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

2596 lines
74 KiB
Plaintext

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-23083092
// Cuda compilation tools, release 9.1, V9.1.85
// Based on LLVM 3.4svn
//
.version 6.1
.target sm_30
.address_size 64
// .globl _Z6oxMainv
.global .align 8 .b8 pixelID[8];
.global .align 8 .b8 resolution[8];
.global .align 4 .b8 normal[12];
.global .align 4 .b8 camPos[12];
.global .align 4 .b8 root[4];
.global .align 4 .u32 imageEnabled;
.global .texref lightmap;
.global .align 16 .b8 tileInfo[16];
.global .align 4 .u32 additive;
.global .align 1 .b8 image[1];
.global .align 1 .b8 image_HDR[1];
.global .align 1 .b8 image_HDR2[1];
.global .align 1 .b8 image_Mask[1];
.global .align 1 .b8 image_RNM0[1];
.global .align 1 .b8 image_RNM1[1];
.global .align 1 .b8 image_RNM2[1];
.global .align 1 .b8 uvtangent[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<159>;
.reg .b16 %rs<158>;
.reg .f32 %f<1404>;
.reg .b32 %r<429>;
.reg .b64 %rd<305>;
mov.u64 %rd304, __local_depot0;
cvta.local.u64 %SP, %rd304;
ld.global.u32 %r1, [lightMeshBufferSize];
abs.s32 %r2, %r1;
ld.global.u32 %r3, [samples];
ld.global.v2.u32 {%r111, %r112}, [pixelID];
cvt.u64.u32 %rd28, %r111;
cvt.u64.u32 %rd29, %r112;
mov.u64 %rd32, uvnormal;
cvta.global.u64 %rd27, %rd32;
mov.u32 %r109, 2;
mov.u32 %r110, 4;
mov.u64 %rd31, 0;
// inline asm
call (%rd26), _rt_buffer_get_64, (%rd27, %r109, %r110, %rd28, %rd29, %rd31, %rd31);
// inline asm
ld.u32 %r4, [%rd26];
shr.u32 %r115, %r4, 16;
cvt.u16.u32 %rs1, %r115;
and.b16 %rs6, %rs1, 255;
cvt.u16.u32 %rs7, %r4;
or.b16 %rs8, %rs7, %rs6;
setp.eq.s16 %p7, %rs8, 0;
mov.f32 %f1258, 0f00000000;
mov.f32 %f1259, %f1258;
mov.f32 %f1260, %f1258;
@%p7 bra BB0_2;
ld.u8 %rs9, [%rd26+1];
and.b16 %rs11, %rs7, 255;
cvt.rn.f32.u16 %f354, %rs11;
div.rn.f32 %f355, %f354, 0f437F0000;
fma.rn.f32 %f356, %f355, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f357, %rs9;
div.rn.f32 %f358, %f357, 0f437F0000;
fma.rn.f32 %f359, %f358, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f360, %rs6;
div.rn.f32 %f361, %f360, 0f437F0000;
fma.rn.f32 %f362, %f361, 0f40000000, 0fBF800000;
mul.f32 %f363, %f359, %f359;
fma.rn.f32 %f364, %f356, %f356, %f363;
fma.rn.f32 %f365, %f362, %f362, %f364;
sqrt.rn.f32 %f366, %f365;
rcp.rn.f32 %f367, %f366;
mul.f32 %f1258, %f356, %f367;
mul.f32 %f1259, %f359, %f367;
mul.f32 %f1260, %f362, %f367;
BB0_2:
ld.global.v2.u32 {%r116, %r117}, [pixelID];
ld.global.v2.u32 {%r119, %r120}, [tileInfo];
add.s32 %r5, %r116, %r119;
add.s32 %r6, %r117, %r120;
setp.eq.f32 %p8, %f1259, 0f00000000;
setp.eq.f32 %p9, %f1258, 0f00000000;
and.pred %p10, %p9, %p8;
setp.eq.f32 %p11, %f1260, 0f00000000;
and.pred %p12, %p10, %p11;
@%p12 bra BB0_147;
bra.uni BB0_3;
BB0_147:
ld.global.u32 %r428, [imageEnabled];
and.b32 %r357, %r428, 1;
setp.eq.b32 %p151, %r357, 1;
@!%p151 bra BB0_149;
bra.uni BB0_148;
BB0_148:
cvt.u64.u32 %rd204, %r5;
cvt.u64.u32 %rd205, %r6;
mov.u64 %rd208, image;
cvta.global.u64 %rd203, %rd208;
mov.u64 %rd207, 0;
// inline asm
call (%rd202), _rt_buffer_get_64, (%rd203, %r109, %r110, %rd204, %rd205, %rd207, %rd207);
// inline asm
mov.u16 %rs103, 0;
st.v4.u8 [%rd202], {%rs103, %rs103, %rs103, %rs103};
ld.global.u32 %r428, [imageEnabled];
BB0_149:
and.b32 %r360, %r428, 8;
setp.eq.s32 %p152, %r360, 0;
@%p152 bra BB0_151;
cvt.u64.u32 %rd212, %r6;
cvt.u64.u32 %rd211, %r5;
mov.u64 %rd215, image_Mask;
cvta.global.u64 %rd210, %rd215;
mov.u64 %rd214, 0;
// inline asm
call (%rd209), _rt_buffer_get_64, (%rd210, %r109, %r109, %rd211, %rd212, %rd214, %rd214);
// inline asm
mov.f32 %f1220, 0f00000000;
cvt.rzi.u32.f32 %r363, %f1220;
cvt.u16.u32 %rs104, %r363;
mov.u16 %rs105, 0;
st.v2.u8 [%rd209], {%rs104, %rs105};
ld.global.u32 %r428, [imageEnabled];
BB0_151:
cvt.u64.u32 %rd24, %r5;
cvt.u64.u32 %rd25, %r6;
and.b32 %r364, %r428, 4;
setp.eq.s32 %p153, %r364, 0;
@%p153 bra BB0_155;
ld.global.u32 %r365, [additive];
setp.eq.s32 %p154, %r365, 0;
@%p154 bra BB0_154;
mov.u64 %rd228, image_HDR;
cvta.global.u64 %rd217, %rd228;
mov.u32 %r369, 8;
mov.u64 %rd227, 0;
// inline asm
call (%rd216), _rt_buffer_get_64, (%rd217, %r109, %r369, %rd24, %rd25, %rd227, %rd227);
// inline asm
ld.v4.u16 {%rs112, %rs113, %rs114, %rs115}, [%rd216];
// inline asm
{ cvt.f32.f16 %f1221, %rs112;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1222, %rs113;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1223, %rs114;}
// inline asm
// inline asm
call (%rd222), _rt_buffer_get_64, (%rd217, %r109, %r369, %rd24, %rd25, %rd227, %rd227);
// inline asm
add.f32 %f1224, %f1221, 0f00000000;
add.f32 %f1225, %f1222, 0f00000000;
add.f32 %f1226, %f1223, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs111, %f1226;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs110, %f1225;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs109, %f1224;}
// inline asm
mov.u16 %rs116, 0;
st.v4.u16 [%rd222], {%rs109, %rs110, %rs111, %rs116};
bra.uni BB0_155;
BB0_3:
ld.global.v2.u32 {%r128, %r129}, [pixelID];
cvt.u64.u32 %rd35, %r128;
cvt.u64.u32 %rd36, %r129;
mov.u64 %rd45, uvpos;
cvta.global.u64 %rd34, %rd45;
mov.u32 %r125, 12;
// inline asm
call (%rd33), _rt_buffer_get_64, (%rd34, %r109, %r125, %rd35, %rd36, %rd31, %rd31);
// inline asm
ld.f32 %f9, [%rd33+8];
ld.f32 %f8, [%rd33+4];
ld.f32 %f7, [%rd33];
mul.f32 %f371, %f7, 0f3456BF95;
mul.f32 %f372, %f8, 0f3456BF95;
mul.f32 %f373, %f9, 0f3456BF95;
abs.f32 %f10, %f1258;
div.rn.f32 %f374, %f371, %f10;
abs.f32 %f375, %f1259;
div.rn.f32 %f376, %f372, %f375;
abs.f32 %f11, %f1260;
div.rn.f32 %f377, %f373, %f11;
abs.f32 %f378, %f374;
abs.f32 %f379, %f376;
abs.f32 %f380, %f377;
mov.f32 %f381, 0f38D1B717;
max.f32 %f382, %f378, %f381;
max.f32 %f383, %f379, %f381;
max.f32 %f384, %f380, %f381;
fma.rn.f32 %f12, %f1258, %f382, %f7;
fma.rn.f32 %f13, %f1259, %f383, %f8;
fma.rn.f32 %f14, %f1260, %f384, %f9;
ld.global.v2.u32 {%r132, %r133}, [pixelID];
cvt.u64.u32 %rd41, %r132;
cvt.u64.u32 %rd42, %r133;
mov.u64 %rd46, uvtangent;
cvta.global.u64 %rd40, %rd46;
// inline asm
call (%rd39), _rt_buffer_get_64, (%rd40, %r109, %r110, %rd41, %rd42, %rd31, %rd31);
// inline asm
ld.u32 %r7, [%rd39];
shr.u32 %r8, %r7, 16;
cvt.u16.u32 %rs13, %r8;
and.b16 %rs14, %rs13, 255;
cvt.u16.u32 %rs15, %r7;
or.b16 %rs16, %rs15, %rs14;
setp.eq.s16 %p13, %rs16, 0;
mov.f32 %f49, 0f00000000;
mov.f32 %f1261, %f49;
mov.f32 %f1262, %f49;
mov.f32 %f1263, %f49;
@%p13 bra BB0_5;
ld.u8 %rs17, [%rd39+1];
and.b16 %rs19, %rs15, 255;
cvt.rn.f32.u16 %f385, %rs19;
div.rn.f32 %f386, %f385, 0f437F0000;
fma.rn.f32 %f387, %f386, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f388, %rs17;
div.rn.f32 %f389, %f388, 0f437F0000;
fma.rn.f32 %f390, %f389, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f391, %rs14;
div.rn.f32 %f392, %f391, 0f437F0000;
fma.rn.f32 %f393, %f392, 0f40000000, 0fBF800000;
mul.f32 %f394, %f390, %f390;
fma.rn.f32 %f395, %f387, %f387, %f394;
fma.rn.f32 %f396, %f393, %f393, %f395;
sqrt.rn.f32 %f397, %f396;
rcp.rn.f32 %f398, %f397;
mul.f32 %f1261, %f387, %f398;
mul.f32 %f1262, %f390, %f398;
mul.f32 %f1263, %f393, %f398;
BB0_5:
mul.f32 %f402, %f1260, %f1262;
mul.f32 %f403, %f1259, %f1263;
sub.f32 %f404, %f403, %f402;
mul.f32 %f405, %f1258, %f1263;
mul.f32 %f406, %f1260, %f1261;
sub.f32 %f407, %f406, %f405;
mul.f32 %f408, %f1259, %f1261;
mul.f32 %f409, %f1258, %f1262;
sub.f32 %f410, %f409, %f408;
setp.lt.u32 %p14, %r7, 16777216;
selp.f32 %f411, 0fBF800000, 0f3F800000, %p14;
mul.f32 %f412, %f404, %f411;
mul.f32 %f413, %f407, %f411;
mul.f32 %f414, %f410, %f411;
mul.f32 %f415, %f412, 0f00000000;
mul.f32 %f416, %f413, 0f00000000;
mul.f32 %f417, %f414, 0f00000000;
fma.rn.f32 %f418, %f1261, 0f3F5105EC, %f415;
fma.rn.f32 %f419, %f1262, 0f3F5105EC, %f416;
fma.rn.f32 %f420, %f1263, 0f3F5105EC, %f417;
mul.f32 %f21, %f1258, 0f3F13CD3A;
add.f32 %f22, %f21, %f418;
mul.f32 %f23, %f1259, 0f3F13CD3A;
add.f32 %f24, %f23, %f419;
mul.f32 %f25, %f1260, 0f3F13CD3A;
add.f32 %f26, %f25, %f420;
ld.global.v2.u32 {%r138, %r139}, [pixelID];
cvt.u64.u32 %rd49, %r138;
cvt.u64.u32 %rd50, %r139;
// inline asm
call (%rd47), _rt_buffer_get_64, (%rd40, %r109, %r110, %rd49, %rd50, %rd31, %rd31);
// inline asm
ld.u32 %r9, [%rd47];
shr.u32 %r10, %r9, 16;
cvt.u16.u32 %rs22, %r10;
and.b16 %rs23, %rs22, 255;
cvt.u16.u32 %rs24, %r9;
or.b16 %rs25, %rs24, %rs23;
setp.eq.s16 %p15, %rs25, 0;
mov.f32 %f1264, %f49;
mov.f32 %f1265, %f49;
mov.f32 %f1266, %f49;
@%p15 bra BB0_7;
ld.u8 %rs26, [%rd47+1];
and.b16 %rs28, %rs24, 255;
cvt.rn.f32.u16 %f421, %rs28;
div.rn.f32 %f422, %f421, 0f437F0000;
fma.rn.f32 %f423, %f422, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f424, %rs26;
div.rn.f32 %f425, %f424, 0f437F0000;
fma.rn.f32 %f426, %f425, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f427, %rs23;
div.rn.f32 %f428, %f427, 0f437F0000;
fma.rn.f32 %f429, %f428, 0f40000000, 0fBF800000;
mul.f32 %f430, %f426, %f426;
fma.rn.f32 %f431, %f423, %f423, %f430;
fma.rn.f32 %f432, %f429, %f429, %f431;
sqrt.rn.f32 %f433, %f432;
rcp.rn.f32 %f434, %f433;
mul.f32 %f1264, %f423, %f434;
mul.f32 %f1265, %f426, %f434;
mul.f32 %f1266, %f429, %f434;
BB0_7:
mul.f32 %f438, %f1260, %f1265;
mul.f32 %f439, %f1259, %f1266;
sub.f32 %f440, %f439, %f438;
mul.f32 %f441, %f1258, %f1266;
mul.f32 %f442, %f1260, %f1264;
sub.f32 %f443, %f442, %f441;
mul.f32 %f444, %f1259, %f1264;
mul.f32 %f445, %f1258, %f1265;
sub.f32 %f446, %f445, %f444;
setp.lt.u32 %p16, %r9, 16777216;
selp.f32 %f447, 0fBF800000, 0f3F800000, %p16;
mul.f32 %f448, %f440, %f447;
mul.f32 %f449, %f443, %f447;
mul.f32 %f450, %f446, %f447;
mul.f32 %f451, %f448, 0f3F3504F3;
mul.f32 %f452, %f449, 0f3F3504F3;
mul.f32 %f453, %f450, 0f3F3504F3;
fma.rn.f32 %f454, %f1264, 0fBED105EC, %f451;
fma.rn.f32 %f455, %f1265, 0fBED105EC, %f452;
fma.rn.f32 %f456, %f1266, 0fBED105EC, %f453;
add.f32 %f33, %f21, %f454;
add.f32 %f34, %f23, %f455;
add.f32 %f35, %f25, %f456;
ld.global.v2.u32 {%r144, %r145}, [pixelID];
cvt.u64.u32 %rd56, %r144;
cvt.u64.u32 %rd57, %r145;
// inline asm
call (%rd54), _rt_buffer_get_64, (%rd40, %r109, %r110, %rd56, %rd57, %rd31, %rd31);
// inline asm
ld.u32 %r11, [%rd54];
shr.u32 %r12, %r11, 16;
cvt.u16.u32 %rs31, %r12;
and.b16 %rs32, %rs31, 255;
cvt.u16.u32 %rs33, %r11;
or.b16 %rs34, %rs33, %rs32;
setp.eq.s16 %p17, %rs34, 0;
mov.f32 %f1267, %f49;
mov.f32 %f1268, %f49;
mov.f32 %f1269, %f49;
@%p17 bra BB0_9;
ld.u8 %rs35, [%rd54+1];
and.b16 %rs37, %rs33, 255;
cvt.rn.f32.u16 %f457, %rs37;
div.rn.f32 %f458, %f457, 0f437F0000;
fma.rn.f32 %f459, %f458, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f460, %rs35;
div.rn.f32 %f461, %f460, 0f437F0000;
fma.rn.f32 %f462, %f461, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f463, %rs32;
div.rn.f32 %f464, %f463, 0f437F0000;
fma.rn.f32 %f465, %f464, 0f40000000, 0fBF800000;
mul.f32 %f466, %f462, %f462;
fma.rn.f32 %f467, %f459, %f459, %f466;
fma.rn.f32 %f468, %f465, %f465, %f467;
sqrt.rn.f32 %f469, %f468;
rcp.rn.f32 %f470, %f469;
mul.f32 %f1267, %f459, %f470;
mul.f32 %f1268, %f462, %f470;
mul.f32 %f1269, %f465, %f470;
BB0_9:
mul.f32 %f484, %f1260, %f1268;
mul.f32 %f485, %f1259, %f1269;
sub.f32 %f486, %f485, %f484;
mul.f32 %f487, %f1258, %f1269;
mul.f32 %f488, %f1260, %f1267;
sub.f32 %f489, %f488, %f487;
mul.f32 %f490, %f1259, %f1267;
mul.f32 %f491, %f1258, %f1268;
sub.f32 %f492, %f491, %f490;
setp.lt.u32 %p19, %r11, 16777216;
selp.f32 %f493, 0fBF800000, 0f3F800000, %p19;
mul.f32 %f494, %f486, %f493;
mul.f32 %f495, %f489, %f493;
mul.f32 %f496, %f492, %f493;
mul.f32 %f497, %f494, 0fBF3504F3;
mul.f32 %f498, %f495, 0fBF3504F3;
mul.f32 %f499, %f496, 0fBF3504F3;
fma.rn.f32 %f500, %f1267, 0fBED105EC, %f497;
fma.rn.f32 %f501, %f1268, 0fBED105EC, %f498;
fma.rn.f32 %f502, %f1269, 0fBED105EC, %f499;
add.f32 %f42, %f21, %f500;
add.f32 %f43, %f23, %f501;
add.f32 %f44, %f25, %f502;
ld.global.v2.u32 {%r151, %r152}, [pixelID];
cvt.u64.u32 %rd63, %r151;
cvt.u64.u32 %rd64, %r152;
mov.u64 %rd67, rnd_seeds;
cvta.global.u64 %rd62, %rd67;
// inline asm
call (%rd61), _rt_buffer_get_64, (%rd62, %r109, %r110, %rd63, %rd64, %rd31, %rd31);
// inline asm
ld.u32 %r404, [%rd61];
mov.pred %p18, 0;
mov.u32 %r15, 0;
setp.lt.s32 %p20, %r2, 1;
@%p20 bra BB0_10;
ld.global.f32 %f45, [lightPointSize];
mul.f32 %f46, %f12, 0f3456BF95;
mul.f32 %f47, %f13, 0f3456BF95;
mul.f32 %f48, %f14, 0f3456BF95;
mov.f32 %f49, 0f00000000;
mov.u32 %r397, 0;
abs.f32 %f639, %f47;
abs.f32 %f640, %f46;
max.f32 %f641, %f640, %f639;
abs.f32 %f642, %f48;
max.f32 %f643, %f641, %f642;
mov.u32 %r15, %r397;
mov.f32 %f50, %f49;
mov.f32 %f51, %f49;
mov.f32 %f52, %f49;
mov.f32 %f53, %f49;
mov.f32 %f54, %f49;
mov.f32 %f55, %f49;
mov.f32 %f56, %f49;
mov.f32 %f57, %f49;
mov.f32 %f58, %f49;
mov.f32 %f59, %f49;
mov.f32 %f60, %f49;
mov.f32 %f61, %f49;
BB0_12:
shl.b32 %r16, %r397, 1;
cvt.s64.s32 %rd70, %r16;
mov.u64 %rd74, lightMeshBuffer;
cvta.global.u64 %rd69, %rd74;
mov.u32 %r157, 1;
mov.u64 %rd73, 0;
// inline asm
call (%rd68), _rt_buffer_get_64, (%rd69, %r157, %r125, %rd70, %rd73, %rd73, %rd73);
// inline asm
ld.f32 %f516, [%rd68];
sub.f32 %f517, %f516, %f7;
ld.f32 %f518, [%rd68+4];
sub.f32 %f519, %f518, %f8;
ld.f32 %f520, [%rd68+8];
sub.f32 %f521, %f520, %f9;
mul.f32 %f522, %f519, %f519;
fma.rn.f32 %f523, %f517, %f517, %f522;
fma.rn.f32 %f524, %f521, %f521, %f523;
sqrt.rn.f32 %f62, %f524;
rcp.rn.f32 %f525, %f62;
mul.f32 %f63, %f517, %f525;
mul.f32 %f64, %f519, %f525;
mul.f32 %f65, %f521, %f525;
mul.f32 %f526, %f1259, %f64;
fma.rn.f32 %f527, %f1258, %f63, %f526;
fma.rn.f32 %f66, %f1260, %f65, %f527;
setp.leu.f32 %p21, %f66, 0f00000000;
@%p21 bra BB0_29;
mov.u64 %rd299, 0;
setp.eq.s32 %p22, %r3, 0;
mul.f32 %f528, %f62, %f62;
div.rn.f32 %f67, %f45, %f528;
ld.global.v2.u32 {%r161, %r162}, [pixelID];
cvt.u64.u32 %rd81, %r162;
cvt.u64.u32 %rd82, %r161;
and.b64 %rd77, %rd82, 255;
and.b64 %rd78, %rd81, 255;
mov.u64 %rd83, rnd_seeds2;
cvta.global.u64 %rd76, %rd83;
mov.u32 %r160, 3;
// inline asm
call (%rd75), _rt_buffer_get_64, (%rd76, %r109, %r160, %rd77, %rd78, %rd299, %rd299);
// inline asm
@%p22 bra BB0_15;
ld.u8 %rs40, [%rd75];
cvt.rn.f32.u16 %f529, %rs40;
div.rn.f32 %f530, %f529, 0fC37F0000;
fma.rn.f32 %f531, %f530, 0f3F333333, 0f3F800000;
setp.gt.f32 %p24, %f67, %f531;
mov.pred %p158, -1;
@%p24 bra BB0_31;
BB0_15:
ld.global.f32 %f534, [lightInvCutoff];
mul.f32 %f68, %f62, %f534;
mov.f32 %f538, 0f40800000;
abs.f32 %f70, %f68;
setp.lt.f32 %p25, %f70, 0f00800000;
mul.f32 %f540, %f70, 0f4B800000;
selp.f32 %f541, 0fC3170000, 0fC2FE0000, %p25;
selp.f32 %f542, %f540, %f70, %p25;
mov.b32 %r165, %f542;
and.b32 %r166, %r165, 8388607;
or.b32 %r167, %r166, 1065353216;
mov.b32 %f543, %r167;
shr.u32 %r168, %r165, 23;
cvt.rn.f32.u32 %f544, %r168;
add.f32 %f545, %f541, %f544;
setp.gt.f32 %p26, %f543, 0f3FB504F3;
mul.f32 %f546, %f543, 0f3F000000;
add.f32 %f547, %f545, 0f3F800000;
selp.f32 %f548, %f546, %f543, %p26;
selp.f32 %f549, %f547, %f545, %p26;
add.f32 %f550, %f548, 0fBF800000;
add.f32 %f533, %f548, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f532,%f533;
// inline asm
add.f32 %f551, %f550, %f550;
mul.f32 %f552, %f532, %f551;
mul.f32 %f553, %f552, %f552;
mov.f32 %f554, 0f3C4CAF63;
mov.f32 %f555, 0f3B18F0FE;
fma.rn.f32 %f556, %f555, %f553, %f554;
mov.f32 %f557, 0f3DAAAABD;
fma.rn.f32 %f558, %f556, %f553, %f557;
mul.rn.f32 %f559, %f558, %f553;
mul.rn.f32 %f560, %f559, %f552;
sub.f32 %f561, %f550, %f552;
neg.f32 %f562, %f552;
add.f32 %f563, %f561, %f561;
fma.rn.f32 %f564, %f562, %f550, %f563;
mul.rn.f32 %f565, %f532, %f564;
add.f32 %f566, %f560, %f552;
sub.f32 %f567, %f552, %f566;
add.f32 %f568, %f560, %f567;
add.f32 %f569, %f565, %f568;
add.f32 %f570, %f566, %f569;
sub.f32 %f571, %f566, %f570;
add.f32 %f572, %f569, %f571;
mov.f32 %f573, 0f3F317200;
mul.rn.f32 %f574, %f549, %f573;
mov.f32 %f575, 0f35BFBE8E;
mul.rn.f32 %f576, %f549, %f575;
add.f32 %f577, %f574, %f570;
sub.f32 %f578, %f574, %f577;
add.f32 %f579, %f570, %f578;
add.f32 %f580, %f572, %f579;
add.f32 %f581, %f576, %f580;
add.f32 %f582, %f577, %f581;
sub.f32 %f583, %f577, %f582;
add.f32 %f584, %f581, %f583;
mul.rn.f32 %f585, %f538, %f582;
neg.f32 %f586, %f585;
fma.rn.f32 %f587, %f538, %f582, %f586;
fma.rn.f32 %f588, %f538, %f584, %f587;
mov.f32 %f589, 0f00000000;
fma.rn.f32 %f590, %f589, %f582, %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 %r169, %f591;
setp.eq.s32 %p27, %r169, 1118925336;
add.s32 %r170, %r169, -1;
mov.b32 %f595, %r170;
add.f32 %f596, %f594, 0f37000000;
selp.f32 %f597, %f595, %f591, %p27;
selp.f32 %f71, %f596, %f594, %p27;
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 %p28, %f597, 0fC2D20000;
selp.f32 %f609, 0f00000000, %f608, %p28;
setp.gt.f32 %p29, %f597, 0f42D20000;
selp.f32 %f1283, 0f7F800000, %f609, %p29;
setp.eq.f32 %p30, %f1283, 0f7F800000;
@%p30 bra BB0_17;
fma.rn.f32 %f1283, %f1283, %f71, %f1283;
BB0_17:
mov.f32 %f1254, 0f40000000;
cvt.rzi.f32.f32 %f1253, %f1254;
add.f32 %f1252, %f1253, %f1253;
mov.f32 %f1251, 0f40800000;
sub.f32 %f1250, %f1251, %f1252;
abs.f32 %f1249, %f1250;
setp.lt.f32 %p31, %f68, 0f00000000;
setp.eq.f32 %p32, %f1249, 0f3F800000;
and.pred %p1, %p31, %p32;
mov.b32 %r171, %f1283;
xor.b32 %r172, %r171, -2147483648;
mov.b32 %f610, %r172;
selp.f32 %f1285, %f610, %f1283, %p1;
setp.eq.f32 %p33, %f68, 0f00000000;
@%p33 bra BB0_20;
bra.uni BB0_18;
BB0_20:
add.f32 %f613, %f68, %f68;
selp.f32 %f1285, %f613, 0f00000000, %p32;
bra.uni BB0_21;
BB0_18:
setp.geu.f32 %p34, %f68, 0f00000000;
@%p34 bra BB0_21;
mov.f32 %f1257, 0f40800000;
cvt.rzi.f32.f32 %f612, %f1257;
setp.neu.f32 %p35, %f612, 0f40800000;
selp.f32 %f1285, 0f7FFFFFFF, %f1285, %p35;
BB0_21:
add.f32 %f614, %f70, 0f40800000;
mov.b32 %r173, %f614;
setp.lt.s32 %p37, %r173, 2139095040;
@%p37 bra BB0_26;
setp.gtu.f32 %p38, %f70, 0f7F800000;
@%p38 bra BB0_25;
bra.uni BB0_23;
BB0_25:
add.f32 %f1285, %f68, 0f40800000;
bra.uni BB0_26;
BB0_23:
setp.neu.f32 %p39, %f70, 0f7F800000;
@%p39 bra BB0_26;
selp.f32 %f1285, 0fFF800000, 0f7F800000, %p1;
BB0_26:
mov.u32 %r394, 1;
mov.u64 %rd298, lightMeshBuffer;
cvta.global.u64 %rd297, %rd298;
shl.b32 %r393, %r397, 1;
mov.u64 %rd296, 0;
mov.f32 %f615, 0f3F800000;
sub.f32 %f616, %f615, %f1285;
setp.eq.f32 %p40, %f68, 0f3F800000;
selp.f32 %f617, 0f00000000, %f616, %p40;
cvt.sat.f32.f32 %f618, %f617;
mul.f32 %f619, %f67, %f618;
add.s32 %r176, %r393, 1;
cvt.s64.s32 %rd86, %r176;
// inline asm
call (%rd84), _rt_buffer_get_64, (%rd297, %r394, %r125, %rd86, %rd296, %rd296, %rd296);
// inline asm
ld.f32 %f620, [%rd84];
mul.f32 %f621, %f63, %f620;
ld.f32 %f622, [%rd84+4];
mul.f32 %f623, %f64, %f622;
neg.f32 %f624, %f623;
sub.f32 %f625, %f624, %f621;
ld.f32 %f626, [%rd84+8];
mul.f32 %f627, %f65, %f626;
sub.f32 %f628, %f625, %f627;
cvt.sat.f32.f32 %f629, %f628;
mul.f32 %f82, %f619, %f629;
setp.leu.f32 %p41, %f82, 0f3727C5AC;
@%p41 bra BB0_28;
mov.u32 %r396, 1;
cvt.sat.f32.f32 %f638, %f66;
add.u64 %rd91, %SP, 16;
cvta.to.local.u64 %rd92, %rd91;
max.f32 %f636, %f643, %f381;
sub.f32 %f637, %f62, %f636;
mov.u32 %r180, 1065353216;
st.local.u32 [%rd92+8], %r180;
st.local.u32 [%rd92+4], %r180;
st.local.u32 [%rd92], %r180;
ld.global.u32 %r177, [root];
// inline asm
call _rt_trace_64, (%r177, %f12, %f13, %f14, %f63, %f64, %f65, %r396, %f636, %f637, %rd91, %r125);
// inline asm
ld.local.f32 %f645, [%rd92];
mul.f32 %f646, %f82, 0f3EA2F983;
mul.f32 %f647, %f646, %f645;
ld.local.f32 %f648, [%rd92+4];
mul.f32 %f649, %f646, %f648;
ld.local.f32 %f650, [%rd92+8];
mul.f32 %f651, %f646, %f650;
fma.rn.f32 %f52, %f638, %f647, %f52;
fma.rn.f32 %f51, %f638, %f649, %f51;
fma.rn.f32 %f50, %f638, %f651, %f50;
mul.f32 %f652, %f24, %f64;
fma.rn.f32 %f653, %f22, %f63, %f652;
fma.rn.f32 %f654, %f26, %f65, %f653;
cvt.sat.f32.f32 %f655, %f654;
fma.rn.f32 %f55, %f647, %f655, %f55;
fma.rn.f32 %f54, %f649, %f655, %f54;
fma.rn.f32 %f53, %f651, %f655, %f53;
mul.f32 %f656, %f34, %f64;
fma.rn.f32 %f657, %f33, %f63, %f656;
fma.rn.f32 %f658, %f35, %f65, %f657;
cvt.sat.f32.f32 %f659, %f658;
fma.rn.f32 %f58, %f647, %f659, %f58;
fma.rn.f32 %f57, %f649, %f659, %f57;
fma.rn.f32 %f56, %f651, %f659, %f56;
mul.f32 %f660, %f43, %f64;
fma.rn.f32 %f661, %f42, %f63, %f660;
fma.rn.f32 %f662, %f44, %f65, %f661;
cvt.sat.f32.f32 %f663, %f662;
fma.rn.f32 %f61, %f647, %f663, %f61;
fma.rn.f32 %f60, %f649, %f663, %f60;
fma.rn.f32 %f59, %f651, %f663, %f59;
add.f32 %f49, %f49, %f645;
BB0_28:
add.s32 %r15, %r15, 1;
BB0_29:
add.s32 %r397, %r397, 1;
setp.lt.s32 %p43, %r397, %r2;
@%p43 bra BB0_12;
bra.uni BB0_30;
BB0_10:
mov.f32 %f50, %f49;
mov.f32 %f51, %f49;
mov.f32 %f52, %f49;
mov.f32 %f53, %f49;
mov.f32 %f54, %f49;
mov.f32 %f55, %f49;
mov.f32 %f56, %f49;
mov.f32 %f57, %f49;
mov.f32 %f58, %f49;
mov.f32 %f59, %f49;
mov.f32 %f60, %f49;
mov.f32 %f61, %f49;
BB0_30:
mov.pred %p158, %p18;
BB0_31:
cvt.rn.f32.s32 %f664, %r15;
mov.f32 %f665, 0f3F800000;
max.f32 %f666, %f664, %f665;
rcp.rn.f32 %f667, %f666;
mul.f32 %f1388, %f52, %f667;
mul.f32 %f1389, %f51, %f667;
mul.f32 %f1390, %f50, %f667;
div.rn.f32 %f1391, %f49, %f666;
mul.f32 %f1385, %f55, %f667;
mul.f32 %f1386, %f54, %f667;
mul.f32 %f1387, %f53, %f667;
mul.f32 %f1382, %f58, %f667;
mul.f32 %f1383, %f57, %f667;
mul.f32 %f1384, %f56, %f667;
mul.f32 %f1379, %f61, %f667;
mul.f32 %f1380, %f60, %f667;
mul.f32 %f1381, %f59, %f667;
@!%p158 bra BB0_86;
bra.uni BB0_32;
BB0_32:
abs.f32 %f1256, %f1260;
abs.f32 %f1255, %f1258;
setp.gt.f32 %p44, %f1255, %f1256;
neg.f32 %f681, %f1259;
selp.f32 %f682, %f681, 0f00000000, %p44;
neg.f32 %f683, %f1260;
selp.f32 %f684, %f1258, %f683, %p44;
selp.f32 %f685, 0f00000000, %f1259, %p44;
mul.f32 %f686, %f684, %f684;
fma.rn.f32 %f687, %f682, %f682, %f686;
fma.rn.f32 %f688, %f685, %f685, %f687;
sqrt.rn.f32 %f689, %f688;
rcp.rn.f32 %f690, %f689;
mul.f32 %f148, %f682, %f690;
mul.f32 %f149, %f684, %f690;
mul.f32 %f150, %f685, %f690;
mov.f32 %f1338, 0f00000000;
setp.lt.s32 %p45, %r3, 1;
mov.f32 %f1339, %f1338;
mov.f32 %f1340, %f1338;
mov.f32 %f1341, %f1338;
mov.f32 %f1342, %f1338;
mov.f32 %f1343, %f1338;
mov.f32 %f1344, %f1338;
mov.f32 %f1345, %f1338;
mov.f32 %f1346, %f1338;
mov.f32 %f1347, %f1338;
mov.f32 %f1348, %f1338;
mov.f32 %f1349, %f1338;
mov.f32 %f1350, %f1338;
@%p45 bra BB0_85;
cvt.rn.f32.s32 %f704, %r3;
rcp.rn.f32 %f151, %f704;
add.u64 %rd93, %SP, 28;
cvta.to.local.u64 %rd6, %rd93;
mul.f32 %f152, %f12, 0f3456BF95;
mul.f32 %f153, %f13, 0f3456BF95;
mul.f32 %f154, %f14, 0f3456BF95;
add.u64 %rd94, %SP, 0;
cvta.to.local.u64 %rd7, %rd94;
mul.f32 %f705, %f1258, %f149;
mul.f32 %f706, %f1259, %f148;
sub.f32 %f155, %f706, %f705;
mul.f32 %f707, %f1260, %f148;
mul.f32 %f708, %f1258, %f150;
sub.f32 %f156, %f708, %f707;
mul.f32 %f709, %f1259, %f150;
mul.f32 %f710, %f1260, %f149;
sub.f32 %f157, %f710, %f709;
mov.f32 %f1338, 0f00000000;
mov.u32 %r181, 0;
abs.f32 %f711, %f153;
abs.f32 %f712, %f152;
max.f32 %f713, %f712, %f711;
abs.f32 %f714, %f154;
max.f32 %f715, %f713, %f714;
mov.u32 %r401, %r181;
mov.f32 %f1339, %f1338;
mov.f32 %f1340, %f1338;
mov.f32 %f1341, %f1338;
mov.f32 %f1342, %f1338;
mov.f32 %f1343, %f1338;
mov.f32 %f1344, %f1338;
mov.f32 %f1345, %f1338;
mov.f32 %f1346, %f1338;
mov.f32 %f1347, %f1338;
mov.f32 %f1348, %f1338;
mov.f32 %f1349, %f1338;
mov.f32 %f1350, %f1338;
BB0_34:
cvt.rn.f32.s32 %f171, %r401;
max.f32 %f172, %f715, %f381;
mov.u32 %r403, %r181;
BB0_35:
mad.lo.s32 %r183, %r404, 1664525, 1013904223;
and.b32 %r184, %r183, 16777215;
cvt.rn.f32.u32 %f717, %r184;
fma.rn.f32 %f718, %f717, 0f33800000, %f171;
mul.f32 %f186, %f151, %f718;
mad.lo.s32 %r404, %r183, 1664525, 1013904223;
and.b32 %r185, %r404, 16777215;
cvt.rn.f32.u32 %f719, %r185;
cvt.rn.f32.s32 %f720, %r403;
fma.rn.f32 %f721, %f719, 0f33800000, %f720;
mul.f32 %f722, %f151, %f721;
mul.f32 %f723, %f186, %f186;
sub.f32 %f725, %f665, %f723;
mov.f32 %f1363, 0f00000000;
max.f32 %f727, %f1363, %f725;
sqrt.rn.f32 %f187, %f727;
mul.f32 %f1357, %f722, 0f40C90FDB;
abs.f32 %f189, %f1357;
setp.neu.f32 %p46, %f189, 0f7F800000;
mov.f32 %f1351, %f1357;
@%p46 bra BB0_37;
mov.f32 %f728, 0f00000000;
mul.rn.f32 %f1351, %f1357, %f728;
BB0_37:
mul.f32 %f729, %f1351, 0f3F22F983;
cvt.rni.s32.f32 %r414, %f729;
cvt.rn.f32.s32 %f730, %r414;
neg.f32 %f731, %f730;
mov.f32 %f732, 0f3FC90FDA;
fma.rn.f32 %f733, %f731, %f732, %f1351;
mov.f32 %f734, 0f33A22168;
fma.rn.f32 %f735, %f731, %f734, %f733;
mov.f32 %f736, 0f27C234C5;
fma.rn.f32 %f1352, %f731, %f736, %f735;
abs.f32 %f737, %f1351;
setp.leu.f32 %p47, %f737, 0f47CE4780;
@%p47 bra BB0_48;
mov.b32 %r27, %f1351;
shr.u32 %r28, %r27, 23;
shl.b32 %r188, %r27, 8;
or.b32 %r29, %r188, -2147483648;
mov.u32 %r405, 0;
mov.u64 %rd301, 0;
mov.u64 %rd300, %rd6;
mov.u32 %r406, %r405;
BB0_39:
.pragma "nounroll";
shl.b64 %rd96, %rd301, 2;
mov.u64 %rd97, __cudart_i2opi_f;
add.s64 %rd98, %rd97, %rd96;
ld.const.u32 %r191, [%rd98];
// inline asm
{
mad.lo.cc.u32 %r189, %r191, %r29, %r406;
madc.hi.u32 %r406, %r191, %r29, 0;
}
// inline asm
st.local.u32 [%rd300], %r189;
add.s32 %r405, %r405, 1;
cvt.s64.s32 %rd301, %r405;
mul.wide.s32 %rd99, %r405, 4;
add.s64 %rd300, %rd6, %rd99;
setp.ne.s32 %p48, %r405, 6;
@%p48 bra BB0_39;
and.b32 %r194, %r28, 255;
add.s32 %r195, %r194, -128;
shr.u32 %r196, %r195, 5;
and.b32 %r34, %r27, -2147483648;
st.local.u32 [%rd6+24], %r406;
mov.u32 %r197, 6;
sub.s32 %r198, %r197, %r196;
mul.wide.s32 %rd100, %r198, 4;
add.s64 %rd13, %rd6, %rd100;
ld.local.u32 %r407, [%rd13];
ld.local.u32 %r408, [%rd13+-4];
and.b32 %r37, %r28, 31;
setp.eq.s32 %p49, %r37, 0;
@%p49 bra BB0_42;
mov.u32 %r199, 32;
sub.s32 %r200, %r199, %r37;
shr.u32 %r201, %r408, %r200;
shl.b32 %r202, %r407, %r37;
add.s32 %r407, %r201, %r202;
ld.local.u32 %r203, [%rd13+-8];
shr.u32 %r204, %r203, %r200;
shl.b32 %r205, %r408, %r37;
add.s32 %r408, %r204, %r205;
BB0_42:
shr.u32 %r206, %r408, 30;
shl.b32 %r207, %r407, 2;
add.s32 %r409, %r206, %r207;
shl.b32 %r43, %r408, 2;
shr.u32 %r208, %r409, 31;
shr.u32 %r209, %r407, 30;
add.s32 %r44, %r208, %r209;
setp.eq.s32 %p50, %r208, 0;
@%p50 bra BB0_43;
bra.uni BB0_44;
BB0_43:
mov.u32 %r410, %r34;
mov.u32 %r411, %r43;
bra.uni BB0_45;
BB0_44:
not.b32 %r210, %r409;
neg.s32 %r411, %r43;
setp.eq.s32 %p51, %r43, 0;
selp.u32 %r211, 1, 0, %p51;
add.s32 %r409, %r211, %r210;
xor.b32 %r410, %r34, -2147483648;
BB0_45:
clz.b32 %r413, %r409;
setp.eq.s32 %p52, %r413, 0;
shl.b32 %r212, %r409, %r413;
mov.u32 %r213, 32;
sub.s32 %r214, %r213, %r413;
shr.u32 %r215, %r411, %r214;
add.s32 %r216, %r215, %r212;
selp.b32 %r52, %r409, %r216, %p52;
mov.u32 %r217, -921707870;
mul.hi.u32 %r412, %r52, %r217;
setp.eq.s32 %p53, %r34, 0;
neg.s32 %r218, %r44;
selp.b32 %r414, %r44, %r218, %p53;
setp.lt.s32 %p54, %r412, 1;
@%p54 bra BB0_47;
mul.lo.s32 %r219, %r52, -921707870;
shr.u32 %r220, %r219, 31;
shl.b32 %r221, %r412, 1;
add.s32 %r412, %r220, %r221;
add.s32 %r413, %r413, 1;
BB0_47:
mov.u32 %r222, 126;
sub.s32 %r223, %r222, %r413;
shl.b32 %r224, %r223, 23;
add.s32 %r225, %r412, 1;
shr.u32 %r226, %r225, 7;
add.s32 %r227, %r226, 1;
shr.u32 %r228, %r227, 1;
add.s32 %r229, %r228, %r224;
or.b32 %r230, %r229, %r410;
mov.b32 %f1352, %r230;
BB0_48:
mul.rn.f32 %f195, %f1352, %f1352;
add.s32 %r60, %r414, 1;
and.b32 %r61, %r60, 1;
setp.eq.s32 %p55, %r61, 0;
@%p55 bra BB0_50;
bra.uni BB0_49;
BB0_50:
mov.f32 %f740, 0f3C08839E;
mov.f32 %f741, 0fB94CA1F9;
fma.rn.f32 %f1353, %f741, %f195, %f740;
bra.uni BB0_51;
BB0_49:
mov.f32 %f738, 0fBAB6061A;
mov.f32 %f739, 0f37CCF5CE;
fma.rn.f32 %f1353, %f739, %f195, %f738;
BB0_51:
@%p55 bra BB0_53;
bra.uni BB0_52;
BB0_53:
mov.f32 %f745, 0fBE2AAAA3;
fma.rn.f32 %f746, %f1353, %f195, %f745;
mov.f32 %f747, 0f00000000;
fma.rn.f32 %f1354, %f746, %f195, %f747;
bra.uni BB0_54;
BB0_52:
mov.f32 %f742, 0f3D2AAAA5;
fma.rn.f32 %f743, %f1353, %f195, %f742;
mov.f32 %f744, 0fBF000000;
fma.rn.f32 %f1354, %f743, %f195, %f744;
BB0_54:
fma.rn.f32 %f1355, %f1354, %f1352, %f1352;
@%p55 bra BB0_56;
fma.rn.f32 %f1355, %f1354, %f195, %f665;
BB0_56:
and.b32 %r231, %r60, 2;
setp.eq.s32 %p58, %r231, 0;
@%p58 bra BB0_58;
mov.f32 %f750, 0fBF800000;
fma.rn.f32 %f1355, %f1355, %f750, %f1363;
BB0_58:
@%p46 bra BB0_60;
mov.f32 %f751, 0f00000000;
mul.rn.f32 %f1357, %f1357, %f751;
BB0_60:
mul.f32 %f752, %f1357, 0f3F22F983;
cvt.rni.s32.f32 %r424, %f752;
cvt.rn.f32.s32 %f753, %r424;
neg.f32 %f754, %f753;
fma.rn.f32 %f756, %f754, %f732, %f1357;
fma.rn.f32 %f758, %f754, %f734, %f756;
fma.rn.f32 %f1358, %f754, %f736, %f758;
abs.f32 %f760, %f1357;
setp.leu.f32 %p60, %f760, 0f47CE4780;
@%p60 bra BB0_71;
mov.b32 %r63, %f1357;
shr.u32 %r64, %r63, 23;
shl.b32 %r234, %r63, 8;
or.b32 %r65, %r234, -2147483648;
mov.u32 %r415, 0;
mov.u64 %rd302, %rd6;
mov.u64 %rd303, %rd31;
mov.u32 %r416, %r415;
BB0_62:
.pragma "nounroll";
shl.b64 %rd102, %rd303, 2;
mov.u64 %rd103, __cudart_i2opi_f;
add.s64 %rd104, %rd103, %rd102;
ld.const.u32 %r237, [%rd104];
// inline asm
{
mad.lo.cc.u32 %r235, %r237, %r65, %r416;
madc.hi.u32 %r416, %r237, %r65, 0;
}
// inline asm
st.local.u32 [%rd302], %r235;
add.s32 %r415, %r415, 1;
cvt.s64.s32 %rd303, %r415;
mul.wide.s32 %rd105, %r415, 4;
add.s64 %rd302, %rd6, %rd105;
setp.ne.s32 %p61, %r415, 6;
@%p61 bra BB0_62;
and.b32 %r240, %r64, 255;
add.s32 %r241, %r240, -128;
shr.u32 %r242, %r241, 5;
and.b32 %r70, %r63, -2147483648;
st.local.u32 [%rd6+24], %r416;
mov.u32 %r243, 6;
sub.s32 %r244, %r243, %r242;
mul.wide.s32 %rd106, %r244, 4;
add.s64 %rd19, %rd6, %rd106;
ld.local.u32 %r417, [%rd19];
ld.local.u32 %r418, [%rd19+-4];
and.b32 %r73, %r64, 31;
setp.eq.s32 %p62, %r73, 0;
@%p62 bra BB0_65;
mov.u32 %r245, 32;
sub.s32 %r246, %r245, %r73;
shr.u32 %r247, %r418, %r246;
shl.b32 %r248, %r417, %r73;
add.s32 %r417, %r247, %r248;
ld.local.u32 %r249, [%rd19+-8];
shr.u32 %r250, %r249, %r246;
shl.b32 %r251, %r418, %r73;
add.s32 %r418, %r250, %r251;
BB0_65:
shr.u32 %r252, %r418, 30;
shl.b32 %r253, %r417, 2;
add.s32 %r419, %r252, %r253;
shl.b32 %r79, %r418, 2;
shr.u32 %r254, %r419, 31;
shr.u32 %r255, %r417, 30;
add.s32 %r80, %r254, %r255;
setp.eq.s32 %p63, %r254, 0;
@%p63 bra BB0_66;
bra.uni BB0_67;
BB0_66:
mov.u32 %r420, %r70;
mov.u32 %r421, %r79;
bra.uni BB0_68;
BB0_67:
not.b32 %r256, %r419;
neg.s32 %r421, %r79;
setp.eq.s32 %p64, %r79, 0;
selp.u32 %r257, 1, 0, %p64;
add.s32 %r419, %r257, %r256;
xor.b32 %r420, %r70, -2147483648;
BB0_68:
clz.b32 %r423, %r419;
setp.eq.s32 %p65, %r423, 0;
shl.b32 %r258, %r419, %r423;
mov.u32 %r259, 32;
sub.s32 %r260, %r259, %r423;
shr.u32 %r261, %r421, %r260;
add.s32 %r262, %r261, %r258;
selp.b32 %r88, %r419, %r262, %p65;
mov.u32 %r263, -921707870;
mul.hi.u32 %r422, %r88, %r263;
setp.eq.s32 %p66, %r70, 0;
neg.s32 %r264, %r80;
selp.b32 %r424, %r80, %r264, %p66;
setp.lt.s32 %p67, %r422, 1;
@%p67 bra BB0_70;
mul.lo.s32 %r265, %r88, -921707870;
shr.u32 %r266, %r265, 31;
shl.b32 %r267, %r422, 1;
add.s32 %r422, %r266, %r267;
add.s32 %r423, %r423, 1;
BB0_70:
mov.u32 %r268, 126;
sub.s32 %r269, %r268, %r423;
shl.b32 %r270, %r269, 23;
add.s32 %r271, %r422, 1;
shr.u32 %r272, %r271, 7;
add.s32 %r273, %r272, 1;
shr.u32 %r274, %r273, 1;
add.s32 %r275, %r274, %r270;
or.b32 %r276, %r275, %r420;
mov.b32 %f1358, %r276;
BB0_71:
mul.rn.f32 %f212, %f1358, %f1358;
and.b32 %r96, %r424, 1;
setp.eq.s32 %p68, %r96, 0;
@%p68 bra BB0_73;
bra.uni BB0_72;
BB0_73:
mov.f32 %f763, 0f3C08839E;
mov.f32 %f764, 0fB94CA1F9;
fma.rn.f32 %f1359, %f764, %f212, %f763;
bra.uni BB0_74;
BB0_72:
mov.f32 %f761, 0fBAB6061A;
mov.f32 %f762, 0f37CCF5CE;
fma.rn.f32 %f1359, %f762, %f212, %f761;
BB0_74:
@%p68 bra BB0_76;
bra.uni BB0_75;
BB0_76:
mov.f32 %f768, 0fBE2AAAA3;
fma.rn.f32 %f769, %f1359, %f212, %f768;
mov.f32 %f770, 0f00000000;
fma.rn.f32 %f1360, %f769, %f212, %f770;
bra.uni BB0_77;
BB0_75:
mov.f32 %f765, 0f3D2AAAA5;
fma.rn.f32 %f766, %f1359, %f212, %f765;
mov.f32 %f767, 0fBF000000;
fma.rn.f32 %f1360, %f766, %f212, %f767;
BB0_77:
fma.rn.f32 %f1361, %f1360, %f1358, %f1358;
@%p68 bra BB0_79;
fma.rn.f32 %f1361, %f1360, %f212, %f665;
BB0_79:
and.b32 %r277, %r424, 2;
setp.eq.s32 %p71, %r277, 0;
@%p71 bra BB0_81;
mov.f32 %f773, 0fBF800000;
fma.rn.f32 %f1361, %f1361, %f773, %f1363;
BB0_81:
mul.f32 %f785, %f187, %f1355;
mul.f32 %f786, %f187, %f1361;
mul.f32 %f787, %f148, %f786;
mul.f32 %f788, %f149, %f786;
mul.f32 %f789, %f150, %f786;
fma.rn.f32 %f790, %f157, %f785, %f787;
fma.rn.f32 %f791, %f156, %f785, %f788;
fma.rn.f32 %f792, %f155, %f785, %f789;
fma.rn.f32 %f224, %f1258, %f186, %f790;
fma.rn.f32 %f225, %f1259, %f186, %f791;
fma.rn.f32 %f226, %f1260, %f186, %f792;
mov.f32 %f793, 0f7F7FFFFF;
st.local.v4.f32 [%rd7], {%f665, %f665, %f665, %f793};
ld.global.u32 %r278, [root];
mov.u32 %r279, 0;
mov.f32 %f781, 0f6C4ECB8F;
mov.u32 %r280, 16;
// inline asm
call _rt_trace_64, (%r278, %f12, %f13, %f14, %f224, %f225, %f226, %r279, %f172, %f781, %rd94, %r280);
// inline asm
ld.local.f32 %f227, [%rd7+12];
setp.ge.f32 %p72, %f227, 0f00000000;
mov.f32 %f1364, %f1363;
mov.f32 %f1365, %f1363;
@%p72 bra BB0_83;
ld.local.v4.f32 {%f1365, %f1364, %f1363, %f798}, [%rd7];
BB0_83:
setp.ltu.f32 %p73, %f227, 0f00000000;
selp.f32 %f799, 0f3F800000, 0f00000000, %p73;
add.f32 %f1338, %f1338, %f799;
mul.f32 %f800, %f24, %f225;
fma.rn.f32 %f801, %f22, %f224, %f800;
fma.rn.f32 %f802, %f26, %f226, %f801;
cvt.sat.f32.f32 %f803, %f802;
fma.rn.f32 %f1344, %f1365, %f803, %f1344;
fma.rn.f32 %f1343, %f1364, %f803, %f1343;
fma.rn.f32 %f1342, %f1363, %f803, %f1342;
mul.f32 %f804, %f34, %f225;
fma.rn.f32 %f805, %f33, %f224, %f804;
fma.rn.f32 %f806, %f35, %f226, %f805;
cvt.sat.f32.f32 %f807, %f806;
fma.rn.f32 %f1347, %f1365, %f807, %f1347;
fma.rn.f32 %f1346, %f1364, %f807, %f1346;
fma.rn.f32 %f1345, %f1363, %f807, %f1345;
mul.f32 %f808, %f43, %f225;
fma.rn.f32 %f809, %f42, %f224, %f808;
fma.rn.f32 %f810, %f44, %f226, %f809;
cvt.sat.f32.f32 %f811, %f810;
fma.rn.f32 %f1350, %f1365, %f811, %f1350;
fma.rn.f32 %f1349, %f1364, %f811, %f1349;
fma.rn.f32 %f1348, %f1363, %f811, %f1348;
mul.f32 %f812, %f1259, %f225;
fma.rn.f32 %f813, %f1258, %f224, %f812;
fma.rn.f32 %f814, %f1260, %f226, %f813;
cvt.sat.f32.f32 %f815, %f814;
fma.rn.f32 %f1341, %f1365, %f815, %f1341;
fma.rn.f32 %f1340, %f1364, %f815, %f1340;
fma.rn.f32 %f1339, %f1363, %f815, %f1339;
add.s32 %r403, %r403, 1;
setp.lt.s32 %p74, %r403, %r3;
@%p74 bra BB0_35;
add.s32 %r401, %r401, 1;
setp.lt.s32 %p75, %r401, %r3;
@%p75 bra BB0_34;
BB0_85:
mul.lo.s32 %r281, %r3, %r3;
cvt.rn.f32.s32 %f816, %r281;
rcp.rn.f32 %f817, %f816;
mul.f32 %f818, %f1341, %f817;
mul.f32 %f819, %f1340, %f817;
mul.f32 %f820, %f1339, %f817;
div.rn.f32 %f1391, %f1338, %f816;
mul.f32 %f1385, %f1344, %f817;
mul.f32 %f1386, %f1343, %f817;
mul.f32 %f1387, %f1342, %f817;
mul.f32 %f1382, %f1347, %f817;
mul.f32 %f1383, %f1346, %f817;
mul.f32 %f1384, %f1345, %f817;
mul.f32 %f1379, %f1350, %f817;
mul.f32 %f1380, %f1349, %f817;
mul.f32 %f1381, %f1348, %f817;
fma.rn.f32 %f1388, %f1341, %f817, %f818;
fma.rn.f32 %f1389, %f1340, %f817, %f819;
fma.rn.f32 %f1390, %f1339, %f817, %f820;
BB0_86:
setp.gt.s32 %p76, %r1, -1;
selp.f32 %f286, %f1391, %f1388, %p76;
ld.global.u32 %r426, [imageEnabled];
and.b32 %r282, %r426, 8;
setp.eq.s32 %p77, %r282, 0;
@%p77 bra BB0_99;
cvt.u64.u32 %rd110, %r5;
cvt.u64.u32 %rd111, %r6;
mov.u64 %rd114, image_Mask;
cvta.global.u64 %rd109, %rd114;
// inline asm
call (%rd108), _rt_buffer_get_64, (%rd109, %r109, %r109, %rd110, %rd111, %rd31, %rd31);
// inline asm
mov.f32 %f823, 0f3E68BA2E;
cvt.rzi.f32.f32 %f824, %f823;
fma.rn.f32 %f825, %f824, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f287, %f825;
abs.f32 %f288, %f286;
setp.lt.f32 %p78, %f288, 0f00800000;
mul.f32 %f826, %f288, 0f4B800000;
selp.f32 %f827, 0fC3170000, 0fC2FE0000, %p78;
selp.f32 %f828, %f826, %f288, %p78;
mov.b32 %r285, %f828;
and.b32 %r286, %r285, 8388607;
or.b32 %r287, %r286, 1065353216;
mov.b32 %f829, %r287;
shr.u32 %r288, %r285, 23;
cvt.rn.f32.u32 %f830, %r288;
add.f32 %f831, %f827, %f830;
setp.gt.f32 %p79, %f829, 0f3FB504F3;
mul.f32 %f832, %f829, 0f3F000000;
add.f32 %f833, %f831, 0f3F800000;
selp.f32 %f834, %f832, %f829, %p79;
selp.f32 %f835, %f833, %f831, %p79;
add.f32 %f836, %f834, 0fBF800000;
add.f32 %f822, %f834, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f821,%f822;
// inline asm
add.f32 %f837, %f836, %f836;
mul.f32 %f838, %f821, %f837;
mul.f32 %f839, %f838, %f838;
mov.f32 %f840, 0f3C4CAF63;
mov.f32 %f841, 0f3B18F0FE;
fma.rn.f32 %f842, %f841, %f839, %f840;
mov.f32 %f843, 0f3DAAAABD;
fma.rn.f32 %f844, %f842, %f839, %f843;
mul.rn.f32 %f845, %f844, %f839;
mul.rn.f32 %f846, %f845, %f838;
sub.f32 %f847, %f836, %f838;
neg.f32 %f848, %f838;
add.f32 %f849, %f847, %f847;
fma.rn.f32 %f850, %f848, %f836, %f849;
mul.rn.f32 %f851, %f821, %f850;
add.f32 %f852, %f846, %f838;
sub.f32 %f853, %f838, %f852;
add.f32 %f854, %f846, %f853;
add.f32 %f855, %f851, %f854;
add.f32 %f856, %f852, %f855;
sub.f32 %f857, %f852, %f856;
add.f32 %f858, %f855, %f857;
mov.f32 %f859, 0f3F317200;
mul.rn.f32 %f860, %f835, %f859;
mov.f32 %f861, 0f35BFBE8E;
mul.rn.f32 %f862, %f835, %f861;
add.f32 %f863, %f860, %f856;
sub.f32 %f864, %f860, %f863;
add.f32 %f865, %f856, %f864;
add.f32 %f866, %f858, %f865;
add.f32 %f867, %f862, %f866;
add.f32 %f868, %f863, %f867;
sub.f32 %f869, %f863, %f868;
add.f32 %f870, %f867, %f869;
mov.f32 %f871, 0f3EE8BA2E;
mul.rn.f32 %f872, %f871, %f868;
neg.f32 %f873, %f872;
fma.rn.f32 %f874, %f871, %f868, %f873;
fma.rn.f32 %f875, %f871, %f870, %f874;
mov.f32 %f876, 0f00000000;
fma.rn.f32 %f877, %f876, %f868, %f875;
add.rn.f32 %f878, %f872, %f877;
neg.f32 %f879, %f878;
add.rn.f32 %f880, %f872, %f879;
add.rn.f32 %f881, %f880, %f877;
mov.b32 %r289, %f878;
setp.eq.s32 %p80, %r289, 1118925336;
add.s32 %r290, %r289, -1;
mov.b32 %f882, %r290;
add.f32 %f883, %f881, 0f37000000;
selp.f32 %f884, %f882, %f878, %p80;
selp.f32 %f289, %f883, %f881, %p80;
mul.f32 %f885, %f884, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f886, %f885;
mov.f32 %f887, 0fBF317200;
fma.rn.f32 %f888, %f886, %f887, %f884;
mov.f32 %f889, 0fB5BFBE8E;
fma.rn.f32 %f890, %f886, %f889, %f888;
mul.f32 %f891, %f890, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f892, %f891;
add.f32 %f893, %f886, 0f00000000;
ex2.approx.f32 %f894, %f893;
mul.f32 %f895, %f892, %f894;
setp.lt.f32 %p81, %f884, 0fC2D20000;
selp.f32 %f896, 0f00000000, %f895, %p81;
setp.gt.f32 %p82, %f884, 0f42D20000;
selp.f32 %f1392, 0f7F800000, %f896, %p82;
setp.eq.f32 %p83, %f1392, 0f7F800000;
@%p83 bra BB0_89;
fma.rn.f32 %f1392, %f1392, %f289, %f1392;
BB0_89:
setp.lt.f32 %p84, %f286, 0f00000000;
setp.eq.f32 %p85, %f287, 0f3F800000;
and.pred %p3, %p84, %p85;
mov.b32 %r291, %f1392;
xor.b32 %r292, %r291, -2147483648;
mov.b32 %f897, %r292;
selp.f32 %f1394, %f897, %f1392, %p3;
setp.eq.f32 %p86, %f286, 0f00000000;
@%p86 bra BB0_92;
bra.uni BB0_90;
BB0_92:
add.f32 %f900, %f286, %f286;
selp.f32 %f1394, %f900, 0f00000000, %p85;
bra.uni BB0_93;
BB0_154:
mov.u64 %rd235, image_HDR;
cvta.global.u64 %rd230, %rd235;
mov.u32 %r371, 8;
mov.u64 %rd234, 0;
// inline asm
call (%rd229), _rt_buffer_get_64, (%rd230, %r109, %r371, %rd24, %rd25, %rd234, %rd234);
// inline asm
mov.f32 %f1227, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs117, %f1227;}
// inline asm
mov.u16 %rs118, 0;
st.v4.u16 [%rd229], {%rs117, %rs117, %rs117, %rs118};
BB0_155:
ld.global.u32 %r372, [additive];
setp.eq.s32 %p155, %r372, 0;
@%p155 bra BB0_157;
mov.u64 %rd248, image_RNM0;
cvta.global.u64 %rd237, %rd248;
mov.u32 %r376, 8;
mov.u64 %rd247, 0;
// inline asm
call (%rd236), _rt_buffer_get_64, (%rd237, %r109, %r376, %rd24, %rd25, %rd247, %rd247);
// inline asm
ld.v4.u16 {%rs125, %rs126, %rs127, %rs128}, [%rd236];
// inline asm
{ cvt.f32.f16 %f1228, %rs125;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1229, %rs126;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1230, %rs127;}
// inline asm
// inline asm
call (%rd242), _rt_buffer_get_64, (%rd237, %r109, %r376, %rd24, %rd25, %rd247, %rd247);
// inline asm
add.f32 %f1231, %f1228, 0f00000000;
add.f32 %f1232, %f1229, 0f00000000;
add.f32 %f1233, %f1230, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs124, %f1233;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs123, %f1232;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs122, %f1231;}
// inline asm
mov.u16 %rs129, 0;
st.v4.u16 [%rd242], {%rs122, %rs123, %rs124, %rs129};
bra.uni BB0_158;
BB0_157:
mov.u64 %rd255, image_RNM0;
cvta.global.u64 %rd250, %rd255;
mov.u32 %r378, 8;
mov.u64 %rd254, 0;
// inline asm
call (%rd249), _rt_buffer_get_64, (%rd250, %r109, %r378, %rd24, %rd25, %rd254, %rd254);
// inline asm
mov.f32 %f1234, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs130, %f1234;}
// inline asm
mov.u16 %rs131, 0;
st.v4.u16 [%rd249], {%rs130, %rs130, %rs130, %rs131};
BB0_158:
ld.global.u32 %r379, [additive];
setp.eq.s32 %p156, %r379, 0;
@%p156 bra BB0_160;
mov.u64 %rd268, image_RNM1;
cvta.global.u64 %rd257, %rd268;
mov.u32 %r383, 8;
mov.u64 %rd267, 0;
// inline asm
call (%rd256), _rt_buffer_get_64, (%rd257, %r109, %r383, %rd24, %rd25, %rd267, %rd267);
// inline asm
ld.v4.u16 {%rs138, %rs139, %rs140, %rs141}, [%rd256];
// inline asm
{ cvt.f32.f16 %f1235, %rs138;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1236, %rs139;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1237, %rs140;}
// inline asm
// inline asm
call (%rd262), _rt_buffer_get_64, (%rd257, %r109, %r383, %rd24, %rd25, %rd267, %rd267);
// inline asm
add.f32 %f1238, %f1235, 0f00000000;
add.f32 %f1239, %f1236, 0f00000000;
add.f32 %f1240, %f1237, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs137, %f1240;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs136, %f1239;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs135, %f1238;}
// inline asm
mov.u16 %rs142, 0;
st.v4.u16 [%rd262], {%rs135, %rs136, %rs137, %rs142};
bra.uni BB0_161;
BB0_160:
mov.u64 %rd275, image_RNM1;
cvta.global.u64 %rd270, %rd275;
mov.u32 %r385, 8;
mov.u64 %rd274, 0;
// inline asm
call (%rd269), _rt_buffer_get_64, (%rd270, %r109, %r385, %rd24, %rd25, %rd274, %rd274);
// inline asm
mov.f32 %f1241, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs143, %f1241;}
// inline asm
mov.u16 %rs144, 0;
st.v4.u16 [%rd269], {%rs143, %rs143, %rs143, %rs144};
BB0_161:
ld.global.u32 %r386, [additive];
setp.eq.s32 %p157, %r386, 0;
@%p157 bra BB0_163;
mov.u64 %rd288, image_RNM2;
cvta.global.u64 %rd277, %rd288;
mov.u32 %r390, 8;
mov.u64 %rd287, 0;
// inline asm
call (%rd276), _rt_buffer_get_64, (%rd277, %r109, %r390, %rd24, %rd25, %rd287, %rd287);
// inline asm
ld.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd276];
// inline asm
{ cvt.f32.f16 %f1242, %rs151;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1243, %rs152;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1244, %rs153;}
// inline asm
// inline asm
call (%rd282), _rt_buffer_get_64, (%rd277, %r109, %r390, %rd24, %rd25, %rd287, %rd287);
// inline asm
add.f32 %f1245, %f1242, 0f00000000;
add.f32 %f1246, %f1243, 0f00000000;
add.f32 %f1247, %f1244, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs150, %f1247;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs149, %f1246;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs148, %f1245;}
// inline asm
mov.u16 %rs155, 0;
st.v4.u16 [%rd282], {%rs148, %rs149, %rs150, %rs155};
bra.uni BB0_164;
BB0_163:
mov.u64 %rd295, image_RNM2;
cvta.global.u64 %rd290, %rd295;
mov.u32 %r392, 8;
mov.u64 %rd294, 0;
// inline asm
call (%rd289), _rt_buffer_get_64, (%rd290, %r109, %r392, %rd24, %rd25, %rd294, %rd294);
// inline asm
mov.f32 %f1248, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs156, %f1248;}
// inline asm
mov.u16 %rs157, 0;
st.v4.u16 [%rd289], {%rs156, %rs156, %rs156, %rs157};
bra.uni BB0_164;
BB0_90:
setp.geu.f32 %p87, %f286, 0f00000000;
@%p87 bra BB0_93;
cvt.rzi.f32.f32 %f899, %f871;
setp.neu.f32 %p88, %f899, 0f3EE8BA2E;
selp.f32 %f1394, 0f7FFFFFFF, %f1394, %p88;
BB0_93:
add.f32 %f901, %f288, 0f3EE8BA2E;
mov.b32 %r293, %f901;
setp.lt.s32 %p90, %r293, 2139095040;
@%p90 bra BB0_98;
setp.gtu.f32 %p91, %f288, 0f7F800000;
@%p91 bra BB0_97;
bra.uni BB0_95;
BB0_97:
add.f32 %f1394, %f286, 0f3EE8BA2E;
bra.uni BB0_98;
BB0_95:
setp.neu.f32 %p92, %f288, 0f7F800000;
@%p92 bra BB0_98;
selp.f32 %f1394, 0fFF800000, 0f7F800000, %p3;
BB0_98:
mul.f32 %f902, %f1394, 0f437F0000;
setp.eq.f32 %p93, %f286, 0f3F800000;
selp.f32 %f903, 0f437F0000, %f902, %p93;
cvt.rzi.u32.f32 %r294, %f903;
cvt.u16.u32 %rs41, %r294;
mov.u16 %rs42, 255;
st.v2.u8 [%rd108], {%rs41, %rs42};
ld.global.u32 %r426, [imageEnabled];
BB0_99:
ld.global.f32 %f904, [lightColor];
mul.f32 %f300, %f1388, %f904;
ld.global.f32 %f905, [lightColor+4];
mul.f32 %f301, %f1389, %f905;
ld.global.f32 %f906, [lightColor+8];
mul.f32 %f302, %f1390, %f906;
and.b32 %r295, %r426, 1;
setp.eq.b32 %p94, %r295, 1;
@!%p94 bra BB0_134;
bra.uni BB0_100;
BB0_100:
mov.f32 %f909, 0f3E666666;
cvt.rzi.f32.f32 %f910, %f909;
fma.rn.f32 %f911, %f910, 0fC0000000, 0f3EE66666;
abs.f32 %f303, %f911;
abs.f32 %f304, %f300;
setp.lt.f32 %p95, %f304, 0f00800000;
mul.f32 %f912, %f304, 0f4B800000;
selp.f32 %f913, 0fC3170000, 0fC2FE0000, %p95;
selp.f32 %f914, %f912, %f304, %p95;
mov.b32 %r296, %f914;
and.b32 %r297, %r296, 8388607;
or.b32 %r298, %r297, 1065353216;
mov.b32 %f915, %r298;
shr.u32 %r299, %r296, 23;
cvt.rn.f32.u32 %f916, %r299;
add.f32 %f917, %f913, %f916;
setp.gt.f32 %p96, %f915, 0f3FB504F3;
mul.f32 %f918, %f915, 0f3F000000;
add.f32 %f919, %f917, 0f3F800000;
selp.f32 %f920, %f918, %f915, %p96;
selp.f32 %f921, %f919, %f917, %p96;
add.f32 %f922, %f920, 0fBF800000;
add.f32 %f908, %f920, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f907,%f908;
// inline asm
add.f32 %f923, %f922, %f922;
mul.f32 %f924, %f907, %f923;
mul.f32 %f925, %f924, %f924;
mov.f32 %f926, 0f3C4CAF63;
mov.f32 %f927, 0f3B18F0FE;
fma.rn.f32 %f928, %f927, %f925, %f926;
mov.f32 %f929, 0f3DAAAABD;
fma.rn.f32 %f930, %f928, %f925, %f929;
mul.rn.f32 %f931, %f930, %f925;
mul.rn.f32 %f932, %f931, %f924;
sub.f32 %f933, %f922, %f924;
neg.f32 %f934, %f924;
add.f32 %f935, %f933, %f933;
fma.rn.f32 %f936, %f934, %f922, %f935;
mul.rn.f32 %f937, %f907, %f936;
add.f32 %f938, %f932, %f924;
sub.f32 %f939, %f924, %f938;
add.f32 %f940, %f932, %f939;
add.f32 %f941, %f937, %f940;
add.f32 %f942, %f938, %f941;
sub.f32 %f943, %f938, %f942;
add.f32 %f944, %f941, %f943;
mov.f32 %f945, 0f3F317200;
mul.rn.f32 %f946, %f921, %f945;
mov.f32 %f947, 0f35BFBE8E;
mul.rn.f32 %f948, %f921, %f947;
add.f32 %f949, %f946, %f942;
sub.f32 %f950, %f946, %f949;
add.f32 %f951, %f942, %f950;
add.f32 %f952, %f944, %f951;
add.f32 %f953, %f948, %f952;
add.f32 %f954, %f949, %f953;
sub.f32 %f955, %f949, %f954;
add.f32 %f956, %f953, %f955;
mov.f32 %f957, 0f3EE66666;
mul.rn.f32 %f958, %f957, %f954;
neg.f32 %f959, %f958;
fma.rn.f32 %f960, %f957, %f954, %f959;
fma.rn.f32 %f961, %f957, %f956, %f960;
mov.f32 %f962, 0f00000000;
fma.rn.f32 %f963, %f962, %f954, %f961;
add.rn.f32 %f964, %f958, %f963;
neg.f32 %f965, %f964;
add.rn.f32 %f966, %f958, %f965;
add.rn.f32 %f967, %f966, %f963;
mov.b32 %r300, %f964;
setp.eq.s32 %p97, %r300, 1118925336;
add.s32 %r301, %r300, -1;
mov.b32 %f968, %r301;
add.f32 %f969, %f967, 0f37000000;
selp.f32 %f970, %f968, %f964, %p97;
selp.f32 %f305, %f969, %f967, %p97;
mul.f32 %f971, %f970, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f972, %f971;
mov.f32 %f973, 0fBF317200;
fma.rn.f32 %f974, %f972, %f973, %f970;
mov.f32 %f975, 0fB5BFBE8E;
fma.rn.f32 %f976, %f972, %f975, %f974;
mul.f32 %f977, %f976, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f978, %f977;
add.f32 %f979, %f972, 0f00000000;
ex2.approx.f32 %f980, %f979;
mul.f32 %f981, %f978, %f980;
setp.lt.f32 %p98, %f970, 0fC2D20000;
selp.f32 %f982, 0f00000000, %f981, %p98;
setp.gt.f32 %p99, %f970, 0f42D20000;
selp.f32 %f1395, 0f7F800000, %f982, %p99;
setp.eq.f32 %p100, %f1395, 0f7F800000;
@%p100 bra BB0_102;
fma.rn.f32 %f1395, %f1395, %f305, %f1395;
BB0_102:
setp.lt.f32 %p101, %f300, 0f00000000;
setp.eq.f32 %p102, %f303, 0f3F800000;
and.pred %p4, %p101, %p102;
mov.b32 %r302, %f1395;
xor.b32 %r303, %r302, -2147483648;
mov.b32 %f983, %r303;
selp.f32 %f1397, %f983, %f1395, %p4;
setp.eq.f32 %p103, %f300, 0f00000000;
@%p103 bra BB0_105;
bra.uni BB0_103;
BB0_105:
add.f32 %f986, %f300, %f300;
selp.f32 %f1397, %f986, 0f00000000, %p102;
bra.uni BB0_106;
BB0_103:
setp.geu.f32 %p104, %f300, 0f00000000;
@%p104 bra BB0_106;
cvt.rzi.f32.f32 %f985, %f957;
setp.neu.f32 %p105, %f985, 0f3EE66666;
selp.f32 %f1397, 0f7FFFFFFF, %f1397, %p105;
BB0_106:
add.f32 %f987, %f304, 0f3EE66666;
mov.b32 %r304, %f987;
setp.lt.s32 %p107, %r304, 2139095040;
@%p107 bra BB0_111;
setp.gtu.f32 %p108, %f304, 0f7F800000;
@%p108 bra BB0_110;
bra.uni BB0_108;
BB0_110:
add.f32 %f1397, %f300, 0f3EE66666;
bra.uni BB0_111;
BB0_108:
setp.neu.f32 %p109, %f304, 0f7F800000;
@%p109 bra BB0_111;
selp.f32 %f1397, 0fFF800000, 0f7F800000, %p4;
BB0_111:
setp.eq.f32 %p110, %f300, 0f3F800000;
selp.f32 %f316, 0f3F800000, %f1397, %p110;
abs.f32 %f317, %f301;
setp.lt.f32 %p111, %f317, 0f00800000;
mul.f32 %f990, %f317, 0f4B800000;
selp.f32 %f991, 0fC3170000, 0fC2FE0000, %p111;
selp.f32 %f992, %f990, %f317, %p111;
mov.b32 %r305, %f992;
and.b32 %r306, %r305, 8388607;
or.b32 %r307, %r306, 1065353216;
mov.b32 %f993, %r307;
shr.u32 %r308, %r305, 23;
cvt.rn.f32.u32 %f994, %r308;
add.f32 %f995, %f991, %f994;
setp.gt.f32 %p112, %f993, 0f3FB504F3;
mul.f32 %f996, %f993, 0f3F000000;
add.f32 %f997, %f995, 0f3F800000;
selp.f32 %f998, %f996, %f993, %p112;
selp.f32 %f999, %f997, %f995, %p112;
add.f32 %f1000, %f998, 0fBF800000;
add.f32 %f989, %f998, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f988,%f989;
// inline asm
add.f32 %f1001, %f1000, %f1000;
mul.f32 %f1002, %f988, %f1001;
mul.f32 %f1003, %f1002, %f1002;
fma.rn.f32 %f1006, %f927, %f1003, %f926;
fma.rn.f32 %f1008, %f1006, %f1003, %f929;
mul.rn.f32 %f1009, %f1008, %f1003;
mul.rn.f32 %f1010, %f1009, %f1002;
sub.f32 %f1011, %f1000, %f1002;
neg.f32 %f1012, %f1002;
add.f32 %f1013, %f1011, %f1011;
fma.rn.f32 %f1014, %f1012, %f1000, %f1013;
mul.rn.f32 %f1015, %f988, %f1014;
add.f32 %f1016, %f1010, %f1002;
sub.f32 %f1017, %f1002, %f1016;
add.f32 %f1018, %f1010, %f1017;
add.f32 %f1019, %f1015, %f1018;
add.f32 %f1020, %f1016, %f1019;
sub.f32 %f1021, %f1016, %f1020;
add.f32 %f1022, %f1019, %f1021;
mul.rn.f32 %f1024, %f999, %f945;
mul.rn.f32 %f1026, %f999, %f947;
add.f32 %f1027, %f1024, %f1020;
sub.f32 %f1028, %f1024, %f1027;
add.f32 %f1029, %f1020, %f1028;
add.f32 %f1030, %f1022, %f1029;
add.f32 %f1031, %f1026, %f1030;
add.f32 %f1032, %f1027, %f1031;
sub.f32 %f1033, %f1027, %f1032;
add.f32 %f1034, %f1031, %f1033;
mul.rn.f32 %f1036, %f957, %f1032;
neg.f32 %f1037, %f1036;
fma.rn.f32 %f1038, %f957, %f1032, %f1037;
fma.rn.f32 %f1039, %f957, %f1034, %f1038;
fma.rn.f32 %f1041, %f962, %f1032, %f1039;
add.rn.f32 %f1042, %f1036, %f1041;
neg.f32 %f1043, %f1042;
add.rn.f32 %f1044, %f1036, %f1043;
add.rn.f32 %f1045, %f1044, %f1041;
mov.b32 %r309, %f1042;
setp.eq.s32 %p113, %r309, 1118925336;
add.s32 %r310, %r309, -1;
mov.b32 %f1046, %r310;
add.f32 %f1047, %f1045, 0f37000000;
selp.f32 %f1048, %f1046, %f1042, %p113;
selp.f32 %f318, %f1047, %f1045, %p113;
mul.f32 %f1049, %f1048, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1050, %f1049;
fma.rn.f32 %f1052, %f1050, %f973, %f1048;
fma.rn.f32 %f1054, %f1050, %f975, %f1052;
mul.f32 %f1055, %f1054, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1056, %f1055;
add.f32 %f1057, %f1050, 0f00000000;
ex2.approx.f32 %f1058, %f1057;
mul.f32 %f1059, %f1056, %f1058;
setp.lt.f32 %p114, %f1048, 0fC2D20000;
selp.f32 %f1060, 0f00000000, %f1059, %p114;
setp.gt.f32 %p115, %f1048, 0f42D20000;
selp.f32 %f1398, 0f7F800000, %f1060, %p115;
setp.eq.f32 %p116, %f1398, 0f7F800000;
@%p116 bra BB0_113;
fma.rn.f32 %f1398, %f1398, %f318, %f1398;
BB0_113:
setp.lt.f32 %p117, %f301, 0f00000000;
and.pred %p5, %p117, %p102;
mov.b32 %r311, %f1398;
xor.b32 %r312, %r311, -2147483648;
mov.b32 %f1061, %r312;
selp.f32 %f1400, %f1061, %f1398, %p5;
setp.eq.f32 %p119, %f301, 0f00000000;
@%p119 bra BB0_116;
bra.uni BB0_114;
BB0_116:
add.f32 %f1064, %f301, %f301;
selp.f32 %f1400, %f1064, 0f00000000, %p102;
bra.uni BB0_117;
BB0_114:
setp.geu.f32 %p120, %f301, 0f00000000;
@%p120 bra BB0_117;
cvt.rzi.f32.f32 %f1063, %f957;
setp.neu.f32 %p121, %f1063, 0f3EE66666;
selp.f32 %f1400, 0f7FFFFFFF, %f1400, %p121;
BB0_117:
add.f32 %f1065, %f317, 0f3EE66666;
mov.b32 %r313, %f1065;
setp.lt.s32 %p123, %r313, 2139095040;
@%p123 bra BB0_122;
setp.gtu.f32 %p124, %f317, 0f7F800000;
@%p124 bra BB0_121;
bra.uni BB0_119;
BB0_121:
add.f32 %f1400, %f301, 0f3EE66666;
bra.uni BB0_122;
BB0_119:
setp.neu.f32 %p125, %f317, 0f7F800000;
@%p125 bra BB0_122;
selp.f32 %f1400, 0fFF800000, 0f7F800000, %p5;
BB0_122:
setp.eq.f32 %p126, %f301, 0f3F800000;
selp.f32 %f329, 0f3F800000, %f1400, %p126;
abs.f32 %f330, %f302;
setp.lt.f32 %p127, %f330, 0f00800000;
mul.f32 %f1068, %f330, 0f4B800000;
selp.f32 %f1069, 0fC3170000, 0fC2FE0000, %p127;
selp.f32 %f1070, %f1068, %f330, %p127;
mov.b32 %r314, %f1070;
and.b32 %r315, %r314, 8388607;
or.b32 %r316, %r315, 1065353216;
mov.b32 %f1071, %r316;
shr.u32 %r317, %r314, 23;
cvt.rn.f32.u32 %f1072, %r317;
add.f32 %f1073, %f1069, %f1072;
setp.gt.f32 %p128, %f1071, 0f3FB504F3;
mul.f32 %f1074, %f1071, 0f3F000000;
add.f32 %f1075, %f1073, 0f3F800000;
selp.f32 %f1076, %f1074, %f1071, %p128;
selp.f32 %f1077, %f1075, %f1073, %p128;
add.f32 %f1078, %f1076, 0fBF800000;
add.f32 %f1067, %f1076, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1066,%f1067;
// inline asm
add.f32 %f1079, %f1078, %f1078;
mul.f32 %f1080, %f1066, %f1079;
mul.f32 %f1081, %f1080, %f1080;
fma.rn.f32 %f1084, %f927, %f1081, %f926;
fma.rn.f32 %f1086, %f1084, %f1081, %f929;
mul.rn.f32 %f1087, %f1086, %f1081;
mul.rn.f32 %f1088, %f1087, %f1080;
sub.f32 %f1089, %f1078, %f1080;
neg.f32 %f1090, %f1080;
add.f32 %f1091, %f1089, %f1089;
fma.rn.f32 %f1092, %f1090, %f1078, %f1091;
mul.rn.f32 %f1093, %f1066, %f1092;
add.f32 %f1094, %f1088, %f1080;
sub.f32 %f1095, %f1080, %f1094;
add.f32 %f1096, %f1088, %f1095;
add.f32 %f1097, %f1093, %f1096;
add.f32 %f1098, %f1094, %f1097;
sub.f32 %f1099, %f1094, %f1098;
add.f32 %f1100, %f1097, %f1099;
mul.rn.f32 %f1102, %f1077, %f945;
mul.rn.f32 %f1104, %f1077, %f947;
add.f32 %f1105, %f1102, %f1098;
sub.f32 %f1106, %f1102, %f1105;
add.f32 %f1107, %f1098, %f1106;
add.f32 %f1108, %f1100, %f1107;
add.f32 %f1109, %f1104, %f1108;
add.f32 %f1110, %f1105, %f1109;
sub.f32 %f1111, %f1105, %f1110;
add.f32 %f1112, %f1109, %f1111;
mul.rn.f32 %f1114, %f957, %f1110;
neg.f32 %f1115, %f1114;
fma.rn.f32 %f1116, %f957, %f1110, %f1115;
fma.rn.f32 %f1117, %f957, %f1112, %f1116;
fma.rn.f32 %f1119, %f962, %f1110, %f1117;
add.rn.f32 %f1120, %f1114, %f1119;
neg.f32 %f1121, %f1120;
add.rn.f32 %f1122, %f1114, %f1121;
add.rn.f32 %f1123, %f1122, %f1119;
mov.b32 %r318, %f1120;
setp.eq.s32 %p129, %r318, 1118925336;
add.s32 %r319, %r318, -1;
mov.b32 %f1124, %r319;
add.f32 %f1125, %f1123, 0f37000000;
selp.f32 %f1126, %f1124, %f1120, %p129;
selp.f32 %f331, %f1125, %f1123, %p129;
mul.f32 %f1127, %f1126, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1128, %f1127;
fma.rn.f32 %f1130, %f1128, %f973, %f1126;
fma.rn.f32 %f1132, %f1128, %f975, %f1130;
mul.f32 %f1133, %f1132, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1134, %f1133;
add.f32 %f1135, %f1128, 0f00000000;
ex2.approx.f32 %f1136, %f1135;
mul.f32 %f1137, %f1134, %f1136;
setp.lt.f32 %p130, %f1126, 0fC2D20000;
selp.f32 %f1138, 0f00000000, %f1137, %p130;
setp.gt.f32 %p131, %f1126, 0f42D20000;
selp.f32 %f1401, 0f7F800000, %f1138, %p131;
setp.eq.f32 %p132, %f1401, 0f7F800000;
@%p132 bra BB0_124;
fma.rn.f32 %f1401, %f1401, %f331, %f1401;
BB0_124:
setp.lt.f32 %p133, %f302, 0f00000000;
and.pred %p6, %p133, %p102;
mov.b32 %r320, %f1401;
xor.b32 %r321, %r320, -2147483648;
mov.b32 %f1139, %r321;
selp.f32 %f1403, %f1139, %f1401, %p6;
setp.eq.f32 %p135, %f302, 0f00000000;
@%p135 bra BB0_127;
bra.uni BB0_125;
BB0_127:
add.f32 %f1142, %f302, %f302;
selp.f32 %f1403, %f1142, 0f00000000, %p102;
bra.uni BB0_128;
BB0_125:
setp.geu.f32 %p136, %f302, 0f00000000;
@%p136 bra BB0_128;
cvt.rzi.f32.f32 %f1141, %f957;
setp.neu.f32 %p137, %f1141, 0f3EE66666;
selp.f32 %f1403, 0f7FFFFFFF, %f1403, %p137;
BB0_128:
add.f32 %f1143, %f330, 0f3EE66666;
mov.b32 %r322, %f1143;
setp.lt.s32 %p139, %r322, 2139095040;
@%p139 bra BB0_133;
setp.gtu.f32 %p140, %f330, 0f7F800000;
@%p140 bra BB0_132;
bra.uni BB0_130;
BB0_132:
add.f32 %f1403, %f302, 0f3EE66666;
bra.uni BB0_133;
BB0_130:
setp.neu.f32 %p141, %f330, 0f7F800000;
@%p141 bra BB0_133;
selp.f32 %f1403, 0fFF800000, 0f7F800000, %p6;
BB0_133:
mov.u32 %r395, 4;
setp.eq.f32 %p142, %f302, 0f3F800000;
selp.f32 %f1144, 0f3F800000, %f1403, %p142;
cvt.u64.u32 %rd118, %r6;
cvt.u64.u32 %rd117, %r5;
mov.u64 %rd121, image;
cvta.global.u64 %rd116, %rd121;
// inline asm
call (%rd115), _rt_buffer_get_64, (%rd116, %r109, %r395, %rd117, %rd118, %rd31, %rd31);
// inline asm
cvt.sat.f32.f32 %f1145, %f1144;
mul.f32 %f1146, %f1145, 0f437FFD71;
cvt.rzi.u32.f32 %r325, %f1146;
cvt.sat.f32.f32 %f1147, %f329;
mul.f32 %f1148, %f1147, 0f437FFD71;
cvt.rzi.u32.f32 %r326, %f1148;
cvt.sat.f32.f32 %f1149, %f316;
mul.f32 %f1150, %f1149, 0f437FFD71;
cvt.rzi.u32.f32 %r327, %f1150;
cvt.u16.u32 %rs43, %r325;
cvt.u16.u32 %rs44, %r327;
cvt.u16.u32 %rs45, %r326;
mov.u16 %rs46, 255;
st.v4.u8 [%rd115], {%rs43, %rs45, %rs44, %rs46};
ld.global.u32 %r426, [imageEnabled];
BB0_134:
cvt.u64.u32 %rd22, %r5;
cvt.u64.u32 %rd23, %r6;
and.b32 %r328, %r426, 4;
setp.eq.s32 %p143, %r328, 0;
@%p143 bra BB0_138;
ld.global.u32 %r329, [additive];
setp.eq.s32 %p144, %r329, 0;
// inline asm
{ cvt.rn.f16.f32 %rs47, %f665;}
// inline asm
@%p144 bra BB0_137;
mov.u64 %rd134, image_HDR;
cvta.global.u64 %rd123, %rd134;
mov.u32 %r333, 8;
// inline asm
call (%rd122), _rt_buffer_get_64, (%rd123, %r109, %r333, %rd22, %rd23, %rd31, %rd31);
// inline asm
ld.v4.u16 {%rs54, %rs55, %rs56, %rs57}, [%rd122];
// inline asm
{ cvt.f32.f16 %f1152, %rs54;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1153, %rs55;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1154, %rs56;}
// inline asm
// inline asm
call (%rd128), _rt_buffer_get_64, (%rd123, %r109, %r333, %rd22, %rd23, %rd31, %rd31);
// inline asm
add.f32 %f1155, %f300, %f1152;
add.f32 %f1156, %f301, %f1153;
add.f32 %f1157, %f302, %f1154;
// inline asm
{ cvt.rn.f16.f32 %rs53, %f1157;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs52, %f1156;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs51, %f1155;}
// inline asm
st.v4.u16 [%rd128], {%rs51, %rs52, %rs53, %rs47};
bra.uni BB0_138;
BB0_137:
mov.u64 %rd141, image_HDR;
cvta.global.u64 %rd136, %rd141;
mov.u32 %r335, 8;
// inline asm
call (%rd135), _rt_buffer_get_64, (%rd136, %r109, %r335, %rd22, %rd23, %rd31, %rd31);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs60, %f302;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs59, %f301;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs58, %f300;}
// inline asm
st.v4.u16 [%rd135], {%rs58, %rs59, %rs60, %rs47};
BB0_138:
add.f32 %f1162, %f1382, %f1385;
add.f32 %f1163, %f1379, %f1162;
add.f32 %f1164, %f1383, %f1386;
add.f32 %f1165, %f1380, %f1164;
add.f32 %f1166, %f1384, %f1387;
add.f32 %f1167, %f1381, %f1166;
mul.f32 %f1168, %f1163, 0f3F13CD3A;
mul.f32 %f1169, %f1165, 0f3F13CD3A;
mul.f32 %f1170, %f1167, 0f3F13CD3A;
div.rn.f32 %f1171, %f1388, %f1168;
div.rn.f32 %f1172, %f1389, %f1169;
div.rn.f32 %f1173, %f1390, %f1170;
setp.eq.f32 %p145, %f1388, 0f00000000;
selp.f32 %f1174, 0f00000000, %f1171, %p145;
setp.eq.f32 %p146, %f1389, 0f00000000;
selp.f32 %f1175, 0f00000000, %f1172, %p146;
setp.eq.f32 %p147, %f1390, 0f00000000;
selp.f32 %f1176, 0f00000000, %f1173, %p147;
mul.f32 %f1177, %f1385, %f1174;
mul.f32 %f1178, %f1386, %f1175;
mul.f32 %f1179, %f1387, %f1176;
mul.f32 %f1180, %f1382, %f1174;
mul.f32 %f1181, %f1383, %f1175;
mul.f32 %f1182, %f1384, %f1176;
mul.f32 %f1183, %f1379, %f1174;
mul.f32 %f1184, %f1380, %f1175;
mul.f32 %f1185, %f1381, %f1176;
ld.global.f32 %f1186, [lightColor];
mul.f32 %f342, %f1177, %f1186;
ld.global.f32 %f1187, [lightColor+4];
mul.f32 %f343, %f1178, %f1187;
ld.global.f32 %f1188, [lightColor+8];
mul.f32 %f344, %f1179, %f1188;
mul.f32 %f345, %f1180, %f1186;
mul.f32 %f346, %f1181, %f1187;
mul.f32 %f347, %f1182, %f1188;
mul.f32 %f348, %f1183, %f1186;
mul.f32 %f349, %f1184, %f1187;
mul.f32 %f350, %f1185, %f1188;
ld.global.u32 %r336, [additive];
setp.eq.s32 %p148, %r336, 0;
// inline asm
{ cvt.rn.f16.f32 %rs61, %f665;}
// inline asm
@%p148 bra BB0_140;
mov.u64 %rd154, image_RNM0;
cvta.global.u64 %rd143, %rd154;
mov.u32 %r340, 8;
// inline asm
call (%rd142), _rt_buffer_get_64, (%rd143, %r109, %r340, %rd22, %rd23, %rd31, %rd31);
// inline asm
ld.v4.u16 {%rs68, %rs69, %rs70, %rs71}, [%rd142];
// inline asm
{ cvt.f32.f16 %f1189, %rs68;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1190, %rs69;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1191, %rs70;}
// inline asm
// inline asm
call (%rd148), _rt_buffer_get_64, (%rd143, %r109, %r340, %rd22, %rd23, %rd31, %rd31);
// inline asm
add.f32 %f1192, %f342, %f1189;
add.f32 %f1193, %f343, %f1190;
add.f32 %f1194, %f344, %f1191;
// inline asm
{ cvt.rn.f16.f32 %rs67, %f1194;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs66, %f1193;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs65, %f1192;}
// inline asm
st.v4.u16 [%rd148], {%rs65, %rs66, %rs67, %rs61};
bra.uni BB0_141;
BB0_140:
mov.u64 %rd161, image_RNM0;
cvta.global.u64 %rd156, %rd161;
mov.u32 %r342, 8;
// inline asm
call (%rd155), _rt_buffer_get_64, (%rd156, %r109, %r342, %rd22, %rd23, %rd31, %rd31);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs74, %f344;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs73, %f343;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs72, %f342;}
// inline asm
st.v4.u16 [%rd155], {%rs72, %rs73, %rs74, %rs61};
BB0_141:
ld.global.u32 %r343, [additive];
setp.eq.s32 %p149, %r343, 0;
// inline asm
{ cvt.rn.f16.f32 %rs75, %f665;}
// inline asm
@%p149 bra BB0_143;
mov.u64 %rd174, image_RNM1;
cvta.global.u64 %rd163, %rd174;
mov.u32 %r347, 8;
// inline asm
call (%rd162), _rt_buffer_get_64, (%rd163, %r109, %r347, %rd22, %rd23, %rd31, %rd31);
// inline asm
ld.v4.u16 {%rs82, %rs83, %rs84, %rs85}, [%rd162];
// inline asm
{ cvt.f32.f16 %f1199, %rs82;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1200, %rs83;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1201, %rs84;}
// inline asm
// inline asm
call (%rd168), _rt_buffer_get_64, (%rd163, %r109, %r347, %rd22, %rd23, %rd31, %rd31);
// inline asm
add.f32 %f1202, %f345, %f1199;
add.f32 %f1203, %f346, %f1200;
add.f32 %f1204, %f347, %f1201;
// inline asm
{ cvt.rn.f16.f32 %rs81, %f1204;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs80, %f1203;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs79, %f1202;}
// inline asm
st.v4.u16 [%rd168], {%rs79, %rs80, %rs81, %rs75};
bra.uni BB0_144;
BB0_143:
mov.u64 %rd181, image_RNM1;
cvta.global.u64 %rd176, %rd181;
mov.u32 %r349, 8;
// inline asm
call (%rd175), _rt_buffer_get_64, (%rd176, %r109, %r349, %rd22, %rd23, %rd31, %rd31);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs88, %f347;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs87, %f346;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs86, %f345;}
// inline asm
st.v4.u16 [%rd175], {%rs86, %rs87, %rs88, %rs75};
BB0_144:
ld.global.u32 %r350, [additive];
setp.eq.s32 %p150, %r350, 0;
// inline asm
{ cvt.rn.f16.f32 %rs89, %f665;}
// inline asm
@%p150 bra BB0_146;
mov.u64 %rd194, image_RNM2;
cvta.global.u64 %rd183, %rd194;
mov.u32 %r354, 8;
// inline asm
call (%rd182), _rt_buffer_get_64, (%rd183, %r109, %r354, %rd22, %rd23, %rd31, %rd31);
// inline asm
ld.v4.u16 {%rs96, %rs97, %rs98, %rs99}, [%rd182];
// inline asm
{ cvt.f32.f16 %f1209, %rs96;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1210, %rs97;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1211, %rs98;}
// inline asm
// inline asm
call (%rd188), _rt_buffer_get_64, (%rd183, %r109, %r354, %rd22, %rd23, %rd31, %rd31);
// inline asm
add.f32 %f1212, %f348, %f1209;
add.f32 %f1213, %f349, %f1210;
add.f32 %f1214, %f350, %f1211;
// inline asm
{ cvt.rn.f16.f32 %rs95, %f1214;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs94, %f1213;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs93, %f1212;}
// inline asm
st.v4.u16 [%rd188], {%rs93, %rs94, %rs95, %rs89};
bra.uni BB0_164;
BB0_146:
mov.u64 %rd201, image_RNM2;
cvta.global.u64 %rd196, %rd201;
mov.u32 %r356, 8;
// inline asm
call (%rd195), _rt_buffer_get_64, (%rd196, %r109, %r356, %rd22, %rd23, %rd31, %rd31);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs102, %f350;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs101, %f349;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs100, %f348;}
// inline asm
st.v4.u16 [%rd195], {%rs100, %rs101, %rs102, %rs89};
BB0_164:
ret;
}