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

1235 lines
35 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_Mask[1];
.global .align 1 .b8 image_HDR[1];
.global .align 1 .b8 image_HDR2[1];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 1 .b8 rnd_seeds[1];
.global .align 4 .u32 samples;
.global .align 4 .f32 lightRadius;
.global .align 4 .f32 intensity;
.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_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo11lightRadiusE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9intensityE[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 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename11lightRadiusE[6] = {102, 108, 111, 97, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename9intensityE[6] = {102, 108, 111, 97, 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_typeenum7samplesE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum11lightRadiusE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9intensityE = 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_semantic7samplesE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic11lightRadiusE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9intensityE[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_annotation7samplesE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation11lightRadiusE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9intensityE[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[32];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<91>;
.reg .b16 %rs<47>;
.reg .f32 %f<453>;
.reg .b32 %r<294>;
.reg .b64 %rd<112>;
mov.u64 %rd111, __local_depot0;
cvta.local.u64 %SP, %rd111;
ld.global.v2.u32 {%r91, %r92}, [pixelID];
cvt.u64.u32 %rd22, %r91;
cvt.u64.u32 %rd23, %r92;
mov.u64 %rd26, uvnormal;
cvta.global.u64 %rd21, %rd26;
mov.u32 %r89, 2;
mov.u32 %r90, 4;
mov.u64 %rd25, 0;
// inline asm
call (%rd20), _rt_buffer_get_64, (%rd21, %r89, %r90, %rd22, %rd23, %rd25, %rd25);
// inline asm
ld.u32 %r1, [%rd20];
shr.u32 %r95, %r1, 16;
cvt.u16.u32 %rs1, %r95;
and.b16 %rs3, %rs1, 255;
cvt.u16.u32 %rs4, %r1;
or.b16 %rs5, %rs4, %rs3;
setp.eq.s16 %p3, %rs5, 0;
mov.f32 %f427, 0f00000000;
mov.f32 %f428, %f427;
mov.f32 %f429, %f427;
@%p3 bra BB0_2;
ld.u8 %rs6, [%rd20+1];
and.b16 %rs8, %rs4, 255;
cvt.rn.f32.u16 %f102, %rs8;
div.rn.f32 %f103, %f102, 0f437F0000;
fma.rn.f32 %f104, %f103, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f105, %rs6;
div.rn.f32 %f106, %f105, 0f437F0000;
fma.rn.f32 %f107, %f106, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f108, %rs3;
div.rn.f32 %f109, %f108, 0f437F0000;
fma.rn.f32 %f110, %f109, 0f40000000, 0fBF800000;
mul.f32 %f111, %f107, %f107;
fma.rn.f32 %f112, %f104, %f104, %f111;
fma.rn.f32 %f113, %f110, %f110, %f112;
sqrt.rn.f32 %f114, %f113;
rcp.rn.f32 %f115, %f114;
mul.f32 %f427, %f104, %f115;
mul.f32 %f428, %f107, %f115;
mul.f32 %f429, %f110, %f115;
BB0_2:
ld.global.v2.u32 {%r96, %r97}, [pixelID];
ld.global.v2.u32 {%r99, %r100}, [tileInfo];
add.s32 %r2, %r96, %r99;
add.s32 %r3, %r97, %r100;
setp.eq.f32 %p4, %f428, 0f00000000;
setp.eq.f32 %p5, %f427, 0f00000000;
and.pred %p6, %p5, %p4;
setp.eq.f32 %p7, %f429, 0f00000000;
and.pred %p8, %p6, %p7;
@%p8 bra BB0_90;
bra.uni BB0_3;
BB0_90:
ld.global.u8 %rs28, [imageEnabled];
and.b16 %rs29, %rs28, 4;
setp.eq.s16 %p88, %rs29, 0;
@%p88 bra BB0_94;
ld.global.u32 %r256, [additive];
setp.eq.s32 %p89, %r256, 0;
cvt.u64.u32 %rd18, %r2;
cvt.u64.u32 %rd19, %r3;
@%p89 bra BB0_93;
mov.u64 %rd92, image_HDR;
cvta.global.u64 %rd81, %rd92;
mov.u32 %r260, 8;
// inline asm
call (%rd80), _rt_buffer_get_64, (%rd81, %r89, %r260, %rd18, %rd19, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs36, %rs37, %rs38, %rs39}, [%rd80];
// inline asm
{ cvt.f32.f16 %f417, %rs36;}
// inline asm
// inline asm
{ cvt.f32.f16 %f418, %rs37;}
// inline asm
// inline asm
{ cvt.f32.f16 %f419, %rs38;}
// inline asm
// inline asm
call (%rd86), _rt_buffer_get_64, (%rd81, %r89, %r260, %rd18, %rd19, %rd25, %rd25);
// inline asm
add.f32 %f420, %f417, 0f00000000;
add.f32 %f421, %f418, 0f00000000;
add.f32 %f422, %f419, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs35, %f422;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs34, %f421;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs33, %f420;}
// inline asm
mov.u16 %rs40, 0;
st.v4.u16 [%rd86], {%rs33, %rs34, %rs35, %rs40};
bra.uni BB0_94;
BB0_3:
ld.global.v2.u32 {%r108, %r109}, [pixelID];
cvt.u64.u32 %rd29, %r108;
cvt.u64.u32 %rd30, %r109;
mov.u64 %rd39, uvpos;
cvta.global.u64 %rd28, %rd39;
mov.u32 %r105, 12;
// inline asm
call (%rd27), _rt_buffer_get_64, (%rd28, %r89, %r105, %rd29, %rd30, %rd25, %rd25);
// inline asm
ld.f32 %f117, [%rd27+8];
ld.f32 %f118, [%rd27+4];
ld.f32 %f119, [%rd27];
mul.f32 %f120, %f119, 0f3456BF95;
mul.f32 %f121, %f118, 0f3456BF95;
mul.f32 %f122, %f117, 0f3456BF95;
abs.f32 %f123, %f427;
div.rn.f32 %f124, %f120, %f123;
abs.f32 %f125, %f428;
div.rn.f32 %f126, %f121, %f125;
abs.f32 %f127, %f429;
div.rn.f32 %f128, %f122, %f127;
abs.f32 %f129, %f124;
abs.f32 %f130, %f126;
abs.f32 %f131, %f128;
mov.f32 %f132, 0f38D1B717;
max.f32 %f133, %f129, %f132;
max.f32 %f134, %f130, %f132;
max.f32 %f135, %f131, %f132;
fma.rn.f32 %f7, %f427, %f133, %f119;
fma.rn.f32 %f8, %f428, %f134, %f118;
fma.rn.f32 %f9, %f429, %f135, %f117;
setp.gt.f32 %p9, %f123, %f127;
neg.f32 %f136, %f428;
selp.f32 %f137, %f136, 0f00000000, %p9;
neg.f32 %f138, %f429;
selp.f32 %f139, %f427, %f138, %p9;
selp.f32 %f140, 0f00000000, %f428, %p9;
mul.f32 %f141, %f139, %f139;
fma.rn.f32 %f142, %f137, %f137, %f141;
fma.rn.f32 %f143, %f140, %f140, %f142;
sqrt.rn.f32 %f144, %f143;
rcp.rn.f32 %f145, %f144;
mul.f32 %f10, %f137, %f145;
mul.f32 %f11, %f139, %f145;
mul.f32 %f12, %f140, %f145;
ld.global.v2.u32 {%r112, %r113}, [pixelID];
cvt.u64.u32 %rd35, %r112;
cvt.u64.u32 %rd36, %r113;
mov.u64 %rd40, rnd_seeds;
cvta.global.u64 %rd34, %rd40;
// inline asm
call (%rd33), _rt_buffer_get_64, (%rd34, %r89, %r90, %rd35, %rd36, %rd25, %rd25);
// inline asm
ld.global.u32 %r266, [samples];
mov.f32 %f444, 0f00000000;
setp.lt.s32 %p10, %r266, 1;
@%p10 bra BB0_55;
cvt.rn.f32.s32 %f147, %r266;
rcp.rn.f32 %f13, %f147;
ld.u32 %r292, [%rd33];
mul.f32 %f14, %f7, 0f3456BF95;
mul.f32 %f15, %f8, 0f3456BF95;
mul.f32 %f16, %f9, 0f3456BF95;
mul.f32 %f148, %f427, %f11;
mul.f32 %f149, %f428, %f10;
sub.f32 %f17, %f149, %f148;
mul.f32 %f150, %f429, %f10;
mul.f32 %f151, %f427, %f12;
sub.f32 %f18, %f151, %f150;
mul.f32 %f152, %f428, %f12;
mul.f32 %f153, %f429, %f11;
sub.f32 %f19, %f153, %f152;
mov.f32 %f444, 0f00000000;
mov.u32 %r267, 0;
abs.f32 %f154, %f15;
abs.f32 %f155, %f14;
max.f32 %f156, %f155, %f154;
abs.f32 %f157, %f16;
max.f32 %f158, %f156, %f157;
BB0_5:
setp.lt.s32 %p11, %r266, 1;
@%p11 bra BB0_54;
cvt.rn.f32.s32 %f21, %r267;
max.f32 %f22, %f158, %f132;
mov.u32 %r269, 0;
BB0_7:
mad.lo.s32 %r118, %r292, 1664525, 1013904223;
and.b32 %r119, %r118, 16777215;
cvt.rn.f32.u32 %f160, %r119;
fma.rn.f32 %f161, %f160, 0f33800000, %f21;
mul.f32 %f162, %f13, %f161;
mad.lo.s32 %r292, %r118, 1664525, 1013904223;
and.b32 %r120, %r292, 16777215;
cvt.rn.f32.u32 %f163, %r120;
cvt.rn.f32.s32 %f164, %r269;
fma.rn.f32 %f165, %f163, 0f33800000, %f164;
mul.f32 %f166, %f13, %f165;
sqrt.rn.f32 %f24, %f162;
mul.f32 %f438, %f166, 0f40C90FDB;
abs.f32 %f26, %f438;
setp.neu.f32 %p12, %f26, 0f7F800000;
mov.f32 %f432, %f438;
@%p12 bra BB0_9;
mov.f32 %f167, 0f00000000;
mul.rn.f32 %f432, %f438, %f167;
BB0_9:
mul.f32 %f168, %f432, 0f3F22F983;
cvt.rni.s32.f32 %r280, %f168;
cvt.rn.f32.s32 %f169, %r280;
neg.f32 %f170, %f169;
mov.f32 %f171, 0f3FC90FDA;
fma.rn.f32 %f172, %f170, %f171, %f432;
mov.f32 %f173, 0f33A22168;
fma.rn.f32 %f174, %f170, %f173, %f172;
mov.f32 %f175, 0f27C234C5;
fma.rn.f32 %f433, %f170, %f175, %f174;
abs.f32 %f176, %f432;
setp.leu.f32 %p13, %f176, 0f47CE4780;
@%p13 bra BB0_20;
mov.b32 %r13, %f432;
shr.u32 %r14, %r13, 23;
shl.b32 %r123, %r13, 8;
or.b32 %r15, %r123, -2147483648;
add.u64 %rd42, %SP, 0;
cvta.to.local.u64 %rd108, %rd42;
mov.u32 %r272, 0;
mov.u64 %rd107, __cudart_i2opi_f;
mov.u32 %r271, -6;
BB0_11:
.pragma "nounroll";
ld.const.u32 %r126, [%rd107];
// inline asm
{
mad.lo.cc.u32 %r124, %r126, %r15, %r272;
madc.hi.u32 %r272, %r126, %r15, 0;
}
// inline asm
st.local.u32 [%rd108], %r124;
add.s64 %rd108, %rd108, 4;
add.s64 %rd107, %rd107, 4;
add.s32 %r271, %r271, 1;
setp.ne.s32 %p14, %r271, 0;
@%p14 bra BB0_11;
and.b32 %r129, %r14, 255;
add.s32 %r130, %r129, -128;
shr.u32 %r131, %r130, 5;
and.b32 %r20, %r13, -2147483648;
cvta.to.local.u64 %rd44, %rd42;
st.local.u32 [%rd44+24], %r272;
mov.u32 %r132, 6;
sub.s32 %r133, %r132, %r131;
mul.wide.s32 %rd45, %r133, 4;
add.s64 %rd8, %rd44, %rd45;
ld.local.u32 %r273, [%rd8];
ld.local.u32 %r274, [%rd8+-4];
and.b32 %r23, %r14, 31;
setp.eq.s32 %p15, %r23, 0;
@%p15 bra BB0_14;
mov.u32 %r134, 32;
sub.s32 %r135, %r134, %r23;
shr.u32 %r136, %r274, %r135;
shl.b32 %r137, %r273, %r23;
add.s32 %r273, %r136, %r137;
ld.local.u32 %r138, [%rd8+-8];
shr.u32 %r139, %r138, %r135;
shl.b32 %r140, %r274, %r23;
add.s32 %r274, %r139, %r140;
BB0_14:
shr.u32 %r141, %r274, 30;
shl.b32 %r142, %r273, 2;
add.s32 %r275, %r141, %r142;
shl.b32 %r29, %r274, 2;
shr.u32 %r143, %r275, 31;
shr.u32 %r144, %r273, 30;
add.s32 %r30, %r143, %r144;
setp.eq.s32 %p16, %r143, 0;
@%p16 bra BB0_15;
bra.uni BB0_16;
BB0_15:
mov.u32 %r276, %r20;
mov.u32 %r277, %r29;
bra.uni BB0_17;
BB0_16:
not.b32 %r145, %r275;
neg.s32 %r277, %r29;
setp.eq.s32 %p17, %r29, 0;
selp.u32 %r146, 1, 0, %p17;
add.s32 %r275, %r146, %r145;
xor.b32 %r276, %r20, -2147483648;
BB0_17:
clz.b32 %r279, %r275;
setp.eq.s32 %p18, %r279, 0;
shl.b32 %r147, %r275, %r279;
mov.u32 %r148, 32;
sub.s32 %r149, %r148, %r279;
shr.u32 %r150, %r277, %r149;
add.s32 %r151, %r150, %r147;
selp.b32 %r38, %r275, %r151, %p18;
mov.u32 %r152, -921707870;
mul.hi.u32 %r278, %r38, %r152;
setp.eq.s32 %p19, %r20, 0;
neg.s32 %r153, %r30;
selp.b32 %r280, %r30, %r153, %p19;
setp.lt.s32 %p20, %r278, 1;
@%p20 bra BB0_19;
mul.lo.s32 %r154, %r38, -921707870;
shr.u32 %r155, %r154, 31;
shl.b32 %r156, %r278, 1;
add.s32 %r278, %r155, %r156;
add.s32 %r279, %r279, 1;
BB0_19:
mov.u32 %r157, 126;
sub.s32 %r158, %r157, %r279;
shl.b32 %r159, %r158, 23;
add.s32 %r160, %r278, 1;
shr.u32 %r161, %r160, 7;
add.s32 %r162, %r161, 1;
shr.u32 %r163, %r162, 1;
add.s32 %r164, %r163, %r159;
or.b32 %r165, %r164, %r276;
mov.b32 %f433, %r165;
BB0_20:
mul.rn.f32 %f32, %f433, %f433;
add.s32 %r46, %r280, 1;
and.b32 %r47, %r46, 1;
setp.eq.s32 %p21, %r47, 0;
@%p21 bra BB0_22;
bra.uni BB0_21;
BB0_22:
mov.f32 %f179, 0f3C08839E;
mov.f32 %f180, 0fB94CA1F9;
fma.rn.f32 %f434, %f180, %f32, %f179;
bra.uni BB0_23;
BB0_21:
mov.f32 %f177, 0fBAB6061A;
mov.f32 %f178, 0f37CCF5CE;
fma.rn.f32 %f434, %f178, %f32, %f177;
BB0_23:
@%p21 bra BB0_25;
bra.uni BB0_24;
BB0_25:
mov.f32 %f184, 0fBE2AAAA3;
fma.rn.f32 %f185, %f434, %f32, %f184;
mov.f32 %f186, 0f00000000;
fma.rn.f32 %f435, %f185, %f32, %f186;
bra.uni BB0_26;
BB0_24:
mov.f32 %f181, 0f3D2AAAA5;
fma.rn.f32 %f182, %f434, %f32, %f181;
mov.f32 %f183, 0fBF000000;
fma.rn.f32 %f435, %f182, %f32, %f183;
BB0_26:
fma.rn.f32 %f436, %f435, %f433, %f433;
@%p21 bra BB0_28;
mov.f32 %f187, 0f3F800000;
fma.rn.f32 %f436, %f435, %f32, %f187;
BB0_28:
and.b32 %r166, %r46, 2;
setp.eq.s32 %p24, %r166, 0;
@%p24 bra BB0_30;
mov.f32 %f188, 0f00000000;
mov.f32 %f189, 0fBF800000;
fma.rn.f32 %f436, %f436, %f189, %f188;
BB0_30:
@%p12 bra BB0_32;
mov.f32 %f190, 0f00000000;
mul.rn.f32 %f438, %f438, %f190;
BB0_32:
mul.f32 %f191, %f438, 0f3F22F983;
cvt.rni.s32.f32 %r290, %f191;
cvt.rn.f32.s32 %f192, %r290;
neg.f32 %f193, %f192;
fma.rn.f32 %f195, %f193, %f171, %f438;
fma.rn.f32 %f197, %f193, %f173, %f195;
fma.rn.f32 %f439, %f193, %f175, %f197;
abs.f32 %f199, %f438;
setp.leu.f32 %p26, %f199, 0f47CE4780;
@%p26 bra BB0_43;
mov.b32 %r49, %f438;
shr.u32 %r50, %r49, 23;
shl.b32 %r169, %r49, 8;
or.b32 %r51, %r169, -2147483648;
add.u64 %rd47, %SP, 0;
cvta.to.local.u64 %rd110, %rd47;
mov.u32 %r282, 0;
mov.u64 %rd109, __cudart_i2opi_f;
mov.u32 %r281, -6;
BB0_34:
.pragma "nounroll";
ld.const.u32 %r172, [%rd109];
// inline asm
{
mad.lo.cc.u32 %r170, %r172, %r51, %r282;
madc.hi.u32 %r282, %r172, %r51, 0;
}
// inline asm
st.local.u32 [%rd110], %r170;
add.s64 %rd110, %rd110, 4;
add.s64 %rd109, %rd109, 4;
add.s32 %r281, %r281, 1;
setp.ne.s32 %p27, %r281, 0;
@%p27 bra BB0_34;
and.b32 %r175, %r50, 255;
add.s32 %r176, %r175, -128;
shr.u32 %r177, %r176, 5;
and.b32 %r56, %r49, -2147483648;
cvta.to.local.u64 %rd49, %rd47;
st.local.u32 [%rd49+24], %r282;
mov.u32 %r178, 6;
sub.s32 %r179, %r178, %r177;
mul.wide.s32 %rd50, %r179, 4;
add.s64 %rd14, %rd49, %rd50;
ld.local.u32 %r283, [%rd14];
ld.local.u32 %r284, [%rd14+-4];
and.b32 %r59, %r50, 31;
setp.eq.s32 %p28, %r59, 0;
@%p28 bra BB0_37;
mov.u32 %r180, 32;
sub.s32 %r181, %r180, %r59;
shr.u32 %r182, %r284, %r181;
shl.b32 %r183, %r283, %r59;
add.s32 %r283, %r182, %r183;
ld.local.u32 %r184, [%rd14+-8];
shr.u32 %r185, %r184, %r181;
shl.b32 %r186, %r284, %r59;
add.s32 %r284, %r185, %r186;
BB0_37:
shr.u32 %r187, %r284, 30;
shl.b32 %r188, %r283, 2;
add.s32 %r285, %r187, %r188;
shl.b32 %r65, %r284, 2;
shr.u32 %r189, %r285, 31;
shr.u32 %r190, %r283, 30;
add.s32 %r66, %r189, %r190;
setp.eq.s32 %p29, %r189, 0;
@%p29 bra BB0_38;
bra.uni BB0_39;
BB0_38:
mov.u32 %r286, %r56;
mov.u32 %r287, %r65;
bra.uni BB0_40;
BB0_39:
not.b32 %r191, %r285;
neg.s32 %r287, %r65;
setp.eq.s32 %p30, %r65, 0;
selp.u32 %r192, 1, 0, %p30;
add.s32 %r285, %r192, %r191;
xor.b32 %r286, %r56, -2147483648;
BB0_40:
clz.b32 %r289, %r285;
setp.eq.s32 %p31, %r289, 0;
shl.b32 %r193, %r285, %r289;
mov.u32 %r194, 32;
sub.s32 %r195, %r194, %r289;
shr.u32 %r196, %r287, %r195;
add.s32 %r197, %r196, %r193;
selp.b32 %r74, %r285, %r197, %p31;
mov.u32 %r198, -921707870;
mul.hi.u32 %r288, %r74, %r198;
setp.eq.s32 %p32, %r56, 0;
neg.s32 %r199, %r66;
selp.b32 %r290, %r66, %r199, %p32;
setp.lt.s32 %p33, %r288, 1;
@%p33 bra BB0_42;
mul.lo.s32 %r200, %r74, -921707870;
shr.u32 %r201, %r200, 31;
shl.b32 %r202, %r288, 1;
add.s32 %r288, %r201, %r202;
add.s32 %r289, %r289, 1;
BB0_42:
mov.u32 %r203, 126;
sub.s32 %r204, %r203, %r289;
shl.b32 %r205, %r204, 23;
add.s32 %r206, %r288, 1;
shr.u32 %r207, %r206, 7;
add.s32 %r208, %r207, 1;
shr.u32 %r209, %r208, 1;
add.s32 %r210, %r209, %r205;
or.b32 %r211, %r210, %r286;
mov.b32 %f439, %r211;
BB0_43:
mul.rn.f32 %f49, %f439, %f439;
and.b32 %r82, %r290, 1;
setp.eq.s32 %p34, %r82, 0;
@%p34 bra BB0_45;
bra.uni BB0_44;
BB0_45:
mov.f32 %f202, 0f3C08839E;
mov.f32 %f203, 0fB94CA1F9;
fma.rn.f32 %f440, %f203, %f49, %f202;
bra.uni BB0_46;
BB0_44:
mov.f32 %f200, 0fBAB6061A;
mov.f32 %f201, 0f37CCF5CE;
fma.rn.f32 %f440, %f201, %f49, %f200;
BB0_46:
@%p34 bra BB0_48;
bra.uni BB0_47;
BB0_48:
mov.f32 %f207, 0fBE2AAAA3;
fma.rn.f32 %f208, %f440, %f49, %f207;
mov.f32 %f209, 0f00000000;
fma.rn.f32 %f441, %f208, %f49, %f209;
bra.uni BB0_49;
BB0_47:
mov.f32 %f204, 0f3D2AAAA5;
fma.rn.f32 %f205, %f440, %f49, %f204;
mov.f32 %f206, 0fBF000000;
fma.rn.f32 %f441, %f205, %f49, %f206;
BB0_49:
fma.rn.f32 %f442, %f441, %f439, %f439;
@%p34 bra BB0_51;
mov.f32 %f210, 0f3F800000;
fma.rn.f32 %f442, %f441, %f49, %f210;
BB0_51:
and.b32 %r212, %r290, 2;
setp.eq.s32 %p37, %r212, 0;
@%p37 bra BB0_53;
mov.f32 %f211, 0f00000000;
mov.f32 %f212, 0fBF800000;
fma.rn.f32 %f442, %f442, %f212, %f211;
BB0_53:
mul.f32 %f221, %f24, %f436;
add.u64 %rd51, %SP, 28;
cvta.to.local.u64 %rd52, %rd51;
mul.f32 %f222, %f221, %f221;
mov.f32 %f223, 0f3F800000;
sub.f32 %f224, %f223, %f222;
mul.f32 %f225, %f24, %f442;
mul.f32 %f226, %f225, %f225;
sub.f32 %f227, %f224, %f226;
mov.f32 %f228, 0f00000000;
max.f32 %f229, %f228, %f227;
sqrt.rn.f32 %f230, %f229;
mul.f32 %f231, %f10, %f225;
mul.f32 %f232, %f11, %f225;
mul.f32 %f233, %f12, %f225;
fma.rn.f32 %f234, %f19, %f221, %f231;
fma.rn.f32 %f235, %f18, %f221, %f232;
fma.rn.f32 %f236, %f17, %f221, %f233;
fma.rn.f32 %f216, %f427, %f230, %f234;
fma.rn.f32 %f217, %f428, %f230, %f235;
fma.rn.f32 %f218, %f429, %f230, %f236;
ld.global.f32 %f220, [lightRadius];
mov.u32 %r216, 1065353216;
st.local.u32 [%rd52], %r216;
ld.global.u32 %r213, [root];
mov.u32 %r214, 1;
// inline asm
call _rt_trace_64, (%r213, %f7, %f8, %f9, %f216, %f217, %f218, %r214, %f22, %f220, %rd51, %r90);
// inline asm
ld.local.f32 %f237, [%rd52];
add.f32 %f444, %f444, %f237;
ld.global.u32 %r266, [samples];
add.s32 %r269, %r269, 1;
setp.lt.s32 %p38, %r269, %r266;
@%p38 bra BB0_7;
BB0_54:
add.s32 %r267, %r267, 1;
setp.lt.s32 %p39, %r267, %r266;
@%p39 bra BB0_5;
BB0_55:
mul.lo.s32 %r217, %r266, %r266;
cvt.rn.f32.s32 %f238, %r217;
div.rn.f32 %f64, %f444, %f238;
ld.global.f32 %f65, [intensity];
setp.lt.f32 %p40, %f65, 0f00000000;
@%p40 bra BB0_57;
bra.uni BB0_56;
BB0_57:
neg.f32 %f67, %f65;
setp.lt.f32 %p41, %f65, 0fBF800000;
@%p41 bra BB0_59;
bra.uni BB0_58;
BB0_59:
mul.f32 %f245, %f65, 0fBF000000;
cvt.rzi.f32.f32 %f246, %f245;
fma.rn.f32 %f247, %f246, 0fC0000000, %f67;
abs.f32 %f69, %f247;
abs.f32 %f70, %f64;
setp.lt.f32 %p42, %f70, 0f00800000;
mul.f32 %f248, %f70, 0f4B800000;
selp.f32 %f249, 0fC3170000, 0fC2FE0000, %p42;
selp.f32 %f250, %f248, %f70, %p42;
mov.b32 %r218, %f250;
and.b32 %r219, %r218, 8388607;
or.b32 %r220, %r219, 1065353216;
mov.b32 %f251, %r220;
shr.u32 %r221, %r218, 23;
cvt.rn.f32.u32 %f252, %r221;
add.f32 %f253, %f249, %f252;
setp.gt.f32 %p43, %f251, 0f3FB504F3;
mul.f32 %f254, %f251, 0f3F000000;
add.f32 %f255, %f253, 0f3F800000;
selp.f32 %f256, %f254, %f251, %p43;
selp.f32 %f257, %f255, %f253, %p43;
add.f32 %f258, %f256, 0fBF800000;
add.f32 %f244, %f256, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f243,%f244;
// inline asm
add.f32 %f259, %f258, %f258;
mul.f32 %f260, %f243, %f259;
mul.f32 %f261, %f260, %f260;
mov.f32 %f262, 0f3C4CAF63;
mov.f32 %f263, 0f3B18F0FE;
fma.rn.f32 %f264, %f263, %f261, %f262;
mov.f32 %f265, 0f3DAAAABD;
fma.rn.f32 %f266, %f264, %f261, %f265;
mul.rn.f32 %f267, %f266, %f261;
mul.rn.f32 %f268, %f267, %f260;
sub.f32 %f269, %f258, %f260;
neg.f32 %f270, %f260;
add.f32 %f271, %f269, %f269;
fma.rn.f32 %f272, %f270, %f258, %f271;
mul.rn.f32 %f273, %f243, %f272;
add.f32 %f274, %f268, %f260;
sub.f32 %f275, %f260, %f274;
add.f32 %f276, %f268, %f275;
add.f32 %f277, %f273, %f276;
add.f32 %f278, %f274, %f277;
sub.f32 %f279, %f274, %f278;
add.f32 %f280, %f277, %f279;
mov.f32 %f281, 0f3F317200;
mul.rn.f32 %f282, %f257, %f281;
mov.f32 %f283, 0f35BFBE8E;
mul.rn.f32 %f284, %f257, %f283;
add.f32 %f285, %f282, %f278;
sub.f32 %f286, %f282, %f285;
add.f32 %f287, %f278, %f286;
add.f32 %f288, %f280, %f287;
add.f32 %f289, %f284, %f288;
add.f32 %f290, %f285, %f289;
sub.f32 %f291, %f285, %f290;
add.f32 %f292, %f289, %f291;
abs.f32 %f71, %f67;
setp.gt.f32 %p44, %f71, 0f77F684DF;
mul.f32 %f293, %f65, 0fB9000000;
selp.f32 %f294, %f293, %f67, %p44;
mul.rn.f32 %f295, %f294, %f290;
neg.f32 %f296, %f295;
fma.rn.f32 %f297, %f294, %f290, %f296;
fma.rn.f32 %f298, %f294, %f292, %f297;
mov.f32 %f299, 0f00000000;
fma.rn.f32 %f300, %f299, %f290, %f298;
add.rn.f32 %f301, %f295, %f300;
neg.f32 %f302, %f301;
add.rn.f32 %f303, %f295, %f302;
add.rn.f32 %f304, %f303, %f300;
mov.b32 %r222, %f301;
setp.eq.s32 %p45, %r222, 1118925336;
add.s32 %r223, %r222, -1;
mov.b32 %f305, %r223;
add.f32 %f306, %f304, 0f37000000;
selp.f32 %f307, %f305, %f301, %p45;
selp.f32 %f72, %f306, %f304, %p45;
mul.f32 %f308, %f307, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f309, %f308;
mov.f32 %f310, 0fBF317200;
fma.rn.f32 %f311, %f309, %f310, %f307;
mov.f32 %f312, 0fB5BFBE8E;
fma.rn.f32 %f313, %f309, %f312, %f311;
mul.f32 %f314, %f313, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f315, %f314;
add.f32 %f316, %f309, 0f00000000;
ex2.approx.f32 %f317, %f316;
mul.f32 %f318, %f315, %f317;
setp.lt.f32 %p46, %f307, 0fC2D20000;
selp.f32 %f319, 0f00000000, %f318, %p46;
setp.gt.f32 %p47, %f307, 0f42D20000;
selp.f32 %f446, 0f7F800000, %f319, %p47;
setp.eq.f32 %p48, %f446, 0f7F800000;
@%p48 bra BB0_61;
fma.rn.f32 %f446, %f446, %f72, %f446;
BB0_61:
setp.lt.f32 %p49, %f64, 0f00000000;
setp.eq.f32 %p50, %f69, 0f3F800000;
and.pred %p1, %p49, %p50;
mov.b32 %r224, %f446;
xor.b32 %r225, %r224, -2147483648;
mov.b32 %f320, %r225;
selp.f32 %f448, %f320, %f446, %p1;
setp.eq.f32 %p51, %f64, 0f00000000;
@%p51 bra BB0_64;
bra.uni BB0_62;
BB0_64:
add.f32 %f322, %f64, %f64;
mov.b32 %r226, %f322;
selp.b32 %r227, %r226, 0, %p50;
or.b32 %r228, %r227, 2139095040;
setp.gt.f32 %p55, %f65, 0f80000000;
selp.b32 %r229, %r228, %r227, %p55;
mov.b32 %f448, %r229;
bra.uni BB0_65;
BB0_56:
add.f32 %f239, %f64, 0fBF800000;
fma.rn.f32 %f240, %f65, %f239, 0f3F800000;
cvt.sat.f32.f32 %f449, %f240;
bra.uni BB0_73;
BB0_58:
add.f32 %f241, %f64, 0fBF800000;
fma.rn.f32 %f242, %f241, %f67, 0f3F800000;
cvt.sat.f32.f32 %f449, %f242;
bra.uni BB0_73;
BB0_93:
mov.u64 %rd99, image_HDR;
cvta.global.u64 %rd94, %rd99;
mov.u32 %r262, 8;
// inline asm
call (%rd93), _rt_buffer_get_64, (%rd94, %r89, %r262, %rd18, %rd19, %rd25, %rd25);
// inline asm
mov.f32 %f423, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs41, %f423;}
// inline asm
mov.u16 %rs42, 0;
st.v4.u16 [%rd93], {%rs41, %rs41, %rs41, %rs42};
BB0_94:
ld.global.u8 %rs43, [imageEnabled];
and.b16 %rs44, %rs43, 8;
setp.eq.s16 %p90, %rs44, 0;
@%p90 bra BB0_96;
cvt.u64.u32 %rd102, %r2;
cvt.u64.u32 %rd103, %r3;
mov.u64 %rd106, image_Mask;
cvta.global.u64 %rd101, %rd106;
// inline asm
call (%rd100), _rt_buffer_get_64, (%rd101, %r89, %r89, %rd102, %rd103, %rd25, %rd25);
// inline asm
mov.f32 %f426, 0f00000000;
cvt.rzi.u32.f32 %r265, %f426;
cvt.u16.u32 %rs45, %r265;
mov.u16 %rs46, 0;
st.v2.u8 [%rd100], {%rs45, %rs46};
bra.uni BB0_96;
BB0_62:
setp.geu.f32 %p52, %f64, 0f00000000;
@%p52 bra BB0_65;
cvt.rzi.f32.f32 %f321, %f67;
setp.neu.f32 %p53, %f321, %f67;
selp.f32 %f448, 0f7FFFFFFF, %f448, %p53;
BB0_65:
add.f32 %f323, %f70, %f71;
mov.b32 %r230, %f323;
setp.lt.s32 %p56, %r230, 2139095040;
@%p56 bra BB0_72;
setp.gtu.f32 %p57, %f70, 0f7F800000;
setp.gtu.f32 %p58, %f71, 0f7F800000;
or.pred %p59, %p57, %p58;
@%p59 bra BB0_71;
bra.uni BB0_67;
BB0_71:
sub.f32 %f448, %f64, %f65;
bra.uni BB0_72;
BB0_67:
setp.eq.f32 %p60, %f71, 0f7F800000;
@%p60 bra BB0_70;
bra.uni BB0_68;
BB0_70:
setp.gt.f32 %p63, %f70, 0f3F800000;
selp.b32 %r234, 2139095040, 0, %p63;
xor.b32 %r235, %r234, 2139095040;
setp.gt.f32 %p64, %f65, 0f80000000;
selp.b32 %r236, %r235, %r234, %p64;
mov.b32 %f324, %r236;
setp.eq.f32 %p65, %f64, 0fBF800000;
selp.f32 %f448, 0f3F800000, %f324, %p65;
bra.uni BB0_72;
BB0_68:
setp.neu.f32 %p61, %f70, 0f7F800000;
@%p61 bra BB0_72;
setp.gtu.f32 %p62, %f65, 0f80000000;
selp.b32 %r231, 0, 2139095040, %p62;
or.b32 %r232, %r231, -2147483648;
selp.b32 %r233, %r232, %r231, %p1;
mov.b32 %f448, %r233;
BB0_72:
setp.eq.f32 %p66, %f65, 0f80000000;
setp.eq.f32 %p67, %f64, 0f3F800000;
or.pred %p68, %p67, %p66;
selp.f32 %f325, 0f3F800000, %f448, %p68;
cvt.sat.f32.f32 %f449, %f325;
BB0_73:
ld.global.u8 %rs10, [imageEnabled];
and.b16 %rs11, %rs10, 4;
setp.eq.s16 %p69, %rs11, 0;
@%p69 bra BB0_77;
ld.global.u32 %r237, [additive];
setp.eq.s32 %p70, %r237, 0;
cvt.u64.u32 %rd15, %r2;
cvt.u64.u32 %rd16, %r3;
mov.f32 %f326, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs12, %f326;}
// inline asm
@%p70 bra BB0_76;
mov.u64 %rd65, image_HDR;
cvta.global.u64 %rd54, %rd65;
mov.u32 %r241, 8;
// inline asm
call (%rd53), _rt_buffer_get_64, (%rd54, %r89, %r241, %rd15, %rd16, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs19, %rs20, %rs21, %rs22}, [%rd53];
// inline asm
{ cvt.f32.f16 %f327, %rs19;}
// inline asm
// inline asm
{ cvt.f32.f16 %f328, %rs20;}
// inline asm
// inline asm
{ cvt.f32.f16 %f329, %rs21;}
// inline asm
// inline asm
call (%rd59), _rt_buffer_get_64, (%rd54, %r89, %r241, %rd15, %rd16, %rd25, %rd25);
// inline asm
add.f32 %f330, %f449, %f327;
add.f32 %f331, %f449, %f328;
add.f32 %f332, %f449, %f329;
// inline asm
{ cvt.rn.f16.f32 %rs18, %f332;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs17, %f331;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs16, %f330;}
// inline asm
st.v4.u16 [%rd59], {%rs16, %rs17, %rs18, %rs12};
bra.uni BB0_77;
BB0_76:
mov.u64 %rd72, image_HDR;
cvta.global.u64 %rd67, %rd72;
mov.u32 %r243, 8;
// inline asm
call (%rd66), _rt_buffer_get_64, (%rd67, %r89, %r243, %rd15, %rd16, %rd25, %rd25);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs23, %f449;}
// inline asm
st.v4.u16 [%rd66], {%rs23, %rs23, %rs23, %rs12};
BB0_77:
ld.global.u8 %rs24, [imageEnabled];
and.b16 %rs25, %rs24, 8;
setp.eq.s16 %p71, %rs25, 0;
@%p71 bra BB0_96;
cvt.u64.u32 %rd75, %r2;
cvt.u64.u32 %rd76, %r3;
mov.u64 %rd79, image_Mask;
cvta.global.u64 %rd74, %rd79;
// inline asm
call (%rd73), _rt_buffer_get_64, (%rd74, %r89, %r89, %rd75, %rd76, %rd25, %rd25);
// inline asm
mov.f32 %f336, 0f3E68BA2E;
cvt.rzi.f32.f32 %f337, %f336;
fma.rn.f32 %f338, %f337, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f86, %f338;
abs.f32 %f87, %f449;
setp.lt.f32 %p72, %f87, 0f00800000;
mul.f32 %f339, %f87, 0f4B800000;
selp.f32 %f340, 0fC3170000, 0fC2FE0000, %p72;
selp.f32 %f341, %f339, %f87, %p72;
mov.b32 %r246, %f341;
and.b32 %r247, %r246, 8388607;
or.b32 %r248, %r247, 1065353216;
mov.b32 %f342, %r248;
shr.u32 %r249, %r246, 23;
cvt.rn.f32.u32 %f343, %r249;
add.f32 %f344, %f340, %f343;
setp.gt.f32 %p73, %f342, 0f3FB504F3;
mul.f32 %f345, %f342, 0f3F000000;
add.f32 %f346, %f344, 0f3F800000;
selp.f32 %f347, %f345, %f342, %p73;
selp.f32 %f348, %f346, %f344, %p73;
add.f32 %f349, %f347, 0fBF800000;
add.f32 %f335, %f347, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f334,%f335;
// inline asm
add.f32 %f350, %f349, %f349;
mul.f32 %f351, %f334, %f350;
mul.f32 %f352, %f351, %f351;
mov.f32 %f353, 0f3C4CAF63;
mov.f32 %f354, 0f3B18F0FE;
fma.rn.f32 %f355, %f354, %f352, %f353;
mov.f32 %f356, 0f3DAAAABD;
fma.rn.f32 %f357, %f355, %f352, %f356;
mul.rn.f32 %f358, %f357, %f352;
mul.rn.f32 %f359, %f358, %f351;
sub.f32 %f360, %f349, %f351;
neg.f32 %f361, %f351;
add.f32 %f362, %f360, %f360;
fma.rn.f32 %f363, %f361, %f349, %f362;
mul.rn.f32 %f364, %f334, %f363;
add.f32 %f365, %f359, %f351;
sub.f32 %f366, %f351, %f365;
add.f32 %f367, %f359, %f366;
add.f32 %f368, %f364, %f367;
add.f32 %f369, %f365, %f368;
sub.f32 %f370, %f365, %f369;
add.f32 %f371, %f368, %f370;
mov.f32 %f372, 0f3F317200;
mul.rn.f32 %f373, %f348, %f372;
mov.f32 %f374, 0f35BFBE8E;
mul.rn.f32 %f375, %f348, %f374;
add.f32 %f376, %f373, %f369;
sub.f32 %f377, %f373, %f376;
add.f32 %f378, %f369, %f377;
add.f32 %f379, %f371, %f378;
add.f32 %f380, %f375, %f379;
add.f32 %f381, %f376, %f380;
sub.f32 %f382, %f376, %f381;
add.f32 %f383, %f380, %f382;
mov.f32 %f384, 0f3EE8BA2E;
mul.rn.f32 %f385, %f384, %f381;
neg.f32 %f386, %f385;
fma.rn.f32 %f387, %f384, %f381, %f386;
fma.rn.f32 %f388, %f384, %f383, %f387;
mov.f32 %f389, 0f00000000;
fma.rn.f32 %f390, %f389, %f381, %f388;
add.rn.f32 %f391, %f385, %f390;
neg.f32 %f392, %f391;
add.rn.f32 %f393, %f385, %f392;
add.rn.f32 %f394, %f393, %f390;
mov.b32 %r250, %f391;
setp.eq.s32 %p74, %r250, 1118925336;
add.s32 %r251, %r250, -1;
mov.b32 %f395, %r251;
add.f32 %f396, %f394, 0f37000000;
selp.f32 %f397, %f395, %f391, %p74;
selp.f32 %f88, %f396, %f394, %p74;
mul.f32 %f398, %f397, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f399, %f398;
mov.f32 %f400, 0fBF317200;
fma.rn.f32 %f401, %f399, %f400, %f397;
mov.f32 %f402, 0fB5BFBE8E;
fma.rn.f32 %f403, %f399, %f402, %f401;
mul.f32 %f404, %f403, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f405, %f404;
add.f32 %f406, %f399, 0f00000000;
ex2.approx.f32 %f407, %f406;
mul.f32 %f408, %f405, %f407;
setp.lt.f32 %p75, %f397, 0fC2D20000;
selp.f32 %f409, 0f00000000, %f408, %p75;
setp.gt.f32 %p76, %f397, 0f42D20000;
selp.f32 %f450, 0f7F800000, %f409, %p76;
setp.eq.f32 %p77, %f450, 0f7F800000;
@%p77 bra BB0_80;
fma.rn.f32 %f450, %f450, %f88, %f450;
BB0_80:
setp.lt.f32 %p78, %f449, 0f00000000;
setp.eq.f32 %p79, %f86, 0f3F800000;
and.pred %p2, %p78, %p79;
mov.b32 %r252, %f450;
xor.b32 %r253, %r252, -2147483648;
mov.b32 %f410, %r253;
selp.f32 %f452, %f410, %f450, %p2;
setp.eq.f32 %p80, %f449, 0f00000000;
@%p80 bra BB0_83;
bra.uni BB0_81;
BB0_83:
add.f32 %f413, %f449, %f449;
selp.f32 %f452, %f413, 0f00000000, %p79;
bra.uni BB0_84;
BB0_81:
setp.geu.f32 %p81, %f449, 0f00000000;
@%p81 bra BB0_84;
cvt.rzi.f32.f32 %f412, %f384;
setp.neu.f32 %p82, %f412, 0f3EE8BA2E;
selp.f32 %f452, 0f7FFFFFFF, %f452, %p82;
BB0_84:
add.f32 %f414, %f87, 0f3EE8BA2E;
mov.b32 %r254, %f414;
setp.lt.s32 %p84, %r254, 2139095040;
@%p84 bra BB0_89;
setp.gtu.f32 %p85, %f87, 0f7F800000;
@%p85 bra BB0_88;
bra.uni BB0_86;
BB0_88:
add.f32 %f452, %f449, 0f3EE8BA2E;
bra.uni BB0_89;
BB0_86:
setp.neu.f32 %p86, %f87, 0f7F800000;
@%p86 bra BB0_89;
selp.f32 %f452, 0fFF800000, 0f7F800000, %p2;
BB0_89:
mul.f32 %f415, %f452, 0f437F0000;
setp.eq.f32 %p87, %f449, 0f3F800000;
selp.f32 %f416, 0f437F0000, %f415, %p87;
cvt.rzi.u32.f32 %r255, %f416;
cvt.u16.u32 %rs26, %r255;
mov.u16 %rs27, 255;
st.v2.u8 [%rd73], {%rs26, %rs27};
BB0_96:
ret;
}