DeltaVR/Assets/Editor/x64/Bakery/lmAreaLight.ptx
2020-12-03 20:09:51 +02:00

1808 lines
52 KiB
Plaintext

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-23083092
// Cuda compilation tools, release 9.1, V9.1.85
// Based on LLVM 3.4svn
//
.version 6.1
.target sm_30
.address_size 64
// .globl _Z6oxMainv
.global .align 8 .b8 pixelID[8];
.global .align 8 .b8 resolution[8];
.global .align 4 .b8 normal[12];
.global .align 4 .b8 camPos[12];
.global .align 4 .b8 root[4];
.global .align 4 .u32 imageEnabled;
.global .texref lightmap;
.global .align 16 .b8 tileInfo[16];
.global .align 4 .u32 additive;
.global .align 1 .b8 image[1];
.global .align 1 .b8 image_HDR[1];
.global .align 1 .b8 image_HDR2[1];
.global .align 1 .b8 image_Mask[1];
.global .align 1 .b8 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 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 4 .b8 __local_depot0[36];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<143>;
.reg .b16 %rs<46>;
.reg .f32 %f<853>;
.reg .b32 %r<363>;
.reg .b64 %rd<138>;
mov.u64 %rd137, __local_depot0;
cvta.local.u64 %SP, %rd137;
ld.global.u32 %r1, [samples];
ld.global.v2.u32 {%r103, %r104}, [pixelID];
cvt.u64.u32 %rd21, %r103;
cvt.u64.u32 %rd22, %r104;
mov.u64 %rd25, uvnormal;
cvta.global.u64 %rd20, %rd25;
mov.u32 %r101, 2;
mov.u32 %r102, 4;
mov.u64 %rd24, 0;
// inline asm
call (%rd19), _rt_buffer_get_64, (%rd20, %r101, %r102, %rd21, %rd22, %rd24, %rd24);
// inline asm
ld.u32 %r2, [%rd19];
shr.u32 %r107, %r2, 16;
cvt.u16.u32 %rs1, %r107;
and.b16 %rs3, %rs1, 255;
cvt.u16.u32 %rs4, %r2;
or.b16 %rs5, %rs4, %rs3;
setp.eq.s16 %p7, %rs5, 0;
mov.f32 %f807, 0f00000000;
mov.f32 %f808, %f807;
mov.f32 %f809, %f807;
@%p7 bra BB0_2;
ld.u8 %rs6, [%rd19+1];
and.b16 %rs8, %rs4, 255;
cvt.rn.f32.u16 %f173, %rs8;
div.rn.f32 %f174, %f173, 0f437F0000;
fma.rn.f32 %f175, %f174, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f176, %rs6;
div.rn.f32 %f177, %f176, 0f437F0000;
fma.rn.f32 %f178, %f177, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f179, %rs3;
div.rn.f32 %f180, %f179, 0f437F0000;
fma.rn.f32 %f181, %f180, 0f40000000, 0fBF800000;
mul.f32 %f182, %f178, %f178;
fma.rn.f32 %f183, %f175, %f175, %f182;
fma.rn.f32 %f184, %f181, %f181, %f183;
sqrt.rn.f32 %f185, %f184;
rcp.rn.f32 %f186, %f185;
mul.f32 %f807, %f175, %f186;
mul.f32 %f808, %f178, %f186;
mul.f32 %f809, %f181, %f186;
BB0_2:
ld.global.v2.u32 {%r108, %r109}, [pixelID];
ld.global.v2.u32 {%r111, %r112}, [tileInfo];
add.s32 %r3, %r108, %r111;
add.s32 %r4, %r109, %r112;
setp.eq.f32 %p8, %f808, 0f00000000;
setp.eq.f32 %p9, %f807, 0f00000000;
and.pred %p10, %p9, %p8;
setp.eq.f32 %p11, %f809, 0f00000000;
and.pred %p12, %p10, %p11;
@%p12 bra BB0_129;
bra.uni BB0_3;
BB0_129:
ld.global.u32 %r362, [imageEnabled];
and.b32 %r308, %r362, 1;
setp.eq.b32 %p138, %r308, 1;
@!%p138 bra BB0_131;
bra.uni BB0_130;
BB0_130:
cvt.u64.u32 %rd99, %r3;
cvt.u64.u32 %rd100, %r4;
mov.u64 %rd103, image;
cvta.global.u64 %rd98, %rd103;
// inline asm
call (%rd97), _rt_buffer_get_64, (%rd98, %r101, %r102, %rd99, %rd100, %rd24, %rd24);
// inline asm
mov.u16 %rs30, 0;
st.v4.u8 [%rd97], {%rs30, %rs30, %rs30, %rs30};
ld.global.u32 %r362, [imageEnabled];
BB0_131:
and.b32 %r311, %r362, 8;
setp.eq.s32 %p139, %r311, 0;
@%p139 bra BB0_133;
cvt.u64.u32 %rd106, %r3;
cvt.u64.u32 %rd107, %r4;
mov.u64 %rd110, image_Mask;
cvta.global.u64 %rd105, %rd110;
// inline asm
call (%rd104), _rt_buffer_get_64, (%rd105, %r101, %r101, %rd106, %rd107, %rd24, %rd24);
// inline asm
mov.f32 %f790, 0f00000000;
cvt.rzi.u32.f32 %r314, %f790;
cvt.u16.u32 %rs31, %r314;
mov.u16 %rs32, 0;
st.v2.u8 [%rd104], {%rs31, %rs32};
ld.global.u32 %r362, [imageEnabled];
BB0_133:
and.b32 %r315, %r362, 4;
setp.eq.s32 %p140, %r315, 0;
@%p140 bra BB0_137;
ld.global.u32 %r316, [additive];
setp.eq.s32 %p141, %r316, 0;
cvt.u64.u32 %rd17, %r3;
cvt.u64.u32 %rd18, %r4;
@%p141 bra BB0_136;
mov.u64 %rd123, image_HDR;
cvta.global.u64 %rd112, %rd123;
mov.u32 %r320, 8;
// inline asm
call (%rd111), _rt_buffer_get_64, (%rd112, %r101, %r320, %rd17, %rd18, %rd24, %rd24);
// inline asm
ld.v4.u16 {%rs39, %rs40, %rs41, %rs42}, [%rd111];
// inline asm
{ cvt.f32.f16 %f791, %rs39;}
// inline asm
// inline asm
{ cvt.f32.f16 %f792, %rs40;}
// inline asm
// inline asm
{ cvt.f32.f16 %f793, %rs41;}
// inline asm
// inline asm
call (%rd117), _rt_buffer_get_64, (%rd112, %r101, %r320, %rd17, %rd18, %rd24, %rd24);
// inline asm
add.f32 %f794, %f791, 0f00000000;
add.f32 %f795, %f792, 0f00000000;
add.f32 %f796, %f793, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs38, %f796;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs37, %f795;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs36, %f794;}
// inline asm
mov.u16 %rs43, 0;
st.v4.u16 [%rd117], {%rs36, %rs37, %rs38, %rs43};
bra.uni BB0_137;
BB0_3:
ld.global.v2.u32 {%r121, %r122}, [pixelID];
cvt.u64.u32 %rd28, %r121;
cvt.u64.u32 %rd29, %r122;
mov.u64 %rd38, uvpos;
cvta.global.u64 %rd27, %rd38;
mov.u32 %r117, 12;
// inline asm
call (%rd26), _rt_buffer_get_64, (%rd27, %r101, %r117, %rd28, %rd29, %rd24, %rd24);
// inline asm
ld.f32 %f9, [%rd26+8];
ld.f32 %f8, [%rd26+4];
ld.f32 %f7, [%rd26];
mul.f32 %f189, %f7, 0f3456BF95;
mul.f32 %f190, %f8, 0f3456BF95;
mul.f32 %f191, %f9, 0f3456BF95;
abs.f32 %f10, %f807;
div.rn.f32 %f192, %f189, %f10;
abs.f32 %f193, %f808;
div.rn.f32 %f194, %f190, %f193;
abs.f32 %f11, %f809;
div.rn.f32 %f195, %f191, %f11;
abs.f32 %f196, %f192;
abs.f32 %f197, %f194;
abs.f32 %f198, %f195;
mov.f32 %f199, 0f38D1B717;
max.f32 %f200, %f196, %f199;
max.f32 %f201, %f197, %f199;
max.f32 %f202, %f198, %f199;
fma.rn.f32 %f12, %f807, %f200, %f7;
fma.rn.f32 %f13, %f808, %f201, %f8;
fma.rn.f32 %f14, %f809, %f202, %f9;
ld.global.v2.u32 {%r125, %r126}, [pixelID];
cvt.u64.u32 %rd34, %r125;
cvt.u64.u32 %rd35, %r126;
mov.u64 %rd39, rnd_seeds;
cvta.global.u64 %rd33, %rd39;
// inline asm
call (%rd32), _rt_buffer_get_64, (%rd33, %r101, %r102, %rd34, %rd35, %rd24, %rd24);
// inline asm
ld.u32 %r129, [%rd32];
mad.lo.s32 %r5, %r129, 1664525, 1013904223;
ld.global.u32 %r130, [lightMeshBufferSize];
setp.eq.s32 %p14, %r130, 0;
mov.pred %p13, 0;
mov.f32 %f20, 0f00000000;
mov.u32 %r7, 0;
@%p14 bra BB0_4;
ld.global.f32 %f15, [lightPointSize];
mul.f32 %f16, %f12, 0f3456BF95;
mul.f32 %f17, %f13, 0f3456BF95;
mul.f32 %f18, %f14, 0f3456BF95;
and.b32 %r133, %r5, 16777215;
cvt.rn.f32.u32 %f205, %r133;
mul.f32 %f206, %f205, 0fB3800000;
fma.rn.f32 %f19, %f206, 0f3F333333, 0f3F800000;
mov.f32 %f20, 0f00000000;
mov.u32 %r331, 0;
abs.f32 %f329, %f17;
abs.f32 %f330, %f16;
max.f32 %f331, %f330, %f329;
abs.f32 %f332, %f18;
max.f32 %f333, %f331, %f332;
mov.u32 %r7, %r331;
mov.f32 %f21, %f20;
BB0_6:
shl.b32 %r8, %r331, 1;
cvt.s64.s32 %rd42, %r8;
mov.u64 %rd46, lightMeshBuffer;
cvta.global.u64 %rd41, %rd46;
mov.u32 %r134, 1;
// inline asm
call (%rd40), _rt_buffer_get_64, (%rd41, %r134, %r117, %rd42, %rd24, %rd24, %rd24);
// inline asm
ld.f32 %f207, [%rd40];
sub.f32 %f208, %f207, %f7;
ld.f32 %f209, [%rd40+4];
sub.f32 %f210, %f209, %f8;
ld.f32 %f211, [%rd40+8];
sub.f32 %f212, %f211, %f9;
mul.f32 %f213, %f210, %f210;
fma.rn.f32 %f214, %f208, %f208, %f213;
fma.rn.f32 %f215, %f212, %f212, %f214;
sqrt.rn.f32 %f22, %f215;
rcp.rn.f32 %f216, %f22;
mul.f32 %f23, %f208, %f216;
mul.f32 %f24, %f210, %f216;
mul.f32 %f25, %f212, %f216;
mul.f32 %f217, %f808, %f24;
fma.rn.f32 %f218, %f807, %f23, %f217;
fma.rn.f32 %f26, %f809, %f25, %f218;
setp.leu.f32 %p15, %f26, 0f00000000;
@%p15 bra BB0_22;
setp.ne.s32 %p17, %r1, 0;
mul.f32 %f219, %f22, %f22;
mul.f32 %f220, %f219, 0f40C90FDB;
div.rn.f32 %f221, %f15, %f220;
add.f32 %f27, %f221, %f221;
setp.gt.f32 %p18, %f27, %f19;
and.pred %p19, %p17, %p18;
mov.pred %p142, -1;
@%p19 bra BB0_24;
ld.global.f32 %f224, [lightInvCutoff];
mul.f32 %f28, %f22, %f224;
mov.f32 %f228, 0f40800000;
abs.f32 %f30, %f28;
setp.lt.f32 %p20, %f30, 0f00800000;
mul.f32 %f230, %f30, 0f4B800000;
selp.f32 %f231, 0fC3170000, 0fC2FE0000, %p20;
selp.f32 %f232, %f230, %f30, %p20;
mov.b32 %r136, %f232;
and.b32 %r137, %r136, 8388607;
or.b32 %r138, %r137, 1065353216;
mov.b32 %f233, %r138;
shr.u32 %r139, %r136, 23;
cvt.rn.f32.u32 %f234, %r139;
add.f32 %f235, %f231, %f234;
setp.gt.f32 %p21, %f233, 0f3FB504F3;
mul.f32 %f236, %f233, 0f3F000000;
add.f32 %f237, %f235, 0f3F800000;
selp.f32 %f238, %f236, %f233, %p21;
selp.f32 %f239, %f237, %f235, %p21;
add.f32 %f240, %f238, 0fBF800000;
add.f32 %f223, %f238, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f222,%f223;
// inline asm
add.f32 %f241, %f240, %f240;
mul.f32 %f242, %f222, %f241;
mul.f32 %f243, %f242, %f242;
mov.f32 %f244, 0f3C4CAF63;
mov.f32 %f245, 0f3B18F0FE;
fma.rn.f32 %f246, %f245, %f243, %f244;
mov.f32 %f247, 0f3DAAAABD;
fma.rn.f32 %f248, %f246, %f243, %f247;
mul.rn.f32 %f249, %f248, %f243;
mul.rn.f32 %f250, %f249, %f242;
sub.f32 %f251, %f240, %f242;
neg.f32 %f252, %f242;
add.f32 %f253, %f251, %f251;
fma.rn.f32 %f254, %f252, %f240, %f253;
mul.rn.f32 %f255, %f222, %f254;
add.f32 %f256, %f250, %f242;
sub.f32 %f257, %f242, %f256;
add.f32 %f258, %f250, %f257;
add.f32 %f259, %f255, %f258;
add.f32 %f260, %f256, %f259;
sub.f32 %f261, %f256, %f260;
add.f32 %f262, %f259, %f261;
mov.f32 %f263, 0f3F317200;
mul.rn.f32 %f264, %f239, %f263;
mov.f32 %f265, 0f35BFBE8E;
mul.rn.f32 %f266, %f239, %f265;
add.f32 %f267, %f264, %f260;
sub.f32 %f268, %f264, %f267;
add.f32 %f269, %f260, %f268;
add.f32 %f270, %f262, %f269;
add.f32 %f271, %f266, %f270;
add.f32 %f272, %f267, %f271;
sub.f32 %f273, %f267, %f272;
add.f32 %f274, %f271, %f273;
mul.rn.f32 %f275, %f228, %f272;
neg.f32 %f276, %f275;
fma.rn.f32 %f277, %f228, %f272, %f276;
fma.rn.f32 %f278, %f228, %f274, %f277;
mov.f32 %f279, 0f00000000;
fma.rn.f32 %f280, %f279, %f272, %f278;
add.rn.f32 %f281, %f275, %f280;
neg.f32 %f282, %f281;
add.rn.f32 %f283, %f275, %f282;
add.rn.f32 %f284, %f283, %f280;
mov.b32 %r140, %f281;
setp.eq.s32 %p22, %r140, 1118925336;
add.s32 %r141, %r140, -1;
mov.b32 %f285, %r141;
add.f32 %f286, %f284, 0f37000000;
selp.f32 %f287, %f285, %f281, %p22;
selp.f32 %f31, %f286, %f284, %p22;
mul.f32 %f288, %f287, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f289, %f288;
mov.f32 %f290, 0fBF317200;
fma.rn.f32 %f291, %f289, %f290, %f287;
mov.f32 %f292, 0fB5BFBE8E;
fma.rn.f32 %f293, %f289, %f292, %f291;
mul.f32 %f294, %f293, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f295, %f294;
add.f32 %f296, %f289, 0f00000000;
ex2.approx.f32 %f297, %f296;
mul.f32 %f298, %f295, %f297;
setp.lt.f32 %p23, %f287, 0fC2D20000;
selp.f32 %f299, 0f00000000, %f298, %p23;
setp.gt.f32 %p24, %f287, 0f42D20000;
selp.f32 %f812, 0f7F800000, %f299, %p24;
setp.eq.f32 %p25, %f812, 0f7F800000;
@%p25 bra BB0_10;
fma.rn.f32 %f812, %f812, %f31, %f812;
BB0_10:
mov.f32 %f803, 0f40000000;
cvt.rzi.f32.f32 %f802, %f803;
add.f32 %f801, %f802, %f802;
mov.f32 %f800, 0f40800000;
sub.f32 %f799, %f800, %f801;
abs.f32 %f798, %f799;
setp.lt.f32 %p26, %f28, 0f00000000;
setp.eq.f32 %p27, %f798, 0f3F800000;
and.pred %p1, %p26, %p27;
mov.b32 %r142, %f812;
xor.b32 %r143, %r142, -2147483648;
mov.b32 %f300, %r143;
selp.f32 %f814, %f300, %f812, %p1;
setp.eq.f32 %p28, %f28, 0f00000000;
@%p28 bra BB0_13;
bra.uni BB0_11;
BB0_13:
add.f32 %f303, %f28, %f28;
selp.f32 %f814, %f303, 0f00000000, %p27;
bra.uni BB0_14;
BB0_11:
setp.geu.f32 %p29, %f28, 0f00000000;
@%p29 bra BB0_14;
mov.f32 %f806, 0f40800000;
cvt.rzi.f32.f32 %f302, %f806;
setp.neu.f32 %p30, %f302, 0f40800000;
selp.f32 %f814, 0f7FFFFFFF, %f814, %p30;
BB0_14:
add.f32 %f304, %f30, 0f40800000;
mov.b32 %r144, %f304;
setp.lt.s32 %p32, %r144, 2139095040;
@%p32 bra BB0_19;
setp.gtu.f32 %p33, %f30, 0f7F800000;
@%p33 bra BB0_18;
bra.uni BB0_16;
BB0_18:
add.f32 %f814, %f28, 0f40800000;
bra.uni BB0_19;
BB0_16:
setp.neu.f32 %p34, %f30, 0f7F800000;
@%p34 bra BB0_19;
selp.f32 %f814, 0fFF800000, 0f7F800000, %p1;
BB0_19:
mov.u32 %r324, 1;
mov.u64 %rd132, lightMeshBuffer;
cvta.global.u64 %rd131, %rd132;
shl.b32 %r323, %r331, 1;
mov.f32 %f305, 0f3F800000;
sub.f32 %f306, %f305, %f814;
setp.eq.f32 %p35, %f28, 0f3F800000;
selp.f32 %f307, 0f00000000, %f306, %p35;
cvt.sat.f32.f32 %f308, %f307;
mul.f32 %f309, %f27, %f308;
add.s32 %r147, %r323, 1;
cvt.s64.s32 %rd49, %r147;
// inline asm
call (%rd47), _rt_buffer_get_64, (%rd131, %r324, %r117, %rd49, %rd24, %rd24, %rd24);
// inline asm
ld.f32 %f310, [%rd47];
mul.f32 %f311, %f23, %f310;
ld.f32 %f312, [%rd47+4];
mul.f32 %f313, %f24, %f312;
neg.f32 %f314, %f313;
sub.f32 %f315, %f314, %f311;
ld.f32 %f316, [%rd47+8];
mul.f32 %f317, %f25, %f316;
sub.f32 %f318, %f315, %f317;
cvt.sat.f32.f32 %f319, %f318;
mul.f32 %f42, %f309, %f319;
cvt.sat.f32.f32 %f43, %f26;
mul.f32 %f320, %f42, %f43;
setp.leu.f32 %p36, %f320, 0f3727C5AC;
@%p36 bra BB0_21;
mov.u32 %r330, 1;
add.u64 %rd54, %SP, 28;
cvta.to.local.u64 %rd55, %rd54;
max.f32 %f327, %f333, %f199;
sub.f32 %f328, %f22, %f327;
mov.u32 %r151, 1065353216;
st.local.u32 [%rd55], %r151;
ld.global.u32 %r148, [root];
// inline asm
call _rt_trace_64, (%r148, %f12, %f13, %f14, %f23, %f24, %f25, %r330, %f327, %f328, %rd54, %r102);
// inline asm
ld.local.f32 %f335, [%rd55];
mul.f32 %f336, %f42, %f335;
fma.rn.f32 %f21, %f43, %f336, %f21;
add.f32 %f20, %f20, %f335;
BB0_21:
add.s32 %r7, %r7, 1;
BB0_22:
ld.global.u32 %r152, [lightMeshBufferSize];
add.s32 %r331, %r331, 1;
setp.lt.u32 %p38, %r331, %r152;
@%p38 bra BB0_6;
bra.uni BB0_23;
BB0_4:
mov.f32 %f21, %f20;
BB0_23:
mov.pred %p142, %p13;
BB0_24:
cvt.rn.f32.s32 %f337, %r7;
mov.f32 %f338, 0f3F800000;
max.f32 %f339, %f337, %f338;
div.rn.f32 %f839, %f21, %f339;
div.rn.f32 %f840, %f20, %f339;
@!%p142 bra BB0_77;
bra.uni BB0_25;
BB0_25:
abs.f32 %f805, %f809;
abs.f32 %f804, %f807;
setp.gt.f32 %p39, %f804, %f805;
neg.f32 %f342, %f808;
selp.f32 %f343, %f342, 0f00000000, %p39;
neg.f32 %f344, %f809;
selp.f32 %f345, %f807, %f344, %p39;
selp.f32 %f346, 0f00000000, %f808, %p39;
mul.f32 %f347, %f345, %f345;
fma.rn.f32 %f348, %f343, %f343, %f347;
fma.rn.f32 %f349, %f346, %f346, %f348;
sqrt.rn.f32 %f350, %f349;
rcp.rn.f32 %f351, %f350;
mul.f32 %f54, %f343, %f351;
mul.f32 %f55, %f345, %f351;
mul.f32 %f56, %f346, %f351;
mov.f32 %f823, 0f00000000;
setp.lt.s32 %p40, %r1, 1;
mov.f32 %f824, %f823;
@%p40 bra BB0_76;
mad.lo.s32 %r338, %r129, 1664525, 1013904223;
cvt.rn.f32.s32 %f354, %r1;
rcp.rn.f32 %f57, %f354;
add.u64 %rd56, %SP, 0;
cvta.to.local.u64 %rd2, %rd56;
mul.f32 %f58, %f12, 0f3456BF95;
mul.f32 %f59, %f13, 0f3456BF95;
mul.f32 %f60, %f14, 0f3456BF95;
add.u64 %rd57, %SP, 32;
cvta.to.local.u64 %rd3, %rd57;
mul.f32 %f355, %f807, %f55;
mul.f32 %f356, %f808, %f54;
sub.f32 %f61, %f356, %f355;
mul.f32 %f357, %f809, %f54;
mul.f32 %f358, %f807, %f56;
sub.f32 %f62, %f358, %f357;
mul.f32 %f359, %f808, %f56;
mul.f32 %f360, %f809, %f55;
sub.f32 %f63, %f360, %f359;
mov.f32 %f823, 0f00000000;
mov.u32 %r153, 0;
abs.f32 %f361, %f59;
abs.f32 %f362, %f58;
max.f32 %f363, %f362, %f361;
abs.f32 %f364, %f60;
max.f32 %f365, %f363, %f364;
mov.u32 %r335, %r153;
mov.f32 %f824, %f823;
BB0_27:
cvt.rn.f32.s32 %f66, %r335;
max.f32 %f67, %f365, %f199;
mov.u32 %r337, %r153;
BB0_28:
mad.lo.s32 %r155, %r338, 1664525, 1013904223;
and.b32 %r156, %r155, 16777215;
cvt.rn.f32.u32 %f367, %r156;
fma.rn.f32 %f368, %f367, 0f33800000, %f66;
mul.f32 %f369, %f57, %f368;
mad.lo.s32 %r338, %r155, 1664525, 1013904223;
and.b32 %r157, %r338, 16777215;
cvt.rn.f32.u32 %f370, %r157;
cvt.rn.f32.s32 %f371, %r337;
fma.rn.f32 %f372, %f370, 0f33800000, %f371;
mul.f32 %f373, %f57, %f372;
sqrt.rn.f32 %f70, %f369;
mul.f32 %f831, %f373, 0f40C90FDB;
abs.f32 %f72, %f831;
setp.neu.f32 %p41, %f72, 0f7F800000;
mov.f32 %f825, %f831;
@%p41 bra BB0_30;
mov.f32 %f374, 0f00000000;
mul.rn.f32 %f825, %f831, %f374;
BB0_30:
mul.f32 %f375, %f825, 0f3F22F983;
cvt.rni.s32.f32 %r348, %f375;
cvt.rn.f32.s32 %f376, %r348;
neg.f32 %f377, %f376;
mov.f32 %f378, 0f3FC90FDA;
fma.rn.f32 %f379, %f377, %f378, %f825;
mov.f32 %f380, 0f33A22168;
fma.rn.f32 %f381, %f377, %f380, %f379;
mov.f32 %f382, 0f27C234C5;
fma.rn.f32 %f826, %f377, %f382, %f381;
abs.f32 %f383, %f825;
setp.leu.f32 %p42, %f383, 0f47CE4780;
@%p42 bra BB0_41;
mov.b32 %r19, %f825;
shr.u32 %r20, %r19, 23;
shl.b32 %r160, %r19, 8;
or.b32 %r21, %r160, -2147483648;
mov.u32 %r340, 0;
mov.u64 %rd133, __cudart_i2opi_f;
mov.u32 %r339, -6;
mov.u64 %rd134, %rd2;
BB0_32:
.pragma "nounroll";
ld.const.u32 %r163, [%rd133];
// inline asm
{
mad.lo.cc.u32 %r161, %r163, %r21, %r340;
madc.hi.u32 %r340, %r163, %r21, 0;
}
// inline asm
st.local.u32 [%rd134], %r161;
add.s64 %rd134, %rd134, 4;
add.s64 %rd133, %rd133, 4;
add.s32 %r339, %r339, 1;
setp.ne.s32 %p43, %r339, 0;
@%p43 bra BB0_32;
and.b32 %r166, %r20, 255;
add.s32 %r167, %r166, -128;
shr.u32 %r168, %r167, 5;
and.b32 %r26, %r19, -2147483648;
st.local.u32 [%rd2+24], %r340;
mov.u32 %r169, 6;
sub.s32 %r170, %r169, %r168;
mul.wide.s32 %rd59, %r170, 4;
add.s64 %rd8, %rd2, %rd59;
ld.local.u32 %r341, [%rd8];
ld.local.u32 %r342, [%rd8+-4];
and.b32 %r29, %r20, 31;
setp.eq.s32 %p44, %r29, 0;
@%p44 bra BB0_35;
mov.u32 %r171, 32;
sub.s32 %r172, %r171, %r29;
shr.u32 %r173, %r342, %r172;
shl.b32 %r174, %r341, %r29;
add.s32 %r341, %r173, %r174;
ld.local.u32 %r175, [%rd8+-8];
shr.u32 %r176, %r175, %r172;
shl.b32 %r177, %r342, %r29;
add.s32 %r342, %r176, %r177;
BB0_35:
shr.u32 %r178, %r342, 30;
shl.b32 %r179, %r341, 2;
add.s32 %r343, %r178, %r179;
shl.b32 %r35, %r342, 2;
shr.u32 %r180, %r343, 31;
shr.u32 %r181, %r341, 30;
add.s32 %r36, %r180, %r181;
setp.eq.s32 %p45, %r180, 0;
@%p45 bra BB0_36;
bra.uni BB0_37;
BB0_36:
mov.u32 %r344, %r26;
mov.u32 %r345, %r35;
bra.uni BB0_38;
BB0_37:
not.b32 %r182, %r343;
neg.s32 %r345, %r35;
setp.eq.s32 %p46, %r35, 0;
selp.u32 %r183, 1, 0, %p46;
add.s32 %r343, %r183, %r182;
xor.b32 %r344, %r26, -2147483648;
BB0_38:
clz.b32 %r347, %r343;
setp.eq.s32 %p47, %r347, 0;
shl.b32 %r184, %r343, %r347;
mov.u32 %r185, 32;
sub.s32 %r186, %r185, %r347;
shr.u32 %r187, %r345, %r186;
add.s32 %r188, %r187, %r184;
selp.b32 %r44, %r343, %r188, %p47;
mov.u32 %r189, -921707870;
mul.hi.u32 %r346, %r44, %r189;
setp.eq.s32 %p48, %r26, 0;
neg.s32 %r190, %r36;
selp.b32 %r348, %r36, %r190, %p48;
setp.lt.s32 %p49, %r346, 1;
@%p49 bra BB0_40;
mul.lo.s32 %r191, %r44, -921707870;
shr.u32 %r192, %r191, 31;
shl.b32 %r193, %r346, 1;
add.s32 %r346, %r192, %r193;
add.s32 %r347, %r347, 1;
BB0_40:
mov.u32 %r194, 126;
sub.s32 %r195, %r194, %r347;
shl.b32 %r196, %r195, 23;
add.s32 %r197, %r346, 1;
shr.u32 %r198, %r197, 7;
add.s32 %r199, %r198, 1;
shr.u32 %r200, %r199, 1;
add.s32 %r201, %r200, %r196;
or.b32 %r202, %r201, %r344;
mov.b32 %f826, %r202;
BB0_41:
mul.rn.f32 %f78, %f826, %f826;
add.s32 %r52, %r348, 1;
and.b32 %r53, %r52, 1;
setp.eq.s32 %p50, %r53, 0;
@%p50 bra BB0_43;
bra.uni BB0_42;
BB0_43:
mov.f32 %f386, 0f3C08839E;
mov.f32 %f387, 0fB94CA1F9;
fma.rn.f32 %f827, %f387, %f78, %f386;
bra.uni BB0_44;
BB0_42:
mov.f32 %f384, 0fBAB6061A;
mov.f32 %f385, 0f37CCF5CE;
fma.rn.f32 %f827, %f385, %f78, %f384;
BB0_44:
@%p50 bra BB0_46;
bra.uni BB0_45;
BB0_46:
mov.f32 %f391, 0fBE2AAAA3;
fma.rn.f32 %f392, %f827, %f78, %f391;
mov.f32 %f393, 0f00000000;
fma.rn.f32 %f828, %f392, %f78, %f393;
bra.uni BB0_47;
BB0_45:
mov.f32 %f388, 0f3D2AAAA5;
fma.rn.f32 %f389, %f827, %f78, %f388;
mov.f32 %f390, 0fBF000000;
fma.rn.f32 %f828, %f389, %f78, %f390;
BB0_47:
fma.rn.f32 %f829, %f828, %f826, %f826;
@%p50 bra BB0_49;
fma.rn.f32 %f829, %f828, %f78, %f338;
BB0_49:
and.b32 %r203, %r52, 2;
setp.eq.s32 %p53, %r203, 0;
@%p53 bra BB0_51;
mov.f32 %f395, 0f00000000;
mov.f32 %f396, 0fBF800000;
fma.rn.f32 %f829, %f829, %f396, %f395;
BB0_51:
@%p41 bra BB0_53;
mov.f32 %f397, 0f00000000;
mul.rn.f32 %f831, %f831, %f397;
BB0_53:
mul.f32 %f398, %f831, 0f3F22F983;
cvt.rni.s32.f32 %r358, %f398;
cvt.rn.f32.s32 %f399, %r358;
neg.f32 %f400, %f399;
fma.rn.f32 %f402, %f400, %f378, %f831;
fma.rn.f32 %f404, %f400, %f380, %f402;
fma.rn.f32 %f832, %f400, %f382, %f404;
abs.f32 %f406, %f831;
setp.leu.f32 %p55, %f406, 0f47CE4780;
@%p55 bra BB0_64;
mov.b32 %r55, %f831;
shr.u32 %r56, %r55, 23;
shl.b32 %r206, %r55, 8;
or.b32 %r57, %r206, -2147483648;
mov.u32 %r350, 0;
mov.u64 %rd135, __cudart_i2opi_f;
mov.u32 %r349, -6;
mov.u64 %rd136, %rd2;
BB0_55:
.pragma "nounroll";
ld.const.u32 %r209, [%rd135];
// inline asm
{
mad.lo.cc.u32 %r207, %r209, %r57, %r350;
madc.hi.u32 %r350, %r209, %r57, 0;
}
// inline asm
st.local.u32 [%rd136], %r207;
add.s64 %rd136, %rd136, 4;
add.s64 %rd135, %rd135, 4;
add.s32 %r349, %r349, 1;
setp.ne.s32 %p56, %r349, 0;
@%p56 bra BB0_55;
and.b32 %r212, %r56, 255;
add.s32 %r213, %r212, -128;
shr.u32 %r214, %r213, 5;
and.b32 %r62, %r55, -2147483648;
st.local.u32 [%rd2+24], %r350;
mov.u32 %r215, 6;
sub.s32 %r216, %r215, %r214;
mul.wide.s32 %rd61, %r216, 4;
add.s64 %rd13, %rd2, %rd61;
ld.local.u32 %r351, [%rd13];
ld.local.u32 %r352, [%rd13+-4];
and.b32 %r65, %r56, 31;
setp.eq.s32 %p57, %r65, 0;
@%p57 bra BB0_58;
mov.u32 %r217, 32;
sub.s32 %r218, %r217, %r65;
shr.u32 %r219, %r352, %r218;
shl.b32 %r220, %r351, %r65;
add.s32 %r351, %r219, %r220;
ld.local.u32 %r221, [%rd13+-8];
shr.u32 %r222, %r221, %r218;
shl.b32 %r223, %r352, %r65;
add.s32 %r352, %r222, %r223;
BB0_58:
shr.u32 %r224, %r352, 30;
shl.b32 %r225, %r351, 2;
add.s32 %r353, %r224, %r225;
shl.b32 %r71, %r352, 2;
shr.u32 %r226, %r353, 31;
shr.u32 %r227, %r351, 30;
add.s32 %r72, %r226, %r227;
setp.eq.s32 %p58, %r226, 0;
@%p58 bra BB0_59;
bra.uni BB0_60;
BB0_59:
mov.u32 %r354, %r62;
mov.u32 %r355, %r71;
bra.uni BB0_61;
BB0_60:
not.b32 %r228, %r353;
neg.s32 %r355, %r71;
setp.eq.s32 %p59, %r71, 0;
selp.u32 %r229, 1, 0, %p59;
add.s32 %r353, %r229, %r228;
xor.b32 %r354, %r62, -2147483648;
BB0_61:
clz.b32 %r357, %r353;
setp.eq.s32 %p60, %r357, 0;
shl.b32 %r230, %r353, %r357;
mov.u32 %r231, 32;
sub.s32 %r232, %r231, %r357;
shr.u32 %r233, %r355, %r232;
add.s32 %r234, %r233, %r230;
selp.b32 %r80, %r353, %r234, %p60;
mov.u32 %r235, -921707870;
mul.hi.u32 %r356, %r80, %r235;
setp.eq.s32 %p61, %r62, 0;
neg.s32 %r236, %r72;
selp.b32 %r358, %r72, %r236, %p61;
setp.lt.s32 %p62, %r356, 1;
@%p62 bra BB0_63;
mul.lo.s32 %r237, %r80, -921707870;
shr.u32 %r238, %r237, 31;
shl.b32 %r239, %r356, 1;
add.s32 %r356, %r238, %r239;
add.s32 %r357, %r357, 1;
BB0_63:
mov.u32 %r240, 126;
sub.s32 %r241, %r240, %r357;
shl.b32 %r242, %r241, 23;
add.s32 %r243, %r356, 1;
shr.u32 %r244, %r243, 7;
add.s32 %r245, %r244, 1;
shr.u32 %r246, %r245, 1;
add.s32 %r247, %r246, %r242;
or.b32 %r248, %r247, %r354;
mov.b32 %f832, %r248;
BB0_64:
mul.rn.f32 %f95, %f832, %f832;
and.b32 %r88, %r358, 1;
setp.eq.s32 %p63, %r88, 0;
@%p63 bra BB0_66;
bra.uni BB0_65;
BB0_66:
mov.f32 %f409, 0f3C08839E;
mov.f32 %f410, 0fB94CA1F9;
fma.rn.f32 %f833, %f410, %f95, %f409;
bra.uni BB0_67;
BB0_65:
mov.f32 %f407, 0fBAB6061A;
mov.f32 %f408, 0f37CCF5CE;
fma.rn.f32 %f833, %f408, %f95, %f407;
BB0_67:
@%p63 bra BB0_69;
bra.uni BB0_68;
BB0_69:
mov.f32 %f414, 0fBE2AAAA3;
fma.rn.f32 %f415, %f833, %f95, %f414;
mov.f32 %f416, 0f00000000;
fma.rn.f32 %f834, %f415, %f95, %f416;
bra.uni BB0_70;
BB0_68:
mov.f32 %f411, 0f3D2AAAA5;
fma.rn.f32 %f412, %f833, %f95, %f411;
mov.f32 %f413, 0fBF000000;
fma.rn.f32 %f834, %f412, %f95, %f413;
BB0_70:
fma.rn.f32 %f835, %f834, %f832, %f832;
@%p63 bra BB0_72;
fma.rn.f32 %f835, %f834, %f95, %f338;
BB0_72:
and.b32 %r249, %r358, 2;
setp.eq.s32 %p66, %r249, 0;
@%p66 bra BB0_74;
mov.f32 %f418, 0f00000000;
mov.f32 %f419, 0fBF800000;
fma.rn.f32 %f835, %f835, %f419, %f418;
BB0_74:
mul.f32 %f428, %f70, %f829;
mul.f32 %f429, %f428, %f428;
sub.f32 %f431, %f338, %f429;
mul.f32 %f432, %f70, %f835;
mul.f32 %f433, %f432, %f432;
sub.f32 %f434, %f431, %f433;
mov.f32 %f435, 0f00000000;
max.f32 %f436, %f435, %f434;
sqrt.rn.f32 %f437, %f436;
mul.f32 %f438, %f54, %f432;
mul.f32 %f439, %f55, %f432;
mul.f32 %f440, %f56, %f432;
fma.rn.f32 %f441, %f63, %f428, %f438;
fma.rn.f32 %f442, %f62, %f428, %f439;
fma.rn.f32 %f443, %f61, %f428, %f440;
fma.rn.f32 %f423, %f807, %f437, %f441;
fma.rn.f32 %f424, %f808, %f437, %f442;
fma.rn.f32 %f425, %f809, %f437, %f443;
mov.u32 %r251, 0;
st.local.u32 [%rd3], %r251;
ld.global.u32 %r250, [root];
mov.f32 %f427, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r250, %f12, %f13, %f14, %f423, %f424, %f425, %r251, %f67, %f427, %rd57, %r102);
// inline asm
ld.local.f32 %f444, [%rd3];
setp.lt.f32 %p67, %f444, 0f00000000;
selp.f32 %f445, 0f00000000, %f444, %p67;
selp.f32 %f446, 0f00000000, 0f3F800000, %p67;
add.f32 %f823, %f823, %f446;
add.f32 %f824, %f824, %f445;
add.s32 %r337, %r337, 1;
setp.lt.s32 %p68, %r337, %r1;
@%p68 bra BB0_28;
add.s32 %r335, %r335, 1;
setp.lt.s32 %p69, %r335, %r1;
@%p69 bra BB0_27;
BB0_76:
mul.lo.s32 %r253, %r1, %r1;
cvt.rn.f32.s32 %f447, %r253;
div.rn.f32 %f839, %f824, %f447;
div.rn.f32 %f840, %f823, %f447;
BB0_77:
ld.global.u32 %r360, [imageEnabled];
and.b32 %r254, %r360, 8;
setp.eq.s32 %p70, %r254, 0;
@%p70 bra BB0_90;
mov.u32 %r325, 2;
cvt.u64.u32 %rd65, %r3;
cvt.u64.u32 %rd66, %r4;
mov.u64 %rd69, image_Mask;
cvta.global.u64 %rd64, %rd69;
// inline asm
call (%rd63), _rt_buffer_get_64, (%rd64, %r325, %r325, %rd65, %rd66, %rd24, %rd24);
// inline asm
mov.f32 %f450, 0f3E68BA2E;
cvt.rzi.f32.f32 %f451, %f450;
fma.rn.f32 %f452, %f451, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f115, %f452;
abs.f32 %f116, %f840;
setp.lt.f32 %p71, %f116, 0f00800000;
mul.f32 %f453, %f116, 0f4B800000;
selp.f32 %f454, 0fC3170000, 0fC2FE0000, %p71;
selp.f32 %f455, %f453, %f116, %p71;
mov.b32 %r257, %f455;
and.b32 %r258, %r257, 8388607;
or.b32 %r259, %r258, 1065353216;
mov.b32 %f456, %r259;
shr.u32 %r260, %r257, 23;
cvt.rn.f32.u32 %f457, %r260;
add.f32 %f458, %f454, %f457;
setp.gt.f32 %p72, %f456, 0f3FB504F3;
mul.f32 %f459, %f456, 0f3F000000;
add.f32 %f460, %f458, 0f3F800000;
selp.f32 %f461, %f459, %f456, %p72;
selp.f32 %f462, %f460, %f458, %p72;
add.f32 %f463, %f461, 0fBF800000;
add.f32 %f449, %f461, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f448,%f449;
// inline asm
add.f32 %f464, %f463, %f463;
mul.f32 %f465, %f448, %f464;
mul.f32 %f466, %f465, %f465;
mov.f32 %f467, 0f3C4CAF63;
mov.f32 %f468, 0f3B18F0FE;
fma.rn.f32 %f469, %f468, %f466, %f467;
mov.f32 %f470, 0f3DAAAABD;
fma.rn.f32 %f471, %f469, %f466, %f470;
mul.rn.f32 %f472, %f471, %f466;
mul.rn.f32 %f473, %f472, %f465;
sub.f32 %f474, %f463, %f465;
neg.f32 %f475, %f465;
add.f32 %f476, %f474, %f474;
fma.rn.f32 %f477, %f475, %f463, %f476;
mul.rn.f32 %f478, %f448, %f477;
add.f32 %f479, %f473, %f465;
sub.f32 %f480, %f465, %f479;
add.f32 %f481, %f473, %f480;
add.f32 %f482, %f478, %f481;
add.f32 %f483, %f479, %f482;
sub.f32 %f484, %f479, %f483;
add.f32 %f485, %f482, %f484;
mov.f32 %f486, 0f3F317200;
mul.rn.f32 %f487, %f462, %f486;
mov.f32 %f488, 0f35BFBE8E;
mul.rn.f32 %f489, %f462, %f488;
add.f32 %f490, %f487, %f483;
sub.f32 %f491, %f487, %f490;
add.f32 %f492, %f483, %f491;
add.f32 %f493, %f485, %f492;
add.f32 %f494, %f489, %f493;
add.f32 %f495, %f490, %f494;
sub.f32 %f496, %f490, %f495;
add.f32 %f497, %f494, %f496;
mov.f32 %f498, 0f3EE8BA2E;
mul.rn.f32 %f499, %f498, %f495;
neg.f32 %f500, %f499;
fma.rn.f32 %f501, %f498, %f495, %f500;
fma.rn.f32 %f502, %f498, %f497, %f501;
mov.f32 %f503, 0f00000000;
fma.rn.f32 %f504, %f503, %f495, %f502;
add.rn.f32 %f505, %f499, %f504;
neg.f32 %f506, %f505;
add.rn.f32 %f507, %f499, %f506;
add.rn.f32 %f508, %f507, %f504;
mov.b32 %r261, %f505;
setp.eq.s32 %p73, %r261, 1118925336;
add.s32 %r262, %r261, -1;
mov.b32 %f509, %r262;
add.f32 %f510, %f508, 0f37000000;
selp.f32 %f511, %f509, %f505, %p73;
selp.f32 %f117, %f510, %f508, %p73;
mul.f32 %f512, %f511, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f513, %f512;
mov.f32 %f514, 0fBF317200;
fma.rn.f32 %f515, %f513, %f514, %f511;
mov.f32 %f516, 0fB5BFBE8E;
fma.rn.f32 %f517, %f513, %f516, %f515;
mul.f32 %f518, %f517, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f519, %f518;
add.f32 %f520, %f513, 0f00000000;
ex2.approx.f32 %f521, %f520;
mul.f32 %f522, %f519, %f521;
setp.lt.f32 %p74, %f511, 0fC2D20000;
selp.f32 %f523, 0f00000000, %f522, %p74;
setp.gt.f32 %p75, %f511, 0f42D20000;
selp.f32 %f841, 0f7F800000, %f523, %p75;
setp.eq.f32 %p76, %f841, 0f7F800000;
@%p76 bra BB0_80;
fma.rn.f32 %f841, %f841, %f117, %f841;
BB0_80:
setp.lt.f32 %p77, %f840, 0f00000000;
setp.eq.f32 %p78, %f115, 0f3F800000;
and.pred %p3, %p77, %p78;
mov.b32 %r263, %f841;
xor.b32 %r264, %r263, -2147483648;
mov.b32 %f524, %r264;
selp.f32 %f843, %f524, %f841, %p3;
setp.eq.f32 %p79, %f840, 0f00000000;
@%p79 bra BB0_83;
bra.uni BB0_81;
BB0_83:
add.f32 %f527, %f840, %f840;
selp.f32 %f843, %f527, 0f00000000, %p78;
bra.uni BB0_84;
BB0_136:
mov.u64 %rd130, image_HDR;
cvta.global.u64 %rd125, %rd130;
mov.u32 %r322, 8;
// inline asm
call (%rd124), _rt_buffer_get_64, (%rd125, %r101, %r322, %rd17, %rd18, %rd24, %rd24);
// inline asm
mov.f32 %f797, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs44, %f797;}
// inline asm
mov.u16 %rs45, 0;
st.v4.u16 [%rd124], {%rs44, %rs44, %rs44, %rs45};
bra.uni BB0_137;
BB0_81:
setp.geu.f32 %p80, %f840, 0f00000000;
@%p80 bra BB0_84;
cvt.rzi.f32.f32 %f526, %f498;
setp.neu.f32 %p81, %f526, 0f3EE8BA2E;
selp.f32 %f843, 0f7FFFFFFF, %f843, %p81;
BB0_84:
add.f32 %f528, %f116, 0f3EE8BA2E;
mov.b32 %r265, %f528;
setp.lt.s32 %p83, %r265, 2139095040;
@%p83 bra BB0_89;
setp.gtu.f32 %p84, %f116, 0f7F800000;
@%p84 bra BB0_88;
bra.uni BB0_86;
BB0_88:
add.f32 %f843, %f840, 0f3EE8BA2E;
bra.uni BB0_89;
BB0_86:
setp.neu.f32 %p85, %f116, 0f7F800000;
@%p85 bra BB0_89;
selp.f32 %f843, 0fFF800000, 0f7F800000, %p3;
BB0_89:
mul.f32 %f529, %f843, 0f437F0000;
setp.eq.f32 %p86, %f840, 0f3F800000;
selp.f32 %f530, 0f437F0000, %f529, %p86;
cvt.rzi.u32.f32 %r266, %f530;
cvt.u16.u32 %rs10, %r266;
mov.u16 %rs11, 255;
st.v2.u8 [%rd63], {%rs10, %rs11};
ld.global.u32 %r360, [imageEnabled];
BB0_90:
ld.global.f32 %f531, [lightColor];
mul.f32 %f128, %f839, %f531;
ld.global.f32 %f532, [lightColor+4];
mul.f32 %f129, %f839, %f532;
ld.global.f32 %f533, [lightColor+8];
mul.f32 %f130, %f839, %f533;
and.b32 %r267, %r360, 1;
setp.eq.b32 %p87, %r267, 1;
@!%p87 bra BB0_125;
bra.uni BB0_91;
BB0_91:
mov.f32 %f536, 0f3E666666;
cvt.rzi.f32.f32 %f537, %f536;
fma.rn.f32 %f538, %f537, 0fC0000000, 0f3EE66666;
abs.f32 %f131, %f538;
abs.f32 %f132, %f128;
setp.lt.f32 %p88, %f132, 0f00800000;
mul.f32 %f539, %f132, 0f4B800000;
selp.f32 %f540, 0fC3170000, 0fC2FE0000, %p88;
selp.f32 %f541, %f539, %f132, %p88;
mov.b32 %r268, %f541;
and.b32 %r269, %r268, 8388607;
or.b32 %r270, %r269, 1065353216;
mov.b32 %f542, %r270;
shr.u32 %r271, %r268, 23;
cvt.rn.f32.u32 %f543, %r271;
add.f32 %f544, %f540, %f543;
setp.gt.f32 %p89, %f542, 0f3FB504F3;
mul.f32 %f545, %f542, 0f3F000000;
add.f32 %f546, %f544, 0f3F800000;
selp.f32 %f547, %f545, %f542, %p89;
selp.f32 %f548, %f546, %f544, %p89;
add.f32 %f549, %f547, 0fBF800000;
add.f32 %f535, %f547, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f534,%f535;
// inline asm
add.f32 %f550, %f549, %f549;
mul.f32 %f551, %f534, %f550;
mul.f32 %f552, %f551, %f551;
mov.f32 %f553, 0f3C4CAF63;
mov.f32 %f554, 0f3B18F0FE;
fma.rn.f32 %f555, %f554, %f552, %f553;
mov.f32 %f556, 0f3DAAAABD;
fma.rn.f32 %f557, %f555, %f552, %f556;
mul.rn.f32 %f558, %f557, %f552;
mul.rn.f32 %f559, %f558, %f551;
sub.f32 %f560, %f549, %f551;
neg.f32 %f561, %f551;
add.f32 %f562, %f560, %f560;
fma.rn.f32 %f563, %f561, %f549, %f562;
mul.rn.f32 %f564, %f534, %f563;
add.f32 %f565, %f559, %f551;
sub.f32 %f566, %f551, %f565;
add.f32 %f567, %f559, %f566;
add.f32 %f568, %f564, %f567;
add.f32 %f569, %f565, %f568;
sub.f32 %f570, %f565, %f569;
add.f32 %f571, %f568, %f570;
mov.f32 %f572, 0f3F317200;
mul.rn.f32 %f573, %f548, %f572;
mov.f32 %f574, 0f35BFBE8E;
mul.rn.f32 %f575, %f548, %f574;
add.f32 %f576, %f573, %f569;
sub.f32 %f577, %f573, %f576;
add.f32 %f578, %f569, %f577;
add.f32 %f579, %f571, %f578;
add.f32 %f580, %f575, %f579;
add.f32 %f581, %f576, %f580;
sub.f32 %f582, %f576, %f581;
add.f32 %f583, %f580, %f582;
mov.f32 %f584, 0f3EE66666;
mul.rn.f32 %f585, %f584, %f581;
neg.f32 %f586, %f585;
fma.rn.f32 %f587, %f584, %f581, %f586;
fma.rn.f32 %f588, %f584, %f583, %f587;
mov.f32 %f589, 0f00000000;
fma.rn.f32 %f590, %f589, %f581, %f588;
add.rn.f32 %f591, %f585, %f590;
neg.f32 %f592, %f591;
add.rn.f32 %f593, %f585, %f592;
add.rn.f32 %f594, %f593, %f590;
mov.b32 %r272, %f591;
setp.eq.s32 %p90, %r272, 1118925336;
add.s32 %r273, %r272, -1;
mov.b32 %f595, %r273;
add.f32 %f596, %f594, 0f37000000;
selp.f32 %f597, %f595, %f591, %p90;
selp.f32 %f133, %f596, %f594, %p90;
mul.f32 %f598, %f597, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f599, %f598;
mov.f32 %f600, 0fBF317200;
fma.rn.f32 %f601, %f599, %f600, %f597;
mov.f32 %f602, 0fB5BFBE8E;
fma.rn.f32 %f603, %f599, %f602, %f601;
mul.f32 %f604, %f603, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f605, %f604;
add.f32 %f606, %f599, 0f00000000;
ex2.approx.f32 %f607, %f606;
mul.f32 %f608, %f605, %f607;
setp.lt.f32 %p91, %f597, 0fC2D20000;
selp.f32 %f609, 0f00000000, %f608, %p91;
setp.gt.f32 %p92, %f597, 0f42D20000;
selp.f32 %f844, 0f7F800000, %f609, %p92;
setp.eq.f32 %p93, %f844, 0f7F800000;
@%p93 bra BB0_93;
fma.rn.f32 %f844, %f844, %f133, %f844;
BB0_93:
setp.lt.f32 %p94, %f128, 0f00000000;
setp.eq.f32 %p95, %f131, 0f3F800000;
and.pred %p4, %p94, %p95;
mov.b32 %r274, %f844;
xor.b32 %r275, %r274, -2147483648;
mov.b32 %f610, %r275;
selp.f32 %f846, %f610, %f844, %p4;
setp.eq.f32 %p96, %f128, 0f00000000;
@%p96 bra BB0_96;
bra.uni BB0_94;
BB0_96:
add.f32 %f613, %f128, %f128;
selp.f32 %f846, %f613, 0f00000000, %p95;
bra.uni BB0_97;
BB0_94:
setp.geu.f32 %p97, %f128, 0f00000000;
@%p97 bra BB0_97;
cvt.rzi.f32.f32 %f612, %f584;
setp.neu.f32 %p98, %f612, 0f3EE66666;
selp.f32 %f846, 0f7FFFFFFF, %f846, %p98;
BB0_97:
add.f32 %f614, %f132, 0f3EE66666;
mov.b32 %r276, %f614;
setp.lt.s32 %p100, %r276, 2139095040;
@%p100 bra BB0_102;
setp.gtu.f32 %p101, %f132, 0f7F800000;
@%p101 bra BB0_101;
bra.uni BB0_99;
BB0_101:
add.f32 %f846, %f128, 0f3EE66666;
bra.uni BB0_102;
BB0_99:
setp.neu.f32 %p102, %f132, 0f7F800000;
@%p102 bra BB0_102;
selp.f32 %f846, 0fFF800000, 0f7F800000, %p4;
BB0_102:
setp.eq.f32 %p103, %f128, 0f3F800000;
selp.f32 %f144, 0f3F800000, %f846, %p103;
abs.f32 %f145, %f129;
setp.lt.f32 %p104, %f145, 0f00800000;
mul.f32 %f617, %f145, 0f4B800000;
selp.f32 %f618, 0fC3170000, 0fC2FE0000, %p104;
selp.f32 %f619, %f617, %f145, %p104;
mov.b32 %r277, %f619;
and.b32 %r278, %r277, 8388607;
or.b32 %r279, %r278, 1065353216;
mov.b32 %f620, %r279;
shr.u32 %r280, %r277, 23;
cvt.rn.f32.u32 %f621, %r280;
add.f32 %f622, %f618, %f621;
setp.gt.f32 %p105, %f620, 0f3FB504F3;
mul.f32 %f623, %f620, 0f3F000000;
add.f32 %f624, %f622, 0f3F800000;
selp.f32 %f625, %f623, %f620, %p105;
selp.f32 %f626, %f624, %f622, %p105;
add.f32 %f627, %f625, 0fBF800000;
add.f32 %f616, %f625, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f615,%f616;
// inline asm
add.f32 %f628, %f627, %f627;
mul.f32 %f629, %f615, %f628;
mul.f32 %f630, %f629, %f629;
fma.rn.f32 %f633, %f554, %f630, %f553;
fma.rn.f32 %f635, %f633, %f630, %f556;
mul.rn.f32 %f636, %f635, %f630;
mul.rn.f32 %f637, %f636, %f629;
sub.f32 %f638, %f627, %f629;
neg.f32 %f639, %f629;
add.f32 %f640, %f638, %f638;
fma.rn.f32 %f641, %f639, %f627, %f640;
mul.rn.f32 %f642, %f615, %f641;
add.f32 %f643, %f637, %f629;
sub.f32 %f644, %f629, %f643;
add.f32 %f645, %f637, %f644;
add.f32 %f646, %f642, %f645;
add.f32 %f647, %f643, %f646;
sub.f32 %f648, %f643, %f647;
add.f32 %f649, %f646, %f648;
mul.rn.f32 %f651, %f626, %f572;
mul.rn.f32 %f653, %f626, %f574;
add.f32 %f654, %f651, %f647;
sub.f32 %f655, %f651, %f654;
add.f32 %f656, %f647, %f655;
add.f32 %f657, %f649, %f656;
add.f32 %f658, %f653, %f657;
add.f32 %f659, %f654, %f658;
sub.f32 %f660, %f654, %f659;
add.f32 %f661, %f658, %f660;
mul.rn.f32 %f663, %f584, %f659;
neg.f32 %f664, %f663;
fma.rn.f32 %f665, %f584, %f659, %f664;
fma.rn.f32 %f666, %f584, %f661, %f665;
fma.rn.f32 %f668, %f589, %f659, %f666;
add.rn.f32 %f669, %f663, %f668;
neg.f32 %f670, %f669;
add.rn.f32 %f671, %f663, %f670;
add.rn.f32 %f672, %f671, %f668;
mov.b32 %r281, %f669;
setp.eq.s32 %p106, %r281, 1118925336;
add.s32 %r282, %r281, -1;
mov.b32 %f673, %r282;
add.f32 %f674, %f672, 0f37000000;
selp.f32 %f675, %f673, %f669, %p106;
selp.f32 %f146, %f674, %f672, %p106;
mul.f32 %f676, %f675, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f677, %f676;
fma.rn.f32 %f679, %f677, %f600, %f675;
fma.rn.f32 %f681, %f677, %f602, %f679;
mul.f32 %f682, %f681, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f683, %f682;
add.f32 %f684, %f677, 0f00000000;
ex2.approx.f32 %f685, %f684;
mul.f32 %f686, %f683, %f685;
setp.lt.f32 %p107, %f675, 0fC2D20000;
selp.f32 %f687, 0f00000000, %f686, %p107;
setp.gt.f32 %p108, %f675, 0f42D20000;
selp.f32 %f847, 0f7F800000, %f687, %p108;
setp.eq.f32 %p109, %f847, 0f7F800000;
@%p109 bra BB0_104;
fma.rn.f32 %f847, %f847, %f146, %f847;
BB0_104:
setp.lt.f32 %p110, %f129, 0f00000000;
and.pred %p5, %p110, %p95;
mov.b32 %r283, %f847;
xor.b32 %r284, %r283, -2147483648;
mov.b32 %f688, %r284;
selp.f32 %f849, %f688, %f847, %p5;
setp.eq.f32 %p112, %f129, 0f00000000;
@%p112 bra BB0_107;
bra.uni BB0_105;
BB0_107:
add.f32 %f691, %f129, %f129;
selp.f32 %f849, %f691, 0f00000000, %p95;
bra.uni BB0_108;
BB0_105:
setp.geu.f32 %p113, %f129, 0f00000000;
@%p113 bra BB0_108;
cvt.rzi.f32.f32 %f690, %f584;
setp.neu.f32 %p114, %f690, 0f3EE66666;
selp.f32 %f849, 0f7FFFFFFF, %f849, %p114;
BB0_108:
add.f32 %f692, %f145, 0f3EE66666;
mov.b32 %r285, %f692;
setp.lt.s32 %p116, %r285, 2139095040;
@%p116 bra BB0_113;
setp.gtu.f32 %p117, %f145, 0f7F800000;
@%p117 bra BB0_112;
bra.uni BB0_110;
BB0_112:
add.f32 %f849, %f129, 0f3EE66666;
bra.uni BB0_113;
BB0_110:
setp.neu.f32 %p118, %f145, 0f7F800000;
@%p118 bra BB0_113;
selp.f32 %f849, 0fFF800000, 0f7F800000, %p5;
BB0_113:
setp.eq.f32 %p119, %f129, 0f3F800000;
selp.f32 %f157, 0f3F800000, %f849, %p119;
abs.f32 %f158, %f130;
setp.lt.f32 %p120, %f158, 0f00800000;
mul.f32 %f695, %f158, 0f4B800000;
selp.f32 %f696, 0fC3170000, 0fC2FE0000, %p120;
selp.f32 %f697, %f695, %f158, %p120;
mov.b32 %r286, %f697;
and.b32 %r287, %r286, 8388607;
or.b32 %r288, %r287, 1065353216;
mov.b32 %f698, %r288;
shr.u32 %r289, %r286, 23;
cvt.rn.f32.u32 %f699, %r289;
add.f32 %f700, %f696, %f699;
setp.gt.f32 %p121, %f698, 0f3FB504F3;
mul.f32 %f701, %f698, 0f3F000000;
add.f32 %f702, %f700, 0f3F800000;
selp.f32 %f703, %f701, %f698, %p121;
selp.f32 %f704, %f702, %f700, %p121;
add.f32 %f705, %f703, 0fBF800000;
add.f32 %f694, %f703, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f693,%f694;
// inline asm
add.f32 %f706, %f705, %f705;
mul.f32 %f707, %f693, %f706;
mul.f32 %f708, %f707, %f707;
fma.rn.f32 %f711, %f554, %f708, %f553;
fma.rn.f32 %f713, %f711, %f708, %f556;
mul.rn.f32 %f714, %f713, %f708;
mul.rn.f32 %f715, %f714, %f707;
sub.f32 %f716, %f705, %f707;
neg.f32 %f717, %f707;
add.f32 %f718, %f716, %f716;
fma.rn.f32 %f719, %f717, %f705, %f718;
mul.rn.f32 %f720, %f693, %f719;
add.f32 %f721, %f715, %f707;
sub.f32 %f722, %f707, %f721;
add.f32 %f723, %f715, %f722;
add.f32 %f724, %f720, %f723;
add.f32 %f725, %f721, %f724;
sub.f32 %f726, %f721, %f725;
add.f32 %f727, %f724, %f726;
mul.rn.f32 %f729, %f704, %f572;
mul.rn.f32 %f731, %f704, %f574;
add.f32 %f732, %f729, %f725;
sub.f32 %f733, %f729, %f732;
add.f32 %f734, %f725, %f733;
add.f32 %f735, %f727, %f734;
add.f32 %f736, %f731, %f735;
add.f32 %f737, %f732, %f736;
sub.f32 %f738, %f732, %f737;
add.f32 %f739, %f736, %f738;
mul.rn.f32 %f741, %f584, %f737;
neg.f32 %f742, %f741;
fma.rn.f32 %f743, %f584, %f737, %f742;
fma.rn.f32 %f744, %f584, %f739, %f743;
fma.rn.f32 %f746, %f589, %f737, %f744;
add.rn.f32 %f747, %f741, %f746;
neg.f32 %f748, %f747;
add.rn.f32 %f749, %f741, %f748;
add.rn.f32 %f750, %f749, %f746;
mov.b32 %r290, %f747;
setp.eq.s32 %p122, %r290, 1118925336;
add.s32 %r291, %r290, -1;
mov.b32 %f751, %r291;
add.f32 %f752, %f750, 0f37000000;
selp.f32 %f753, %f751, %f747, %p122;
selp.f32 %f159, %f752, %f750, %p122;
mul.f32 %f754, %f753, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f755, %f754;
fma.rn.f32 %f757, %f755, %f600, %f753;
fma.rn.f32 %f759, %f755, %f602, %f757;
mul.f32 %f760, %f759, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f761, %f760;
add.f32 %f762, %f755, 0f00000000;
ex2.approx.f32 %f763, %f762;
mul.f32 %f764, %f761, %f763;
setp.lt.f32 %p123, %f753, 0fC2D20000;
selp.f32 %f765, 0f00000000, %f764, %p123;
setp.gt.f32 %p124, %f753, 0f42D20000;
selp.f32 %f850, 0f7F800000, %f765, %p124;
setp.eq.f32 %p125, %f850, 0f7F800000;
@%p125 bra BB0_115;
fma.rn.f32 %f850, %f850, %f159, %f850;
BB0_115:
setp.lt.f32 %p126, %f130, 0f00000000;
and.pred %p6, %p126, %p95;
mov.b32 %r292, %f850;
xor.b32 %r293, %r292, -2147483648;
mov.b32 %f766, %r293;
selp.f32 %f852, %f766, %f850, %p6;
setp.eq.f32 %p128, %f130, 0f00000000;
@%p128 bra BB0_118;
bra.uni BB0_116;
BB0_118:
add.f32 %f769, %f130, %f130;
selp.f32 %f852, %f769, 0f00000000, %p95;
bra.uni BB0_119;
BB0_116:
setp.geu.f32 %p129, %f130, 0f00000000;
@%p129 bra BB0_119;
cvt.rzi.f32.f32 %f768, %f584;
setp.neu.f32 %p130, %f768, 0f3EE66666;
selp.f32 %f852, 0f7FFFFFFF, %f852, %p130;
BB0_119:
add.f32 %f770, %f158, 0f3EE66666;
mov.b32 %r294, %f770;
setp.lt.s32 %p132, %r294, 2139095040;
@%p132 bra BB0_124;
setp.gtu.f32 %p133, %f158, 0f7F800000;
@%p133 bra BB0_123;
bra.uni BB0_121;
BB0_123:
add.f32 %f852, %f130, 0f3EE66666;
bra.uni BB0_124;
BB0_121:
setp.neu.f32 %p134, %f158, 0f7F800000;
@%p134 bra BB0_124;
selp.f32 %f852, 0fFF800000, 0f7F800000, %p6;
BB0_124:
mov.u32 %r326, 2;
setp.eq.f32 %p135, %f130, 0f3F800000;
selp.f32 %f771, 0f3F800000, %f852, %p135;
cvt.u64.u32 %rd73, %r4;
cvt.u64.u32 %rd72, %r3;
mov.u64 %rd76, image;
cvta.global.u64 %rd71, %rd76;
// inline asm
call (%rd70), _rt_buffer_get_64, (%rd71, %r326, %r102, %rd72, %rd73, %rd24, %rd24);
// inline asm
cvt.sat.f32.f32 %f772, %f771;
mul.f32 %f773, %f772, 0f437FFD71;
cvt.rzi.u32.f32 %r297, %f773;
cvt.sat.f32.f32 %f774, %f157;
mul.f32 %f775, %f774, 0f437FFD71;
cvt.rzi.u32.f32 %r298, %f775;
cvt.sat.f32.f32 %f776, %f144;
mul.f32 %f777, %f776, 0f437FFD71;
cvt.rzi.u32.f32 %r299, %f777;
cvt.u16.u32 %rs12, %r297;
cvt.u16.u32 %rs13, %r299;
cvt.u16.u32 %rs14, %r298;
mov.u16 %rs15, 255;
st.v4.u8 [%rd70], {%rs12, %rs14, %rs13, %rs15};
ld.global.u32 %r360, [imageEnabled];
BB0_125:
and.b32 %r300, %r360, 4;
setp.eq.s32 %p136, %r300, 0;
@%p136 bra BB0_137;
ld.global.u32 %r301, [additive];
setp.eq.s32 %p137, %r301, 0;
cvt.u64.u32 %rd15, %r3;
cvt.u64.u32 %rd16, %r4;
// inline asm
{ cvt.rn.f16.f32 %rs16, %f338;}
// inline asm
@%p137 bra BB0_128;
mov.u32 %r327, 2;
mov.u64 %rd89, image_HDR;
cvta.global.u64 %rd78, %rd89;
mov.u32 %r305, 8;
// inline asm
call (%rd77), _rt_buffer_get_64, (%rd78, %r327, %r305, %rd15, %rd16, %rd24, %rd24);
// inline asm
ld.v4.u16 {%rs23, %rs24, %rs25, %rs26}, [%rd77];
// inline asm
{ cvt.f32.f16 %f779, %rs23;}
// inline asm
// inline asm
{ cvt.f32.f16 %f780, %rs24;}
// inline asm
// inline asm
{ cvt.f32.f16 %f781, %rs25;}
// inline asm
// inline asm
call (%rd83), _rt_buffer_get_64, (%rd78, %r327, %r305, %rd15, %rd16, %rd24, %rd24);
// inline asm
add.f32 %f782, %f128, %f779;
add.f32 %f783, %f129, %f780;
add.f32 %f784, %f130, %f781;
// inline asm
{ cvt.rn.f16.f32 %rs22, %f784;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs21, %f783;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs20, %f782;}
// inline asm
st.v4.u16 [%rd83], {%rs20, %rs21, %rs22, %rs16};
bra.uni BB0_137;
BB0_128:
mov.u32 %r328, 2;
mov.u64 %rd96, image_HDR;
cvta.global.u64 %rd91, %rd96;
mov.u32 %r307, 8;
// inline asm
call (%rd90), _rt_buffer_get_64, (%rd91, %r328, %r307, %rd15, %rd16, %rd24, %rd24);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs29, %f130;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs28, %f129;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs27, %f128;}
// inline asm
st.v4.u16 [%rd90], {%rs27, %rs28, %rs29, %rs16};
BB0_137:
ret;
}