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

2179 lines
60 KiB
Plaintext

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-23083092
// Cuda compilation tools, release 9.1, V9.1.85
// Based on LLVM 3.4svn
//
.version 6.1
.target sm_30
.address_size 64
// .globl _Z6oxMainv
.global .align 8 .b8 pixelID[8];
.global .align 8 .b8 resolution[8];
.global .align 4 .b8 normal[12];
.global .align 4 .b8 camPos[12];
.global .align 4 .b8 root[4];
.global .align 4 .u32 imageEnabled;
.global .texref lightmap;
.global .align 16 .b8 tileInfo[16];
.global .align 4 .u32 additive;
.global .align 1 .b8 image[1];
.global .align 1 .b8 image_HDR[1];
.global .align 1 .b8 image_HDR2[1];
.global .align 1 .b8 image_Mask[1];
.global .align 1 .b8 image_RNM0[1];
.global .align 1 .b8 image_RNM1[1];
.global .align 1 .b8 image_RNM2[1];
.global .align 1 .b8 image_RNM3[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 .u32 lightCookie;
.global .align 16 .b8 lightTilingOffset[16];
.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 4 .b8 _ZN21rti_internal_typeinfo11lightCookieE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo17lightTilingOffsetE[8] = {82, 97, 121, 0, 16, 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 .b8 _ZN21rti_internal_typename11lightCookieE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename17lightTilingOffsetE[7] = {102, 108, 111, 97, 116, 52, 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 4 .u32 _ZN21rti_internal_typeenum11lightCookieE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum17lightTilingOffsetE = 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 _ZN21rti_internal_semantic11lightCookieE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic17lightTilingOffsetE[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];
.global .align 1 .b8 _ZN23rti_internal_annotation11lightCookieE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation17lightTilingOffsetE[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[40];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<124>;
.reg .b16 %rs<152>;
.reg .f32 %f<853>;
.reg .b32 %r<390>;
.reg .b64 %rd<297>;
mov.u64 %rd296, __local_depot0;
cvta.local.u64 %SP, %rd296;
ld.global.v2.u32 {%r101, %r102}, [pixelID];
cvt.u64.u32 %rd23, %r101;
cvt.u64.u32 %rd24, %r102;
mov.u64 %rd27, uvnormal;
cvta.global.u64 %rd22, %rd27;
mov.u32 %r99, 2;
mov.u32 %r100, 4;
mov.u64 %rd26, 0;
// inline asm
call (%rd21), _rt_buffer_get_64, (%rd22, %r99, %r100, %rd23, %rd24, %rd26, %rd26);
// inline asm
ld.u32 %r1, [%rd21];
shr.u32 %r105, %r1, 16;
cvt.u16.u32 %rs1, %r105;
and.b16 %rs7, %rs1, 255;
cvt.u16.u32 %rs8, %r1;
or.b16 %rs9, %rs8, %rs7;
setp.eq.s16 %p5, %rs9, 0;
mov.f32 %f811, 0f00000000;
mov.f32 %f812, %f811;
mov.f32 %f813, %f811;
@%p5 bra BB0_2;
ld.u8 %rs10, [%rd21+1];
and.b16 %rs12, %rs8, 255;
cvt.rn.f32.u16 %f152, %rs12;
div.rn.f32 %f153, %f152, 0f437F0000;
fma.rn.f32 %f154, %f153, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f155, %rs10;
div.rn.f32 %f156, %f155, 0f437F0000;
fma.rn.f32 %f157, %f156, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f158, %rs7;
div.rn.f32 %f159, %f158, 0f437F0000;
fma.rn.f32 %f160, %f159, 0f40000000, 0fBF800000;
mul.f32 %f161, %f157, %f157;
fma.rn.f32 %f162, %f154, %f154, %f161;
fma.rn.f32 %f163, %f160, %f160, %f162;
sqrt.rn.f32 %f164, %f163;
rcp.rn.f32 %f165, %f164;
mul.f32 %f811, %f154, %f165;
mul.f32 %f812, %f157, %f165;
mul.f32 %f813, %f160, %f165;
BB0_2:
ld.global.v2.u32 {%r106, %r107}, [pixelID];
ld.global.v2.u32 {%r109, %r110}, [tileInfo];
add.s32 %r2, %r106, %r109;
add.s32 %r3, %r107, %r110;
setp.eq.f32 %p6, %f812, 0f00000000;
setp.eq.f32 %p7, %f811, 0f00000000;
and.pred %p8, %p7, %p6;
setp.eq.f32 %p9, %f813, 0f00000000;
and.pred %p10, %p8, %p9;
@%p10 bra BB0_122;
bra.uni BB0_3;
BB0_122:
ld.global.u32 %r389, [imageEnabled];
and.b32 %r314, %r389, 1;
setp.eq.b32 %p116, %r314, 1;
@!%p116 bra BB0_124;
bra.uni BB0_123;
BB0_123:
cvt.u64.u32 %rd180, %r2;
cvt.u64.u32 %rd181, %r3;
mov.u64 %rd184, image;
cvta.global.u64 %rd179, %rd184;
mov.u64 %rd183, 0;
// inline asm
call (%rd178), _rt_buffer_get_64, (%rd179, %r99, %r100, %rd180, %rd181, %rd183, %rd183);
// inline asm
mov.u16 %rs84, 0;
st.v4.u8 [%rd178], {%rs84, %rs84, %rs84, %rs84};
ld.global.u32 %r389, [imageEnabled];
BB0_124:
and.b32 %r317, %r389, 8;
setp.eq.s32 %p117, %r317, 0;
@%p117 bra BB0_126;
cvt.u64.u32 %rd188, %r3;
cvt.u64.u32 %rd187, %r2;
mov.u64 %rd191, image_Mask;
cvta.global.u64 %rd186, %rd191;
mov.u64 %rd190, 0;
// inline asm
call (%rd185), _rt_buffer_get_64, (%rd186, %r99, %r99, %rd187, %rd188, %rd190, %rd190);
// inline asm
mov.f32 %f732, 0f00000000;
cvt.rzi.u32.f32 %r320, %f732;
cvt.u16.u32 %rs85, %r320;
mov.u16 %rs86, 0;
st.v2.u8 [%rd185], {%rs85, %rs86};
ld.global.u32 %r389, [imageEnabled];
BB0_126:
cvt.u64.u32 %rd19, %r2;
cvt.u64.u32 %rd20, %r3;
and.b32 %r321, %r389, 4;
setp.eq.s32 %p118, %r321, 0;
@%p118 bra BB0_130;
ld.global.u32 %r322, [additive];
setp.eq.s32 %p119, %r322, 0;
@%p119 bra BB0_129;
mov.u64 %rd204, image_HDR;
cvta.global.u64 %rd193, %rd204;
mov.u32 %r326, 8;
mov.u64 %rd203, 0;
// inline asm
call (%rd192), _rt_buffer_get_64, (%rd193, %r99, %r326, %rd19, %rd20, %rd203, %rd203);
// inline asm
ld.v4.u16 {%rs93, %rs94, %rs95, %rs96}, [%rd192];
// inline asm
{ cvt.f32.f16 %f733, %rs93;}
// inline asm
// inline asm
{ cvt.f32.f16 %f734, %rs94;}
// inline asm
// inline asm
{ cvt.f32.f16 %f735, %rs95;}
// inline asm
// inline asm
call (%rd198), _rt_buffer_get_64, (%rd193, %r99, %r326, %rd19, %rd20, %rd203, %rd203);
// inline asm
add.f32 %f736, %f733, 0f00000000;
add.f32 %f737, %f734, 0f00000000;
add.f32 %f738, %f735, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs92, %f738;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs91, %f737;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs90, %f736;}
// inline asm
mov.u16 %rs97, 0;
st.v4.u16 [%rd198], {%rs90, %rs91, %rs92, %rs97};
bra.uni BB0_130;
BB0_3:
ld.global.f32 %f9, [directDir+8];
ld.global.f32 %f8, [directDir+4];
ld.global.f32 %f7, [directDir];
mul.f32 %f169, %f811, %f7;
mul.f32 %f170, %f812, %f8;
neg.f32 %f171, %f170;
sub.f32 %f172, %f171, %f169;
mul.f32 %f173, %f813, %f9;
sub.f32 %f10, %f172, %f173;
ld.global.v2.u32 {%r118, %r119}, [pixelID];
cvt.u64.u32 %rd30, %r118;
cvt.u64.u32 %rd31, %r119;
mov.u64 %rd40, uvpos;
cvta.global.u64 %rd29, %rd40;
mov.u32 %r115, 12;
// inline asm
call (%rd28), _rt_buffer_get_64, (%rd29, %r99, %r115, %rd30, %rd31, %rd26, %rd26);
// inline asm
ld.f32 %f13, [%rd28+8];
ld.f32 %f12, [%rd28+4];
ld.f32 %f11, [%rd28];
mul.f32 %f174, %f11, 0f3456BF95;
mul.f32 %f175, %f12, 0f3456BF95;
mul.f32 %f176, %f13, 0f3456BF95;
abs.f32 %f177, %f811;
div.rn.f32 %f178, %f174, %f177;
abs.f32 %f179, %f812;
div.rn.f32 %f180, %f175, %f179;
abs.f32 %f181, %f813;
div.rn.f32 %f182, %f176, %f181;
abs.f32 %f183, %f178;
abs.f32 %f184, %f180;
abs.f32 %f185, %f182;
mov.f32 %f186, 0f38D1B717;
max.f32 %f187, %f183, %f186;
max.f32 %f188, %f184, %f186;
max.f32 %f189, %f185, %f186;
fma.rn.f32 %f14, %f811, %f187, %f11;
fma.rn.f32 %f15, %f812, %f188, %f12;
fma.rn.f32 %f16, %f813, %f189, %f13;
abs.f32 %f190, %f7;
abs.f32 %f191, %f9;
setp.gt.f32 %p11, %f190, %f191;
neg.f32 %f192, %f8;
neg.f32 %f193, %f9;
selp.f32 %f194, %f192, 0f00000000, %p11;
selp.f32 %f195, %f7, %f193, %p11;
selp.f32 %f196, 0f00000000, %f8, %p11;
mul.f32 %f197, %f195, %f195;
fma.rn.f32 %f198, %f194, %f194, %f197;
fma.rn.f32 %f199, %f196, %f196, %f198;
sqrt.rn.f32 %f200, %f199;
rcp.rn.f32 %f201, %f200;
mul.f32 %f17, %f194, %f201;
mul.f32 %f18, %f195, %f201;
mul.f32 %f19, %f196, %f201;
mul.f32 %f202, %f9, %f18;
mul.f32 %f203, %f8, %f19;
sub.f32 %f20, %f202, %f203;
mul.f32 %f204, %f7, %f19;
mul.f32 %f205, %f9, %f17;
sub.f32 %f21, %f204, %f205;
mul.f32 %f206, %f8, %f17;
mul.f32 %f207, %f7, %f18;
sub.f32 %f22, %f206, %f207;
ld.global.v2.u32 {%r122, %r123}, [pixelID];
cvt.u64.u32 %rd36, %r122;
cvt.u64.u32 %rd37, %r123;
mov.u64 %rd41, rnd_seeds;
cvta.global.u64 %rd35, %rd41;
// inline asm
call (%rd34), _rt_buffer_get_64, (%rd35, %r99, %r100, %rd36, %rd37, %rd26, %rd26);
// inline asm
ld.global.u32 %r358, [samples];
mov.f32 %f832, 0f00000000;
setp.lt.s32 %p12, %r358, 1;
@%p12 bra BB0_4;
cvt.rn.f32.s32 %f211, %r358;
rcp.rn.f32 %f23, %f211;
ld.u32 %r384, [%rd34];
mul.f32 %f24, %f14, 0f3456BF95;
mul.f32 %f25, %f15, 0f3456BF95;
mul.f32 %f26, %f16, 0f3456BF95;
mov.f32 %f832, 0f00000000;
mov.u32 %r359, 0;
abs.f32 %f212, %f25;
abs.f32 %f213, %f24;
max.f32 %f214, %f213, %f212;
abs.f32 %f215, %f26;
max.f32 %f216, %f214, %f215;
mov.f32 %f833, %f832;
mov.f32 %f834, %f832;
BB0_6:
setp.lt.s32 %p13, %r358, 1;
@%p13 bra BB0_55;
cvt.rn.f32.s32 %f30, %r359;
max.f32 %f31, %f216, %f186;
mov.u32 %r361, 0;
BB0_8:
mad.lo.s32 %r128, %r384, 1664525, 1013904223;
and.b32 %r129, %r128, 16777215;
cvt.rn.f32.u32 %f218, %r129;
fma.rn.f32 %f219, %f218, 0f33800000, %f30;
mul.f32 %f220, %f23, %f219;
mad.lo.s32 %r384, %r128, 1664525, 1013904223;
and.b32 %r130, %r384, 16777215;
cvt.rn.f32.u32 %f221, %r130;
cvt.rn.f32.s32 %f222, %r361;
fma.rn.f32 %f223, %f221, 0f33800000, %f222;
mul.f32 %f224, %f23, %f223;
sqrt.rn.f32 %f35, %f220;
mul.f32 %f826, %f224, 0f40C90FDB;
abs.f32 %f37, %f826;
setp.neu.f32 %p14, %f37, 0f7F800000;
mov.f32 %f820, %f826;
@%p14 bra BB0_10;
mov.f32 %f225, 0f00000000;
mul.rn.f32 %f820, %f826, %f225;
BB0_10:
mul.f32 %f226, %f820, 0f3F22F983;
cvt.rni.s32.f32 %r372, %f226;
cvt.rn.f32.s32 %f227, %r372;
neg.f32 %f228, %f227;
mov.f32 %f229, 0f3FC90FDA;
fma.rn.f32 %f230, %f228, %f229, %f820;
mov.f32 %f231, 0f33A22168;
fma.rn.f32 %f232, %f228, %f231, %f230;
mov.f32 %f233, 0f27C234C5;
fma.rn.f32 %f821, %f228, %f233, %f232;
abs.f32 %f234, %f820;
setp.leu.f32 %p15, %f234, 0f47CE4780;
@%p15 bra BB0_21;
add.u64 %rd43, %SP, 12;
cvta.to.local.u64 %rd3, %rd43;
mov.b32 %r13, %f820;
shr.u32 %r14, %r13, 23;
shl.b32 %r133, %r13, 8;
or.b32 %r15, %r133, -2147483648;
mov.u32 %r363, 0;
mov.u64 %rd293, 0;
mov.u64 %rd292, %rd3;
mov.u32 %r364, %r363;
BB0_12:
.pragma "nounroll";
shl.b64 %rd44, %rd293, 2;
mov.u64 %rd45, __cudart_i2opi_f;
add.s64 %rd46, %rd45, %rd44;
ld.const.u32 %r136, [%rd46];
// inline asm
{
mad.lo.cc.u32 %r134, %r136, %r15, %r364;
madc.hi.u32 %r364, %r136, %r15, 0;
}
// inline asm
st.local.u32 [%rd292], %r134;
add.s32 %r363, %r363, 1;
cvt.s64.s32 %rd293, %r363;
mul.wide.s32 %rd49, %r363, 4;
add.s64 %rd292, %rd3, %rd49;
setp.ne.s32 %p16, %r363, 6;
@%p16 bra BB0_12;
and.b32 %r139, %r14, 255;
add.s32 %r140, %r139, -128;
shr.u32 %r141, %r140, 5;
and.b32 %r20, %r13, -2147483648;
cvta.to.local.u64 %rd51, %rd43;
st.local.u32 [%rd51+24], %r364;
mov.u32 %r142, 6;
sub.s32 %r143, %r142, %r141;
mul.wide.s32 %rd52, %r143, 4;
add.s64 %rd8, %rd51, %rd52;
ld.local.u32 %r365, [%rd8];
ld.local.u32 %r366, [%rd8+-4];
and.b32 %r23, %r14, 31;
setp.eq.s32 %p17, %r23, 0;
@%p17 bra BB0_15;
mov.u32 %r144, 32;
sub.s32 %r145, %r144, %r23;
shr.u32 %r146, %r366, %r145;
shl.b32 %r147, %r365, %r23;
add.s32 %r365, %r146, %r147;
ld.local.u32 %r148, [%rd8+-8];
shr.u32 %r149, %r148, %r145;
shl.b32 %r150, %r366, %r23;
add.s32 %r366, %r149, %r150;
BB0_15:
shr.u32 %r151, %r366, 30;
shl.b32 %r152, %r365, 2;
add.s32 %r367, %r151, %r152;
shl.b32 %r29, %r366, 2;
shr.u32 %r153, %r367, 31;
shr.u32 %r154, %r365, 30;
add.s32 %r30, %r153, %r154;
setp.eq.s32 %p18, %r153, 0;
@%p18 bra BB0_16;
bra.uni BB0_17;
BB0_16:
mov.u32 %r368, %r20;
mov.u32 %r369, %r29;
bra.uni BB0_18;
BB0_17:
not.b32 %r155, %r367;
neg.s32 %r369, %r29;
setp.eq.s32 %p19, %r29, 0;
selp.u32 %r156, 1, 0, %p19;
add.s32 %r367, %r156, %r155;
xor.b32 %r368, %r20, -2147483648;
BB0_18:
clz.b32 %r371, %r367;
setp.eq.s32 %p20, %r371, 0;
shl.b32 %r157, %r367, %r371;
mov.u32 %r158, 32;
sub.s32 %r159, %r158, %r371;
shr.u32 %r160, %r369, %r159;
add.s32 %r161, %r160, %r157;
selp.b32 %r38, %r367, %r161, %p20;
mov.u32 %r162, -921707870;
mul.hi.u32 %r370, %r38, %r162;
setp.eq.s32 %p21, %r20, 0;
neg.s32 %r163, %r30;
selp.b32 %r372, %r30, %r163, %p21;
setp.lt.s32 %p22, %r370, 1;
@%p22 bra BB0_20;
mul.lo.s32 %r164, %r38, -921707870;
shr.u32 %r165, %r164, 31;
shl.b32 %r166, %r370, 1;
add.s32 %r370, %r165, %r166;
add.s32 %r371, %r371, 1;
BB0_20:
mov.u32 %r167, 126;
sub.s32 %r168, %r167, %r371;
shl.b32 %r169, %r168, 23;
add.s32 %r170, %r370, 1;
shr.u32 %r171, %r170, 7;
add.s32 %r172, %r171, 1;
shr.u32 %r173, %r172, 1;
add.s32 %r174, %r173, %r169;
or.b32 %r175, %r174, %r368;
mov.b32 %f821, %r175;
BB0_21:
mul.rn.f32 %f43, %f821, %f821;
add.s32 %r46, %r372, 1;
and.b32 %r47, %r46, 1;
setp.eq.s32 %p23, %r47, 0;
@%p23 bra BB0_23;
bra.uni BB0_22;
BB0_23:
mov.f32 %f237, 0f3C08839E;
mov.f32 %f238, 0fB94CA1F9;
fma.rn.f32 %f822, %f238, %f43, %f237;
bra.uni BB0_24;
BB0_22:
mov.f32 %f235, 0fBAB6061A;
mov.f32 %f236, 0f37CCF5CE;
fma.rn.f32 %f822, %f236, %f43, %f235;
BB0_24:
@%p23 bra BB0_26;
bra.uni BB0_25;
BB0_26:
mov.f32 %f242, 0fBE2AAAA3;
fma.rn.f32 %f243, %f822, %f43, %f242;
mov.f32 %f244, 0f00000000;
fma.rn.f32 %f823, %f243, %f43, %f244;
bra.uni BB0_27;
BB0_25:
mov.f32 %f239, 0f3D2AAAA5;
fma.rn.f32 %f240, %f822, %f43, %f239;
mov.f32 %f241, 0fBF000000;
fma.rn.f32 %f823, %f240, %f43, %f241;
BB0_27:
fma.rn.f32 %f824, %f823, %f821, %f821;
@%p23 bra BB0_29;
mov.f32 %f245, 0f3F800000;
fma.rn.f32 %f824, %f823, %f43, %f245;
BB0_29:
and.b32 %r176, %r46, 2;
setp.eq.s32 %p26, %r176, 0;
@%p26 bra BB0_31;
mov.f32 %f246, 0f00000000;
mov.f32 %f247, 0fBF800000;
fma.rn.f32 %f824, %f824, %f247, %f246;
BB0_31:
@%p14 bra BB0_33;
mov.f32 %f248, 0f00000000;
mul.rn.f32 %f826, %f826, %f248;
BB0_33:
mul.f32 %f249, %f826, 0f3F22F983;
cvt.rni.s32.f32 %r382, %f249;
cvt.rn.f32.s32 %f250, %r382;
neg.f32 %f251, %f250;
fma.rn.f32 %f253, %f251, %f229, %f826;
fma.rn.f32 %f255, %f251, %f231, %f253;
fma.rn.f32 %f827, %f251, %f233, %f255;
abs.f32 %f257, %f826;
setp.leu.f32 %p28, %f257, 0f47CE4780;
@%p28 bra BB0_44;
add.u64 %rd54, %SP, 12;
cvta.to.local.u64 %rd9, %rd54;
mov.b32 %r49, %f826;
shr.u32 %r50, %r49, 23;
shl.b32 %r179, %r49, 8;
or.b32 %r51, %r179, -2147483648;
mov.u32 %r373, 0;
mov.u64 %rd294, %rd9;
mov.u64 %rd295, %rd26;
mov.u32 %r374, %r373;
BB0_35:
.pragma "nounroll";
shl.b64 %rd55, %rd295, 2;
mov.u64 %rd56, __cudart_i2opi_f;
add.s64 %rd57, %rd56, %rd55;
ld.const.u32 %r182, [%rd57];
// inline asm
{
mad.lo.cc.u32 %r180, %r182, %r51, %r374;
madc.hi.u32 %r374, %r182, %r51, 0;
}
// inline asm
st.local.u32 [%rd294], %r180;
add.s32 %r373, %r373, 1;
cvt.s64.s32 %rd295, %r373;
mul.wide.s32 %rd58, %r373, 4;
add.s64 %rd294, %rd9, %rd58;
setp.ne.s32 %p29, %r373, 6;
@%p29 bra BB0_35;
and.b32 %r185, %r50, 255;
add.s32 %r186, %r185, -128;
shr.u32 %r187, %r186, 5;
and.b32 %r56, %r49, -2147483648;
cvta.to.local.u64 %rd60, %rd54;
st.local.u32 [%rd60+24], %r374;
mov.u32 %r188, 6;
sub.s32 %r189, %r188, %r187;
mul.wide.s32 %rd61, %r189, 4;
add.s64 %rd15, %rd60, %rd61;
ld.local.u32 %r375, [%rd15];
ld.local.u32 %r376, [%rd15+-4];
and.b32 %r59, %r50, 31;
setp.eq.s32 %p30, %r59, 0;
@%p30 bra BB0_38;
mov.u32 %r190, 32;
sub.s32 %r191, %r190, %r59;
shr.u32 %r192, %r376, %r191;
shl.b32 %r193, %r375, %r59;
add.s32 %r375, %r192, %r193;
ld.local.u32 %r194, [%rd15+-8];
shr.u32 %r195, %r194, %r191;
shl.b32 %r196, %r376, %r59;
add.s32 %r376, %r195, %r196;
BB0_38:
shr.u32 %r197, %r376, 30;
shl.b32 %r198, %r375, 2;
add.s32 %r377, %r197, %r198;
shl.b32 %r65, %r376, 2;
shr.u32 %r199, %r377, 31;
shr.u32 %r200, %r375, 30;
add.s32 %r66, %r199, %r200;
setp.eq.s32 %p31, %r199, 0;
@%p31 bra BB0_39;
bra.uni BB0_40;
BB0_39:
mov.u32 %r378, %r56;
mov.u32 %r379, %r65;
bra.uni BB0_41;
BB0_40:
not.b32 %r201, %r377;
neg.s32 %r379, %r65;
setp.eq.s32 %p32, %r65, 0;
selp.u32 %r202, 1, 0, %p32;
add.s32 %r377, %r202, %r201;
xor.b32 %r378, %r56, -2147483648;
BB0_41:
clz.b32 %r381, %r377;
setp.eq.s32 %p33, %r381, 0;
shl.b32 %r203, %r377, %r381;
mov.u32 %r204, 32;
sub.s32 %r205, %r204, %r381;
shr.u32 %r206, %r379, %r205;
add.s32 %r207, %r206, %r203;
selp.b32 %r74, %r377, %r207, %p33;
mov.u32 %r208, -921707870;
mul.hi.u32 %r380, %r74, %r208;
setp.eq.s32 %p34, %r56, 0;
neg.s32 %r209, %r66;
selp.b32 %r382, %r66, %r209, %p34;
setp.lt.s32 %p35, %r380, 1;
@%p35 bra BB0_43;
mul.lo.s32 %r210, %r74, -921707870;
shr.u32 %r211, %r210, 31;
shl.b32 %r212, %r380, 1;
add.s32 %r380, %r211, %r212;
add.s32 %r381, %r381, 1;
BB0_43:
mov.u32 %r213, 126;
sub.s32 %r214, %r213, %r381;
shl.b32 %r215, %r214, 23;
add.s32 %r216, %r380, 1;
shr.u32 %r217, %r216, 7;
add.s32 %r218, %r217, 1;
shr.u32 %r219, %r218, 1;
add.s32 %r220, %r219, %r215;
or.b32 %r221, %r220, %r378;
mov.b32 %f827, %r221;
BB0_44:
mul.rn.f32 %f60, %f827, %f827;
and.b32 %r82, %r382, 1;
setp.eq.s32 %p36, %r82, 0;
@%p36 bra BB0_46;
bra.uni BB0_45;
BB0_46:
mov.f32 %f260, 0f3C08839E;
mov.f32 %f261, 0fB94CA1F9;
fma.rn.f32 %f828, %f261, %f60, %f260;
bra.uni BB0_47;
BB0_45:
mov.f32 %f258, 0fBAB6061A;
mov.f32 %f259, 0f37CCF5CE;
fma.rn.f32 %f828, %f259, %f60, %f258;
BB0_47:
@%p36 bra BB0_49;
bra.uni BB0_48;
BB0_49:
mov.f32 %f265, 0fBE2AAAA3;
fma.rn.f32 %f266, %f828, %f60, %f265;
mov.f32 %f267, 0f00000000;
fma.rn.f32 %f829, %f266, %f60, %f267;
bra.uni BB0_50;
BB0_48:
mov.f32 %f262, 0f3D2AAAA5;
fma.rn.f32 %f263, %f828, %f60, %f262;
mov.f32 %f264, 0fBF000000;
fma.rn.f32 %f829, %f263, %f60, %f264;
BB0_50:
fma.rn.f32 %f830, %f829, %f827, %f827;
@%p36 bra BB0_52;
mov.f32 %f268, 0f3F800000;
fma.rn.f32 %f830, %f829, %f60, %f268;
BB0_52:
and.b32 %r222, %r382, 2;
setp.eq.s32 %p39, %r222, 0;
@%p39 bra BB0_54;
mov.f32 %f269, 0f00000000;
mov.f32 %f270, 0fBF800000;
fma.rn.f32 %f830, %f830, %f270, %f269;
BB0_54:
mul.f32 %f279, %f35, %f824;
add.u64 %rd62, %SP, 0;
cvta.to.local.u64 %rd63, %rd62;
mul.f32 %f280, %f279, %f279;
mov.f32 %f281, 0f3F800000;
sub.f32 %f282, %f281, %f280;
mul.f32 %f283, %f35, %f830;
mul.f32 %f284, %f283, %f283;
sub.f32 %f285, %f282, %f284;
mov.f32 %f286, 0f00000000;
max.f32 %f287, %f286, %f285;
sqrt.rn.f32 %f288, %f287;
mul.f32 %f289, %f17, %f283;
mul.f32 %f290, %f18, %f283;
mul.f32 %f291, %f19, %f283;
fma.rn.f32 %f292, %f20, %f279, %f289;
fma.rn.f32 %f293, %f21, %f279, %f290;
fma.rn.f32 %f294, %f22, %f279, %f291;
fma.rn.f32 %f295, %f7, %f288, %f292;
fma.rn.f32 %f296, %f8, %f288, %f293;
fma.rn.f32 %f297, %f9, %f288, %f294;
add.f32 %f298, %f7, %f295;
add.f32 %f299, %f8, %f296;
add.f32 %f300, %f9, %f297;
ld.global.f32 %f301, [shadowSpread];
mul.f32 %f302, %f301, %f298;
mul.f32 %f303, %f301, %f299;
mul.f32 %f304, %f301, %f300;
sub.f32 %f305, %f302, %f7;
sub.f32 %f306, %f303, %f8;
sub.f32 %f307, %f304, %f9;
mul.f32 %f308, %f306, %f306;
fma.rn.f32 %f309, %f305, %f305, %f308;
fma.rn.f32 %f310, %f307, %f307, %f309;
sqrt.rn.f32 %f311, %f310;
rcp.rn.f32 %f312, %f311;
mul.f32 %f274, %f312, %f305;
mul.f32 %f275, %f312, %f306;
mul.f32 %f276, %f312, %f307;
ld.global.u32 %r226, [imageEnabled];
and.b32 %r227, %r226, 32;
setp.eq.s32 %p40, %r227, 0;
selp.f32 %f313, 0f3F800000, 0f41200000, %p40;
mul.f32 %f277, %f313, %f31;
mov.u32 %r228, 1065353216;
st.local.u32 [%rd63+8], %r228;
st.local.u32 [%rd63+4], %r228;
st.local.u32 [%rd63], %r228;
ld.global.u32 %r223, [root];
mov.u32 %r224, 1;
mov.f32 %f278, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r223, %f14, %f15, %f16, %f274, %f275, %f276, %r224, %f277, %f278, %rd62, %r115);
// inline asm
ld.local.f32 %f314, [%rd63];
add.f32 %f834, %f834, %f314;
ld.local.f32 %f315, [%rd63+4];
add.f32 %f833, %f833, %f315;
ld.local.f32 %f316, [%rd63+8];
add.f32 %f832, %f832, %f316;
ld.global.u32 %r358, [samples];
add.s32 %r361, %r361, 1;
setp.lt.s32 %p41, %r361, %r358;
@%p41 bra BB0_8;
BB0_55:
add.s32 %r359, %r359, 1;
setp.lt.s32 %p42, %r359, %r358;
@%p42 bra BB0_6;
bra.uni BB0_56;
BB0_4:
mov.f32 %f833, %f832;
mov.f32 %f834, %f832;
BB0_56:
setp.eq.s32 %p43, %r358, 0;
mov.f32 %f838, 0f3F800000;
mov.f32 %f839, %f838;
mov.f32 %f840, %f838;
@%p43 bra BB0_58;
mul.lo.s32 %r229, %r358, %r358;
cvt.rn.f32.s32 %f320, %r229;
rcp.rn.f32 %f321, %f320;
mul.f32 %f838, %f834, %f321;
mul.f32 %f839, %f833, %f321;
mul.f32 %f840, %f832, %f321;
BB0_58:
mul.f32 %f330, %f12, %f21;
fma.rn.f32 %f331, %f11, %f20, %f330;
fma.rn.f32 %f332, %f13, %f22, %f331;
ld.global.v4.f32 {%f333, %f334, %f335, %f336}, [lightTilingOffset];
fma.rn.f32 %f326, %f332, %f333, %f335;
mul.f32 %f339, %f12, %f18;
fma.rn.f32 %f340, %f11, %f17, %f339;
fma.rn.f32 %f341, %f13, %f19, %f340;
fma.rn.f32 %f327, %f341, %f334, %f336;
ld.global.u32 %r230, [lightCookie];
mov.f32 %f329, 0f00000000;
// inline asm
call (%f322, %f323, %f324, %f325), _rt_texture_get_f_id, (%r230, %r99, %f326, %f327, %f329, %f329);
// inline asm
min.f32 %f344, %f322, %f323;
min.f32 %f345, %f344, %f324;
mul.f32 %f87, %f838, %f345;
ld.global.f32 %f346, [directColor];
mul.f32 %f347, %f838, %f346;
ld.global.f32 %f348, [directColor+4];
mul.f32 %f349, %f839, %f348;
ld.global.f32 %f350, [directColor+8];
mul.f32 %f351, %f840, %f350;
mul.f32 %f352, %f322, %f347;
mul.f32 %f353, %f323, %f349;
mul.f32 %f354, %f324, %f351;
cvt.sat.f32.f32 %f355, %f10;
mul.f32 %f88, %f352, %f355;
mul.f32 %f89, %f353, %f355;
mul.f32 %f90, %f355, %f354;
mul.f32 %f91, %f352, 0f3E800000;
mul.f32 %f92, %f353, 0f3E800000;
mul.f32 %f93, %f354, 0f3E800000;
ld.global.u32 %r387, [imageEnabled];
and.b32 %r232, %r387, 8;
setp.eq.s32 %p44, %r232, 0;
@%p44 bra BB0_71;
cvt.u64.u32 %rd66, %r2;
cvt.u64.u32 %rd67, %r3;
mov.u64 %rd70, image_Mask;
cvta.global.u64 %rd65, %rd70;
// inline asm
call (%rd64), _rt_buffer_get_64, (%rd65, %r99, %r99, %rd66, %rd67, %rd26, %rd26);
// inline asm
abs.f32 %f95, %f87;
setp.lt.f32 %p45, %f95, 0f00800000;
mul.f32 %f361, %f95, 0f4B800000;
selp.f32 %f362, 0fC3170000, 0fC2FE0000, %p45;
selp.f32 %f363, %f361, %f95, %p45;
mov.b32 %r235, %f363;
and.b32 %r236, %r235, 8388607;
or.b32 %r237, %r236, 1065353216;
mov.b32 %f364, %r237;
shr.u32 %r238, %r235, 23;
cvt.rn.f32.u32 %f365, %r238;
add.f32 %f366, %f362, %f365;
setp.gt.f32 %p46, %f364, 0f3FB504F3;
mul.f32 %f367, %f364, 0f3F000000;
add.f32 %f368, %f366, 0f3F800000;
selp.f32 %f369, %f367, %f364, %p46;
selp.f32 %f370, %f368, %f366, %p46;
add.f32 %f371, %f369, 0fBF800000;
add.f32 %f357, %f369, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f356,%f357;
// inline asm
add.f32 %f372, %f371, %f371;
mul.f32 %f373, %f356, %f372;
mul.f32 %f374, %f373, %f373;
mov.f32 %f375, 0f3C4CAF63;
mov.f32 %f376, 0f3B18F0FE;
fma.rn.f32 %f377, %f376, %f374, %f375;
mov.f32 %f378, 0f3DAAAABD;
fma.rn.f32 %f379, %f377, %f374, %f378;
mul.rn.f32 %f380, %f379, %f374;
mul.rn.f32 %f381, %f380, %f373;
sub.f32 %f382, %f371, %f373;
neg.f32 %f383, %f373;
add.f32 %f384, %f382, %f382;
fma.rn.f32 %f385, %f383, %f371, %f384;
mul.rn.f32 %f386, %f356, %f385;
add.f32 %f387, %f381, %f373;
sub.f32 %f388, %f373, %f387;
add.f32 %f389, %f381, %f388;
add.f32 %f390, %f386, %f389;
add.f32 %f391, %f387, %f390;
sub.f32 %f392, %f387, %f391;
add.f32 %f393, %f390, %f392;
mov.f32 %f394, 0f3F317200;
mul.rn.f32 %f395, %f370, %f394;
mov.f32 %f396, 0f35BFBE8E;
mul.rn.f32 %f397, %f370, %f396;
add.f32 %f398, %f395, %f391;
sub.f32 %f399, %f395, %f398;
add.f32 %f400, %f391, %f399;
add.f32 %f401, %f393, %f400;
add.f32 %f402, %f397, %f401;
add.f32 %f403, %f398, %f402;
sub.f32 %f404, %f398, %f403;
add.f32 %f405, %f402, %f404;
mov.f32 %f406, 0f3EE8BA2E;
mul.rn.f32 %f407, %f406, %f403;
neg.f32 %f408, %f407;
fma.rn.f32 %f409, %f406, %f403, %f408;
fma.rn.f32 %f410, %f406, %f405, %f409;
fma.rn.f32 %f412, %f329, %f403, %f410;
add.rn.f32 %f413, %f407, %f412;
neg.f32 %f414, %f413;
add.rn.f32 %f415, %f407, %f414;
add.rn.f32 %f416, %f415, %f412;
mov.b32 %r239, %f413;
setp.eq.s32 %p47, %r239, 1118925336;
add.s32 %r240, %r239, -1;
mov.b32 %f417, %r240;
add.f32 %f418, %f416, 0f37000000;
selp.f32 %f419, %f417, %f413, %p47;
selp.f32 %f96, %f418, %f416, %p47;
mul.f32 %f420, %f419, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f421, %f420;
mov.f32 %f422, 0fBF317200;
fma.rn.f32 %f423, %f421, %f422, %f419;
mov.f32 %f424, 0fB5BFBE8E;
fma.rn.f32 %f425, %f421, %f424, %f423;
mul.f32 %f426, %f425, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f427, %f426;
add.f32 %f428, %f421, 0f00000000;
ex2.approx.f32 %f429, %f428;
mul.f32 %f430, %f427, %f429;
setp.lt.f32 %p48, %f419, 0fC2D20000;
selp.f32 %f431, 0f00000000, %f430, %p48;
setp.gt.f32 %p49, %f419, 0f42D20000;
selp.f32 %f841, 0f7F800000, %f431, %p49;
setp.eq.f32 %p50, %f841, 0f7F800000;
@%p50 bra BB0_61;
fma.rn.f32 %f841, %f841, %f96, %f841;
BB0_61:
mov.f32 %f802, 0f3E68BA2E;
cvt.rzi.f32.f32 %f801, %f802;
fma.rn.f32 %f800, %f801, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f799, %f800;
setp.lt.f32 %p51, %f87, 0f00000000;
setp.eq.f32 %p52, %f799, 0f3F800000;
and.pred %p1, %p51, %p52;
mov.b32 %r241, %f841;
xor.b32 %r242, %r241, -2147483648;
mov.b32 %f432, %r242;
selp.f32 %f843, %f432, %f841, %p1;
setp.eq.f32 %p53, %f87, 0f00000000;
@%p53 bra BB0_64;
bra.uni BB0_62;
BB0_64:
add.f32 %f435, %f87, %f87;
selp.f32 %f843, %f435, 0f00000000, %p52;
bra.uni BB0_65;
BB0_129:
mov.u64 %rd211, image_HDR;
cvta.global.u64 %rd206, %rd211;
mov.u32 %r328, 8;
mov.u64 %rd210, 0;
// inline asm
call (%rd205), _rt_buffer_get_64, (%rd206, %r99, %r328, %rd19, %rd20, %rd210, %rd210);
// inline asm
mov.f32 %f739, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs98, %f739;}
// inline asm
mov.u16 %rs99, 0;
st.v4.u16 [%rd205], {%rs98, %rs98, %rs98, %rs99};
BB0_130:
ld.global.u32 %r329, [additive];
setp.eq.s32 %p120, %r329, 0;
@%p120 bra BB0_132;
mov.u64 %rd224, image_RNM0;
cvta.global.u64 %rd213, %rd224;
mov.u32 %r333, 8;
mov.u64 %rd223, 0;
// inline asm
call (%rd212), _rt_buffer_get_64, (%rd213, %r99, %r333, %rd19, %rd20, %rd223, %rd223);
// inline asm
ld.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd212];
// inline asm
{ cvt.f32.f16 %f740, %rs106;}
// inline asm
// inline asm
{ cvt.f32.f16 %f741, %rs107;}
// inline asm
// inline asm
{ cvt.f32.f16 %f742, %rs108;}
// inline asm
// inline asm
call (%rd218), _rt_buffer_get_64, (%rd213, %r99, %r333, %rd19, %rd20, %rd223, %rd223);
// inline asm
add.f32 %f743, %f740, 0f00000000;
add.f32 %f744, %f741, 0f00000000;
add.f32 %f745, %f742, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs105, %f745;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs104, %f744;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs103, %f743;}
// inline asm
mov.u16 %rs110, 0;
st.v4.u16 [%rd218], {%rs103, %rs104, %rs105, %rs110};
bra.uni BB0_133;
BB0_132:
mov.u64 %rd231, image_RNM0;
cvta.global.u64 %rd226, %rd231;
mov.u32 %r335, 8;
mov.u64 %rd230, 0;
// inline asm
call (%rd225), _rt_buffer_get_64, (%rd226, %r99, %r335, %rd19, %rd20, %rd230, %rd230);
// inline asm
mov.f32 %f746, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs111, %f746;}
// inline asm
mov.u16 %rs112, 0;
st.v4.u16 [%rd225], {%rs111, %rs111, %rs111, %rs112};
BB0_133:
ld.global.u32 %r336, [additive];
setp.eq.s32 %p121, %r336, 0;
@%p121 bra BB0_135;
mov.u64 %rd244, image_RNM1;
cvta.global.u64 %rd233, %rd244;
mov.u32 %r340, 8;
mov.u64 %rd243, 0;
// inline asm
call (%rd232), _rt_buffer_get_64, (%rd233, %r99, %r340, %rd19, %rd20, %rd243, %rd243);
// inline asm
ld.v4.u16 {%rs119, %rs120, %rs121, %rs122}, [%rd232];
// inline asm
{ cvt.f32.f16 %f747, %rs119;}
// inline asm
// inline asm
{ cvt.f32.f16 %f748, %rs120;}
// inline asm
// inline asm
{ cvt.f32.f16 %f749, %rs121;}
// inline asm
// inline asm
call (%rd238), _rt_buffer_get_64, (%rd233, %r99, %r340, %rd19, %rd20, %rd243, %rd243);
// inline asm
add.f32 %f750, %f747, 0f00000000;
add.f32 %f751, %f748, 0f00000000;
add.f32 %f752, %f749, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs118, %f752;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs117, %f751;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs116, %f750;}
// inline asm
mov.u16 %rs123, 0;
st.v4.u16 [%rd238], {%rs116, %rs117, %rs118, %rs123};
bra.uni BB0_136;
BB0_135:
mov.u64 %rd251, image_RNM1;
cvta.global.u64 %rd246, %rd251;
mov.u32 %r342, 8;
mov.u64 %rd250, 0;
// inline asm
call (%rd245), _rt_buffer_get_64, (%rd246, %r99, %r342, %rd19, %rd20, %rd250, %rd250);
// inline asm
mov.f32 %f753, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs124, %f753;}
// inline asm
mov.u16 %rs125, 0;
st.v4.u16 [%rd245], {%rs124, %rs124, %rs124, %rs125};
BB0_136:
ld.global.u32 %r343, [additive];
setp.eq.s32 %p122, %r343, 0;
@%p122 bra BB0_138;
mov.u64 %rd264, image_RNM2;
cvta.global.u64 %rd253, %rd264;
mov.u32 %r347, 8;
mov.u64 %rd263, 0;
// inline asm
call (%rd252), _rt_buffer_get_64, (%rd253, %r99, %r347, %rd19, %rd20, %rd263, %rd263);
// inline asm
ld.v4.u16 {%rs132, %rs133, %rs134, %rs135}, [%rd252];
// inline asm
{ cvt.f32.f16 %f754, %rs132;}
// inline asm
// inline asm
{ cvt.f32.f16 %f755, %rs133;}
// inline asm
// inline asm
{ cvt.f32.f16 %f756, %rs134;}
// inline asm
// inline asm
call (%rd258), _rt_buffer_get_64, (%rd253, %r99, %r347, %rd19, %rd20, %rd263, %rd263);
// inline asm
add.f32 %f757, %f754, 0f00000000;
add.f32 %f758, %f755, 0f00000000;
add.f32 %f759, %f756, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs131, %f759;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs130, %f758;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs129, %f757;}
// inline asm
mov.u16 %rs136, 0;
st.v4.u16 [%rd258], {%rs129, %rs130, %rs131, %rs136};
bra.uni BB0_139;
BB0_138:
mov.u64 %rd271, image_RNM2;
cvta.global.u64 %rd266, %rd271;
mov.u32 %r349, 8;
mov.u64 %rd270, 0;
// inline asm
call (%rd265), _rt_buffer_get_64, (%rd266, %r99, %r349, %rd19, %rd20, %rd270, %rd270);
// inline asm
mov.f32 %f760, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs137, %f760;}
// inline asm
mov.u16 %rs138, 0;
st.v4.u16 [%rd265], {%rs137, %rs137, %rs137, %rs138};
BB0_139:
ld.global.u32 %r350, [additive];
setp.eq.s32 %p123, %r350, 0;
@%p123 bra BB0_141;
mov.u64 %rd284, image_RNM3;
cvta.global.u64 %rd273, %rd284;
mov.u32 %r354, 8;
mov.u64 %rd283, 0;
// inline asm
call (%rd272), _rt_buffer_get_64, (%rd273, %r99, %r354, %rd19, %rd20, %rd283, %rd283);
// inline asm
ld.v4.u16 {%rs145, %rs146, %rs147, %rs148}, [%rd272];
// inline asm
{ cvt.f32.f16 %f761, %rs145;}
// inline asm
// inline asm
{ cvt.f32.f16 %f762, %rs146;}
// inline asm
// inline asm
{ cvt.f32.f16 %f763, %rs147;}
// inline asm
// inline asm
call (%rd278), _rt_buffer_get_64, (%rd273, %r99, %r354, %rd19, %rd20, %rd283, %rd283);
// inline asm
add.f32 %f764, %f761, 0f00000000;
add.f32 %f765, %f762, 0f00000000;
add.f32 %f766, %f763, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs144, %f766;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs143, %f765;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs142, %f764;}
// inline asm
mov.u16 %rs149, 0;
st.v4.u16 [%rd278], {%rs142, %rs143, %rs144, %rs149};
bra.uni BB0_142;
BB0_141:
mov.u64 %rd291, image_RNM3;
cvta.global.u64 %rd286, %rd291;
mov.u32 %r356, 8;
mov.u64 %rd290, 0;
// inline asm
call (%rd285), _rt_buffer_get_64, (%rd286, %r99, %r356, %rd19, %rd20, %rd290, %rd290);
// inline asm
mov.f32 %f767, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs150, %f767;}
// inline asm
mov.u16 %rs151, 0;
st.v4.u16 [%rd285], {%rs150, %rs150, %rs150, %rs151};
bra.uni BB0_142;
BB0_62:
setp.geu.f32 %p54, %f87, 0f00000000;
@%p54 bra BB0_65;
mov.f32 %f807, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f434, %f807;
setp.neu.f32 %p55, %f434, 0f3EE8BA2E;
selp.f32 %f843, 0f7FFFFFFF, %f843, %p55;
BB0_65:
abs.f32 %f803, %f87;
add.f32 %f436, %f803, 0f3EE8BA2E;
mov.b32 %r243, %f436;
setp.lt.s32 %p57, %r243, 2139095040;
@%p57 bra BB0_70;
abs.f32 %f805, %f87;
setp.gtu.f32 %p58, %f805, 0f7F800000;
@%p58 bra BB0_69;
bra.uni BB0_67;
BB0_69:
add.f32 %f843, %f87, 0f3EE8BA2E;
bra.uni BB0_70;
BB0_67:
abs.f32 %f806, %f87;
setp.neu.f32 %p59, %f806, 0f7F800000;
@%p59 bra BB0_70;
selp.f32 %f843, 0fFF800000, 0f7F800000, %p1;
BB0_70:
mul.f32 %f437, %f843, 0f437F0000;
setp.eq.f32 %p60, %f87, 0f3F800000;
selp.f32 %f438, 0f437F0000, %f437, %p60;
cvt.rzi.u32.f32 %r244, %f438;
cvt.u16.u32 %rs14, %r244;
mov.u16 %rs15, 255;
st.v2.u8 [%rd64], {%rs14, %rs15};
ld.global.u32 %r387, [imageEnabled];
BB0_71:
and.b32 %r245, %r387, 1;
setp.eq.b32 %p61, %r245, 1;
@!%p61 bra BB0_106;
bra.uni BB0_72;
BB0_72:
mov.f32 %f804, 0f00000000;
abs.f32 %f108, %f88;
setp.lt.f32 %p62, %f108, 0f00800000;
mul.f32 %f444, %f108, 0f4B800000;
selp.f32 %f445, 0fC3170000, 0fC2FE0000, %p62;
selp.f32 %f446, %f444, %f108, %p62;
mov.b32 %r246, %f446;
and.b32 %r247, %r246, 8388607;
or.b32 %r248, %r247, 1065353216;
mov.b32 %f447, %r248;
shr.u32 %r249, %r246, 23;
cvt.rn.f32.u32 %f448, %r249;
add.f32 %f449, %f445, %f448;
setp.gt.f32 %p63, %f447, 0f3FB504F3;
mul.f32 %f450, %f447, 0f3F000000;
add.f32 %f451, %f449, 0f3F800000;
selp.f32 %f452, %f450, %f447, %p63;
selp.f32 %f453, %f451, %f449, %p63;
add.f32 %f454, %f452, 0fBF800000;
add.f32 %f440, %f452, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f439,%f440;
// inline asm
add.f32 %f455, %f454, %f454;
mul.f32 %f456, %f439, %f455;
mul.f32 %f457, %f456, %f456;
mov.f32 %f458, 0f3C4CAF63;
mov.f32 %f459, 0f3B18F0FE;
fma.rn.f32 %f460, %f459, %f457, %f458;
mov.f32 %f461, 0f3DAAAABD;
fma.rn.f32 %f462, %f460, %f457, %f461;
mul.rn.f32 %f463, %f462, %f457;
mul.rn.f32 %f464, %f463, %f456;
sub.f32 %f465, %f454, %f456;
neg.f32 %f466, %f456;
add.f32 %f467, %f465, %f465;
fma.rn.f32 %f468, %f466, %f454, %f467;
mul.rn.f32 %f469, %f439, %f468;
add.f32 %f470, %f464, %f456;
sub.f32 %f471, %f456, %f470;
add.f32 %f472, %f464, %f471;
add.f32 %f473, %f469, %f472;
add.f32 %f474, %f470, %f473;
sub.f32 %f475, %f470, %f474;
add.f32 %f476, %f473, %f475;
mov.f32 %f477, 0f3F317200;
mul.rn.f32 %f478, %f453, %f477;
mov.f32 %f479, 0f35BFBE8E;
mul.rn.f32 %f480, %f453, %f479;
add.f32 %f481, %f478, %f474;
sub.f32 %f482, %f478, %f481;
add.f32 %f483, %f474, %f482;
add.f32 %f484, %f476, %f483;
add.f32 %f485, %f480, %f484;
add.f32 %f486, %f481, %f485;
sub.f32 %f487, %f481, %f486;
add.f32 %f488, %f485, %f487;
mov.f32 %f489, 0f3EE66666;
mul.rn.f32 %f490, %f489, %f486;
neg.f32 %f491, %f490;
fma.rn.f32 %f492, %f489, %f486, %f491;
fma.rn.f32 %f493, %f489, %f488, %f492;
fma.rn.f32 %f495, %f804, %f486, %f493;
add.rn.f32 %f496, %f490, %f495;
neg.f32 %f497, %f496;
add.rn.f32 %f498, %f490, %f497;
add.rn.f32 %f499, %f498, %f495;
mov.b32 %r250, %f496;
setp.eq.s32 %p64, %r250, 1118925336;
add.s32 %r251, %r250, -1;
mov.b32 %f500, %r251;
add.f32 %f501, %f499, 0f37000000;
selp.f32 %f502, %f500, %f496, %p64;
selp.f32 %f109, %f501, %f499, %p64;
mul.f32 %f503, %f502, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f504, %f503;
mov.f32 %f505, 0fBF317200;
fma.rn.f32 %f506, %f504, %f505, %f502;
mov.f32 %f507, 0fB5BFBE8E;
fma.rn.f32 %f508, %f504, %f507, %f506;
mul.f32 %f509, %f508, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f510, %f509;
add.f32 %f511, %f504, 0f00000000;
ex2.approx.f32 %f512, %f511;
mul.f32 %f513, %f510, %f512;
setp.lt.f32 %p65, %f502, 0fC2D20000;
selp.f32 %f514, 0f00000000, %f513, %p65;
setp.gt.f32 %p66, %f502, 0f42D20000;
selp.f32 %f844, 0f7F800000, %f514, %p66;
setp.eq.f32 %p67, %f844, 0f7F800000;
@%p67 bra BB0_74;
fma.rn.f32 %f844, %f844, %f109, %f844;
BB0_74:
mov.f32 %f771, 0f3E666666;
cvt.rzi.f32.f32 %f770, %f771;
fma.rn.f32 %f769, %f770, 0fC0000000, 0f3EE66666;
abs.f32 %f768, %f769;
setp.lt.f32 %p68, %f88, 0f00000000;
setp.eq.f32 %p69, %f768, 0f3F800000;
and.pred %p2, %p68, %p69;
mov.b32 %r252, %f844;
xor.b32 %r253, %r252, -2147483648;
mov.b32 %f515, %r253;
selp.f32 %f846, %f515, %f844, %p2;
setp.eq.f32 %p70, %f88, 0f00000000;
@%p70 bra BB0_77;
bra.uni BB0_75;
BB0_77:
add.f32 %f518, %f88, %f88;
selp.f32 %f846, %f518, 0f00000000, %p69;
bra.uni BB0_78;
BB0_75:
setp.geu.f32 %p71, %f88, 0f00000000;
@%p71 bra BB0_78;
mov.f32 %f795, 0f3EE66666;
cvt.rzi.f32.f32 %f517, %f795;
setp.neu.f32 %p72, %f517, 0f3EE66666;
selp.f32 %f846, 0f7FFFFFFF, %f846, %p72;
BB0_78:
abs.f32 %f772, %f88;
add.f32 %f519, %f772, 0f3EE66666;
mov.b32 %r254, %f519;
setp.lt.s32 %p74, %r254, 2139095040;
@%p74 bra BB0_83;
abs.f32 %f793, %f88;
setp.gtu.f32 %p75, %f793, 0f7F800000;
@%p75 bra BB0_82;
bra.uni BB0_80;
BB0_82:
add.f32 %f846, %f88, 0f3EE66666;
bra.uni BB0_83;
BB0_80:
abs.f32 %f794, %f88;
setp.neu.f32 %p76, %f794, 0f7F800000;
@%p76 bra BB0_83;
selp.f32 %f846, 0fFF800000, 0f7F800000, %p2;
BB0_83:
mov.f32 %f781, 0fB5BFBE8E;
mov.f32 %f780, 0fBF317200;
mov.f32 %f779, 0f35BFBE8E;
mov.f32 %f778, 0f3F317200;
mov.f32 %f777, 0f3DAAAABD;
mov.f32 %f776, 0f3C4CAF63;
mov.f32 %f775, 0f3B18F0FE;
mov.f32 %f774, 0f3EE66666;
mov.f32 %f773, 0f00000000;
setp.eq.f32 %p77, %f88, 0f3F800000;
selp.f32 %f120, 0f3F800000, %f846, %p77;
abs.f32 %f121, %f89;
setp.lt.f32 %p78, %f121, 0f00800000;
mul.f32 %f522, %f121, 0f4B800000;
selp.f32 %f523, 0fC3170000, 0fC2FE0000, %p78;
selp.f32 %f524, %f522, %f121, %p78;
mov.b32 %r255, %f524;
and.b32 %r256, %r255, 8388607;
or.b32 %r257, %r256, 1065353216;
mov.b32 %f525, %r257;
shr.u32 %r258, %r255, 23;
cvt.rn.f32.u32 %f526, %r258;
add.f32 %f527, %f523, %f526;
setp.gt.f32 %p79, %f525, 0f3FB504F3;
mul.f32 %f528, %f525, 0f3F000000;
add.f32 %f529, %f527, 0f3F800000;
selp.f32 %f530, %f528, %f525, %p79;
selp.f32 %f531, %f529, %f527, %p79;
add.f32 %f532, %f530, 0fBF800000;
add.f32 %f521, %f530, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f520,%f521;
// inline asm
add.f32 %f533, %f532, %f532;
mul.f32 %f534, %f520, %f533;
mul.f32 %f535, %f534, %f534;
fma.rn.f32 %f538, %f775, %f535, %f776;
fma.rn.f32 %f540, %f538, %f535, %f777;
mul.rn.f32 %f541, %f540, %f535;
mul.rn.f32 %f542, %f541, %f534;
sub.f32 %f543, %f532, %f534;
neg.f32 %f544, %f534;
add.f32 %f545, %f543, %f543;
fma.rn.f32 %f546, %f544, %f532, %f545;
mul.rn.f32 %f547, %f520, %f546;
add.f32 %f548, %f542, %f534;
sub.f32 %f549, %f534, %f548;
add.f32 %f550, %f542, %f549;
add.f32 %f551, %f547, %f550;
add.f32 %f552, %f548, %f551;
sub.f32 %f553, %f548, %f552;
add.f32 %f554, %f551, %f553;
mul.rn.f32 %f556, %f531, %f778;
mul.rn.f32 %f558, %f531, %f779;
add.f32 %f559, %f556, %f552;
sub.f32 %f560, %f556, %f559;
add.f32 %f561, %f552, %f560;
add.f32 %f562, %f554, %f561;
add.f32 %f563, %f558, %f562;
add.f32 %f564, %f559, %f563;
sub.f32 %f565, %f559, %f564;
add.f32 %f566, %f563, %f565;
mul.rn.f32 %f568, %f774, %f564;
neg.f32 %f569, %f568;
fma.rn.f32 %f570, %f774, %f564, %f569;
fma.rn.f32 %f571, %f774, %f566, %f570;
fma.rn.f32 %f573, %f773, %f564, %f571;
add.rn.f32 %f574, %f568, %f573;
neg.f32 %f575, %f574;
add.rn.f32 %f576, %f568, %f575;
add.rn.f32 %f577, %f576, %f573;
mov.b32 %r259, %f574;
setp.eq.s32 %p80, %r259, 1118925336;
add.s32 %r260, %r259, -1;
mov.b32 %f578, %r260;
add.f32 %f579, %f577, 0f37000000;
selp.f32 %f580, %f578, %f574, %p80;
selp.f32 %f122, %f579, %f577, %p80;
mul.f32 %f581, %f580, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f582, %f581;
fma.rn.f32 %f584, %f582, %f780, %f580;
fma.rn.f32 %f586, %f582, %f781, %f584;
mul.f32 %f587, %f586, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f588, %f587;
add.f32 %f589, %f582, 0f00000000;
ex2.approx.f32 %f590, %f589;
mul.f32 %f591, %f588, %f590;
setp.lt.f32 %p81, %f580, 0fC2D20000;
selp.f32 %f592, 0f00000000, %f591, %p81;
setp.gt.f32 %p82, %f580, 0f42D20000;
selp.f32 %f847, 0f7F800000, %f592, %p82;
setp.eq.f32 %p83, %f847, 0f7F800000;
@%p83 bra BB0_85;
fma.rn.f32 %f847, %f847, %f122, %f847;
BB0_85:
setp.lt.f32 %p84, %f89, 0f00000000;
and.pred %p3, %p84, %p69;
mov.b32 %r261, %f847;
xor.b32 %r262, %r261, -2147483648;
mov.b32 %f593, %r262;
selp.f32 %f849, %f593, %f847, %p3;
setp.eq.f32 %p86, %f89, 0f00000000;
@%p86 bra BB0_88;
bra.uni BB0_86;
BB0_88:
add.f32 %f596, %f89, %f89;
selp.f32 %f849, %f596, 0f00000000, %p69;
bra.uni BB0_89;
BB0_86:
setp.geu.f32 %p87, %f89, 0f00000000;
@%p87 bra BB0_89;
mov.f32 %f792, 0f3EE66666;
cvt.rzi.f32.f32 %f595, %f792;
setp.neu.f32 %p88, %f595, 0f3EE66666;
selp.f32 %f849, 0f7FFFFFFF, %f849, %p88;
BB0_89:
abs.f32 %f796, %f89;
add.f32 %f597, %f796, 0f3EE66666;
mov.b32 %r263, %f597;
setp.lt.s32 %p90, %r263, 2139095040;
@%p90 bra BB0_94;
abs.f32 %f797, %f89;
setp.gtu.f32 %p91, %f797, 0f7F800000;
@%p91 bra BB0_93;
bra.uni BB0_91;
BB0_93:
add.f32 %f849, %f89, 0f3EE66666;
bra.uni BB0_94;
BB0_91:
abs.f32 %f798, %f89;
setp.neu.f32 %p92, %f798, 0f7F800000;
@%p92 bra BB0_94;
selp.f32 %f849, 0fFF800000, 0f7F800000, %p3;
BB0_94:
mov.f32 %f790, 0fB5BFBE8E;
mov.f32 %f789, 0fBF317200;
mov.f32 %f788, 0f35BFBE8E;
mov.f32 %f787, 0f3F317200;
mov.f32 %f786, 0f3DAAAABD;
mov.f32 %f785, 0f3C4CAF63;
mov.f32 %f784, 0f3B18F0FE;
mov.f32 %f783, 0f3EE66666;
mov.f32 %f782, 0f00000000;
setp.eq.f32 %p93, %f89, 0f3F800000;
selp.f32 %f133, 0f3F800000, %f849, %p93;
abs.f32 %f134, %f90;
setp.lt.f32 %p94, %f134, 0f00800000;
mul.f32 %f600, %f134, 0f4B800000;
selp.f32 %f601, 0fC3170000, 0fC2FE0000, %p94;
selp.f32 %f602, %f600, %f134, %p94;
mov.b32 %r264, %f602;
and.b32 %r265, %r264, 8388607;
or.b32 %r266, %r265, 1065353216;
mov.b32 %f603, %r266;
shr.u32 %r267, %r264, 23;
cvt.rn.f32.u32 %f604, %r267;
add.f32 %f605, %f601, %f604;
setp.gt.f32 %p95, %f603, 0f3FB504F3;
mul.f32 %f606, %f603, 0f3F000000;
add.f32 %f607, %f605, 0f3F800000;
selp.f32 %f608, %f606, %f603, %p95;
selp.f32 %f609, %f607, %f605, %p95;
add.f32 %f610, %f608, 0fBF800000;
add.f32 %f599, %f608, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f598,%f599;
// inline asm
add.f32 %f611, %f610, %f610;
mul.f32 %f612, %f598, %f611;
mul.f32 %f613, %f612, %f612;
fma.rn.f32 %f616, %f784, %f613, %f785;
fma.rn.f32 %f618, %f616, %f613, %f786;
mul.rn.f32 %f619, %f618, %f613;
mul.rn.f32 %f620, %f619, %f612;
sub.f32 %f621, %f610, %f612;
neg.f32 %f622, %f612;
add.f32 %f623, %f621, %f621;
fma.rn.f32 %f624, %f622, %f610, %f623;
mul.rn.f32 %f625, %f598, %f624;
add.f32 %f626, %f620, %f612;
sub.f32 %f627, %f612, %f626;
add.f32 %f628, %f620, %f627;
add.f32 %f629, %f625, %f628;
add.f32 %f630, %f626, %f629;
sub.f32 %f631, %f626, %f630;
add.f32 %f632, %f629, %f631;
mul.rn.f32 %f634, %f609, %f787;
mul.rn.f32 %f636, %f609, %f788;
add.f32 %f637, %f634, %f630;
sub.f32 %f638, %f634, %f637;
add.f32 %f639, %f630, %f638;
add.f32 %f640, %f632, %f639;
add.f32 %f641, %f636, %f640;
add.f32 %f642, %f637, %f641;
sub.f32 %f643, %f637, %f642;
add.f32 %f644, %f641, %f643;
mul.rn.f32 %f646, %f783, %f642;
neg.f32 %f647, %f646;
fma.rn.f32 %f648, %f783, %f642, %f647;
fma.rn.f32 %f649, %f783, %f644, %f648;
fma.rn.f32 %f651, %f782, %f642, %f649;
add.rn.f32 %f652, %f646, %f651;
neg.f32 %f653, %f652;
add.rn.f32 %f654, %f646, %f653;
add.rn.f32 %f655, %f654, %f651;
mov.b32 %r268, %f652;
setp.eq.s32 %p96, %r268, 1118925336;
add.s32 %r269, %r268, -1;
mov.b32 %f656, %r269;
add.f32 %f657, %f655, 0f37000000;
selp.f32 %f658, %f656, %f652, %p96;
selp.f32 %f135, %f657, %f655, %p96;
mul.f32 %f659, %f658, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f660, %f659;
fma.rn.f32 %f662, %f660, %f789, %f658;
fma.rn.f32 %f664, %f660, %f790, %f662;
mul.f32 %f665, %f664, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f666, %f665;
add.f32 %f667, %f660, 0f00000000;
ex2.approx.f32 %f668, %f667;
mul.f32 %f669, %f666, %f668;
setp.lt.f32 %p97, %f658, 0fC2D20000;
selp.f32 %f670, 0f00000000, %f669, %p97;
setp.gt.f32 %p98, %f658, 0f42D20000;
selp.f32 %f850, 0f7F800000, %f670, %p98;
setp.eq.f32 %p99, %f850, 0f7F800000;
@%p99 bra BB0_96;
fma.rn.f32 %f850, %f850, %f135, %f850;
BB0_96:
setp.lt.f32 %p100, %f90, 0f00000000;
and.pred %p4, %p100, %p69;
mov.b32 %r270, %f850;
xor.b32 %r271, %r270, -2147483648;
mov.b32 %f671, %r271;
selp.f32 %f852, %f671, %f850, %p4;
setp.eq.f32 %p102, %f90, 0f00000000;
@%p102 bra BB0_99;
bra.uni BB0_97;
BB0_99:
add.f32 %f674, %f90, %f90;
selp.f32 %f852, %f674, 0f00000000, %p69;
bra.uni BB0_100;
BB0_97:
setp.geu.f32 %p103, %f90, 0f00000000;
@%p103 bra BB0_100;
mov.f32 %f791, 0f3EE66666;
cvt.rzi.f32.f32 %f673, %f791;
setp.neu.f32 %p104, %f673, 0f3EE66666;
selp.f32 %f852, 0f7FFFFFFF, %f852, %p104;
BB0_100:
abs.f32 %f808, %f90;
add.f32 %f675, %f808, 0f3EE66666;
mov.b32 %r272, %f675;
setp.lt.s32 %p106, %r272, 2139095040;
@%p106 bra BB0_105;
abs.f32 %f809, %f90;
setp.gtu.f32 %p107, %f809, 0f7F800000;
@%p107 bra BB0_104;
bra.uni BB0_102;
BB0_104:
add.f32 %f852, %f90, 0f3EE66666;
bra.uni BB0_105;
BB0_102:
abs.f32 %f810, %f90;
setp.neu.f32 %p108, %f810, 0f7F800000;
@%p108 bra BB0_105;
selp.f32 %f852, 0fFF800000, 0f7F800000, %p4;
BB0_105:
mov.u32 %r357, 4;
setp.eq.f32 %p109, %f90, 0f3F800000;
selp.f32 %f676, 0f3F800000, %f852, %p109;
cvt.u64.u32 %rd74, %r3;
cvt.u64.u32 %rd73, %r2;
mov.u64 %rd77, image;
cvta.global.u64 %rd72, %rd77;
// inline asm
call (%rd71), _rt_buffer_get_64, (%rd72, %r99, %r357, %rd73, %rd74, %rd26, %rd26);
// inline asm
cvt.sat.f32.f32 %f677, %f676;
mul.f32 %f678, %f677, 0f437FFD71;
cvt.rzi.u32.f32 %r275, %f678;
cvt.sat.f32.f32 %f679, %f133;
mul.f32 %f680, %f679, 0f437FFD71;
cvt.rzi.u32.f32 %r276, %f680;
cvt.sat.f32.f32 %f681, %f120;
mul.f32 %f682, %f681, 0f437FFD71;
cvt.rzi.u32.f32 %r277, %f682;
cvt.u16.u32 %rs16, %r275;
cvt.u16.u32 %rs17, %r277;
cvt.u16.u32 %rs18, %r276;
mov.u16 %rs19, 255;
st.v4.u8 [%rd71], {%rs16, %rs18, %rs17, %rs19};
ld.global.u32 %r387, [imageEnabled];
BB0_106:
cvt.u64.u32 %rd17, %r2;
cvt.u64.u32 %rd18, %r3;
and.b32 %r278, %r387, 4;
setp.eq.s32 %p110, %r278, 0;
@%p110 bra BB0_110;
ld.global.u32 %r279, [additive];
setp.eq.s32 %p111, %r279, 0;
mov.f32 %f683, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs20, %f683;}
// inline asm
@%p111 bra BB0_109;
mov.u64 %rd90, image_HDR;
cvta.global.u64 %rd79, %rd90;
mov.u32 %r283, 8;
// inline asm
call (%rd78), _rt_buffer_get_64, (%rd79, %r99, %r283, %rd17, %rd18, %rd26, %rd26);
// inline asm
ld.v4.u16 {%rs27, %rs28, %rs29, %rs30}, [%rd78];
// inline asm
{ cvt.f32.f16 %f684, %rs27;}
// inline asm
// inline asm
{ cvt.f32.f16 %f685, %rs28;}
// inline asm
// inline asm
{ cvt.f32.f16 %f686, %rs29;}
// inline asm
// inline asm
call (%rd84), _rt_buffer_get_64, (%rd79, %r99, %r283, %rd17, %rd18, %rd26, %rd26);
// inline asm
add.f32 %f687, %f88, %f684;
add.f32 %f688, %f89, %f685;
add.f32 %f689, %f90, %f686;
// inline asm
{ cvt.rn.f16.f32 %rs26, %f689;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs25, %f688;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs24, %f687;}
// inline asm
st.v4.u16 [%rd84], {%rs24, %rs25, %rs26, %rs20};
bra.uni BB0_110;
BB0_109:
mov.u64 %rd97, image_HDR;
cvta.global.u64 %rd92, %rd97;
mov.u32 %r285, 8;
// inline asm
call (%rd91), _rt_buffer_get_64, (%rd92, %r99, %r285, %rd17, %rd18, %rd26, %rd26);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs33, %f90;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs32, %f89;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs31, %f88;}
// inline asm
st.v4.u16 [%rd91], {%rs31, %rs32, %rs33, %rs20};
BB0_110:
ld.global.u32 %r286, [additive];
setp.eq.s32 %p112, %r286, 0;
mov.f32 %f693, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs34, %f693;}
// inline asm
@%p112 bra BB0_112;
mov.u64 %rd110, image_RNM0;
cvta.global.u64 %rd99, %rd110;
mov.u32 %r290, 8;
// inline asm
call (%rd98), _rt_buffer_get_64, (%rd99, %r99, %r290, %rd17, %rd18, %rd26, %rd26);
// inline asm
ld.v4.u16 {%rs41, %rs42, %rs43, %rs44}, [%rd98];
// inline asm
{ cvt.f32.f16 %f694, %rs41;}
// inline asm
// inline asm
{ cvt.f32.f16 %f695, %rs42;}
// inline asm
// inline asm
{ cvt.f32.f16 %f696, %rs43;}
// inline asm
// inline asm
call (%rd104), _rt_buffer_get_64, (%rd99, %r99, %r290, %rd17, %rd18, %rd26, %rd26);
// inline asm
add.f32 %f697, %f91, %f694;
add.f32 %f698, %f92, %f695;
add.f32 %f699, %f93, %f696;
// inline asm
{ cvt.rn.f16.f32 %rs40, %f699;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs39, %f698;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs38, %f697;}
// inline asm
st.v4.u16 [%rd104], {%rs38, %rs39, %rs40, %rs34};
bra.uni BB0_113;
BB0_112:
mov.u64 %rd117, image_RNM0;
cvta.global.u64 %rd112, %rd117;
mov.u32 %r292, 8;
// inline asm
call (%rd111), _rt_buffer_get_64, (%rd112, %r99, %r292, %rd17, %rd18, %rd26, %rd26);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs47, %f93;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs46, %f92;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs45, %f91;}
// inline asm
st.v4.u16 [%rd111], {%rs45, %rs46, %rs47, %rs34};
BB0_113:
ld.global.f32 %f704, [directDir];
fma.rn.f32 %f146, %f704, 0fBF000000, 0f3F000000;
ld.global.f32 %f705, [directDir+4];
fma.rn.f32 %f147, %f705, 0fBF000000, 0f3F000000;
ld.global.f32 %f706, [directDir+8];
fma.rn.f32 %f148, %f706, 0fBF000000, 0f3F000000;
ld.global.u32 %r293, [additive];
setp.eq.s32 %p113, %r293, 0;
// inline asm
{ cvt.rn.f16.f32 %rs48, %f693;}
// inline asm
@%p113 bra BB0_115;
mov.u64 %rd130, image_RNM1;
cvta.global.u64 %rd119, %rd130;
mov.u32 %r297, 8;
// inline asm
call (%rd118), _rt_buffer_get_64, (%rd119, %r99, %r297, %rd17, %rd18, %rd26, %rd26);
// inline asm
ld.v4.u16 {%rs55, %rs56, %rs57, %rs58}, [%rd118];
// inline asm
{ cvt.f32.f16 %f707, %rs55;}
// inline asm
// inline asm
{ cvt.f32.f16 %f708, %rs56;}
// inline asm
// inline asm
{ cvt.f32.f16 %f709, %rs57;}
// inline asm
// inline asm
call (%rd124), _rt_buffer_get_64, (%rd119, %r99, %r297, %rd17, %rd18, %rd26, %rd26);
// inline asm
add.f32 %f710, %f146, %f707;
add.f32 %f711, %f146, %f708;
add.f32 %f712, %f146, %f709;
// inline asm
{ cvt.rn.f16.f32 %rs54, %f712;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs53, %f711;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs52, %f710;}
// inline asm
st.v4.u16 [%rd124], {%rs52, %rs53, %rs54, %rs48};
bra.uni BB0_116;
BB0_115:
mov.u64 %rd137, image_RNM1;
cvta.global.u64 %rd132, %rd137;
mov.u32 %r299, 8;
// inline asm
call (%rd131), _rt_buffer_get_64, (%rd132, %r99, %r299, %rd17, %rd18, %rd26, %rd26);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs59, %f146;}
// inline asm
st.v4.u16 [%rd131], {%rs59, %rs59, %rs59, %rs48};
BB0_116:
ld.global.u32 %r300, [additive];
setp.eq.s32 %p114, %r300, 0;
// inline asm
{ cvt.rn.f16.f32 %rs60, %f693;}
// inline asm
@%p114 bra BB0_118;
mov.u64 %rd150, image_RNM2;
cvta.global.u64 %rd139, %rd150;
mov.u32 %r304, 8;
// inline asm
call (%rd138), _rt_buffer_get_64, (%rd139, %r99, %r304, %rd17, %rd18, %rd26, %rd26);
// inline asm
ld.v4.u16 {%rs67, %rs68, %rs69, %rs70}, [%rd138];
// inline asm
{ cvt.f32.f16 %f715, %rs67;}
// inline asm
// inline asm
{ cvt.f32.f16 %f716, %rs68;}
// inline asm
// inline asm
{ cvt.f32.f16 %f717, %rs69;}
// inline asm
// inline asm
call (%rd144), _rt_buffer_get_64, (%rd139, %r99, %r304, %rd17, %rd18, %rd26, %rd26);
// inline asm
add.f32 %f718, %f147, %f715;
add.f32 %f719, %f147, %f716;
add.f32 %f720, %f147, %f717;
// inline asm
{ cvt.rn.f16.f32 %rs66, %f720;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs65, %f719;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs64, %f718;}
// inline asm
st.v4.u16 [%rd144], {%rs64, %rs65, %rs66, %rs60};
bra.uni BB0_119;
BB0_118:
mov.u64 %rd157, image_RNM2;
cvta.global.u64 %rd152, %rd157;
mov.u32 %r306, 8;
// inline asm
call (%rd151), _rt_buffer_get_64, (%rd152, %r99, %r306, %rd17, %rd18, %rd26, %rd26);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs71, %f147;}
// inline asm
st.v4.u16 [%rd151], {%rs71, %rs71, %rs71, %rs60};
BB0_119:
ld.global.u32 %r307, [additive];
setp.eq.s32 %p115, %r307, 0;
// inline asm
{ cvt.rn.f16.f32 %rs72, %f693;}
// inline asm
@%p115 bra BB0_121;
mov.u64 %rd170, image_RNM3;
cvta.global.u64 %rd159, %rd170;
mov.u32 %r311, 8;
// inline asm
call (%rd158), _rt_buffer_get_64, (%rd159, %r99, %r311, %rd17, %rd18, %rd26, %rd26);
// inline asm
ld.v4.u16 {%rs79, %rs80, %rs81, %rs82}, [%rd158];
// inline asm
{ cvt.f32.f16 %f723, %rs79;}
// inline asm
// inline asm
{ cvt.f32.f16 %f724, %rs80;}
// inline asm
// inline asm
{ cvt.f32.f16 %f725, %rs81;}
// inline asm
// inline asm
call (%rd164), _rt_buffer_get_64, (%rd159, %r99, %r311, %rd17, %rd18, %rd26, %rd26);
// inline asm
add.f32 %f726, %f148, %f723;
add.f32 %f727, %f148, %f724;
add.f32 %f728, %f148, %f725;
// inline asm
{ cvt.rn.f16.f32 %rs78, %f728;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs77, %f727;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs76, %f726;}
// inline asm
st.v4.u16 [%rd164], {%rs76, %rs77, %rs78, %rs72};
bra.uni BB0_142;
BB0_121:
mov.u64 %rd177, image_RNM3;
cvta.global.u64 %rd172, %rd177;
mov.u32 %r313, 8;
// inline asm
call (%rd171), _rt_buffer_get_64, (%rd172, %r99, %r313, %rd17, %rd18, %rd26, %rd26);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs83, %f148;}
// inline asm
st.v4.u16 [%rd171], {%rs83, %rs83, %rs83, %rs72};
BB0_142:
ret;
}