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

1866 lines
54 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 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 1 .b8 lightMeshBuffer[1];
.global .align 4 .u32 lightMeshBufferSize;
.global .align 4 .f32 lightInvCutoff;
.global .align 4 .f32 lightPointSize;
.global .align 4 .b8 lightColor[12];
.global .align 1 .b8 rnd_seeds[1];
.global .align 1 .b8 rnd_seeds2[1];
.global .align 4 .u32 samples;
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
.global .align 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum19lightMeshBufferSizeE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic19lightMeshBufferSizeE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation19lightMeshBufferSizeE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
.visible .entry _Z6oxMainv(
)
{
.local .align 16 .b8 __local_depot0[64];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<144>;
.reg .b16 %rs<47>;
.reg .f32 %f<922>;
.reg .b32 %r<363>;
.reg .b64 %rd<161>;
mov.u64 %rd160, __local_depot0;
cvta.local.u64 %SP, %rd160;
ld.global.u32 %r1, [lightMeshBufferSize];
abs.s32 %r2, %r1;
ld.global.u32 %r3, [samples];
ld.global.v2.u32 {%r105, %r106}, [pixelID];
cvt.u64.u32 %rd25, %r105;
cvt.u64.u32 %rd26, %r106;
mov.u64 %rd29, uvnormal;
cvta.global.u64 %rd24, %rd29;
mov.u32 %r103, 2;
mov.u32 %r104, 4;
mov.u64 %rd28, 0;
// inline asm
call (%rd23), _rt_buffer_get_64, (%rd24, %r103, %r104, %rd25, %rd26, %rd28, %rd28);
// inline asm
ld.u32 %r4, [%rd23];
shr.u32 %r109, %r4, 16;
cvt.u16.u32 %rs1, %r109;
and.b16 %rs3, %rs1, 255;
cvt.u16.u32 %rs4, %r4;
or.b16 %rs5, %rs4, %rs3;
setp.eq.s16 %p7, %rs5, 0;
mov.f32 %f857, 0f00000000;
mov.f32 %f858, %f857;
mov.f32 %f859, %f857;
@%p7 bra BB0_2;
ld.u8 %rs6, [%rd23+1];
and.b16 %rs8, %rs4, 255;
cvt.rn.f32.u16 %f204, %rs8;
div.rn.f32 %f205, %f204, 0f437F0000;
fma.rn.f32 %f206, %f205, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f207, %rs6;
div.rn.f32 %f208, %f207, 0f437F0000;
fma.rn.f32 %f209, %f208, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f210, %rs3;
div.rn.f32 %f211, %f210, 0f437F0000;
fma.rn.f32 %f212, %f211, 0f40000000, 0fBF800000;
mul.f32 %f213, %f209, %f209;
fma.rn.f32 %f214, %f206, %f206, %f213;
fma.rn.f32 %f215, %f212, %f212, %f214;
sqrt.rn.f32 %f216, %f215;
rcp.rn.f32 %f217, %f216;
mul.f32 %f857, %f206, %f217;
mul.f32 %f858, %f209, %f217;
mul.f32 %f859, %f212, %f217;
BB0_2:
ld.global.v2.u32 {%r110, %r111}, [pixelID];
ld.global.v2.u32 {%r113, %r114}, [tileInfo];
add.s32 %r5, %r110, %r113;
add.s32 %r6, %r111, %r114;
setp.eq.f32 %p8, %f858, 0f00000000;
setp.eq.f32 %p9, %f857, 0f00000000;
and.pred %p10, %p9, %p8;
setp.eq.f32 %p11, %f859, 0f00000000;
and.pred %p12, %p10, %p11;
@%p12 bra BB0_132;
bra.uni BB0_3;
BB0_132:
ld.global.u32 %r362, [imageEnabled];
and.b32 %r312, %r362, 1;
setp.eq.b32 %p139, %r312, 1;
@!%p139 bra BB0_134;
bra.uni BB0_133;
BB0_133:
cvt.u64.u32 %rd120, %r5;
cvt.u64.u32 %rd121, %r6;
mov.u64 %rd124, image;
cvta.global.u64 %rd119, %rd124;
mov.u64 %rd123, 0;
// inline asm
call (%rd118), _rt_buffer_get_64, (%rd119, %r103, %r104, %rd120, %rd121, %rd123, %rd123);
// inline asm
mov.u16 %rs31, 0;
st.v4.u8 [%rd118], {%rs31, %rs31, %rs31, %rs31};
ld.global.u32 %r362, [imageEnabled];
BB0_134:
and.b32 %r315, %r362, 8;
setp.eq.s32 %p140, %r315, 0;
@%p140 bra BB0_136;
cvt.u64.u32 %rd127, %r5;
cvt.u64.u32 %rd128, %r6;
mov.u64 %rd131, image_Mask;
cvta.global.u64 %rd126, %rd131;
mov.u64 %rd130, 0;
// inline asm
call (%rd125), _rt_buffer_get_64, (%rd126, %r103, %r103, %rd127, %rd128, %rd130, %rd130);
// inline asm
mov.f32 %f840, 0f00000000;
cvt.rzi.u32.f32 %r318, %f840;
cvt.u16.u32 %rs32, %r318;
mov.u16 %rs33, 0;
st.v2.u8 [%rd125], {%rs32, %rs33};
ld.global.u32 %r362, [imageEnabled];
BB0_136:
and.b32 %r319, %r362, 4;
setp.eq.s32 %p141, %r319, 0;
@%p141 bra BB0_140;
ld.global.u32 %r320, [additive];
setp.eq.s32 %p142, %r320, 0;
cvt.u64.u32 %rd21, %r5;
cvt.u64.u32 %rd22, %r6;
@%p142 bra BB0_139;
mov.u64 %rd144, image_HDR;
cvta.global.u64 %rd133, %rd144;
mov.u32 %r324, 8;
mov.u64 %rd143, 0;
// inline asm
call (%rd132), _rt_buffer_get_64, (%rd133, %r103, %r324, %rd21, %rd22, %rd143, %rd143);
// inline asm
ld.v4.u16 {%rs40, %rs41, %rs42, %rs43}, [%rd132];
// inline asm
{ cvt.f32.f16 %f841, %rs40;}
// inline asm
// inline asm
{ cvt.f32.f16 %f842, %rs41;}
// inline asm
// inline asm
{ cvt.f32.f16 %f843, %rs42;}
// inline asm
// inline asm
call (%rd138), _rt_buffer_get_64, (%rd133, %r103, %r324, %rd21, %rd22, %rd143, %rd143);
// inline asm
add.f32 %f844, %f841, 0f00000000;
add.f32 %f845, %f842, 0f00000000;
add.f32 %f846, %f843, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs39, %f846;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs38, %f845;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs37, %f844;}
// inline asm
mov.u16 %rs44, 0;
st.v4.u16 [%rd138], {%rs37, %rs38, %rs39, %rs44};
bra.uni BB0_140;
BB0_3:
ld.global.v2.u32 {%r123, %r124}, [pixelID];
cvt.u64.u32 %rd32, %r123;
cvt.u64.u32 %rd33, %r124;
mov.u64 %rd42, uvpos;
cvta.global.u64 %rd31, %rd42;
mov.u32 %r119, 12;
// inline asm
call (%rd30), _rt_buffer_get_64, (%rd31, %r103, %r119, %rd32, %rd33, %rd28, %rd28);
// inline asm
ld.f32 %f9, [%rd30+8];
ld.f32 %f8, [%rd30+4];
ld.f32 %f7, [%rd30];
mul.f32 %f222, %f7, 0f3456BF95;
mul.f32 %f223, %f8, 0f3456BF95;
mul.f32 %f224, %f9, 0f3456BF95;
abs.f32 %f10, %f857;
div.rn.f32 %f225, %f222, %f10;
abs.f32 %f226, %f858;
div.rn.f32 %f227, %f223, %f226;
abs.f32 %f11, %f859;
div.rn.f32 %f228, %f224, %f11;
abs.f32 %f229, %f225;
abs.f32 %f230, %f227;
abs.f32 %f231, %f228;
mov.f32 %f232, 0f38D1B717;
max.f32 %f233, %f229, %f232;
max.f32 %f234, %f230, %f232;
max.f32 %f235, %f231, %f232;
fma.rn.f32 %f12, %f857, %f233, %f7;
fma.rn.f32 %f13, %f858, %f234, %f8;
fma.rn.f32 %f14, %f859, %f235, %f9;
ld.global.v2.u32 {%r127, %r128}, [pixelID];
cvt.u64.u32 %rd38, %r127;
cvt.u64.u32 %rd39, %r128;
mov.u64 %rd43, rnd_seeds;
cvta.global.u64 %rd37, %rd43;
// inline asm
call (%rd36), _rt_buffer_get_64, (%rd37, %r103, %r104, %rd38, %rd39, %rd28, %rd28);
// inline asm
ld.u32 %r338, [%rd36];
mov.pred %p13, 0;
mov.f32 %f19, 0f00000000;
mov.u32 %r9, 0;
setp.lt.s32 %p14, %r2, 1;
@%p14 bra BB0_4;
ld.global.f32 %f15, [lightPointSize];
mul.f32 %f16, %f12, 0f3456BF95;
mul.f32 %f17, %f13, 0f3456BF95;
mul.f32 %f18, %f14, 0f3456BF95;
mov.f32 %f19, 0f00000000;
mov.u32 %r331, 0;
abs.f32 %f363, %f17;
abs.f32 %f364, %f16;
max.f32 %f365, %f364, %f363;
abs.f32 %f366, %f18;
max.f32 %f367, %f365, %f366;
mov.u32 %r9, %r331;
mov.f32 %f20, %f19;
mov.f32 %f21, %f19;
mov.f32 %f22, %f19;
BB0_6:
shl.b32 %r10, %r331, 1;
cvt.s64.s32 %rd46, %r10;
mov.u64 %rd50, lightMeshBuffer;
cvta.global.u64 %rd45, %rd50;
mov.u32 %r133, 1;
mov.u64 %rd49, 0;
// inline asm
call (%rd44), _rt_buffer_get_64, (%rd45, %r133, %r119, %rd46, %rd49, %rd49, %rd49);
// inline asm
ld.f32 %f240, [%rd44];
sub.f32 %f241, %f240, %f7;
ld.f32 %f242, [%rd44+4];
sub.f32 %f243, %f242, %f8;
ld.f32 %f244, [%rd44+8];
sub.f32 %f245, %f244, %f9;
mul.f32 %f246, %f243, %f243;
fma.rn.f32 %f247, %f241, %f241, %f246;
fma.rn.f32 %f248, %f245, %f245, %f247;
sqrt.rn.f32 %f23, %f248;
rcp.rn.f32 %f249, %f23;
mul.f32 %f24, %f241, %f249;
mul.f32 %f25, %f243, %f249;
mul.f32 %f26, %f245, %f249;
mul.f32 %f250, %f858, %f25;
fma.rn.f32 %f251, %f857, %f24, %f250;
fma.rn.f32 %f27, %f859, %f26, %f251;
setp.leu.f32 %p15, %f27, 0f00000000;
@%p15 bra BB0_23;
mov.u64 %rd155, 0;
setp.eq.s32 %p16, %r3, 0;
mul.f32 %f252, %f23, %f23;
div.rn.f32 %f28, %f15, %f252;
ld.global.v2.u32 {%r137, %r138}, [pixelID];
cvt.u64.u32 %rd57, %r138;
cvt.u64.u32 %rd58, %r137;
and.b64 %rd53, %rd58, 255;
and.b64 %rd54, %rd57, 255;
mov.u64 %rd59, rnd_seeds2;
cvta.global.u64 %rd52, %rd59;
mov.u32 %r136, 3;
// inline asm
call (%rd51), _rt_buffer_get_64, (%rd52, %r103, %r136, %rd53, %rd54, %rd155, %rd155);
// inline asm
@%p16 bra BB0_9;
ld.u8 %rs10, [%rd51];
cvt.rn.f32.u16 %f253, %rs10;
div.rn.f32 %f254, %f253, 0fC37F0000;
fma.rn.f32 %f255, %f254, 0f3F333333, 0f3F800000;
setp.gt.f32 %p18, %f28, %f255;
mov.pred %p143, -1;
@%p18 bra BB0_25;
BB0_9:
ld.global.f32 %f258, [lightInvCutoff];
mul.f32 %f29, %f23, %f258;
mov.f32 %f262, 0f40800000;
abs.f32 %f31, %f29;
setp.lt.f32 %p19, %f31, 0f00800000;
mul.f32 %f264, %f31, 0f4B800000;
selp.f32 %f265, 0fC3170000, 0fC2FE0000, %p19;
selp.f32 %f266, %f264, %f31, %p19;
mov.b32 %r141, %f266;
and.b32 %r142, %r141, 8388607;
or.b32 %r143, %r142, 1065353216;
mov.b32 %f267, %r143;
shr.u32 %r144, %r141, 23;
cvt.rn.f32.u32 %f268, %r144;
add.f32 %f269, %f265, %f268;
setp.gt.f32 %p20, %f267, 0f3FB504F3;
mul.f32 %f270, %f267, 0f3F000000;
add.f32 %f271, %f269, 0f3F800000;
selp.f32 %f272, %f270, %f267, %p20;
selp.f32 %f273, %f271, %f269, %p20;
add.f32 %f274, %f272, 0fBF800000;
add.f32 %f257, %f272, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f256,%f257;
// inline asm
add.f32 %f275, %f274, %f274;
mul.f32 %f276, %f256, %f275;
mul.f32 %f277, %f276, %f276;
mov.f32 %f278, 0f3C4CAF63;
mov.f32 %f279, 0f3B18F0FE;
fma.rn.f32 %f280, %f279, %f277, %f278;
mov.f32 %f281, 0f3DAAAABD;
fma.rn.f32 %f282, %f280, %f277, %f281;
mul.rn.f32 %f283, %f282, %f277;
mul.rn.f32 %f284, %f283, %f276;
sub.f32 %f285, %f274, %f276;
neg.f32 %f286, %f276;
add.f32 %f287, %f285, %f285;
fma.rn.f32 %f288, %f286, %f274, %f287;
mul.rn.f32 %f289, %f256, %f288;
add.f32 %f290, %f284, %f276;
sub.f32 %f291, %f276, %f290;
add.f32 %f292, %f284, %f291;
add.f32 %f293, %f289, %f292;
add.f32 %f294, %f290, %f293;
sub.f32 %f295, %f290, %f294;
add.f32 %f296, %f293, %f295;
mov.f32 %f297, 0f3F317200;
mul.rn.f32 %f298, %f273, %f297;
mov.f32 %f299, 0f35BFBE8E;
mul.rn.f32 %f300, %f273, %f299;
add.f32 %f301, %f298, %f294;
sub.f32 %f302, %f298, %f301;
add.f32 %f303, %f294, %f302;
add.f32 %f304, %f296, %f303;
add.f32 %f305, %f300, %f304;
add.f32 %f306, %f301, %f305;
sub.f32 %f307, %f301, %f306;
add.f32 %f308, %f305, %f307;
mul.rn.f32 %f309, %f262, %f306;
neg.f32 %f310, %f309;
fma.rn.f32 %f311, %f262, %f306, %f310;
fma.rn.f32 %f312, %f262, %f308, %f311;
mov.f32 %f313, 0f00000000;
fma.rn.f32 %f314, %f313, %f306, %f312;
add.rn.f32 %f315, %f309, %f314;
neg.f32 %f316, %f315;
add.rn.f32 %f317, %f309, %f316;
add.rn.f32 %f318, %f317, %f314;
mov.b32 %r145, %f315;
setp.eq.s32 %p21, %r145, 1118925336;
add.s32 %r146, %r145, -1;
mov.b32 %f319, %r146;
add.f32 %f320, %f318, 0f37000000;
selp.f32 %f321, %f319, %f315, %p21;
selp.f32 %f32, %f320, %f318, %p21;
mul.f32 %f322, %f321, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f323, %f322;
mov.f32 %f324, 0fBF317200;
fma.rn.f32 %f325, %f323, %f324, %f321;
mov.f32 %f326, 0fB5BFBE8E;
fma.rn.f32 %f327, %f323, %f326, %f325;
mul.f32 %f328, %f327, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f329, %f328;
add.f32 %f330, %f323, 0f00000000;
ex2.approx.f32 %f331, %f330;
mul.f32 %f332, %f329, %f331;
setp.lt.f32 %p22, %f321, 0fC2D20000;
selp.f32 %f333, 0f00000000, %f332, %p22;
setp.gt.f32 %p23, %f321, 0f42D20000;
selp.f32 %f864, 0f7F800000, %f333, %p23;
setp.eq.f32 %p24, %f864, 0f7F800000;
@%p24 bra BB0_11;
fma.rn.f32 %f864, %f864, %f32, %f864;
BB0_11:
mov.f32 %f853, 0f40000000;
cvt.rzi.f32.f32 %f852, %f853;
add.f32 %f851, %f852, %f852;
mov.f32 %f850, 0f40800000;
sub.f32 %f849, %f850, %f851;
abs.f32 %f848, %f849;
setp.lt.f32 %p25, %f29, 0f00000000;
setp.eq.f32 %p26, %f848, 0f3F800000;
and.pred %p1, %p25, %p26;
mov.b32 %r147, %f864;
xor.b32 %r148, %r147, -2147483648;
mov.b32 %f334, %r148;
selp.f32 %f866, %f334, %f864, %p1;
setp.eq.f32 %p27, %f29, 0f00000000;
@%p27 bra BB0_14;
bra.uni BB0_12;
BB0_14:
add.f32 %f337, %f29, %f29;
selp.f32 %f866, %f337, 0f00000000, %p26;
bra.uni BB0_15;
BB0_12:
setp.geu.f32 %p28, %f29, 0f00000000;
@%p28 bra BB0_15;
mov.f32 %f856, 0f40800000;
cvt.rzi.f32.f32 %f336, %f856;
setp.neu.f32 %p29, %f336, 0f40800000;
selp.f32 %f866, 0f7FFFFFFF, %f866, %p29;
BB0_15:
add.f32 %f338, %f31, 0f40800000;
mov.b32 %r149, %f338;
setp.lt.s32 %p31, %r149, 2139095040;
@%p31 bra BB0_20;
setp.gtu.f32 %p32, %f31, 0f7F800000;
@%p32 bra BB0_19;
bra.uni BB0_17;
BB0_19:
add.f32 %f866, %f29, 0f40800000;
bra.uni BB0_20;
BB0_17:
setp.neu.f32 %p33, %f31, 0f7F800000;
@%p33 bra BB0_20;
selp.f32 %f866, 0fFF800000, 0f7F800000, %p1;
BB0_20:
mov.u32 %r328, 1;
mov.u64 %rd154, lightMeshBuffer;
cvta.global.u64 %rd153, %rd154;
shl.b32 %r327, %r331, 1;
mov.u64 %rd152, 0;
mov.f32 %f339, 0f3F800000;
sub.f32 %f340, %f339, %f866;
setp.eq.f32 %p34, %f29, 0f3F800000;
selp.f32 %f341, 0f00000000, %f340, %p34;
cvt.sat.f32.f32 %f342, %f341;
mul.f32 %f343, %f28, %f342;
add.s32 %r152, %r327, 1;
cvt.s64.s32 %rd62, %r152;
// inline asm
call (%rd60), _rt_buffer_get_64, (%rd153, %r328, %r119, %rd62, %rd152, %rd152, %rd152);
// inline asm
ld.f32 %f344, [%rd60];
mul.f32 %f345, %f24, %f344;
ld.f32 %f346, [%rd60+4];
mul.f32 %f347, %f25, %f346;
neg.f32 %f348, %f347;
sub.f32 %f349, %f348, %f345;
ld.f32 %f350, [%rd60+8];
mul.f32 %f351, %f26, %f350;
sub.f32 %f352, %f349, %f351;
cvt.sat.f32.f32 %f353, %f352;
mul.f32 %f43, %f343, %f353;
cvt.sat.f32.f32 %f44, %f27;
mul.f32 %f354, %f43, %f44;
setp.leu.f32 %p35, %f354, 0f3727C5AC;
@%p35 bra BB0_22;
mov.u32 %r330, 1;
add.u64 %rd67, %SP, 28;
cvta.to.local.u64 %rd68, %rd67;
max.f32 %f361, %f367, %f232;
sub.f32 %f362, %f23, %f361;
mov.u32 %r156, 1065353216;
st.local.u32 [%rd68+8], %r156;
st.local.u32 [%rd68+4], %r156;
st.local.u32 [%rd68], %r156;
ld.global.u32 %r153, [root];
// inline asm
call _rt_trace_64, (%r153, %f12, %f13, %f14, %f24, %f25, %f26, %r330, %f361, %f362, %rd67, %r119);
// inline asm
ld.local.f32 %f369, [%rd68];
mul.f32 %f370, %f43, 0f3EA2F983;
mul.f32 %f371, %f370, %f369;
ld.local.f32 %f372, [%rd68+4];
mul.f32 %f373, %f370, %f372;
ld.local.f32 %f374, [%rd68+8];
mul.f32 %f375, %f370, %f374;
fma.rn.f32 %f22, %f44, %f371, %f22;
fma.rn.f32 %f21, %f44, %f373, %f21;
fma.rn.f32 %f20, %f44, %f375, %f20;
add.f32 %f19, %f19, %f369;
BB0_22:
add.s32 %r9, %r9, 1;
BB0_23:
add.s32 %r331, %r331, 1;
setp.lt.s32 %p37, %r331, %r2;
@%p37 bra BB0_6;
bra.uni BB0_24;
BB0_4:
mov.f32 %f20, %f19;
mov.f32 %f21, %f19;
mov.f32 %f22, %f19;
BB0_24:
mov.pred %p143, %p13;
BB0_25:
cvt.rn.f32.s32 %f376, %r9;
mov.f32 %f377, 0f3F800000;
max.f32 %f378, %f376, %f377;
rcp.rn.f32 %f379, %f378;
mul.f32 %f906, %f22, %f379;
mul.f32 %f907, %f21, %f379;
mul.f32 %f908, %f20, %f379;
div.rn.f32 %f909, %f19, %f378;
@!%p143 bra BB0_80;
bra.uni BB0_26;
BB0_26:
abs.f32 %f855, %f859;
abs.f32 %f854, %f857;
setp.gt.f32 %p38, %f854, %f855;
neg.f32 %f384, %f858;
selp.f32 %f385, %f384, 0f00000000, %p38;
neg.f32 %f386, %f859;
selp.f32 %f387, %f857, %f386, %p38;
selp.f32 %f388, 0f00000000, %f858, %p38;
mul.f32 %f389, %f387, %f387;
fma.rn.f32 %f390, %f385, %f385, %f389;
fma.rn.f32 %f391, %f388, %f388, %f390;
sqrt.rn.f32 %f392, %f391;
rcp.rn.f32 %f393, %f392;
mul.f32 %f65, %f385, %f393;
mul.f32 %f66, %f387, %f393;
mul.f32 %f67, %f388, %f393;
mov.f32 %f883, 0f00000000;
setp.lt.s32 %p39, %r3, 1;
mov.f32 %f884, %f883;
mov.f32 %f885, %f883;
mov.f32 %f886, %f883;
@%p39 bra BB0_79;
cvt.rn.f32.s32 %f398, %r3;
rcp.rn.f32 %f68, %f398;
add.u64 %rd69, %SP, 0;
cvta.to.local.u64 %rd3, %rd69;
mul.f32 %f69, %f12, 0f3456BF95;
mul.f32 %f70, %f13, 0f3456BF95;
mul.f32 %f71, %f14, 0f3456BF95;
add.u64 %rd70, %SP, 48;
cvta.to.local.u64 %rd4, %rd70;
mul.f32 %f399, %f857, %f66;
mul.f32 %f400, %f858, %f65;
sub.f32 %f72, %f400, %f399;
mul.f32 %f401, %f859, %f65;
mul.f32 %f402, %f857, %f67;
sub.f32 %f73, %f402, %f401;
mul.f32 %f403, %f858, %f67;
mul.f32 %f404, %f859, %f66;
sub.f32 %f74, %f404, %f403;
mov.f32 %f883, 0f00000000;
mov.u32 %r157, 0;
abs.f32 %f405, %f70;
abs.f32 %f406, %f69;
max.f32 %f407, %f406, %f405;
abs.f32 %f408, %f71;
max.f32 %f409, %f407, %f408;
mov.u32 %r335, %r157;
mov.f32 %f884, %f883;
mov.f32 %f885, %f883;
mov.f32 %f886, %f883;
BB0_28:
cvt.rn.f32.s32 %f79, %r335;
max.f32 %f80, %f409, %f232;
mov.u32 %r337, %r157;
BB0_29:
mad.lo.s32 %r159, %r338, 1664525, 1013904223;
and.b32 %r160, %r159, 16777215;
cvt.rn.f32.u32 %f411, %r160;
fma.rn.f32 %f412, %f411, 0f33800000, %f79;
mul.f32 %f413, %f68, %f412;
mad.lo.s32 %r338, %r159, 1664525, 1013904223;
and.b32 %r161, %r338, 16777215;
cvt.rn.f32.u32 %f414, %r161;
cvt.rn.f32.s32 %f415, %r337;
fma.rn.f32 %f416, %f414, 0f33800000, %f415;
mul.f32 %f417, %f68, %f416;
sqrt.rn.f32 %f85, %f413;
mul.f32 %f893, %f417, 0f40C90FDB;
abs.f32 %f87, %f893;
setp.neu.f32 %p40, %f87, 0f7F800000;
mov.f32 %f887, %f893;
@%p40 bra BB0_31;
mov.f32 %f418, 0f00000000;
mul.rn.f32 %f887, %f893, %f418;
BB0_31:
mul.f32 %f419, %f887, 0f3F22F983;
cvt.rni.s32.f32 %r348, %f419;
cvt.rn.f32.s32 %f420, %r348;
neg.f32 %f421, %f420;
mov.f32 %f422, 0f3FC90FDA;
fma.rn.f32 %f423, %f421, %f422, %f887;
mov.f32 %f424, 0f33A22168;
fma.rn.f32 %f425, %f421, %f424, %f423;
mov.f32 %f426, 0f27C234C5;
fma.rn.f32 %f888, %f421, %f426, %f425;
abs.f32 %f427, %f887;
setp.leu.f32 %p41, %f427, 0f47CE4780;
@%p41 bra BB0_42;
mov.b32 %r21, %f887;
shr.u32 %r22, %r21, 23;
shl.b32 %r164, %r21, 8;
or.b32 %r23, %r164, -2147483648;
mov.u32 %r339, 0;
mov.u64 %rd157, 0;
mov.u64 %rd156, %rd3;
mov.u32 %r340, %r339;
BB0_33:
.pragma "nounroll";
shl.b64 %rd72, %rd157, 2;
mov.u64 %rd73, __cudart_i2opi_f;
add.s64 %rd74, %rd73, %rd72;
ld.const.u32 %r167, [%rd74];
// inline asm
{
mad.lo.cc.u32 %r165, %r167, %r23, %r340;
madc.hi.u32 %r340, %r167, %r23, 0;
}
// inline asm
st.local.u32 [%rd156], %r165;
add.s32 %r339, %r339, 1;
cvt.s64.s32 %rd157, %r339;
mul.wide.s32 %rd75, %r339, 4;
add.s64 %rd156, %rd3, %rd75;
setp.ne.s32 %p42, %r339, 6;
@%p42 bra BB0_33;
and.b32 %r170, %r22, 255;
add.s32 %r171, %r170, -128;
shr.u32 %r172, %r171, 5;
and.b32 %r28, %r21, -2147483648;
st.local.u32 [%rd3+24], %r340;
mov.u32 %r173, 6;
sub.s32 %r174, %r173, %r172;
mul.wide.s32 %rd76, %r174, 4;
add.s64 %rd10, %rd3, %rd76;
ld.local.u32 %r341, [%rd10];
ld.local.u32 %r342, [%rd10+-4];
and.b32 %r31, %r22, 31;
setp.eq.s32 %p43, %r31, 0;
@%p43 bra BB0_36;
mov.u32 %r175, 32;
sub.s32 %r176, %r175, %r31;
shr.u32 %r177, %r342, %r176;
shl.b32 %r178, %r341, %r31;
add.s32 %r341, %r177, %r178;
ld.local.u32 %r179, [%rd10+-8];
shr.u32 %r180, %r179, %r176;
shl.b32 %r181, %r342, %r31;
add.s32 %r342, %r180, %r181;
BB0_36:
shr.u32 %r182, %r342, 30;
shl.b32 %r183, %r341, 2;
add.s32 %r343, %r182, %r183;
shl.b32 %r37, %r342, 2;
shr.u32 %r184, %r343, 31;
shr.u32 %r185, %r341, 30;
add.s32 %r38, %r184, %r185;
setp.eq.s32 %p44, %r184, 0;
@%p44 bra BB0_37;
bra.uni BB0_38;
BB0_37:
mov.u32 %r344, %r28;
mov.u32 %r345, %r37;
bra.uni BB0_39;
BB0_38:
not.b32 %r186, %r343;
neg.s32 %r345, %r37;
setp.eq.s32 %p45, %r37, 0;
selp.u32 %r187, 1, 0, %p45;
add.s32 %r343, %r187, %r186;
xor.b32 %r344, %r28, -2147483648;
BB0_39:
clz.b32 %r347, %r343;
setp.eq.s32 %p46, %r347, 0;
shl.b32 %r188, %r343, %r347;
mov.u32 %r189, 32;
sub.s32 %r190, %r189, %r347;
shr.u32 %r191, %r345, %r190;
add.s32 %r192, %r191, %r188;
selp.b32 %r46, %r343, %r192, %p46;
mov.u32 %r193, -921707870;
mul.hi.u32 %r346, %r46, %r193;
setp.eq.s32 %p47, %r28, 0;
neg.s32 %r194, %r38;
selp.b32 %r348, %r38, %r194, %p47;
setp.lt.s32 %p48, %r346, 1;
@%p48 bra BB0_41;
mul.lo.s32 %r195, %r46, -921707870;
shr.u32 %r196, %r195, 31;
shl.b32 %r197, %r346, 1;
add.s32 %r346, %r196, %r197;
add.s32 %r347, %r347, 1;
BB0_41:
mov.u32 %r198, 126;
sub.s32 %r199, %r198, %r347;
shl.b32 %r200, %r199, 23;
add.s32 %r201, %r346, 1;
shr.u32 %r202, %r201, 7;
add.s32 %r203, %r202, 1;
shr.u32 %r204, %r203, 1;
add.s32 %r205, %r204, %r200;
or.b32 %r206, %r205, %r344;
mov.b32 %f888, %r206;
BB0_42:
mul.rn.f32 %f93, %f888, %f888;
add.s32 %r54, %r348, 1;
and.b32 %r55, %r54, 1;
setp.eq.s32 %p49, %r55, 0;
@%p49 bra BB0_44;
bra.uni BB0_43;
BB0_44:
mov.f32 %f430, 0f3C08839E;
mov.f32 %f431, 0fB94CA1F9;
fma.rn.f32 %f889, %f431, %f93, %f430;
bra.uni BB0_45;
BB0_43:
mov.f32 %f428, 0fBAB6061A;
mov.f32 %f429, 0f37CCF5CE;
fma.rn.f32 %f889, %f429, %f93, %f428;
BB0_45:
@%p49 bra BB0_47;
bra.uni BB0_46;
BB0_47:
mov.f32 %f435, 0fBE2AAAA3;
fma.rn.f32 %f436, %f889, %f93, %f435;
mov.f32 %f437, 0f00000000;
fma.rn.f32 %f890, %f436, %f93, %f437;
bra.uni BB0_48;
BB0_46:
mov.f32 %f432, 0f3D2AAAA5;
fma.rn.f32 %f433, %f889, %f93, %f432;
mov.f32 %f434, 0fBF000000;
fma.rn.f32 %f890, %f433, %f93, %f434;
BB0_48:
fma.rn.f32 %f891, %f890, %f888, %f888;
@%p49 bra BB0_50;
fma.rn.f32 %f891, %f890, %f93, %f377;
BB0_50:
and.b32 %r207, %r54, 2;
setp.eq.s32 %p52, %r207, 0;
@%p52 bra BB0_52;
mov.f32 %f439, 0f00000000;
mov.f32 %f440, 0fBF800000;
fma.rn.f32 %f891, %f891, %f440, %f439;
BB0_52:
@%p40 bra BB0_54;
mov.f32 %f441, 0f00000000;
mul.rn.f32 %f893, %f893, %f441;
BB0_54:
mul.f32 %f442, %f893, 0f3F22F983;
cvt.rni.s32.f32 %r358, %f442;
cvt.rn.f32.s32 %f443, %r358;
neg.f32 %f444, %f443;
fma.rn.f32 %f446, %f444, %f422, %f893;
fma.rn.f32 %f448, %f444, %f424, %f446;
fma.rn.f32 %f894, %f444, %f426, %f448;
abs.f32 %f450, %f893;
setp.leu.f32 %p54, %f450, 0f47CE4780;
@%p54 bra BB0_65;
mov.b32 %r57, %f893;
shr.u32 %r58, %r57, 23;
shl.b32 %r210, %r57, 8;
or.b32 %r59, %r210, -2147483648;
mov.u32 %r349, 0;
mov.u64 %rd158, %rd3;
mov.u64 %rd159, %rd28;
mov.u32 %r350, %r349;
BB0_56:
.pragma "nounroll";
shl.b64 %rd78, %rd159, 2;
mov.u64 %rd79, __cudart_i2opi_f;
add.s64 %rd80, %rd79, %rd78;
ld.const.u32 %r213, [%rd80];
// inline asm
{
mad.lo.cc.u32 %r211, %r213, %r59, %r350;
madc.hi.u32 %r350, %r213, %r59, 0;
}
// inline asm
st.local.u32 [%rd158], %r211;
add.s32 %r349, %r349, 1;
cvt.s64.s32 %rd159, %r349;
mul.wide.s32 %rd81, %r349, 4;
add.s64 %rd158, %rd3, %rd81;
setp.ne.s32 %p55, %r349, 6;
@%p55 bra BB0_56;
and.b32 %r216, %r58, 255;
add.s32 %r217, %r216, -128;
shr.u32 %r218, %r217, 5;
and.b32 %r64, %r57, -2147483648;
st.local.u32 [%rd3+24], %r350;
mov.u32 %r219, 6;
sub.s32 %r220, %r219, %r218;
mul.wide.s32 %rd82, %r220, 4;
add.s64 %rd16, %rd3, %rd82;
ld.local.u32 %r351, [%rd16];
ld.local.u32 %r352, [%rd16+-4];
and.b32 %r67, %r58, 31;
setp.eq.s32 %p56, %r67, 0;
@%p56 bra BB0_59;
mov.u32 %r221, 32;
sub.s32 %r222, %r221, %r67;
shr.u32 %r223, %r352, %r222;
shl.b32 %r224, %r351, %r67;
add.s32 %r351, %r223, %r224;
ld.local.u32 %r225, [%rd16+-8];
shr.u32 %r226, %r225, %r222;
shl.b32 %r227, %r352, %r67;
add.s32 %r352, %r226, %r227;
BB0_59:
shr.u32 %r228, %r352, 30;
shl.b32 %r229, %r351, 2;
add.s32 %r353, %r228, %r229;
shl.b32 %r73, %r352, 2;
shr.u32 %r230, %r353, 31;
shr.u32 %r231, %r351, 30;
add.s32 %r74, %r230, %r231;
setp.eq.s32 %p57, %r230, 0;
@%p57 bra BB0_60;
bra.uni BB0_61;
BB0_60:
mov.u32 %r354, %r64;
mov.u32 %r355, %r73;
bra.uni BB0_62;
BB0_61:
not.b32 %r232, %r353;
neg.s32 %r355, %r73;
setp.eq.s32 %p58, %r73, 0;
selp.u32 %r233, 1, 0, %p58;
add.s32 %r353, %r233, %r232;
xor.b32 %r354, %r64, -2147483648;
BB0_62:
clz.b32 %r357, %r353;
setp.eq.s32 %p59, %r357, 0;
shl.b32 %r234, %r353, %r357;
mov.u32 %r235, 32;
sub.s32 %r236, %r235, %r357;
shr.u32 %r237, %r355, %r236;
add.s32 %r238, %r237, %r234;
selp.b32 %r82, %r353, %r238, %p59;
mov.u32 %r239, -921707870;
mul.hi.u32 %r356, %r82, %r239;
setp.eq.s32 %p60, %r64, 0;
neg.s32 %r240, %r74;
selp.b32 %r358, %r74, %r240, %p60;
setp.lt.s32 %p61, %r356, 1;
@%p61 bra BB0_64;
mul.lo.s32 %r241, %r82, -921707870;
shr.u32 %r242, %r241, 31;
shl.b32 %r243, %r356, 1;
add.s32 %r356, %r242, %r243;
add.s32 %r357, %r357, 1;
BB0_64:
mov.u32 %r244, 126;
sub.s32 %r245, %r244, %r357;
shl.b32 %r246, %r245, 23;
add.s32 %r247, %r356, 1;
shr.u32 %r248, %r247, 7;
add.s32 %r249, %r248, 1;
shr.u32 %r250, %r249, 1;
add.s32 %r251, %r250, %r246;
or.b32 %r252, %r251, %r354;
mov.b32 %f894, %r252;
BB0_65:
mul.rn.f32 %f110, %f894, %f894;
and.b32 %r90, %r358, 1;
setp.eq.s32 %p62, %r90, 0;
@%p62 bra BB0_67;
bra.uni BB0_66;
BB0_67:
mov.f32 %f453, 0f3C08839E;
mov.f32 %f454, 0fB94CA1F9;
fma.rn.f32 %f895, %f454, %f110, %f453;
bra.uni BB0_68;
BB0_66:
mov.f32 %f451, 0fBAB6061A;
mov.f32 %f452, 0f37CCF5CE;
fma.rn.f32 %f895, %f452, %f110, %f451;
BB0_68:
@%p62 bra BB0_70;
bra.uni BB0_69;
BB0_70:
mov.f32 %f458, 0fBE2AAAA3;
fma.rn.f32 %f459, %f895, %f110, %f458;
mov.f32 %f460, 0f00000000;
fma.rn.f32 %f896, %f459, %f110, %f460;
bra.uni BB0_71;
BB0_69:
mov.f32 %f455, 0f3D2AAAA5;
fma.rn.f32 %f456, %f895, %f110, %f455;
mov.f32 %f457, 0fBF000000;
fma.rn.f32 %f896, %f456, %f110, %f457;
BB0_71:
fma.rn.f32 %f897, %f896, %f894, %f894;
@%p62 bra BB0_73;
fma.rn.f32 %f897, %f896, %f110, %f377;
BB0_73:
and.b32 %r253, %r358, 2;
setp.eq.s32 %p65, %r253, 0;
@%p65 bra BB0_75;
mov.f32 %f462, 0f00000000;
mov.f32 %f463, 0fBF800000;
fma.rn.f32 %f897, %f897, %f463, %f462;
BB0_75:
mul.f32 %f475, %f85, %f891;
mul.f32 %f476, %f475, %f475;
sub.f32 %f478, %f377, %f476;
mul.f32 %f479, %f85, %f897;
mul.f32 %f480, %f479, %f479;
sub.f32 %f481, %f478, %f480;
mov.f32 %f899, 0f00000000;
max.f32 %f482, %f899, %f481;
sqrt.rn.f32 %f483, %f482;
mul.f32 %f484, %f65, %f479;
mul.f32 %f485, %f66, %f479;
mul.f32 %f486, %f67, %f479;
fma.rn.f32 %f487, %f74, %f475, %f484;
fma.rn.f32 %f488, %f73, %f475, %f485;
fma.rn.f32 %f489, %f72, %f475, %f486;
fma.rn.f32 %f467, %f857, %f483, %f487;
fma.rn.f32 %f468, %f858, %f483, %f488;
fma.rn.f32 %f469, %f859, %f483, %f489;
mov.f32 %f490, 0f7F7FFFFF;
st.local.v4.f32 [%rd4], {%f377, %f377, %f377, %f490};
ld.global.u32 %r254, [root];
mov.u32 %r255, 0;
mov.f32 %f471, 0f6C4ECB8F;
mov.u32 %r256, 16;
// inline asm
call _rt_trace_64, (%r254, %f12, %f13, %f14, %f467, %f468, %f469, %r255, %f80, %f471, %rd70, %r256);
// inline asm
ld.local.f32 %f122, [%rd4+12];
setp.ge.f32 %p66, %f122, 0f00000000;
mov.f32 %f900, %f899;
mov.f32 %f901, %f899;
@%p66 bra BB0_77;
ld.local.v4.f32 {%f901, %f900, %f899, %f494}, [%rd4];
BB0_77:
setp.ltu.f32 %p67, %f122, 0f00000000;
selp.f32 %f495, 0f3F800000, 0f00000000, %p67;
add.f32 %f883, %f883, %f495;
add.f32 %f886, %f886, %f901;
add.f32 %f885, %f885, %f900;
add.f32 %f884, %f884, %f899;
add.s32 %r337, %r337, 1;
setp.lt.s32 %p68, %r337, %r3;
@%p68 bra BB0_29;
add.s32 %r335, %r335, 1;
setp.lt.s32 %p69, %r335, %r3;
@%p69 bra BB0_28;
BB0_79:
mul.lo.s32 %r257, %r3, %r3;
cvt.rn.f32.s32 %f496, %r257;
rcp.rn.f32 %f497, %f496;
mul.f32 %f906, %f886, %f497;
mul.f32 %f907, %f885, %f497;
mul.f32 %f908, %f884, %f497;
div.rn.f32 %f909, %f883, %f496;
BB0_80:
setp.gt.s32 %p70, %r1, -1;
selp.f32 %f145, %f909, %f906, %p70;
ld.global.u32 %r360, [imageEnabled];
and.b32 %r258, %r360, 8;
setp.eq.s32 %p71, %r258, 0;
@%p71 bra BB0_93;
cvt.u64.u32 %rd86, %r5;
cvt.u64.u32 %rd87, %r6;
mov.u64 %rd90, image_Mask;
cvta.global.u64 %rd85, %rd90;
// inline asm
call (%rd84), _rt_buffer_get_64, (%rd85, %r103, %r103, %rd86, %rd87, %rd28, %rd28);
// inline asm
mov.f32 %f500, 0f3E68BA2E;
cvt.rzi.f32.f32 %f501, %f500;
fma.rn.f32 %f502, %f501, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f146, %f502;
abs.f32 %f147, %f145;
setp.lt.f32 %p72, %f147, 0f00800000;
mul.f32 %f503, %f147, 0f4B800000;
selp.f32 %f504, 0fC3170000, 0fC2FE0000, %p72;
selp.f32 %f505, %f503, %f147, %p72;
mov.b32 %r261, %f505;
and.b32 %r262, %r261, 8388607;
or.b32 %r263, %r262, 1065353216;
mov.b32 %f506, %r263;
shr.u32 %r264, %r261, 23;
cvt.rn.f32.u32 %f507, %r264;
add.f32 %f508, %f504, %f507;
setp.gt.f32 %p73, %f506, 0f3FB504F3;
mul.f32 %f509, %f506, 0f3F000000;
add.f32 %f510, %f508, 0f3F800000;
selp.f32 %f511, %f509, %f506, %p73;
selp.f32 %f512, %f510, %f508, %p73;
add.f32 %f513, %f511, 0fBF800000;
add.f32 %f499, %f511, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f498,%f499;
// inline asm
add.f32 %f514, %f513, %f513;
mul.f32 %f515, %f498, %f514;
mul.f32 %f516, %f515, %f515;
mov.f32 %f517, 0f3C4CAF63;
mov.f32 %f518, 0f3B18F0FE;
fma.rn.f32 %f519, %f518, %f516, %f517;
mov.f32 %f520, 0f3DAAAABD;
fma.rn.f32 %f521, %f519, %f516, %f520;
mul.rn.f32 %f522, %f521, %f516;
mul.rn.f32 %f523, %f522, %f515;
sub.f32 %f524, %f513, %f515;
neg.f32 %f525, %f515;
add.f32 %f526, %f524, %f524;
fma.rn.f32 %f527, %f525, %f513, %f526;
mul.rn.f32 %f528, %f498, %f527;
add.f32 %f529, %f523, %f515;
sub.f32 %f530, %f515, %f529;
add.f32 %f531, %f523, %f530;
add.f32 %f532, %f528, %f531;
add.f32 %f533, %f529, %f532;
sub.f32 %f534, %f529, %f533;
add.f32 %f535, %f532, %f534;
mov.f32 %f536, 0f3F317200;
mul.rn.f32 %f537, %f512, %f536;
mov.f32 %f538, 0f35BFBE8E;
mul.rn.f32 %f539, %f512, %f538;
add.f32 %f540, %f537, %f533;
sub.f32 %f541, %f537, %f540;
add.f32 %f542, %f533, %f541;
add.f32 %f543, %f535, %f542;
add.f32 %f544, %f539, %f543;
add.f32 %f545, %f540, %f544;
sub.f32 %f546, %f540, %f545;
add.f32 %f547, %f544, %f546;
mov.f32 %f548, 0f3EE8BA2E;
mul.rn.f32 %f549, %f548, %f545;
neg.f32 %f550, %f549;
fma.rn.f32 %f551, %f548, %f545, %f550;
fma.rn.f32 %f552, %f548, %f547, %f551;
mov.f32 %f553, 0f00000000;
fma.rn.f32 %f554, %f553, %f545, %f552;
add.rn.f32 %f555, %f549, %f554;
neg.f32 %f556, %f555;
add.rn.f32 %f557, %f549, %f556;
add.rn.f32 %f558, %f557, %f554;
mov.b32 %r265, %f555;
setp.eq.s32 %p74, %r265, 1118925336;
add.s32 %r266, %r265, -1;
mov.b32 %f559, %r266;
add.f32 %f560, %f558, 0f37000000;
selp.f32 %f561, %f559, %f555, %p74;
selp.f32 %f148, %f560, %f558, %p74;
mul.f32 %f562, %f561, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f563, %f562;
mov.f32 %f564, 0fBF317200;
fma.rn.f32 %f565, %f563, %f564, %f561;
mov.f32 %f566, 0fB5BFBE8E;
fma.rn.f32 %f567, %f563, %f566, %f565;
mul.f32 %f568, %f567, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f569, %f568;
add.f32 %f570, %f563, 0f00000000;
ex2.approx.f32 %f571, %f570;
mul.f32 %f572, %f569, %f571;
setp.lt.f32 %p75, %f561, 0fC2D20000;
selp.f32 %f573, 0f00000000, %f572, %p75;
setp.gt.f32 %p76, %f561, 0f42D20000;
selp.f32 %f910, 0f7F800000, %f573, %p76;
setp.eq.f32 %p77, %f910, 0f7F800000;
@%p77 bra BB0_83;
fma.rn.f32 %f910, %f910, %f148, %f910;
BB0_83:
setp.lt.f32 %p78, %f145, 0f00000000;
setp.eq.f32 %p79, %f146, 0f3F800000;
and.pred %p3, %p78, %p79;
mov.b32 %r267, %f910;
xor.b32 %r268, %r267, -2147483648;
mov.b32 %f574, %r268;
selp.f32 %f912, %f574, %f910, %p3;
setp.eq.f32 %p80, %f145, 0f00000000;
@%p80 bra BB0_86;
bra.uni BB0_84;
BB0_86:
add.f32 %f577, %f145, %f145;
selp.f32 %f912, %f577, 0f00000000, %p79;
bra.uni BB0_87;
BB0_139:
mov.u64 %rd151, image_HDR;
cvta.global.u64 %rd146, %rd151;
mov.u32 %r326, 8;
mov.u64 %rd150, 0;
// inline asm
call (%rd145), _rt_buffer_get_64, (%rd146, %r103, %r326, %rd21, %rd22, %rd150, %rd150);
// inline asm
mov.f32 %f847, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs45, %f847;}
// inline asm
mov.u16 %rs46, 0;
st.v4.u16 [%rd145], {%rs45, %rs45, %rs45, %rs46};
bra.uni BB0_140;
BB0_84:
setp.geu.f32 %p81, %f145, 0f00000000;
@%p81 bra BB0_87;
cvt.rzi.f32.f32 %f576, %f548;
setp.neu.f32 %p82, %f576, 0f3EE8BA2E;
selp.f32 %f912, 0f7FFFFFFF, %f912, %p82;
BB0_87:
add.f32 %f578, %f147, 0f3EE8BA2E;
mov.b32 %r269, %f578;
setp.lt.s32 %p84, %r269, 2139095040;
@%p84 bra BB0_92;
setp.gtu.f32 %p85, %f147, 0f7F800000;
@%p85 bra BB0_91;
bra.uni BB0_89;
BB0_91:
add.f32 %f912, %f145, 0f3EE8BA2E;
bra.uni BB0_92;
BB0_89:
setp.neu.f32 %p86, %f147, 0f7F800000;
@%p86 bra BB0_92;
selp.f32 %f912, 0fFF800000, 0f7F800000, %p3;
BB0_92:
mul.f32 %f579, %f912, 0f437F0000;
setp.eq.f32 %p87, %f145, 0f3F800000;
selp.f32 %f580, 0f437F0000, %f579, %p87;
cvt.rzi.u32.f32 %r270, %f580;
cvt.u16.u32 %rs11, %r270;
mov.u16 %rs12, 255;
st.v2.u8 [%rd84], {%rs11, %rs12};
ld.global.u32 %r360, [imageEnabled];
BB0_93:
ld.global.f32 %f581, [lightColor];
mul.f32 %f159, %f906, %f581;
ld.global.f32 %f582, [lightColor+4];
mul.f32 %f160, %f907, %f582;
ld.global.f32 %f583, [lightColor+8];
mul.f32 %f161, %f908, %f583;
and.b32 %r271, %r360, 1;
setp.eq.b32 %p88, %r271, 1;
@!%p88 bra BB0_128;
bra.uni BB0_94;
BB0_94:
mov.f32 %f586, 0f3E666666;
cvt.rzi.f32.f32 %f587, %f586;
fma.rn.f32 %f588, %f587, 0fC0000000, 0f3EE66666;
abs.f32 %f162, %f588;
abs.f32 %f163, %f159;
setp.lt.f32 %p89, %f163, 0f00800000;
mul.f32 %f589, %f163, 0f4B800000;
selp.f32 %f590, 0fC3170000, 0fC2FE0000, %p89;
selp.f32 %f591, %f589, %f163, %p89;
mov.b32 %r272, %f591;
and.b32 %r273, %r272, 8388607;
or.b32 %r274, %r273, 1065353216;
mov.b32 %f592, %r274;
shr.u32 %r275, %r272, 23;
cvt.rn.f32.u32 %f593, %r275;
add.f32 %f594, %f590, %f593;
setp.gt.f32 %p90, %f592, 0f3FB504F3;
mul.f32 %f595, %f592, 0f3F000000;
add.f32 %f596, %f594, 0f3F800000;
selp.f32 %f597, %f595, %f592, %p90;
selp.f32 %f598, %f596, %f594, %p90;
add.f32 %f599, %f597, 0fBF800000;
add.f32 %f585, %f597, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f584,%f585;
// inline asm
add.f32 %f600, %f599, %f599;
mul.f32 %f601, %f584, %f600;
mul.f32 %f602, %f601, %f601;
mov.f32 %f603, 0f3C4CAF63;
mov.f32 %f604, 0f3B18F0FE;
fma.rn.f32 %f605, %f604, %f602, %f603;
mov.f32 %f606, 0f3DAAAABD;
fma.rn.f32 %f607, %f605, %f602, %f606;
mul.rn.f32 %f608, %f607, %f602;
mul.rn.f32 %f609, %f608, %f601;
sub.f32 %f610, %f599, %f601;
neg.f32 %f611, %f601;
add.f32 %f612, %f610, %f610;
fma.rn.f32 %f613, %f611, %f599, %f612;
mul.rn.f32 %f614, %f584, %f613;
add.f32 %f615, %f609, %f601;
sub.f32 %f616, %f601, %f615;
add.f32 %f617, %f609, %f616;
add.f32 %f618, %f614, %f617;
add.f32 %f619, %f615, %f618;
sub.f32 %f620, %f615, %f619;
add.f32 %f621, %f618, %f620;
mov.f32 %f622, 0f3F317200;
mul.rn.f32 %f623, %f598, %f622;
mov.f32 %f624, 0f35BFBE8E;
mul.rn.f32 %f625, %f598, %f624;
add.f32 %f626, %f623, %f619;
sub.f32 %f627, %f623, %f626;
add.f32 %f628, %f619, %f627;
add.f32 %f629, %f621, %f628;
add.f32 %f630, %f625, %f629;
add.f32 %f631, %f626, %f630;
sub.f32 %f632, %f626, %f631;
add.f32 %f633, %f630, %f632;
mov.f32 %f634, 0f3EE66666;
mul.rn.f32 %f635, %f634, %f631;
neg.f32 %f636, %f635;
fma.rn.f32 %f637, %f634, %f631, %f636;
fma.rn.f32 %f638, %f634, %f633, %f637;
mov.f32 %f639, 0f00000000;
fma.rn.f32 %f640, %f639, %f631, %f638;
add.rn.f32 %f641, %f635, %f640;
neg.f32 %f642, %f641;
add.rn.f32 %f643, %f635, %f642;
add.rn.f32 %f644, %f643, %f640;
mov.b32 %r276, %f641;
setp.eq.s32 %p91, %r276, 1118925336;
add.s32 %r277, %r276, -1;
mov.b32 %f645, %r277;
add.f32 %f646, %f644, 0f37000000;
selp.f32 %f647, %f645, %f641, %p91;
selp.f32 %f164, %f646, %f644, %p91;
mul.f32 %f648, %f647, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f649, %f648;
mov.f32 %f650, 0fBF317200;
fma.rn.f32 %f651, %f649, %f650, %f647;
mov.f32 %f652, 0fB5BFBE8E;
fma.rn.f32 %f653, %f649, %f652, %f651;
mul.f32 %f654, %f653, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f655, %f654;
add.f32 %f656, %f649, 0f00000000;
ex2.approx.f32 %f657, %f656;
mul.f32 %f658, %f655, %f657;
setp.lt.f32 %p92, %f647, 0fC2D20000;
selp.f32 %f659, 0f00000000, %f658, %p92;
setp.gt.f32 %p93, %f647, 0f42D20000;
selp.f32 %f913, 0f7F800000, %f659, %p93;
setp.eq.f32 %p94, %f913, 0f7F800000;
@%p94 bra BB0_96;
fma.rn.f32 %f913, %f913, %f164, %f913;
BB0_96:
setp.lt.f32 %p95, %f159, 0f00000000;
setp.eq.f32 %p96, %f162, 0f3F800000;
and.pred %p4, %p95, %p96;
mov.b32 %r278, %f913;
xor.b32 %r279, %r278, -2147483648;
mov.b32 %f660, %r279;
selp.f32 %f915, %f660, %f913, %p4;
setp.eq.f32 %p97, %f159, 0f00000000;
@%p97 bra BB0_99;
bra.uni BB0_97;
BB0_99:
add.f32 %f663, %f159, %f159;
selp.f32 %f915, %f663, 0f00000000, %p96;
bra.uni BB0_100;
BB0_97:
setp.geu.f32 %p98, %f159, 0f00000000;
@%p98 bra BB0_100;
cvt.rzi.f32.f32 %f662, %f634;
setp.neu.f32 %p99, %f662, 0f3EE66666;
selp.f32 %f915, 0f7FFFFFFF, %f915, %p99;
BB0_100:
add.f32 %f664, %f163, 0f3EE66666;
mov.b32 %r280, %f664;
setp.lt.s32 %p101, %r280, 2139095040;
@%p101 bra BB0_105;
setp.gtu.f32 %p102, %f163, 0f7F800000;
@%p102 bra BB0_104;
bra.uni BB0_102;
BB0_104:
add.f32 %f915, %f159, 0f3EE66666;
bra.uni BB0_105;
BB0_102:
setp.neu.f32 %p103, %f163, 0f7F800000;
@%p103 bra BB0_105;
selp.f32 %f915, 0fFF800000, 0f7F800000, %p4;
BB0_105:
setp.eq.f32 %p104, %f159, 0f3F800000;
selp.f32 %f175, 0f3F800000, %f915, %p104;
abs.f32 %f176, %f160;
setp.lt.f32 %p105, %f176, 0f00800000;
mul.f32 %f667, %f176, 0f4B800000;
selp.f32 %f668, 0fC3170000, 0fC2FE0000, %p105;
selp.f32 %f669, %f667, %f176, %p105;
mov.b32 %r281, %f669;
and.b32 %r282, %r281, 8388607;
or.b32 %r283, %r282, 1065353216;
mov.b32 %f670, %r283;
shr.u32 %r284, %r281, 23;
cvt.rn.f32.u32 %f671, %r284;
add.f32 %f672, %f668, %f671;
setp.gt.f32 %p106, %f670, 0f3FB504F3;
mul.f32 %f673, %f670, 0f3F000000;
add.f32 %f674, %f672, 0f3F800000;
selp.f32 %f675, %f673, %f670, %p106;
selp.f32 %f676, %f674, %f672, %p106;
add.f32 %f677, %f675, 0fBF800000;
add.f32 %f666, %f675, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f665,%f666;
// inline asm
add.f32 %f678, %f677, %f677;
mul.f32 %f679, %f665, %f678;
mul.f32 %f680, %f679, %f679;
fma.rn.f32 %f683, %f604, %f680, %f603;
fma.rn.f32 %f685, %f683, %f680, %f606;
mul.rn.f32 %f686, %f685, %f680;
mul.rn.f32 %f687, %f686, %f679;
sub.f32 %f688, %f677, %f679;
neg.f32 %f689, %f679;
add.f32 %f690, %f688, %f688;
fma.rn.f32 %f691, %f689, %f677, %f690;
mul.rn.f32 %f692, %f665, %f691;
add.f32 %f693, %f687, %f679;
sub.f32 %f694, %f679, %f693;
add.f32 %f695, %f687, %f694;
add.f32 %f696, %f692, %f695;
add.f32 %f697, %f693, %f696;
sub.f32 %f698, %f693, %f697;
add.f32 %f699, %f696, %f698;
mul.rn.f32 %f701, %f676, %f622;
mul.rn.f32 %f703, %f676, %f624;
add.f32 %f704, %f701, %f697;
sub.f32 %f705, %f701, %f704;
add.f32 %f706, %f697, %f705;
add.f32 %f707, %f699, %f706;
add.f32 %f708, %f703, %f707;
add.f32 %f709, %f704, %f708;
sub.f32 %f710, %f704, %f709;
add.f32 %f711, %f708, %f710;
mul.rn.f32 %f713, %f634, %f709;
neg.f32 %f714, %f713;
fma.rn.f32 %f715, %f634, %f709, %f714;
fma.rn.f32 %f716, %f634, %f711, %f715;
fma.rn.f32 %f718, %f639, %f709, %f716;
add.rn.f32 %f719, %f713, %f718;
neg.f32 %f720, %f719;
add.rn.f32 %f721, %f713, %f720;
add.rn.f32 %f722, %f721, %f718;
mov.b32 %r285, %f719;
setp.eq.s32 %p107, %r285, 1118925336;
add.s32 %r286, %r285, -1;
mov.b32 %f723, %r286;
add.f32 %f724, %f722, 0f37000000;
selp.f32 %f725, %f723, %f719, %p107;
selp.f32 %f177, %f724, %f722, %p107;
mul.f32 %f726, %f725, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f727, %f726;
fma.rn.f32 %f729, %f727, %f650, %f725;
fma.rn.f32 %f731, %f727, %f652, %f729;
mul.f32 %f732, %f731, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f733, %f732;
add.f32 %f734, %f727, 0f00000000;
ex2.approx.f32 %f735, %f734;
mul.f32 %f736, %f733, %f735;
setp.lt.f32 %p108, %f725, 0fC2D20000;
selp.f32 %f737, 0f00000000, %f736, %p108;
setp.gt.f32 %p109, %f725, 0f42D20000;
selp.f32 %f916, 0f7F800000, %f737, %p109;
setp.eq.f32 %p110, %f916, 0f7F800000;
@%p110 bra BB0_107;
fma.rn.f32 %f916, %f916, %f177, %f916;
BB0_107:
setp.lt.f32 %p111, %f160, 0f00000000;
and.pred %p5, %p111, %p96;
mov.b32 %r287, %f916;
xor.b32 %r288, %r287, -2147483648;
mov.b32 %f738, %r288;
selp.f32 %f918, %f738, %f916, %p5;
setp.eq.f32 %p113, %f160, 0f00000000;
@%p113 bra BB0_110;
bra.uni BB0_108;
BB0_110:
add.f32 %f741, %f160, %f160;
selp.f32 %f918, %f741, 0f00000000, %p96;
bra.uni BB0_111;
BB0_108:
setp.geu.f32 %p114, %f160, 0f00000000;
@%p114 bra BB0_111;
cvt.rzi.f32.f32 %f740, %f634;
setp.neu.f32 %p115, %f740, 0f3EE66666;
selp.f32 %f918, 0f7FFFFFFF, %f918, %p115;
BB0_111:
add.f32 %f742, %f176, 0f3EE66666;
mov.b32 %r289, %f742;
setp.lt.s32 %p117, %r289, 2139095040;
@%p117 bra BB0_116;
setp.gtu.f32 %p118, %f176, 0f7F800000;
@%p118 bra BB0_115;
bra.uni BB0_113;
BB0_115:
add.f32 %f918, %f160, 0f3EE66666;
bra.uni BB0_116;
BB0_113:
setp.neu.f32 %p119, %f176, 0f7F800000;
@%p119 bra BB0_116;
selp.f32 %f918, 0fFF800000, 0f7F800000, %p5;
BB0_116:
setp.eq.f32 %p120, %f160, 0f3F800000;
selp.f32 %f188, 0f3F800000, %f918, %p120;
abs.f32 %f189, %f161;
setp.lt.f32 %p121, %f189, 0f00800000;
mul.f32 %f745, %f189, 0f4B800000;
selp.f32 %f746, 0fC3170000, 0fC2FE0000, %p121;
selp.f32 %f747, %f745, %f189, %p121;
mov.b32 %r290, %f747;
and.b32 %r291, %r290, 8388607;
or.b32 %r292, %r291, 1065353216;
mov.b32 %f748, %r292;
shr.u32 %r293, %r290, 23;
cvt.rn.f32.u32 %f749, %r293;
add.f32 %f750, %f746, %f749;
setp.gt.f32 %p122, %f748, 0f3FB504F3;
mul.f32 %f751, %f748, 0f3F000000;
add.f32 %f752, %f750, 0f3F800000;
selp.f32 %f753, %f751, %f748, %p122;
selp.f32 %f754, %f752, %f750, %p122;
add.f32 %f755, %f753, 0fBF800000;
add.f32 %f744, %f753, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f743,%f744;
// inline asm
add.f32 %f756, %f755, %f755;
mul.f32 %f757, %f743, %f756;
mul.f32 %f758, %f757, %f757;
fma.rn.f32 %f761, %f604, %f758, %f603;
fma.rn.f32 %f763, %f761, %f758, %f606;
mul.rn.f32 %f764, %f763, %f758;
mul.rn.f32 %f765, %f764, %f757;
sub.f32 %f766, %f755, %f757;
neg.f32 %f767, %f757;
add.f32 %f768, %f766, %f766;
fma.rn.f32 %f769, %f767, %f755, %f768;
mul.rn.f32 %f770, %f743, %f769;
add.f32 %f771, %f765, %f757;
sub.f32 %f772, %f757, %f771;
add.f32 %f773, %f765, %f772;
add.f32 %f774, %f770, %f773;
add.f32 %f775, %f771, %f774;
sub.f32 %f776, %f771, %f775;
add.f32 %f777, %f774, %f776;
mul.rn.f32 %f779, %f754, %f622;
mul.rn.f32 %f781, %f754, %f624;
add.f32 %f782, %f779, %f775;
sub.f32 %f783, %f779, %f782;
add.f32 %f784, %f775, %f783;
add.f32 %f785, %f777, %f784;
add.f32 %f786, %f781, %f785;
add.f32 %f787, %f782, %f786;
sub.f32 %f788, %f782, %f787;
add.f32 %f789, %f786, %f788;
mul.rn.f32 %f791, %f634, %f787;
neg.f32 %f792, %f791;
fma.rn.f32 %f793, %f634, %f787, %f792;
fma.rn.f32 %f794, %f634, %f789, %f793;
fma.rn.f32 %f796, %f639, %f787, %f794;
add.rn.f32 %f797, %f791, %f796;
neg.f32 %f798, %f797;
add.rn.f32 %f799, %f791, %f798;
add.rn.f32 %f800, %f799, %f796;
mov.b32 %r294, %f797;
setp.eq.s32 %p123, %r294, 1118925336;
add.s32 %r295, %r294, -1;
mov.b32 %f801, %r295;
add.f32 %f802, %f800, 0f37000000;
selp.f32 %f803, %f801, %f797, %p123;
selp.f32 %f190, %f802, %f800, %p123;
mul.f32 %f804, %f803, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f805, %f804;
fma.rn.f32 %f807, %f805, %f650, %f803;
fma.rn.f32 %f809, %f805, %f652, %f807;
mul.f32 %f810, %f809, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f811, %f810;
add.f32 %f812, %f805, 0f00000000;
ex2.approx.f32 %f813, %f812;
mul.f32 %f814, %f811, %f813;
setp.lt.f32 %p124, %f803, 0fC2D20000;
selp.f32 %f815, 0f00000000, %f814, %p124;
setp.gt.f32 %p125, %f803, 0f42D20000;
selp.f32 %f919, 0f7F800000, %f815, %p125;
setp.eq.f32 %p126, %f919, 0f7F800000;
@%p126 bra BB0_118;
fma.rn.f32 %f919, %f919, %f190, %f919;
BB0_118:
setp.lt.f32 %p127, %f161, 0f00000000;
and.pred %p6, %p127, %p96;
mov.b32 %r296, %f919;
xor.b32 %r297, %r296, -2147483648;
mov.b32 %f816, %r297;
selp.f32 %f921, %f816, %f919, %p6;
setp.eq.f32 %p129, %f161, 0f00000000;
@%p129 bra BB0_121;
bra.uni BB0_119;
BB0_121:
add.f32 %f819, %f161, %f161;
selp.f32 %f921, %f819, 0f00000000, %p96;
bra.uni BB0_122;
BB0_119:
setp.geu.f32 %p130, %f161, 0f00000000;
@%p130 bra BB0_122;
cvt.rzi.f32.f32 %f818, %f634;
setp.neu.f32 %p131, %f818, 0f3EE66666;
selp.f32 %f921, 0f7FFFFFFF, %f921, %p131;
BB0_122:
add.f32 %f820, %f189, 0f3EE66666;
mov.b32 %r298, %f820;
setp.lt.s32 %p133, %r298, 2139095040;
@%p133 bra BB0_127;
setp.gtu.f32 %p134, %f189, 0f7F800000;
@%p134 bra BB0_126;
bra.uni BB0_124;
BB0_126:
add.f32 %f921, %f161, 0f3EE66666;
bra.uni BB0_127;
BB0_124:
setp.neu.f32 %p135, %f189, 0f7F800000;
@%p135 bra BB0_127;
selp.f32 %f921, 0fFF800000, 0f7F800000, %p6;
BB0_127:
mov.u32 %r329, 4;
setp.eq.f32 %p136, %f161, 0f3F800000;
selp.f32 %f821, 0f3F800000, %f921, %p136;
cvt.u64.u32 %rd94, %r6;
cvt.u64.u32 %rd93, %r5;
mov.u64 %rd97, image;
cvta.global.u64 %rd92, %rd97;
// inline asm
call (%rd91), _rt_buffer_get_64, (%rd92, %r103, %r329, %rd93, %rd94, %rd28, %rd28);
// inline asm
cvt.sat.f32.f32 %f822, %f821;
mul.f32 %f823, %f822, 0f437FFD71;
cvt.rzi.u32.f32 %r301, %f823;
cvt.sat.f32.f32 %f824, %f188;
mul.f32 %f825, %f824, 0f437FFD71;
cvt.rzi.u32.f32 %r302, %f825;
cvt.sat.f32.f32 %f826, %f175;
mul.f32 %f827, %f826, 0f437FFD71;
cvt.rzi.u32.f32 %r303, %f827;
cvt.u16.u32 %rs13, %r301;
cvt.u16.u32 %rs14, %r303;
cvt.u16.u32 %rs15, %r302;
mov.u16 %rs16, 255;
st.v4.u8 [%rd91], {%rs13, %rs15, %rs14, %rs16};
ld.global.u32 %r360, [imageEnabled];
BB0_128:
and.b32 %r304, %r360, 4;
setp.eq.s32 %p137, %r304, 0;
@%p137 bra BB0_140;
ld.global.u32 %r305, [additive];
setp.eq.s32 %p138, %r305, 0;
cvt.u64.u32 %rd19, %r5;
cvt.u64.u32 %rd20, %r6;
// inline asm
{ cvt.rn.f16.f32 %rs17, %f377;}
// inline asm
@%p138 bra BB0_131;
mov.u64 %rd110, image_HDR;
cvta.global.u64 %rd99, %rd110;
mov.u32 %r309, 8;
// inline asm
call (%rd98), _rt_buffer_get_64, (%rd99, %r103, %r309, %rd19, %rd20, %rd28, %rd28);
// inline asm
ld.v4.u16 {%rs24, %rs25, %rs26, %rs27}, [%rd98];
// inline asm
{ cvt.f32.f16 %f829, %rs24;}
// inline asm
// inline asm
{ cvt.f32.f16 %f830, %rs25;}
// inline asm
// inline asm
{ cvt.f32.f16 %f831, %rs26;}
// inline asm
// inline asm
call (%rd104), _rt_buffer_get_64, (%rd99, %r103, %r309, %rd19, %rd20, %rd28, %rd28);
// inline asm
add.f32 %f832, %f159, %f829;
add.f32 %f833, %f160, %f830;
add.f32 %f834, %f161, %f831;
// inline asm
{ cvt.rn.f16.f32 %rs23, %f834;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs22, %f833;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs21, %f832;}
// inline asm
st.v4.u16 [%rd104], {%rs21, %rs22, %rs23, %rs17};
bra.uni BB0_140;
BB0_131:
mov.u64 %rd117, image_HDR;
cvta.global.u64 %rd112, %rd117;
mov.u32 %r311, 8;
// inline asm
call (%rd111), _rt_buffer_get_64, (%rd112, %r103, %r311, %rd19, %rd20, %rd28, %rd28);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs30, %f161;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs29, %f160;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs28, %f159;}
// inline asm
st.v4.u16 [%rd111], {%rs28, %rs29, %rs30, %rs17};
BB0_140:
ret;
}