000, %f758, %p99;
abs.f32 %f127, %f86;
setp.lt.f32 %p100, %f127, 0f00800000;
mul.f32 %f569, %f127, 0f4B800000;
selp.f32 %f570, 0fC3170000, 0fC2FE0000, %p100;
selp.f32 %f571, %f569, %f127, %p100;
mov.b32 %r270, %f571;
and.b32 %r271, %r270, 8388607;
or.b32 %r272, %r271, 1065353216;
mov.b32 %f572, %r272;
shr.u32 %r273, %r270, 23;
cvt.rn.f32.u32 %f573, %r273;
add.f32 %f574, %f570, %f573;
setp.gt.f32 %p101, %f572, 0f3FB504F3;
mul.f32 %f575, %f572, 0f3F000000;
add.f32 %f576, %f574, 0f3F800000;
selp.f32 %f577, %f575, %f572, %p101;
selp.f32 %f578, %f576, %f574, %p101;
add.f32 %f579, %f577, 0fBF800000;
add.f32 %f568, %f577, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f567,%f568;
// inline asm
add.f32 %f580, %f579, %f579;
mul.f32 %f581, %f567, %f580;
mul.f32 %f582, %f581, %f581;
fma.rn.f32 %f585, %f701, %f582, %f702;
fma.rn.f32 %f587, %f585, %f582, %f703;
mul.rn.f32 %f588, %f587, %f582;
mul.rn.f32 %f589, %f588, %f581;
sub.f32 %f590, %f579, %f581;
neg.f32 %f591, %f581;
add.f32 %f592, %f590, %f590;
fma.rn.f32 %f593, %f591, %f579, %f592;
mul.rn.f32 %f594, %f567, %f593;
add.f32 %f595, %f589, %f581;
sub.f32 %f596, %f581, %f595;
add.f32 %f597, %f589, %f596;
add.f32 %f598, %f594, %f597;
add.f32 %f599, %f595, %f598;
sub.f32 %f600, %f595, %f599;
add.f32 %f601, %f598, %f600;
mul.rn.f32 %f603, %f578, %f704;
mul.rn.f32 %f605, %f578, %f705;
add.f32 %f606, %f603, %f599;
sub.f32 %f607, %f603, %f606;
add.f32 %f608, %f599, %f607;
add.f32 %f609, %f601, %f608;
add.f32 %f610, %f605, %f609;
add.f32 %f611, %f606, %f610;
sub.f32 %f612, %f606, %f611;
add.f32 %f613, %f610, %f612;
mul.rn.f32 %f615, %f712, %f611;
neg.f32 %f616, %f615;
fma.rn.f32 %f617, %f712, %f611, %f616;
fma.rn.f32 %f618, %f712, %f613, %f617;
fma.rn.f32 %f620, %f706, %f611, %f618;
add.rn.f32 %f621, %f615, %f620;
neg.f32 %f622, %f621;
add.rn.f32 %f623, %f615, %f622;
add.rn.f32 %f624, %f623, %f620;
mov.b32 %r274, %f621;
setp.eq.s32 %p102, %r274, 1118925336;
add.s32 %r275, %r274, -1;
mov.b32 %f625, %r275;
add.f32 %f626, %f624, 0f37000000;
selp.f32 %f627, %f625, %f621, %p102;
selp.f32 %f128, %f626, %f624, %p102;
mul.f32 %f628, %f627, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f629, %f628;
fma.rn.f32 %f631, %f629, %f707, %f627;
fma.rn.f32 %f633, %f629, %f708, %f631;
mul.f32 %f634, %f633, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f635, %f634;
add.f32 %f636, %f629, 0f00000000;
ex2.approx.f32 %f637, %f636;
mul.f32 %f638, %f635, %f637;
setp.lt.f32 %p103, %f627, 0fC2D20000;
selp.f32 %f639, 0f00000000, %f638, %p103;
setp.gt.f32 %p104, %f627, 0f42D20000;
selp.f32 %f759, 0f7F800000, %f639, %p104;
setp.eq.f32 %p105, %f759, 0f7F800000;
@%p105 bra BB0_97;
fma.rn.f32 %f759, %f759, %f128, %f759;
BB0_97:
setp.lt.f32 %p106, %f86, 0f00000000;
and.pred %p4, %p106, %p75;
mov.b32 %r276, %f759;
xor.b32 %r277, %r276, -2147483648;
mov.b32 %f640, %r277;
selp.f32 %f761, %f640, %f759, %p4;
setp.eq.f32 %p108, %f86, 0f00000000;
@%p108 bra BB0_100;
bra.uni BB0_98;
BB0_100:
add.f32 %f643, %f86, %f86;
selp.f32 %f761, %f643, 0f00000000, %p75;
bra.uni BB0_101;
BB0_98:
setp.geu.f32 %p109, %f86, 0f00000000;
@%p109 bra BB0_101;
mov.f32 %f713, 0f3EE66666;
cvt.rzi.f32.f32 %f642, %f713;
setp.neu.f32 %p110, %f642, 0f3EE66666;
selp.f32 %f761, 0f7FFFFFFF, %f761, %p110;
BB0_101:
abs.f32 %f717, %f86;
add.f32 %f644, %f717, 0f3EE66666;
mov.b32 %r278, %f644;
setp.lt.s32 %p112, %r278, 2139095040;
@%p112 bra BB0_106;
abs.f32 %f718, %f86;
setp.gtu.f32 %p113, %f718, 0f7F800000;
@%p113 bra BB0_105;
bra.uni BB0_103;
BB0_105:
add.f32 %f761, %f86, 0f3EE66666;
bra.uni BB0_106;
BB0_103:
abs.f32 %f719, %f86;
setp.neu.f32 %p114, %f719, 0f7F800000;
@%p114 bra BB0_106;
selp.f32 %f761, 0fFF800000, 0f7F800000, %p4;
BB0_106:
mov.u32 %r332, 4;
mov.u64 %rd189, 0;
mov.u32 %r331, 2;
setp.eq.f32 %p115, %f86, 0f3F800000;
selp.f32 %f645, 0f3F800000, %f761, %p115;
cvt.u64.u32 %rd76, %r3;
cvt.u64.u32 %rd75, %r2;
mov.u64 %rd79, image;
cvta.global.u64 %rd74, %rd79;
// inline asm
call (%rd73), _rt_buffer_get_64, (%rd74, %r331, %r332, %rd75, %rd76, %rd189, %rd189);
// inline asm
cvt.sat.f32.f32 %f646, %f645;
mul.f32 %f647, %f646, 0f437FFD71;
cvt.rzi.u32.f32 %r281, %f647;
cvt.sat.f32.f32 %f648, %f126;
mul.f32 %f649, %f648, 0f437FFD71;
cvt.rzi.u32.f32 %r282, %f649;
cvt.sat.f32.f32 %f650, %f113;
mul.f32 %f651, %f650, 0f437FFD71;
cvt.rzi.u32.f32 %r283, %f651;
cvt.u16.u32 %rs13, %r281;
cvt.u16.u32 %rs14, %r283;
cvt.u16.u32 %rs15, %r282;
mov.u16 %rs16, 255;
st.v4.u8 [%rd73], {%rs13, %rs15, %rs14, %rs16};
ld.global.u32 %r366, [imageEnabled];
BB0_107:
and.b32 %r284, %r366, 4;
setp.eq.s32 %p116, %r284, 0;
@%p116 bra BB0_111;
ld.global.u32 %r285, [additive];
setp.eq.s32 %p117, %r285, 0;
cvt.u64.u32 %rd17, %r2;
cvt.u64.u32 %rd18, %r3;
mov.f32 %f652, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs17, %f652;}
// inline asm
@%p117 bra BB0_110;
mov.u64 %rd190, 0;
mov.u32 %r333, 2;
mov.u64 %rd92, image_HDR;
cvta.global.u64 %rd81, %rd92;
mov.u32 %r289, 8;
// inline asm
call (%rd80), _rt_buffer_get_64, (%rd81, %r333, %r289, %rd17, %rd18, %rd190, %rd190);
// inline asm
ld.v4.u16 {%rs24, %rs25, %rs26, %rs27}, [%rd80];
// inline asm
{ cvt.f32.f16 %f653, %rs24;}
// inline asm
// inline asm
{ cvt.f32.f16 %f654, %rs25;}
// inline asm
// inline asm
{ cvt.f32.f16 %f655, %rs26;}
// inline asm
// inline asm
call (%rd86), _rt_buffer_get_64, (%rd81, %r333, %r289, %rd17, %rd18, %rd190, %rd190);
// inline asm
add.f32 %f656, %f84, %f653;
add.f32 %f657, %f85, %f654;
add.f32 %f658, %f86, %f655;
// inline asm
{ cvt.rn.f16.f32 %rs23, %f658;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs22, %f657;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs21, %f656;}
// inline asm
st.v4.u16 [%rd86], {%rs21, %rs22, %rs23, %rs17};
bra.uni BB0_111;
BB0_110:
mov.u64 %rd191, 0;
mov.u32 %r334, 2;
mov.u64 %rd99, image_HDR;
cvta.global.u64 %rd94, %rd99;
mov.u32 %r291, 8;
// inline asm
call (%rd93), _rt_buffer_get_64, (%rd94, %r334, %r291, %rd17, %rd18, %rd191, %rd191);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs30, %f86;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs29, %f85;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs28, %f84;}
// inline asm
st.v4.u16 [%rd93], {%rs28, %rs29, %rs30, %rs17};
BB0_111:
ld.global.u8 %rs31, [imageEnabled];
and.b16 %rs32, %rs31, 64;
setp.eq.s16 %p118, %rs32, 0;
@%p118 bra BB0_133;
mov.u32 %r336, 4;
mov.u64 %rd192, 0;
mov.u32 %r335, 2;
ld.global.f32 %f662, [directDir];
ld.global.f32 %f663, [directDir+4];
ld.global.f32 %f664, [directDir+8];
cvt.u64.u32 %rd103, %r3;
cvt.u64.u32 %rd102, %r2;
mov.u64 %rd106, image_Dir;
cvta.global.u64 %rd101, %rd106;
// inline asm
call (%rd100), _rt_buffer_get_64, (%rd101, %r335, %r336, %rd102, %rd103, %rd192, %rd192);
// inline asm
fma.rn.f32 %f665, %f662, 0fBF000000, 0f3F000000;
mul.f32 %f666, %f665, 0f437F0000;
cvt.rzi.u32.f32 %r294, %f666;
fma.rn.f32 %f667, %f663, 0fBF000000, 0f3F000000;
mul.f32 %f668, %f667, 0f437F0000;
cvt.rzi.u32.f32 %r295, %f668;
fma.rn.f32 %f669, %f664, 0fBF000000, 0f3F000000;
mul.f32 %f670, %f669, 0f437F0000;
cvt.rzi.u32.f32 %r296, %f670;
cvt.u16.u32 %rs33, %r296;
cvt.u16.u32 %rs34, %r295;
cvt.u16.u32 %rs35, %r294;
mov.u16 %rs36, 255;
st.v4.u8 [%rd100], {%rs35, %rs34, %rs33, %rs36};
BB0_133:
ret;
}