AsanoRema/FORT-BO
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
173bbb165cc3674ff53b9deab99020956412808e
FORT-BO/Assets/Editor/x64/Bakery/lmSunSuperSampleTranslucent.ptx
T
AsanoRema d9d7ee49d2 Added many other
2026-08-06 18:56:32 +03:00

2646 lines
79 KiB
Plaintext
Raw Blame History

//
// 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_Dir[1];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 1 .b8 rnd_seeds[1];
.global .align 4 .b8 directDir[12];
.global .align 4 .b8 directColor[12];
.global .align 4 .f32 shadowSpread;
.global .align 4 .u32 samples;
.global .align 4 .u32 ignoreNormal;
.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_typeinfo9directDirE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo11directColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12shadowSpreadE[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_typeinfo12ignoreNormalE[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 8 .b8 _ZN21rti_internal_typename9directDirE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename11directColorE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename12shadowSpreadE[6] = {102, 108, 111, 97, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[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_typeenum9directDirE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum11directColorE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12shadowSpreadE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 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_semantic9directDirE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic11directColorE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12shadowSpreadE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[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_annotation9directDirE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation11directColorE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12shadowSpreadE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[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[52];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<176>;
.reg .b16 %rs<105>;
.reg .f32 %f<1468>;
.reg .b32 %r<487>;
.reg .b64 %rd<223>;
mov.u64 %rd222, __local_depot0;
cvta.local.u64 %SP, %rd222;
ld.global.v2.u32 {%r114, %r115}, [pixelID];
cvt.u64.u32 %rd26, %r114;
cvt.u64.u32 %rd27, %r115;
mov.u64 %rd30, uvnormal;
cvta.global.u64 %rd25, %rd30;
mov.u32 %r112, 2;
mov.u32 %r113, 4;
mov.u64 %rd29, 0;
// inline asm
call (%rd24), _rt_buffer_get_64, (%rd25, %r112, %r113, %rd26, %rd27, %rd29, %rd29);
// inline asm
ld.u32 %r1, [%rd24];
shr.u32 %r118, %r1, 16;
cvt.u16.u32 %rs1, %r118;
and.b16 %rs12, %rs1, 255;
cvt.u16.u32 %rs13, %r1;
or.b16 %rs14, %rs13, %rs12;
setp.eq.s16 %p5, %rs14, 0;
mov.f32 %f1386, 0f00000000;
mov.f32 %f1387, %f1386;
mov.f32 %f1388, %f1386;
@%p5 bra BB0_2;
ld.u8 %rs15, [%rd24+1];
and.b16 %rs17, %rs13, 255;
cvt.rn.f32.u16 %f250, %rs17;
div.rn.f32 %f251, %f250, 0f437F0000;
fma.rn.f32 %f252, %f251, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f253, %rs15;
div.rn.f32 %f254, %f253, 0f437F0000;
fma.rn.f32 %f255, %f254, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f256, %rs12;
div.rn.f32 %f257, %f256, 0f437F0000;
fma.rn.f32 %f258, %f257, 0f40000000, 0fBF800000;
mul.f32 %f259, %f255, %f255;
fma.rn.f32 %f260, %f252, %f252, %f259;
fma.rn.f32 %f261, %f258, %f258, %f260;
sqrt.rn.f32 %f262, %f261;
rcp.rn.f32 %f263, %f262;
mul.f32 %f1386, %f252, %f263;
mul.f32 %f1387, %f255, %f263;
mul.f32 %f1388, %f258, %f263;
BB0_2:
ld.global.v2.u32 {%r119, %r120}, [pixelID];
ld.global.v2.u32 {%r122, %r123}, [tileInfo];
add.s32 %r2, %r119, %r122;
add.s32 %r3, %r120, %r123;
setp.eq.f32 %p6, %f1387, 0f00000000;
setp.eq.f32 %p7, %f1386, 0f00000000;
and.pred %p8, %p7, %p6;
setp.eq.f32 %p9, %f1388, 0f00000000;
and.pred %p10, %p8, %p9;
@%p10 bra BB0_157;
bra.uni BB0_3;
BB0_157:
ld.global.u32 %r486, [imageEnabled];
and.b32 %r429, %r486, 1;
setp.eq.b32 %p171, %r429, 1;
@!%p171 bra BB0_159;
bra.uni BB0_158;
BB0_158:
cvt.u64.u32 %rd179, %r2;
cvt.u64.u32 %rd180, %r3;
mov.u64 %rd183, image;
cvta.global.u64 %rd178, %rd183;
mov.u64 %rd182, 0;
// inline asm
call (%rd177), _rt_buffer_get_64, (%rd178, %r112, %r113, %rd179, %rd180, %rd182, %rd182);
// inline asm
mov.u16 %rs76, 0;
st.v4.u8 [%rd177], {%rs76, %rs76, %rs76, %rs76};
ld.global.u32 %r486, [imageEnabled];
BB0_159:
and.b32 %r432, %r486, 8;
setp.eq.s32 %p172, %r432, 0;
@%p172 bra BB0_161;
cvt.u64.u32 %rd186, %r2;
cvt.u64.u32 %rd187, %r3;
mov.u64 %rd190, image_Mask;
cvta.global.u64 %rd185, %rd190;
mov.u64 %rd189, 0;
// inline asm
call (%rd184), _rt_buffer_get_64, (%rd185, %r112, %r112, %rd186, %rd187, %rd189, %rd189);
// inline asm
mov.f32 %f1376, 0f00000000;
cvt.rzi.u32.f32 %r435, %f1376;
cvt.u16.u32 %rs77, %r435;
mov.u16 %rs78, 0;
st.v2.u8 [%rd184], {%rs77, %rs78};
ld.global.u32 %r486, [imageEnabled];
BB0_161:
and.b32 %r436, %r486, 4;
setp.eq.s32 %p173, %r436, 0;
@%p173 bra BB0_165;
ld.global.u32 %r437, [additive];
setp.eq.s32 %p174, %r437, 0;
cvt.u64.u32 %rd22, %r2;
cvt.u64.u32 %rd23, %r3;
@%p174 bra BB0_164;
mov.u64 %rd203, image_HDR;
cvta.global.u64 %rd192, %rd203;
mov.u32 %r441, 8;
mov.u64 %rd202, 0;
// inline asm
call (%rd191), _rt_buffer_get_64, (%rd192, %r112, %r441, %rd22, %rd23, %rd202, %rd202);
// inline asm
ld.v4.u16 {%rs85, %rs86, %rs87, %rs88}, [%rd191];
// inline asm
{ cvt.f32.f16 %f1377, %rs85;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1378, %rs86;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1379, %rs87;}
// inline asm
// inline asm
call (%rd197), _rt_buffer_get_64, (%rd192, %r112, %r441, %rd22, %rd23, %rd202, %rd202);
// inline asm
add.f32 %f1380, %f1377, 0f00000000;
add.f32 %f1381, %f1378, 0f00000000;
add.f32 %f1382, %f1379, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs84, %f1382;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs83, %f1381;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs82, %f1380;}
// inline asm
mov.u16 %rs89, 0;
st.v4.u16 [%rd197], {%rs82, %rs83, %rs84, %rs89};
bra.uni BB0_165;
BB0_3:
ld.global.f32 %f9, [directDir+8];
ld.global.f32 %f8, [directDir+4];
ld.global.f32 %f7, [directDir];
mul.f32 %f264, %f1386, %f7;
mul.f32 %f265, %f1387, %f8;
neg.f32 %f266, %f265;
sub.f32 %f267, %f266, %f264;
mul.f32 %f268, %f1388, %f9;
sub.f32 %f10, %f267, %f268;
ld.global.u32 %r127, [ignoreNormal];
setp.eq.s32 %p11, %r127, 0;
setp.le.f32 %p12, %f10, 0f00000000;
and.pred %p13, %p11, %p12;
ld.global.u32 %r484, [imageEnabled];
and.b32 %r128, %r484, 32;
setp.eq.s32 %p14, %r128, 0;
and.pred %p15, %p13, %p14;
@%p15 bra BB0_147;
bra.uni BB0_4;
BB0_147:
and.b32 %r412, %r484, 1;
setp.eq.b32 %p166, %r412, 1;
@!%p166 bra BB0_149;
bra.uni BB0_148;
BB0_148:
cvt.u64.u32 %rd138, %r2;
cvt.u64.u32 %rd139, %r3;
mov.u64 %rd142, image;
cvta.global.u64 %rd137, %rd142;
mov.u64 %rd141, 0;
// inline asm
call (%rd136), _rt_buffer_get_64, (%rd137, %r112, %r113, %rd138, %rd139, %rd141, %rd141);
// inline asm
mov.u16 %rs56, 255;
mov.u16 %rs57, 0;
st.v4.u8 [%rd136], {%rs57, %rs57, %rs57, %rs56};
ld.global.u32 %r484, [imageEnabled];
BB0_149:
and.b32 %r415, %r484, 8;
setp.eq.s32 %p167, %r415, 0;
@%p167 bra BB0_151;
cvt.u64.u32 %rd145, %r2;
cvt.u64.u32 %rd146, %r3;
mov.u64 %rd149, image_Mask;
cvta.global.u64 %rd144, %rd149;
mov.u64 %rd148, 0;
// inline asm
call (%rd143), _rt_buffer_get_64, (%rd144, %r112, %r112, %rd145, %rd146, %rd148, %rd148);
// inline asm
mov.f32 %f1365, 0f00000000;
cvt.rzi.u32.f32 %r418, %f1365;
cvt.u16.u32 %rs58, %r418;
mov.u16 %rs59, 255;
st.v2.u8 [%rd143], {%rs58, %rs59};
ld.global.u32 %r484, [imageEnabled];
BB0_151:
and.b32 %r419, %r484, 4;
setp.eq.s32 %p168, %r419, 0;
@%p168 bra BB0_155;
ld.global.u32 %r420, [additive];
setp.eq.s32 %p169, %r420, 0;
cvt.u64.u32 %rd20, %r2;
cvt.u64.u32 %rd21, %r3;
mov.f32 %f1366, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs60, %f1366;}
// inline asm
@%p169 bra BB0_154;
mov.u64 %rd162, image_HDR;
cvta.global.u64 %rd151, %rd162;
mov.u32 %r424, 8;
mov.u64 %rd161, 0;
// inline asm
call (%rd150), _rt_buffer_get_64, (%rd151, %r112, %r424, %rd20, %rd21, %rd161, %rd161);
// inline asm
ld.v4.u16 {%rs67, %rs68, %rs69, %rs70}, [%rd150];
// inline asm
{ cvt.f32.f16 %f1367, %rs67;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1368, %rs68;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1369, %rs69;}
// inline asm
// inline asm
call (%rd156), _rt_buffer_get_64, (%rd151, %r112, %r424, %rd20, %rd21, %rd161, %rd161);
// inline asm
add.f32 %f1370, %f1367, 0f00000000;
add.f32 %f1371, %f1368, 0f00000000;
add.f32 %f1372, %f1369, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs66, %f1372;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs65, %f1371;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs64, %f1370;}
// inline asm
st.v4.u16 [%rd156], {%rs64, %rs65, %rs66, %rs60};
bra.uni BB0_155;
BB0_4:
ld.global.v2.u32 {%r137, %r138}, [pixelID];
cvt.u64.u32 %rd33, %r137;
cvt.u64.u32 %rd34, %r138;
mov.u64 %rd55, uvpos;
cvta.global.u64 %rd32, %rd55;
mov.u32 %r134, 12;
// inline asm
call (%rd31), _rt_buffer_get_64, (%rd32, %r112, %r134, %rd33, %rd34, %rd29, %rd29);
// inline asm
ld.f32 %f273, [%rd31+8];
ld.f32 %f274, [%rd31+4];
ld.f32 %f275, [%rd31];
mul.f32 %f276, %f275, 0f3456BF95;
mul.f32 %f277, %f274, 0f3456BF95;
mul.f32 %f278, %f273, 0f3456BF95;
abs.f32 %f279, %f1386;
div.rn.f32 %f280, %f276, %f279;
abs.f32 %f281, %f1387;
div.rn.f32 %f282, %f277, %f281;
abs.f32 %f283, %f1388;
div.rn.f32 %f284, %f278, %f283;
abs.f32 %f285, %f280;
abs.f32 %f286, %f282;
abs.f32 %f287, %f284;
mov.f32 %f288, 0f38D1B717;
max.f32 %f289, %f285, %f288;
max.f32 %f290, %f286, %f288;
max.f32 %f291, %f287, %f288;
fma.rn.f32 %f11, %f1386, %f289, %f275;
fma.rn.f32 %f12, %f1387, %f290, %f274;
fma.rn.f32 %f13, %f1388, %f291, %f273;
setp.gt.f32 %p16, %f279, %f283;
neg.f32 %f292, %f1387;
selp.f32 %f293, %f292, 0f00000000, %p16;
neg.f32 %f294, %f1388;
selp.f32 %f295, %f1386, %f294, %p16;
selp.f32 %f296, 0f00000000, %f1387, %p16;
mul.f32 %f297, %f295, %f295;
fma.rn.f32 %f298, %f293, %f293, %f297;
fma.rn.f32 %f299, %f296, %f296, %f298;
sqrt.rn.f32 %f300, %f299;
rcp.rn.f32 %f301, %f300;
mul.f32 %f14, %f293, %f301;
mul.f32 %f15, %f295, %f301;
mul.f32 %f16, %f296, %f301;
mul.f32 %f302, %f1388, %f15;
mul.f32 %f303, %f1387, %f16;
sub.f32 %f17, %f302, %f303;
mul.f32 %f304, %f1386, %f16;
mul.f32 %f305, %f1388, %f14;
sub.f32 %f18, %f304, %f305;
mul.f32 %f306, %f1387, %f14;
mul.f32 %f307, %f1386, %f15;
sub.f32 %f19, %f306, %f307;
ld.global.v2.u32 {%r141, %r142}, [pixelID];
cvt.u64.u32 %rd40, %r142;
add.s32 %r145, %r141, 1;
ld.global.u32 %r146, [resolution];
add.s32 %r147, %r146, -1;
min.u32 %r148, %r145, %r147;
cvt.u64.u32 %rd39, %r148;
// inline asm
call (%rd37), _rt_buffer_get_64, (%rd32, %r112, %r134, %rd39, %rd40, %rd29, %rd29);
// inline asm
ld.f32 %f308, [%rd37+8];
ld.f32 %f309, [%rd37+4];
ld.f32 %f310, [%rd37];
setp.eq.f32 %p17, %f310, 0f00000000;
setp.eq.f32 %p18, %f309, 0f00000000;
and.pred %p19, %p17, %p18;
setp.eq.f32 %p20, %f308, 0f00000000;
and.pred %p21, %p19, %p20;
selp.f32 %f311, %f275, %f310, %p21;
selp.f32 %f312, %f274, %f309, %p21;
selp.f32 %f313, %f273, %f308, %p21;
ld.global.v2.u32 {%r149, %r150}, [pixelID];
add.s32 %r153, %r150, 1;
ld.global.u32 %r154, [resolution+4];
add.s32 %r155, %r154, -1;
min.u32 %r156, %r153, %r155;
cvt.u64.u32 %rd45, %r149;
cvt.u64.u32 %rd46, %r156;
// inline asm
call (%rd43), _rt_buffer_get_64, (%rd32, %r112, %r134, %rd45, %rd46, %rd29, %rd29);
// inline asm
ld.f32 %f314, [%rd43+8];
ld.f32 %f315, [%rd43+4];
ld.f32 %f316, [%rd43];
setp.eq.f32 %p22, %f316, 0f00000000;
setp.eq.f32 %p23, %f315, 0f00000000;
and.pred %p24, %p22, %p23;
setp.eq.f32 %p25, %f314, 0f00000000;
and.pred %p26, %p24, %p25;
selp.f32 %f317, %f275, %f316, %p26;
selp.f32 %f318, %f274, %f315, %p26;
selp.f32 %f319, %f273, %f314, %p26;
sub.f32 %f320, %f311, %f275;
sub.f32 %f321, %f312, %f274;
sub.f32 %f322, %f313, %f273;
sub.f32 %f323, %f317, %f275;
sub.f32 %f324, %f318, %f274;
sub.f32 %f325, %f319, %f273;
abs.f32 %f326, %f320;
abs.f32 %f327, %f323;
max.f32 %f328, %f326, %f327;
abs.f32 %f329, %f321;
abs.f32 %f330, %f324;
max.f32 %f331, %f329, %f330;
abs.f32 %f332, %f322;
abs.f32 %f333, %f325;
max.f32 %f334, %f332, %f333;
max.f32 %f335, %f328, %f331;
max.f32 %f336, %f335, %f334;
setp.neu.f32 %p27, %f320, 0f00000000;
setp.neu.f32 %p28, %f321, 0f00000000;
or.pred %p29, %p27, %p28;
setp.neu.f32 %p30, %f322, 0f00000000;
setp.neu.f32 %p31, %f323, 0f00000000;
setp.neu.f32 %p32, %f324, 0f00000000;
or.pred %p33, %p31, %p32;
setp.neu.f32 %p34, %f325, 0f00000000;
or.pred %p35, %p33, %p34;
or.pred %p36, %p29, %p30;
and.pred %p37, %p36, %p35;
selp.f32 %f20, %f336, 0f00000000, %p37;
abs.f32 %f337, %f9;
abs.f32 %f338, %f7;
setp.gt.f32 %p38, %f338, %f337;
neg.f32 %f339, %f8;
selp.f32 %f340, %f339, 0f00000000, %p38;
neg.f32 %f341, %f9;
selp.f32 %f342, %f7, %f341, %p38;
selp.f32 %f343, 0f00000000, %f8, %p38;
mul.f32 %f344, %f342, %f342;
fma.rn.f32 %f345, %f340, %f340, %f344;
fma.rn.f32 %f346, %f343, %f343, %f345;
sqrt.rn.f32 %f347, %f346;
rcp.rn.f32 %f348, %f347;
mul.f32 %f21, %f340, %f348;
mul.f32 %f22, %f342, %f348;
mul.f32 %f23, %f343, %f348;
ld.global.v2.u32 {%r157, %r158}, [pixelID];
cvt.u64.u32 %rd51, %r157;
cvt.u64.u32 %rd52, %r158;
mov.u64 %rd56, rnd_seeds;
cvta.global.u64 %rd50, %rd56;
// inline asm
call (%rd49), _rt_buffer_get_64, (%rd50, %r112, %r113, %rd51, %rd52, %rd29, %rd29);
// inline asm
ld.global.u32 %r453, [samples];
mov.f32 %f37, 0f00000000;
setp.lt.s32 %p39, %r453, 1;
@%p39 bra BB0_5;
cvt.rn.f32.s32 %f353, %r453;
rcp.rn.f32 %f24, %f353;
ld.u32 %r479, [%rd49];
mul.f32 %f25, %f11, 0f3456BF95;
mul.f32 %f26, %f12, 0f3456BF95;
mul.f32 %f27, %f13, 0f3456BF95;
mul.f32 %f354, %f7, %f22;
mul.f32 %f355, %f8, %f21;
sub.f32 %f28, %f355, %f354;
mul.f32 %f356, %f9, %f21;
mul.f32 %f357, %f7, %f23;
sub.f32 %f29, %f357, %f356;
mul.f32 %f358, %f8, %f23;
mul.f32 %f359, %f9, %f22;
sub.f32 %f30, %f359, %f358;
mov.f32 %f37, 0f00000000;
mov.u32 %r454, 0;
abs.f32 %f360, %f26;
abs.f32 %f361, %f25;
max.f32 %f362, %f361, %f360;
abs.f32 %f363, %f27;
max.f32 %f364, %f362, %f363;
mov.f32 %f38, %f37;
mov.f32 %f39, %f37;
mov.f32 %f40, %f37;
BB0_7:
setp.lt.s32 %p40, %r453, 1;
@%p40 bra BB0_90;
mov.u32 %r456, 0;
BB0_9:
cvt.rn.f32.s32 %f1384, %r454;
mad.lo.s32 %r163, %r479, 1664525, 1013904223;
and.b32 %r164, %r163, 16777215;
cvt.rn.f32.u32 %f366, %r164;
fma.rn.f32 %f367, %f366, 0f33800000, %f1384;
mul.f32 %f368, %f24, %f367;
mad.lo.s32 %r12, %r163, 1664525, 1013904223;
and.b32 %r165, %r12, 16777215;
cvt.rn.f32.u32 %f369, %r165;
cvt.rn.f32.s32 %f370, %r456;
fma.rn.f32 %f371, %f369, 0f33800000, %f370;
mul.f32 %f372, %f24, %f371;
sqrt.rn.f32 %f41, %f368;
mul.f32 %f1403, %f372, 0f40C90FDB;
abs.f32 %f43, %f1403;
setp.neu.f32 %p41, %f43, 0f7F800000;
mov.f32 %f1397, %f1403;
@%p41 bra BB0_11;
mov.f32 %f373, 0f00000000;
mul.rn.f32 %f1397, %f1403, %f373;
BB0_11:
mul.f32 %f374, %f1397, 0f3F22F983;
cvt.rni.s32.f32 %r467, %f374;
cvt.rn.f32.s32 %f375, %r467;
neg.f32 %f376, %f375;
mov.f32 %f377, 0f3FC90FDA;
fma.rn.f32 %f378, %f376, %f377, %f1397;
mov.f32 %f379, 0f33A22168;
fma.rn.f32 %f380, %f376, %f379, %f378;
mov.f32 %f381, 0f27C234C5;
fma.rn.f32 %f1398, %f376, %f381, %f380;
abs.f32 %f382, %f1397;
setp.leu.f32 %p42, %f382, 0f47CE4780;
@%p42 bra BB0_22;
add.u64 %rd58, %SP, 0;
cvta.to.local.u64 %rd3, %rd58;
mov.b32 %r14, %f1397;
shr.u32 %r15, %r14, 23;
shl.b32 %r168, %r14, 8;
or.b32 %r16, %r168, -2147483648;
mov.u32 %r458, 0;
mov.u64 %rd219, 0;
mov.u64 %rd218, %rd3;
mov.u32 %r459, %r458;
BB0_13:
.pragma "nounroll";
shl.b64 %rd59, %rd219, 2;
mov.u64 %rd60, __cudart_i2opi_f;
add.s64 %rd61, %rd60, %rd59;
ld.const.u32 %r171, [%rd61];
// inline asm
{
mad.lo.cc.u32 %r169, %r171, %r16, %r459;
madc.hi.u32 %r459, %r171, %r16, 0;
}
// inline asm
st.local.u32 [%rd218], %r169;
add.s32 %r458, %r458, 1;
cvt.s64.s32 %rd219, %r458;
mul.wide.s32 %rd64, %r458, 4;
add.s64 %rd218, %rd3, %rd64;
setp.ne.s32 %p43, %r458, 6;
@%p43 bra BB0_13;
and.b32 %r174, %r15, 255;
add.s32 %r175, %r174, -128;
shr.u32 %r176, %r175, 5;
and.b32 %r21, %r14, -2147483648;
cvta.to.local.u64 %rd66, %rd58;
st.local.u32 [%rd66+24], %r459;
mov.u32 %r177, 6;
sub.s32 %r178, %r177, %r176;
mul.wide.s32 %rd67, %r178, 4;
add.s64 %rd8, %rd66, %rd67;
ld.local.u32 %r460, [%rd8];
ld.local.u32 %r461, [%rd8+-4];
and.b32 %r24, %r15, 31;
setp.eq.s32 %p44, %r24, 0;
@%p44 bra BB0_16;
mov.u32 %r179, 32;
sub.s32 %r180, %r179, %r24;
shr.u32 %r181, %r461, %r180;
shl.b32 %r182, %r460, %r24;
add.s32 %r460, %r181, %r182;
ld.local.u32 %r183, [%rd8+-8];
shr.u32 %r184, %r183, %r180;
shl.b32 %r185, %r461, %r24;
add.s32 %r461, %r184, %r185;
BB0_16:
shr.u32 %r186, %r461, 30;
shl.b32 %r187, %r460, 2;
add.s32 %r462, %r186, %r187;
shl.b32 %r30, %r461, 2;
shr.u32 %r188, %r462, 31;
shr.u32 %r189, %r460, 30;
add.s32 %r31, %r188, %r189;
setp.eq.s32 %p45, %r188, 0;
@%p45 bra BB0_17;
bra.uni BB0_18;
BB0_17:
mov.u32 %r463, %r21;
mov.u32 %r464, %r30;
bra.uni BB0_19;
BB0_18:
not.b32 %r190, %r462;
neg.s32 %r464, %r30;
setp.eq.s32 %p46, %r30, 0;
selp.u32 %r191, 1, 0, %p46;
add.s32 %r462, %r191, %r190;
xor.b32 %r463, %r21, -2147483648;
BB0_19:
clz.b32 %r466, %r462;
setp.eq.s32 %p47, %r466, 0;
shl.b32 %r192, %r462, %r466;
mov.u32 %r193, 32;
sub.s32 %r194, %r193, %r466;
shr.u32 %r195, %r464, %r194;
add.s32 %r196, %r195, %r192;
selp.b32 %r39, %r462, %r196, %p47;
mov.u32 %r197, -921707870;
mul.hi.u32 %r465, %r39, %r197;
setp.eq.s32 %p48, %r21, 0;
neg.s32 %r198, %r31;
selp.b32 %r467, %r31, %r198, %p48;
setp.lt.s32 %p49, %r465, 1;
@%p49 bra BB0_21;
mul.lo.s32 %r199, %r39, -921707870;
shr.u32 %r200, %r199, 31;
shl.b32 %r201, %r465, 1;
add.s32 %r465, %r200, %r201;
add.s32 %r466, %r466, 1;
BB0_21:
mov.u32 %r202, 126;
sub.s32 %r203, %r202, %r466;
shl.b32 %r204, %r203, 23;
add.s32 %r205, %r465, 1;
shr.u32 %r206, %r205, 7;
add.s32 %r207, %r206, 1;
shr.u32 %r208, %r207, 1;
add.s32 %r209, %r208, %r204;
or.b32 %r210, %r209, %r463;
mov.b32 %f1398, %r210;
BB0_22:
mul.rn.f32 %f49, %f1398, %f1398;
add.s32 %r47, %r467, 1;
and.b32 %r48, %r47, 1;
setp.eq.s32 %p50, %r48, 0;
@%p50 bra BB0_24;
bra.uni BB0_23;
BB0_24:
mov.f32 %f385, 0f3C08839E;
mov.f32 %f386, 0fB94CA1F9;
fma.rn.f32 %f1399, %f386, %f49, %f385;
bra.uni BB0_25;
BB0_23:
mov.f32 %f383, 0fBAB6061A;
mov.f32 %f384, 0f37CCF5CE;
fma.rn.f32 %f1399, %f384, %f49, %f383;
BB0_25:
@%p50 bra BB0_27;
bra.uni BB0_26;
BB0_27:
mov.f32 %f390, 0fBE2AAAA3;
fma.rn.f32 %f391, %f1399, %f49, %f390;
mov.f32 %f392, 0f00000000;
fma.rn.f32 %f1400, %f391, %f49, %f392;
bra.uni BB0_28;
BB0_26:
mov.f32 %f387, 0f3D2AAAA5;
fma.rn.f32 %f388, %f1399, %f49, %f387;
mov.f32 %f389, 0fBF000000;
fma.rn.f32 %f1400, %f388, %f49, %f389;
BB0_28:
fma.rn.f32 %f1401, %f1400, %f1398, %f1398;
@%p50 bra BB0_30;
mov.f32 %f393, 0f3F800000;
fma.rn.f32 %f1401, %f1400, %f49, %f393;
BB0_30:
and.b32 %r211, %r47, 2;
setp.eq.s32 %p53, %r211, 0;
@%p53 bra BB0_32;
mov.f32 %f394, 0f00000000;
mov.f32 %f395, 0fBF800000;
fma.rn.f32 %f1401, %f1401, %f395, %f394;
BB0_32:
@%p41 bra BB0_34;
mov.f32 %f396, 0f00000000;
mul.rn.f32 %f1403, %f1403, %f396;
BB0_34:
mul.f32 %f397, %f1403, 0f3F22F983;
cvt.rni.s32.f32 %r477, %f397;
cvt.rn.f32.s32 %f398, %r477;
neg.f32 %f399, %f398;
fma.rn.f32 %f401, %f399, %f377, %f1403;
fma.rn.f32 %f403, %f399, %f379, %f401;
fma.rn.f32 %f1404, %f399, %f381, %f403;
abs.f32 %f405, %f1403;
setp.leu.f32 %p55, %f405, 0f47CE4780;
@%p55 bra BB0_45;
add.u64 %rd69, %SP, 0;
cvta.to.local.u64 %rd9, %rd69;
mov.b32 %r50, %f1403;
shr.u32 %r51, %r50, 23;
shl.b32 %r214, %r50, 8;
or.b32 %r52, %r214, -2147483648;
mov.u32 %r468, 0;
mov.u64 %rd220, %rd9;
mov.u64 %rd221, %rd29;
mov.u32 %r469, %r468;
BB0_36:
.pragma "nounroll";
shl.b64 %rd70, %rd221, 2;
mov.u64 %rd71, __cudart_i2opi_f;
add.s64 %rd72, %rd71, %rd70;
ld.const.u32 %r217, [%rd72];
// inline asm
{
mad.lo.cc.u32 %r215, %r217, %r52, %r469;
madc.hi.u32 %r469, %r217, %r52, 0;
}
// inline asm
st.local.u32 [%rd220], %r215;
add.s32 %r468, %r468, 1;
cvt.s64.s32 %rd221, %r468;
mul.wide.s32 %rd73, %r468, 4;
add.s64 %rd220, %rd9, %rd73;
setp.ne.s32 %p56, %r468, 6;
@%p56 bra BB0_36;
and.b32 %r220, %r51, 255;
add.s32 %r221, %r220, -128;
shr.u32 %r222, %r221, 5;
and.b32 %r57, %r50, -2147483648;
cvta.to.local.u64 %rd75, %rd69;
st.local.u32 [%rd75+24], %r469;
mov.u32 %r223, 6;
sub.s32 %r224, %r223, %r222;
mul.wide.s32 %rd76, %r224, 4;
add.s64 %rd15, %rd75, %rd76;
ld.local.u32 %r470, [%rd15];
ld.local.u32 %r471, [%rd15+-4];
and.b32 %r60, %r51, 31;
setp.eq.s32 %p57, %r60, 0;
@%p57 bra BB0_39;
mov.u32 %r225, 32;
sub.s32 %r226, %r225, %r60;
shr.u32 %r227, %r471, %r226;
shl.b32 %r228, %r470, %r60;
add.s32 %r470, %r227, %r228;
ld.local.u32 %r229, [%rd15+-8];
shr.u32 %r230, %r229, %r226;
shl.b32 %r231, %r471, %r60;
add.s32 %r471, %r230, %r231;
BB0_39:
shr.u32 %r232, %r471, 30;
shl.b32 %r233, %r470, 2;
add.s32 %r472, %r232, %r233;
shl.b32 %r66, %r471, 2;
shr.u32 %r234, %r472, 31;
shr.u32 %r235, %r470, 30;
add.s32 %r67, %r234, %r235;
setp.eq.s32 %p58, %r234, 0;
@%p58 bra BB0_40;
bra.uni BB0_41;
BB0_40:
mov.u32 %r473, %r57;
mov.u32 %r474, %r66;
bra.uni BB0_42;
BB0_41:
not.b32 %r236, %r472;
neg.s32 %r474, %r66;
setp.eq.s32 %p59, %r66, 0;
selp.u32 %r237, 1, 0, %p59;
add.s32 %r472, %r237, %r236;
xor.b32 %r473, %r57, -2147483648;
BB0_42:
clz.b32 %r476, %r472;
setp.eq.s32 %p60, %r476, 0;
shl.b32 %r238, %r472, %r476;
mov.u32 %r239, 32;
sub.s32 %r240, %r239, %r476;
shr.u32 %r241, %r474, %r240;
add.s32 %r242, %r241, %r238;
selp.b32 %r75, %r472, %r242, %p60;
mov.u32 %r243, -921707870;
mul.hi.u32 %r475, %r75, %r243;
setp.eq.s32 %p61, %r57, 0;
neg.s32 %r244, %r67;
selp.b32 %r477, %r67, %r244, %p61;
setp.lt.s32 %p62, %r475, 1;
@%p62 bra BB0_44;
mul.lo.s32 %r245, %r75, -921707870;
shr.u32 %r246, %r245, 31;
shl.b32 %r247, %r475, 1;
add.s32 %r475, %r246, %r247;
add.s32 %r476, %r476, 1;
BB0_44:
mov.u32 %r248, 126;
sub.s32 %r249, %r248, %r476;
shl.b32 %r250, %r249, 23;
add.s32 %r251, %r475, 1;
shr.u32 %r252, %r251, 7;
add.s32 %r253, %r252, 1;
shr.u32 %r254, %r253, 1;
add.s32 %r255, %r254, %r250;
or.b32 %r256, %r255, %r473;
mov.b32 %f1404, %r256;
BB0_45:
mul.rn.f32 %f66, %f1404, %f1404;
and.b32 %r83, %r477, 1;
setp.eq.s32 %p63, %r83, 0;
@%p63 bra BB0_47;
bra.uni BB0_46;
BB0_47:
mov.f32 %f408, 0f3C08839E;
mov.f32 %f409, 0fB94CA1F9;
fma.rn.f32 %f1405, %f409, %f66, %f408;
bra.uni BB0_48;
BB0_46:
mov.f32 %f406, 0fBAB6061A;
mov.f32 %f407, 0f37CCF5CE;
fma.rn.f32 %f1405, %f407, %f66, %f406;
BB0_48:
@%p63 bra BB0_50;
bra.uni BB0_49;
BB0_50:
mov.f32 %f413, 0fBE2AAAA3;
fma.rn.f32 %f414, %f1405, %f66, %f413;
mov.f32 %f415, 0f00000000;
fma.rn.f32 %f1406, %f414, %f66, %f415;
bra.uni BB0_51;
BB0_49:
mov.f32 %f410, 0f3D2AAAA5;
fma.rn.f32 %f411, %f1405, %f66, %f410;
mov.f32 %f412, 0fBF000000;
fma.rn.f32 %f1406, %f411, %f66, %f412;
BB0_51:
fma.rn.f32 %f1407, %f1406, %f1404, %f1404;
@%p63 bra BB0_53;
mov.f32 %f416, 0f3F800000;
fma.rn.f32 %f1407, %f1406, %f66, %f416;
BB0_53:
and.b32 %r257, %r477, 2;
setp.eq.s32 %p66, %r257, 0;
@%p66 bra BB0_55;
mov.f32 %f417, 0f00000000;
mov.f32 %f418, 0fBF800000;
fma.rn.f32 %f1407, %f1407, %f418, %f417;
BB0_55:
max.f32 %f1385, %f364, %f288;
mul.f32 %f427, %f41, %f1401;
add.u64 %rd77, %SP, 28;
cvta.to.local.u64 %rd16, %rd77;
mul.f32 %f428, %f427, %f427;
mov.f32 %f429, 0f3F800000;
sub.f32 %f430, %f429, %f428;
mul.f32 %f431, %f41, %f1407;
mul.f32 %f432, %f431, %f431;
sub.f32 %f433, %f430, %f432;
mov.f32 %f434, 0f00000000;
max.f32 %f435, %f434, %f433;
sqrt.rn.f32 %f436, %f435;
mul.f32 %f437, %f21, %f431;
mul.f32 %f438, %f22, %f431;
mul.f32 %f439, %f23, %f431;
fma.rn.f32 %f440, %f30, %f427, %f437;
fma.rn.f32 %f441, %f29, %f427, %f438;
fma.rn.f32 %f442, %f28, %f427, %f439;
fma.rn.f32 %f443, %f7, %f436, %f440;
fma.rn.f32 %f444, %f8, %f436, %f441;
fma.rn.f32 %f445, %f9, %f436, %f442;
add.f32 %f446, %f7, %f443;
add.f32 %f447, %f8, %f444;
add.f32 %f448, %f9, %f445;
ld.global.f32 %f449, [shadowSpread];
mul.f32 %f450, %f449, %f446;
mul.f32 %f451, %f449, %f447;
mul.f32 %f452, %f449, %f448;
sub.f32 %f453, %f450, %f7;
sub.f32 %f454, %f451, %f8;
sub.f32 %f455, %f452, %f9;
mul.f32 %f456, %f454, %f454;
fma.rn.f32 %f457, %f453, %f453, %f456;
fma.rn.f32 %f458, %f455, %f455, %f457;
sqrt.rn.f32 %f459, %f458;
rcp.rn.f32 %f460, %f459;
mul.f32 %f78, %f460, %f453;
mul.f32 %f79, %f460, %f454;
mul.f32 %f80, %f460, %f455;
ld.global.u32 %r261, [imageEnabled];
and.b32 %r262, %r261, 32;
setp.eq.s32 %p67, %r262, 0;
selp.f32 %f461, 0f3F800000, 0f41200000, %p67;
mul.f32 %f425, %f461, %f1385;
mad.lo.s32 %r263, %r12, 1664525, 1013904223;
and.b32 %r264, %r263, 16777215;
cvt.rn.f32.u32 %f462, %r264;
mul.f32 %f463, %f462, 0f33800000;
fma.rn.f32 %f464, %f463, 0f40000000, 0fBF800000;
mad.lo.s32 %r84, %r263, 1664525, 1013904223;
and.b32 %r265, %r84, 16777215;
cvt.rn.f32.u32 %f465, %r265;
mul.f32 %f466, %f465, 0f33800000;
fma.rn.f32 %f467, %f466, 0f40000000, 0fBF800000;
mul.f32 %f468, %f14, %f467;
mul.f32 %f469, %f15, %f467;
mul.f32 %f470, %f16, %f467;
fma.rn.f32 %f471, %f17, %f464, %f468;
fma.rn.f32 %f472, %f18, %f464, %f469;
fma.rn.f32 %f473, %f19, %f464, %f470;
mul.f32 %f474, %f472, %f472;
fma.rn.f32 %f475, %f471, %f471, %f474;
fma.rn.f32 %f476, %f473, %f473, %f475;
sqrt.rn.f32 %f477, %f476;
rcp.rn.f32 %f478, %f477;
mul.f32 %f479, %f471, %f478;
mul.f32 %f480, %f472, %f478;
mul.f32 %f481, %f473, %f478;
mul.f32 %f482, %f1387, %f481;
mul.f32 %f483, %f1388, %f480;
sub.f32 %f484, %f482, %f483;
mul.f32 %f485, %f1388, %f479;
mul.f32 %f486, %f1386, %f481;
sub.f32 %f487, %f485, %f486;
mul.f32 %f488, %f1386, %f480;
mul.f32 %f489, %f1387, %f479;
sub.f32 %f490, %f488, %f489;
mul.f32 %f491, %f479, 0f3E24A490;
mul.f32 %f492, %f480, 0f3E24A490;
mul.f32 %f493, %f481, 0f3E24A490;
fma.rn.f32 %f494, %f20, %f491, %f11;
fma.rn.f32 %f495, %f20, %f492, %f12;
fma.rn.f32 %f496, %f20, %f493, %f13;
mul.f32 %f497, %f484, 0fBECECEDE;
mul.f32 %f498, %f487, 0fBECECEDE;
mul.f32 %f499, %f490, 0fBECECEDE;
fma.rn.f32 %f82, %f20, %f497, %f494;
fma.rn.f32 %f83, %f20, %f498, %f495;
fma.rn.f32 %f84, %f20, %f499, %f496;
sub.f32 %f500, %f82, %f11;
sub.f32 %f501, %f83, %f12;
sub.f32 %f502, %f84, %f13;
mul.f32 %f503, %f501, %f501;
fma.rn.f32 %f504, %f500, %f500, %f503;
fma.rn.f32 %f505, %f502, %f502, %f504;
sqrt.rn.f32 %f426, %f505;
rcp.rn.f32 %f506, %f426;
mul.f32 %f422, %f506, %f500;
mul.f32 %f423, %f506, %f501;
mul.f32 %f424, %f506, %f502;
mov.u32 %r266, 1065353216;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r258, [root];
mov.u32 %r259, 1;
// inline asm
call _rt_trace_64, (%r258, %f11, %f12, %f13, %f422, %f423, %f424, %r259, %f425, %f426, %rd77, %r134);
// inline asm
ld.local.f32 %f507, [%rd16];
mov.u16 %rs98, 0;
setp.neu.f32 %p68, %f507, 0f3F800000;
@%p68 bra BB0_59;
mov.u16 %rs98, 0;
ld.local.f32 %f508, [%rd16+4];
setp.neu.f32 %p69, %f508, 0f3F800000;
@%p69 bra BB0_59;
mov.u16 %rs98, 0;
ld.local.f32 %f509, [%rd16+8];
setp.neu.f32 %p70, %f509, 0f3F800000;
@%p70 bra BB0_59;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r267, [root];
mov.f32 %f517, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r267, %f82, %f83, %f84, %f78, %f79, %f80, %r259, %f425, %f517, %rd77, %r134);
// inline asm
ld.local.f32 %f518, [%rd16];
add.f32 %f40, %f40, %f518;
ld.local.f32 %f519, [%rd16+4];
add.f32 %f39, %f39, %f519;
ld.local.f32 %f520, [%rd16+8];
add.f32 %f38, %f38, %f520;
add.f32 %f37, %f37, 0f3F800000;
mov.u16 %rs98, 1;
BB0_59:
mad.lo.s32 %r274, %r84, 1664525, 1013904223;
and.b32 %r275, %r274, 16777215;
cvt.rn.f32.u32 %f529, %r275;
mul.f32 %f530, %f529, 0f33800000;
fma.rn.f32 %f531, %f530, 0f40000000, 0fBF800000;
mad.lo.s32 %r85, %r274, 1664525, 1013904223;
and.b32 %r276, %r85, 16777215;
cvt.rn.f32.u32 %f532, %r276;
mul.f32 %f533, %f532, 0f33800000;
fma.rn.f32 %f534, %f533, 0f40000000, 0fBF800000;
mul.f32 %f535, %f14, %f534;
mul.f32 %f536, %f15, %f534;
mul.f32 %f537, %f16, %f534;
fma.rn.f32 %f538, %f17, %f531, %f535;
fma.rn.f32 %f539, %f18, %f531, %f536;
fma.rn.f32 %f540, %f19, %f531, %f537;
mul.f32 %f541, %f539, %f539;
fma.rn.f32 %f542, %f538, %f538, %f541;
fma.rn.f32 %f543, %f540, %f540, %f542;
sqrt.rn.f32 %f544, %f543;
rcp.rn.f32 %f545, %f544;
mul.f32 %f546, %f538, %f545;
mul.f32 %f547, %f539, %f545;
mul.f32 %f548, %f540, %f545;
mul.f32 %f549, %f1387, %f548;
mul.f32 %f550, %f1388, %f547;
sub.f32 %f551, %f549, %f550;
mul.f32 %f552, %f1388, %f546;
mul.f32 %f553, %f1386, %f548;
sub.f32 %f554, %f552, %f553;
mul.f32 %f555, %f1386, %f547;
mul.f32 %f556, %f1387, %f546;
sub.f32 %f557, %f555, %f556;
mul.f32 %f558, %f546, 0fBE24A490;
mul.f32 %f559, %f547, 0fBE24A490;
mul.f32 %f560, %f548, 0fBE24A490;
fma.rn.f32 %f561, %f20, %f558, %f11;
fma.rn.f32 %f562, %f20, %f559, %f12;
fma.rn.f32 %f563, %f20, %f560, %f13;
mul.f32 %f564, %f551, 0f3ECACABC;
mul.f32 %f565, %f554, 0f3ECACABC;
mul.f32 %f566, %f557, 0f3ECACABC;
fma.rn.f32 %f93, %f20, %f564, %f561;
fma.rn.f32 %f94, %f20, %f565, %f562;
fma.rn.f32 %f95, %f20, %f566, %f563;
sub.f32 %f567, %f93, %f11;
sub.f32 %f568, %f94, %f12;
sub.f32 %f569, %f95, %f13;
mul.f32 %f570, %f568, %f568;
fma.rn.f32 %f571, %f567, %f567, %f570;
fma.rn.f32 %f572, %f569, %f569, %f571;
sqrt.rn.f32 %f528, %f572;
rcp.rn.f32 %f573, %f528;
mul.f32 %f524, %f573, %f567;
mul.f32 %f525, %f573, %f568;
mul.f32 %f526, %f573, %f569;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r271, [root];
// inline asm
call _rt_trace_64, (%r271, %f11, %f12, %f13, %f524, %f525, %f526, %r259, %f425, %f528, %rd77, %r134);
// inline asm
ld.local.f32 %f574, [%rd16];
setp.neu.f32 %p71, %f574, 0f3F800000;
@%p71 bra BB0_63;
ld.local.f32 %f575, [%rd16+4];
setp.neu.f32 %p72, %f575, 0f3F800000;
@%p72 bra BB0_63;
ld.local.f32 %f576, [%rd16+8];
setp.neu.f32 %p73, %f576, 0f3F800000;
@%p73 bra BB0_63;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r278, [root];
mov.f32 %f584, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r278, %f93, %f94, %f95, %f78, %f79, %f80, %r259, %f425, %f584, %rd77, %r134);
// inline asm
ld.local.f32 %f585, [%rd16];
add.f32 %f40, %f40, %f585;
ld.local.f32 %f586, [%rd16+4];
add.f32 %f39, %f39, %f586;
ld.local.f32 %f587, [%rd16+8];
add.f32 %f38, %f38, %f587;
add.f32 %f37, %f37, 0f3F800000;
mov.u16 %rs98, 1;
BB0_63:
mad.lo.s32 %r285, %r85, 1664525, 1013904223;
and.b32 %r286, %r285, 16777215;
cvt.rn.f32.u32 %f596, %r286;
mul.f32 %f597, %f596, 0f33800000;
fma.rn.f32 %f598, %f597, 0f40000000, 0fBF800000;
mad.lo.s32 %r86, %r285, 1664525, 1013904223;
and.b32 %r287, %r86, 16777215;
cvt.rn.f32.u32 %f599, %r287;
mul.f32 %f600, %f599, 0f33800000;
fma.rn.f32 %f601, %f600, 0f40000000, 0fBF800000;
mul.f32 %f602, %f14, %f601;
mul.f32 %f603, %f15, %f601;
mul.f32 %f604, %f16, %f601;
fma.rn.f32 %f605, %f17, %f598, %f602;
fma.rn.f32 %f606, %f18, %f598, %f603;
fma.rn.f32 %f607, %f19, %f598, %f604;
mul.f32 %f608, %f606, %f606;
fma.rn.f32 %f609, %f605, %f605, %f608;
fma.rn.f32 %f610, %f607, %f607, %f609;
sqrt.rn.f32 %f611, %f610;
rcp.rn.f32 %f612, %f611;
mul.f32 %f613, %f605, %f612;
mul.f32 %f614, %f606, %f612;
mul.f32 %f615, %f607, %f612;
mul.f32 %f616, %f1387, %f615;
mul.f32 %f617, %f1388, %f614;
sub.f32 %f618, %f616, %f617;
mul.f32 %f619, %f1388, %f613;
mul.f32 %f620, %f1386, %f615;
sub.f32 %f621, %f619, %f620;
mul.f32 %f622, %f1386, %f614;
mul.f32 %f623, %f1387, %f613;
sub.f32 %f624, %f622, %f623;
mul.f32 %f625, %f613, 0f3F23A3A8;
mul.f32 %f626, %f614, 0f3F23A3A8;
mul.f32 %f627, %f615, 0f3F23A3A8;
fma.rn.f32 %f628, %f20, %f625, %f11;
fma.rn.f32 %f629, %f20, %f626, %f12;
fma.rn.f32 %f630, %f20, %f627, %f13;
mul.f32 %f631, %f618, 0f3E24A490;
mul.f32 %f632, %f621, 0f3E24A490;
mul.f32 %f633, %f624, 0f3E24A490;
fma.rn.f32 %f104, %f20, %f631, %f628;
fma.rn.f32 %f105, %f20, %f632, %f629;
fma.rn.f32 %f106, %f20, %f633, %f630;
sub.f32 %f634, %f104, %f11;
sub.f32 %f635, %f105, %f12;
sub.f32 %f636, %f106, %f13;
mul.f32 %f637, %f635, %f635;
fma.rn.f32 %f638, %f634, %f634, %f637;
fma.rn.f32 %f639, %f636, %f636, %f638;
sqrt.rn.f32 %f595, %f639;
rcp.rn.f32 %f640, %f595;
mul.f32 %f591, %f640, %f634;
mul.f32 %f592, %f640, %f635;
mul.f32 %f593, %f640, %f636;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r282, [root];
// inline asm
call _rt_trace_64, (%r282, %f11, %f12, %f13, %f591, %f592, %f593, %r259, %f425, %f595, %rd77, %r134);
// inline asm
ld.local.f32 %f641, [%rd16];
setp.neu.f32 %p74, %f641, 0f3F800000;
@%p74 bra BB0_67;
ld.local.f32 %f642, [%rd16+4];
setp.neu.f32 %p75, %f642, 0f3F800000;
@%p75 bra BB0_67;
ld.local.f32 %f643, [%rd16+8];
setp.neu.f32 %p76, %f643, 0f3F800000;
@%p76 bra BB0_67;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r289, [root];
mov.f32 %f651, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r289, %f104, %f105, %f106, %f78, %f79, %f80, %r259, %f425, %f651, %rd77, %r134);
// inline asm
ld.local.f32 %f652, [%rd16];
add.f32 %f40, %f40, %f652;
ld.local.f32 %f653, [%rd16+4];
add.f32 %f39, %f39, %f653;
ld.local.f32 %f654, [%rd16+8];
add.f32 %f38, %f38, %f654;
add.f32 %f37, %f37, 0f3F800000;
mov.u16 %rs98, 1;
BB0_67:
mad.lo.s32 %r296, %r86, 1664525, 1013904223;
and.b32 %r297, %r296, 16777215;
cvt.rn.f32.u32 %f663, %r297;
mul.f32 %f664, %f663, 0f33800000;
fma.rn.f32 %f665, %f664, 0f40000000, 0fBF800000;
mad.lo.s32 %r87, %r296, 1664525, 1013904223;
and.b32 %r298, %r87, 16777215;
cvt.rn.f32.u32 %f666, %r298;
mul.f32 %f667, %f666, 0f33800000;
fma.rn.f32 %f668, %f667, 0f40000000, 0fBF800000;
mul.f32 %f669, %f14, %f668;
mul.f32 %f670, %f15, %f668;
mul.f32 %f671, %f16, %f668;
fma.rn.f32 %f672, %f17, %f665, %f669;
fma.rn.f32 %f673, %f18, %f665, %f670;
fma.rn.f32 %f674, %f19, %f665, %f671;
mul.f32 %f675, %f673, %f673;
fma.rn.f32 %f676, %f672, %f672, %f675;
fma.rn.f32 %f677, %f674, %f674, %f676;
sqrt.rn.f32 %f678, %f677;
rcp.rn.f32 %f679, %f678;
mul.f32 %f680, %f672, %f679;
mul.f32 %f681, %f673, %f679;
mul.f32 %f682, %f674, %f679;
mul.f32 %f683, %f1387, %f682;
mul.f32 %f684, %f1388, %f681;
sub.f32 %f685, %f683, %f684;
mul.f32 %f686, %f1388, %f680;
mul.f32 %f687, %f1386, %f682;
sub.f32 %f688, %f686, %f687;
mul.f32 %f689, %f1386, %f681;
mul.f32 %f690, %f1387, %f680;
sub.f32 %f691, %f689, %f690;
mul.f32 %f692, %f680, 0fBECECEDE;
mul.f32 %f693, %f681, 0fBECECEDE;
mul.f32 %f694, %f682, 0fBECECEDE;
fma.rn.f32 %f695, %f20, %f692, %f11;
fma.rn.f32 %f696, %f20, %f693, %f12;
fma.rn.f32 %f697, %f20, %f694, %f13;
mul.f32 %f698, %f685, 0fBF23A3A9;
mul.f32 %f699, %f688, 0fBF23A3A9;
mul.f32 %f700, %f691, 0fBF23A3A9;
fma.rn.f32 %f115, %f20, %f698, %f695;
fma.rn.f32 %f116, %f20, %f699, %f696;
fma.rn.f32 %f117, %f20, %f700, %f697;
sub.f32 %f701, %f115, %f11;
sub.f32 %f702, %f116, %f12;
sub.f32 %f703, %f117, %f13;
mul.f32 %f704, %f702, %f702;
fma.rn.f32 %f705, %f701, %f701, %f704;
fma.rn.f32 %f706, %f703, %f703, %f705;
sqrt.rn.f32 %f662, %f706;
rcp.rn.f32 %f707, %f662;
mul.f32 %f658, %f707, %f701;
mul.f32 %f659, %f707, %f702;
mul.f32 %f660, %f707, %f703;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r293, [root];
// inline asm
call _rt_trace_64, (%r293, %f11, %f12, %f13, %f658, %f659, %f660, %r259, %f425, %f662, %rd77, %r134);
// inline asm
ld.local.f32 %f708, [%rd16];
setp.neu.f32 %p77, %f708, 0f3F800000;
@%p77 bra BB0_71;
ld.local.f32 %f709, [%rd16+4];
setp.neu.f32 %p78, %f709, 0f3F800000;
@%p78 bra BB0_71;
ld.local.f32 %f710, [%rd16+8];
setp.neu.f32 %p79, %f710, 0f3F800000;
@%p79 bra BB0_71;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r300, [root];
mov.f32 %f718, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r300, %f115, %f116, %f117, %f78, %f79, %f80, %r259, %f425, %f718, %rd77, %r134);
// inline asm
ld.local.f32 %f719, [%rd16];
add.f32 %f40, %f40, %f719;
ld.local.f32 %f720, [%rd16+4];
add.f32 %f39, %f39, %f720;
ld.local.f32 %f721, [%rd16+8];
add.f32 %f38, %f38, %f721;
add.f32 %f37, %f37, 0f3F800000;
mov.u16 %rs98, 1;
BB0_71:
mad.lo.s32 %r307, %r87, 1664525, 1013904223;
and.b32 %r308, %r307, 16777215;
cvt.rn.f32.u32 %f730, %r308;
mul.f32 %f731, %f730, 0f33800000;
fma.rn.f32 %f732, %f731, 0f40000000, 0fBF800000;
mad.lo.s32 %r88, %r307, 1664525, 1013904223;
and.b32 %r309, %r88, 16777215;
cvt.rn.f32.u32 %f733, %r309;
mul.f32 %f734, %f733, 0f33800000;
fma.rn.f32 %f735, %f734, 0f40000000, 0fBF800000;
mul.f32 %f736, %f14, %f735;
mul.f32 %f737, %f15, %f735;
mul.f32 %f738, %f16, %f735;
fma.rn.f32 %f739, %f17, %f732, %f736;
fma.rn.f32 %f740, %f18, %f732, %f737;
fma.rn.f32 %f741, %f19, %f732, %f738;
mul.f32 %f742, %f740, %f740;
fma.rn.f32 %f743, %f739, %f739, %f742;
fma.rn.f32 %f744, %f741, %f741, %f743;
sqrt.rn.f32 %f745, %f744;
rcp.rn.f32 %f746, %f745;
mul.f32 %f747, %f739, %f746;
mul.f32 %f748, %f740, %f746;
mul.f32 %f749, %f741, %f746;
mul.f32 %f750, %f1387, %f749;
mul.f32 %f751, %f1388, %f748;
sub.f32 %f752, %f750, %f751;
mul.f32 %f753, %f1388, %f747;
mul.f32 %f754, %f1386, %f749;
sub.f32 %f755, %f753, %f754;
mul.f32 %f756, %f1386, %f748;
mul.f32 %f757, %f1387, %f747;
sub.f32 %f758, %f756, %f757;
mul.f32 %f759, %f747, 0fBF23A3A9;
mul.f32 %f760, %f748, 0fBF23A3A9;
mul.f32 %f761, %f749, 0fBF23A3A9;
fma.rn.f32 %f762, %f20, %f759, %f11;
fma.rn.f32 %f763, %f20, %f760, %f12;
fma.rn.f32 %f764, %f20, %f761, %f13;
mul.f32 %f765, %f752, 0f3F23A3A8;
mul.f32 %f766, %f755, 0f3F23A3A8;
mul.f32 %f767, %f758, 0f3F23A3A8;
fma.rn.f32 %f126, %f20, %f765, %f762;
fma.rn.f32 %f127, %f20, %f766, %f763;
fma.rn.f32 %f128, %f20, %f767, %f764;
sub.f32 %f768, %f126, %f11;
sub.f32 %f769, %f127, %f12;
sub.f32 %f770, %f128, %f13;
mul.f32 %f771, %f769, %f769;
fma.rn.f32 %f772, %f768, %f768, %f771;
fma.rn.f32 %f773, %f770, %f770, %f772;
sqrt.rn.f32 %f729, %f773;
rcp.rn.f32 %f774, %f729;
mul.f32 %f725, %f774, %f768;
mul.f32 %f726, %f774, %f769;
mul.f32 %f727, %f774, %f770;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r304, [root];
// inline asm
call _rt_trace_64, (%r304, %f11, %f12, %f13, %f725, %f726, %f727, %r259, %f425, %f729, %rd77, %r134);
// inline asm
ld.local.f32 %f775, [%rd16];
setp.neu.f32 %p80, %f775, 0f3F800000;
@%p80 bra BB0_75;
ld.local.f32 %f776, [%rd16+4];
setp.neu.f32 %p81, %f776, 0f3F800000;
@%p81 bra BB0_75;
ld.local.f32 %f777, [%rd16+8];
setp.neu.f32 %p82, %f777, 0f3F800000;
@%p82 bra BB0_75;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r311, [root];
mov.f32 %f785, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r311, %f126, %f127, %f128, %f78, %f79, %f80, %r259, %f425, %f785, %rd77, %r134);
// inline asm
ld.local.f32 %f786, [%rd16];
add.f32 %f40, %f40, %f786;
ld.local.f32 %f787, [%rd16+4];
add.f32 %f39, %f39, %f787;
ld.local.f32 %f788, [%rd16+8];
add.f32 %f38, %f38, %f788;
add.f32 %f37, %f37, 0f3F800000;
mov.u16 %rs98, 1;
BB0_75:
mad.lo.s32 %r318, %r88, 1664525, 1013904223;
and.b32 %r319, %r318, 16777215;
cvt.rn.f32.u32 %f797, %r319;
mul.f32 %f798, %f797, 0f33800000;
fma.rn.f32 %f799, %f798, 0f40000000, 0fBF800000;
mad.lo.s32 %r89, %r318, 1664525, 1013904223;
and.b32 %r320, %r89, 16777215;
cvt.rn.f32.u32 %f800, %r320;
mul.f32 %f801, %f800, 0f33800000;
fma.rn.f32 %f802, %f801, 0f40000000, 0fBF800000;
mul.f32 %f803, %f14, %f802;
mul.f32 %f804, %f15, %f802;
mul.f32 %f805, %f16, %f802;
fma.rn.f32 %f806, %f17, %f799, %f803;
fma.rn.f32 %f807, %f18, %f799, %f804;
fma.rn.f32 %f808, %f19, %f799, %f805;
mul.f32 %f809, %f807, %f807;
fma.rn.f32 %f810, %f806, %f806, %f809;
fma.rn.f32 %f811, %f808, %f808, %f810;
sqrt.rn.f32 %f812, %f811;
rcp.rn.f32 %f813, %f812;
mul.f32 %f814, %f806, %f813;
mul.f32 %f815, %f807, %f813;
mul.f32 %f816, %f808, %f813;
mul.f32 %f817, %f1387, %f816;
mul.f32 %f818, %f1388, %f815;
sub.f32 %f819, %f817, %f818;
mul.f32 %f820, %f1388, %f814;
mul.f32 %f821, %f1386, %f816;
sub.f32 %f822, %f820, %f821;
mul.f32 %f823, %f1386, %f815;
mul.f32 %f824, %f1387, %f814;
sub.f32 %f825, %f823, %f824;
mul.f32 %f826, %f814, 0fBF61E1E3;
mul.f32 %f827, %f815, 0fBF61E1E3;
mul.f32 %f828, %f816, 0fBF61E1E3;
fma.rn.f32 %f829, %f20, %f826, %f11;
fma.rn.f32 %f830, %f20, %f827, %f12;
fma.rn.f32 %f831, %f20, %f828, %f13;
mul.f32 %f832, %f819, 0fBE24A490;
mul.f32 %f833, %f822, 0fBE24A490;
mul.f32 %f834, %f825, 0fBE24A490;
fma.rn.f32 %f137, %f20, %f832, %f829;
fma.rn.f32 %f138, %f20, %f833, %f830;
fma.rn.f32 %f139, %f20, %f834, %f831;
sub.f32 %f835, %f137, %f11;
sub.f32 %f836, %f138, %f12;
sub.f32 %f837, %f139, %f13;
mul.f32 %f838, %f836, %f836;
fma.rn.f32 %f839, %f835, %f835, %f838;
fma.rn.f32 %f840, %f837, %f837, %f839;
sqrt.rn.f32 %f796, %f840;
rcp.rn.f32 %f841, %f796;
mul.f32 %f792, %f841, %f835;
mul.f32 %f793, %f841, %f836;
mul.f32 %f794, %f841, %f837;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r315, [root];
// inline asm
call _rt_trace_64, (%r315, %f11, %f12, %f13, %f792, %f793, %f794, %r259, %f425, %f796, %rd77, %r134);
// inline asm
ld.local.f32 %f842, [%rd16];
setp.neu.f32 %p83, %f842, 0f3F800000;
@%p83 bra BB0_79;
ld.local.f32 %f843, [%rd16+4];
setp.neu.f32 %p84, %f843, 0f3F800000;
@%p84 bra BB0_79;
ld.local.f32 %f844, [%rd16+8];
setp.neu.f32 %p85, %f844, 0f3F800000;
@%p85 bra BB0_79;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r322, [root];
mov.f32 %f852, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r322, %f137, %f138, %f139, %f78, %f79, %f80, %r259, %f425, %f852, %rd77, %r134);
// inline asm
ld.local.f32 %f853, [%rd16];
add.f32 %f40, %f40, %f853;
ld.local.f32 %f854, [%rd16+4];
add.f32 %f39, %f39, %f854;
ld.local.f32 %f855, [%rd16+8];
add.f32 %f38, %f38, %f855;
add.f32 %f37, %f37, 0f3F800000;
mov.u16 %rs98, 1;
BB0_79:
mad.lo.s32 %r329, %r89, 1664525, 1013904223;
and.b32 %r330, %r329, 16777215;
cvt.rn.f32.u32 %f864, %r330;
mul.f32 %f865, %f864, 0f33800000;
fma.rn.f32 %f866, %f865, 0f40000000, 0fBF800000;
mad.lo.s32 %r90, %r329, 1664525, 1013904223;
and.b32 %r331, %r90, 16777215;
cvt.rn.f32.u32 %f867, %r331;
mul.f32 %f868, %f867, 0f33800000;
fma.rn.f32 %f869, %f868, 0f40000000, 0fBF800000;
mul.f32 %f870, %f14, %f869;
mul.f32 %f871, %f15, %f869;
mul.f32 %f872, %f16, %f869;
fma.rn.f32 %f873, %f17, %f866, %f870;
fma.rn.f32 %f874, %f18, %f866, %f871;
fma.rn.f32 %f875, %f19, %f866, %f872;
mul.f32 %f876, %f874, %f874;
fma.rn.f32 %f877, %f873, %f873, %f876;
fma.rn.f32 %f878, %f875, %f875, %f877;
sqrt.rn.f32 %f879, %f878;
rcp.rn.f32 %f880, %f879;
mul.f32 %f881, %f873, %f880;
mul.f32 %f882, %f874, %f880;
mul.f32 %f883, %f875, %f880;
mul.f32 %f884, %f1387, %f883;
mul.f32 %f885, %f1388, %f882;
sub.f32 %f886, %f884, %f885;
mul.f32 %f887, %f1388, %f881;
mul.f32 %f888, %f1386, %f883;
sub.f32 %f889, %f887, %f888;
mul.f32 %f890, %f1386, %f882;
mul.f32 %f891, %f1387, %f881;
sub.f32 %f892, %f890, %f891;
mul.f32 %f893, %f881, 0f3ECACABC;
mul.f32 %f894, %f882, 0f3ECACABC;
mul.f32 %f895, %f883, 0f3ECACABC;
fma.rn.f32 %f896, %f20, %f893, %f11;
fma.rn.f32 %f897, %f20, %f894, %f12;
fma.rn.f32 %f898, %f20, %f895, %f13;
mul.f32 %f899, %f886, 0f3F61E1D2;
mul.f32 %f900, %f889, 0f3F61E1D2;
mul.f32 %f901, %f892, 0f3F61E1D2;
fma.rn.f32 %f148, %f20, %f899, %f896;
fma.rn.f32 %f149, %f20, %f900, %f897;
fma.rn.f32 %f150, %f20, %f901, %f898;
sub.f32 %f902, %f148, %f11;
sub.f32 %f903, %f149, %f12;
sub.f32 %f904, %f150, %f13;
mul.f32 %f905, %f903, %f903;
fma.rn.f32 %f906, %f902, %f902, %f905;
fma.rn.f32 %f907, %f904, %f904, %f906;
sqrt.rn.f32 %f863, %f907;
rcp.rn.f32 %f908, %f863;
mul.f32 %f859, %f908, %f902;
mul.f32 %f860, %f908, %f903;
mul.f32 %f861, %f908, %f904;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r326, [root];
// inline asm
call _rt_trace_64, (%r326, %f11, %f12, %f13, %f859, %f860, %f861, %r259, %f425, %f863, %rd77, %r134);
// inline asm
ld.local.f32 %f909, [%rd16];
setp.neu.f32 %p86, %f909, 0f3F800000;
@%p86 bra BB0_83;
ld.local.f32 %f910, [%rd16+4];
setp.neu.f32 %p87, %f910, 0f3F800000;
@%p87 bra BB0_83;
ld.local.f32 %f911, [%rd16+8];
setp.neu.f32 %p88, %f911, 0f3F800000;
@%p88 bra BB0_83;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r333, [root];
mov.f32 %f919, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r333, %f148, %f149, %f150, %f78, %f79, %f80, %r259, %f425, %f919, %rd77, %r134);
// inline asm
ld.local.f32 %f920, [%rd16];
add.f32 %f40, %f40, %f920;
ld.local.f32 %f921, [%rd16+4];
add.f32 %f39, %f39, %f921;
ld.local.f32 %f922, [%rd16+8];
add.f32 %f38, %f38, %f922;
add.f32 %f37, %f37, 0f3F800000;
mov.u16 %rs98, 1;
BB0_83:
mad.lo.s32 %r340, %r90, 1664525, 1013904223;
and.b32 %r341, %r340, 16777215;
cvt.rn.f32.u32 %f931, %r341;
mul.f32 %f932, %f931, 0f33800000;
fma.rn.f32 %f933, %f932, 0f40000000, 0fBF800000;
mad.lo.s32 %r479, %r340, 1664525, 1013904223;
and.b32 %r342, %r479, 16777215;
cvt.rn.f32.u32 %f934, %r342;
mul.f32 %f935, %f934, 0f33800000;
fma.rn.f32 %f936, %f935, 0f40000000, 0fBF800000;
mul.f32 %f937, %f14, %f936;
mul.f32 %f938, %f15, %f936;
mul.f32 %f939, %f16, %f936;
fma.rn.f32 %f940, %f17, %f933, %f937;
fma.rn.f32 %f941, %f18, %f933, %f938;
fma.rn.f32 %f942, %f19, %f933, %f939;
mul.f32 %f943, %f941, %f941;
fma.rn.f32 %f944, %f940, %f940, %f943;
fma.rn.f32 %f945, %f942, %f942, %f944;
sqrt.rn.f32 %f946, %f945;
rcp.rn.f32 %f947, %f946;
mul.f32 %f948, %f940, %f947;
mul.f32 %f949, %f941, %f947;
mul.f32 %f950, %f942, %f947;
mul.f32 %f951, %f1387, %f950;
mul.f32 %f952, %f1388, %f949;
sub.f32 %f953, %f951, %f952;
mul.f32 %f954, %f1388, %f948;
mul.f32 %f955, %f1386, %f950;
sub.f32 %f956, %f954, %f955;
mul.f32 %f957, %f1386, %f949;
mul.f32 %f958, %f1387, %f948;
sub.f32 %f959, %f957, %f958;
mul.f32 %f960, %f948, 0f3F61E1D2;
mul.f32 %f961, %f949, 0f3F61E1D2;
mul.f32 %f962, %f950, 0f3F61E1D2;
fma.rn.f32 %f963, %f20, %f960, %f11;
fma.rn.f32 %f964, %f20, %f961, %f12;
fma.rn.f32 %f965, %f20, %f962, %f13;
mul.f32 %f966, %f953, 0fBF61E1E3;
mul.f32 %f967, %f956, 0fBF61E1E3;
mul.f32 %f968, %f959, 0fBF61E1E3;
fma.rn.f32 %f159, %f20, %f966, %f963;
fma.rn.f32 %f160, %f20, %f967, %f964;
fma.rn.f32 %f161, %f20, %f968, %f965;
sub.f32 %f969, %f159, %f11;
sub.f32 %f970, %f160, %f12;
sub.f32 %f971, %f161, %f13;
mul.f32 %f972, %f970, %f970;
fma.rn.f32 %f973, %f969, %f969, %f972;
fma.rn.f32 %f974, %f971, %f971, %f973;
sqrt.rn.f32 %f930, %f974;
rcp.rn.f32 %f975, %f930;
mul.f32 %f926, %f975, %f969;
mul.f32 %f927, %f975, %f970;
mul.f32 %f928, %f975, %f971;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r337, [root];
// inline asm
call _rt_trace_64, (%r337, %f11, %f12, %f13, %f926, %f927, %f928, %r259, %f425, %f930, %rd77, %r134);
// inline asm
ld.local.f32 %f976, [%rd16];
setp.neu.f32 %p89, %f976, 0f3F800000;
@%p89 bra BB0_87;
ld.local.f32 %f977, [%rd16+4];
setp.neu.f32 %p90, %f977, 0f3F800000;
@%p90 bra BB0_87;
ld.local.f32 %f978, [%rd16+8];
setp.neu.f32 %p91, %f978, 0f3F800000;
@%p91 bra BB0_87;
st.local.u32 [%rd16+8], %r266;
st.local.u32 [%rd16+4], %r266;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r344, [root];
mov.f32 %f986, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r344, %f159, %f160, %f161, %f78, %f79, %f80, %r259, %f425, %f986, %rd77, %r134);
// inline asm
ld.local.f32 %f987, [%rd16];
add.f32 %f40, %f40, %f987;
ld.local.f32 %f988, [%rd16+4];
add.f32 %f39, %f39, %f988;
ld.local.f32 %f989, [%rd16+8];
add.f32 %f38, %f38, %f989;
add.f32 %f37, %f37, 0f3F800000;
mov.u16 %rs98, 1;
BB0_87:
setp.ne.s16 %p92, %rs98, 0;
@%p92 bra BB0_89;
add.u64 %rd93, %SP, 40;
cvta.to.local.u64 %rd94, %rd93;
st.local.u32 [%rd94+8], %r266;
st.local.u32 [%rd94+4], %r266;
st.local.u32 [%rd94], %r266;
ld.global.u32 %r348, [root];
mov.f32 %f997, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r348, %f11, %f12, %f13, %f78, %f79, %f80, %r259, %f425, %f997, %rd93, %r134);
// inline asm
ld.local.f32 %f998, [%rd94];
add.f32 %f40, %f40, %f998;
ld.local.f32 %f999, [%rd94+4];
add.f32 %f39, %f39, %f999;
ld.local.f32 %f1000, [%rd94+8];
add.f32 %f38, %f38, %f1000;
add.f32 %f37, %f37, 0f3F800000;
BB0_89:
ld.global.u32 %r453, [samples];
add.s32 %r456, %r456, 1;
setp.lt.s32 %p93, %r456, %r453;
@%p93 bra BB0_9;
BB0_90:
add.s32 %r454, %r454, 1;
setp.lt.s32 %p94, %r454, %r453;
@%p94 bra BB0_7;
bra.uni BB0_91;
BB0_164:
mov.u64 %rd210, image_HDR;
cvta.global.u64 %rd205, %rd210;
mov.u32 %r443, 8;
mov.u64 %rd209, 0;
// inline asm
call (%rd204), _rt_buffer_get_64, (%rd205, %r112, %r443, %rd22, %rd23, %rd209, %rd209);
// inline asm
mov.f32 %f1383, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs90, %f1383;}
// inline asm
mov.u16 %rs91, 0;
st.v4.u16 [%rd204], {%rs90, %rs90, %rs90, %rs91};
BB0_165:
ld.global.u8 %rs92, [imageEnabled];
and.b16 %rs93, %rs92, 64;
setp.eq.s16 %p175, %rs93, 0;
@%p175 bra BB0_167;
cvt.u64.u32 %rd213, %r2;
cvt.u64.u32 %rd214, %r3;
mov.u64 %rd217, image_Dir;
cvta.global.u64 %rd212, %rd217;
mov.u64 %rd216, 0;
// inline asm
call (%rd211), _rt_buffer_get_64, (%rd212, %r112, %r113, %rd213, %rd214, %rd216, %rd216);
// inline asm
mov.u16 %rs94, 0;
st.v4.u8 [%rd211], {%rs94, %rs94, %rs94, %rs94};
bra.uni BB0_167;
BB0_5:
mov.f32 %f38, %f37;
mov.f32 %f39, %f37;
mov.f32 %f40, %f37;
BB0_91:
setp.eq.s32 %p95, %r453, 0;
mov.f32 %f1453, 0f3F800000;
mov.f32 %f1454, %f1453;
mov.f32 %f1455, %f1453;
@%p95 bra BB0_93;
mov.f32 %f1004, 0f3F800000;
max.f32 %f1005, %f37, %f1004;
rcp.rn.f32 %f1006, %f1005;
mul.f32 %f1453, %f40, %f1006;
mul.f32 %f1454, %f39, %f1006;
mul.f32 %f1455, %f38, %f1006;
BB0_93:
ld.global.f32 %f1007, [directColor];
mul.f32 %f1008, %f1453, %f1007;
ld.global.f32 %f1009, [directColor+4];
mul.f32 %f1010, %f1454, %f1009;
ld.global.f32 %f1011, [directColor+8];
mul.f32 %f1012, %f1455, %f1011;
cvt.sat.f32.f32 %f1013, %f10;
mul.f32 %f1014, %f1008, %f1013;
mul.f32 %f1015, %f1010, %f1013;
mul.f32 %f1016, %f1012, %f1013;
ld.global.u32 %r352, [ignoreNormal];
setp.eq.s32 %p96, %r352, 0;
selp.f32 %f192, %f1014, %f1008, %p96;
selp.f32 %f193, %f1015, %f1010, %p96;
selp.f32 %f194, %f1016, %f1012, %p96;
ld.global.u32 %r482, [imageEnabled];
and.b32 %r353, %r482, 8;
setp.eq.s32 %p97, %r353, 0;
@%p97 bra BB0_106;
mov.u32 %r446, 2;
cvt.u64.u32 %rd97, %r2;
cvt.u64.u32 %rd98, %r3;
mov.u64 %rd101, image_Mask;
cvta.global.u64 %rd96, %rd101;
// inline asm
call (%rd95), _rt_buffer_get_64, (%rd96, %r446, %r446, %rd97, %rd98, %rd29, %rd29);
// inline asm
mov.f32 %f1019, 0f3E68BA2E;
cvt.rzi.f32.f32 %f1020, %f1019;
fma.rn.f32 %f1021, %f1020, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f195, %f1021;
abs.f32 %f196, %f1453;
setp.lt.f32 %p98, %f196, 0f00800000;
mul.f32 %f1022, %f196, 0f4B800000;
selp.f32 %f1023, 0fC3170000, 0fC2FE0000, %p98;
selp.f32 %f1024, %f1022, %f196, %p98;
mov.b32 %r356, %f1024;
and.b32 %r357, %r356, 8388607;
or.b32 %r358, %r357, 1065353216;
mov.b32 %f1025, %r358;
shr.u32 %r359, %r356, 23;
cvt.rn.f32.u32 %f1026, %r359;
add.f32 %f1027, %f1023, %f1026;
setp.gt.f32 %p99, %f1025, 0f3FB504F3;
mul.f32 %f1028, %f1025, 0f3F000000;
add.f32 %f1029, %f1027, 0f3F800000;
selp.f32 %f1030, %f1028, %f1025, %p99;
selp.f32 %f1031, %f1029, %f1027, %p99;
add.f32 %f1032, %f1030, 0fBF800000;
add.f32 %f1018, %f1030, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1017,%f1018;
// inline asm
add.f32 %f1033, %f1032, %f1032;
mul.f32 %f1034, %f1017, %f1033;
mul.f32 %f1035, %f1034, %f1034;
mov.f32 %f1036, 0f3C4CAF63;
mov.f32 %f1037, 0f3B18F0FE;
fma.rn.f32 %f1038, %f1037, %f1035, %f1036;
mov.f32 %f1039, 0f3DAAAABD;
fma.rn.f32 %f1040, %f1038, %f1035, %f1039;
mul.rn.f32 %f1041, %f1040, %f1035;
mul.rn.f32 %f1042, %f1041, %f1034;
sub.f32 %f1043, %f1032, %f1034;
neg.f32 %f1044, %f1034;
add.f32 %f1045, %f1043, %f1043;
fma.rn.f32 %f1046, %f1044, %f1032, %f1045;
mul.rn.f32 %f1047, %f1017, %f1046;
add.f32 %f1048, %f1042, %f1034;
sub.f32 %f1049, %f1034, %f1048;
add.f32 %f1050, %f1042, %f1049;
add.f32 %f1051, %f1047, %f1050;
add.f32 %f1052, %f1048, %f1051;
sub.f32 %f1053, %f1048, %f1052;
add.f32 %f1054, %f1051, %f1053;
mov.f32 %f1055, 0f3F317200;
mul.rn.f32 %f1056, %f1031, %f1055;
mov.f32 %f1057, 0f35BFBE8E;
mul.rn.f32 %f1058, %f1031, %f1057;
add.f32 %f1059, %f1056, %f1052;
sub.f32 %f1060, %f1056, %f1059;
add.f32 %f1061, %f1052, %f1060;
add.f32 %f1062, %f1054, %f1061;
add.f32 %f1063, %f1058, %f1062;
add.f32 %f1064, %f1059, %f1063;
sub.f32 %f1065, %f1059, %f1064;
add.f32 %f1066, %f1063, %f1065;
mov.f32 %f1067, 0f3EE8BA2E;
mul.rn.f32 %f1068, %f1067, %f1064;
neg.f32 %f1069, %f1068;
fma.rn.f32 %f1070, %f1067, %f1064, %f1069;
fma.rn.f32 %f1071, %f1067, %f1066, %f1070;
mov.f32 %f1072, 0f00000000;
fma.rn.f32 %f1073, %f1072, %f1064, %f1071;
add.rn.f32 %f1074, %f1068, %f1073;
neg.f32 %f1075, %f1074;
add.rn.f32 %f1076, %f1068, %f1075;
add.rn.f32 %f1077, %f1076, %f1073;
mov.b32 %r360, %f1074;
setp.eq.s32 %p100, %r360, 1118925336;
add.s32 %r361, %r360, -1;
mov.b32 %f1078, %r361;
add.f32 %f1079, %f1077, 0f37000000;
selp.f32 %f1080, %f1078, %f1074, %p100;
selp.f32 %f197, %f1079, %f1077, %p100;
mul.f32 %f1081, %f1080, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1082, %f1081;
mov.f32 %f1083, 0fBF317200;
fma.rn.f32 %f1084, %f1082, %f1083, %f1080;
mov.f32 %f1085, 0fB5BFBE8E;
fma.rn.f32 %f1086, %f1082, %f1085, %f1084;
mul.f32 %f1087, %f1086, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1088, %f1087;
add.f32 %f1089, %f1082, 0f00000000;
ex2.approx.f32 %f1090, %f1089;
mul.f32 %f1091, %f1088, %f1090;
setp.lt.f32 %p101, %f1080, 0fC2D20000;
selp.f32 %f1092, 0f00000000, %f1091, %p101;
setp.gt.f32 %p102, %f1080, 0f42D20000;
selp.f32 %f1456, 0f7F800000, %f1092, %p102;
setp.eq.f32 %p103, %f1456, 0f7F800000;
@%p103 bra BB0_96;
fma.rn.f32 %f1456, %f1456, %f197, %f1456;
BB0_96:
setp.lt.f32 %p104, %f1453, 0f00000000;
setp.eq.f32 %p105, %f195, 0f3F800000;
and.pred %p1, %p104, %p105;
mov.b32 %r362, %f1456;
xor.b32 %r363, %r362, -2147483648;
mov.b32 %f1093, %r363;
selp.f32 %f1458, %f1093, %f1456, %p1;
setp.eq.f32 %p106, %f1453, 0f00000000;
@%p106 bra BB0_99;
bra.uni BB0_97;
BB0_99:
add.f32 %f1096, %f1453, %f1453;
selp.f32 %f1458, %f1096, 0f00000000, %p105;
bra.uni BB0_100;
BB0_154:
mov.u64 %rd169, image_HDR;
cvta.global.u64 %rd164, %rd169;
mov.u32 %r426, 8;
mov.u64 %rd168, 0;
// inline asm
call (%rd163), _rt_buffer_get_64, (%rd164, %r112, %r426, %rd20, %rd21, %rd168, %rd168);
// inline asm
mov.f32 %f1373, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs71, %f1373;}
// inline asm
st.v4.u16 [%rd163], {%rs71, %rs71, %rs71, %rs60};
BB0_155:
ld.global.u8 %rs72, [imageEnabled];
and.b16 %rs73, %rs72, 64;
setp.eq.s16 %p170, %rs73, 0;
@%p170 bra BB0_167;
cvt.u64.u32 %rd172, %r2;
cvt.u64.u32 %rd173, %r3;
mov.u64 %rd176, image_Dir;
cvta.global.u64 %rd171, %rd176;
mov.u64 %rd175, 0;
// inline asm
call (%rd170), _rt_buffer_get_64, (%rd171, %r112, %r113, %rd172, %rd173, %rd175, %rd175);
// inline asm
mov.u16 %rs74, 255;
mov.u16 %rs75, 0;
st.v4.u8 [%rd170], {%rs75, %rs75, %rs75, %rs74};
bra.uni BB0_167;
BB0_97:
setp.geu.f32 %p107, %f1453, 0f00000000;
@%p107 bra BB0_100;
cvt.rzi.f32.f32 %f1095, %f1067;
setp.neu.f32 %p108, %f1095, 0f3EE8BA2E;
selp.f32 %f1458, 0f7FFFFFFF, %f1458, %p108;
BB0_100:
add.f32 %f1097, %f196, 0f3EE8BA2E;
mov.b32 %r364, %f1097;
setp.lt.s32 %p110, %r364, 2139095040;
@%p110 bra BB0_105;
setp.gtu.f32 %p111, %f196, 0f7F800000;
@%p111 bra BB0_104;
bra.uni BB0_102;
BB0_104:
add.f32 %f1458, %f1453, 0f3EE8BA2E;
bra.uni BB0_105;
BB0_102:
setp.neu.f32 %p112, %f196, 0f7F800000;
@%p112 bra BB0_105;
selp.f32 %f1458, 0fFF800000, 0f7F800000, %p1;
BB0_105:
mul.f32 %f1098, %f1458, 0f437F0000;
setp.eq.f32 %p113, %f1453, 0f3F800000;
selp.f32 %f1099, 0f437F0000, %f1098, %p113;
cvt.rzi.u32.f32 %r365, %f1099;
cvt.u16.u32 %rs30, %r365;
mov.u16 %rs31, 255;
st.v2.u8 [%rd95], {%rs30, %rs31};
ld.global.u32 %r482, [imageEnabled];
BB0_106:
and.b32 %r366, %r482, 1;
setp.eq.b32 %p114, %r366, 1;
@!%p114 bra BB0_141;
bra.uni BB0_107;
BB0_107:
mov.f32 %f1102, 0f3E666666;
cvt.rzi.f32.f32 %f1103, %f1102;
fma.rn.f32 %f1104, %f1103, 0fC0000000, 0f3EE66666;
abs.f32 %f208, %f1104;
abs.f32 %f209, %f192;
setp.lt.f32 %p115, %f209, 0f00800000;
mul.f32 %f1105, %f209, 0f4B800000;
selp.f32 %f1106, 0fC3170000, 0fC2FE0000, %p115;
selp.f32 %f1107, %f1105, %f209, %p115;
mov.b32 %r367, %f1107;
and.b32 %r368, %r367, 8388607;
or.b32 %r369, %r368, 1065353216;
mov.b32 %f1108, %r369;
shr.u32 %r370, %r367, 23;
cvt.rn.f32.u32 %f1109, %r370;
add.f32 %f1110, %f1106, %f1109;
setp.gt.f32 %p116, %f1108, 0f3FB504F3;
mul.f32 %f1111, %f1108, 0f3F000000;
add.f32 %f1112, %f1110, 0f3F800000;
selp.f32 %f1113, %f1111, %f1108, %p116;
selp.f32 %f1114, %f1112, %f1110, %p116;
add.f32 %f1115, %f1113, 0fBF800000;
add.f32 %f1101, %f1113, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1100,%f1101;
// inline asm
add.f32 %f1116, %f1115, %f1115;
mul.f32 %f1117, %f1100, %f1116;
mul.f32 %f1118, %f1117, %f1117;
mov.f32 %f1119, 0f3C4CAF63;
mov.f32 %f1120, 0f3B18F0FE;
fma.rn.f32 %f1121, %f1120, %f1118, %f1119;
mov.f32 %f1122, 0f3DAAAABD;
fma.rn.f32 %f1123, %f1121, %f1118, %f1122;
mul.rn.f32 %f1124, %f1123, %f1118;
mul.rn.f32 %f1125, %f1124, %f1117;
sub.f32 %f1126, %f1115, %f1117;
neg.f32 %f1127, %f1117;
add.f32 %f1128, %f1126, %f1126;
fma.rn.f32 %f1129, %f1127, %f1115, %f1128;
mul.rn.f32 %f1130, %f1100, %f1129;
add.f32 %f1131, %f1125, %f1117;
sub.f32 %f1132, %f1117, %f1131;
add.f32 %f1133, %f1125, %f1132;
add.f32 %f1134, %f1130, %f1133;
add.f32 %f1135, %f1131, %f1134;
sub.f32 %f1136, %f1131, %f1135;
add.f32 %f1137, %f1134, %f1136;
mov.f32 %f1138, 0f3F317200;
mul.rn.f32 %f1139, %f1114, %f1138;
mov.f32 %f1140, 0f35BFBE8E;
mul.rn.f32 %f1141, %f1114, %f1140;
add.f32 %f1142, %f1139, %f1135;
sub.f32 %f1143, %f1139, %f1142;
add.f32 %f1144, %f1135, %f1143;
add.f32 %f1145, %f1137, %f1144;
add.f32 %f1146, %f1141, %f1145;
add.f32 %f1147, %f1142, %f1146;
sub.f32 %f1148, %f1142, %f1147;
add.f32 %f1149, %f1146, %f1148;
mov.f32 %f1150, 0f3EE66666;
mul.rn.f32 %f1151, %f1150, %f1147;
neg.f32 %f1152, %f1151;
fma.rn.f32 %f1153, %f1150, %f1147, %f1152;
fma.rn.f32 %f1154, %f1150, %f1149, %f1153;
mov.f32 %f1155, 0f00000000;
fma.rn.f32 %f1156, %f1155, %f1147, %f1154;
add.rn.f32 %f1157, %f1151, %f1156;
neg.f32 %f1158, %f1157;
add.rn.f32 %f1159, %f1151, %f1158;
add.rn.f32 %f1160, %f1159, %f1156;
mov.b32 %r371, %f1157;
setp.eq.s32 %p117, %r371, 1118925336;
add.s32 %r372, %r371, -1;
mov.b32 %f1161, %r372;
add.f32 %f1162, %f1160, 0f37000000;
selp.f32 %f1163, %f1161, %f1157, %p117;
selp.f32 %f210, %f1162, %f1160, %p117;
mul.f32 %f1164, %f1163, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1165, %f1164;
mov.f32 %f1166, 0fBF317200;
fma.rn.f32 %f1167, %f1165, %f1166, %f1163;
mov.f32 %f1168, 0fB5BFBE8E;
fma.rn.f32 %f1169, %f1165, %f1168, %f1167;
mul.f32 %f1170, %f1169, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1171, %f1170;
add.f32 %f1172, %f1165, 0f00000000;
ex2.approx.f32 %f1173, %f1172;
mul.f32 %f1174, %f1171, %f1173;
setp.lt.f32 %p118, %f1163, 0fC2D20000;
selp.f32 %f1175, 0f00000000, %f1174, %p118;
setp.gt.f32 %p119, %f1163, 0f42D20000;
selp.f32 %f1459, 0f7F800000, %f1175, %p119;
setp.eq.f32 %p120, %f1459, 0f7F800000;
@%p120 bra BB0_109;
fma.rn.f32 %f1459, %f1459, %f210, %f1459;
BB0_109:
setp.lt.f32 %p121, %f192, 0f00000000;
setp.eq.f32 %p122, %f208, 0f3F800000;
and.pred %p2, %p121, %p122;
mov.b32 %r373, %f1459;
xor.b32 %r374, %r373, -2147483648;
mov.b32 %f1176, %r374;
selp.f32 %f1461, %f1176, %f1459, %p2;
setp.eq.f32 %p123, %f192, 0f00000000;
@%p123 bra BB0_112;
bra.uni BB0_110;
BB0_112:
add.f32 %f1179, %f192, %f192;
selp.f32 %f1461, %f1179, 0f00000000, %p122;
bra.uni BB0_113;
BB0_110:
setp.geu.f32 %p124, %f192, 0f00000000;
@%p124 bra BB0_113;
cvt.rzi.f32.f32 %f1178, %f1150;
setp.neu.f32 %p125, %f1178, 0f3EE66666;
selp.f32 %f1461, 0f7FFFFFFF, %f1461, %p125;
BB0_113:
add.f32 %f1180, %f209, 0f3EE66666;
mov.b32 %r375, %f1180;
setp.lt.s32 %p127, %r375, 2139095040;
@%p127 bra BB0_118;
setp.gtu.f32 %p128, %f209, 0f7F800000;
@%p128 bra BB0_117;
bra.uni BB0_115;
BB0_117:
add.f32 %f1461, %f192, 0f3EE66666;
bra.uni BB0_118;
BB0_115:
setp.neu.f32 %p129, %f209, 0f7F800000;
@%p129 bra BB0_118;
selp.f32 %f1461, 0fFF800000, 0f7F800000, %p2;
BB0_118:
setp.eq.f32 %p130, %f192, 0f3F800000;
selp.f32 %f221, 0f3F800000, %f1461, %p130;
abs.f32 %f222, %f193;
setp.lt.f32 %p131, %f222, 0f00800000;
mul.f32 %f1183, %f222, 0f4B800000;
selp.f32 %f1184, 0fC3170000, 0fC2FE0000, %p131;
selp.f32 %f1185, %f1183, %f222, %p131;
mov.b32 %r376, %f1185;
and.b32 %r377, %r376, 8388607;
or.b32 %r378, %r377, 1065353216;
mov.b32 %f1186, %r378;
shr.u32 %r379, %r376, 23;
cvt.rn.f32.u32 %f1187, %r379;
add.f32 %f1188, %f1184, %f1187;
setp.gt.f32 %p132, %f1186, 0f3FB504F3;
mul.f32 %f1189, %f1186, 0f3F000000;
add.f32 %f1190, %f1188, 0f3F800000;
selp.f32 %f1191, %f1189, %f1186, %p132;
selp.f32 %f1192, %f1190, %f1188, %p132;
add.f32 %f1193, %f1191, 0fBF800000;
add.f32 %f1182, %f1191, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1181,%f1182;
// inline asm
add.f32 %f1194, %f1193, %f1193;
mul.f32 %f1195, %f1181, %f1194;
mul.f32 %f1196, %f1195, %f1195;
fma.rn.f32 %f1199, %f1120, %f1196, %f1119;
fma.rn.f32 %f1201, %f1199, %f1196, %f1122;
mul.rn.f32 %f1202, %f1201, %f1196;
mul.rn.f32 %f1203, %f1202, %f1195;
sub.f32 %f1204, %f1193, %f1195;
neg.f32 %f1205, %f1195;
add.f32 %f1206, %f1204, %f1204;
fma.rn.f32 %f1207, %f1205, %f1193, %f1206;
mul.rn.f32 %f1208, %f1181, %f1207;
add.f32 %f1209, %f1203, %f1195;
sub.f32 %f1210, %f1195, %f1209;
add.f32 %f1211, %f1203, %f1210;
add.f32 %f1212, %f1208, %f1211;
add.f32 %f1213, %f1209, %f1212;
sub.f32 %f1214, %f1209, %f1213;
add.f32 %f1215, %f1212, %f1214;
mul.rn.f32 %f1217, %f1192, %f1138;
mul.rn.f32 %f1219, %f1192, %f1140;
add.f32 %f1220, %f1217, %f1213;
sub.f32 %f1221, %f1217, %f1220;
add.f32 %f1222, %f1213, %f1221;
add.f32 %f1223, %f1215, %f1222;
add.f32 %f1224, %f1219, %f1223;
add.f32 %f1225, %f1220, %f1224;
sub.f32 %f1226, %f1220, %f1225;
add.f32 %f1227, %f1224, %f1226;
mul.rn.f32 %f1229, %f1150, %f1225;
neg.f32 %f1230, %f1229;
fma.rn.f32 %f1231, %f1150, %f1225, %f1230;
fma.rn.f32 %f1232, %f1150, %f1227, %f1231;
fma.rn.f32 %f1234, %f1155, %f1225, %f1232;
add.rn.f32 %f1235, %f1229, %f1234;
neg.f32 %f1236, %f1235;
add.rn.f32 %f1237, %f1229, %f1236;
add.rn.f32 %f1238, %f1237, %f1234;
mov.b32 %r380, %f1235;
setp.eq.s32 %p133, %r380, 1118925336;
add.s32 %r381, %r380, -1;
mov.b32 %f1239, %r381;
add.f32 %f1240, %f1238, 0f37000000;
selp.f32 %f1241, %f1239, %f1235, %p133;
selp.f32 %f223, %f1240, %f1238, %p133;
mul.f32 %f1242, %f1241, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1243, %f1242;
fma.rn.f32 %f1245, %f1243, %f1166, %f1241;
fma.rn.f32 %f1247, %f1243, %f1168, %f1245;
mul.f32 %f1248, %f1247, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1249, %f1248;
add.f32 %f1250, %f1243, 0f00000000;
ex2.approx.f32 %f1251, %f1250;
mul.f32 %f1252, %f1249, %f1251;
setp.lt.f32 %p134, %f1241, 0fC2D20000;
selp.f32 %f1253, 0f00000000, %f1252, %p134;
setp.gt.f32 %p135, %f1241, 0f42D20000;
selp.f32 %f1462, 0f7F800000, %f1253, %p135;
setp.eq.f32 %p136, %f1462, 0f7F800000;
@%p136 bra BB0_120;
fma.rn.f32 %f1462, %f1462, %f223, %f1462;
BB0_120:
setp.lt.f32 %p137, %f193, 0f00000000;
and.pred %p3, %p137, %p122;
mov.b32 %r382, %f1462;
xor.b32 %r383, %r382, -2147483648;
mov.b32 %f1254, %r383;
selp.f32 %f1464, %f1254, %f1462, %p3;
setp.eq.f32 %p139, %f193, 0f00000000;
@%p139 bra BB0_123;
bra.uni BB0_121;
BB0_123:
add.f32 %f1257, %f193, %f193;
selp.f32 %f1464, %f1257, 0f00000000, %p122;
bra.uni BB0_124;
BB0_121:
setp.geu.f32 %p140, %f193, 0f00000000;
@%p140 bra BB0_124;
cvt.rzi.f32.f32 %f1256, %f1150;
setp.neu.f32 %p141, %f1256, 0f3EE66666;
selp.f32 %f1464, 0f7FFFFFFF, %f1464, %p141;
BB0_124:
add.f32 %f1258, %f222, 0f3EE66666;
mov.b32 %r384, %f1258;
setp.lt.s32 %p143, %r384, 2139095040;
@%p143 bra BB0_129;
setp.gtu.f32 %p144, %f222, 0f7F800000;
@%p144 bra BB0_128;
bra.uni BB0_126;
BB0_128:
add.f32 %f1464, %f193, 0f3EE66666;
bra.uni BB0_129;
BB0_126:
setp.neu.f32 %p145, %f222, 0f7F800000;
@%p145 bra BB0_129;
selp.f32 %f1464, 0fFF800000, 0f7F800000, %p3;
BB0_129:
setp.eq.f32 %p146, %f193, 0f3F800000;
selp.f32 %f234, 0f3F800000, %f1464, %p146;
abs.f32 %f235, %f194;
setp.lt.f32 %p147, %f235, 0f00800000;
mul.f32 %f1261, %f235, 0f4B800000;
selp.f32 %f1262, 0fC3170000, 0fC2FE0000, %p147;
selp.f32 %f1263, %f1261, %f235, %p147;
mov.b32 %r385, %f1263;
and.b32 %r386, %r385, 8388607;
or.b32 %r387, %r386, 1065353216;
mov.b32 %f1264, %r387;
shr.u32 %r388, %r385, 23;
cvt.rn.f32.u32 %f1265, %r388;
add.f32 %f1266, %f1262, %f1265;
setp.gt.f32 %p148, %f1264, 0f3FB504F3;
mul.f32 %f1267, %f1264, 0f3F000000;
add.f32 %f1268, %f1266, 0f3F800000;
selp.f32 %f1269, %f1267, %f1264, %p148;
selp.f32 %f1270, %f1268, %f1266, %p148;
add.f32 %f1271, %f1269, 0fBF800000;
add.f32 %f1260, %f1269, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1259,%f1260;
// inline asm
add.f32 %f1272, %f1271, %f1271;
mul.f32 %f1273, %f1259, %f1272;
mul.f32 %f1274, %f1273, %f1273;
fma.rn.f32 %f1277, %f1120, %f1274, %f1119;
fma.rn.f32 %f1279, %f1277, %f1274, %f1122;
mul.rn.f32 %f1280, %f1279, %f1274;
mul.rn.f32 %f1281, %f1280, %f1273;
sub.f32 %f1282, %f1271, %f1273;
neg.f32 %f1283, %f1273;
add.f32 %f1284, %f1282, %f1282;
fma.rn.f32 %f1285, %f1283, %f1271, %f1284;
mul.rn.f32 %f1286, %f1259, %f1285;
add.f32 %f1287, %f1281, %f1273;
sub.f32 %f1288, %f1273, %f1287;
add.f32 %f1289, %f1281, %f1288;
add.f32 %f1290, %f1286, %f1289;
add.f32 %f1291, %f1287, %f1290;
sub.f32 %f1292, %f1287, %f1291;
add.f32 %f1293, %f1290, %f1292;
mul.rn.f32 %f1295, %f1270, %f1138;
mul.rn.f32 %f1297, %f1270, %f1140;
add.f32 %f1298, %f1295, %f1291;
sub.f32 %f1299, %f1295, %f1298;
add.f32 %f1300, %f1291, %f1299;
add.f32 %f1301, %f1293, %f1300;
add.f32 %f1302, %f1297, %f1301;
add.f32 %f1303, %f1298, %f1302;
sub.f32 %f1304, %f1298, %f1303;
add.f32 %f1305, %f1302, %f1304;
mul.rn.f32 %f1307, %f1150, %f1303;
neg.f32 %f1308, %f1307;
fma.rn.f32 %f1309, %f1150, %f1303, %f1308;
fma.rn.f32 %f1310, %f1150, %f1305, %f1309;
fma.rn.f32 %f1312, %f1155, %f1303, %f1310;
add.rn.f32 %f1313, %f1307, %f1312;
neg.f32 %f1314, %f1313;
add.rn.f32 %f1315, %f1307, %f1314;
add.rn.f32 %f1316, %f1315, %f1312;
mov.b32 %r389, %f1313;
setp.eq.s32 %p149, %r389, 1118925336;
add.s32 %r390, %r389, -1;
mov.b32 %f1317, %r390;
add.f32 %f1318, %f1316, 0f37000000;
selp.f32 %f1319, %f1317, %f1313, %p149;
selp.f32 %f236, %f1318, %f1316, %p149;
mul.f32 %f1320, %f1319, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1321, %f1320;
fma.rn.f32 %f1323, %f1321, %f1166, %f1319;
fma.rn.f32 %f1325, %f1321, %f1168, %f1323;
mul.f32 %f1326, %f1325, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1327, %f1326;
add.f32 %f1328, %f1321, 0f00000000;
ex2.approx.f32 %f1329, %f1328;
mul.f32 %f1330, %f1327, %f1329;
setp.lt.f32 %p150, %f1319, 0fC2D20000;
selp.f32 %f1331, 0f00000000, %f1330, %p150;
setp.gt.f32 %p151, %f1319, 0f42D20000;
selp.f32 %f1465, 0f7F800000, %f1331, %p151;
setp.eq.f32 %p152, %f1465, 0f7F800000;
@%p152 bra BB0_131;
fma.rn.f32 %f1465, %f1465, %f236, %f1465;
BB0_131:
setp.lt.f32 %p153, %f194, 0f00000000;
and.pred %p4, %p153, %p122;
mov.b32 %r391, %f1465;
xor.b32 %r392, %r391, -2147483648;
mov.b32 %f1332, %r392;
selp.f32 %f1467, %f1332, %f1465, %p4;
setp.eq.f32 %p155, %f194, 0f00000000;
@%p155 bra BB0_134;
bra.uni BB0_132;
BB0_134:
add.f32 %f1335, %f194, %f194;
selp.f32 %f1467, %f1335, 0f00000000, %p122;
bra.uni BB0_135;
BB0_132:
setp.geu.f32 %p156, %f194, 0f00000000;
@%p156 bra BB0_135;
cvt.rzi.f32.f32 %f1334, %f1150;
setp.neu.f32 %p157, %f1334, 0f3EE66666;
selp.f32 %f1467, 0f7FFFFFFF, %f1467, %p157;
BB0_135:
add.f32 %f1336, %f235, 0f3EE66666;
mov.b32 %r393, %f1336;
setp.lt.s32 %p159, %r393, 2139095040;
@%p159 bra BB0_140;
setp.gtu.f32 %p160, %f235, 0f7F800000;
@%p160 bra BB0_139;
bra.uni BB0_137;
BB0_139:
add.f32 %f1467, %f194, 0f3EE66666;
bra.uni BB0_140;
BB0_137:
setp.neu.f32 %p161, %f235, 0f7F800000;
@%p161 bra BB0_140;
selp.f32 %f1467, 0fFF800000, 0f7F800000, %p4;
BB0_140:
mov.u32 %r448, 4;
mov.u32 %r447, 2;
setp.eq.f32 %p162, %f194, 0f3F800000;
selp.f32 %f1337, 0f3F800000, %f1467, %p162;
cvt.u64.u32 %rd105, %r3;
cvt.u64.u32 %rd104, %r2;
mov.u64 %rd108, image;
cvta.global.u64 %rd103, %rd108;
// inline asm
call (%rd102), _rt_buffer_get_64, (%rd103, %r447, %r448, %rd104, %rd105, %rd29, %rd29);
// inline asm
cvt.sat.f32.f32 %f1338, %f1337;
mul.f32 %f1339, %f1338, 0f437FFD71;
cvt.rzi.u32.f32 %r396, %f1339;
cvt.sat.f32.f32 %f1340, %f234;
mul.f32 %f1341, %f1340, 0f437FFD71;
cvt.rzi.u32.f32 %r397, %f1341;
cvt.sat.f32.f32 %f1342, %f221;
mul.f32 %f1343, %f1342, 0f437FFD71;
cvt.rzi.u32.f32 %r398, %f1343;
cvt.u16.u32 %rs32, %r396;
cvt.u16.u32 %rs33, %r398;
cvt.u16.u32 %rs34, %r397;
mov.u16 %rs35, 255;
st.v4.u8 [%rd102], {%rs32, %rs34, %rs33, %rs35};
ld.global.u32 %r482, [imageEnabled];
BB0_141:
and.b32 %r399, %r482, 4;
setp.eq.s32 %p163, %r399, 0;
@%p163 bra BB0_145;
ld.global.u32 %r400, [additive];
setp.eq.s32 %p164, %r400, 0;
cvt.u64.u32 %rd18, %r2;
cvt.u64.u32 %rd19, %r3;
mov.f32 %f1344, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs36, %f1344;}
// inline asm
@%p164 bra BB0_144;
mov.u32 %r449, 2;
mov.u64 %rd121, image_HDR;
cvta.global.u64 %rd110, %rd121;
mov.u32 %r404, 8;
// inline asm
call (%rd109), _rt_buffer_get_64, (%rd110, %r449, %r404, %rd18, %rd19, %rd29, %rd29);
// inline asm
ld.v4.u16 {%rs43, %rs44, %rs45, %rs46}, [%rd109];
// inline asm
{ cvt.f32.f16 %f1345, %rs43;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1346, %rs44;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1347, %rs45;}
// inline asm
// inline asm
call (%rd115), _rt_buffer_get_64, (%rd110, %r449, %r404, %rd18, %rd19, %rd29, %rd29);
// inline asm
add.f32 %f1348, %f192, %f1345;
add.f32 %f1349, %f193, %f1346;
add.f32 %f1350, %f194, %f1347;
// inline asm
{ cvt.rn.f16.f32 %rs42, %f1350;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs41, %f1349;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs40, %f1348;}
// inline asm
st.v4.u16 [%rd115], {%rs40, %rs41, %rs42, %rs36};
bra.uni BB0_145;
BB0_144:
mov.u32 %r450, 2;
mov.u64 %rd128, image_HDR;
cvta.global.u64 %rd123, %rd128;
mov.u32 %r406, 8;
// inline asm
call (%rd122), _rt_buffer_get_64, (%rd123, %r450, %r406, %rd18, %rd19, %rd29, %rd29);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs49, %f194;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs48, %f193;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs47, %f192;}
// inline asm
st.v4.u16 [%rd122], {%rs47, %rs48, %rs49, %rs36};
BB0_145:
ld.global.u8 %rs50, [imageEnabled];
and.b16 %rs51, %rs50, 64;
setp.eq.s16 %p165, %rs51, 0;
@%p165 bra BB0_167;
mov.u32 %r452, 4;
mov.u32 %r451, 2;
ld.global.f32 %f1354, [directDir];
ld.global.f32 %f1355, [directDir+4];
ld.global.f32 %f1356, [directDir+8];
cvt.u64.u32 %rd132, %r3;
cvt.u64.u32 %rd131, %r2;
mov.u64 %rd135, image_Dir;
cvta.global.u64 %rd130, %rd135;
// inline asm
call (%rd129), _rt_buffer_get_64, (%rd130, %r451, %r452, %rd131, %rd132, %rd29, %rd29);
// inline asm
fma.rn.f32 %f1357, %f1354, 0fBF000000, 0f3F000000;
mul.f32 %f1358, %f1357, 0f437F0000;
cvt.rzi.u32.f32 %r409, %f1358;
fma.rn.f32 %f1359, %f1355, 0fBF000000, 0f3F000000;
mul.f32 %f1360, %f1359, 0f437F0000;
cvt.rzi.u32.f32 %r410, %f1360;
fma.rn.f32 %f1361, %f1356, 0fBF000000, 0f3F000000;
mul.f32 %f1362, %f1361, 0f437F0000;
cvt.rzi.u32.f32 %r411, %f1362;
cvt.u16.u32 %rs52, %r411;
cvt.u16.u32 %rs53, %r410;
cvt.u16.u32 %rs54, %r409;
mov.u16 %rs55, 255;
st.v4.u8 [%rd129], {%rs54, %rs53, %rs52, %rs55};
BB0_167:
ret;
}
Reference in New Issue View Git Blame Copy Permalink