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

1855 lines
53 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 .u32 lightCookie;
.global .align 16 .b8 lightTilingOffset[16];
.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 4 .b8 _ZN21rti_internal_typeinfo11lightCookieE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo17lightTilingOffsetE[8] = {82, 97, 121, 0, 16, 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 .b8 _ZN21rti_internal_typename11lightCookieE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename17lightTilingOffsetE[7] = {102, 108, 111, 97, 116, 52, 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 4 .u32 _ZN21rti_internal_typeenum11lightCookieE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum17lightTilingOffsetE = 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 _ZN21rti_internal_semantic11lightCookieE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic17lightTilingOffsetE[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];
.global .align 1 .b8 _ZN23rti_internal_annotation11lightCookieE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation17lightTilingOffsetE[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<789>;
.reg .b32 %r<373>;
.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 %f747, 0f00000000;
mov.f32 %f748, %f747;
mov.f32 %f749, %f747;
@%p5 bra BB0_2;
ld.u8 %rs7, [%rd23+1];
and.b16 %rs9, %rs5, 255;
cvt.rn.f32.u16 %f146, %rs9;
div.rn.f32 %f147, %f146, 0f437F0000;
fma.rn.f32 %f148, %f147, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f149, %rs7;
div.rn.f32 %f150, %f149, 0f437F0000;
fma.rn.f32 %f151, %f150, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f152, %rs4;
div.rn.f32 %f153, %f152, 0f437F0000;
fma.rn.f32 %f154, %f153, 0f40000000, 0fBF800000;
mul.f32 %f155, %f151, %f151;
fma.rn.f32 %f156, %f148, %f148, %f155;
fma.rn.f32 %f157, %f154, %f154, %f156;
sqrt.rn.f32 %f158, %f157;
rcp.rn.f32 %f159, %f158;
mul.f32 %f747, %f148, %f159;
mul.f32 %f748, %f151, %f159;
mul.f32 %f749, %f154, %f159;
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, %f748, 0f00000000;
setp.eq.f32 %p7, %f747, 0f00000000;
and.pred %p8, %p7, %p6;
setp.eq.f32 %p9, %f749, 0f00000000;
and.pred %p10, %p8, %p9;
@%p10 bra BB0_123;
bra.uni BB0_3;
BB0_123:
ld.global.u32 %r372, [imageEnabled];
and.b32 %r316, %r372, 1;
setp.eq.b32 %p124, %r316, 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 %r372, [imageEnabled];
BB0_125:
and.b32 %r319, %r372, 8;
setp.eq.s32 %p125, %r319, 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 %f712, 0f00000000;
cvt.rzi.u32.f32 %r322, %f712;
cvt.u16.u32 %rs58, %r322;
mov.u16 %rs59, 0;
st.v2.u8 [%rd155], {%rs58, %rs59};
ld.global.u32 %r372, [imageEnabled];
BB0_127:
and.b32 %r323, %r372, 4;
setp.eq.s32 %p126, %r323, 0;
@%p126 bra BB0_131;
ld.global.u32 %r324, [additive];
setp.eq.s32 %p127, %r324, 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 %r328, 8;
mov.u64 %rd173, 0;
// inline asm
call (%rd162), _rt_buffer_get_64, (%rd163, %r104, %r328, %rd21, %rd22, %rd173, %rd173);
// inline asm
ld.v4.u16 {%rs66, %rs67, %rs68, %rs69}, [%rd162];
// inline asm
{ cvt.f32.f16 %f713, %rs66;}
// inline asm
// inline asm
{ cvt.f32.f16 %f714, %rs67;}
// inline asm
// inline asm
{ cvt.f32.f16 %f715, %rs68;}
// inline asm
// inline asm
call (%rd168), _rt_buffer_get_64, (%rd163, %r104, %r328, %rd21, %rd22, %rd173, %rd173);
// inline asm
add.f32 %f716, %f713, 0f00000000;
add.f32 %f717, %f714, 0f00000000;
add.f32 %f718, %f715, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs65, %f718;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs64, %f717;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs63, %f716;}
// 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 %f160, %f747, %f7;
mul.f32 %f161, %f748, %f8;
neg.f32 %f162, %f161;
sub.f32 %f163, %f162, %f160;
mul.f32 %f164, %f749, %f9;
sub.f32 %f10, %f163, %f164;
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 %r370, [imageEnabled];
and.b32 %r120, %r370, 32;
setp.eq.s32 %p14, %r120, 0;
and.pred %p15, %p13, %p14;
@%p15 bra BB0_113;
bra.uni BB0_4;
BB0_113:
and.b32 %r299, %r370, 1;
setp.eq.b32 %p119, %r299, 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 %r370, [imageEnabled];
BB0_115:
and.b32 %r302, %r370, 8;
setp.eq.s32 %p120, %r302, 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 %f701, 0f00000000;
cvt.rzi.u32.f32 %r305, %f701;
cvt.u16.u32 %rs39, %r305;
mov.u16 %rs40, 255;
st.v2.u8 [%rd114], {%rs39, %rs40};
ld.global.u32 %r370, [imageEnabled];
BB0_117:
and.b32 %r306, %r370, 4;
setp.eq.s32 %p121, %r306, 0;
@%p121 bra BB0_121;
ld.global.u32 %r307, [additive];
setp.eq.s32 %p122, %r307, 0;
cvt.u64.u32 %rd19, %r2;
cvt.u64.u32 %rd20, %r3;
mov.f32 %f702, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs41, %f702;}
// inline asm
@%p122 bra BB0_120;
mov.u64 %rd133, image_HDR;
cvta.global.u64 %rd122, %rd133;
mov.u32 %r311, 8;
mov.u64 %rd132, 0;
// inline asm
call (%rd121), _rt_buffer_get_64, (%rd122, %r104, %r311, %rd19, %rd20, %rd132, %rd132);
// inline asm
ld.v4.u16 {%rs48, %rs49, %rs50, %rs51}, [%rd121];
// inline asm
{ cvt.f32.f16 %f703, %rs48;}
// inline asm
// inline asm
{ cvt.f32.f16 %f704, %rs49;}
// inline asm
// inline asm
{ cvt.f32.f16 %f705, %rs50;}
// inline asm
// inline asm
call (%rd127), _rt_buffer_get_64, (%rd122, %r104, %r311, %rd19, %rd20, %rd132, %rd132);
// inline asm
add.f32 %f706, %f703, 0f00000000;
add.f32 %f707, %f704, 0f00000000;
add.f32 %f708, %f705, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs47, %f708;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs46, %f707;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs45, %f706;}
// 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 %f13, [%rd30+8];
ld.f32 %f12, [%rd30+4];
ld.f32 %f11, [%rd30];
mul.f32 %f168, %f11, 0f3456BF95;
mul.f32 %f169, %f12, 0f3456BF95;
mul.f32 %f170, %f13, 0f3456BF95;
abs.f32 %f171, %f747;
div.rn.f32 %f172, %f168, %f171;
abs.f32 %f173, %f748;
div.rn.f32 %f174, %f169, %f173;
abs.f32 %f175, %f749;
div.rn.f32 %f176, %f170, %f175;
abs.f32 %f177, %f172;
abs.f32 %f178, %f174;
abs.f32 %f179, %f176;
mov.f32 %f180, 0f38D1B717;
max.f32 %f181, %f177, %f180;
max.f32 %f182, %f178, %f180;
max.f32 %f183, %f179, %f180;
fma.rn.f32 %f14, %f747, %f181, %f11;
fma.rn.f32 %f15, %f748, %f182, %f12;
fma.rn.f32 %f16, %f749, %f183, %f13;
abs.f32 %f184, %f9;
abs.f32 %f185, %f7;
setp.gt.f32 %p16, %f185, %f184;
neg.f32 %f186, %f8;
selp.f32 %f187, %f186, 0f00000000, %p16;
neg.f32 %f188, %f9;
selp.f32 %f189, %f7, %f188, %p16;
selp.f32 %f190, 0f00000000, %f8, %p16;
mul.f32 %f191, %f189, %f189;
fma.rn.f32 %f192, %f187, %f187, %f191;
fma.rn.f32 %f193, %f190, %f190, %f192;
sqrt.rn.f32 %f194, %f193;
rcp.rn.f32 %f195, %f194;
mul.f32 %f17, %f187, %f195;
mul.f32 %f18, %f189, %f195;
mul.f32 %f19, %f190, %f195;
mul.f32 %f196, %f9, %f18;
mul.f32 %f197, %f8, %f19;
sub.f32 %f20, %f196, %f197;
mul.f32 %f198, %f7, %f19;
mul.f32 %f199, %f9, %f17;
sub.f32 %f21, %f198, %f199;
mul.f32 %f200, %f8, %f17;
mul.f32 %f201, %f7, %f18;
sub.f32 %f22, %f200, %f201;
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 %r339, [samples];
mov.f32 %f768, 0f00000000;
setp.lt.s32 %p17, %r339, 1;
@%p17 bra BB0_5;
cvt.rn.f32.s32 %f205, %r339;
rcp.rn.f32 %f23, %f205;
ld.u32 %r365, [%rd36];
mul.f32 %f24, %f14, 0f3456BF95;
mul.f32 %f25, %f15, 0f3456BF95;
mul.f32 %f26, %f16, 0f3456BF95;
mov.f32 %f768, 0f00000000;
mov.u32 %r340, 0;
abs.f32 %f206, %f25;
abs.f32 %f207, %f24;
max.f32 %f208, %f207, %f206;
abs.f32 %f209, %f26;
max.f32 %f210, %f208, %f209;
mov.f32 %f769, %f768;
mov.f32 %f770, %f768;
BB0_7:
setp.lt.s32 %p18, %r339, 1;
@%p18 bra BB0_56;
cvt.rn.f32.s32 %f30, %r340;
max.f32 %f31, %f210, %f180;
mov.u32 %r342, 0;
BB0_9:
mad.lo.s32 %r135, %r365, 1664525, 1013904223;
and.b32 %r136, %r135, 16777215;
cvt.rn.f32.u32 %f212, %r136;
fma.rn.f32 %f213, %f212, 0f33800000, %f30;
mul.f32 %f214, %f23, %f213;
mad.lo.s32 %r365, %r135, 1664525, 1013904223;
and.b32 %r137, %r365, 16777215;
cvt.rn.f32.u32 %f215, %r137;
cvt.rn.f32.s32 %f216, %r342;
fma.rn.f32 %f217, %f215, 0f33800000, %f216;
mul.f32 %f218, %f23, %f217;
sqrt.rn.f32 %f35, %f214;
mul.f32 %f762, %f218, 0f40C90FDB;
abs.f32 %f37, %f762;
setp.neu.f32 %p19, %f37, 0f7F800000;
mov.f32 %f756, %f762;
@%p19 bra BB0_11;
mov.f32 %f219, 0f00000000;
mul.rn.f32 %f756, %f762, %f219;
BB0_11:
mul.f32 %f220, %f756, 0f3F22F983;
cvt.rni.s32.f32 %r353, %f220;
cvt.rn.f32.s32 %f221, %r353;
neg.f32 %f222, %f221;
mov.f32 %f223, 0f3FC90FDA;
fma.rn.f32 %f224, %f222, %f223, %f756;
mov.f32 %f225, 0f33A22168;
fma.rn.f32 %f226, %f222, %f225, %f224;
mov.f32 %f227, 0f27C234C5;
fma.rn.f32 %f757, %f222, %f227, %f226;
abs.f32 %f228, %f756;
setp.leu.f32 %p20, %f228, 0f47CE4780;
@%p20 bra BB0_22;
add.u64 %rd45, %SP, 0;
cvta.to.local.u64 %rd3, %rd45;
mov.b32 %r14, %f756;
shr.u32 %r15, %r14, 23;
shl.b32 %r140, %r14, 8;
or.b32 %r16, %r140, -2147483648;
mov.u32 %r344, 0;
mov.u64 %rd194, 0;
mov.u64 %rd193, %rd3;
mov.u32 %r345, %r344;
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, %r345;
madc.hi.u32 %r345, %r143, %r16, 0;
}
// inline asm
st.local.u32 [%rd193], %r141;
add.s32 %r344, %r344, 1;
cvt.s64.s32 %rd194, %r344;
mul.wide.s32 %rd51, %r344, 4;
add.s64 %rd193, %rd3, %rd51;
setp.ne.s32 %p21, %r344, 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], %r345;
mov.u32 %r149, 6;
sub.s32 %r150, %r149, %r148;
mul.wide.s32 %rd54, %r150, 4;
add.s64 %rd8, %rd53, %rd54;
ld.local.u32 %r346, [%rd8];
ld.local.u32 %r347, [%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, %r347, %r152;
shl.b32 %r154, %r346, %r24;
add.s32 %r346, %r153, %r154;
ld.local.u32 %r155, [%rd8+-8];
shr.u32 %r156, %r155, %r152;
shl.b32 %r157, %r347, %r24;
add.s32 %r347, %r156, %r157;
BB0_16:
shr.u32 %r158, %r347, 30;
shl.b32 %r159, %r346, 2;
add.s32 %r348, %r158, %r159;
shl.b32 %r30, %r347, 2;
shr.u32 %r160, %r348, 31;
shr.u32 %r161, %r346, 30;
add.s32 %r31, %r160, %r161;
setp.eq.s32 %p23, %r160, 0;
@%p23 bra BB0_17;
bra.uni BB0_18;
BB0_17:
mov.u32 %r349, %r21;
mov.u32 %r350, %r30;
bra.uni BB0_19;
BB0_18:
not.b32 %r162, %r348;
neg.s32 %r350, %r30;
setp.eq.s32 %p24, %r30, 0;
selp.u32 %r163, 1, 0, %p24;
add.s32 %r348, %r163, %r162;
xor.b32 %r349, %r21, -2147483648;
BB0_19:
clz.b32 %r352, %r348;
setp.eq.s32 %p25, %r352, 0;
shl.b32 %r164, %r348, %r352;
mov.u32 %r165, 32;
sub.s32 %r166, %r165, %r352;
shr.u32 %r167, %r350, %r166;
add.s32 %r168, %r167, %r164;
selp.b32 %r39, %r348, %r168, %p25;
mov.u32 %r169, -921707870;
mul.hi.u32 %r351, %r39, %r169;
setp.eq.s32 %p26, %r21, 0;
neg.s32 %r170, %r31;
selp.b32 %r353, %r31, %r170, %p26;
setp.lt.s32 %p27, %r351, 1;
@%p27 bra BB0_21;
mul.lo.s32 %r171, %r39, -921707870;
shr.u32 %r172, %r171, 31;
shl.b32 %r173, %r351, 1;
add.s32 %r351, %r172, %r173;
add.s32 %r352, %r352, 1;
BB0_21:
mov.u32 %r174, 126;
sub.s32 %r175, %r174, %r352;
shl.b32 %r176, %r175, 23;
add.s32 %r177, %r351, 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, %r349;
mov.b32 %f757, %r182;
BB0_22:
mul.rn.f32 %f43, %f757, %f757;
add.s32 %r47, %r353, 1;
and.b32 %r48, %r47, 1;
setp.eq.s32 %p28, %r48, 0;
@%p28 bra BB0_24;
bra.uni BB0_23;
BB0_24:
mov.f32 %f231, 0f3C08839E;
mov.f32 %f232, 0fB94CA1F9;
fma.rn.f32 %f758, %f232, %f43, %f231;
bra.uni BB0_25;
BB0_23:
mov.f32 %f229, 0fBAB6061A;
mov.f32 %f230, 0f37CCF5CE;
fma.rn.f32 %f758, %f230, %f43, %f229;
BB0_25:
@%p28 bra BB0_27;
bra.uni BB0_26;
BB0_27:
mov.f32 %f236, 0fBE2AAAA3;
fma.rn.f32 %f237, %f758, %f43, %f236;
mov.f32 %f238, 0f00000000;
fma.rn.f32 %f759, %f237, %f43, %f238;
bra.uni BB0_28;
BB0_26:
mov.f32 %f233, 0f3D2AAAA5;
fma.rn.f32 %f234, %f758, %f43, %f233;
mov.f32 %f235, 0fBF000000;
fma.rn.f32 %f759, %f234, %f43, %f235;
BB0_28:
fma.rn.f32 %f760, %f759, %f757, %f757;
@%p28 bra BB0_30;
mov.f32 %f239, 0f3F800000;
fma.rn.f32 %f760, %f759, %f43, %f239;
BB0_30:
and.b32 %r183, %r47, 2;
setp.eq.s32 %p31, %r183, 0;
@%p31 bra BB0_32;
mov.f32 %f240, 0f00000000;
mov.f32 %f241, 0fBF800000;
fma.rn.f32 %f760, %f760, %f241, %f240;
BB0_32:
@%p19 bra BB0_34;
mov.f32 %f242, 0f00000000;
mul.rn.f32 %f762, %f762, %f242;
BB0_34:
mul.f32 %f243, %f762, 0f3F22F983;
cvt.rni.s32.f32 %r363, %f243;
cvt.rn.f32.s32 %f244, %r363;
neg.f32 %f245, %f244;
fma.rn.f32 %f247, %f245, %f223, %f762;
fma.rn.f32 %f249, %f245, %f225, %f247;
fma.rn.f32 %f763, %f245, %f227, %f249;
abs.f32 %f251, %f762;
setp.leu.f32 %p33, %f251, 0f47CE4780;
@%p33 bra BB0_45;
add.u64 %rd56, %SP, 0;
cvta.to.local.u64 %rd9, %rd56;
mov.b32 %r50, %f762;
shr.u32 %r51, %r50, 23;
shl.b32 %r186, %r50, 8;
or.b32 %r52, %r186, -2147483648;
mov.u32 %r354, 0;
mov.u64 %rd195, %rd9;
mov.u64 %rd196, %rd28;
mov.u32 %r355, %r354;
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, %r355;
madc.hi.u32 %r355, %r189, %r52, 0;
}
// inline asm
st.local.u32 [%rd195], %r187;
add.s32 %r354, %r354, 1;
cvt.s64.s32 %rd196, %r354;
mul.wide.s32 %rd60, %r354, 4;
add.s64 %rd195, %rd9, %rd60;
setp.ne.s32 %p34, %r354, 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], %r355;
mov.u32 %r195, 6;
sub.s32 %r196, %r195, %r194;
mul.wide.s32 %rd63, %r196, 4;
add.s64 %rd15, %rd62, %rd63;
ld.local.u32 %r356, [%rd15];
ld.local.u32 %r357, [%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, %r357, %r198;
shl.b32 %r200, %r356, %r60;
add.s32 %r356, %r199, %r200;
ld.local.u32 %r201, [%rd15+-8];
shr.u32 %r202, %r201, %r198;
shl.b32 %r203, %r357, %r60;
add.s32 %r357, %r202, %r203;
BB0_39:
shr.u32 %r204, %r357, 30;
shl.b32 %r205, %r356, 2;
add.s32 %r358, %r204, %r205;
shl.b32 %r66, %r357, 2;
shr.u32 %r206, %r358, 31;
shr.u32 %r207, %r356, 30;
add.s32 %r67, %r206, %r207;
setp.eq.s32 %p36, %r206, 0;
@%p36 bra BB0_40;
bra.uni BB0_41;
BB0_40:
mov.u32 %r359, %r57;
mov.u32 %r360, %r66;
bra.uni BB0_42;
BB0_41:
not.b32 %r208, %r358;
neg.s32 %r360, %r66;
setp.eq.s32 %p37, %r66, 0;
selp.u32 %r209, 1, 0, %p37;
add.s32 %r358, %r209, %r208;
xor.b32 %r359, %r57, -2147483648;
BB0_42:
clz.b32 %r362, %r358;
setp.eq.s32 %p38, %r362, 0;
shl.b32 %r210, %r358, %r362;
mov.u32 %r211, 32;
sub.s32 %r212, %r211, %r362;
shr.u32 %r213, %r360, %r212;
add.s32 %r214, %r213, %r210;
selp.b32 %r75, %r358, %r214, %p38;
mov.u32 %r215, -921707870;
mul.hi.u32 %r361, %r75, %r215;
setp.eq.s32 %p39, %r57, 0;
neg.s32 %r216, %r67;
selp.b32 %r363, %r67, %r216, %p39;
setp.lt.s32 %p40, %r361, 1;
@%p40 bra BB0_44;
mul.lo.s32 %r217, %r75, -921707870;
shr.u32 %r218, %r217, 31;
shl.b32 %r219, %r361, 1;
add.s32 %r361, %r218, %r219;
add.s32 %r362, %r362, 1;
BB0_44:
mov.u32 %r220, 126;
sub.s32 %r221, %r220, %r362;
shl.b32 %r222, %r221, 23;
add.s32 %r223, %r361, 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, %r359;
mov.b32 %f763, %r228;
BB0_45:
mul.rn.f32 %f60, %f763, %f763;
and.b32 %r83, %r363, 1;
setp.eq.s32 %p41, %r83, 0;
@%p41 bra BB0_47;
bra.uni BB0_46;
BB0_47:
mov.f32 %f254, 0f3C08839E;
mov.f32 %f255, 0fB94CA1F9;
fma.rn.f32 %f764, %f255, %f60, %f254;
bra.uni BB0_48;
BB0_46:
mov.f32 %f252, 0fBAB6061A;
mov.f32 %f253, 0f37CCF5CE;
fma.rn.f32 %f764, %f253, %f60, %f252;
BB0_48:
@%p41 bra BB0_50;
bra.uni BB0_49;
BB0_50:
mov.f32 %f259, 0fBE2AAAA3;
fma.rn.f32 %f260, %f764, %f60, %f259;
mov.f32 %f261, 0f00000000;
fma.rn.f32 %f765, %f260, %f60, %f261;
bra.uni BB0_51;
BB0_49:
mov.f32 %f256, 0f3D2AAAA5;
fma.rn.f32 %f257, %f764, %f60, %f256;
mov.f32 %f258, 0fBF000000;
fma.rn.f32 %f765, %f257, %f60, %f258;
BB0_51:
fma.rn.f32 %f766, %f765, %f763, %f763;
@%p41 bra BB0_53;
mov.f32 %f262, 0f3F800000;
fma.rn.f32 %f766, %f765, %f60, %f262;
BB0_53:
and.b32 %r229, %r363, 2;
setp.eq.s32 %p44, %r229, 0;
@%p44 bra BB0_55;
mov.f32 %f263, 0f00000000;
mov.f32 %f264, 0fBF800000;
fma.rn.f32 %f766, %f766, %f264, %f263;
BB0_55:
mul.f32 %f273, %f35, %f760;
add.u64 %rd64, %SP, 28;
cvta.to.local.u64 %rd65, %rd64;
mul.f32 %f274, %f273, %f273;
mov.f32 %f275, 0f3F800000;
sub.f32 %f276, %f275, %f274;
mul.f32 %f277, %f35, %f766;
mul.f32 %f278, %f277, %f277;
sub.f32 %f279, %f276, %f278;
mov.f32 %f280, 0f00000000;
max.f32 %f281, %f280, %f279;
sqrt.rn.f32 %f282, %f281;
mul.f32 %f283, %f17, %f277;
mul.f32 %f284, %f18, %f277;
mul.f32 %f285, %f19, %f277;
fma.rn.f32 %f286, %f20, %f273, %f283;
fma.rn.f32 %f287, %f21, %f273, %f284;
fma.rn.f32 %f288, %f22, %f273, %f285;
fma.rn.f32 %f289, %f7, %f282, %f286;
fma.rn.f32 %f290, %f8, %f282, %f287;
fma.rn.f32 %f291, %f9, %f282, %f288;
add.f32 %f292, %f7, %f289;
add.f32 %f293, %f8, %f290;
add.f32 %f294, %f9, %f291;
ld.global.f32 %f295, [shadowSpread];
mul.f32 %f296, %f295, %f292;
mul.f32 %f297, %f295, %f293;
mul.f32 %f298, %f295, %f294;
sub.f32 %f299, %f296, %f7;
sub.f32 %f300, %f297, %f8;
sub.f32 %f301, %f298, %f9;
mul.f32 %f302, %f300, %f300;
fma.rn.f32 %f303, %f299, %f299, %f302;
fma.rn.f32 %f304, %f301, %f301, %f303;
sqrt.rn.f32 %f305, %f304;
rcp.rn.f32 %f306, %f305;
mul.f32 %f268, %f306, %f299;
mul.f32 %f269, %f306, %f300;
mul.f32 %f270, %f306, %f301;
ld.global.u32 %r233, [imageEnabled];
and.b32 %r234, %r233, 32;
setp.eq.s32 %p45, %r234, 0;
selp.f32 %f307, 0f3F800000, 0f41200000, %p45;
mul.f32 %f271, %f307, %f31;
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 %f272, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r230, %f14, %f15, %f16, %f268, %f269, %f270, %r231, %f271, %f272, %rd64, %r122);
// inline asm
ld.local.f32 %f308, [%rd65];
add.f32 %f770, %f770, %f308;
ld.local.f32 %f309, [%rd65+4];
add.f32 %f769, %f769, %f309;
ld.local.f32 %f310, [%rd65+8];
add.f32 %f768, %f768, %f310;
ld.global.u32 %r339, [samples];
add.s32 %r342, %r342, 1;
setp.lt.s32 %p46, %r342, %r339;
@%p46 bra BB0_9;
BB0_56:
add.s32 %r340, %r340, 1;
setp.lt.s32 %p47, %r340, %r339;
@%p47 bra BB0_7;
bra.uni BB0_57;
BB0_130:
mov.u64 %rd181, image_HDR;
cvta.global.u64 %rd176, %rd181;
mov.u32 %r330, 8;
mov.u64 %rd180, 0;
// inline asm
call (%rd175), _rt_buffer_get_64, (%rd176, %r104, %r330, %rd21, %rd22, %rd180, %rd180);
// inline asm
mov.f32 %f719, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs71, %f719;}
// 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 %f769, %f768;
mov.f32 %f770, %f768;
BB0_57:
setp.eq.s32 %p48, %r339, 0;
mov.f32 %f774, 0f3F800000;
mov.f32 %f775, %f774;
mov.f32 %f776, %f774;
@%p48 bra BB0_59;
mul.lo.s32 %r236, %r339, %r339;
cvt.rn.f32.s32 %f314, %r236;
rcp.rn.f32 %f315, %f314;
mul.f32 %f774, %f770, %f315;
mul.f32 %f775, %f769, %f315;
mul.f32 %f776, %f768, %f315;
BB0_59:
mul.f32 %f324, %f12, %f21;
fma.rn.f32 %f325, %f11, %f20, %f324;
fma.rn.f32 %f326, %f13, %f22, %f325;
ld.global.v4.f32 {%f327, %f328, %f329, %f330}, [lightTilingOffset];
fma.rn.f32 %f320, %f326, %f327, %f329;
mul.f32 %f333, %f12, %f18;
fma.rn.f32 %f334, %f11, %f17, %f333;
fma.rn.f32 %f335, %f13, %f19, %f334;
fma.rn.f32 %f321, %f335, %f328, %f330;
ld.global.u32 %r237, [lightCookie];
mov.f32 %f323, 0f00000000;
// inline asm
call (%f316, %f317, %f318, %f319), _rt_texture_get_f_id, (%r237, %r104, %f320, %f321, %f323, %f323);
// inline asm
min.f32 %f338, %f316, %f317;
min.f32 %f339, %f338, %f318;
mul.f32 %f87, %f774, %f339;
ld.global.f32 %f340, [directColor];
mul.f32 %f341, %f774, %f340;
ld.global.f32 %f342, [directColor+4];
mul.f32 %f343, %f775, %f342;
ld.global.f32 %f344, [directColor+8];
mul.f32 %f345, %f776, %f344;
mul.f32 %f346, %f316, %f341;
mul.f32 %f347, %f317, %f343;
mul.f32 %f348, %f318, %f345;
cvt.sat.f32.f32 %f349, %f10;
mul.f32 %f350, %f346, %f349;
mul.f32 %f351, %f347, %f349;
mul.f32 %f352, %f349, %f348;
ld.global.u32 %r239, [ignoreNormal];
setp.eq.s32 %p49, %r239, 0;
selp.f32 %f88, %f350, %f346, %p49;
selp.f32 %f89, %f351, %f347, %p49;
selp.f32 %f90, %f352, %f348, %p49;
ld.global.u32 %r368, [imageEnabled];
and.b32 %r240, %r368, 8;
setp.eq.s32 %p50, %r240, 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 %f355, 0f3E68BA2E;
cvt.rzi.f32.f32 %f356, %f355;
fma.rn.f32 %f357, %f356, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f91, %f357;
abs.f32 %f92, %f87;
setp.lt.f32 %p51, %f92, 0f00800000;
mul.f32 %f358, %f92, 0f4B800000;
selp.f32 %f359, 0fC3170000, 0fC2FE0000, %p51;
selp.f32 %f360, %f358, %f92, %p51;
mov.b32 %r243, %f360;
and.b32 %r244, %r243, 8388607;
or.b32 %r245, %r244, 1065353216;
mov.b32 %f361, %r245;
shr.u32 %r246, %r243, 23;
cvt.rn.f32.u32 %f362, %r246;
add.f32 %f363, %f359, %f362;
setp.gt.f32 %p52, %f361, 0f3FB504F3;
mul.f32 %f364, %f361, 0f3F000000;
add.f32 %f365, %f363, 0f3F800000;
selp.f32 %f366, %f364, %f361, %p52;
selp.f32 %f367, %f365, %f363, %p52;
add.f32 %f368, %f366, 0fBF800000;
add.f32 %f354, %f366, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f353,%f354;
// inline asm
add.f32 %f369, %f368, %f368;
mul.f32 %f370, %f353, %f369;
mul.f32 %f371, %f370, %f370;
mov.f32 %f372, 0f3C4CAF63;
mov.f32 %f373, 0f3B18F0FE;
fma.rn.f32 %f374, %f373, %f371, %f372;
mov.f32 %f375, 0f3DAAAABD;
fma.rn.f32 %f376, %f374, %f371, %f375;
mul.rn.f32 %f377, %f376, %f371;
mul.rn.f32 %f378, %f377, %f370;
sub.f32 %f379, %f368, %f370;
neg.f32 %f380, %f370;
add.f32 %f381, %f379, %f379;
fma.rn.f32 %f382, %f380, %f368, %f381;
mul.rn.f32 %f383, %f353, %f382;
add.f32 %f384, %f378, %f370;
sub.f32 %f385, %f370, %f384;
add.f32 %f386, %f378, %f385;
add.f32 %f387, %f383, %f386;
add.f32 %f388, %f384, %f387;
sub.f32 %f389, %f384, %f388;
add.f32 %f390, %f387, %f389;
mov.f32 %f391, 0f3F317200;
mul.rn.f32 %f392, %f367, %f391;
mov.f32 %f393, 0f35BFBE8E;
mul.rn.f32 %f394, %f367, %f393;
add.f32 %f395, %f392, %f388;
sub.f32 %f396, %f392, %f395;
add.f32 %f397, %f388, %f396;
add.f32 %f398, %f390, %f397;
add.f32 %f399, %f394, %f398;
add.f32 %f400, %f395, %f399;
sub.f32 %f401, %f395, %f400;
add.f32 %f402, %f399, %f401;
mov.f32 %f403, 0f3EE8BA2E;
mul.rn.f32 %f404, %f403, %f400;
neg.f32 %f405, %f404;
fma.rn.f32 %f406, %f403, %f400, %f405;
fma.rn.f32 %f407, %f403, %f402, %f406;
fma.rn.f32 %f409, %f323, %f400, %f407;
add.rn.f32 %f410, %f404, %f409;
neg.f32 %f411, %f410;
add.rn.f32 %f412, %f404, %f411;
add.rn.f32 %f413, %f412, %f409;
mov.b32 %r247, %f410;
setp.eq.s32 %p53, %r247, 1118925336;
add.s32 %r248, %r247, -1;
mov.b32 %f414, %r248;
add.f32 %f415, %f413, 0f37000000;
selp.f32 %f416, %f414, %f410, %p53;
selp.f32 %f93, %f415, %f413, %p53;
mul.f32 %f417, %f416, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f418, %f417;
mov.f32 %f419, 0fBF317200;
fma.rn.f32 %f420, %f418, %f419, %f416;
mov.f32 %f421, 0fB5BFBE8E;
fma.rn.f32 %f422, %f418, %f421, %f420;
mul.f32 %f423, %f422, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f424, %f423;
add.f32 %f425, %f418, 0f00000000;
ex2.approx.f32 %f426, %f425;
mul.f32 %f427, %f424, %f426;
setp.lt.f32 %p54, %f416, 0fC2D20000;
selp.f32 %f428, 0f00000000, %f427, %p54;
setp.gt.f32 %p55, %f416, 0f42D20000;
selp.f32 %f777, 0f7F800000, %f428, %p55;
setp.eq.f32 %p56, %f777, 0f7F800000;
@%p56 bra BB0_62;
fma.rn.f32 %f777, %f777, %f93, %f777;
BB0_62:
setp.lt.f32 %p57, %f87, 0f00000000;
setp.eq.f32 %p58, %f91, 0f3F800000;
and.pred %p1, %p57, %p58;
mov.b32 %r249, %f777;
xor.b32 %r250, %r249, -2147483648;
mov.b32 %f429, %r250;
selp.f32 %f779, %f429, %f777, %p1;
setp.eq.f32 %p59, %f87, 0f00000000;
@%p59 bra BB0_65;
bra.uni BB0_63;
BB0_65:
add.f32 %f432, %f87, %f87;
selp.f32 %f779, %f432, 0f00000000, %p58;
bra.uni BB0_66;
BB0_120:
mov.u64 %rd140, image_HDR;
cvta.global.u64 %rd135, %rd140;
mov.u32 %r313, 8;
mov.u64 %rd139, 0;
// inline asm
call (%rd134), _rt_buffer_get_64, (%rd135, %r104, %r313, %rd19, %rd20, %rd139, %rd139);
// inline asm
mov.f32 %f709, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs52, %f709;}
// 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, %f87, 0f00000000;
@%p60 bra BB0_66;
mov.f32 %f746, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f431, %f746;
setp.neu.f32 %p61, %f431, 0f3EE8BA2E;
selp.f32 %f779, 0f7FFFFFFF, %f779, %p61;
BB0_66:
add.f32 %f433, %f92, 0f3EE8BA2E;
mov.b32 %r251, %f433;
setp.lt.s32 %p63, %r251, 2139095040;
@%p63 bra BB0_71;
setp.gtu.f32 %p64, %f92, 0f7F800000;
@%p64 bra BB0_70;
bra.uni BB0_68;
BB0_70:
add.f32 %f779, %f87, 0f3EE8BA2E;
bra.uni BB0_71;
BB0_68:
setp.neu.f32 %p65, %f92, 0f7F800000;
@%p65 bra BB0_71;
selp.f32 %f779, 0fFF800000, 0f7F800000, %p1;
BB0_71:
mul.f32 %f434, %f779, 0f437F0000;
setp.eq.f32 %p66, %f87, 0f3F800000;
selp.f32 %f435, 0f437F0000, %f434, %p66;
cvt.rzi.u32.f32 %r252, %f435;
cvt.u16.u32 %rs11, %r252;
mov.u16 %rs12, 255;
st.v2.u8 [%rd66], {%rs11, %rs12};
ld.global.u32 %r368, [imageEnabled];
BB0_72:
and.b32 %r253, %r368, 1;
setp.eq.b32 %p67, %r253, 1;
@!%p67 bra BB0_107;
bra.uni BB0_73;
BB0_73:
mov.f32 %f745, 0f00000000;
mov.f32 %f438, 0f3E666666;
cvt.rzi.f32.f32 %f439, %f438;
fma.rn.f32 %f440, %f439, 0fC0000000, 0f3EE66666;
abs.f32 %f104, %f440;
abs.f32 %f105, %f88;
setp.lt.f32 %p68, %f105, 0f00800000;
mul.f32 %f441, %f105, 0f4B800000;
selp.f32 %f442, 0fC3170000, 0fC2FE0000, %p68;
selp.f32 %f443, %f441, %f105, %p68;
mov.b32 %r254, %f443;
and.b32 %r255, %r254, 8388607;
or.b32 %r256, %r255, 1065353216;
mov.b32 %f444, %r256;
shr.u32 %r257, %r254, 23;
cvt.rn.f32.u32 %f445, %r257;
add.f32 %f446, %f442, %f445;
setp.gt.f32 %p69, %f444, 0f3FB504F3;
mul.f32 %f447, %f444, 0f3F000000;
add.f32 %f448, %f446, 0f3F800000;
selp.f32 %f449, %f447, %f444, %p69;
selp.f32 %f450, %f448, %f446, %p69;
add.f32 %f451, %f449, 0fBF800000;
add.f32 %f437, %f449, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f436,%f437;
// inline asm
add.f32 %f452, %f451, %f451;
mul.f32 %f453, %f436, %f452;
mul.f32 %f454, %f453, %f453;
mov.f32 %f455, 0f3C4CAF63;
mov.f32 %f456, 0f3B18F0FE;
fma.rn.f32 %f457, %f456, %f454, %f455;
mov.f32 %f458, 0f3DAAAABD;
fma.rn.f32 %f459, %f457, %f454, %f458;
mul.rn.f32 %f460, %f459, %f454;
mul.rn.f32 %f461, %f460, %f453;
sub.f32 %f462, %f451, %f453;
neg.f32 %f463, %f453;
add.f32 %f464, %f462, %f462;
fma.rn.f32 %f465, %f463, %f451, %f464;
mul.rn.f32 %f466, %f436, %f465;
add.f32 %f467, %f461, %f453;
sub.f32 %f468, %f453, %f467;
add.f32 %f469, %f461, %f468;
add.f32 %f470, %f466, %f469;
add.f32 %f471, %f467, %f470;
sub.f32 %f472, %f467, %f471;
add.f32 %f473, %f470, %f472;
mov.f32 %f474, 0f3F317200;
mul.rn.f32 %f475, %f450, %f474;
mov.f32 %f476, 0f35BFBE8E;
mul.rn.f32 %f477, %f450, %f476;
add.f32 %f478, %f475, %f471;
sub.f32 %f479, %f475, %f478;
add.f32 %f480, %f471, %f479;
add.f32 %f481, %f473, %f480;
add.f32 %f482, %f477, %f481;
add.f32 %f483, %f478, %f482;
sub.f32 %f484, %f478, %f483;
add.f32 %f485, %f482, %f484;
mov.f32 %f486, 0f3EE66666;
mul.rn.f32 %f487, %f486, %f483;
neg.f32 %f488, %f487;
fma.rn.f32 %f489, %f486, %f483, %f488;
fma.rn.f32 %f490, %f486, %f485, %f489;
fma.rn.f32 %f492, %f745, %f483, %f490;
add.rn.f32 %f493, %f487, %f492;
neg.f32 %f494, %f493;
add.rn.f32 %f495, %f487, %f494;
add.rn.f32 %f496, %f495, %f492;
mov.b32 %r258, %f493;
setp.eq.s32 %p70, %r258, 1118925336;
add.s32 %r259, %r258, -1;
mov.b32 %f497, %r259;
add.f32 %f498, %f496, 0f37000000;
selp.f32 %f499, %f497, %f493, %p70;
selp.f32 %f106, %f498, %f496, %p70;
mul.f32 %f500, %f499, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f501, %f500;
mov.f32 %f502, 0fBF317200;
fma.rn.f32 %f503, %f501, %f502, %f499;
mov.f32 %f504, 0fB5BFBE8E;
fma.rn.f32 %f505, %f501, %f504, %f503;
mul.f32 %f506, %f505, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f507, %f506;
add.f32 %f508, %f501, 0f00000000;
ex2.approx.f32 %f509, %f508;
mul.f32 %f510, %f507, %f509;
setp.lt.f32 %p71, %f499, 0fC2D20000;
selp.f32 %f511, 0f00000000, %f510, %p71;
setp.gt.f32 %p72, %f499, 0f42D20000;
selp.f32 %f780, 0f7F800000, %f511, %p72;
setp.eq.f32 %p73, %f780, 0f7F800000;
@%p73 bra BB0_75;
fma.rn.f32 %f780, %f780, %f106, %f780;
BB0_75:
setp.lt.f32 %p74, %f88, 0f00000000;
setp.eq.f32 %p75, %f104, 0f3F800000;
and.pred %p2, %p74, %p75;
mov.b32 %r260, %f780;
xor.b32 %r261, %r260, -2147483648;
mov.b32 %f512, %r261;
selp.f32 %f782, %f512, %f780, %p2;
setp.eq.f32 %p76, %f88, 0f00000000;
@%p76 bra BB0_78;
bra.uni BB0_76;
BB0_78:
add.f32 %f515, %f88, %f88;
selp.f32 %f782, %f515, 0f00000000, %p75;
bra.uni BB0_79;
BB0_76:
setp.geu.f32 %p77, %f88, 0f00000000;
@%p77 bra BB0_79;
cvt.rzi.f32.f32 %f514, %f486;
setp.neu.f32 %p78, %f514, 0f3EE66666;
selp.f32 %f782, 0f7FFFFFFF, %f782, %p78;
BB0_79:
abs.f32 %f720, %f88;
add.f32 %f516, %f720, 0f3EE66666;
mov.b32 %r262, %f516;
setp.lt.s32 %p80, %r262, 2139095040;
@%p80 bra BB0_84;
abs.f32 %f737, %f88;
setp.gtu.f32 %p81, %f737, 0f7F800000;
@%p81 bra BB0_83;
bra.uni BB0_81;
BB0_83:
add.f32 %f782, %f88, 0f3EE66666;
bra.uni BB0_84;
BB0_81:
abs.f32 %f738, %f88;
setp.neu.f32 %p82, %f738, 0f7F800000;
@%p82 bra BB0_84;
selp.f32 %f782, 0fFF800000, 0f7F800000, %p2;
BB0_84:
mov.f32 %f728, 0fB5BFBE8E;
mov.f32 %f727, 0fBF317200;
mov.f32 %f726, 0f35BFBE8E;
mov.f32 %f725, 0f3F317200;
mov.f32 %f724, 0f3DAAAABD;
mov.f32 %f723, 0f3C4CAF63;
mov.f32 %f722, 0f3B18F0FE;
mov.f32 %f721, 0f00000000;
setp.eq.f32 %p83, %f88, 0f3F800000;
selp.f32 %f117, 0f3F800000, %f782, %p83;
abs.f32 %f118, %f89;
setp.lt.f32 %p84, %f118, 0f00800000;
mul.f32 %f519, %f118, 0f4B800000;
selp.f32 %f520, 0fC3170000, 0fC2FE0000, %p84;
selp.f32 %f521, %f519, %f118, %p84;
mov.b32 %r263, %f521;
and.b32 %r264, %r263, 8388607;
or.b32 %r265, %r264, 1065353216;
mov.b32 %f522, %r265;
shr.u32 %r266, %r263, 23;
cvt.rn.f32.u32 %f523, %r266;
add.f32 %f524, %f520, %f523;
setp.gt.f32 %p85, %f522, 0f3FB504F3;
mul.f32 %f525, %f522, 0f3F000000;
add.f32 %f526, %f524, 0f3F800000;
selp.f32 %f527, %f525, %f522, %p85;
selp.f32 %f528, %f526, %f524, %p85;
add.f32 %f529, %f527, 0fBF800000;
add.f32 %f518, %f527, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f517,%f518;
// inline asm
add.f32 %f530, %f529, %f529;
mul.f32 %f531, %f517, %f530;
mul.f32 %f532, %f531, %f531;
fma.rn.f32 %f535, %f722, %f532, %f723;
fma.rn.f32 %f537, %f535, %f532, %f724;
mul.rn.f32 %f538, %f537, %f532;
mul.rn.f32 %f539, %f538, %f531;
sub.f32 %f540, %f529, %f531;
neg.f32 %f541, %f531;
add.f32 %f542, %f540, %f540;
fma.rn.f32 %f543, %f541, %f529, %f542;
mul.rn.f32 %f544, %f517, %f543;
add.f32 %f545, %f539, %f531;
sub.f32 %f546, %f531, %f545;
add.f32 %f547, %f539, %f546;
add.f32 %f548, %f544, %f547;
add.f32 %f549, %f545, %f548;
sub.f32 %f550, %f545, %f549;
add.f32 %f551, %f548, %f550;
mul.rn.f32 %f553, %f528, %f725;
mul.rn.f32 %f555, %f528, %f726;
add.f32 %f556, %f553, %f549;
sub.f32 %f557, %f553, %f556;
add.f32 %f558, %f549, %f557;
add.f32 %f559, %f551, %f558;
add.f32 %f560, %f555, %f559;
add.f32 %f561, %f556, %f560;
sub.f32 %f562, %f556, %f561;
add.f32 %f563, %f560, %f562;
mul.rn.f32 %f565, %f486, %f561;
neg.f32 %f566, %f565;
fma.rn.f32 %f567, %f486, %f561, %f566;
fma.rn.f32 %f568, %f486, %f563, %f567;
fma.rn.f32 %f570, %f721, %f561, %f568;
add.rn.f32 %f571, %f565, %f570;
neg.f32 %f572, %f571;
add.rn.f32 %f573, %f565, %f572;
add.rn.f32 %f574, %f573, %f570;
mov.b32 %r267, %f571;
setp.eq.s32 %p86, %r267, 1118925336;
add.s32 %r268, %r267, -1;
mov.b32 %f575, %r268;
add.f32 %f576, %f574, 0f37000000;
selp.f32 %f577, %f575, %f571, %p86;
selp.f32 %f119, %f576, %f574, %p86;
mul.f32 %f578, %f577, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f579, %f578;
fma.rn.f32 %f581, %f579, %f727, %f577;
fma.rn.f32 %f583, %f579, %f728, %f581;
mul.f32 %f584, %f583, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f585, %f584;
add.f32 %f586, %f579, 0f00000000;
ex2.approx.f32 %f587, %f586;
mul.f32 %f588, %f585, %f587;
setp.lt.f32 %p87, %f577, 0fC2D20000;
selp.f32 %f589, 0f00000000, %f588, %p87;
setp.gt.f32 %p88, %f577, 0f42D20000;
selp.f32 %f783, 0f7F800000, %f589, %p88;
setp.eq.f32 %p89, %f783, 0f7F800000;
@%p89 bra BB0_86;
fma.rn.f32 %f783, %f783, %f119, %f783;
BB0_86:
setp.lt.f32 %p90, %f89, 0f00000000;
and.pred %p3, %p90, %p75;
mov.b32 %r269, %f783;
xor.b32 %r270, %r269, -2147483648;
mov.b32 %f590, %r270;
selp.f32 %f785, %f590, %f783, %p3;
setp.eq.f32 %p92, %f89, 0f00000000;
@%p92 bra BB0_89;
bra.uni BB0_87;
BB0_89:
add.f32 %f593, %f89, %f89;
selp.f32 %f785, %f593, 0f00000000, %p75;
bra.uni BB0_90;
BB0_87:
setp.geu.f32 %p93, %f89, 0f00000000;
@%p93 bra BB0_90;
mov.f32 %f744, 0f3EE66666;
cvt.rzi.f32.f32 %f592, %f744;
setp.neu.f32 %p94, %f592, 0f3EE66666;
selp.f32 %f785, 0f7FFFFFFF, %f785, %p94;
BB0_90:
abs.f32 %f739, %f89;
add.f32 %f594, %f739, 0f3EE66666;
mov.b32 %r271, %f594;
setp.lt.s32 %p96, %r271, 2139095040;
@%p96 bra BB0_95;
abs.f32 %f742, %f89;
setp.gtu.f32 %p97, %f742, 0f7F800000;
@%p97 bra BB0_94;
bra.uni BB0_92;
BB0_94:
add.f32 %f785, %f89, 0f3EE66666;
bra.uni BB0_95;
BB0_92:
abs.f32 %f743, %f89;
setp.neu.f32 %p98, %f743, 0f7F800000;
@%p98 bra BB0_95;
selp.f32 %f785, 0fFF800000, 0f7F800000, %p3;
BB0_95:
mov.f32 %f740, 0f3EE66666;
mov.f32 %f736, 0fB5BFBE8E;
mov.f32 %f735, 0fBF317200;
mov.f32 %f734, 0f35BFBE8E;
mov.f32 %f733, 0f3F317200;
mov.f32 %f732, 0f3DAAAABD;
mov.f32 %f731, 0f3C4CAF63;
mov.f32 %f730, 0f3B18F0FE;
mov.f32 %f729, 0f00000000;
setp.eq.f32 %p99, %f89, 0f3F800000;
selp.f32 %f130, 0f3F800000, %f785, %p99;
abs.f32 %f131, %f90;
setp.lt.f32 %p100, %f131, 0f00800000;
mul.f32 %f597, %f131, 0f4B800000;
selp.f32 %f598, 0fC3170000, 0fC2FE0000, %p100;
selp.f32 %f599, %f597, %f131, %p100;
mov.b32 %r272, %f599;
and.b32 %r273, %r272, 8388607;
or.b32 %r274, %r273, 1065353216;
mov.b32 %f600, %r274;
shr.u32 %r275, %r272, 23;
cvt.rn.f32.u32 %f601, %r275;
add.f32 %f602, %f598, %f601;
setp.gt.f32 %p101, %f600, 0f3FB504F3;
mul.f32 %f603, %f600, 0f3F000000;
add.f32 %f604, %f602, 0f3F800000;
selp.f32 %f605, %f603, %f600, %p101;
selp.f32 %f606, %f604, %f602, %p101;
add.f32 %f607, %f605, 0fBF800000;
add.f32 %f596, %f605, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f595,%f596;
// inline asm
add.f32 %f608, %f607, %f607;
mul.f32 %f609, %f595, %f608;
mul.f32 %f610, %f609, %f609;
fma.rn.f32 %f613, %f730, %f610, %f731;
fma.rn.f32 %f615, %f613, %f610, %f732;
mul.rn.f32 %f616, %f615, %f610;
mul.rn.f32 %f617, %f616, %f609;
sub.f32 %f618, %f607, %f609;
neg.f32 %f619, %f609;
add.f32 %f620, %f618, %f618;
fma.rn.f32 %f621, %f619, %f607, %f620;
mul.rn.f32 %f622, %f595, %f621;
add.f32 %f623, %f617, %f609;
sub.f32 %f624, %f609, %f623;
add.f32 %f625, %f617, %f624;
add.f32 %f626, %f622, %f625;
add.f32 %f627, %f623, %f626;
sub.f32 %f628, %f623, %f627;
add.f32 %f629, %f626, %f628;
mul.rn.f32 %f631, %f606, %f733;
mul.rn.f32 %f633, %f606, %f734;
add.f32 %f634, %f631, %f627;
sub.f32 %f635, %f631, %f634;
add.f32 %f636, %f627, %f635;
add.f32 %f637, %f629, %f636;
add.f32 %f638, %f633, %f637;
add.f32 %f639, %f634, %f638;
sub.f32 %f640, %f634, %f639;
add.f32 %f641, %f638, %f640;
mul.rn.f32 %f643, %f740, %f639;
neg.f32 %f644, %f643;
fma.rn.f32 %f645, %f740, %f639, %f644;
fma.rn.f32 %f646, %f740, %f641, %f645;
fma.rn.f32 %f648, %f729, %f639, %f646;
add.rn.f32 %f649, %f643, %f648;
neg.f32 %f650, %f649;
add.rn.f32 %f651, %f643, %f650;
add.rn.f32 %f652, %f651, %f648;
mov.b32 %r276, %f649;
setp.eq.s32 %p102, %r276, 1118925336;
add.s32 %r277, %r276, -1;
mov.b32 %f653, %r277;
add.f32 %f654, %f652, 0f37000000;
selp.f32 %f655, %f653, %f649, %p102;
selp.f32 %f132, %f654, %f652, %p102;
mul.f32 %f656, %f655, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f657, %f656;
fma.rn.f32 %f659, %f657, %f735, %f655;
fma.rn.f32 %f661, %f657, %f736, %f659;
mul.f32 %f662, %f661, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f663, %f662;
add.f32 %f664, %f657, 0f00000000;
ex2.approx.f32 %f665, %f664;
mul.f32 %f666, %f663, %f665;
setp.lt.f32 %p103, %f655, 0fC2D20000;
selp.f32 %f667, 0f00000000, %f666, %p103;
setp.gt.f32 %p104, %f655, 0f42D20000;
selp.f32 %f786, 0f7F800000, %f667, %p104;
setp.eq.f32 %p105, %f786, 0f7F800000;
@%p105 bra BB0_97;
fma.rn.f32 %f786, %f786, %f132, %f786;
BB0_97:
setp.lt.f32 %p106, %f90, 0f00000000;
and.pred %p4, %p106, %p75;
mov.b32 %r278, %f786;
xor.b32 %r279, %r278, -2147483648;
mov.b32 %f668, %r279;
selp.f32 %f788, %f668, %f786, %p4;
setp.eq.f32 %p108, %f90, 0f00000000;
@%p108 bra BB0_100;
bra.uni BB0_98;
BB0_100:
add.f32 %f671, %f90, %f90;
selp.f32 %f788, %f671, 0f00000000, %p75;
bra.uni BB0_101;
BB0_98:
setp.geu.f32 %p109, %f90, 0f00000000;
@%p109 bra BB0_101;
mov.f32 %f741, 0f3EE66666;
cvt.rzi.f32.f32 %f670, %f741;
setp.neu.f32 %p110, %f670, 0f3EE66666;
selp.f32 %f788, 0f7FFFFFFF, %f788, %p110;
BB0_101:
add.f32 %f672, %f131, 0f3EE66666;
mov.b32 %r280, %f672;
setp.lt.s32 %p112, %r280, 2139095040;
@%p112 bra BB0_106;
setp.gtu.f32 %p113, %f131, 0f7F800000;
@%p113 bra BB0_105;
bra.uni BB0_103;
BB0_105:
add.f32 %f788, %f90, 0f3EE66666;
bra.uni BB0_106;
BB0_103:
setp.neu.f32 %p114, %f131, 0f7F800000;
@%p114 bra BB0_106;
selp.f32 %f788, 0fFF800000, 0f7F800000, %p4;
BB0_106:
mov.u32 %r334, 4;
mov.u64 %rd189, 0;
mov.u32 %r333, 2;
setp.eq.f32 %p115, %f90, 0f3F800000;
selp.f32 %f673, 0f3F800000, %f788, %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, %r333, %r334, %rd75, %rd76, %rd189, %rd189);
// inline asm
cvt.sat.f32.f32 %f674, %f673;
mul.f32 %f675, %f674, 0f437FFD71;
cvt.rzi.u32.f32 %r283, %f675;
cvt.sat.f32.f32 %f676, %f130;
mul.f32 %f677, %f676, 0f437FFD71;
cvt.rzi.u32.f32 %r284, %f677;
cvt.sat.f32.f32 %f678, %f117;
mul.f32 %f679, %f678, 0f437FFD71;
cvt.rzi.u32.f32 %r285, %f679;
cvt.u16.u32 %rs13, %r283;
cvt.u16.u32 %rs14, %r285;
cvt.u16.u32 %rs15, %r284;
mov.u16 %rs16, 255;
st.v4.u8 [%rd73], {%rs13, %rs15, %rs14, %rs16};
ld.global.u32 %r368, [imageEnabled];
BB0_107:
and.b32 %r286, %r368, 4;
setp.eq.s32 %p116, %r286, 0;
@%p116 bra BB0_111;
ld.global.u32 %r287, [additive];
setp.eq.s32 %p117, %r287, 0;
cvt.u64.u32 %rd17, %r2;
cvt.u64.u32 %rd18, %r3;
mov.f32 %f680, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs17, %f680;}
// inline asm
@%p117 bra BB0_110;
mov.u64 %rd190, 0;
mov.u32 %r335, 2;
mov.u64 %rd92, image_HDR;
cvta.global.u64 %rd81, %rd92;
mov.u32 %r291, 8;
// inline asm
call (%rd80), _rt_buffer_get_64, (%rd81, %r335, %r291, %rd17, %rd18, %rd190, %rd190);
// inline asm
ld.v4.u16 {%rs24, %rs25, %rs26, %rs27}, [%rd80];
// inline asm
{ cvt.f32.f16 %f681, %rs24;}
// inline asm
// inline asm
{ cvt.f32.f16 %f682, %rs25;}
// inline asm
// inline asm
{ cvt.f32.f16 %f683, %rs26;}
// inline asm
// inline asm
call (%rd86), _rt_buffer_get_64, (%rd81, %r335, %r291, %rd17, %rd18, %rd190, %rd190);
// inline asm
add.f32 %f684, %f88, %f681;
add.f32 %f685, %f89, %f682;
add.f32 %f686, %f90, %f683;
// inline asm
{ cvt.rn.f16.f32 %rs23, %f686;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs22, %f685;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs21, %f684;}
// inline asm
st.v4.u16 [%rd86], {%rs21, %rs22, %rs23, %rs17};
bra.uni BB0_111;
BB0_110:
mov.u64 %rd191, 0;
mov.u32 %r336, 2;
mov.u64 %rd99, image_HDR;
cvta.global.u64 %rd94, %rd99;
mov.u32 %r293, 8;
// inline asm
call (%rd93), _rt_buffer_get_64, (%rd94, %r336, %r293, %rd17, %rd18, %rd191, %rd191);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs30, %f90;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs29, %f89;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs28, %f88;}
// 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 %r338, 4;
mov.u64 %rd192, 0;
mov.u32 %r337, 2;
ld.global.f32 %f690, [directDir];
ld.global.f32 %f691, [directDir+4];
ld.global.f32 %f692, [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, %r337, %r338, %rd102, %rd103, %rd192, %rd192);
// inline asm
fma.rn.f32 %f693, %f690, 0fBF000000, 0f3F000000;
mul.f32 %f694, %f693, 0f437F0000;
cvt.rzi.u32.f32 %r296, %f694;
fma.rn.f32 %f695, %f691, 0fBF000000, 0f3F000000;
mul.f32 %f696, %f695, 0f437F0000;
cvt.rzi.u32.f32 %r297, %f696;
fma.rn.f32 %f697, %f692, 0fBF000000, 0f3F000000;
mul.f32 %f698, %f697, 0f437F0000;
cvt.rzi.u32.f32 %r298, %f698;
cvt.u16.u32 %rs33, %r298;
cvt.u16.u32 %rs34, %r297;
cvt.u16.u32 %rs35, %r296;
mov.u16 %rs36, 255;
st.v4.u8 [%rd100], {%rs35, %rs34, %rs33, %rs36};
BB0_133:
ret;
}