Files
2026-08-06 18:56:32 +03:00

1301 lines
37 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 image_Dir[1];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 1 .b8 rnd_seeds[1];
.global .align 4 .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<93>;
.reg .b16 %rs<48>;
.reg .f32 %f<500>;
.reg .b32 %r<314>;
.reg .b64 %rd<126>;
mov.u64 %rd125, __local_depot0;
cvta.local.u64 %SP, %rd125;
ld.global.v2.u32 {%r97, %r98}, [pixelID];
cvt.u64.u32 %rd22, %r97;
cvt.u64.u32 %rd23, %r98;
mov.u64 %rd26, uvnormal;
cvta.global.u64 %rd21, %rd26;
mov.u32 %r95, 2;
mov.u32 %r96, 4;
mov.u64 %rd25, 0;
// inline asm
call (%rd20), _rt_buffer_get_64, (%rd21, %r95, %r96, %rd22, %rd23, %rd25, %rd25);
// inline asm
ld.u32 %r1, [%rd20];
shr.u32 %r101, %r1, 16;
cvt.u16.u32 %rs1, %r101;
and.b16 %rs3, %rs1, 255;
cvt.u16.u32 %rs4, %r1;
or.b16 %rs5, %rs4, %rs3;
setp.eq.s16 %p3, %rs5, 0;
mov.f32 %f462, 0f00000000;
mov.f32 %f463, %f462;
mov.f32 %f464, %f462;
@%p3 bra BB0_2;
ld.u8 %rs6, [%rd20+1];
and.b16 %rs8, %rs4, 255;
cvt.rn.f32.u16 %f117, %rs8;
div.rn.f32 %f118, %f117, 0f437F0000;
fma.rn.f32 %f119, %f118, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f120, %rs6;
div.rn.f32 %f121, %f120, 0f437F0000;
fma.rn.f32 %f122, %f121, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f123, %rs3;
div.rn.f32 %f124, %f123, 0f437F0000;
fma.rn.f32 %f125, %f124, 0f40000000, 0fBF800000;
mul.f32 %f126, %f122, %f122;
fma.rn.f32 %f127, %f119, %f119, %f126;
fma.rn.f32 %f128, %f125, %f125, %f127;
sqrt.rn.f32 %f129, %f128;
rcp.rn.f32 %f130, %f129;
mul.f32 %f462, %f119, %f130;
mul.f32 %f463, %f122, %f130;
mul.f32 %f464, %f125, %f130;
BB0_2:
ld.global.v2.u32 {%r102, %r103}, [pixelID];
ld.global.v2.u32 {%r105, %r106}, [tileInfo];
add.s32 %r2, %r102, %r105;
add.s32 %r3, %r103, %r106;
setp.eq.f32 %p4, %f463, 0f00000000;
setp.eq.f32 %p5, %f462, 0f00000000;
and.pred %p6, %p5, %p4;
setp.eq.f32 %p7, %f464, 0f00000000;
and.pred %p8, %p6, %p7;
@%p8 bra BB0_93;
bra.uni BB0_3;
BB0_93:
ld.global.u8 %rs30, [imageEnabled];
and.b16 %rs31, %rs30, 4;
setp.eq.s16 %p89, %rs31, 0;
@%p89 bra BB0_97;
ld.global.u32 %r269, [additive];
setp.eq.s32 %p90, %r269, 0;
cvt.u64.u32 %rd18, %r2;
cvt.u64.u32 %rd19, %r3;
@%p90 bra BB0_96;
mov.u64 %rd99, image_HDR;
cvta.global.u64 %rd88, %rd99;
mov.u32 %r273, 8;
// inline asm
call (%rd87), _rt_buffer_get_64, (%rd88, %r95, %r273, %rd18, %rd19, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs38, %rs39, %rs40, %rs41}, [%rd87];
// inline asm
{ cvt.f32.f16 %f452, %rs38;}
// inline asm
// inline asm
{ cvt.f32.f16 %f453, %rs39;}
// inline asm
// inline asm
{ cvt.f32.f16 %f454, %rs40;}
// inline asm
// inline asm
call (%rd93), _rt_buffer_get_64, (%rd88, %r95, %r273, %rd18, %rd19, %rd25, %rd25);
// inline asm
add.f32 %f455, %f452, 0f00000000;
add.f32 %f456, %f453, 0f00000000;
add.f32 %f457, %f454, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs37, %f457;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs36, %f456;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs35, %f455;}
// inline asm
mov.u16 %rs42, 0;
st.v4.u16 [%rd93], {%rs35, %rs36, %rs37, %rs42};
bra.uni BB0_97;
BB0_3:
ld.global.v2.u32 {%r114, %r115}, [pixelID];
cvt.u64.u32 %rd29, %r114;
cvt.u64.u32 %rd30, %r115;
mov.u64 %rd39, uvpos;
cvta.global.u64 %rd28, %rd39;
mov.u32 %r111, 12;
// inline asm
call (%rd27), _rt_buffer_get_64, (%rd28, %r95, %r111, %rd29, %rd30, %rd25, %rd25);
// inline asm
ld.f32 %f135, [%rd27+8];
ld.f32 %f136, [%rd27+4];
ld.f32 %f137, [%rd27];
mul.f32 %f138, %f137, 0f3456BF95;
mul.f32 %f139, %f136, 0f3456BF95;
mul.f32 %f140, %f135, 0f3456BF95;
abs.f32 %f141, %f462;
div.rn.f32 %f142, %f138, %f141;
abs.f32 %f143, %f463;
div.rn.f32 %f144, %f139, %f143;
abs.f32 %f145, %f464;
div.rn.f32 %f146, %f140, %f145;
abs.f32 %f147, %f142;
abs.f32 %f148, %f144;
abs.f32 %f149, %f146;
mov.f32 %f150, 0f38D1B717;
max.f32 %f151, %f147, %f150;
max.f32 %f152, %f148, %f150;
max.f32 %f153, %f149, %f150;
fma.rn.f32 %f7, %f462, %f151, %f137;
fma.rn.f32 %f8, %f463, %f152, %f136;
fma.rn.f32 %f9, %f464, %f153, %f135;
setp.gt.f32 %p9, %f141, %f145;
neg.f32 %f154, %f463;
selp.f32 %f155, %f154, 0f00000000, %p9;
neg.f32 %f156, %f464;
selp.f32 %f157, %f462, %f156, %p9;
selp.f32 %f158, 0f00000000, %f463, %p9;
mul.f32 %f159, %f157, %f157;
fma.rn.f32 %f160, %f155, %f155, %f159;
fma.rn.f32 %f161, %f158, %f158, %f160;
sqrt.rn.f32 %f162, %f161;
rcp.rn.f32 %f163, %f162;
mul.f32 %f10, %f155, %f163;
mul.f32 %f11, %f157, %f163;
mul.f32 %f12, %f158, %f163;
ld.global.v2.u32 {%r118, %r119}, [pixelID];
cvt.u64.u32 %rd35, %r118;
cvt.u64.u32 %rd36, %r119;
mov.u64 %rd40, rnd_seeds;
cvta.global.u64 %rd34, %rd40;
// inline asm
call (%rd33), _rt_buffer_get_64, (%rd34, %r95, %r96, %rd35, %rd36, %rd25, %rd25);
// inline asm
ld.global.u32 %r284, [samples];
mov.f32 %f485, 0f00000000;
setp.lt.s32 %p10, %r284, 1;
@%p10 bra BB0_4;
cvt.rn.f32.s32 %f168, %r284;
rcp.rn.f32 %f13, %f168;
ld.u32 %r310, [%rd33];
mul.f32 %f14, %f7, 0f3456BF95;
mul.f32 %f15, %f8, 0f3456BF95;
mul.f32 %f16, %f9, 0f3456BF95;
mul.f32 %f169, %f462, %f11;
mul.f32 %f170, %f463, %f10;
sub.f32 %f17, %f170, %f169;
mul.f32 %f171, %f464, %f10;
mul.f32 %f172, %f462, %f12;
sub.f32 %f18, %f172, %f171;
mul.f32 %f173, %f463, %f12;
mul.f32 %f174, %f464, %f11;
sub.f32 %f19, %f174, %f173;
mov.f32 %f485, 0f00000000;
mov.u32 %r285, 0;
abs.f32 %f175, %f15;
abs.f32 %f176, %f14;
max.f32 %f177, %f176, %f175;
abs.f32 %f178, %f16;
max.f32 %f179, %f177, %f178;
mov.f32 %f486, %f485;
mov.f32 %f487, %f485;
mov.f32 %f488, %f485;
BB0_6:
setp.lt.s32 %p11, %r284, 1;
@%p11 bra BB0_55;
cvt.rn.f32.s32 %f24, %r285;
max.f32 %f25, %f179, %f150;
mov.u32 %r287, 0;
BB0_8:
mad.lo.s32 %r124, %r310, 1664525, 1013904223;
and.b32 %r125, %r124, 16777215;
cvt.rn.f32.u32 %f181, %r125;
fma.rn.f32 %f182, %f181, 0f33800000, %f24;
mul.f32 %f183, %f13, %f182;
mad.lo.s32 %r310, %r124, 1664525, 1013904223;
and.b32 %r126, %r310, 16777215;
cvt.rn.f32.u32 %f184, %r126;
cvt.rn.f32.s32 %f185, %r287;
fma.rn.f32 %f186, %f184, 0f33800000, %f185;
mul.f32 %f187, %f13, %f186;
sqrt.rn.f32 %f30, %f183;
mul.f32 %f479, %f187, 0f40C90FDB;
abs.f32 %f32, %f479;
setp.neu.f32 %p12, %f32, 0f7F800000;
mov.f32 %f473, %f479;
@%p12 bra BB0_10;
mov.f32 %f188, 0f00000000;
mul.rn.f32 %f473, %f479, %f188;
BB0_10:
mul.f32 %f189, %f473, 0f3F22F983;
cvt.rni.s32.f32 %r298, %f189;
cvt.rn.f32.s32 %f190, %r298;
neg.f32 %f191, %f190;
mov.f32 %f192, 0f3FC90FDA;
fma.rn.f32 %f193, %f191, %f192, %f473;
mov.f32 %f194, 0f33A22168;
fma.rn.f32 %f195, %f191, %f194, %f193;
mov.f32 %f196, 0f27C234C5;
fma.rn.f32 %f474, %f191, %f196, %f195;
abs.f32 %f197, %f473;
setp.leu.f32 %p13, %f197, 0f47CE4780;
@%p13 bra BB0_21;
mov.b32 %r13, %f473;
shr.u32 %r14, %r13, 23;
shl.b32 %r129, %r13, 8;
or.b32 %r15, %r129, -2147483648;
add.u64 %rd42, %SP, 4;
cvta.to.local.u64 %rd122, %rd42;
mov.u32 %r290, 0;
mov.u64 %rd121, __cudart_i2opi_f;
mov.u32 %r289, -6;
BB0_12:
.pragma "nounroll";
ld.const.u32 %r132, [%rd121];
// inline asm
{
mad.lo.cc.u32 %r130, %r132, %r15, %r290;
madc.hi.u32 %r290, %r132, %r15, 0;
}
// inline asm
st.local.u32 [%rd122], %r130;
add.s64 %rd122, %rd122, 4;
add.s64 %rd121, %rd121, 4;
add.s32 %r289, %r289, 1;
setp.ne.s32 %p14, %r289, 0;
@%p14 bra BB0_12;
and.b32 %r135, %r14, 255;
add.s32 %r136, %r135, -128;
shr.u32 %r137, %r136, 5;
and.b32 %r20, %r13, -2147483648;
cvta.to.local.u64 %rd44, %rd42;
st.local.u32 [%rd44+24], %r290;
mov.u32 %r138, 6;
sub.s32 %r139, %r138, %r137;
mul.wide.s32 %rd45, %r139, 4;
add.s64 %rd8, %rd44, %rd45;
ld.local.u32 %r291, [%rd8];
ld.local.u32 %r292, [%rd8+-4];
and.b32 %r23, %r14, 31;
setp.eq.s32 %p15, %r23, 0;
@%p15 bra BB0_15;
mov.u32 %r140, 32;
sub.s32 %r141, %r140, %r23;
shr.u32 %r142, %r292, %r141;
shl.b32 %r143, %r291, %r23;
add.s32 %r291, %r142, %r143;
ld.local.u32 %r144, [%rd8+-8];
shr.u32 %r145, %r144, %r141;
shl.b32 %r146, %r292, %r23;
add.s32 %r292, %r145, %r146;
BB0_15:
shr.u32 %r147, %r292, 30;
shl.b32 %r148, %r291, 2;
add.s32 %r293, %r147, %r148;
shl.b32 %r29, %r292, 2;
shr.u32 %r149, %r293, 31;
shr.u32 %r150, %r291, 30;
add.s32 %r30, %r149, %r150;
setp.eq.s32 %p16, %r149, 0;
@%p16 bra BB0_16;
bra.uni BB0_17;
BB0_16:
mov.u32 %r294, %r20;
mov.u32 %r295, %r29;
bra.uni BB0_18;
BB0_17:
not.b32 %r151, %r293;
neg.s32 %r295, %r29;
setp.eq.s32 %p17, %r29, 0;
selp.u32 %r152, 1, 0, %p17;
add.s32 %r293, %r152, %r151;
xor.b32 %r294, %r20, -2147483648;
BB0_18:
clz.b32 %r297, %r293;
setp.eq.s32 %p18, %r297, 0;
shl.b32 %r153, %r293, %r297;
mov.u32 %r154, 32;
sub.s32 %r155, %r154, %r297;
shr.u32 %r156, %r295, %r155;
add.s32 %r157, %r156, %r153;
selp.b32 %r38, %r293, %r157, %p18;
mov.u32 %r158, -921707870;
mul.hi.u32 %r296, %r38, %r158;
setp.eq.s32 %p19, %r20, 0;
neg.s32 %r159, %r30;
selp.b32 %r298, %r30, %r159, %p19;
setp.lt.s32 %p20, %r296, 1;
@%p20 bra BB0_20;
mul.lo.s32 %r160, %r38, -921707870;
shr.u32 %r161, %r160, 31;
shl.b32 %r162, %r296, 1;
add.s32 %r296, %r161, %r162;
add.s32 %r297, %r297, 1;
BB0_20:
mov.u32 %r163, 126;
sub.s32 %r164, %r163, %r297;
shl.b32 %r165, %r164, 23;
add.s32 %r166, %r296, 1;
shr.u32 %r167, %r166, 7;
add.s32 %r168, %r167, 1;
shr.u32 %r169, %r168, 1;
add.s32 %r170, %r169, %r165;
or.b32 %r171, %r170, %r294;
mov.b32 %f474, %r171;
BB0_21:
mul.rn.f32 %f38, %f474, %f474;
add.s32 %r46, %r298, 1;
and.b32 %r47, %r46, 1;
setp.eq.s32 %p21, %r47, 0;
@%p21 bra BB0_23;
bra.uni BB0_22;
BB0_23:
mov.f32 %f200, 0f3C08839E;
mov.f32 %f201, 0fB94CA1F9;
fma.rn.f32 %f475, %f201, %f38, %f200;
bra.uni BB0_24;
BB0_22:
mov.f32 %f198, 0fBAB6061A;
mov.f32 %f199, 0f37CCF5CE;
fma.rn.f32 %f475, %f199, %f38, %f198;
BB0_24:
@%p21 bra BB0_26;
bra.uni BB0_25;
BB0_26:
mov.f32 %f205, 0fBE2AAAA3;
fma.rn.f32 %f206, %f475, %f38, %f205;
mov.f32 %f207, 0f00000000;
fma.rn.f32 %f476, %f206, %f38, %f207;
bra.uni BB0_27;
BB0_25:
mov.f32 %f202, 0f3D2AAAA5;
fma.rn.f32 %f203, %f475, %f38, %f202;
mov.f32 %f204, 0fBF000000;
fma.rn.f32 %f476, %f203, %f38, %f204;
BB0_27:
fma.rn.f32 %f477, %f476, %f474, %f474;
@%p21 bra BB0_29;
mov.f32 %f208, 0f3F800000;
fma.rn.f32 %f477, %f476, %f38, %f208;
BB0_29:
and.b32 %r172, %r46, 2;
setp.eq.s32 %p24, %r172, 0;
@%p24 bra BB0_31;
mov.f32 %f209, 0f00000000;
mov.f32 %f210, 0fBF800000;
fma.rn.f32 %f477, %f477, %f210, %f209;
BB0_31:
@%p12 bra BB0_33;
mov.f32 %f211, 0f00000000;
mul.rn.f32 %f479, %f479, %f211;
BB0_33:
mul.f32 %f212, %f479, 0f3F22F983;
cvt.rni.s32.f32 %r308, %f212;
cvt.rn.f32.s32 %f213, %r308;
neg.f32 %f214, %f213;
fma.rn.f32 %f216, %f214, %f192, %f479;
fma.rn.f32 %f218, %f214, %f194, %f216;
fma.rn.f32 %f480, %f214, %f196, %f218;
abs.f32 %f220, %f479;
setp.leu.f32 %p26, %f220, 0f47CE4780;
@%p26 bra BB0_44;
mov.b32 %r49, %f479;
shr.u32 %r50, %r49, 23;
shl.b32 %r175, %r49, 8;
or.b32 %r51, %r175, -2147483648;
add.u64 %rd47, %SP, 4;
cvta.to.local.u64 %rd124, %rd47;
mov.u32 %r300, 0;
mov.u64 %rd123, __cudart_i2opi_f;
mov.u32 %r299, -6;
BB0_35:
.pragma "nounroll";
ld.const.u32 %r178, [%rd123];
// inline asm
{
mad.lo.cc.u32 %r176, %r178, %r51, %r300;
madc.hi.u32 %r300, %r178, %r51, 0;
}
// inline asm
st.local.u32 [%rd124], %r176;
add.s64 %rd124, %rd124, 4;
add.s64 %rd123, %rd123, 4;
add.s32 %r299, %r299, 1;
setp.ne.s32 %p27, %r299, 0;
@%p27 bra BB0_35;
and.b32 %r181, %r50, 255;
add.s32 %r182, %r181, -128;
shr.u32 %r183, %r182, 5;
and.b32 %r56, %r49, -2147483648;
cvta.to.local.u64 %rd49, %rd47;
st.local.u32 [%rd49+24], %r300;
mov.u32 %r184, 6;
sub.s32 %r185, %r184, %r183;
mul.wide.s32 %rd50, %r185, 4;
add.s64 %rd14, %rd49, %rd50;
ld.local.u32 %r301, [%rd14];
ld.local.u32 %r302, [%rd14+-4];
and.b32 %r59, %r50, 31;
setp.eq.s32 %p28, %r59, 0;
@%p28 bra BB0_38;
mov.u32 %r186, 32;
sub.s32 %r187, %r186, %r59;
shr.u32 %r188, %r302, %r187;
shl.b32 %r189, %r301, %r59;
add.s32 %r301, %r188, %r189;
ld.local.u32 %r190, [%rd14+-8];
shr.u32 %r191, %r190, %r187;
shl.b32 %r192, %r302, %r59;
add.s32 %r302, %r191, %r192;
BB0_38:
shr.u32 %r193, %r302, 30;
shl.b32 %r194, %r301, 2;
add.s32 %r303, %r193, %r194;
shl.b32 %r65, %r302, 2;
shr.u32 %r195, %r303, 31;
shr.u32 %r196, %r301, 30;
add.s32 %r66, %r195, %r196;
setp.eq.s32 %p29, %r195, 0;
@%p29 bra BB0_39;
bra.uni BB0_40;
BB0_39:
mov.u32 %r304, %r56;
mov.u32 %r305, %r65;
bra.uni BB0_41;
BB0_40:
not.b32 %r197, %r303;
neg.s32 %r305, %r65;
setp.eq.s32 %p30, %r65, 0;
selp.u32 %r198, 1, 0, %p30;
add.s32 %r303, %r198, %r197;
xor.b32 %r304, %r56, -2147483648;
BB0_41:
clz.b32 %r307, %r303;
setp.eq.s32 %p31, %r307, 0;
shl.b32 %r199, %r303, %r307;
mov.u32 %r200, 32;
sub.s32 %r201, %r200, %r307;
shr.u32 %r202, %r305, %r201;
add.s32 %r203, %r202, %r199;
selp.b32 %r74, %r303, %r203, %p31;
mov.u32 %r204, -921707870;
mul.hi.u32 %r306, %r74, %r204;
setp.eq.s32 %p32, %r56, 0;
neg.s32 %r205, %r66;
selp.b32 %r308, %r66, %r205, %p32;
setp.lt.s32 %p33, %r306, 1;
@%p33 bra BB0_43;
mul.lo.s32 %r206, %r74, -921707870;
shr.u32 %r207, %r206, 31;
shl.b32 %r208, %r306, 1;
add.s32 %r306, %r207, %r208;
add.s32 %r307, %r307, 1;
BB0_43:
mov.u32 %r209, 126;
sub.s32 %r210, %r209, %r307;
shl.b32 %r211, %r210, 23;
add.s32 %r212, %r306, 1;
shr.u32 %r213, %r212, 7;
add.s32 %r214, %r213, 1;
shr.u32 %r215, %r214, 1;
add.s32 %r216, %r215, %r211;
or.b32 %r217, %r216, %r304;
mov.b32 %f480, %r217;
BB0_44:
mul.rn.f32 %f55, %f480, %f480;
and.b32 %r82, %r308, 1;
setp.eq.s32 %p34, %r82, 0;
@%p34 bra BB0_46;
bra.uni BB0_45;
BB0_46:
mov.f32 %f223, 0f3C08839E;
mov.f32 %f224, 0fB94CA1F9;
fma.rn.f32 %f481, %f224, %f55, %f223;
bra.uni BB0_47;
BB0_45:
mov.f32 %f221, 0fBAB6061A;
mov.f32 %f222, 0f37CCF5CE;
fma.rn.f32 %f481, %f222, %f55, %f221;
BB0_47:
@%p34 bra BB0_49;
bra.uni BB0_48;
BB0_49:
mov.f32 %f228, 0fBE2AAAA3;
fma.rn.f32 %f229, %f481, %f55, %f228;
mov.f32 %f230, 0f00000000;
fma.rn.f32 %f482, %f229, %f55, %f230;
bra.uni BB0_50;
BB0_48:
mov.f32 %f225, 0f3D2AAAA5;
fma.rn.f32 %f226, %f481, %f55, %f225;
mov.f32 %f227, 0fBF000000;
fma.rn.f32 %f482, %f226, %f55, %f227;
BB0_50:
fma.rn.f32 %f483, %f482, %f480, %f480;
@%p34 bra BB0_52;
mov.f32 %f231, 0f3F800000;
fma.rn.f32 %f483, %f482, %f55, %f231;
BB0_52:
and.b32 %r218, %r308, 2;
setp.eq.s32 %p37, %r218, 0;
@%p37 bra BB0_54;
mov.f32 %f232, 0f00000000;
mov.f32 %f233, 0fBF800000;
fma.rn.f32 %f483, %f483, %f233, %f232;
BB0_54:
mul.f32 %f242, %f30, %f477;
add.u64 %rd51, %SP, 0;
cvta.to.local.u64 %rd52, %rd51;
mul.f32 %f243, %f242, %f242;
mov.f32 %f244, 0f3F800000;
sub.f32 %f245, %f244, %f243;
mul.f32 %f246, %f30, %f483;
mul.f32 %f247, %f246, %f246;
sub.f32 %f248, %f245, %f247;
mov.f32 %f249, 0f00000000;
max.f32 %f250, %f249, %f248;
sqrt.rn.f32 %f251, %f250;
mul.f32 %f252, %f10, %f246;
mul.f32 %f253, %f11, %f246;
mul.f32 %f254, %f12, %f246;
fma.rn.f32 %f255, %f19, %f242, %f252;
fma.rn.f32 %f256, %f18, %f242, %f253;
fma.rn.f32 %f257, %f17, %f242, %f254;
fma.rn.f32 %f237, %f462, %f251, %f255;
fma.rn.f32 %f238, %f463, %f251, %f256;
fma.rn.f32 %f239, %f464, %f251, %f257;
ld.global.f32 %f241, [lightRadius];
mov.u32 %r222, 1065353216;
st.local.u32 [%rd52], %r222;
ld.global.u32 %r219, [root];
mov.u32 %r220, 1;
// inline asm
call _rt_trace_64, (%r219, %f7, %f8, %f9, %f237, %f238, %f239, %r220, %f25, %f241, %rd51, %r96);
// inline asm
ld.local.f32 %f258, [%rd52];
add.f32 %f485, %f485, %f258;
fma.rn.f32 %f488, %f237, %f258, %f488;
fma.rn.f32 %f487, %f238, %f258, %f487;
fma.rn.f32 %f486, %f239, %f258, %f486;
ld.global.u32 %r284, [samples];
add.s32 %r287, %r287, 1;
setp.lt.s32 %p38, %r287, %r284;
@%p38 bra BB0_8;
BB0_55:
add.s32 %r285, %r285, 1;
setp.lt.s32 %p39, %r285, %r284;
@%p39 bra BB0_6;
bra.uni BB0_56;
BB0_4:
mov.f32 %f486, %f485;
mov.f32 %f487, %f485;
mov.f32 %f488, %f485;
BB0_56:
mul.lo.s32 %r223, %r284, %r284;
cvt.rn.f32.s32 %f259, %r223;
div.rn.f32 %f79, %f485, %f259;
ld.global.f32 %f80, [intensity];
setp.lt.f32 %p40, %f80, 0f00000000;
@%p40 bra BB0_58;
bra.uni BB0_57;
BB0_58:
neg.f32 %f82, %f80;
setp.lt.f32 %p41, %f80, 0fBF800000;
@%p41 bra BB0_60;
bra.uni BB0_59;
BB0_60:
mul.f32 %f266, %f80, 0fBF000000;
cvt.rzi.f32.f32 %f267, %f266;
fma.rn.f32 %f268, %f267, 0fC0000000, %f82;
abs.f32 %f84, %f268;
abs.f32 %f85, %f79;
setp.lt.f32 %p42, %f85, 0f00800000;
mul.f32 %f269, %f85, 0f4B800000;
selp.f32 %f270, 0fC3170000, 0fC2FE0000, %p42;
selp.f32 %f271, %f269, %f85, %p42;
mov.b32 %r224, %f271;
and.b32 %r225, %r224, 8388607;
or.b32 %r226, %r225, 1065353216;
mov.b32 %f272, %r226;
shr.u32 %r227, %r224, 23;
cvt.rn.f32.u32 %f273, %r227;
add.f32 %f274, %f270, %f273;
setp.gt.f32 %p43, %f272, 0f3FB504F3;
mul.f32 %f275, %f272, 0f3F000000;
add.f32 %f276, %f274, 0f3F800000;
selp.f32 %f277, %f275, %f272, %p43;
selp.f32 %f278, %f276, %f274, %p43;
add.f32 %f279, %f277, 0fBF800000;
add.f32 %f265, %f277, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f264,%f265;
// inline asm
add.f32 %f280, %f279, %f279;
mul.f32 %f281, %f264, %f280;
mul.f32 %f282, %f281, %f281;
mov.f32 %f283, 0f3C4CAF63;
mov.f32 %f284, 0f3B18F0FE;
fma.rn.f32 %f285, %f284, %f282, %f283;
mov.f32 %f286, 0f3DAAAABD;
fma.rn.f32 %f287, %f285, %f282, %f286;
mul.rn.f32 %f288, %f287, %f282;
mul.rn.f32 %f289, %f288, %f281;
sub.f32 %f290, %f279, %f281;
neg.f32 %f291, %f281;
add.f32 %f292, %f290, %f290;
fma.rn.f32 %f293, %f291, %f279, %f292;
mul.rn.f32 %f294, %f264, %f293;
add.f32 %f295, %f289, %f281;
sub.f32 %f296, %f281, %f295;
add.f32 %f297, %f289, %f296;
add.f32 %f298, %f294, %f297;
add.f32 %f299, %f295, %f298;
sub.f32 %f300, %f295, %f299;
add.f32 %f301, %f298, %f300;
mov.f32 %f302, 0f3F317200;
mul.rn.f32 %f303, %f278, %f302;
mov.f32 %f304, 0f35BFBE8E;
mul.rn.f32 %f305, %f278, %f304;
add.f32 %f306, %f303, %f299;
sub.f32 %f307, %f303, %f306;
add.f32 %f308, %f299, %f307;
add.f32 %f309, %f301, %f308;
add.f32 %f310, %f305, %f309;
add.f32 %f311, %f306, %f310;
sub.f32 %f312, %f306, %f311;
add.f32 %f313, %f310, %f312;
abs.f32 %f86, %f82;
setp.gt.f32 %p44, %f86, 0f77F684DF;
mul.f32 %f314, %f80, 0fB9000000;
selp.f32 %f315, %f314, %f82, %p44;
mul.rn.f32 %f316, %f315, %f311;
neg.f32 %f317, %f316;
fma.rn.f32 %f318, %f315, %f311, %f317;
fma.rn.f32 %f319, %f315, %f313, %f318;
mov.f32 %f320, 0f00000000;
fma.rn.f32 %f321, %f320, %f311, %f319;
add.rn.f32 %f322, %f316, %f321;
neg.f32 %f323, %f322;
add.rn.f32 %f324, %f316, %f323;
add.rn.f32 %f325, %f324, %f321;
mov.b32 %r228, %f322;
setp.eq.s32 %p45, %r228, 1118925336;
add.s32 %r229, %r228, -1;
mov.b32 %f326, %r229;
add.f32 %f327, %f325, 0f37000000;
selp.f32 %f328, %f326, %f322, %p45;
selp.f32 %f87, %f327, %f325, %p45;
mul.f32 %f329, %f328, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f330, %f329;
mov.f32 %f331, 0fBF317200;
fma.rn.f32 %f332, %f330, %f331, %f328;
mov.f32 %f333, 0fB5BFBE8E;
fma.rn.f32 %f334, %f330, %f333, %f332;
mul.f32 %f335, %f334, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f336, %f335;
add.f32 %f337, %f330, 0f00000000;
ex2.approx.f32 %f338, %f337;
mul.f32 %f339, %f336, %f338;
setp.lt.f32 %p46, %f328, 0fC2D20000;
selp.f32 %f340, 0f00000000, %f339, %p46;
setp.gt.f32 %p47, %f328, 0f42D20000;
selp.f32 %f493, 0f7F800000, %f340, %p47;
setp.eq.f32 %p48, %f493, 0f7F800000;
@%p48 bra BB0_62;
fma.rn.f32 %f493, %f493, %f87, %f493;
BB0_62:
setp.lt.f32 %p49, %f79, 0f00000000;
setp.eq.f32 %p50, %f84, 0f3F800000;
and.pred %p1, %p49, %p50;
mov.b32 %r230, %f493;
xor.b32 %r231, %r230, -2147483648;
mov.b32 %f341, %r231;
selp.f32 %f495, %f341, %f493, %p1;
setp.eq.f32 %p51, %f79, 0f00000000;
@%p51 bra BB0_65;
bra.uni BB0_63;
BB0_65:
add.f32 %f343, %f79, %f79;
mov.b32 %r232, %f343;
selp.b32 %r233, %r232, 0, %p50;
or.b32 %r234, %r233, 2139095040;
setp.gt.f32 %p55, %f80, 0f80000000;
selp.b32 %r235, %r234, %r233, %p55;
mov.b32 %f495, %r235;
bra.uni BB0_66;
BB0_57:
add.f32 %f260, %f79, 0fBF800000;
fma.rn.f32 %f261, %f80, %f260, 0f3F800000;
cvt.sat.f32.f32 %f496, %f261;
bra.uni BB0_74;
BB0_59:
add.f32 %f262, %f79, 0fBF800000;
fma.rn.f32 %f263, %f262, %f82, 0f3F800000;
cvt.sat.f32.f32 %f496, %f263;
bra.uni BB0_74;
BB0_96:
mov.u64 %rd106, image_HDR;
cvta.global.u64 %rd101, %rd106;
mov.u32 %r275, 8;
// inline asm
call (%rd100), _rt_buffer_get_64, (%rd101, %r95, %r275, %rd18, %rd19, %rd25, %rd25);
// inline asm
mov.f32 %f458, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs43, %f458;}
// inline asm
mov.u16 %rs44, 0;
st.v4.u16 [%rd100], {%rs43, %rs43, %rs43, %rs44};
BB0_97:
ld.global.u32 %r313, [imageEnabled];
and.b32 %r276, %r313, 8;
setp.eq.s32 %p91, %r276, 0;
@%p91 bra BB0_99;
cvt.u64.u32 %rd109, %r2;
cvt.u64.u32 %rd110, %r3;
mov.u64 %rd113, image_Mask;
cvta.global.u64 %rd108, %rd113;
// inline asm
call (%rd107), _rt_buffer_get_64, (%rd108, %r95, %r95, %rd109, %rd110, %rd25, %rd25);
// inline asm
mov.f32 %f461, 0f00000000;
cvt.rzi.u32.f32 %r279, %f461;
cvt.u16.u32 %rs45, %r279;
mov.u16 %rs46, 0;
st.v2.u8 [%rd107], {%rs45, %rs46};
ld.global.u32 %r313, [imageEnabled];
BB0_99:
and.b32 %r280, %r313, 64;
setp.eq.s32 %p92, %r280, 0;
@%p92 bra BB0_101;
cvt.u64.u32 %rd116, %r2;
cvt.u64.u32 %rd117, %r3;
mov.u64 %rd120, image_Dir;
cvta.global.u64 %rd115, %rd120;
// inline asm
call (%rd114), _rt_buffer_get_64, (%rd115, %r95, %r96, %rd116, %rd117, %rd25, %rd25);
// inline asm
mov.u16 %rs47, 0;
st.v4.u8 [%rd114], {%rs47, %rs47, %rs47, %rs47};
bra.uni BB0_101;
BB0_63:
setp.geu.f32 %p52, %f79, 0f00000000;
@%p52 bra BB0_66;
cvt.rzi.f32.f32 %f342, %f82;
setp.neu.f32 %p53, %f342, %f82;
selp.f32 %f495, 0f7FFFFFFF, %f495, %p53;
BB0_66:
add.f32 %f344, %f85, %f86;
mov.b32 %r236, %f344;
setp.lt.s32 %p56, %r236, 2139095040;
@%p56 bra BB0_73;
setp.gtu.f32 %p57, %f85, 0f7F800000;
setp.gtu.f32 %p58, %f86, 0f7F800000;
or.pred %p59, %p57, %p58;
@%p59 bra BB0_72;
bra.uni BB0_68;
BB0_72:
sub.f32 %f495, %f79, %f80;
bra.uni BB0_73;
BB0_68:
setp.eq.f32 %p60, %f86, 0f7F800000;
@%p60 bra BB0_71;
bra.uni BB0_69;
BB0_71:
setp.gt.f32 %p63, %f85, 0f3F800000;
selp.b32 %r240, 2139095040, 0, %p63;
xor.b32 %r241, %r240, 2139095040;
setp.gt.f32 %p64, %f80, 0f80000000;
selp.b32 %r242, %r241, %r240, %p64;
mov.b32 %f345, %r242;
setp.eq.f32 %p65, %f79, 0fBF800000;
selp.f32 %f495, 0f3F800000, %f345, %p65;
bra.uni BB0_73;
BB0_69:
setp.neu.f32 %p61, %f85, 0f7F800000;
@%p61 bra BB0_73;
setp.gtu.f32 %p62, %f80, 0f80000000;
selp.b32 %r237, 0, 2139095040, %p62;
or.b32 %r238, %r237, -2147483648;
selp.b32 %r239, %r238, %r237, %p1;
mov.b32 %f495, %r239;
BB0_73:
setp.eq.f32 %p66, %f80, 0f80000000;
setp.eq.f32 %p67, %f79, 0f3F800000;
or.pred %p68, %p67, %p66;
selp.f32 %f346, 0f3F800000, %f495, %p68;
cvt.sat.f32.f32 %f496, %f346;
BB0_74:
ld.global.u8 %rs10, [imageEnabled];
and.b16 %rs11, %rs10, 4;
setp.eq.s16 %p69, %rs11, 0;
@%p69 bra BB0_78;
ld.global.u32 %r243, [additive];
setp.eq.s32 %p70, %r243, 0;
cvt.u64.u32 %rd15, %r2;
cvt.u64.u32 %rd16, %r3;
mov.f32 %f347, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs12, %f347;}
// inline asm
@%p70 bra BB0_77;
mov.u64 %rd65, image_HDR;
cvta.global.u64 %rd54, %rd65;
mov.u32 %r247, 8;
// inline asm
call (%rd53), _rt_buffer_get_64, (%rd54, %r95, %r247, %rd15, %rd16, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs19, %rs20, %rs21, %rs22}, [%rd53];
// inline asm
{ cvt.f32.f16 %f348, %rs19;}
// inline asm
// inline asm
{ cvt.f32.f16 %f349, %rs20;}
// inline asm
// inline asm
{ cvt.f32.f16 %f350, %rs21;}
// inline asm
// inline asm
call (%rd59), _rt_buffer_get_64, (%rd54, %r95, %r247, %rd15, %rd16, %rd25, %rd25);
// inline asm
add.f32 %f351, %f496, %f348;
add.f32 %f352, %f496, %f349;
add.f32 %f353, %f496, %f350;
// inline asm
{ cvt.rn.f16.f32 %rs18, %f353;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs17, %f352;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs16, %f351;}
// inline asm
st.v4.u16 [%rd59], {%rs16, %rs17, %rs18, %rs12};
bra.uni BB0_78;
BB0_77:
mov.u64 %rd72, image_HDR;
cvta.global.u64 %rd67, %rd72;
mov.u32 %r249, 8;
// inline asm
call (%rd66), _rt_buffer_get_64, (%rd67, %r95, %r249, %rd15, %rd16, %rd25, %rd25);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs23, %f496;}
// inline asm
st.v4.u16 [%rd66], {%rs23, %rs23, %rs23, %rs12};
BB0_78:
ld.global.u32 %r312, [imageEnabled];
and.b32 %r250, %r312, 8;
setp.eq.s32 %p71, %r250, 0;
@%p71 bra BB0_91;
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, %r95, %r95, %rd75, %rd76, %rd25, %rd25);
// inline asm
mov.f32 %f357, 0f3E68BA2E;
cvt.rzi.f32.f32 %f358, %f357;
fma.rn.f32 %f359, %f358, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f101, %f359;
abs.f32 %f102, %f496;
setp.lt.f32 %p72, %f102, 0f00800000;
mul.f32 %f360, %f102, 0f4B800000;
selp.f32 %f361, 0fC3170000, 0fC2FE0000, %p72;
selp.f32 %f362, %f360, %f102, %p72;
mov.b32 %r253, %f362;
and.b32 %r254, %r253, 8388607;
or.b32 %r255, %r254, 1065353216;
mov.b32 %f363, %r255;
shr.u32 %r256, %r253, 23;
cvt.rn.f32.u32 %f364, %r256;
add.f32 %f365, %f361, %f364;
setp.gt.f32 %p73, %f363, 0f3FB504F3;
mul.f32 %f366, %f363, 0f3F000000;
add.f32 %f367, %f365, 0f3F800000;
selp.f32 %f368, %f366, %f363, %p73;
selp.f32 %f369, %f367, %f365, %p73;
add.f32 %f370, %f368, 0fBF800000;
add.f32 %f356, %f368, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f355,%f356;
// inline asm
add.f32 %f371, %f370, %f370;
mul.f32 %f372, %f355, %f371;
mul.f32 %f373, %f372, %f372;
mov.f32 %f374, 0f3C4CAF63;
mov.f32 %f375, 0f3B18F0FE;
fma.rn.f32 %f376, %f375, %f373, %f374;
mov.f32 %f377, 0f3DAAAABD;
fma.rn.f32 %f378, %f376, %f373, %f377;
mul.rn.f32 %f379, %f378, %f373;
mul.rn.f32 %f380, %f379, %f372;
sub.f32 %f381, %f370, %f372;
neg.f32 %f382, %f372;
add.f32 %f383, %f381, %f381;
fma.rn.f32 %f384, %f382, %f370, %f383;
mul.rn.f32 %f385, %f355, %f384;
add.f32 %f386, %f380, %f372;
sub.f32 %f387, %f372, %f386;
add.f32 %f388, %f380, %f387;
add.f32 %f389, %f385, %f388;
add.f32 %f390, %f386, %f389;
sub.f32 %f391, %f386, %f390;
add.f32 %f392, %f389, %f391;
mov.f32 %f393, 0f3F317200;
mul.rn.f32 %f394, %f369, %f393;
mov.f32 %f395, 0f35BFBE8E;
mul.rn.f32 %f396, %f369, %f395;
add.f32 %f397, %f394, %f390;
sub.f32 %f398, %f394, %f397;
add.f32 %f399, %f390, %f398;
add.f32 %f400, %f392, %f399;
add.f32 %f401, %f396, %f400;
add.f32 %f402, %f397, %f401;
sub.f32 %f403, %f397, %f402;
add.f32 %f404, %f401, %f403;
mov.f32 %f405, 0f3EE8BA2E;
mul.rn.f32 %f406, %f405, %f402;
neg.f32 %f407, %f406;
fma.rn.f32 %f408, %f405, %f402, %f407;
fma.rn.f32 %f409, %f405, %f404, %f408;
mov.f32 %f410, 0f00000000;
fma.rn.f32 %f411, %f410, %f402, %f409;
add.rn.f32 %f412, %f406, %f411;
neg.f32 %f413, %f412;
add.rn.f32 %f414, %f406, %f413;
add.rn.f32 %f415, %f414, %f411;
mov.b32 %r257, %f412;
setp.eq.s32 %p74, %r257, 1118925336;
add.s32 %r258, %r257, -1;
mov.b32 %f416, %r258;
add.f32 %f417, %f415, 0f37000000;
selp.f32 %f418, %f416, %f412, %p74;
selp.f32 %f103, %f417, %f415, %p74;
mul.f32 %f419, %f418, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f420, %f419;
mov.f32 %f421, 0fBF317200;
fma.rn.f32 %f422, %f420, %f421, %f418;
mov.f32 %f423, 0fB5BFBE8E;
fma.rn.f32 %f424, %f420, %f423, %f422;
mul.f32 %f425, %f424, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f426, %f425;
add.f32 %f427, %f420, 0f00000000;
ex2.approx.f32 %f428, %f427;
mul.f32 %f429, %f426, %f428;
setp.lt.f32 %p75, %f418, 0fC2D20000;
selp.f32 %f430, 0f00000000, %f429, %p75;
setp.gt.f32 %p76, %f418, 0f42D20000;
selp.f32 %f497, 0f7F800000, %f430, %p76;
setp.eq.f32 %p77, %f497, 0f7F800000;
@%p77 bra BB0_81;
fma.rn.f32 %f497, %f497, %f103, %f497;
BB0_81:
setp.lt.f32 %p78, %f496, 0f00000000;
setp.eq.f32 %p79, %f101, 0f3F800000;
and.pred %p2, %p78, %p79;
mov.b32 %r259, %f497;
xor.b32 %r260, %r259, -2147483648;
mov.b32 %f431, %r260;
selp.f32 %f499, %f431, %f497, %p2;
setp.eq.f32 %p80, %f496, 0f00000000;
@%p80 bra BB0_84;
bra.uni BB0_82;
BB0_84:
add.f32 %f434, %f496, %f496;
selp.f32 %f499, %f434, 0f00000000, %p79;
bra.uni BB0_85;
BB0_82:
setp.geu.f32 %p81, %f496, 0f00000000;
@%p81 bra BB0_85;
cvt.rzi.f32.f32 %f433, %f405;
setp.neu.f32 %p82, %f433, 0f3EE8BA2E;
selp.f32 %f499, 0f7FFFFFFF, %f499, %p82;
BB0_85:
add.f32 %f435, %f102, 0f3EE8BA2E;
mov.b32 %r261, %f435;
setp.lt.s32 %p84, %r261, 2139095040;
@%p84 bra BB0_90;
setp.gtu.f32 %p85, %f102, 0f7F800000;
@%p85 bra BB0_89;
bra.uni BB0_87;
BB0_89:
add.f32 %f499, %f496, 0f3EE8BA2E;
bra.uni BB0_90;
BB0_87:
setp.neu.f32 %p86, %f102, 0f7F800000;
@%p86 bra BB0_90;
selp.f32 %f499, 0fFF800000, 0f7F800000, %p2;
BB0_90:
mul.f32 %f436, %f499, 0f437F0000;
setp.eq.f32 %p87, %f496, 0f3F800000;
selp.f32 %f437, 0f437F0000, %f436, %p87;
cvt.rzi.u32.f32 %r262, %f437;
cvt.u16.u32 %rs24, %r262;
mov.u16 %rs25, 255;
st.v2.u8 [%rd73], {%rs24, %rs25};
ld.global.u32 %r312, [imageEnabled];
BB0_91:
and.b32 %r263, %r312, 64;
setp.eq.s32 %p88, %r263, 0;
@%p88 bra BB0_101;
mov.u32 %r283, 4;
mul.f32 %f438, %f487, %f487;
fma.rn.f32 %f439, %f488, %f488, %f438;
fma.rn.f32 %f440, %f486, %f486, %f439;
sqrt.rn.f32 %f441, %f440;
rcp.rn.f32 %f442, %f441;
mul.f32 %f443, %f488, %f442;
mul.f32 %f444, %f487, %f442;
mul.f32 %f445, %f486, %f442;
cvt.u64.u32 %rd83, %r3;
cvt.u64.u32 %rd82, %r2;
mov.u64 %rd86, image_Dir;
cvta.global.u64 %rd81, %rd86;
// inline asm
call (%rd80), _rt_buffer_get_64, (%rd81, %r95, %r283, %rd82, %rd83, %rd25, %rd25);
// inline asm
fma.rn.f32 %f446, %f443, 0f3F000000, 0f3F000000;
mul.f32 %f447, %f446, 0f437F0000;
cvt.rzi.u32.f32 %r266, %f447;
fma.rn.f32 %f448, %f444, 0f3F000000, 0f3F000000;
mul.f32 %f449, %f448, 0f437F0000;
cvt.rzi.u32.f32 %r267, %f449;
fma.rn.f32 %f450, %f445, 0f3F000000, 0f3F000000;
mul.f32 %f451, %f450, 0f437F0000;
cvt.rzi.u32.f32 %r268, %f451;
cvt.u16.u32 %rs26, %r268;
cvt.u16.u32 %rs27, %r267;
cvt.u16.u32 %rs28, %r266;
mov.u16 %rs29, 255;
st.v4.u8 [%rd80], {%rs28, %rs27, %rs26, %rs29};
BB0_101:
ret;
}