saad
/
ArabDesert
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
ArabDesert/Assets/Editor/x64/Bakery/lmSunSuperSample.ptx

2499 lines
75 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[36];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<160>;
.reg .b16 %rs<101>;
.reg .f32 %f<1351>;
.reg .b32 %r<485>;
.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 %f1297, 0f00000000;
mov.f32 %f1298, %f1297;
mov.f32 %f1299, %f1297;
@%p5 bra BB0_2;
ld.u8 %rs15, [%rd24+1];
and.b16 %rs17, %rs13, 255;
cvt.rn.f32.u16 %f202, %rs17;
div.rn.f32 %f203, %f202, 0f437F0000;
fma.rn.f32 %f204, %f203, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f205, %rs15;
div.rn.f32 %f206, %f205, 0f437F0000;
fma.rn.f32 %f207, %f206, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f208, %rs12;
div.rn.f32 %f209, %f208, 0f437F0000;
fma.rn.f32 %f210, %f209, 0f40000000, 0fBF800000;
mul.f32 %f211, %f207, %f207;
fma.rn.f32 %f212, %f204, %f204, %f211;
fma.rn.f32 %f213, %f210, %f210, %f212;
sqrt.rn.f32 %f214, %f213;
rcp.rn.f32 %f215, %f214;
mul.f32 %f1297, %f204, %f215;
mul.f32 %f1298, %f207, %f215;
mul.f32 %f1299, %f210, %f215;
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, %f1298, 0f00000000;
setp.eq.f32 %p7, %f1297, 0f00000000;
and.pred %p8, %p7, %p6;
setp.eq.f32 %p9, %f1299, 0f00000000;
and.pred %p10, %p8, %p9;
@%p10 bra BB0_141;
bra.uni BB0_3;
BB0_141:
ld.global.u32 %r484, [imageEnabled];
and.b32 %r429, %r484, 1;
setp.eq.b32 %p155, %r429, 1;
@!%p155 bra BB0_143;
bra.uni BB0_142;
BB0_142:
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 %rs74, 0;
st.v4.u8 [%rd177], {%rs74, %rs74, %rs74, %rs74};
ld.global.u32 %r484, [imageEnabled];
BB0_143:
and.b32 %r432, %r484, 8;
setp.eq.s32 %p156, %r432, 0;
@%p156 bra BB0_145;
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 %f1287, 0f00000000;
cvt.rzi.u32.f32 %r435, %f1287;
cvt.u16.u32 %rs75, %r435;
mov.u16 %rs76, 0;
st.v2.u8 [%rd184], {%rs75, %rs76};
ld.global.u32 %r484, [imageEnabled];
BB0_145:
and.b32 %r436, %r484, 4;
setp.eq.s32 %p157, %r436, 0;
@%p157 bra BB0_149;
ld.global.u32 %r437, [additive];
setp.eq.s32 %p158, %r437, 0;
cvt.u64.u32 %rd22, %r2;
cvt.u64.u32 %rd23, %r3;
@%p158 bra BB0_148;
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 {%rs83, %rs84, %rs85, %rs86}, [%rd191];
// inline asm
{ cvt.f32.f16 %f1288, %rs83;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1289, %rs84;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1290, %rs85;}
// inline asm
// inline asm
call (%rd197), _rt_buffer_get_64, (%rd192, %r112, %r441, %rd22, %rd23, %rd202, %rd202);
// inline asm
add.f32 %f1291, %f1288, 0f00000000;
add.f32 %f1292, %f1289, 0f00000000;
add.f32 %f1293, %f1290, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs82, %f1293;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs81, %f1292;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs80, %f1291;}
// inline asm
mov.u16 %rs87, 0;
st.v4.u16 [%rd197], {%rs80, %rs81, %rs82, %rs87};
bra.uni BB0_149;
BB0_3:
ld.global.f32 %f9, [directDir+8];
ld.global.f32 %f8, [directDir+4];
ld.global.f32 %f7, [directDir];
mul.f32 %f216, %f1297, %f7;
mul.f32 %f217, %f1298, %f8;
neg.f32 %f218, %f217;
sub.f32 %f219, %f218, %f216;
mul.f32 %f220, %f1299, %f9;
sub.f32 %f10, %f219, %f220;
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 %r482, [imageEnabled];
and.b32 %r128, %r482, 32;
setp.eq.s32 %p14, %r128, 0;
and.pred %p15, %p13, %p14;
@%p15 bra BB0_131;
bra.uni BB0_4;
BB0_131:
and.b32 %r412, %r482, 1;
setp.eq.b32 %p150, %r412, 1;
@!%p150 bra BB0_133;
bra.uni BB0_132;
BB0_132:
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 %rs54, 255;
mov.u16 %rs55, 0;
st.v4.u8 [%rd136], {%rs55, %rs55, %rs55, %rs54};
ld.global.u32 %r482, [imageEnabled];
BB0_133:
and.b32 %r415, %r482, 8;
setp.eq.s32 %p151, %r415, 0;
@%p151 bra BB0_135;
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 %f1276, 0f00000000;
cvt.rzi.u32.f32 %r418, %f1276;
cvt.u16.u32 %rs56, %r418;
mov.u16 %rs57, 255;
st.v2.u8 [%rd143], {%rs56, %rs57};
ld.global.u32 %r482, [imageEnabled];
BB0_135:
and.b32 %r419, %r482, 4;
setp.eq.s32 %p152, %r419, 0;
@%p152 bra BB0_139;
ld.global.u32 %r420, [additive];
setp.eq.s32 %p153, %r420, 0;
cvt.u64.u32 %rd20, %r2;
cvt.u64.u32 %rd21, %r3;
mov.f32 %f1277, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs58, %f1277;}
// inline asm
@%p153 bra BB0_138;
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 {%rs65, %rs66, %rs67, %rs68}, [%rd150];
// inline asm
{ cvt.f32.f16 %f1278, %rs65;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1279, %rs66;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1280, %rs67;}
// inline asm
// inline asm
call (%rd156), _rt_buffer_get_64, (%rd151, %r112, %r424, %rd20, %rd21, %rd161, %rd161);
// inline asm
add.f32 %f1281, %f1278, 0f00000000;
add.f32 %f1282, %f1279, 0f00000000;
add.f32 %f1283, %f1280, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs64, %f1283;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs63, %f1282;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs62, %f1281;}
// inline asm
st.v4.u16 [%rd156], {%rs62, %rs63, %rs64, %rs58};
bra.uni BB0_139;
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 %f223, [%rd31+8];
ld.f32 %f224, [%rd31+4];
ld.f32 %f225, [%rd31];
mul.f32 %f226, %f225, 0f3456BF95;
mul.f32 %f227, %f224, 0f3456BF95;
mul.f32 %f228, %f223, 0f3456BF95;
abs.f32 %f229, %f1297;
div.rn.f32 %f230, %f226, %f229;
abs.f32 %f231, %f1298;
div.rn.f32 %f232, %f227, %f231;
abs.f32 %f233, %f1299;
div.rn.f32 %f234, %f228, %f233;
abs.f32 %f235, %f230;
abs.f32 %f236, %f232;
abs.f32 %f237, %f234;
mov.f32 %f238, 0f38D1B717;
max.f32 %f239, %f235, %f238;
max.f32 %f240, %f236, %f238;
max.f32 %f241, %f237, %f238;
fma.rn.f32 %f11, %f1297, %f239, %f225;
fma.rn.f32 %f12, %f1298, %f240, %f224;
fma.rn.f32 %f13, %f1299, %f241, %f223;
setp.gt.f32 %p16, %f229, %f233;
neg.f32 %f242, %f1298;
selp.f32 %f243, %f242, 0f00000000, %p16;
neg.f32 %f244, %f1299;
selp.f32 %f245, %f1297, %f244, %p16;
selp.f32 %f246, 0f00000000, %f1298, %p16;
mul.f32 %f247, %f245, %f245;
fma.rn.f32 %f248, %f243, %f243, %f247;
fma.rn.f32 %f249, %f246, %f246, %f248;
sqrt.rn.f32 %f250, %f249;
rcp.rn.f32 %f251, %f250;
mul.f32 %f14, %f243, %f251;
mul.f32 %f15, %f245, %f251;
mul.f32 %f16, %f246, %f251;
mul.f32 %f252, %f1299, %f15;
mul.f32 %f253, %f1298, %f16;
sub.f32 %f17, %f252, %f253;
mul.f32 %f254, %f1297, %f16;
mul.f32 %f255, %f1299, %f14;
sub.f32 %f18, %f254, %f255;
mul.f32 %f256, %f1298, %f14;
mul.f32 %f257, %f1297, %f15;
sub.f32 %f19, %f256, %f257;
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 %f258, [%rd37+8];
ld.f32 %f259, [%rd37+4];
ld.f32 %f260, [%rd37];
setp.eq.f32 %p17, %f260, 0f00000000;
setp.eq.f32 %p18, %f259, 0f00000000;
and.pred %p19, %p17, %p18;
setp.eq.f32 %p20, %f258, 0f00000000;
and.pred %p21, %p19, %p20;
selp.f32 %f261, %f225, %f260, %p21;
selp.f32 %f262, %f224, %f259, %p21;
selp.f32 %f263, %f223, %f258, %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 %f264, [%rd43+8];
ld.f32 %f265, [%rd43+4];
ld.f32 %f266, [%rd43];
setp.eq.f32 %p22, %f266, 0f00000000;
setp.eq.f32 %p23, %f265, 0f00000000;
and.pred %p24, %p22, %p23;
setp.eq.f32 %p25, %f264, 0f00000000;
and.pred %p26, %p24, %p25;
selp.f32 %f267, %f225, %f266, %p26;
selp.f32 %f268, %f224, %f265, %p26;
selp.f32 %f269, %f223, %f264, %p26;
sub.f32 %f270, %f261, %f225;
sub.f32 %f271, %f262, %f224;
sub.f32 %f272, %f263, %f223;
sub.f32 %f273, %f267, %f225;
sub.f32 %f274, %f268, %f224;
sub.f32 %f275, %f269, %f223;
abs.f32 %f276, %f270;
abs.f32 %f277, %f273;
max.f32 %f278, %f276, %f277;
abs.f32 %f279, %f271;
abs.f32 %f280, %f274;
max.f32 %f281, %f279, %f280;
abs.f32 %f282, %f272;
abs.f32 %f283, %f275;
max.f32 %f284, %f282, %f283;
max.f32 %f285, %f278, %f281;
max.f32 %f286, %f285, %f284;
setp.neu.f32 %p27, %f270, 0f00000000;
setp.neu.f32 %p28, %f271, 0f00000000;
or.pred %p29, %p27, %p28;
setp.neu.f32 %p30, %f272, 0f00000000;
setp.neu.f32 %p31, %f273, 0f00000000;
setp.neu.f32 %p32, %f274, 0f00000000;
or.pred %p33, %p31, %p32;
setp.neu.f32 %p34, %f275, 0f00000000;
or.pred %p35, %p33, %p34;
or.pred %p36, %p29, %p30;
and.pred %p37, %p36, %p35;
selp.f32 %f20, %f286, 0f00000000, %p37;
abs.f32 %f287, %f9;
abs.f32 %f288, %f7;
setp.gt.f32 %p38, %f288, %f287;
neg.f32 %f289, %f8;
selp.f32 %f290, %f289, 0f00000000, %p38;
neg.f32 %f291, %f9;
selp.f32 %f292, %f7, %f291, %p38;
selp.f32 %f293, 0f00000000, %f8, %p38;
mul.f32 %f294, %f292, %f292;
fma.rn.f32 %f295, %f290, %f290, %f294;
fma.rn.f32 %f296, %f293, %f293, %f295;
sqrt.rn.f32 %f297, %f296;
rcp.rn.f32 %f298, %f297;
mul.f32 %f21, %f290, %f298;
mul.f32 %f22, %f292, %f298;
mul.f32 %f23, %f293, %f298;
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 %r451, [samples];
mov.f32 %f1302, 0f00000000;
setp.lt.s32 %p39, %r451, 1;
@%p39 bra BB0_5;
cvt.rn.f32.s32 %f301, %r451;
rcp.rn.f32 %f24, %f301;
ld.u32 %r477, [%rd49];
mul.f32 %f25, %f11, 0f3456BF95;
mul.f32 %f26, %f12, 0f3456BF95;
mul.f32 %f27, %f13, 0f3456BF95;
mul.f32 %f302, %f7, %f22;
mul.f32 %f303, %f8, %f21;
sub.f32 %f28, %f303, %f302;
mul.f32 %f304, %f9, %f21;
mul.f32 %f305, %f7, %f23;
sub.f32 %f29, %f305, %f304;
mul.f32 %f306, %f8, %f23;
mul.f32 %f307, %f9, %f22;
sub.f32 %f30, %f307, %f306;
mov.f32 %f1302, 0f00000000;
mov.u32 %r452, 0;
abs.f32 %f308, %f26;
abs.f32 %f309, %f25;
max.f32 %f310, %f309, %f308;
abs.f32 %f311, %f27;
max.f32 %f312, %f310, %f311;
mov.f32 %f1303, %f1302;
BB0_7:
setp.lt.s32 %p40, %r451, 1;
@%p40 bra BB0_74;
mov.u32 %r454, 0;
BB0_9:
cvt.rn.f32.s32 %f1295, %r452;
mad.lo.s32 %r163, %r477, 1664525, 1013904223;
and.b32 %r164, %r163, 16777215;
cvt.rn.f32.u32 %f314, %r164;
fma.rn.f32 %f315, %f314, 0f33800000, %f1295;
mul.f32 %f316, %f24, %f315;
mad.lo.s32 %r12, %r163, 1664525, 1013904223;
and.b32 %r165, %r12, 16777215;
cvt.rn.f32.u32 %f317, %r165;
cvt.rn.f32.s32 %f318, %r454;
fma.rn.f32 %f319, %f317, 0f33800000, %f318;
mul.f32 %f320, %f24, %f319;
sqrt.rn.f32 %f37, %f316;
mul.f32 %f1310, %f320, 0f40C90FDB;
abs.f32 %f39, %f1310;
setp.neu.f32 %p41, %f39, 0f7F800000;
mov.f32 %f1304, %f1310;
@%p41 bra BB0_11;
mov.f32 %f321, 0f00000000;
mul.rn.f32 %f1304, %f1310, %f321;
BB0_11:
mul.f32 %f322, %f1304, 0f3F22F983;
cvt.rni.s32.f32 %r465, %f322;
cvt.rn.f32.s32 %f323, %r465;
neg.f32 %f324, %f323;
mov.f32 %f325, 0f3FC90FDA;
fma.rn.f32 %f326, %f324, %f325, %f1304;
mov.f32 %f327, 0f33A22168;
fma.rn.f32 %f328, %f324, %f327, %f326;
mov.f32 %f329, 0f27C234C5;
fma.rn.f32 %f1305, %f324, %f329, %f328;
abs.f32 %f330, %f1304;
setp.leu.f32 %p42, %f330, 0f47CE4780;
@%p42 bra BB0_22;
add.u64 %rd58, %SP, 0;
cvta.to.local.u64 %rd3, %rd58;
mov.b32 %r14, %f1304;
shr.u32 %r15, %r14, 23;
shl.b32 %r168, %r14, 8;
or.b32 %r16, %r168, -2147483648;
mov.u32 %r456, 0;
mov.u64 %rd219, 0;
mov.u64 %rd218, %rd3;
mov.u32 %r457, %r456;
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, %r457;
madc.hi.u32 %r457, %r171, %r16, 0;
}
// inline asm
st.local.u32 [%rd218], %r169;
add.s32 %r456, %r456, 1;
cvt.s64.s32 %rd219, %r456;
mul.wide.s32 %rd64, %r456, 4;
add.s64 %rd218, %rd3, %rd64;
setp.ne.s32 %p43, %r456, 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], %r457;
mov.u32 %r177, 6;
sub.s32 %r178, %r177, %r176;
mul.wide.s32 %rd67, %r178, 4;
add.s64 %rd8, %rd66, %rd67;
ld.local.u32 %r458, [%rd8];
ld.local.u32 %r459, [%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, %r459, %r180;
shl.b32 %r182, %r458, %r24;
add.s32 %r458, %r181, %r182;
ld.local.u32 %r183, [%rd8+-8];
shr.u32 %r184, %r183, %r180;
shl.b32 %r185, %r459, %r24;
add.s32 %r459, %r184, %r185;
BB0_16:
shr.u32 %r186, %r459, 30;
shl.b32 %r187, %r458, 2;
add.s32 %r460, %r186, %r187;
shl.b32 %r30, %r459, 2;
shr.u32 %r188, %r460, 31;
shr.u32 %r189, %r458, 30;
add.s32 %r31, %r188, %r189;
setp.eq.s32 %p45, %r188, 0;
@%p45 bra BB0_17;
bra.uni BB0_18;
BB0_17:
mov.u32 %r461, %r21;
mov.u32 %r462, %r30;
bra.uni BB0_19;
BB0_18:
not.b32 %r190, %r460;
neg.s32 %r462, %r30;
setp.eq.s32 %p46, %r30, 0;
selp.u32 %r191, 1, 0, %p46;
add.s32 %r460, %r191, %r190;
xor.b32 %r461, %r21, -2147483648;
BB0_19:
clz.b32 %r464, %r460;
setp.eq.s32 %p47, %r464, 0;
shl.b32 %r192, %r460, %r464;
mov.u32 %r193, 32;
sub.s32 %r194, %r193, %r464;
shr.u32 %r195, %r462, %r194;
add.s32 %r196, %r195, %r192;
selp.b32 %r39, %r460, %r196, %p47;
mov.u32 %r197, -921707870;
mul.hi.u32 %r463, %r39, %r197;
setp.eq.s32 %p48, %r21, 0;
neg.s32 %r198, %r31;
selp.b32 %r465, %r31, %r198, %p48;
setp.lt.s32 %p49, %r463, 1;
@%p49 bra BB0_21;
mul.lo.s32 %r199, %r39, -921707870;
shr.u32 %r200, %r199, 31;
shl.b32 %r201, %r463, 1;
add.s32 %r463, %r200, %r201;
add.s32 %r464, %r464, 1;
BB0_21:
mov.u32 %r202, 126;
sub.s32 %r203, %r202, %r464;
shl.b32 %r204, %r203, 23;
add.s32 %r205, %r463, 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, %r461;
mov.b32 %f1305, %r210;
BB0_22:
mul.rn.f32 %f45, %f1305, %f1305;
add.s32 %r47, %r465, 1;
and.b32 %r48, %r47, 1;
setp.eq.s32 %p50, %r48, 0;
@%p50 bra BB0_24;
bra.uni BB0_23;
BB0_24:
mov.f32 %f333, 0f3C08839E;
mov.f32 %f334, 0fB94CA1F9;
fma.rn.f32 %f1306, %f334, %f45, %f333;
bra.uni BB0_25;
BB0_23:
mov.f32 %f331, 0fBAB6061A;
mov.f32 %f332, 0f37CCF5CE;
fma.rn.f32 %f1306, %f332, %f45, %f331;
BB0_25:
@%p50 bra BB0_27;
bra.uni BB0_26;
BB0_27:
mov.f32 %f338, 0fBE2AAAA3;
fma.rn.f32 %f339, %f1306, %f45, %f338;
mov.f32 %f340, 0f00000000;
fma.rn.f32 %f1307, %f339, %f45, %f340;
bra.uni BB0_28;
BB0_26:
mov.f32 %f335, 0f3D2AAAA5;
fma.rn.f32 %f336, %f1306, %f45, %f335;
mov.f32 %f337, 0fBF000000;
fma.rn.f32 %f1307, %f336, %f45, %f337;
BB0_28:
fma.rn.f32 %f1308, %f1307, %f1305, %f1305;
@%p50 bra BB0_30;
mov.f32 %f341, 0f3F800000;
fma.rn.f32 %f1308, %f1307, %f45, %f341;
BB0_30:
and.b32 %r211, %r47, 2;
setp.eq.s32 %p53, %r211, 0;
@%p53 bra BB0_32;
mov.f32 %f342, 0f00000000;
mov.f32 %f343, 0fBF800000;
fma.rn.f32 %f1308, %f1308, %f343, %f342;
BB0_32:
@%p41 bra BB0_34;
mov.f32 %f344, 0f00000000;
mul.rn.f32 %f1310, %f1310, %f344;
BB0_34:
mul.f32 %f345, %f1310, 0f3F22F983;
cvt.rni.s32.f32 %r475, %f345;
cvt.rn.f32.s32 %f346, %r475;
neg.f32 %f347, %f346;
fma.rn.f32 %f349, %f347, %f325, %f1310;
fma.rn.f32 %f351, %f347, %f327, %f349;
fma.rn.f32 %f1311, %f347, %f329, %f351;
abs.f32 %f353, %f1310;
setp.leu.f32 %p55, %f353, 0f47CE4780;
@%p55 bra BB0_45;
add.u64 %rd69, %SP, 0;
cvta.to.local.u64 %rd9, %rd69;
mov.b32 %r50, %f1310;
shr.u32 %r51, %r50, 23;
shl.b32 %r214, %r50, 8;
or.b32 %r52, %r214, -2147483648;
mov.u32 %r466, 0;
mov.u64 %rd220, %rd9;
mov.u64 %rd221, %rd29;
mov.u32 %r467, %r466;
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, %r467;
madc.hi.u32 %r467, %r217, %r52, 0;
}
// inline asm
st.local.u32 [%rd220], %r215;
add.s32 %r466, %r466, 1;
cvt.s64.s32 %rd221, %r466;
mul.wide.s32 %rd73, %r466, 4;
add.s64 %rd220, %rd9, %rd73;
setp.ne.s32 %p56, %r466, 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], %r467;
mov.u32 %r223, 6;
sub.s32 %r224, %r223, %r222;
mul.wide.s32 %rd76, %r224, 4;
add.s64 %rd15, %rd75, %rd76;
ld.local.u32 %r468, [%rd15];
ld.local.u32 %r469, [%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, %r469, %r226;
shl.b32 %r228, %r468, %r60;
add.s32 %r468, %r227, %r228;
ld.local.u32 %r229, [%rd15+-8];
shr.u32 %r230, %r229, %r226;
shl.b32 %r231, %r469, %r60;
add.s32 %r469, %r230, %r231;
BB0_39:
shr.u32 %r232, %r469, 30;
shl.b32 %r233, %r468, 2;
add.s32 %r470, %r232, %r233;
shl.b32 %r66, %r469, 2;
shr.u32 %r234, %r470, 31;
shr.u32 %r235, %r468, 30;
add.s32 %r67, %r234, %r235;
setp.eq.s32 %p58, %r234, 0;
@%p58 bra BB0_40;
bra.uni BB0_41;
BB0_40:
mov.u32 %r471, %r57;
mov.u32 %r472, %r66;
bra.uni BB0_42;
BB0_41:
not.b32 %r236, %r470;
neg.s32 %r472, %r66;
setp.eq.s32 %p59, %r66, 0;
selp.u32 %r237, 1, 0, %p59;
add.s32 %r470, %r237, %r236;
xor.b32 %r471, %r57, -2147483648;
BB0_42:
clz.b32 %r474, %r470;
setp.eq.s32 %p60, %r474, 0;
shl.b32 %r238, %r470, %r474;
mov.u32 %r239, 32;
sub.s32 %r240, %r239, %r474;
shr.u32 %r241, %r472, %r240;
add.s32 %r242, %r241, %r238;
selp.b32 %r75, %r470, %r242, %p60;
mov.u32 %r243, -921707870;
mul.hi.u32 %r473, %r75, %r243;
setp.eq.s32 %p61, %r57, 0;
neg.s32 %r244, %r67;
selp.b32 %r475, %r67, %r244, %p61;
setp.lt.s32 %p62, %r473, 1;
@%p62 bra BB0_44;
mul.lo.s32 %r245, %r75, -921707870;
shr.u32 %r246, %r245, 31;
shl.b32 %r247, %r473, 1;
add.s32 %r473, %r246, %r247;
add.s32 %r474, %r474, 1;
BB0_44:
mov.u32 %r248, 126;
sub.s32 %r249, %r248, %r474;
shl.b32 %r250, %r249, 23;
add.s32 %r251, %r473, 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, %r471;
mov.b32 %f1311, %r256;
BB0_45:
mul.rn.f32 %f62, %f1311, %f1311;
and.b32 %r83, %r475, 1;
setp.eq.s32 %p63, %r83, 0;
@%p63 bra BB0_47;
bra.uni BB0_46;
BB0_47:
mov.f32 %f356, 0f3C08839E;
mov.f32 %f357, 0fB94CA1F9;
fma.rn.f32 %f1312, %f357, %f62, %f356;
bra.uni BB0_48;
BB0_46:
mov.f32 %f354, 0fBAB6061A;
mov.f32 %f355, 0f37CCF5CE;
fma.rn.f32 %f1312, %f355, %f62, %f354;
BB0_48:
@%p63 bra BB0_50;
bra.uni BB0_49;
BB0_50:
mov.f32 %f361, 0fBE2AAAA3;
fma.rn.f32 %f362, %f1312, %f62, %f361;
mov.f32 %f363, 0f00000000;
fma.rn.f32 %f1313, %f362, %f62, %f363;
bra.uni BB0_51;
BB0_49:
mov.f32 %f358, 0f3D2AAAA5;
fma.rn.f32 %f359, %f1312, %f62, %f358;
mov.f32 %f360, 0fBF000000;
fma.rn.f32 %f1313, %f359, %f62, %f360;
BB0_51:
fma.rn.f32 %f1314, %f1313, %f1311, %f1311;
@%p63 bra BB0_53;
mov.f32 %f364, 0f3F800000;
fma.rn.f32 %f1314, %f1313, %f62, %f364;
BB0_53:
and.b32 %r257, %r475, 2;
setp.eq.s32 %p66, %r257, 0;
@%p66 bra BB0_55;
mov.f32 %f365, 0f00000000;
mov.f32 %f366, 0fBF800000;
fma.rn.f32 %f1314, %f1314, %f366, %f365;
BB0_55:
max.f32 %f1296, %f312, %f238;
mul.f32 %f375, %f37, %f1308;
add.u64 %rd77, %SP, 28;
cvta.to.local.u64 %rd16, %rd77;
mul.f32 %f376, %f375, %f375;
mov.f32 %f377, 0f3F800000;
sub.f32 %f378, %f377, %f376;
mul.f32 %f379, %f37, %f1314;
mul.f32 %f380, %f379, %f379;
sub.f32 %f381, %f378, %f380;
mov.f32 %f382, 0f00000000;
max.f32 %f383, %f382, %f381;
sqrt.rn.f32 %f384, %f383;
mul.f32 %f385, %f21, %f379;
mul.f32 %f386, %f22, %f379;
mul.f32 %f387, %f23, %f379;
fma.rn.f32 %f388, %f30, %f375, %f385;
fma.rn.f32 %f389, %f29, %f375, %f386;
fma.rn.f32 %f390, %f28, %f375, %f387;
fma.rn.f32 %f391, %f7, %f384, %f388;
fma.rn.f32 %f392, %f8, %f384, %f389;
fma.rn.f32 %f393, %f9, %f384, %f390;
add.f32 %f394, %f7, %f391;
add.f32 %f395, %f8, %f392;
add.f32 %f396, %f9, %f393;
ld.global.f32 %f397, [shadowSpread];
mul.f32 %f398, %f397, %f394;
mul.f32 %f399, %f397, %f395;
mul.f32 %f400, %f397, %f396;
sub.f32 %f401, %f398, %f7;
sub.f32 %f402, %f399, %f8;
sub.f32 %f403, %f400, %f9;
mul.f32 %f404, %f402, %f402;
fma.rn.f32 %f405, %f401, %f401, %f404;
fma.rn.f32 %f406, %f403, %f403, %f405;
sqrt.rn.f32 %f407, %f406;
rcp.rn.f32 %f408, %f407;
mul.f32 %f74, %f408, %f401;
mul.f32 %f75, %f408, %f402;
mul.f32 %f76, %f408, %f403;
ld.global.u32 %r261, [imageEnabled];
and.b32 %r262, %r261, 32;
setp.eq.s32 %p67, %r262, 0;
selp.f32 %f409, 0f3F800000, 0f41200000, %p67;
mul.f32 %f373, %f409, %f1296;
mad.lo.s32 %r263, %r12, 1664525, 1013904223;
and.b32 %r264, %r263, 16777215;
cvt.rn.f32.u32 %f410, %r264;
mul.f32 %f411, %f410, 0f33800000;
fma.rn.f32 %f412, %f411, 0f40000000, 0fBF800000;
mad.lo.s32 %r84, %r263, 1664525, 1013904223;
and.b32 %r265, %r84, 16777215;
cvt.rn.f32.u32 %f413, %r265;
mul.f32 %f414, %f413, 0f33800000;
fma.rn.f32 %f415, %f414, 0f40000000, 0fBF800000;
mul.f32 %f416, %f14, %f415;
mul.f32 %f417, %f15, %f415;
mul.f32 %f418, %f16, %f415;
fma.rn.f32 %f419, %f17, %f412, %f416;
fma.rn.f32 %f420, %f18, %f412, %f417;
fma.rn.f32 %f421, %f19, %f412, %f418;
mul.f32 %f422, %f420, %f420;
fma.rn.f32 %f423, %f419, %f419, %f422;
fma.rn.f32 %f424, %f421, %f421, %f423;
sqrt.rn.f32 %f425, %f424;
rcp.rn.f32 %f426, %f425;
mul.f32 %f427, %f419, %f426;
mul.f32 %f428, %f420, %f426;
mul.f32 %f429, %f421, %f426;
mul.f32 %f430, %f1298, %f429;
mul.f32 %f431, %f1299, %f428;
sub.f32 %f432, %f430, %f431;
mul.f32 %f433, %f1299, %f427;
mul.f32 %f434, %f1297, %f429;
sub.f32 %f435, %f433, %f434;
mul.f32 %f436, %f1297, %f428;
mul.f32 %f437, %f1298, %f427;
sub.f32 %f438, %f436, %f437;
mul.f32 %f439, %f427, 0f3E24A490;
mul.f32 %f440, %f428, 0f3E24A490;
mul.f32 %f441, %f429, 0f3E24A490;
fma.rn.f32 %f442, %f20, %f439, %f11;
fma.rn.f32 %f443, %f20, %f440, %f12;
fma.rn.f32 %f444, %f20, %f441, %f13;
mul.f32 %f445, %f432, 0fBECECEDE;
mul.f32 %f446, %f435, 0fBECECEDE;
mul.f32 %f447, %f438, 0fBECECEDE;
fma.rn.f32 %f78, %f20, %f445, %f442;
fma.rn.f32 %f79, %f20, %f446, %f443;
fma.rn.f32 %f80, %f20, %f447, %f444;
sub.f32 %f448, %f78, %f11;
sub.f32 %f449, %f79, %f12;
sub.f32 %f450, %f80, %f13;
mul.f32 %f451, %f449, %f449;
fma.rn.f32 %f452, %f448, %f448, %f451;
fma.rn.f32 %f453, %f450, %f450, %f452;
sqrt.rn.f32 %f374, %f453;
rcp.rn.f32 %f454, %f374;
mul.f32 %f370, %f454, %f448;
mul.f32 %f371, %f454, %f449;
mul.f32 %f372, %f454, %f450;
mov.u32 %r266, 1065353216;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r258, [root];
mov.u32 %r259, 1;
// inline asm
call _rt_trace_64, (%r258, %f11, %f12, %f13, %f370, %f371, %f372, %r259, %f373, %f374, %rd77, %r113);
// inline asm
ld.local.f32 %f455, [%rd16];
mov.u16 %rs94, 0;
setp.neu.f32 %p68, %f455, 0f3F800000;
@%p68 bra BB0_57;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r267, [root];
mov.f32 %f463, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r267, %f78, %f79, %f80, %f74, %f75, %f76, %r259, %f373, %f463, %rd77, %r113);
// inline asm
ld.local.f32 %f464, [%rd16];
add.f32 %f1303, %f1303, %f464;
add.f32 %f1302, %f1302, 0f3F800000;
mov.u16 %rs94, 1;
BB0_57:
mad.lo.s32 %r274, %r84, 1664525, 1013904223;
and.b32 %r275, %r274, 16777215;
cvt.rn.f32.u32 %f473, %r275;
mul.f32 %f474, %f473, 0f33800000;
fma.rn.f32 %f475, %f474, 0f40000000, 0fBF800000;
mad.lo.s32 %r85, %r274, 1664525, 1013904223;
and.b32 %r276, %r85, 16777215;
cvt.rn.f32.u32 %f476, %r276;
mul.f32 %f477, %f476, 0f33800000;
fma.rn.f32 %f478, %f477, 0f40000000, 0fBF800000;
mul.f32 %f479, %f14, %f478;
mul.f32 %f480, %f15, %f478;
mul.f32 %f481, %f16, %f478;
fma.rn.f32 %f482, %f17, %f475, %f479;
fma.rn.f32 %f483, %f18, %f475, %f480;
fma.rn.f32 %f484, %f19, %f475, %f481;
mul.f32 %f485, %f483, %f483;
fma.rn.f32 %f486, %f482, %f482, %f485;
fma.rn.f32 %f487, %f484, %f484, %f486;
sqrt.rn.f32 %f488, %f487;
rcp.rn.f32 %f489, %f488;
mul.f32 %f490, %f482, %f489;
mul.f32 %f491, %f483, %f489;
mul.f32 %f492, %f484, %f489;
mul.f32 %f493, %f1298, %f492;
mul.f32 %f494, %f1299, %f491;
sub.f32 %f495, %f493, %f494;
mul.f32 %f496, %f1299, %f490;
mul.f32 %f497, %f1297, %f492;
sub.f32 %f498, %f496, %f497;
mul.f32 %f499, %f1297, %f491;
mul.f32 %f500, %f1298, %f490;
sub.f32 %f501, %f499, %f500;
mul.f32 %f502, %f490, 0fBE24A490;
mul.f32 %f503, %f491, 0fBE24A490;
mul.f32 %f504, %f492, 0fBE24A490;
fma.rn.f32 %f505, %f20, %f502, %f11;
fma.rn.f32 %f506, %f20, %f503, %f12;
fma.rn.f32 %f507, %f20, %f504, %f13;
mul.f32 %f508, %f495, 0f3ECACABC;
mul.f32 %f509, %f498, 0f3ECACABC;
mul.f32 %f510, %f501, 0f3ECACABC;
fma.rn.f32 %f85, %f20, %f508, %f505;
fma.rn.f32 %f86, %f20, %f509, %f506;
fma.rn.f32 %f87, %f20, %f510, %f507;
sub.f32 %f511, %f85, %f11;
sub.f32 %f512, %f86, %f12;
sub.f32 %f513, %f87, %f13;
mul.f32 %f514, %f512, %f512;
fma.rn.f32 %f515, %f511, %f511, %f514;
fma.rn.f32 %f516, %f513, %f513, %f515;
sqrt.rn.f32 %f472, %f516;
rcp.rn.f32 %f517, %f472;
mul.f32 %f468, %f517, %f511;
mul.f32 %f469, %f517, %f512;
mul.f32 %f470, %f517, %f513;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r271, [root];
// inline asm
call _rt_trace_64, (%r271, %f11, %f12, %f13, %f468, %f469, %f470, %r259, %f373, %f472, %rd77, %r113);
// inline asm
ld.local.f32 %f518, [%rd16];
setp.neu.f32 %p69, %f518, 0f3F800000;
@%p69 bra BB0_59;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r278, [root];
mov.f32 %f526, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r278, %f85, %f86, %f87, %f74, %f75, %f76, %r259, %f373, %f526, %rd77, %r113);
// inline asm
ld.local.f32 %f527, [%rd16];
add.f32 %f1303, %f1303, %f527;
add.f32 %f1302, %f1302, 0f3F800000;
mov.u16 %rs94, 1;
BB0_59:
mad.lo.s32 %r285, %r85, 1664525, 1013904223;
and.b32 %r286, %r285, 16777215;
cvt.rn.f32.u32 %f536, %r286;
mul.f32 %f537, %f536, 0f33800000;
fma.rn.f32 %f538, %f537, 0f40000000, 0fBF800000;
mad.lo.s32 %r86, %r285, 1664525, 1013904223;
and.b32 %r287, %r86, 16777215;
cvt.rn.f32.u32 %f539, %r287;
mul.f32 %f540, %f539, 0f33800000;
fma.rn.f32 %f541, %f540, 0f40000000, 0fBF800000;
mul.f32 %f542, %f14, %f541;
mul.f32 %f543, %f15, %f541;
mul.f32 %f544, %f16, %f541;
fma.rn.f32 %f545, %f17, %f538, %f542;
fma.rn.f32 %f546, %f18, %f538, %f543;
fma.rn.f32 %f547, %f19, %f538, %f544;
mul.f32 %f548, %f546, %f546;
fma.rn.f32 %f549, %f545, %f545, %f548;
fma.rn.f32 %f550, %f547, %f547, %f549;
sqrt.rn.f32 %f551, %f550;
rcp.rn.f32 %f552, %f551;
mul.f32 %f553, %f545, %f552;
mul.f32 %f554, %f546, %f552;
mul.f32 %f555, %f547, %f552;
mul.f32 %f556, %f1298, %f555;
mul.f32 %f557, %f1299, %f554;
sub.f32 %f558, %f556, %f557;
mul.f32 %f559, %f1299, %f553;
mul.f32 %f560, %f1297, %f555;
sub.f32 %f561, %f559, %f560;
mul.f32 %f562, %f1297, %f554;
mul.f32 %f563, %f1298, %f553;
sub.f32 %f564, %f562, %f563;
mul.f32 %f565, %f553, 0f3F23A3A8;
mul.f32 %f566, %f554, 0f3F23A3A8;
mul.f32 %f567, %f555, 0f3F23A3A8;
fma.rn.f32 %f568, %f20, %f565, %f11;
fma.rn.f32 %f569, %f20, %f566, %f12;
fma.rn.f32 %f570, %f20, %f567, %f13;
mul.f32 %f571, %f558, 0f3E24A490;
mul.f32 %f572, %f561, 0f3E24A490;
mul.f32 %f573, %f564, 0f3E24A490;
fma.rn.f32 %f92, %f20, %f571, %f568;
fma.rn.f32 %f93, %f20, %f572, %f569;
fma.rn.f32 %f94, %f20, %f573, %f570;
sub.f32 %f574, %f92, %f11;
sub.f32 %f575, %f93, %f12;
sub.f32 %f576, %f94, %f13;
mul.f32 %f577, %f575, %f575;
fma.rn.f32 %f578, %f574, %f574, %f577;
fma.rn.f32 %f579, %f576, %f576, %f578;
sqrt.rn.f32 %f535, %f579;
rcp.rn.f32 %f580, %f535;
mul.f32 %f531, %f580, %f574;
mul.f32 %f532, %f580, %f575;
mul.f32 %f533, %f580, %f576;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r282, [root];
// inline asm
call _rt_trace_64, (%r282, %f11, %f12, %f13, %f531, %f532, %f533, %r259, %f373, %f535, %rd77, %r113);
// inline asm
ld.local.f32 %f581, [%rd16];
setp.neu.f32 %p70, %f581, 0f3F800000;
@%p70 bra BB0_61;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r289, [root];
mov.f32 %f589, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r289, %f92, %f93, %f94, %f74, %f75, %f76, %r259, %f373, %f589, %rd77, %r113);
// inline asm
ld.local.f32 %f590, [%rd16];
add.f32 %f1303, %f1303, %f590;
add.f32 %f1302, %f1302, 0f3F800000;
mov.u16 %rs94, 1;
BB0_61:
mad.lo.s32 %r296, %r86, 1664525, 1013904223;
and.b32 %r297, %r296, 16777215;
cvt.rn.f32.u32 %f599, %r297;
mul.f32 %f600, %f599, 0f33800000;
fma.rn.f32 %f601, %f600, 0f40000000, 0fBF800000;
mad.lo.s32 %r87, %r296, 1664525, 1013904223;
and.b32 %r298, %r87, 16777215;
cvt.rn.f32.u32 %f602, %r298;
mul.f32 %f603, %f602, 0f33800000;
fma.rn.f32 %f604, %f603, 0f40000000, 0fBF800000;
mul.f32 %f605, %f14, %f604;
mul.f32 %f606, %f15, %f604;
mul.f32 %f607, %f16, %f604;
fma.rn.f32 %f608, %f17, %f601, %f605;
fma.rn.f32 %f609, %f18, %f601, %f606;
fma.rn.f32 %f610, %f19, %f601, %f607;
mul.f32 %f611, %f609, %f609;
fma.rn.f32 %f612, %f608, %f608, %f611;
fma.rn.f32 %f613, %f610, %f610, %f612;
sqrt.rn.f32 %f614, %f613;
rcp.rn.f32 %f615, %f614;
mul.f32 %f616, %f608, %f615;
mul.f32 %f617, %f609, %f615;
mul.f32 %f618, %f610, %f615;
mul.f32 %f619, %f1298, %f618;
mul.f32 %f620, %f1299, %f617;
sub.f32 %f621, %f619, %f620;
mul.f32 %f622, %f1299, %f616;
mul.f32 %f623, %f1297, %f618;
sub.f32 %f624, %f622, %f623;
mul.f32 %f625, %f1297, %f617;
mul.f32 %f626, %f1298, %f616;
sub.f32 %f627, %f625, %f626;
mul.f32 %f628, %f616, 0fBECECEDE;
mul.f32 %f629, %f617, 0fBECECEDE;
mul.f32 %f630, %f618, 0fBECECEDE;
fma.rn.f32 %f631, %f20, %f628, %f11;
fma.rn.f32 %f632, %f20, %f629, %f12;
fma.rn.f32 %f633, %f20, %f630, %f13;
mul.f32 %f634, %f621, 0fBF23A3A9;
mul.f32 %f635, %f624, 0fBF23A3A9;
mul.f32 %f636, %f627, 0fBF23A3A9;
fma.rn.f32 %f99, %f20, %f634, %f631;
fma.rn.f32 %f100, %f20, %f635, %f632;
fma.rn.f32 %f101, %f20, %f636, %f633;
sub.f32 %f637, %f99, %f11;
sub.f32 %f638, %f100, %f12;
sub.f32 %f639, %f101, %f13;
mul.f32 %f640, %f638, %f638;
fma.rn.f32 %f641, %f637, %f637, %f640;
fma.rn.f32 %f642, %f639, %f639, %f641;
sqrt.rn.f32 %f598, %f642;
rcp.rn.f32 %f643, %f598;
mul.f32 %f594, %f643, %f637;
mul.f32 %f595, %f643, %f638;
mul.f32 %f596, %f643, %f639;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r293, [root];
// inline asm
call _rt_trace_64, (%r293, %f11, %f12, %f13, %f594, %f595, %f596, %r259, %f373, %f598, %rd77, %r113);
// inline asm
ld.local.f32 %f644, [%rd16];
setp.neu.f32 %p71, %f644, 0f3F800000;
@%p71 bra BB0_63;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r300, [root];
mov.f32 %f652, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r300, %f99, %f100, %f101, %f74, %f75, %f76, %r259, %f373, %f652, %rd77, %r113);
// inline asm
ld.local.f32 %f653, [%rd16];
add.f32 %f1303, %f1303, %f653;
add.f32 %f1302, %f1302, 0f3F800000;
mov.u16 %rs94, 1;
BB0_63:
mad.lo.s32 %r307, %r87, 1664525, 1013904223;
and.b32 %r308, %r307, 16777215;
cvt.rn.f32.u32 %f662, %r308;
mul.f32 %f663, %f662, 0f33800000;
fma.rn.f32 %f664, %f663, 0f40000000, 0fBF800000;
mad.lo.s32 %r88, %r307, 1664525, 1013904223;
and.b32 %r309, %r88, 16777215;
cvt.rn.f32.u32 %f665, %r309;
mul.f32 %f666, %f665, 0f33800000;
fma.rn.f32 %f667, %f666, 0f40000000, 0fBF800000;
mul.f32 %f668, %f14, %f667;
mul.f32 %f669, %f15, %f667;
mul.f32 %f670, %f16, %f667;
fma.rn.f32 %f671, %f17, %f664, %f668;
fma.rn.f32 %f672, %f18, %f664, %f669;
fma.rn.f32 %f673, %f19, %f664, %f670;
mul.f32 %f674, %f672, %f672;
fma.rn.f32 %f675, %f671, %f671, %f674;
fma.rn.f32 %f676, %f673, %f673, %f675;
sqrt.rn.f32 %f677, %f676;
rcp.rn.f32 %f678, %f677;
mul.f32 %f679, %f671, %f678;
mul.f32 %f680, %f672, %f678;
mul.f32 %f681, %f673, %f678;
mul.f32 %f682, %f1298, %f681;
mul.f32 %f683, %f1299, %f680;
sub.f32 %f684, %f682, %f683;
mul.f32 %f685, %f1299, %f679;
mul.f32 %f686, %f1297, %f681;
sub.f32 %f687, %f685, %f686;
mul.f32 %f688, %f1297, %f680;
mul.f32 %f689, %f1298, %f679;
sub.f32 %f690, %f688, %f689;
mul.f32 %f691, %f679, 0fBF23A3A9;
mul.f32 %f692, %f680, 0fBF23A3A9;
mul.f32 %f693, %f681, 0fBF23A3A9;
fma.rn.f32 %f694, %f20, %f691, %f11;
fma.rn.f32 %f695, %f20, %f692, %f12;
fma.rn.f32 %f696, %f20, %f693, %f13;
mul.f32 %f697, %f684, 0f3F23A3A8;
mul.f32 %f698, %f687, 0f3F23A3A8;
mul.f32 %f699, %f690, 0f3F23A3A8;
fma.rn.f32 %f106, %f20, %f697, %f694;
fma.rn.f32 %f107, %f20, %f698, %f695;
fma.rn.f32 %f108, %f20, %f699, %f696;
sub.f32 %f700, %f106, %f11;
sub.f32 %f701, %f107, %f12;
sub.f32 %f702, %f108, %f13;
mul.f32 %f703, %f701, %f701;
fma.rn.f32 %f704, %f700, %f700, %f703;
fma.rn.f32 %f705, %f702, %f702, %f704;
sqrt.rn.f32 %f661, %f705;
rcp.rn.f32 %f706, %f661;
mul.f32 %f657, %f706, %f700;
mul.f32 %f658, %f706, %f701;
mul.f32 %f659, %f706, %f702;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r304, [root];
// inline asm
call _rt_trace_64, (%r304, %f11, %f12, %f13, %f657, %f658, %f659, %r259, %f373, %f661, %rd77, %r113);
// inline asm
ld.local.f32 %f707, [%rd16];
setp.neu.f32 %p72, %f707, 0f3F800000;
@%p72 bra BB0_65;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r311, [root];
mov.f32 %f715, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r311, %f106, %f107, %f108, %f74, %f75, %f76, %r259, %f373, %f715, %rd77, %r113);
// inline asm
ld.local.f32 %f716, [%rd16];
add.f32 %f1303, %f1303, %f716;
add.f32 %f1302, %f1302, 0f3F800000;
mov.u16 %rs94, 1;
BB0_65:
mad.lo.s32 %r318, %r88, 1664525, 1013904223;
and.b32 %r319, %r318, 16777215;
cvt.rn.f32.u32 %f725, %r319;
mul.f32 %f726, %f725, 0f33800000;
fma.rn.f32 %f727, %f726, 0f40000000, 0fBF800000;
mad.lo.s32 %r89, %r318, 1664525, 1013904223;
and.b32 %r320, %r89, 16777215;
cvt.rn.f32.u32 %f728, %r320;
mul.f32 %f729, %f728, 0f33800000;
fma.rn.f32 %f730, %f729, 0f40000000, 0fBF800000;
mul.f32 %f731, %f14, %f730;
mul.f32 %f732, %f15, %f730;
mul.f32 %f733, %f16, %f730;
fma.rn.f32 %f734, %f17, %f727, %f731;
fma.rn.f32 %f735, %f18, %f727, %f732;
fma.rn.f32 %f736, %f19, %f727, %f733;
mul.f32 %f737, %f735, %f735;
fma.rn.f32 %f738, %f734, %f734, %f737;
fma.rn.f32 %f739, %f736, %f736, %f738;
sqrt.rn.f32 %f740, %f739;
rcp.rn.f32 %f741, %f740;
mul.f32 %f742, %f734, %f741;
mul.f32 %f743, %f735, %f741;
mul.f32 %f744, %f736, %f741;
mul.f32 %f745, %f1298, %f744;
mul.f32 %f746, %f1299, %f743;
sub.f32 %f747, %f745, %f746;
mul.f32 %f748, %f1299, %f742;
mul.f32 %f749, %f1297, %f744;
sub.f32 %f750, %f748, %f749;
mul.f32 %f751, %f1297, %f743;
mul.f32 %f752, %f1298, %f742;
sub.f32 %f753, %f751, %f752;
mul.f32 %f754, %f742, 0fBF61E1E3;
mul.f32 %f755, %f743, 0fBF61E1E3;
mul.f32 %f756, %f744, 0fBF61E1E3;
fma.rn.f32 %f757, %f20, %f754, %f11;
fma.rn.f32 %f758, %f20, %f755, %f12;
fma.rn.f32 %f759, %f20, %f756, %f13;
mul.f32 %f760, %f747, 0fBE24A490;
mul.f32 %f761, %f750, 0fBE24A490;
mul.f32 %f762, %f753, 0fBE24A490;
fma.rn.f32 %f113, %f20, %f760, %f757;
fma.rn.f32 %f114, %f20, %f761, %f758;
fma.rn.f32 %f115, %f20, %f762, %f759;
sub.f32 %f763, %f113, %f11;
sub.f32 %f764, %f114, %f12;
sub.f32 %f765, %f115, %f13;
mul.f32 %f766, %f764, %f764;
fma.rn.f32 %f767, %f763, %f763, %f766;
fma.rn.f32 %f768, %f765, %f765, %f767;
sqrt.rn.f32 %f724, %f768;
rcp.rn.f32 %f769, %f724;
mul.f32 %f720, %f769, %f763;
mul.f32 %f721, %f769, %f764;
mul.f32 %f722, %f769, %f765;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r315, [root];
// inline asm
call _rt_trace_64, (%r315, %f11, %f12, %f13, %f720, %f721, %f722, %r259, %f373, %f724, %rd77, %r113);
// inline asm
ld.local.f32 %f770, [%rd16];
setp.neu.f32 %p73, %f770, 0f3F800000;
@%p73 bra BB0_67;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r322, [root];
mov.f32 %f778, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r322, %f113, %f114, %f115, %f74, %f75, %f76, %r259, %f373, %f778, %rd77, %r113);
// inline asm
ld.local.f32 %f779, [%rd16];
add.f32 %f1303, %f1303, %f779;
add.f32 %f1302, %f1302, 0f3F800000;
mov.u16 %rs94, 1;
BB0_67:
mad.lo.s32 %r329, %r89, 1664525, 1013904223;
and.b32 %r330, %r329, 16777215;
cvt.rn.f32.u32 %f788, %r330;
mul.f32 %f789, %f788, 0f33800000;
fma.rn.f32 %f790, %f789, 0f40000000, 0fBF800000;
mad.lo.s32 %r90, %r329, 1664525, 1013904223;
and.b32 %r331, %r90, 16777215;
cvt.rn.f32.u32 %f791, %r331;
mul.f32 %f792, %f791, 0f33800000;
fma.rn.f32 %f793, %f792, 0f40000000, 0fBF800000;
mul.f32 %f794, %f14, %f793;
mul.f32 %f795, %f15, %f793;
mul.f32 %f796, %f16, %f793;
fma.rn.f32 %f797, %f17, %f790, %f794;
fma.rn.f32 %f798, %f18, %f790, %f795;
fma.rn.f32 %f799, %f19, %f790, %f796;
mul.f32 %f800, %f798, %f798;
fma.rn.f32 %f801, %f797, %f797, %f800;
fma.rn.f32 %f802, %f799, %f799, %f801;
sqrt.rn.f32 %f803, %f802;
rcp.rn.f32 %f804, %f803;
mul.f32 %f805, %f797, %f804;
mul.f32 %f806, %f798, %f804;
mul.f32 %f807, %f799, %f804;
mul.f32 %f808, %f1298, %f807;
mul.f32 %f809, %f1299, %f806;
sub.f32 %f810, %f808, %f809;
mul.f32 %f811, %f1299, %f805;
mul.f32 %f812, %f1297, %f807;
sub.f32 %f813, %f811, %f812;
mul.f32 %f814, %f1297, %f806;
mul.f32 %f815, %f1298, %f805;
sub.f32 %f816, %f814, %f815;
mul.f32 %f817, %f805, 0f3ECACABC;
mul.f32 %f818, %f806, 0f3ECACABC;
mul.f32 %f819, %f807, 0f3ECACABC;
fma.rn.f32 %f820, %f20, %f817, %f11;
fma.rn.f32 %f821, %f20, %f818, %f12;
fma.rn.f32 %f822, %f20, %f819, %f13;
mul.f32 %f823, %f810, 0f3F61E1D2;
mul.f32 %f824, %f813, 0f3F61E1D2;
mul.f32 %f825, %f816, 0f3F61E1D2;
fma.rn.f32 %f120, %f20, %f823, %f820;
fma.rn.f32 %f121, %f20, %f824, %f821;
fma.rn.f32 %f122, %f20, %f825, %f822;
sub.f32 %f826, %f120, %f11;
sub.f32 %f827, %f121, %f12;
sub.f32 %f828, %f122, %f13;
mul.f32 %f829, %f827, %f827;
fma.rn.f32 %f830, %f826, %f826, %f829;
fma.rn.f32 %f831, %f828, %f828, %f830;
sqrt.rn.f32 %f787, %f831;
rcp.rn.f32 %f832, %f787;
mul.f32 %f783, %f832, %f826;
mul.f32 %f784, %f832, %f827;
mul.f32 %f785, %f832, %f828;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r326, [root];
// inline asm
call _rt_trace_64, (%r326, %f11, %f12, %f13, %f783, %f784, %f785, %r259, %f373, %f787, %rd77, %r113);
// inline asm
ld.local.f32 %f833, [%rd16];
setp.neu.f32 %p74, %f833, 0f3F800000;
@%p74 bra BB0_69;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r333, [root];
mov.f32 %f841, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r333, %f120, %f121, %f122, %f74, %f75, %f76, %r259, %f373, %f841, %rd77, %r113);
// inline asm
ld.local.f32 %f842, [%rd16];
add.f32 %f1303, %f1303, %f842;
add.f32 %f1302, %f1302, 0f3F800000;
mov.u16 %rs94, 1;
BB0_69:
mad.lo.s32 %r340, %r90, 1664525, 1013904223;
and.b32 %r341, %r340, 16777215;
cvt.rn.f32.u32 %f851, %r341;
mul.f32 %f852, %f851, 0f33800000;
fma.rn.f32 %f853, %f852, 0f40000000, 0fBF800000;
mad.lo.s32 %r477, %r340, 1664525, 1013904223;
and.b32 %r342, %r477, 16777215;
cvt.rn.f32.u32 %f854, %r342;
mul.f32 %f855, %f854, 0f33800000;
fma.rn.f32 %f856, %f855, 0f40000000, 0fBF800000;
mul.f32 %f857, %f14, %f856;
mul.f32 %f858, %f15, %f856;
mul.f32 %f859, %f16, %f856;
fma.rn.f32 %f860, %f17, %f853, %f857;
fma.rn.f32 %f861, %f18, %f853, %f858;
fma.rn.f32 %f862, %f19, %f853, %f859;
mul.f32 %f863, %f861, %f861;
fma.rn.f32 %f864, %f860, %f860, %f863;
fma.rn.f32 %f865, %f862, %f862, %f864;
sqrt.rn.f32 %f866, %f865;
rcp.rn.f32 %f867, %f866;
mul.f32 %f868, %f860, %f867;
mul.f32 %f869, %f861, %f867;
mul.f32 %f870, %f862, %f867;
mul.f32 %f871, %f1298, %f870;
mul.f32 %f872, %f1299, %f869;
sub.f32 %f873, %f871, %f872;
mul.f32 %f874, %f1299, %f868;
mul.f32 %f875, %f1297, %f870;
sub.f32 %f876, %f874, %f875;
mul.f32 %f877, %f1297, %f869;
mul.f32 %f878, %f1298, %f868;
sub.f32 %f879, %f877, %f878;
mul.f32 %f880, %f868, 0f3F61E1D2;
mul.f32 %f881, %f869, 0f3F61E1D2;
mul.f32 %f882, %f870, 0f3F61E1D2;
fma.rn.f32 %f883, %f20, %f880, %f11;
fma.rn.f32 %f884, %f20, %f881, %f12;
fma.rn.f32 %f885, %f20, %f882, %f13;
mul.f32 %f886, %f873, 0fBF61E1E3;
mul.f32 %f887, %f876, 0fBF61E1E3;
mul.f32 %f888, %f879, 0fBF61E1E3;
fma.rn.f32 %f127, %f20, %f886, %f883;
fma.rn.f32 %f128, %f20, %f887, %f884;
fma.rn.f32 %f129, %f20, %f888, %f885;
sub.f32 %f889, %f127, %f11;
sub.f32 %f890, %f128, %f12;
sub.f32 %f891, %f129, %f13;
mul.f32 %f892, %f890, %f890;
fma.rn.f32 %f893, %f889, %f889, %f892;
fma.rn.f32 %f894, %f891, %f891, %f893;
sqrt.rn.f32 %f850, %f894;
rcp.rn.f32 %f895, %f850;
mul.f32 %f846, %f895, %f889;
mul.f32 %f847, %f895, %f890;
mul.f32 %f848, %f895, %f891;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r337, [root];
// inline asm
call _rt_trace_64, (%r337, %f11, %f12, %f13, %f846, %f847, %f848, %r259, %f373, %f850, %rd77, %r113);
// inline asm
ld.local.f32 %f896, [%rd16];
setp.neu.f32 %p75, %f896, 0f3F800000;
@%p75 bra BB0_71;
st.local.u32 [%rd16], %r266;
ld.global.u32 %r344, [root];
mov.f32 %f904, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r344, %f127, %f128, %f129, %f74, %f75, %f76, %r259, %f373, %f904, %rd77, %r113);
// inline asm
ld.local.f32 %f905, [%rd16];
add.f32 %f1303, %f1303, %f905;
add.f32 %f1302, %f1302, 0f3F800000;
mov.u16 %rs94, 1;
BB0_71:
setp.ne.s16 %p76, %rs94, 0;
@%p76 bra BB0_73;
add.u64 %rd93, %SP, 32;
cvta.to.local.u64 %rd94, %rd93;
st.local.u32 [%rd94], %r266;
ld.global.u32 %r348, [root];
mov.f32 %f913, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r348, %f11, %f12, %f13, %f74, %f75, %f76, %r259, %f373, %f913, %rd93, %r113);
// inline asm
ld.local.f32 %f914, [%rd94];
add.f32 %f1303, %f1303, %f914;
add.f32 %f1302, %f1302, 0f3F800000;
BB0_73:
ld.global.u32 %r451, [samples];
add.s32 %r454, %r454, 1;
setp.lt.s32 %p77, %r454, %r451;
@%p77 bra BB0_9;
BB0_74:
add.s32 %r452, %r452, 1;
setp.lt.s32 %p78, %r452, %r451;
@%p78 bra BB0_7;
bra.uni BB0_75;
BB0_148:
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 %f1294, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs88, %f1294;}
// inline asm
mov.u16 %rs89, 0;
st.v4.u16 [%rd204], {%rs88, %rs88, %rs88, %rs89};
BB0_149:
ld.global.u8 %rs90, [imageEnabled];
and.b16 %rs91, %rs90, 64;
setp.eq.s16 %p159, %rs91, 0;
@%p159 bra BB0_151;
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 %rs92, 0;
st.v4.u8 [%rd211], {%rs92, %rs92, %rs92, %rs92};
bra.uni BB0_151;
BB0_5:
mov.f32 %f1303, %f1302;
BB0_75:
setp.eq.s32 %p79, %r451, 0;
mov.f32 %f1338, 0f3F800000;
@%p79 bra BB0_77;
mov.f32 %f916, 0f3F800000;
max.f32 %f917, %f1302, %f916;
div.rn.f32 %f1338, %f1303, %f917;
BB0_77:
ld.global.f32 %f918, [directColor];
mul.f32 %f919, %f1338, %f918;
ld.global.f32 %f920, [directColor+4];
mul.f32 %f921, %f1338, %f920;
ld.global.f32 %f922, [directColor+8];
mul.f32 %f923, %f1338, %f922;
cvt.sat.f32.f32 %f924, %f10;
mul.f32 %f925, %f919, %f924;
mul.f32 %f926, %f921, %f924;
mul.f32 %f927, %f923, %f924;
ld.global.u32 %r352, [ignoreNormal];
setp.eq.s32 %p80, %r352, 0;
selp.f32 %f144, %f925, %f919, %p80;
selp.f32 %f145, %f926, %f921, %p80;
selp.f32 %f146, %f927, %f923, %p80;
ld.global.u32 %r480, [imageEnabled];
and.b32 %r353, %r480, 8;
setp.eq.s32 %p81, %r353, 0;
@%p81 bra BB0_90;
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 %f930, 0f3E68BA2E;
cvt.rzi.f32.f32 %f931, %f930;
fma.rn.f32 %f932, %f931, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f147, %f932;
abs.f32 %f148, %f1338;
setp.lt.f32 %p82, %f148, 0f00800000;
mul.f32 %f933, %f148, 0f4B800000;
selp.f32 %f934, 0fC3170000, 0fC2FE0000, %p82;
selp.f32 %f935, %f933, %f148, %p82;
mov.b32 %r356, %f935;
and.b32 %r357, %r356, 8388607;
or.b32 %r358, %r357, 1065353216;
mov.b32 %f936, %r358;
shr.u32 %r359, %r356, 23;
cvt.rn.f32.u32 %f937, %r359;
add.f32 %f938, %f934, %f937;
setp.gt.f32 %p83, %f936, 0f3FB504F3;
mul.f32 %f939, %f936, 0f3F000000;
add.f32 %f940, %f938, 0f3F800000;
selp.f32 %f941, %f939, %f936, %p83;
selp.f32 %f942, %f940, %f938, %p83;
add.f32 %f943, %f941, 0fBF800000;
add.f32 %f929, %f941, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f928,%f929;
// inline asm
add.f32 %f944, %f943, %f943;
mul.f32 %f945, %f928, %f944;
mul.f32 %f946, %f945, %f945;
mov.f32 %f947, 0f3C4CAF63;
mov.f32 %f948, 0f3B18F0FE;
fma.rn.f32 %f949, %f948, %f946, %f947;
mov.f32 %f950, 0f3DAAAABD;
fma.rn.f32 %f951, %f949, %f946, %f950;
mul.rn.f32 %f952, %f951, %f946;
mul.rn.f32 %f953, %f952, %f945;
sub.f32 %f954, %f943, %f945;
neg.f32 %f955, %f945;
add.f32 %f956, %f954, %f954;
fma.rn.f32 %f957, %f955, %f943, %f956;
mul.rn.f32 %f958, %f928, %f957;
add.f32 %f959, %f953, %f945;
sub.f32 %f960, %f945, %f959;
add.f32 %f961, %f953, %f960;
add.f32 %f962, %f958, %f961;
add.f32 %f963, %f959, %f962;
sub.f32 %f964, %f959, %f963;
add.f32 %f965, %f962, %f964;
mov.f32 %f966, 0f3F317200;
mul.rn.f32 %f967, %f942, %f966;
mov.f32 %f968, 0f35BFBE8E;
mul.rn.f32 %f969, %f942, %f968;
add.f32 %f970, %f967, %f963;
sub.f32 %f971, %f967, %f970;
add.f32 %f972, %f963, %f971;
add.f32 %f973, %f965, %f972;
add.f32 %f974, %f969, %f973;
add.f32 %f975, %f970, %f974;
sub.f32 %f976, %f970, %f975;
add.f32 %f977, %f974, %f976;
mov.f32 %f978, 0f3EE8BA2E;
mul.rn.f32 %f979, %f978, %f975;
neg.f32 %f980, %f979;
fma.rn.f32 %f981, %f978, %f975, %f980;
fma.rn.f32 %f982, %f978, %f977, %f981;
mov.f32 %f983, 0f00000000;
fma.rn.f32 %f984, %f983, %f975, %f982;
add.rn.f32 %f985, %f979, %f984;
neg.f32 %f986, %f985;
add.rn.f32 %f987, %f979, %f986;
add.rn.f32 %f988, %f987, %f984;
mov.b32 %r360, %f985;
setp.eq.s32 %p84, %r360, 1118925336;
add.s32 %r361, %r360, -1;
mov.b32 %f989, %r361;
add.f32 %f990, %f988, 0f37000000;
selp.f32 %f991, %f989, %f985, %p84;
selp.f32 %f149, %f990, %f988, %p84;
mul.f32 %f992, %f991, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f993, %f992;
mov.f32 %f994, 0fBF317200;
fma.rn.f32 %f995, %f993, %f994, %f991;
mov.f32 %f996, 0fB5BFBE8E;
fma.rn.f32 %f997, %f993, %f996, %f995;
mul.f32 %f998, %f997, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f999, %f998;
add.f32 %f1000, %f993, 0f00000000;
ex2.approx.f32 %f1001, %f1000;
mul.f32 %f1002, %f999, %f1001;
setp.lt.f32 %p85, %f991, 0fC2D20000;
selp.f32 %f1003, 0f00000000, %f1002, %p85;
setp.gt.f32 %p86, %f991, 0f42D20000;
selp.f32 %f1339, 0f7F800000, %f1003, %p86;
setp.eq.f32 %p87, %f1339, 0f7F800000;
@%p87 bra BB0_80;
fma.rn.f32 %f1339, %f1339, %f149, %f1339;
BB0_80:
setp.lt.f32 %p88, %f1338, 0f00000000;
setp.eq.f32 %p89, %f147, 0f3F800000;
and.pred %p1, %p88, %p89;
mov.b32 %r362, %f1339;
xor.b32 %r363, %r362, -2147483648;
mov.b32 %f1004, %r363;
selp.f32 %f1341, %f1004, %f1339, %p1;
setp.eq.f32 %p90, %f1338, 0f00000000;
@%p90 bra BB0_83;
bra.uni BB0_81;
BB0_83:
add.f32 %f1007, %f1338, %f1338;
selp.f32 %f1341, %f1007, 0f00000000, %p89;
bra.uni BB0_84;
BB0_138:
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 %f1284, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs69, %f1284;}
// inline asm
st.v4.u16 [%rd163], {%rs69, %rs69, %rs69, %rs58};
BB0_139:
ld.global.u8 %rs70, [imageEnabled];
and.b16 %rs71, %rs70, 64;
setp.eq.s16 %p154, %rs71, 0;
@%p154 bra BB0_151;
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 %rs72, 255;
mov.u16 %rs73, 0;
st.v4.u8 [%rd170], {%rs73, %rs73, %rs73, %rs72};
bra.uni BB0_151;
BB0_81:
setp.geu.f32 %p91, %f1338, 0f00000000;
@%p91 bra BB0_84;
cvt.rzi.f32.f32 %f1006, %f978;
setp.neu.f32 %p92, %f1006, 0f3EE8BA2E;
selp.f32 %f1341, 0f7FFFFFFF, %f1341, %p92;
BB0_84:
add.f32 %f1008, %f148, 0f3EE8BA2E;
mov.b32 %r364, %f1008;
setp.lt.s32 %p94, %r364, 2139095040;
@%p94 bra BB0_89;
setp.gtu.f32 %p95, %f148, 0f7F800000;
@%p95 bra BB0_88;
bra.uni BB0_86;
BB0_88:
add.f32 %f1341, %f1338, 0f3EE8BA2E;
bra.uni BB0_89;
BB0_86:
setp.neu.f32 %p96, %f148, 0f7F800000;
@%p96 bra BB0_89;
selp.f32 %f1341, 0fFF800000, 0f7F800000, %p1;
BB0_89:
mul.f32 %f1009, %f1341, 0f437F0000;
setp.eq.f32 %p97, %f1338, 0f3F800000;
selp.f32 %f1010, 0f437F0000, %f1009, %p97;
cvt.rzi.u32.f32 %r365, %f1010;
cvt.u16.u32 %rs28, %r365;
mov.u16 %rs29, 255;
st.v2.u8 [%rd95], {%rs28, %rs29};
ld.global.u32 %r480, [imageEnabled];
BB0_90:
and.b32 %r366, %r480, 1;
setp.eq.b32 %p98, %r366, 1;
@!%p98 bra BB0_125;
bra.uni BB0_91;
BB0_91:
mov.f32 %f1013, 0f3E666666;
cvt.rzi.f32.f32 %f1014, %f1013;
fma.rn.f32 %f1015, %f1014, 0fC0000000, 0f3EE66666;
abs.f32 %f160, %f1015;
abs.f32 %f161, %f144;
setp.lt.f32 %p99, %f161, 0f00800000;
mul.f32 %f1016, %f161, 0f4B800000;
selp.f32 %f1017, 0fC3170000, 0fC2FE0000, %p99;
selp.f32 %f1018, %f1016, %f161, %p99;
mov.b32 %r367, %f1018;
and.b32 %r368, %r367, 8388607;
or.b32 %r369, %r368, 1065353216;
mov.b32 %f1019, %r369;
shr.u32 %r370, %r367, 23;
cvt.rn.f32.u32 %f1020, %r370;
add.f32 %f1021, %f1017, %f1020;
setp.gt.f32 %p100, %f1019, 0f3FB504F3;
mul.f32 %f1022, %f1019, 0f3F000000;
add.f32 %f1023, %f1021, 0f3F800000;
selp.f32 %f1024, %f1022, %f1019, %p100;
selp.f32 %f1025, %f1023, %f1021, %p100;
add.f32 %f1026, %f1024, 0fBF800000;
add.f32 %f1012, %f1024, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1011,%f1012;
// inline asm
add.f32 %f1027, %f1026, %f1026;
mul.f32 %f1028, %f1011, %f1027;
mul.f32 %f1029, %f1028, %f1028;
mov.f32 %f1030, 0f3C4CAF63;
mov.f32 %f1031, 0f3B18F0FE;
fma.rn.f32 %f1032, %f1031, %f1029, %f1030;
mov.f32 %f1033, 0f3DAAAABD;
fma.rn.f32 %f1034, %f1032, %f1029, %f1033;
mul.rn.f32 %f1035, %f1034, %f1029;
mul.rn.f32 %f1036, %f1035, %f1028;
sub.f32 %f1037, %f1026, %f1028;
neg.f32 %f1038, %f1028;
add.f32 %f1039, %f1037, %f1037;
fma.rn.f32 %f1040, %f1038, %f1026, %f1039;
mul.rn.f32 %f1041, %f1011, %f1040;
add.f32 %f1042, %f1036, %f1028;
sub.f32 %f1043, %f1028, %f1042;
add.f32 %f1044, %f1036, %f1043;
add.f32 %f1045, %f1041, %f1044;
add.f32 %f1046, %f1042, %f1045;
sub.f32 %f1047, %f1042, %f1046;
add.f32 %f1048, %f1045, %f1047;
mov.f32 %f1049, 0f3F317200;
mul.rn.f32 %f1050, %f1025, %f1049;
mov.f32 %f1051, 0f35BFBE8E;
mul.rn.f32 %f1052, %f1025, %f1051;
add.f32 %f1053, %f1050, %f1046;
sub.f32 %f1054, %f1050, %f1053;
add.f32 %f1055, %f1046, %f1054;
add.f32 %f1056, %f1048, %f1055;
add.f32 %f1057, %f1052, %f1056;
add.f32 %f1058, %f1053, %f1057;
sub.f32 %f1059, %f1053, %f1058;
add.f32 %f1060, %f1057, %f1059;
mov.f32 %f1061, 0f3EE66666;
mul.rn.f32 %f1062, %f1061, %f1058;
neg.f32 %f1063, %f1062;
fma.rn.f32 %f1064, %f1061, %f1058, %f1063;
fma.rn.f32 %f1065, %f1061, %f1060, %f1064;
mov.f32 %f1066, 0f00000000;
fma.rn.f32 %f1067, %f1066, %f1058, %f1065;
add.rn.f32 %f1068, %f1062, %f1067;
neg.f32 %f1069, %f1068;
add.rn.f32 %f1070, %f1062, %f1069;
add.rn.f32 %f1071, %f1070, %f1067;
mov.b32 %r371, %f1068;
setp.eq.s32 %p101, %r371, 1118925336;
add.s32 %r372, %r371, -1;
mov.b32 %f1072, %r372;
add.f32 %f1073, %f1071, 0f37000000;
selp.f32 %f1074, %f1072, %f1068, %p101;
selp.f32 %f162, %f1073, %f1071, %p101;
mul.f32 %f1075, %f1074, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1076, %f1075;
mov.f32 %f1077, 0fBF317200;
fma.rn.f32 %f1078, %f1076, %f1077, %f1074;
mov.f32 %f1079, 0fB5BFBE8E;
fma.rn.f32 %f1080, %f1076, %f1079, %f1078;
mul.f32 %f1081, %f1080, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1082, %f1081;
add.f32 %f1083, %f1076, 0f00000000;
ex2.approx.f32 %f1084, %f1083;
mul.f32 %f1085, %f1082, %f1084;
setp.lt.f32 %p102, %f1074, 0fC2D20000;
selp.f32 %f1086, 0f00000000, %f1085, %p102;
setp.gt.f32 %p103, %f1074, 0f42D20000;
selp.f32 %f1342, 0f7F800000, %f1086, %p103;
setp.eq.f32 %p104, %f1342, 0f7F800000;
@%p104 bra BB0_93;
fma.rn.f32 %f1342, %f1342, %f162, %f1342;
BB0_93:
setp.lt.f32 %p105, %f144, 0f00000000;
setp.eq.f32 %p106, %f160, 0f3F800000;
and.pred %p2, %p105, %p106;
mov.b32 %r373, %f1342;
xor.b32 %r374, %r373, -2147483648;
mov.b32 %f1087, %r374;
selp.f32 %f1344, %f1087, %f1342, %p2;
setp.eq.f32 %p107, %f144, 0f00000000;
@%p107 bra BB0_96;
bra.uni BB0_94;
BB0_96:
add.f32 %f1090, %f144, %f144;
selp.f32 %f1344, %f1090, 0f00000000, %p106;
bra.uni BB0_97;
BB0_94:
setp.geu.f32 %p108, %f144, 0f00000000;
@%p108 bra BB0_97;
cvt.rzi.f32.f32 %f1089, %f1061;
setp.neu.f32 %p109, %f1089, 0f3EE66666;
selp.f32 %f1344, 0f7FFFFFFF, %f1344, %p109;
BB0_97:
add.f32 %f1091, %f161, 0f3EE66666;
mov.b32 %r375, %f1091;
setp.lt.s32 %p111, %r375, 2139095040;
@%p111 bra BB0_102;
setp.gtu.f32 %p112, %f161, 0f7F800000;
@%p112 bra BB0_101;
bra.uni BB0_99;
BB0_101:
add.f32 %f1344, %f144, 0f3EE66666;
bra.uni BB0_102;
BB0_99:
setp.neu.f32 %p113, %f161, 0f7F800000;
@%p113 bra BB0_102;
selp.f32 %f1344, 0fFF800000, 0f7F800000, %p2;
BB0_102:
setp.eq.f32 %p114, %f144, 0f3F800000;
selp.f32 %f173, 0f3F800000, %f1344, %p114;
abs.f32 %f174, %f145;
setp.lt.f32 %p115, %f174, 0f00800000;
mul.f32 %f1094, %f174, 0f4B800000;
selp.f32 %f1095, 0fC3170000, 0fC2FE0000, %p115;
selp.f32 %f1096, %f1094, %f174, %p115;
mov.b32 %r376, %f1096;
and.b32 %r377, %r376, 8388607;
or.b32 %r378, %r377, 1065353216;
mov.b32 %f1097, %r378;
shr.u32 %r379, %r376, 23;
cvt.rn.f32.u32 %f1098, %r379;
add.f32 %f1099, %f1095, %f1098;
setp.gt.f32 %p116, %f1097, 0f3FB504F3;
mul.f32 %f1100, %f1097, 0f3F000000;
add.f32 %f1101, %f1099, 0f3F800000;
selp.f32 %f1102, %f1100, %f1097, %p116;
selp.f32 %f1103, %f1101, %f1099, %p116;
add.f32 %f1104, %f1102, 0fBF800000;
add.f32 %f1093, %f1102, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1092,%f1093;
// inline asm
add.f32 %f1105, %f1104, %f1104;
mul.f32 %f1106, %f1092, %f1105;
mul.f32 %f1107, %f1106, %f1106;
fma.rn.f32 %f1110, %f1031, %f1107, %f1030;
fma.rn.f32 %f1112, %f1110, %f1107, %f1033;
mul.rn.f32 %f1113, %f1112, %f1107;
mul.rn.f32 %f1114, %f1113, %f1106;
sub.f32 %f1115, %f1104, %f1106;
neg.f32 %f1116, %f1106;
add.f32 %f1117, %f1115, %f1115;
fma.rn.f32 %f1118, %f1116, %f1104, %f1117;
mul.rn.f32 %f1119, %f1092, %f1118;
add.f32 %f1120, %f1114, %f1106;
sub.f32 %f1121, %f1106, %f1120;
add.f32 %f1122, %f1114, %f1121;
add.f32 %f1123, %f1119, %f1122;
add.f32 %f1124, %f1120, %f1123;
sub.f32 %f1125, %f1120, %f1124;
add.f32 %f1126, %f1123, %f1125;
mul.rn.f32 %f1128, %f1103, %f1049;
mul.rn.f32 %f1130, %f1103, %f1051;
add.f32 %f1131, %f1128, %f1124;
sub.f32 %f1132, %f1128, %f1131;
add.f32 %f1133, %f1124, %f1132;
add.f32 %f1134, %f1126, %f1133;
add.f32 %f1135, %f1130, %f1134;
add.f32 %f1136, %f1131, %f1135;
sub.f32 %f1137, %f1131, %f1136;
add.f32 %f1138, %f1135, %f1137;
mul.rn.f32 %f1140, %f1061, %f1136;
neg.f32 %f1141, %f1140;
fma.rn.f32 %f1142, %f1061, %f1136, %f1141;
fma.rn.f32 %f1143, %f1061, %f1138, %f1142;
fma.rn.f32 %f1145, %f1066, %f1136, %f1143;
add.rn.f32 %f1146, %f1140, %f1145;
neg.f32 %f1147, %f1146;
add.rn.f32 %f1148, %f1140, %f1147;
add.rn.f32 %f1149, %f1148, %f1145;
mov.b32 %r380, %f1146;
setp.eq.s32 %p117, %r380, 1118925336;
add.s32 %r381, %r380, -1;
mov.b32 %f1150, %r381;
add.f32 %f1151, %f1149, 0f37000000;
selp.f32 %f1152, %f1150, %f1146, %p117;
selp.f32 %f175, %f1151, %f1149, %p117;
mul.f32 %f1153, %f1152, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1154, %f1153;
fma.rn.f32 %f1156, %f1154, %f1077, %f1152;
fma.rn.f32 %f1158, %f1154, %f1079, %f1156;
mul.f32 %f1159, %f1158, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1160, %f1159;
add.f32 %f1161, %f1154, 0f00000000;
ex2.approx.f32 %f1162, %f1161;
mul.f32 %f1163, %f1160, %f1162;
setp.lt.f32 %p118, %f1152, 0fC2D20000;
selp.f32 %f1164, 0f00000000, %f1163, %p118;
setp.gt.f32 %p119, %f1152, 0f42D20000;
selp.f32 %f1345, 0f7F800000, %f1164, %p119;
setp.eq.f32 %p120, %f1345, 0f7F800000;
@%p120 bra BB0_104;
fma.rn.f32 %f1345, %f1345, %f175, %f1345;
BB0_104:
setp.lt.f32 %p121, %f145, 0f00000000;
and.pred %p3, %p121, %p106;
mov.b32 %r382, %f1345;
xor.b32 %r383, %r382, -2147483648;
mov.b32 %f1165, %r383;
selp.f32 %f1347, %f1165, %f1345, %p3;
setp.eq.f32 %p123, %f145, 0f00000000;
@%p123 bra BB0_107;
bra.uni BB0_105;
BB0_107:
add.f32 %f1168, %f145, %f145;
selp.f32 %f1347, %f1168, 0f00000000, %p106;
bra.uni BB0_108;
BB0_105:
setp.geu.f32 %p124, %f145, 0f00000000;
@%p124 bra BB0_108;
cvt.rzi.f32.f32 %f1167, %f1061;
setp.neu.f32 %p125, %f1167, 0f3EE66666;
selp.f32 %f1347, 0f7FFFFFFF, %f1347, %p125;
BB0_108:
add.f32 %f1169, %f174, 0f3EE66666;
mov.b32 %r384, %f1169;
setp.lt.s32 %p127, %r384, 2139095040;
@%p127 bra BB0_113;
setp.gtu.f32 %p128, %f174, 0f7F800000;
@%p128 bra BB0_112;
bra.uni BB0_110;
BB0_112:
add.f32 %f1347, %f145, 0f3EE66666;
bra.uni BB0_113;
BB0_110:
setp.neu.f32 %p129, %f174, 0f7F800000;
@%p129 bra BB0_113;
selp.f32 %f1347, 0fFF800000, 0f7F800000, %p3;
BB0_113:
setp.eq.f32 %p130, %f145, 0f3F800000;
selp.f32 %f186, 0f3F800000, %f1347, %p130;
abs.f32 %f187, %f146;
setp.lt.f32 %p131, %f187, 0f00800000;
mul.f32 %f1172, %f187, 0f4B800000;
selp.f32 %f1173, 0fC3170000, 0fC2FE0000, %p131;
selp.f32 %f1174, %f1172, %f187, %p131;
mov.b32 %r385, %f1174;
and.b32 %r386, %r385, 8388607;
or.b32 %r387, %r386, 1065353216;
mov.b32 %f1175, %r387;
shr.u32 %r388, %r385, 23;
cvt.rn.f32.u32 %f1176, %r388;
add.f32 %f1177, %f1173, %f1176;
setp.gt.f32 %p132, %f1175, 0f3FB504F3;
mul.f32 %f1178, %f1175, 0f3F000000;
add.f32 %f1179, %f1177, 0f3F800000;
selp.f32 %f1180, %f1178, %f1175, %p132;
selp.f32 %f1181, %f1179, %f1177, %p132;
add.f32 %f1182, %f1180, 0fBF800000;
add.f32 %f1171, %f1180, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1170,%f1171;
// inline asm
add.f32 %f1183, %f1182, %f1182;
mul.f32 %f1184, %f1170, %f1183;
mul.f32 %f1185, %f1184, %f1184;
fma.rn.f32 %f1188, %f1031, %f1185, %f1030;
fma.rn.f32 %f1190, %f1188, %f1185, %f1033;
mul.rn.f32 %f1191, %f1190, %f1185;
mul.rn.f32 %f1192, %f1191, %f1184;
sub.f32 %f1193, %f1182, %f1184;
neg.f32 %f1194, %f1184;
add.f32 %f1195, %f1193, %f1193;
fma.rn.f32 %f1196, %f1194, %f1182, %f1195;
mul.rn.f32 %f1197, %f1170, %f1196;
add.f32 %f1198, %f1192, %f1184;
sub.f32 %f1199, %f1184, %f1198;
add.f32 %f1200, %f1192, %f1199;
add.f32 %f1201, %f1197, %f1200;
add.f32 %f1202, %f1198, %f1201;
sub.f32 %f1203, %f1198, %f1202;
add.f32 %f1204, %f1201, %f1203;
mul.rn.f32 %f1206, %f1181, %f1049;
mul.rn.f32 %f1208, %f1181, %f1051;
add.f32 %f1209, %f1206, %f1202;
sub.f32 %f1210, %f1206, %f1209;
add.f32 %f1211, %f1202, %f1210;
add.f32 %f1212, %f1204, %f1211;
add.f32 %f1213, %f1208, %f1212;
add.f32 %f1214, %f1209, %f1213;
sub.f32 %f1215, %f1209, %f1214;
add.f32 %f1216, %f1213, %f1215;
mul.rn.f32 %f1218, %f1061, %f1214;
neg.f32 %f1219, %f1218;
fma.rn.f32 %f1220, %f1061, %f1214, %f1219;
fma.rn.f32 %f1221, %f1061, %f1216, %f1220;
fma.rn.f32 %f1223, %f1066, %f1214, %f1221;
add.rn.f32 %f1224, %f1218, %f1223;
neg.f32 %f1225, %f1224;
add.rn.f32 %f1226, %f1218, %f1225;
add.rn.f32 %f1227, %f1226, %f1223;
mov.b32 %r389, %f1224;
setp.eq.s32 %p133, %r389, 1118925336;
add.s32 %r390, %r389, -1;
mov.b32 %f1228, %r390;
add.f32 %f1229, %f1227, 0f37000000;
selp.f32 %f1230, %f1228, %f1224, %p133;
selp.f32 %f188, %f1229, %f1227, %p133;
mul.f32 %f1231, %f1230, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1232, %f1231;
fma.rn.f32 %f1234, %f1232, %f1077, %f1230;
fma.rn.f32 %f1236, %f1232, %f1079, %f1234;
mul.f32 %f1237, %f1236, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1238, %f1237;
add.f32 %f1239, %f1232, 0f00000000;
ex2.approx.f32 %f1240, %f1239;
mul.f32 %f1241, %f1238, %f1240;
setp.lt.f32 %p134, %f1230, 0fC2D20000;
selp.f32 %f1242, 0f00000000, %f1241, %p134;
setp.gt.f32 %p135, %f1230, 0f42D20000;
selp.f32 %f1348, 0f7F800000, %f1242, %p135;
setp.eq.f32 %p136, %f1348, 0f7F800000;
@%p136 bra BB0_115;
fma.rn.f32 %f1348, %f1348, %f188, %f1348;
BB0_115:
setp.lt.f32 %p137, %f146, 0f00000000;
and.pred %p4, %p137, %p106;
mov.b32 %r391, %f1348;
xor.b32 %r392, %r391, -2147483648;
mov.b32 %f1243, %r392;
selp.f32 %f1350, %f1243, %f1348, %p4;
setp.eq.f32 %p139, %f146, 0f00000000;
@%p139 bra BB0_118;
bra.uni BB0_116;
BB0_118:
add.f32 %f1246, %f146, %f146;
selp.f32 %f1350, %f1246, 0f00000000, %p106;
bra.uni BB0_119;
BB0_116:
setp.geu.f32 %p140, %f146, 0f00000000;
@%p140 bra BB0_119;
cvt.rzi.f32.f32 %f1245, %f1061;
setp.neu.f32 %p141, %f1245, 0f3EE66666;
selp.f32 %f1350, 0f7FFFFFFF, %f1350, %p141;
BB0_119:
add.f32 %f1247, %f187, 0f3EE66666;
mov.b32 %r393, %f1247;
setp.lt.s32 %p143, %r393, 2139095040;
@%p143 bra BB0_124;
setp.gtu.f32 %p144, %f187, 0f7F800000;
@%p144 bra BB0_123;
bra.uni BB0_121;
BB0_123:
add.f32 %f1350, %f146, 0f3EE66666;
bra.uni BB0_124;
BB0_121:
setp.neu.f32 %p145, %f187, 0f7F800000;
@%p145 bra BB0_124;
selp.f32 %f1350, 0fFF800000, 0f7F800000, %p4;
BB0_124:
mov.u32 %r447, 2;
setp.eq.f32 %p146, %f146, 0f3F800000;
selp.f32 %f1248, 0f3F800000, %f1350, %p146;
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, %r113, %rd104, %rd105, %rd29, %rd29);
// inline asm
cvt.sat.f32.f32 %f1249, %f1248;
mul.f32 %f1250, %f1249, 0f437FFD71;
cvt.rzi.u32.f32 %r396, %f1250;
cvt.sat.f32.f32 %f1251, %f186;
mul.f32 %f1252, %f1251, 0f437FFD71;
cvt.rzi.u32.f32 %r397, %f1252;
cvt.sat.f32.f32 %f1253, %f173;
mul.f32 %f1254, %f1253, 0f437FFD71;
cvt.rzi.u32.f32 %r398, %f1254;
cvt.u16.u32 %rs30, %r396;
cvt.u16.u32 %rs31, %r398;
cvt.u16.u32 %rs32, %r397;
mov.u16 %rs33, 255;
st.v4.u8 [%rd102], {%rs30, %rs32, %rs31, %rs33};
ld.global.u32 %r480, [imageEnabled];
BB0_125:
and.b32 %r399, %r480, 4;
setp.eq.s32 %p147, %r399, 0;
@%p147 bra BB0_129;
ld.global.u32 %r400, [additive];
setp.eq.s32 %p148, %r400, 0;
cvt.u64.u32 %rd18, %r2;
cvt.u64.u32 %rd19, %r3;
mov.f32 %f1255, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs34, %f1255;}
// inline asm
@%p148 bra BB0_128;
mov.u32 %r448, 2;
mov.u64 %rd121, image_HDR;
cvta.global.u64 %rd110, %rd121;
mov.u32 %r404, 8;
// inline asm
call (%rd109), _rt_buffer_get_64, (%rd110, %r448, %r404, %rd18, %rd19, %rd29, %rd29);
// inline asm
ld.v4.u16 {%rs41, %rs42, %rs43, %rs44}, [%rd109];
// inline asm
{ cvt.f32.f16 %f1256, %rs41;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1257, %rs42;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1258, %rs43;}
// inline asm
// inline asm
call (%rd115), _rt_buffer_get_64, (%rd110, %r448, %r404, %rd18, %rd19, %rd29, %rd29);
// inline asm
add.f32 %f1259, %f144, %f1256;
add.f32 %f1260, %f145, %f1257;
add.f32 %f1261, %f146, %f1258;
// inline asm
{ cvt.rn.f16.f32 %rs40, %f1261;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs39, %f1260;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs38, %f1259;}
// inline asm
st.v4.u16 [%rd115], {%rs38, %rs39, %rs40, %rs34};
bra.uni BB0_129;
BB0_128:
mov.u32 %r449, 2;
mov.u64 %rd128, image_HDR;
cvta.global.u64 %rd123, %rd128;
mov.u32 %r406, 8;
// inline asm
call (%rd122), _rt_buffer_get_64, (%rd123, %r449, %r406, %rd18, %rd19, %rd29, %rd29);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs47, %f146;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs46, %f145;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs45, %f144;}
// inline asm
st.v4.u16 [%rd122], {%rs45, %rs46, %rs47, %rs34};
BB0_129:
ld.global.u8 %rs48, [imageEnabled];
and.b16 %rs49, %rs48, 64;
setp.eq.s16 %p149, %rs49, 0;
@%p149 bra BB0_151;
mov.u32 %r450, 2;
ld.global.f32 %f1265, [directDir];
ld.global.f32 %f1266, [directDir+4];
ld.global.f32 %f1267, [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, %r450, %r113, %rd131, %rd132, %rd29, %rd29);
// inline asm
fma.rn.f32 %f1268, %f1265, 0fBF000000, 0f3F000000;
mul.f32 %f1269, %f1268, 0f437F0000;
cvt.rzi.u32.f32 %r409, %f1269;
fma.rn.f32 %f1270, %f1266, 0fBF000000, 0f3F000000;
mul.f32 %f1271, %f1270, 0f437F0000;
cvt.rzi.u32.f32 %r410, %f1271;
fma.rn.f32 %f1272, %f1267, 0fBF000000, 0f3F000000;
mul.f32 %f1273, %f1272, 0f437F0000;
cvt.rzi.u32.f32 %r411, %f1273;
cvt.u16.u32 %rs50, %r411;
cvt.u16.u32 %rs51, %r410;
cvt.u16.u32 %rs52, %r409;
mov.u16 %rs53, 255;
st.v4.u8 [%rd129], {%rs52, %rs51, %rs50, %rs53};
BB0_151:
ret;
}
Reference in New Issue View Git Blame Copy Permalink