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

2970 lines
86 KiB
Plaintext
Raw Permalink 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_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 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo9directDirE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo11directColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12shadowSpreadE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12ignoreNormalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename9directDirE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename11directColorE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename12shadowSpreadE[6] = {102, 108, 111, 97, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[4] = {105, 110, 116, 0};
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum9directDirE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum11directColorE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12shadowSpreadE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 4919;
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic9directDirE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic11directColorE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12shadowSpreadE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation9directDirE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation11directColorE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12shadowSpreadE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[1];
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
.visible .entry _Z6oxMainv(
)
{
.local .align 4 .b8 __local_depot0[52];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<171>;
.reg .b16 %rs<181>;
.reg .f32 %f<1521>;
.reg .b32 %r<503>;
.reg .b64 %rd<326>;
mov.u64 %rd325, __local_depot0;
cvta.local.u64 %SP, %rd325;
ld.global.v2.u32 {%r109, %r110}, [pixelID];
cvt.u64.u32 %rd24, %r109;
cvt.u64.u32 %rd25, %r110;
mov.u64 %rd28, uvnormal;
cvta.global.u64 %rd23, %rd28;
mov.u32 %r107, 2;
mov.u32 %r108, 4;
mov.u64 %rd27, 0;
// inline asm
call (%rd22), _rt_buffer_get_64, (%rd23, %r107, %r108, %rd24, %rd25, %rd27, %rd27);
// inline asm
ld.u32 %r1, [%rd22];
shr.u32 %r113, %r1, 16;
cvt.u16.u32 %rs1, %r113;
and.b16 %rs15, %rs1, 255;
cvt.u16.u32 %rs16, %r1;
or.b16 %rs17, %rs16, %rs15;
setp.eq.s16 %p5, %rs17, 0;
mov.f32 %f1439, 0f00000000;
mov.f32 %f1440, %f1439;
mov.f32 %f1441, %f1439;
@%p5 bra BB0_2;
ld.u8 %rs18, [%rd22+1];
and.b16 %rs20, %rs16, 255;
cvt.rn.f32.u16 %f256, %rs20;
div.rn.f32 %f257, %f256, 0f437F0000;
fma.rn.f32 %f258, %f257, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f259, %rs18;
div.rn.f32 %f260, %f259, 0f437F0000;
fma.rn.f32 %f261, %f260, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f262, %rs15;
div.rn.f32 %f263, %f262, 0f437F0000;
fma.rn.f32 %f264, %f263, 0f40000000, 0fBF800000;
mul.f32 %f265, %f261, %f261;
fma.rn.f32 %f266, %f258, %f258, %f265;
fma.rn.f32 %f267, %f264, %f264, %f266;
sqrt.rn.f32 %f268, %f267;
rcp.rn.f32 %f269, %f268;
mul.f32 %f1439, %f258, %f269;
mul.f32 %f1440, %f261, %f269;
mul.f32 %f1441, %f264, %f269;
BB0_2:
ld.global.v2.u32 {%r114, %r115}, [pixelID];
ld.global.v2.u32 {%r117, %r118}, [tileInfo];
add.s32 %r2, %r114, %r117;
add.s32 %r3, %r115, %r118;
setp.eq.f32 %p6, %f1440, 0f00000000;
setp.eq.f32 %p7, %f1439, 0f00000000;
and.pred %p8, %p7, %p6;
setp.eq.f32 %p9, %f1441, 0f00000000;
and.pred %p10, %p8, %p9;
@%p10 bra BB0_156;
bra.uni BB0_3;
BB0_156:
ld.global.u32 %r502, [imageEnabled];
and.b32 %r427, %r502, 1;
setp.eq.b32 %p163, %r427, 1;
@!%p163 bra BB0_158;
bra.uni BB0_157;
BB0_157:
cvt.u64.u32 %rd209, %r2;
cvt.u64.u32 %rd210, %r3;
mov.u64 %rd213, image;
cvta.global.u64 %rd208, %rd213;
mov.u64 %rd212, 0;
// inline asm
call (%rd207), _rt_buffer_get_64, (%rd208, %r107, %r108, %rd209, %rd210, %rd212, %rd212);
// inline asm
mov.u16 %rs103, 0;
st.v4.u8 [%rd207], {%rs103, %rs103, %rs103, %rs103};
ld.global.u32 %r502, [imageEnabled];
BB0_158:
and.b32 %r430, %r502, 8;
setp.eq.s32 %p164, %r430, 0;
@%p164 bra BB0_160;
cvt.u64.u32 %rd217, %r3;
cvt.u64.u32 %rd216, %r2;
mov.u64 %rd220, image_Mask;
cvta.global.u64 %rd215, %rd220;
mov.u64 %rd219, 0;
// inline asm
call (%rd214), _rt_buffer_get_64, (%rd215, %r107, %r107, %rd216, %rd217, %rd219, %rd219);
// inline asm
mov.f32 %f1401, 0f00000000;
cvt.rzi.u32.f32 %r433, %f1401;
cvt.u16.u32 %rs104, %r433;
mov.u16 %rs105, 0;
st.v2.u8 [%rd214], {%rs104, %rs105};
ld.global.u32 %r502, [imageEnabled];
BB0_160:
cvt.u64.u32 %rd20, %r2;
cvt.u64.u32 %rd21, %r3;
and.b32 %r434, %r502, 4;
setp.eq.s32 %p165, %r434, 0;
@%p165 bra BB0_164;
ld.global.u32 %r435, [additive];
setp.eq.s32 %p166, %r435, 0;
@%p166 bra BB0_163;
mov.u64 %rd233, image_HDR;
cvta.global.u64 %rd222, %rd233;
mov.u32 %r439, 8;
mov.u64 %rd232, 0;
// inline asm
call (%rd221), _rt_buffer_get_64, (%rd222, %r107, %r439, %rd20, %rd21, %rd232, %rd232);
// inline asm
ld.v4.u16 {%rs112, %rs113, %rs114, %rs115}, [%rd221];
// inline asm
{ cvt.f32.f16 %f1402, %rs112;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1403, %rs113;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1404, %rs114;}
// inline asm
// inline asm
call (%rd227), _rt_buffer_get_64, (%rd222, %r107, %r439, %rd20, %rd21, %rd232, %rd232);
// inline asm
add.f32 %f1405, %f1402, 0f00000000;
add.f32 %f1406, %f1403, 0f00000000;
add.f32 %f1407, %f1404, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs111, %f1407;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs110, %f1406;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs109, %f1405;}
// inline asm
mov.u16 %rs116, 0;
st.v4.u16 [%rd227], {%rs109, %rs110, %rs111, %rs116};
bra.uni BB0_164;
BB0_3:
ld.global.f32 %f8, [directDir+4];
ld.global.f32 %f9, [directDir+8];
ld.global.f32 %f7, [directDir];
ld.global.v2.u32 {%r130, %r131}, [pixelID];
cvt.u64.u32 %rd31, %r130;
cvt.u64.u32 %rd32, %r131;
mov.u64 %rd53, uvpos;
cvta.global.u64 %rd30, %rd53;
mov.u32 %r127, 12;
// inline asm
call (%rd29), _rt_buffer_get_64, (%rd30, %r107, %r127, %rd31, %rd32, %rd27, %rd27);
// inline asm
ld.f32 %f274, [%rd29+8];
ld.f32 %f275, [%rd29+4];
ld.f32 %f276, [%rd29];
mul.f32 %f277, %f276, 0f3456BF95;
mul.f32 %f278, %f275, 0f3456BF95;
mul.f32 %f279, %f274, 0f3456BF95;
abs.f32 %f280, %f1439;
div.rn.f32 %f281, %f277, %f280;
abs.f32 %f282, %f1440;
div.rn.f32 %f283, %f278, %f282;
abs.f32 %f284, %f1441;
div.rn.f32 %f285, %f279, %f284;
abs.f32 %f286, %f281;
abs.f32 %f287, %f283;
abs.f32 %f288, %f285;
mov.f32 %f289, 0f38D1B717;
max.f32 %f290, %f286, %f289;
max.f32 %f291, %f287, %f289;
max.f32 %f292, %f288, %f289;
fma.rn.f32 %f10, %f1439, %f290, %f276;
fma.rn.f32 %f11, %f1440, %f291, %f275;
fma.rn.f32 %f12, %f1441, %f292, %f274;
setp.gt.f32 %p11, %f280, %f284;
neg.f32 %f293, %f1440;
selp.f32 %f294, %f293, 0f00000000, %p11;
neg.f32 %f295, %f1441;
selp.f32 %f296, %f1439, %f295, %p11;
selp.f32 %f297, 0f00000000, %f1440, %p11;
mul.f32 %f298, %f296, %f296;
fma.rn.f32 %f299, %f294, %f294, %f298;
fma.rn.f32 %f300, %f297, %f297, %f299;
sqrt.rn.f32 %f301, %f300;
rcp.rn.f32 %f302, %f301;
mul.f32 %f13, %f294, %f302;
mul.f32 %f14, %f296, %f302;
mul.f32 %f15, %f297, %f302;
mul.f32 %f303, %f1441, %f14;
mul.f32 %f304, %f1440, %f15;
sub.f32 %f16, %f303, %f304;
mul.f32 %f305, %f1439, %f15;
mul.f32 %f306, %f1441, %f13;
sub.f32 %f17, %f305, %f306;
mul.f32 %f307, %f1440, %f13;
mul.f32 %f308, %f1439, %f14;
sub.f32 %f18, %f307, %f308;
ld.global.v2.u32 {%r134, %r135}, [pixelID];
cvt.u64.u32 %rd38, %r135;
add.s32 %r138, %r134, 1;
ld.global.u32 %r139, [resolution];
add.s32 %r140, %r139, -1;
min.u32 %r141, %r138, %r140;
cvt.u64.u32 %rd37, %r141;
// inline asm
call (%rd35), _rt_buffer_get_64, (%rd30, %r107, %r127, %rd37, %rd38, %rd27, %rd27);
// inline asm
ld.f32 %f309, [%rd35+8];
ld.f32 %f310, [%rd35+4];
ld.f32 %f311, [%rd35];
setp.eq.f32 %p12, %f311, 0f00000000;
setp.eq.f32 %p13, %f310, 0f00000000;
and.pred %p14, %p12, %p13;
setp.eq.f32 %p15, %f309, 0f00000000;
and.pred %p16, %p14, %p15;
selp.f32 %f312, %f276, %f311, %p16;
selp.f32 %f313, %f275, %f310, %p16;
selp.f32 %f314, %f274, %f309, %p16;
ld.global.v2.u32 {%r142, %r143}, [pixelID];
add.s32 %r146, %r143, 1;
ld.global.u32 %r147, [resolution+4];
add.s32 %r148, %r147, -1;
min.u32 %r149, %r146, %r148;
cvt.u64.u32 %rd43, %r142;
cvt.u64.u32 %rd44, %r149;
// inline asm
call (%rd41), _rt_buffer_get_64, (%rd30, %r107, %r127, %rd43, %rd44, %rd27, %rd27);
// inline asm
ld.f32 %f315, [%rd41+8];
ld.f32 %f316, [%rd41+4];
ld.f32 %f317, [%rd41];
setp.eq.f32 %p17, %f317, 0f00000000;
setp.eq.f32 %p18, %f316, 0f00000000;
and.pred %p19, %p17, %p18;
setp.eq.f32 %p20, %f315, 0f00000000;
and.pred %p21, %p19, %p20;
selp.f32 %f318, %f276, %f317, %p21;
selp.f32 %f319, %f275, %f316, %p21;
selp.f32 %f320, %f274, %f315, %p21;
sub.f32 %f321, %f312, %f276;
sub.f32 %f322, %f313, %f275;
sub.f32 %f323, %f314, %f274;
sub.f32 %f324, %f318, %f276;
sub.f32 %f325, %f319, %f275;
sub.f32 %f326, %f320, %f274;
abs.f32 %f327, %f321;
abs.f32 %f328, %f324;
max.f32 %f329, %f327, %f328;
abs.f32 %f330, %f322;
abs.f32 %f331, %f325;
max.f32 %f332, %f330, %f331;
abs.f32 %f333, %f323;
abs.f32 %f334, %f326;
max.f32 %f335, %f333, %f334;
max.f32 %f336, %f329, %f332;
max.f32 %f337, %f336, %f335;
setp.neu.f32 %p22, %f321, 0f00000000;
setp.neu.f32 %p23, %f322, 0f00000000;
or.pred %p24, %p22, %p23;
setp.neu.f32 %p25, %f323, 0f00000000;
setp.neu.f32 %p26, %f324, 0f00000000;
setp.neu.f32 %p27, %f325, 0f00000000;
or.pred %p28, %p26, %p27;
setp.neu.f32 %p29, %f326, 0f00000000;
or.pred %p30, %p28, %p29;
or.pred %p31, %p24, %p25;
and.pred %p32, %p31, %p30;
selp.f32 %f19, %f337, 0f00000000, %p32;
abs.f32 %f338, %f7;
abs.f32 %f339, %f9;
setp.gt.f32 %p33, %f338, %f339;
neg.f32 %f340, %f8;
neg.f32 %f341, %f9;
selp.f32 %f342, %f340, 0f00000000, %p33;
selp.f32 %f343, %f7, %f341, %p33;
selp.f32 %f344, 0f00000000, %f8, %p33;
mul.f32 %f345, %f343, %f343;
fma.rn.f32 %f346, %f342, %f342, %f345;
fma.rn.f32 %f347, %f344, %f344, %f346;
sqrt.rn.f32 %f348, %f347;
rcp.rn.f32 %f349, %f348;
mul.f32 %f20, %f342, %f349;
mul.f32 %f21, %f343, %f349;
mul.f32 %f22, %f344, %f349;
ld.global.v2.u32 {%r150, %r151}, [pixelID];
cvt.u64.u32 %rd49, %r150;
cvt.u64.u32 %rd50, %r151;
mov.u64 %rd54, rnd_seeds;
cvta.global.u64 %rd48, %rd54;
// inline asm
call (%rd47), _rt_buffer_get_64, (%rd48, %r107, %r108, %rd49, %rd50, %rd27, %rd27);
// inline asm
ld.global.u32 %r471, [samples];
mov.f32 %f36, 0f00000000;
setp.lt.s32 %p34, %r471, 1;
@%p34 bra BB0_4;
cvt.rn.f32.s32 %f354, %r471;
rcp.rn.f32 %f23, %f354;
ld.u32 %r497, [%rd47];
mul.f32 %f24, %f10, 0f3456BF95;
mul.f32 %f25, %f11, 0f3456BF95;
mul.f32 %f26, %f12, 0f3456BF95;
mul.f32 %f355, %f7, %f21;
mul.f32 %f356, %f8, %f20;
sub.f32 %f27, %f356, %f355;
mul.f32 %f357, %f9, %f20;
mul.f32 %f358, %f7, %f22;
sub.f32 %f28, %f358, %f357;
mul.f32 %f359, %f8, %f22;
mul.f32 %f360, %f9, %f21;
sub.f32 %f29, %f360, %f359;
mov.f32 %f36, 0f00000000;
mov.u32 %r472, 0;
abs.f32 %f361, %f25;
abs.f32 %f362, %f24;
max.f32 %f363, %f362, %f361;
abs.f32 %f364, %f26;
max.f32 %f365, %f363, %f364;
mov.f32 %f37, %f36;
mov.f32 %f38, %f36;
mov.f32 %f39, %f36;
BB0_6:
setp.lt.s32 %p35, %r471, 1;
@%p35 bra BB0_89;
mov.u32 %r474, 0;
BB0_8:
cvt.rn.f32.s32 %f1437, %r472;
mad.lo.s32 %r156, %r497, 1664525, 1013904223;
and.b32 %r157, %r156, 16777215;
cvt.rn.f32.u32 %f367, %r157;
fma.rn.f32 %f368, %f367, 0f33800000, %f1437;
mul.f32 %f369, %f23, %f368;
mad.lo.s32 %r11, %r156, 1664525, 1013904223;
and.b32 %r158, %r11, 16777215;
cvt.rn.f32.u32 %f370, %r158;
cvt.rn.f32.s32 %f371, %r474;
fma.rn.f32 %f372, %f370, 0f33800000, %f371;
mul.f32 %f373, %f23, %f372;
sqrt.rn.f32 %f40, %f369;
mul.f32 %f1456, %f373, 0f40C90FDB;
abs.f32 %f42, %f1456;
setp.neu.f32 %p36, %f42, 0f7F800000;
mov.f32 %f1450, %f1456;
@%p36 bra BB0_10;
mov.f32 %f374, 0f00000000;
mul.rn.f32 %f1450, %f1456, %f374;
BB0_10:
mul.f32 %f375, %f1450, 0f3F22F983;
cvt.rni.s32.f32 %r485, %f375;
cvt.rn.f32.s32 %f376, %r485;
neg.f32 %f377, %f376;
mov.f32 %f378, 0f3FC90FDA;
fma.rn.f32 %f379, %f377, %f378, %f1450;
mov.f32 %f380, 0f33A22168;
fma.rn.f32 %f381, %f377, %f380, %f379;
mov.f32 %f382, 0f27C234C5;
fma.rn.f32 %f1451, %f377, %f382, %f381;
abs.f32 %f383, %f1450;
setp.leu.f32 %p37, %f383, 0f47CE4780;
@%p37 bra BB0_21;
add.u64 %rd56, %SP, 24;
cvta.to.local.u64 %rd3, %rd56;
mov.b32 %r13, %f1450;
shr.u32 %r14, %r13, 23;
shl.b32 %r161, %r13, 8;
or.b32 %r15, %r161, -2147483648;
mov.u32 %r476, 0;
mov.u64 %rd322, 0;
mov.u64 %rd321, %rd3;
mov.u32 %r477, %r476;
BB0_12:
.pragma "nounroll";
shl.b64 %rd57, %rd322, 2;
mov.u64 %rd58, __cudart_i2opi_f;
add.s64 %rd59, %rd58, %rd57;
ld.const.u32 %r164, [%rd59];
// inline asm
{
mad.lo.cc.u32 %r162, %r164, %r15, %r477;
madc.hi.u32 %r477, %r164, %r15, 0;
}
// inline asm
st.local.u32 [%rd321], %r162;
add.s32 %r476, %r476, 1;
cvt.s64.s32 %rd322, %r476;
mul.wide.s32 %rd62, %r476, 4;
add.s64 %rd321, %rd3, %rd62;
setp.ne.s32 %p38, %r476, 6;
@%p38 bra BB0_12;
and.b32 %r167, %r14, 255;
add.s32 %r168, %r167, -128;
shr.u32 %r169, %r168, 5;
and.b32 %r20, %r13, -2147483648;
cvta.to.local.u64 %rd64, %rd56;
st.local.u32 [%rd64+24], %r477;
mov.u32 %r170, 6;
sub.s32 %r171, %r170, %r169;
mul.wide.s32 %rd65, %r171, 4;
add.s64 %rd8, %rd64, %rd65;
ld.local.u32 %r478, [%rd8];
ld.local.u32 %r479, [%rd8+-4];
and.b32 %r23, %r14, 31;
setp.eq.s32 %p39, %r23, 0;
@%p39 bra BB0_15;
mov.u32 %r172, 32;
sub.s32 %r173, %r172, %r23;
shr.u32 %r174, %r479, %r173;
shl.b32 %r175, %r478, %r23;
add.s32 %r478, %r174, %r175;
ld.local.u32 %r176, [%rd8+-8];
shr.u32 %r177, %r176, %r173;
shl.b32 %r178, %r479, %r23;
add.s32 %r479, %r177, %r178;
BB0_15:
shr.u32 %r179, %r479, 30;
shl.b32 %r180, %r478, 2;
add.s32 %r480, %r179, %r180;
shl.b32 %r29, %r479, 2;
shr.u32 %r181, %r480, 31;
shr.u32 %r182, %r478, 30;
add.s32 %r30, %r181, %r182;
setp.eq.s32 %p40, %r181, 0;
@%p40 bra BB0_16;
bra.uni BB0_17;
BB0_16:
mov.u32 %r481, %r20;
mov.u32 %r482, %r29;
bra.uni BB0_18;
BB0_17:
not.b32 %r183, %r480;
neg.s32 %r482, %r29;
setp.eq.s32 %p41, %r29, 0;
selp.u32 %r184, 1, 0, %p41;
add.s32 %r480, %r184, %r183;
xor.b32 %r481, %r20, -2147483648;
BB0_18:
clz.b32 %r484, %r480;
setp.eq.s32 %p42, %r484, 0;
shl.b32 %r185, %r480, %r484;
mov.u32 %r186, 32;
sub.s32 %r187, %r186, %r484;
shr.u32 %r188, %r482, %r187;
add.s32 %r189, %r188, %r185;
selp.b32 %r38, %r480, %r189, %p42;
mov.u32 %r190, -921707870;
mul.hi.u32 %r483, %r38, %r190;
setp.eq.s32 %p43, %r20, 0;
neg.s32 %r191, %r30;
selp.b32 %r485, %r30, %r191, %p43;
setp.lt.s32 %p44, %r483, 1;
@%p44 bra BB0_20;
mul.lo.s32 %r192, %r38, -921707870;
shr.u32 %r193, %r192, 31;
shl.b32 %r194, %r483, 1;
add.s32 %r483, %r193, %r194;
add.s32 %r484, %r484, 1;
BB0_20:
mov.u32 %r195, 126;
sub.s32 %r196, %r195, %r484;
shl.b32 %r197, %r196, 23;
add.s32 %r198, %r483, 1;
shr.u32 %r199, %r198, 7;
add.s32 %r200, %r199, 1;
shr.u32 %r201, %r200, 1;
add.s32 %r202, %r201, %r197;
or.b32 %r203, %r202, %r481;
mov.b32 %f1451, %r203;
BB0_21:
mul.rn.f32 %f48, %f1451, %f1451;
add.s32 %r46, %r485, 1;
and.b32 %r47, %r46, 1;
setp.eq.s32 %p45, %r47, 0;
@%p45 bra BB0_23;
bra.uni BB0_22;
BB0_23:
mov.f32 %f386, 0f3C08839E;
mov.f32 %f387, 0fB94CA1F9;
fma.rn.f32 %f1452, %f387, %f48, %f386;
bra.uni BB0_24;
BB0_22:
mov.f32 %f384, 0fBAB6061A;
mov.f32 %f385, 0f37CCF5CE;
fma.rn.f32 %f1452, %f385, %f48, %f384;
BB0_24:
@%p45 bra BB0_26;
bra.uni BB0_25;
BB0_26:
mov.f32 %f391, 0fBE2AAAA3;
fma.rn.f32 %f392, %f1452, %f48, %f391;
mov.f32 %f393, 0f00000000;
fma.rn.f32 %f1453, %f392, %f48, %f393;
bra.uni BB0_27;
BB0_25:
mov.f32 %f388, 0f3D2AAAA5;
fma.rn.f32 %f389, %f1452, %f48, %f388;
mov.f32 %f390, 0fBF000000;
fma.rn.f32 %f1453, %f389, %f48, %f390;
BB0_27:
fma.rn.f32 %f1454, %f1453, %f1451, %f1451;
@%p45 bra BB0_29;
mov.f32 %f394, 0f3F800000;
fma.rn.f32 %f1454, %f1453, %f48, %f394;
BB0_29:
and.b32 %r204, %r46, 2;
setp.eq.s32 %p48, %r204, 0;
@%p48 bra BB0_31;
mov.f32 %f395, 0f00000000;
mov.f32 %f396, 0fBF800000;
fma.rn.f32 %f1454, %f1454, %f396, %f395;
BB0_31:
@%p36 bra BB0_33;
mov.f32 %f397, 0f00000000;
mul.rn.f32 %f1456, %f1456, %f397;
BB0_33:
mul.f32 %f398, %f1456, 0f3F22F983;
cvt.rni.s32.f32 %r495, %f398;
cvt.rn.f32.s32 %f399, %r495;
neg.f32 %f400, %f399;
fma.rn.f32 %f402, %f400, %f378, %f1456;
fma.rn.f32 %f404, %f400, %f380, %f402;
fma.rn.f32 %f1457, %f400, %f382, %f404;
abs.f32 %f406, %f1456;
setp.leu.f32 %p50, %f406, 0f47CE4780;
@%p50 bra BB0_44;
add.u64 %rd67, %SP, 24;
cvta.to.local.u64 %rd9, %rd67;
mov.b32 %r49, %f1456;
shr.u32 %r50, %r49, 23;
shl.b32 %r207, %r49, 8;
or.b32 %r51, %r207, -2147483648;
mov.u32 %r486, 0;
mov.u64 %rd323, %rd9;
mov.u64 %rd324, %rd27;
mov.u32 %r487, %r486;
BB0_35:
.pragma "nounroll";
shl.b64 %rd68, %rd324, 2;
mov.u64 %rd69, __cudart_i2opi_f;
add.s64 %rd70, %rd69, %rd68;
ld.const.u32 %r210, [%rd70];
// inline asm
{
mad.lo.cc.u32 %r208, %r210, %r51, %r487;
madc.hi.u32 %r487, %r210, %r51, 0;
}
// inline asm
st.local.u32 [%rd323], %r208;
add.s32 %r486, %r486, 1;
cvt.s64.s32 %rd324, %r486;
mul.wide.s32 %rd71, %r486, 4;
add.s64 %rd323, %rd9, %rd71;
setp.ne.s32 %p51, %r486, 6;
@%p51 bra BB0_35;
and.b32 %r213, %r50, 255;
add.s32 %r214, %r213, -128;
shr.u32 %r215, %r214, 5;
and.b32 %r56, %r49, -2147483648;
cvta.to.local.u64 %rd73, %rd67;
st.local.u32 [%rd73+24], %r487;
mov.u32 %r216, 6;
sub.s32 %r217, %r216, %r215;
mul.wide.s32 %rd74, %r217, 4;
add.s64 %rd15, %rd73, %rd74;
ld.local.u32 %r488, [%rd15];
ld.local.u32 %r489, [%rd15+-4];
and.b32 %r59, %r50, 31;
setp.eq.s32 %p52, %r59, 0;
@%p52 bra BB0_38;
mov.u32 %r218, 32;
sub.s32 %r219, %r218, %r59;
shr.u32 %r220, %r489, %r219;
shl.b32 %r221, %r488, %r59;
add.s32 %r488, %r220, %r221;
ld.local.u32 %r222, [%rd15+-8];
shr.u32 %r223, %r222, %r219;
shl.b32 %r224, %r489, %r59;
add.s32 %r489, %r223, %r224;
BB0_38:
shr.u32 %r225, %r489, 30;
shl.b32 %r226, %r488, 2;
add.s32 %r490, %r225, %r226;
shl.b32 %r65, %r489, 2;
shr.u32 %r227, %r490, 31;
shr.u32 %r228, %r488, 30;
add.s32 %r66, %r227, %r228;
setp.eq.s32 %p53, %r227, 0;
@%p53 bra BB0_39;
bra.uni BB0_40;
BB0_39:
mov.u32 %r491, %r56;
mov.u32 %r492, %r65;
bra.uni BB0_41;
BB0_40:
not.b32 %r229, %r490;
neg.s32 %r492, %r65;
setp.eq.s32 %p54, %r65, 0;
selp.u32 %r230, 1, 0, %p54;
add.s32 %r490, %r230, %r229;
xor.b32 %r491, %r56, -2147483648;
BB0_41:
clz.b32 %r494, %r490;
setp.eq.s32 %p55, %r494, 0;
shl.b32 %r231, %r490, %r494;
mov.u32 %r232, 32;
sub.s32 %r233, %r232, %r494;
shr.u32 %r234, %r492, %r233;
add.s32 %r235, %r234, %r231;
selp.b32 %r74, %r490, %r235, %p55;
mov.u32 %r236, -921707870;
mul.hi.u32 %r493, %r74, %r236;
setp.eq.s32 %p56, %r56, 0;
neg.s32 %r237, %r66;
selp.b32 %r495, %r66, %r237, %p56;
setp.lt.s32 %p57, %r493, 1;
@%p57 bra BB0_43;
mul.lo.s32 %r238, %r74, -921707870;
shr.u32 %r239, %r238, 31;
shl.b32 %r240, %r493, 1;
add.s32 %r493, %r239, %r240;
add.s32 %r494, %r494, 1;
BB0_43:
mov.u32 %r241, 126;
sub.s32 %r242, %r241, %r494;
shl.b32 %r243, %r242, 23;
add.s32 %r244, %r493, 1;
shr.u32 %r245, %r244, 7;
add.s32 %r246, %r245, 1;
shr.u32 %r247, %r246, 1;
add.s32 %r248, %r247, %r243;
or.b32 %r249, %r248, %r491;
mov.b32 %f1457, %r249;
BB0_44:
mul.rn.f32 %f65, %f1457, %f1457;
and.b32 %r82, %r495, 1;
setp.eq.s32 %p58, %r82, 0;
@%p58 bra BB0_46;
bra.uni BB0_45;
BB0_46:
mov.f32 %f409, 0f3C08839E;
mov.f32 %f410, 0fB94CA1F9;
fma.rn.f32 %f1458, %f410, %f65, %f409;
bra.uni BB0_47;
BB0_45:
mov.f32 %f407, 0fBAB6061A;
mov.f32 %f408, 0f37CCF5CE;
fma.rn.f32 %f1458, %f408, %f65, %f407;
BB0_47:
@%p58 bra BB0_49;
bra.uni BB0_48;
BB0_49:
mov.f32 %f414, 0fBE2AAAA3;
fma.rn.f32 %f415, %f1458, %f65, %f414;
mov.f32 %f416, 0f00000000;
fma.rn.f32 %f1459, %f415, %f65, %f416;
bra.uni BB0_50;
BB0_48:
mov.f32 %f411, 0f3D2AAAA5;
fma.rn.f32 %f412, %f1458, %f65, %f411;
mov.f32 %f413, 0fBF000000;
fma.rn.f32 %f1459, %f412, %f65, %f413;
BB0_50:
fma.rn.f32 %f1460, %f1459, %f1457, %f1457;
@%p58 bra BB0_52;
mov.f32 %f417, 0f3F800000;
fma.rn.f32 %f1460, %f1459, %f65, %f417;
BB0_52:
and.b32 %r250, %r495, 2;
setp.eq.s32 %p61, %r250, 0;
@%p61 bra BB0_54;
mov.f32 %f418, 0f00000000;
mov.f32 %f419, 0fBF800000;
fma.rn.f32 %f1460, %f1460, %f419, %f418;
BB0_54:
max.f32 %f1438, %f365, %f289;
mul.f32 %f428, %f40, %f1454;
add.u64 %rd75, %SP, 12;
cvta.to.local.u64 %rd16, %rd75;
mul.f32 %f429, %f428, %f428;
mov.f32 %f430, 0f3F800000;
sub.f32 %f431, %f430, %f429;
mul.f32 %f432, %f40, %f1460;
mul.f32 %f433, %f432, %f432;
sub.f32 %f434, %f431, %f433;
mov.f32 %f435, 0f00000000;
max.f32 %f436, %f435, %f434;
sqrt.rn.f32 %f437, %f436;
mul.f32 %f438, %f20, %f432;
mul.f32 %f439, %f21, %f432;
mul.f32 %f440, %f22, %f432;
fma.rn.f32 %f441, %f29, %f428, %f438;
fma.rn.f32 %f442, %f28, %f428, %f439;
fma.rn.f32 %f443, %f27, %f428, %f440;
fma.rn.f32 %f444, %f7, %f437, %f441;
fma.rn.f32 %f445, %f8, %f437, %f442;
fma.rn.f32 %f446, %f9, %f437, %f443;
add.f32 %f447, %f7, %f444;
add.f32 %f448, %f8, %f445;
add.f32 %f449, %f9, %f446;
ld.global.f32 %f450, [shadowSpread];
mul.f32 %f451, %f450, %f447;
mul.f32 %f452, %f450, %f448;
mul.f32 %f453, %f450, %f449;
sub.f32 %f454, %f451, %f7;
sub.f32 %f455, %f452, %f8;
sub.f32 %f456, %f453, %f9;
mul.f32 %f457, %f455, %f455;
fma.rn.f32 %f458, %f454, %f454, %f457;
fma.rn.f32 %f459, %f456, %f456, %f458;
sqrt.rn.f32 %f460, %f459;
rcp.rn.f32 %f461, %f460;
mul.f32 %f77, %f461, %f454;
mul.f32 %f78, %f461, %f455;
mul.f32 %f79, %f461, %f456;
ld.global.u32 %r254, [imageEnabled];
and.b32 %r255, %r254, 32;
setp.eq.s32 %p62, %r255, 0;
selp.f32 %f462, 0f3F800000, 0f41200000, %p62;
mul.f32 %f426, %f462, %f1438;
mad.lo.s32 %r256, %r11, 1664525, 1013904223;
and.b32 %r257, %r256, 16777215;
cvt.rn.f32.u32 %f463, %r257;
mul.f32 %f464, %f463, 0f33800000;
fma.rn.f32 %f465, %f464, 0f40000000, 0fBF800000;
mad.lo.s32 %r83, %r256, 1664525, 1013904223;
and.b32 %r258, %r83, 16777215;
cvt.rn.f32.u32 %f466, %r258;
mul.f32 %f467, %f466, 0f33800000;
fma.rn.f32 %f468, %f467, 0f40000000, 0fBF800000;
mul.f32 %f469, %f13, %f468;
mul.f32 %f470, %f14, %f468;
mul.f32 %f471, %f15, %f468;
fma.rn.f32 %f472, %f16, %f465, %f469;
fma.rn.f32 %f473, %f17, %f465, %f470;
fma.rn.f32 %f474, %f18, %f465, %f471;
mul.f32 %f475, %f473, %f473;
fma.rn.f32 %f476, %f472, %f472, %f475;
fma.rn.f32 %f477, %f474, %f474, %f476;
sqrt.rn.f32 %f478, %f477;
rcp.rn.f32 %f479, %f478;
mul.f32 %f480, %f472, %f479;
mul.f32 %f481, %f473, %f479;
mul.f32 %f482, %f474, %f479;
mul.f32 %f483, %f1440, %f482;
mul.f32 %f484, %f1441, %f481;
sub.f32 %f485, %f483, %f484;
mul.f32 %f486, %f1441, %f480;
mul.f32 %f487, %f1439, %f482;
sub.f32 %f488, %f486, %f487;
mul.f32 %f489, %f1439, %f481;
mul.f32 %f490, %f1440, %f480;
sub.f32 %f491, %f489, %f490;
mul.f32 %f492, %f480, 0f3E24A490;
mul.f32 %f493, %f481, 0f3E24A490;
mul.f32 %f494, %f482, 0f3E24A490;
fma.rn.f32 %f495, %f19, %f492, %f10;
fma.rn.f32 %f496, %f19, %f493, %f11;
fma.rn.f32 %f497, %f19, %f494, %f12;
mul.f32 %f498, %f485, 0fBECECEDE;
mul.f32 %f499, %f488, 0fBECECEDE;
mul.f32 %f500, %f491, 0fBECECEDE;
fma.rn.f32 %f81, %f19, %f498, %f495;
fma.rn.f32 %f82, %f19, %f499, %f496;
fma.rn.f32 %f83, %f19, %f500, %f497;
sub.f32 %f501, %f81, %f10;
sub.f32 %f502, %f82, %f11;
sub.f32 %f503, %f83, %f12;
mul.f32 %f504, %f502, %f502;
fma.rn.f32 %f505, %f501, %f501, %f504;
fma.rn.f32 %f506, %f503, %f503, %f505;
sqrt.rn.f32 %f427, %f506;
rcp.rn.f32 %f507, %f427;
mul.f32 %f423, %f507, %f501;
mul.f32 %f424, %f507, %f502;
mul.f32 %f425, %f507, %f503;
mov.u32 %r259, 1065353216;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r251, [root];
mov.u32 %r252, 1;
// inline asm
call _rt_trace_64, (%r251, %f10, %f11, %f12, %f423, %f424, %f425, %r252, %f426, %f427, %rd75, %r127);
// inline asm
ld.local.f32 %f508, [%rd16];
mov.u16 %rs174, 0;
setp.neu.f32 %p63, %f508, 0f3F800000;
@%p63 bra BB0_58;
mov.u16 %rs174, 0;
ld.local.f32 %f509, [%rd16+4];
setp.neu.f32 %p64, %f509, 0f3F800000;
@%p64 bra BB0_58;
mov.u16 %rs174, 0;
ld.local.f32 %f510, [%rd16+8];
setp.neu.f32 %p65, %f510, 0f3F800000;
@%p65 bra BB0_58;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r260, [root];
mov.f32 %f518, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r260, %f81, %f82, %f83, %f77, %f78, %f79, %r252, %f426, %f518, %rd75, %r127);
// inline asm
ld.local.f32 %f519, [%rd16];
add.f32 %f39, %f39, %f519;
ld.local.f32 %f520, [%rd16+4];
add.f32 %f38, %f38, %f520;
ld.local.f32 %f521, [%rd16+8];
add.f32 %f37, %f37, %f521;
add.f32 %f36, %f36, 0f3F800000;
mov.u16 %rs174, 1;
BB0_58:
mad.lo.s32 %r267, %r83, 1664525, 1013904223;
and.b32 %r268, %r267, 16777215;
cvt.rn.f32.u32 %f530, %r268;
mul.f32 %f531, %f530, 0f33800000;
fma.rn.f32 %f532, %f531, 0f40000000, 0fBF800000;
mad.lo.s32 %r84, %r267, 1664525, 1013904223;
and.b32 %r269, %r84, 16777215;
cvt.rn.f32.u32 %f533, %r269;
mul.f32 %f534, %f533, 0f33800000;
fma.rn.f32 %f535, %f534, 0f40000000, 0fBF800000;
mul.f32 %f536, %f13, %f535;
mul.f32 %f537, %f14, %f535;
mul.f32 %f538, %f15, %f535;
fma.rn.f32 %f539, %f16, %f532, %f536;
fma.rn.f32 %f540, %f17, %f532, %f537;
fma.rn.f32 %f541, %f18, %f532, %f538;
mul.f32 %f542, %f540, %f540;
fma.rn.f32 %f543, %f539, %f539, %f542;
fma.rn.f32 %f544, %f541, %f541, %f543;
sqrt.rn.f32 %f545, %f544;
rcp.rn.f32 %f546, %f545;
mul.f32 %f547, %f539, %f546;
mul.f32 %f548, %f540, %f546;
mul.f32 %f549, %f541, %f546;
mul.f32 %f550, %f1440, %f549;
mul.f32 %f551, %f1441, %f548;
sub.f32 %f552, %f550, %f551;
mul.f32 %f553, %f1441, %f547;
mul.f32 %f554, %f1439, %f549;
sub.f32 %f555, %f553, %f554;
mul.f32 %f556, %f1439, %f548;
mul.f32 %f557, %f1440, %f547;
sub.f32 %f558, %f556, %f557;
mul.f32 %f559, %f547, 0fBE24A490;
mul.f32 %f560, %f548, 0fBE24A490;
mul.f32 %f561, %f549, 0fBE24A490;
fma.rn.f32 %f562, %f19, %f559, %f10;
fma.rn.f32 %f563, %f19, %f560, %f11;
fma.rn.f32 %f564, %f19, %f561, %f12;
mul.f32 %f565, %f552, 0f3ECACABC;
mul.f32 %f566, %f555, 0f3ECACABC;
mul.f32 %f567, %f558, 0f3ECACABC;
fma.rn.f32 %f92, %f19, %f565, %f562;
fma.rn.f32 %f93, %f19, %f566, %f563;
fma.rn.f32 %f94, %f19, %f567, %f564;
sub.f32 %f568, %f92, %f10;
sub.f32 %f569, %f93, %f11;
sub.f32 %f570, %f94, %f12;
mul.f32 %f571, %f569, %f569;
fma.rn.f32 %f572, %f568, %f568, %f571;
fma.rn.f32 %f573, %f570, %f570, %f572;
sqrt.rn.f32 %f529, %f573;
rcp.rn.f32 %f574, %f529;
mul.f32 %f525, %f574, %f568;
mul.f32 %f526, %f574, %f569;
mul.f32 %f527, %f574, %f570;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r264, [root];
// inline asm
call _rt_trace_64, (%r264, %f10, %f11, %f12, %f525, %f526, %f527, %r252, %f426, %f529, %rd75, %r127);
// inline asm
ld.local.f32 %f575, [%rd16];
setp.neu.f32 %p66, %f575, 0f3F800000;
@%p66 bra BB0_62;
ld.local.f32 %f576, [%rd16+4];
setp.neu.f32 %p67, %f576, 0f3F800000;
@%p67 bra BB0_62;
ld.local.f32 %f577, [%rd16+8];
setp.neu.f32 %p68, %f577, 0f3F800000;
@%p68 bra BB0_62;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r271, [root];
mov.f32 %f585, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r271, %f92, %f93, %f94, %f77, %f78, %f79, %r252, %f426, %f585, %rd75, %r127);
// inline asm
ld.local.f32 %f586, [%rd16];
add.f32 %f39, %f39, %f586;
ld.local.f32 %f587, [%rd16+4];
add.f32 %f38, %f38, %f587;
ld.local.f32 %f588, [%rd16+8];
add.f32 %f37, %f37, %f588;
add.f32 %f36, %f36, 0f3F800000;
mov.u16 %rs174, 1;
BB0_62:
mad.lo.s32 %r278, %r84, 1664525, 1013904223;
and.b32 %r279, %r278, 16777215;
cvt.rn.f32.u32 %f597, %r279;
mul.f32 %f598, %f597, 0f33800000;
fma.rn.f32 %f599, %f598, 0f40000000, 0fBF800000;
mad.lo.s32 %r85, %r278, 1664525, 1013904223;
and.b32 %r280, %r85, 16777215;
cvt.rn.f32.u32 %f600, %r280;
mul.f32 %f601, %f600, 0f33800000;
fma.rn.f32 %f602, %f601, 0f40000000, 0fBF800000;
mul.f32 %f603, %f13, %f602;
mul.f32 %f604, %f14, %f602;
mul.f32 %f605, %f15, %f602;
fma.rn.f32 %f606, %f16, %f599, %f603;
fma.rn.f32 %f607, %f17, %f599, %f604;
fma.rn.f32 %f608, %f18, %f599, %f605;
mul.f32 %f609, %f607, %f607;
fma.rn.f32 %f610, %f606, %f606, %f609;
fma.rn.f32 %f611, %f608, %f608, %f610;
sqrt.rn.f32 %f612, %f611;
rcp.rn.f32 %f613, %f612;
mul.f32 %f614, %f606, %f613;
mul.f32 %f615, %f607, %f613;
mul.f32 %f616, %f608, %f613;
mul.f32 %f617, %f1440, %f616;
mul.f32 %f618, %f1441, %f615;
sub.f32 %f619, %f617, %f618;
mul.f32 %f620, %f1441, %f614;
mul.f32 %f621, %f1439, %f616;
sub.f32 %f622, %f620, %f621;
mul.f32 %f623, %f1439, %f615;
mul.f32 %f624, %f1440, %f614;
sub.f32 %f625, %f623, %f624;
mul.f32 %f626, %f614, 0f3F23A3A8;
mul.f32 %f627, %f615, 0f3F23A3A8;
mul.f32 %f628, %f616, 0f3F23A3A8;
fma.rn.f32 %f629, %f19, %f626, %f10;
fma.rn.f32 %f630, %f19, %f627, %f11;
fma.rn.f32 %f631, %f19, %f628, %f12;
mul.f32 %f632, %f619, 0f3E24A490;
mul.f32 %f633, %f622, 0f3E24A490;
mul.f32 %f634, %f625, 0f3E24A490;
fma.rn.f32 %f103, %f19, %f632, %f629;
fma.rn.f32 %f104, %f19, %f633, %f630;
fma.rn.f32 %f105, %f19, %f634, %f631;
sub.f32 %f635, %f103, %f10;
sub.f32 %f636, %f104, %f11;
sub.f32 %f637, %f105, %f12;
mul.f32 %f638, %f636, %f636;
fma.rn.f32 %f639, %f635, %f635, %f638;
fma.rn.f32 %f640, %f637, %f637, %f639;
sqrt.rn.f32 %f596, %f640;
rcp.rn.f32 %f641, %f596;
mul.f32 %f592, %f641, %f635;
mul.f32 %f593, %f641, %f636;
mul.f32 %f594, %f641, %f637;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r275, [root];
// inline asm
call _rt_trace_64, (%r275, %f10, %f11, %f12, %f592, %f593, %f594, %r252, %f426, %f596, %rd75, %r127);
// inline asm
ld.local.f32 %f642, [%rd16];
setp.neu.f32 %p69, %f642, 0f3F800000;
@%p69 bra BB0_66;
ld.local.f32 %f643, [%rd16+4];
setp.neu.f32 %p70, %f643, 0f3F800000;
@%p70 bra BB0_66;
ld.local.f32 %f644, [%rd16+8];
setp.neu.f32 %p71, %f644, 0f3F800000;
@%p71 bra BB0_66;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r282, [root];
mov.f32 %f652, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r282, %f103, %f104, %f105, %f77, %f78, %f79, %r252, %f426, %f652, %rd75, %r127);
// inline asm
ld.local.f32 %f653, [%rd16];
add.f32 %f39, %f39, %f653;
ld.local.f32 %f654, [%rd16+4];
add.f32 %f38, %f38, %f654;
ld.local.f32 %f655, [%rd16+8];
add.f32 %f37, %f37, %f655;
add.f32 %f36, %f36, 0f3F800000;
mov.u16 %rs174, 1;
BB0_66:
mad.lo.s32 %r289, %r85, 1664525, 1013904223;
and.b32 %r290, %r289, 16777215;
cvt.rn.f32.u32 %f664, %r290;
mul.f32 %f665, %f664, 0f33800000;
fma.rn.f32 %f666, %f665, 0f40000000, 0fBF800000;
mad.lo.s32 %r86, %r289, 1664525, 1013904223;
and.b32 %r291, %r86, 16777215;
cvt.rn.f32.u32 %f667, %r291;
mul.f32 %f668, %f667, 0f33800000;
fma.rn.f32 %f669, %f668, 0f40000000, 0fBF800000;
mul.f32 %f670, %f13, %f669;
mul.f32 %f671, %f14, %f669;
mul.f32 %f672, %f15, %f669;
fma.rn.f32 %f673, %f16, %f666, %f670;
fma.rn.f32 %f674, %f17, %f666, %f671;
fma.rn.f32 %f675, %f18, %f666, %f672;
mul.f32 %f676, %f674, %f674;
fma.rn.f32 %f677, %f673, %f673, %f676;
fma.rn.f32 %f678, %f675, %f675, %f677;
sqrt.rn.f32 %f679, %f678;
rcp.rn.f32 %f680, %f679;
mul.f32 %f681, %f673, %f680;
mul.f32 %f682, %f674, %f680;
mul.f32 %f683, %f675, %f680;
mul.f32 %f684, %f1440, %f683;
mul.f32 %f685, %f1441, %f682;
sub.f32 %f686, %f684, %f685;
mul.f32 %f687, %f1441, %f681;
mul.f32 %f688, %f1439, %f683;
sub.f32 %f689, %f687, %f688;
mul.f32 %f690, %f1439, %f682;
mul.f32 %f691, %f1440, %f681;
sub.f32 %f692, %f690, %f691;
mul.f32 %f693, %f681, 0fBECECEDE;
mul.f32 %f694, %f682, 0fBECECEDE;
mul.f32 %f695, %f683, 0fBECECEDE;
fma.rn.f32 %f696, %f19, %f693, %f10;
fma.rn.f32 %f697, %f19, %f694, %f11;
fma.rn.f32 %f698, %f19, %f695, %f12;
mul.f32 %f699, %f686, 0fBF23A3A9;
mul.f32 %f700, %f689, 0fBF23A3A9;
mul.f32 %f701, %f692, 0fBF23A3A9;
fma.rn.f32 %f114, %f19, %f699, %f696;
fma.rn.f32 %f115, %f19, %f700, %f697;
fma.rn.f32 %f116, %f19, %f701, %f698;
sub.f32 %f702, %f114, %f10;
sub.f32 %f703, %f115, %f11;
sub.f32 %f704, %f116, %f12;
mul.f32 %f705, %f703, %f703;
fma.rn.f32 %f706, %f702, %f702, %f705;
fma.rn.f32 %f707, %f704, %f704, %f706;
sqrt.rn.f32 %f663, %f707;
rcp.rn.f32 %f708, %f663;
mul.f32 %f659, %f708, %f702;
mul.f32 %f660, %f708, %f703;
mul.f32 %f661, %f708, %f704;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r286, [root];
// inline asm
call _rt_trace_64, (%r286, %f10, %f11, %f12, %f659, %f660, %f661, %r252, %f426, %f663, %rd75, %r127);
// inline asm
ld.local.f32 %f709, [%rd16];
setp.neu.f32 %p72, %f709, 0f3F800000;
@%p72 bra BB0_70;
ld.local.f32 %f710, [%rd16+4];
setp.neu.f32 %p73, %f710, 0f3F800000;
@%p73 bra BB0_70;
ld.local.f32 %f711, [%rd16+8];
setp.neu.f32 %p74, %f711, 0f3F800000;
@%p74 bra BB0_70;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r293, [root];
mov.f32 %f719, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r293, %f114, %f115, %f116, %f77, %f78, %f79, %r252, %f426, %f719, %rd75, %r127);
// inline asm
ld.local.f32 %f720, [%rd16];
add.f32 %f39, %f39, %f720;
ld.local.f32 %f721, [%rd16+4];
add.f32 %f38, %f38, %f721;
ld.local.f32 %f722, [%rd16+8];
add.f32 %f37, %f37, %f722;
add.f32 %f36, %f36, 0f3F800000;
mov.u16 %rs174, 1;
BB0_70:
mad.lo.s32 %r300, %r86, 1664525, 1013904223;
and.b32 %r301, %r300, 16777215;
cvt.rn.f32.u32 %f731, %r301;
mul.f32 %f732, %f731, 0f33800000;
fma.rn.f32 %f733, %f732, 0f40000000, 0fBF800000;
mad.lo.s32 %r87, %r300, 1664525, 1013904223;
and.b32 %r302, %r87, 16777215;
cvt.rn.f32.u32 %f734, %r302;
mul.f32 %f735, %f734, 0f33800000;
fma.rn.f32 %f736, %f735, 0f40000000, 0fBF800000;
mul.f32 %f737, %f13, %f736;
mul.f32 %f738, %f14, %f736;
mul.f32 %f739, %f15, %f736;
fma.rn.f32 %f740, %f16, %f733, %f737;
fma.rn.f32 %f741, %f17, %f733, %f738;
fma.rn.f32 %f742, %f18, %f733, %f739;
mul.f32 %f743, %f741, %f741;
fma.rn.f32 %f744, %f740, %f740, %f743;
fma.rn.f32 %f745, %f742, %f742, %f744;
sqrt.rn.f32 %f746, %f745;
rcp.rn.f32 %f747, %f746;
mul.f32 %f748, %f740, %f747;
mul.f32 %f749, %f741, %f747;
mul.f32 %f750, %f742, %f747;
mul.f32 %f751, %f1440, %f750;
mul.f32 %f752, %f1441, %f749;
sub.f32 %f753, %f751, %f752;
mul.f32 %f754, %f1441, %f748;
mul.f32 %f755, %f1439, %f750;
sub.f32 %f756, %f754, %f755;
mul.f32 %f757, %f1439, %f749;
mul.f32 %f758, %f1440, %f748;
sub.f32 %f759, %f757, %f758;
mul.f32 %f760, %f748, 0fBF23A3A9;
mul.f32 %f761, %f749, 0fBF23A3A9;
mul.f32 %f762, %f750, 0fBF23A3A9;
fma.rn.f32 %f763, %f19, %f760, %f10;
fma.rn.f32 %f764, %f19, %f761, %f11;
fma.rn.f32 %f765, %f19, %f762, %f12;
mul.f32 %f766, %f753, 0f3F23A3A8;
mul.f32 %f767, %f756, 0f3F23A3A8;
mul.f32 %f768, %f759, 0f3F23A3A8;
fma.rn.f32 %f125, %f19, %f766, %f763;
fma.rn.f32 %f126, %f19, %f767, %f764;
fma.rn.f32 %f127, %f19, %f768, %f765;
sub.f32 %f769, %f125, %f10;
sub.f32 %f770, %f126, %f11;
sub.f32 %f771, %f127, %f12;
mul.f32 %f772, %f770, %f770;
fma.rn.f32 %f773, %f769, %f769, %f772;
fma.rn.f32 %f774, %f771, %f771, %f773;
sqrt.rn.f32 %f730, %f774;
rcp.rn.f32 %f775, %f730;
mul.f32 %f726, %f775, %f769;
mul.f32 %f727, %f775, %f770;
mul.f32 %f728, %f775, %f771;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r297, [root];
// inline asm
call _rt_trace_64, (%r297, %f10, %f11, %f12, %f726, %f727, %f728, %r252, %f426, %f730, %rd75, %r127);
// inline asm
ld.local.f32 %f776, [%rd16];
setp.neu.f32 %p75, %f776, 0f3F800000;
@%p75 bra BB0_74;
ld.local.f32 %f777, [%rd16+4];
setp.neu.f32 %p76, %f777, 0f3F800000;
@%p76 bra BB0_74;
ld.local.f32 %f778, [%rd16+8];
setp.neu.f32 %p77, %f778, 0f3F800000;
@%p77 bra BB0_74;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r304, [root];
mov.f32 %f786, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r304, %f125, %f126, %f127, %f77, %f78, %f79, %r252, %f426, %f786, %rd75, %r127);
// inline asm
ld.local.f32 %f787, [%rd16];
add.f32 %f39, %f39, %f787;
ld.local.f32 %f788, [%rd16+4];
add.f32 %f38, %f38, %f788;
ld.local.f32 %f789, [%rd16+8];
add.f32 %f37, %f37, %f789;
add.f32 %f36, %f36, 0f3F800000;
mov.u16 %rs174, 1;
BB0_74:
mad.lo.s32 %r311, %r87, 1664525, 1013904223;
and.b32 %r312, %r311, 16777215;
cvt.rn.f32.u32 %f798, %r312;
mul.f32 %f799, %f798, 0f33800000;
fma.rn.f32 %f800, %f799, 0f40000000, 0fBF800000;
mad.lo.s32 %r88, %r311, 1664525, 1013904223;
and.b32 %r313, %r88, 16777215;
cvt.rn.f32.u32 %f801, %r313;
mul.f32 %f802, %f801, 0f33800000;
fma.rn.f32 %f803, %f802, 0f40000000, 0fBF800000;
mul.f32 %f804, %f13, %f803;
mul.f32 %f805, %f14, %f803;
mul.f32 %f806, %f15, %f803;
fma.rn.f32 %f807, %f16, %f800, %f804;
fma.rn.f32 %f808, %f17, %f800, %f805;
fma.rn.f32 %f809, %f18, %f800, %f806;
mul.f32 %f810, %f808, %f808;
fma.rn.f32 %f811, %f807, %f807, %f810;
fma.rn.f32 %f812, %f809, %f809, %f811;
sqrt.rn.f32 %f813, %f812;
rcp.rn.f32 %f814, %f813;
mul.f32 %f815, %f807, %f814;
mul.f32 %f816, %f808, %f814;
mul.f32 %f817, %f809, %f814;
mul.f32 %f818, %f1440, %f817;
mul.f32 %f819, %f1441, %f816;
sub.f32 %f820, %f818, %f819;
mul.f32 %f821, %f1441, %f815;
mul.f32 %f822, %f1439, %f817;
sub.f32 %f823, %f821, %f822;
mul.f32 %f824, %f1439, %f816;
mul.f32 %f825, %f1440, %f815;
sub.f32 %f826, %f824, %f825;
mul.f32 %f827, %f815, 0fBF61E1E3;
mul.f32 %f828, %f816, 0fBF61E1E3;
mul.f32 %f829, %f817, 0fBF61E1E3;
fma.rn.f32 %f830, %f19, %f827, %f10;
fma.rn.f32 %f831, %f19, %f828, %f11;
fma.rn.f32 %f832, %f19, %f829, %f12;
mul.f32 %f833, %f820, 0fBE24A490;
mul.f32 %f834, %f823, 0fBE24A490;
mul.f32 %f835, %f826, 0fBE24A490;
fma.rn.f32 %f136, %f19, %f833, %f830;
fma.rn.f32 %f137, %f19, %f834, %f831;
fma.rn.f32 %f138, %f19, %f835, %f832;
sub.f32 %f836, %f136, %f10;
sub.f32 %f837, %f137, %f11;
sub.f32 %f838, %f138, %f12;
mul.f32 %f839, %f837, %f837;
fma.rn.f32 %f840, %f836, %f836, %f839;
fma.rn.f32 %f841, %f838, %f838, %f840;
sqrt.rn.f32 %f797, %f841;
rcp.rn.f32 %f842, %f797;
mul.f32 %f793, %f842, %f836;
mul.f32 %f794, %f842, %f837;
mul.f32 %f795, %f842, %f838;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r308, [root];
// inline asm
call _rt_trace_64, (%r308, %f10, %f11, %f12, %f793, %f794, %f795, %r252, %f426, %f797, %rd75, %r127);
// inline asm
ld.local.f32 %f843, [%rd16];
setp.neu.f32 %p78, %f843, 0f3F800000;
@%p78 bra BB0_78;
ld.local.f32 %f844, [%rd16+4];
setp.neu.f32 %p79, %f844, 0f3F800000;
@%p79 bra BB0_78;
ld.local.f32 %f845, [%rd16+8];
setp.neu.f32 %p80, %f845, 0f3F800000;
@%p80 bra BB0_78;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r315, [root];
mov.f32 %f853, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r315, %f136, %f137, %f138, %f77, %f78, %f79, %r252, %f426, %f853, %rd75, %r127);
// inline asm
ld.local.f32 %f854, [%rd16];
add.f32 %f39, %f39, %f854;
ld.local.f32 %f855, [%rd16+4];
add.f32 %f38, %f38, %f855;
ld.local.f32 %f856, [%rd16+8];
add.f32 %f37, %f37, %f856;
add.f32 %f36, %f36, 0f3F800000;
mov.u16 %rs174, 1;
BB0_78:
mad.lo.s32 %r322, %r88, 1664525, 1013904223;
and.b32 %r323, %r322, 16777215;
cvt.rn.f32.u32 %f865, %r323;
mul.f32 %f866, %f865, 0f33800000;
fma.rn.f32 %f867, %f866, 0f40000000, 0fBF800000;
mad.lo.s32 %r89, %r322, 1664525, 1013904223;
and.b32 %r324, %r89, 16777215;
cvt.rn.f32.u32 %f868, %r324;
mul.f32 %f869, %f868, 0f33800000;
fma.rn.f32 %f870, %f869, 0f40000000, 0fBF800000;
mul.f32 %f871, %f13, %f870;
mul.f32 %f872, %f14, %f870;
mul.f32 %f873, %f15, %f870;
fma.rn.f32 %f874, %f16, %f867, %f871;
fma.rn.f32 %f875, %f17, %f867, %f872;
fma.rn.f32 %f876, %f18, %f867, %f873;
mul.f32 %f877, %f875, %f875;
fma.rn.f32 %f878, %f874, %f874, %f877;
fma.rn.f32 %f879, %f876, %f876, %f878;
sqrt.rn.f32 %f880, %f879;
rcp.rn.f32 %f881, %f880;
mul.f32 %f882, %f874, %f881;
mul.f32 %f883, %f875, %f881;
mul.f32 %f884, %f876, %f881;
mul.f32 %f885, %f1440, %f884;
mul.f32 %f886, %f1441, %f883;
sub.f32 %f887, %f885, %f886;
mul.f32 %f888, %f1441, %f882;
mul.f32 %f889, %f1439, %f884;
sub.f32 %f890, %f888, %f889;
mul.f32 %f891, %f1439, %f883;
mul.f32 %f892, %f1440, %f882;
sub.f32 %f893, %f891, %f892;
mul.f32 %f894, %f882, 0f3ECACABC;
mul.f32 %f895, %f883, 0f3ECACABC;
mul.f32 %f896, %f884, 0f3ECACABC;
fma.rn.f32 %f897, %f19, %f894, %f10;
fma.rn.f32 %f898, %f19, %f895, %f11;
fma.rn.f32 %f899, %f19, %f896, %f12;
mul.f32 %f900, %f887, 0f3F61E1D2;
mul.f32 %f901, %f890, 0f3F61E1D2;
mul.f32 %f902, %f893, 0f3F61E1D2;
fma.rn.f32 %f147, %f19, %f900, %f897;
fma.rn.f32 %f148, %f19, %f901, %f898;
fma.rn.f32 %f149, %f19, %f902, %f899;
sub.f32 %f903, %f147, %f10;
sub.f32 %f904, %f148, %f11;
sub.f32 %f905, %f149, %f12;
mul.f32 %f906, %f904, %f904;
fma.rn.f32 %f907, %f903, %f903, %f906;
fma.rn.f32 %f908, %f905, %f905, %f907;
sqrt.rn.f32 %f864, %f908;
rcp.rn.f32 %f909, %f864;
mul.f32 %f860, %f909, %f903;
mul.f32 %f861, %f909, %f904;
mul.f32 %f862, %f909, %f905;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r319, [root];
// inline asm
call _rt_trace_64, (%r319, %f10, %f11, %f12, %f860, %f861, %f862, %r252, %f426, %f864, %rd75, %r127);
// inline asm
ld.local.f32 %f910, [%rd16];
setp.neu.f32 %p81, %f910, 0f3F800000;
@%p81 bra BB0_82;
ld.local.f32 %f911, [%rd16+4];
setp.neu.f32 %p82, %f911, 0f3F800000;
@%p82 bra BB0_82;
ld.local.f32 %f912, [%rd16+8];
setp.neu.f32 %p83, %f912, 0f3F800000;
@%p83 bra BB0_82;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r326, [root];
mov.f32 %f920, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r326, %f147, %f148, %f149, %f77, %f78, %f79, %r252, %f426, %f920, %rd75, %r127);
// inline asm
ld.local.f32 %f921, [%rd16];
add.f32 %f39, %f39, %f921;
ld.local.f32 %f922, [%rd16+4];
add.f32 %f38, %f38, %f922;
ld.local.f32 %f923, [%rd16+8];
add.f32 %f37, %f37, %f923;
add.f32 %f36, %f36, 0f3F800000;
mov.u16 %rs174, 1;
BB0_82:
mad.lo.s32 %r333, %r89, 1664525, 1013904223;
and.b32 %r334, %r333, 16777215;
cvt.rn.f32.u32 %f932, %r334;
mul.f32 %f933, %f932, 0f33800000;
fma.rn.f32 %f934, %f933, 0f40000000, 0fBF800000;
mad.lo.s32 %r497, %r333, 1664525, 1013904223;
and.b32 %r335, %r497, 16777215;
cvt.rn.f32.u32 %f935, %r335;
mul.f32 %f936, %f935, 0f33800000;
fma.rn.f32 %f937, %f936, 0f40000000, 0fBF800000;
mul.f32 %f938, %f13, %f937;
mul.f32 %f939, %f14, %f937;
mul.f32 %f940, %f15, %f937;
fma.rn.f32 %f941, %f16, %f934, %f938;
fma.rn.f32 %f942, %f17, %f934, %f939;
fma.rn.f32 %f943, %f18, %f934, %f940;
mul.f32 %f944, %f942, %f942;
fma.rn.f32 %f945, %f941, %f941, %f944;
fma.rn.f32 %f946, %f943, %f943, %f945;
sqrt.rn.f32 %f947, %f946;
rcp.rn.f32 %f948, %f947;
mul.f32 %f949, %f941, %f948;
mul.f32 %f950, %f942, %f948;
mul.f32 %f951, %f943, %f948;
mul.f32 %f952, %f1440, %f951;
mul.f32 %f953, %f1441, %f950;
sub.f32 %f954, %f952, %f953;
mul.f32 %f955, %f1441, %f949;
mul.f32 %f956, %f1439, %f951;
sub.f32 %f957, %f955, %f956;
mul.f32 %f958, %f1439, %f950;
mul.f32 %f959, %f1440, %f949;
sub.f32 %f960, %f958, %f959;
mul.f32 %f961, %f949, 0f3F61E1D2;
mul.f32 %f962, %f950, 0f3F61E1D2;
mul.f32 %f963, %f951, 0f3F61E1D2;
fma.rn.f32 %f964, %f19, %f961, %f10;
fma.rn.f32 %f965, %f19, %f962, %f11;
fma.rn.f32 %f966, %f19, %f963, %f12;
mul.f32 %f967, %f954, 0fBF61E1E3;
mul.f32 %f968, %f957, 0fBF61E1E3;
mul.f32 %f969, %f960, 0fBF61E1E3;
fma.rn.f32 %f158, %f19, %f967, %f964;
fma.rn.f32 %f159, %f19, %f968, %f965;
fma.rn.f32 %f160, %f19, %f969, %f966;
sub.f32 %f970, %f158, %f10;
sub.f32 %f971, %f159, %f11;
sub.f32 %f972, %f160, %f12;
mul.f32 %f973, %f971, %f971;
fma.rn.f32 %f974, %f970, %f970, %f973;
fma.rn.f32 %f975, %f972, %f972, %f974;
sqrt.rn.f32 %f931, %f975;
rcp.rn.f32 %f976, %f931;
mul.f32 %f927, %f976, %f970;
mul.f32 %f928, %f976, %f971;
mul.f32 %f929, %f976, %f972;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r330, [root];
// inline asm
call _rt_trace_64, (%r330, %f10, %f11, %f12, %f927, %f928, %f929, %r252, %f426, %f931, %rd75, %r127);
// inline asm
ld.local.f32 %f977, [%rd16];
setp.neu.f32 %p84, %f977, 0f3F800000;
@%p84 bra BB0_86;
ld.local.f32 %f978, [%rd16+4];
setp.neu.f32 %p85, %f978, 0f3F800000;
@%p85 bra BB0_86;
ld.local.f32 %f979, [%rd16+8];
setp.neu.f32 %p86, %f979, 0f3F800000;
@%p86 bra BB0_86;
st.local.u32 [%rd16+8], %r259;
st.local.u32 [%rd16+4], %r259;
st.local.u32 [%rd16], %r259;
ld.global.u32 %r337, [root];
mov.f32 %f987, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r337, %f158, %f159, %f160, %f77, %f78, %f79, %r252, %f426, %f987, %rd75, %r127);
// inline asm
ld.local.f32 %f988, [%rd16];
add.f32 %f39, %f39, %f988;
ld.local.f32 %f989, [%rd16+4];
add.f32 %f38, %f38, %f989;
ld.local.f32 %f990, [%rd16+8];
add.f32 %f37, %f37, %f990;
add.f32 %f36, %f36, 0f3F800000;
mov.u16 %rs174, 1;
BB0_86:
setp.ne.s16 %p87, %rs174, 0;
@%p87 bra BB0_88;
add.u64 %rd91, %SP, 0;
cvta.to.local.u64 %rd92, %rd91;
st.local.u32 [%rd92+8], %r259;
st.local.u32 [%rd92+4], %r259;
st.local.u32 [%rd92], %r259;
ld.global.u32 %r341, [root];
mov.f32 %f998, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r341, %f10, %f11, %f12, %f77, %f78, %f79, %r252, %f426, %f998, %rd91, %r127);
// inline asm
ld.local.f32 %f999, [%rd92];
add.f32 %f39, %f39, %f999;
ld.local.f32 %f1000, [%rd92+4];
add.f32 %f38, %f38, %f1000;
ld.local.f32 %f1001, [%rd92+8];
add.f32 %f37, %f37, %f1001;
add.f32 %f36, %f36, 0f3F800000;
BB0_88:
ld.global.u32 %r471, [samples];
add.s32 %r474, %r474, 1;
setp.lt.s32 %p88, %r474, %r471;
@%p88 bra BB0_8;
BB0_89:
add.s32 %r472, %r472, 1;
setp.lt.s32 %p89, %r472, %r471;
@%p89 bra BB0_6;
bra.uni BB0_90;
BB0_4:
mov.f32 %f37, %f36;
mov.f32 %f38, %f36;
mov.f32 %f39, %f36;
BB0_90:
mul.f32 %f1005, %f1441, %f9;
mul.f32 %f1006, %f1440, %f8;
neg.f32 %f1007, %f1006;
mul.f32 %f1008, %f1439, %f7;
sub.f32 %f1009, %f1007, %f1008;
sub.f32 %f185, %f1009, %f1005;
setp.eq.s32 %p90, %r471, 0;
mov.f32 %f1506, 0f3F800000;
mov.f32 %f1507, %f1506;
mov.f32 %f1508, %f1506;
@%p90 bra BB0_92;
mov.f32 %f1010, 0f3F800000;
max.f32 %f1011, %f36, %f1010;
rcp.rn.f32 %f1012, %f1011;
mul.f32 %f1506, %f39, %f1012;
mul.f32 %f1507, %f38, %f1012;
mul.f32 %f1508, %f37, %f1012;
BB0_92:
ld.global.f32 %f1013, [directColor];
mul.f32 %f1014, %f1506, %f1013;
ld.global.f32 %f1015, [directColor+4];
mul.f32 %f1016, %f1507, %f1015;
ld.global.f32 %f1017, [directColor+8];
mul.f32 %f1018, %f1508, %f1017;
cvt.sat.f32.f32 %f1019, %f185;
mul.f32 %f192, %f1014, %f1019;
mul.f32 %f193, %f1016, %f1019;
mul.f32 %f194, %f1018, %f1019;
mul.f32 %f1020, %f185, 0f40800000;
cvt.sat.f32.f32 %f1021, %f1020;
mul.f32 %f1022, %f1014, %f1021;
mul.f32 %f1023, %f1016, %f1021;
mul.f32 %f1024, %f1018, %f1021;
mul.f32 %f195, %f1022, 0f3E800000;
mul.f32 %f196, %f1023, 0f3E800000;
mul.f32 %f197, %f1024, 0f3E800000;
ld.global.u32 %r500, [imageEnabled];
and.b32 %r345, %r500, 8;
setp.eq.s32 %p91, %r345, 0;
@%p91 bra BB0_105;
cvt.u64.u32 %rd95, %r2;
cvt.u64.u32 %rd96, %r3;
mov.u64 %rd99, image_Mask;
cvta.global.u64 %rd94, %rd99;
// inline asm
call (%rd93), _rt_buffer_get_64, (%rd94, %r107, %r107, %rd95, %rd96, %rd27, %rd27);
// inline asm
mov.f32 %f1027, 0f3E68BA2E;
cvt.rzi.f32.f32 %f1028, %f1027;
fma.rn.f32 %f1029, %f1028, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f198, %f1029;
abs.f32 %f199, %f1506;
setp.lt.f32 %p92, %f199, 0f00800000;
mul.f32 %f1030, %f199, 0f4B800000;
selp.f32 %f1031, 0fC3170000, 0fC2FE0000, %p92;
selp.f32 %f1032, %f1030, %f199, %p92;
mov.b32 %r348, %f1032;
and.b32 %r349, %r348, 8388607;
or.b32 %r350, %r349, 1065353216;
mov.b32 %f1033, %r350;
shr.u32 %r351, %r348, 23;
cvt.rn.f32.u32 %f1034, %r351;
add.f32 %f1035, %f1031, %f1034;
setp.gt.f32 %p93, %f1033, 0f3FB504F3;
mul.f32 %f1036, %f1033, 0f3F000000;
add.f32 %f1037, %f1035, 0f3F800000;
selp.f32 %f1038, %f1036, %f1033, %p93;
selp.f32 %f1039, %f1037, %f1035, %p93;
add.f32 %f1040, %f1038, 0fBF800000;
add.f32 %f1026, %f1038, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1025,%f1026;
// inline asm
add.f32 %f1041, %f1040, %f1040;
mul.f32 %f1042, %f1025, %f1041;
mul.f32 %f1043, %f1042, %f1042;
mov.f32 %f1044, 0f3C4CAF63;
mov.f32 %f1045, 0f3B18F0FE;
fma.rn.f32 %f1046, %f1045, %f1043, %f1044;
mov.f32 %f1047, 0f3DAAAABD;
fma.rn.f32 %f1048, %f1046, %f1043, %f1047;
mul.rn.f32 %f1049, %f1048, %f1043;
mul.rn.f32 %f1050, %f1049, %f1042;
sub.f32 %f1051, %f1040, %f1042;
neg.f32 %f1052, %f1042;
add.f32 %f1053, %f1051, %f1051;
fma.rn.f32 %f1054, %f1052, %f1040, %f1053;
mul.rn.f32 %f1055, %f1025, %f1054;
add.f32 %f1056, %f1050, %f1042;
sub.f32 %f1057, %f1042, %f1056;
add.f32 %f1058, %f1050, %f1057;
add.f32 %f1059, %f1055, %f1058;
add.f32 %f1060, %f1056, %f1059;
sub.f32 %f1061, %f1056, %f1060;
add.f32 %f1062, %f1059, %f1061;
mov.f32 %f1063, 0f3F317200;
mul.rn.f32 %f1064, %f1039, %f1063;
mov.f32 %f1065, 0f35BFBE8E;
mul.rn.f32 %f1066, %f1039, %f1065;
add.f32 %f1067, %f1064, %f1060;
sub.f32 %f1068, %f1064, %f1067;
add.f32 %f1069, %f1060, %f1068;
add.f32 %f1070, %f1062, %f1069;
add.f32 %f1071, %f1066, %f1070;
add.f32 %f1072, %f1067, %f1071;
sub.f32 %f1073, %f1067, %f1072;
add.f32 %f1074, %f1071, %f1073;
mov.f32 %f1075, 0f3EE8BA2E;
mul.rn.f32 %f1076, %f1075, %f1072;
neg.f32 %f1077, %f1076;
fma.rn.f32 %f1078, %f1075, %f1072, %f1077;
fma.rn.f32 %f1079, %f1075, %f1074, %f1078;
mov.f32 %f1080, 0f00000000;
fma.rn.f32 %f1081, %f1080, %f1072, %f1079;
add.rn.f32 %f1082, %f1076, %f1081;
neg.f32 %f1083, %f1082;
add.rn.f32 %f1084, %f1076, %f1083;
add.rn.f32 %f1085, %f1084, %f1081;
mov.b32 %r352, %f1082;
setp.eq.s32 %p94, %r352, 1118925336;
add.s32 %r353, %r352, -1;
mov.b32 %f1086, %r353;
add.f32 %f1087, %f1085, 0f37000000;
selp.f32 %f1088, %f1086, %f1082, %p94;
selp.f32 %f200, %f1087, %f1085, %p94;
mul.f32 %f1089, %f1088, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1090, %f1089;
mov.f32 %f1091, 0fBF317200;
fma.rn.f32 %f1092, %f1090, %f1091, %f1088;
mov.f32 %f1093, 0fB5BFBE8E;
fma.rn.f32 %f1094, %f1090, %f1093, %f1092;
mul.f32 %f1095, %f1094, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1096, %f1095;
add.f32 %f1097, %f1090, 0f00000000;
ex2.approx.f32 %f1098, %f1097;
mul.f32 %f1099, %f1096, %f1098;
setp.lt.f32 %p95, %f1088, 0fC2D20000;
selp.f32 %f1100, 0f00000000, %f1099, %p95;
setp.gt.f32 %p96, %f1088, 0f42D20000;
selp.f32 %f1509, 0f7F800000, %f1100, %p96;
setp.eq.f32 %p97, %f1509, 0f7F800000;
@%p97 bra BB0_95;
fma.rn.f32 %f1509, %f1509, %f200, %f1509;
BB0_95:
setp.lt.f32 %p98, %f1506, 0f00000000;
setp.eq.f32 %p99, %f198, 0f3F800000;
and.pred %p1, %p98, %p99;
mov.b32 %r354, %f1509;
xor.b32 %r355, %r354, -2147483648;
mov.b32 %f1101, %r355;
selp.f32 %f1511, %f1101, %f1509, %p1;
setp.eq.f32 %p100, %f1506, 0f00000000;
@%p100 bra BB0_98;
bra.uni BB0_96;
BB0_98:
add.f32 %f1104, %f1506, %f1506;
selp.f32 %f1511, %f1104, 0f00000000, %p99;
bra.uni BB0_99;
BB0_163:
mov.u64 %rd240, image_HDR;
cvta.global.u64 %rd235, %rd240;
mov.u32 %r441, 8;
mov.u64 %rd239, 0;
// inline asm
call (%rd234), _rt_buffer_get_64, (%rd235, %r107, %r441, %rd20, %rd21, %rd239, %rd239);
// inline asm
mov.f32 %f1408, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs117, %f1408;}
// inline asm
mov.u16 %rs118, 0;
st.v4.u16 [%rd234], {%rs117, %rs117, %rs117, %rs118};
BB0_164:
ld.global.u32 %r442, [additive];
setp.eq.s32 %p167, %r442, 0;
@%p167 bra BB0_166;
mov.u64 %rd253, image_RNM0;
cvta.global.u64 %rd242, %rd253;
mov.u32 %r446, 8;
mov.u64 %rd252, 0;
// inline asm
call (%rd241), _rt_buffer_get_64, (%rd242, %r107, %r446, %rd20, %rd21, %rd252, %rd252);
// inline asm
ld.v4.u16 {%rs125, %rs126, %rs127, %rs128}, [%rd241];
// inline asm
{ cvt.f32.f16 %f1409, %rs125;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1410, %rs126;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1411, %rs127;}
// inline asm
// inline asm
call (%rd247), _rt_buffer_get_64, (%rd242, %r107, %r446, %rd20, %rd21, %rd252, %rd252);
// inline asm
add.f32 %f1412, %f1409, 0f00000000;
add.f32 %f1413, %f1410, 0f00000000;
add.f32 %f1414, %f1411, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs124, %f1414;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs123, %f1413;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs122, %f1412;}
// inline asm
mov.u16 %rs129, 0;
st.v4.u16 [%rd247], {%rs122, %rs123, %rs124, %rs129};
bra.uni BB0_167;
BB0_166:
mov.u64 %rd260, image_RNM0;
cvta.global.u64 %rd255, %rd260;
mov.u32 %r448, 8;
mov.u64 %rd259, 0;
// inline asm
call (%rd254), _rt_buffer_get_64, (%rd255, %r107, %r448, %rd20, %rd21, %rd259, %rd259);
// inline asm
mov.f32 %f1415, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs130, %f1415;}
// inline asm
mov.u16 %rs131, 0;
st.v4.u16 [%rd254], {%rs130, %rs130, %rs130, %rs131};
BB0_167:
ld.global.u32 %r449, [additive];
setp.eq.s32 %p168, %r449, 0;
@%p168 bra BB0_169;
mov.u64 %rd273, image_RNM1;
cvta.global.u64 %rd262, %rd273;
mov.u32 %r453, 8;
mov.u64 %rd272, 0;
// inline asm
call (%rd261), _rt_buffer_get_64, (%rd262, %r107, %r453, %rd20, %rd21, %rd272, %rd272);
// inline asm
ld.v4.u16 {%rs138, %rs139, %rs140, %rs141}, [%rd261];
// inline asm
{ cvt.f32.f16 %f1416, %rs138;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1417, %rs139;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1418, %rs140;}
// inline asm
// inline asm
call (%rd267), _rt_buffer_get_64, (%rd262, %r107, %r453, %rd20, %rd21, %rd272, %rd272);
// inline asm
add.f32 %f1419, %f1416, 0f00000000;
add.f32 %f1420, %f1417, 0f00000000;
add.f32 %f1421, %f1418, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs137, %f1421;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs136, %f1420;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs135, %f1419;}
// inline asm
mov.u16 %rs142, 0;
st.v4.u16 [%rd267], {%rs135, %rs136, %rs137, %rs142};
bra.uni BB0_170;
BB0_169:
mov.u64 %rd280, image_RNM1;
cvta.global.u64 %rd275, %rd280;
mov.u32 %r455, 8;
mov.u64 %rd279, 0;
// inline asm
call (%rd274), _rt_buffer_get_64, (%rd275, %r107, %r455, %rd20, %rd21, %rd279, %rd279);
// inline asm
mov.f32 %f1422, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs143, %f1422;}
// inline asm
mov.u16 %rs144, 0;
st.v4.u16 [%rd274], {%rs143, %rs143, %rs143, %rs144};
BB0_170:
ld.global.u32 %r456, [additive];
setp.eq.s32 %p169, %r456, 0;
@%p169 bra BB0_172;
mov.u64 %rd293, image_RNM2;
cvta.global.u64 %rd282, %rd293;
mov.u32 %r460, 8;
mov.u64 %rd292, 0;
// inline asm
call (%rd281), _rt_buffer_get_64, (%rd282, %r107, %r460, %rd20, %rd21, %rd292, %rd292);
// inline asm
ld.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd281];
// inline asm
{ cvt.f32.f16 %f1423, %rs151;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1424, %rs152;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1425, %rs153;}
// inline asm
// inline asm
call (%rd287), _rt_buffer_get_64, (%rd282, %r107, %r460, %rd20, %rd21, %rd292, %rd292);
// inline asm
add.f32 %f1426, %f1423, 0f00000000;
add.f32 %f1427, %f1424, 0f00000000;
add.f32 %f1428, %f1425, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs150, %f1428;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs149, %f1427;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs148, %f1426;}
// inline asm
mov.u16 %rs155, 0;
st.v4.u16 [%rd287], {%rs148, %rs149, %rs150, %rs155};
bra.uni BB0_173;
BB0_172:
mov.u64 %rd300, image_RNM2;
cvta.global.u64 %rd295, %rd300;
mov.u32 %r462, 8;
mov.u64 %rd299, 0;
// inline asm
call (%rd294), _rt_buffer_get_64, (%rd295, %r107, %r462, %rd20, %rd21, %rd299, %rd299);
// inline asm
mov.f32 %f1429, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs156, %f1429;}
// inline asm
mov.u16 %rs157, 0;
st.v4.u16 [%rd294], {%rs156, %rs156, %rs156, %rs157};
BB0_173:
ld.global.u32 %r463, [additive];
setp.eq.s32 %p170, %r463, 0;
@%p170 bra BB0_175;
mov.u64 %rd313, image_RNM3;
cvta.global.u64 %rd302, %rd313;
mov.u32 %r467, 8;
mov.u64 %rd312, 0;
// inline asm
call (%rd301), _rt_buffer_get_64, (%rd302, %r107, %r467, %rd20, %rd21, %rd312, %rd312);
// inline asm
ld.v4.u16 {%rs164, %rs165, %rs166, %rs167}, [%rd301];
// inline asm
{ cvt.f32.f16 %f1430, %rs164;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1431, %rs165;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1432, %rs166;}
// inline asm
// inline asm
call (%rd307), _rt_buffer_get_64, (%rd302, %r107, %r467, %rd20, %rd21, %rd312, %rd312);
// inline asm
add.f32 %f1433, %f1430, 0f00000000;
add.f32 %f1434, %f1431, 0f00000000;
add.f32 %f1435, %f1432, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs163, %f1435;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs162, %f1434;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs161, %f1433;}
// inline asm
mov.u16 %rs168, 0;
st.v4.u16 [%rd307], {%rs161, %rs162, %rs163, %rs168};
bra.uni BB0_176;
BB0_175:
mov.u64 %rd320, image_RNM3;
cvta.global.u64 %rd315, %rd320;
mov.u32 %r469, 8;
mov.u64 %rd319, 0;
// inline asm
call (%rd314), _rt_buffer_get_64, (%rd315, %r107, %r469, %rd20, %rd21, %rd319, %rd319);
// inline asm
mov.f32 %f1436, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs169, %f1436;}
// inline asm
mov.u16 %rs170, 0;
st.v4.u16 [%rd314], {%rs169, %rs169, %rs169, %rs170};
bra.uni BB0_176;
BB0_96:
setp.geu.f32 %p101, %f1506, 0f00000000;
@%p101 bra BB0_99;
cvt.rzi.f32.f32 %f1103, %f1075;
setp.neu.f32 %p102, %f1103, 0f3EE8BA2E;
selp.f32 %f1511, 0f7FFFFFFF, %f1511, %p102;
BB0_99:
add.f32 %f1105, %f199, 0f3EE8BA2E;
mov.b32 %r356, %f1105;
setp.lt.s32 %p104, %r356, 2139095040;
@%p104 bra BB0_104;
setp.gtu.f32 %p105, %f199, 0f7F800000;
@%p105 bra BB0_103;
bra.uni BB0_101;
BB0_103:
add.f32 %f1511, %f1506, 0f3EE8BA2E;
bra.uni BB0_104;
BB0_101:
setp.neu.f32 %p106, %f199, 0f7F800000;
@%p106 bra BB0_104;
selp.f32 %f1511, 0fFF800000, 0f7F800000, %p1;
BB0_104:
mul.f32 %f1106, %f1511, 0f437F0000;
setp.eq.f32 %p107, %f1506, 0f3F800000;
selp.f32 %f1107, 0f437F0000, %f1106, %p107;
cvt.rzi.u32.f32 %r357, %f1107;
cvt.u16.u32 %rs33, %r357;
mov.u16 %rs34, 255;
st.v2.u8 [%rd93], {%rs33, %rs34};
ld.global.u32 %r500, [imageEnabled];
BB0_105:
and.b32 %r358, %r500, 1;
setp.eq.b32 %p108, %r358, 1;
@!%p108 bra BB0_140;
bra.uni BB0_106;
BB0_106:
mov.f32 %f1110, 0f3E666666;
cvt.rzi.f32.f32 %f1111, %f1110;
fma.rn.f32 %f1112, %f1111, 0fC0000000, 0f3EE66666;
abs.f32 %f211, %f1112;
abs.f32 %f212, %f192;
setp.lt.f32 %p109, %f212, 0f00800000;
mul.f32 %f1113, %f212, 0f4B800000;
selp.f32 %f1114, 0fC3170000, 0fC2FE0000, %p109;
selp.f32 %f1115, %f1113, %f212, %p109;
mov.b32 %r359, %f1115;
and.b32 %r360, %r359, 8388607;
or.b32 %r361, %r360, 1065353216;
mov.b32 %f1116, %r361;
shr.u32 %r362, %r359, 23;
cvt.rn.f32.u32 %f1117, %r362;
add.f32 %f1118, %f1114, %f1117;
setp.gt.f32 %p110, %f1116, 0f3FB504F3;
mul.f32 %f1119, %f1116, 0f3F000000;
add.f32 %f1120, %f1118, 0f3F800000;
selp.f32 %f1121, %f1119, %f1116, %p110;
selp.f32 %f1122, %f1120, %f1118, %p110;
add.f32 %f1123, %f1121, 0fBF800000;
add.f32 %f1109, %f1121, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1108,%f1109;
// inline asm
add.f32 %f1124, %f1123, %f1123;
mul.f32 %f1125, %f1108, %f1124;
mul.f32 %f1126, %f1125, %f1125;
mov.f32 %f1127, 0f3C4CAF63;
mov.f32 %f1128, 0f3B18F0FE;
fma.rn.f32 %f1129, %f1128, %f1126, %f1127;
mov.f32 %f1130, 0f3DAAAABD;
fma.rn.f32 %f1131, %f1129, %f1126, %f1130;
mul.rn.f32 %f1132, %f1131, %f1126;
mul.rn.f32 %f1133, %f1132, %f1125;
sub.f32 %f1134, %f1123, %f1125;
neg.f32 %f1135, %f1125;
add.f32 %f1136, %f1134, %f1134;
fma.rn.f32 %f1137, %f1135, %f1123, %f1136;
mul.rn.f32 %f1138, %f1108, %f1137;
add.f32 %f1139, %f1133, %f1125;
sub.f32 %f1140, %f1125, %f1139;
add.f32 %f1141, %f1133, %f1140;
add.f32 %f1142, %f1138, %f1141;
add.f32 %f1143, %f1139, %f1142;
sub.f32 %f1144, %f1139, %f1143;
add.f32 %f1145, %f1142, %f1144;
mov.f32 %f1146, 0f3F317200;
mul.rn.f32 %f1147, %f1122, %f1146;
mov.f32 %f1148, 0f35BFBE8E;
mul.rn.f32 %f1149, %f1122, %f1148;
add.f32 %f1150, %f1147, %f1143;
sub.f32 %f1151, %f1147, %f1150;
add.f32 %f1152, %f1143, %f1151;
add.f32 %f1153, %f1145, %f1152;
add.f32 %f1154, %f1149, %f1153;
add.f32 %f1155, %f1150, %f1154;
sub.f32 %f1156, %f1150, %f1155;
add.f32 %f1157, %f1154, %f1156;
mov.f32 %f1158, 0f3EE66666;
mul.rn.f32 %f1159, %f1158, %f1155;
neg.f32 %f1160, %f1159;
fma.rn.f32 %f1161, %f1158, %f1155, %f1160;
fma.rn.f32 %f1162, %f1158, %f1157, %f1161;
mov.f32 %f1163, 0f00000000;
fma.rn.f32 %f1164, %f1163, %f1155, %f1162;
add.rn.f32 %f1165, %f1159, %f1164;
neg.f32 %f1166, %f1165;
add.rn.f32 %f1167, %f1159, %f1166;
add.rn.f32 %f1168, %f1167, %f1164;
mov.b32 %r363, %f1165;
setp.eq.s32 %p111, %r363, 1118925336;
add.s32 %r364, %r363, -1;
mov.b32 %f1169, %r364;
add.f32 %f1170, %f1168, 0f37000000;
selp.f32 %f1171, %f1169, %f1165, %p111;
selp.f32 %f213, %f1170, %f1168, %p111;
mul.f32 %f1172, %f1171, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1173, %f1172;
mov.f32 %f1174, 0fBF317200;
fma.rn.f32 %f1175, %f1173, %f1174, %f1171;
mov.f32 %f1176, 0fB5BFBE8E;
fma.rn.f32 %f1177, %f1173, %f1176, %f1175;
mul.f32 %f1178, %f1177, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1179, %f1178;
add.f32 %f1180, %f1173, 0f00000000;
ex2.approx.f32 %f1181, %f1180;
mul.f32 %f1182, %f1179, %f1181;
setp.lt.f32 %p112, %f1171, 0fC2D20000;
selp.f32 %f1183, 0f00000000, %f1182, %p112;
setp.gt.f32 %p113, %f1171, 0f42D20000;
selp.f32 %f1512, 0f7F800000, %f1183, %p113;
setp.eq.f32 %p114, %f1512, 0f7F800000;
@%p114 bra BB0_108;
fma.rn.f32 %f1512, %f1512, %f213, %f1512;
BB0_108:
setp.lt.f32 %p115, %f192, 0f00000000;
setp.eq.f32 %p116, %f211, 0f3F800000;
and.pred %p2, %p115, %p116;
mov.b32 %r365, %f1512;
xor.b32 %r366, %r365, -2147483648;
mov.b32 %f1184, %r366;
selp.f32 %f1514, %f1184, %f1512, %p2;
setp.eq.f32 %p117, %f192, 0f00000000;
@%p117 bra BB0_111;
bra.uni BB0_109;
BB0_111:
add.f32 %f1187, %f192, %f192;
selp.f32 %f1514, %f1187, 0f00000000, %p116;
bra.uni BB0_112;
BB0_109:
setp.geu.f32 %p118, %f192, 0f00000000;
@%p118 bra BB0_112;
cvt.rzi.f32.f32 %f1186, %f1158;
setp.neu.f32 %p119, %f1186, 0f3EE66666;
selp.f32 %f1514, 0f7FFFFFFF, %f1514, %p119;
BB0_112:
add.f32 %f1188, %f212, 0f3EE66666;
mov.b32 %r367, %f1188;
setp.lt.s32 %p121, %r367, 2139095040;
@%p121 bra BB0_117;
setp.gtu.f32 %p122, %f212, 0f7F800000;
@%p122 bra BB0_116;
bra.uni BB0_114;
BB0_116:
add.f32 %f1514, %f192, 0f3EE66666;
bra.uni BB0_117;
BB0_114:
setp.neu.f32 %p123, %f212, 0f7F800000;
@%p123 bra BB0_117;
selp.f32 %f1514, 0fFF800000, 0f7F800000, %p2;
BB0_117:
setp.eq.f32 %p124, %f192, 0f3F800000;
selp.f32 %f224, 0f3F800000, %f1514, %p124;
abs.f32 %f225, %f193;
setp.lt.f32 %p125, %f225, 0f00800000;
mul.f32 %f1191, %f225, 0f4B800000;
selp.f32 %f1192, 0fC3170000, 0fC2FE0000, %p125;
selp.f32 %f1193, %f1191, %f225, %p125;
mov.b32 %r368, %f1193;
and.b32 %r369, %r368, 8388607;
or.b32 %r370, %r369, 1065353216;
mov.b32 %f1194, %r370;
shr.u32 %r371, %r368, 23;
cvt.rn.f32.u32 %f1195, %r371;
add.f32 %f1196, %f1192, %f1195;
setp.gt.f32 %p126, %f1194, 0f3FB504F3;
mul.f32 %f1197, %f1194, 0f3F000000;
add.f32 %f1198, %f1196, 0f3F800000;
selp.f32 %f1199, %f1197, %f1194, %p126;
selp.f32 %f1200, %f1198, %f1196, %p126;
add.f32 %f1201, %f1199, 0fBF800000;
add.f32 %f1190, %f1199, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1189,%f1190;
// inline asm
add.f32 %f1202, %f1201, %f1201;
mul.f32 %f1203, %f1189, %f1202;
mul.f32 %f1204, %f1203, %f1203;
fma.rn.f32 %f1207, %f1128, %f1204, %f1127;
fma.rn.f32 %f1209, %f1207, %f1204, %f1130;
mul.rn.f32 %f1210, %f1209, %f1204;
mul.rn.f32 %f1211, %f1210, %f1203;
sub.f32 %f1212, %f1201, %f1203;
neg.f32 %f1213, %f1203;
add.f32 %f1214, %f1212, %f1212;
fma.rn.f32 %f1215, %f1213, %f1201, %f1214;
mul.rn.f32 %f1216, %f1189, %f1215;
add.f32 %f1217, %f1211, %f1203;
sub.f32 %f1218, %f1203, %f1217;
add.f32 %f1219, %f1211, %f1218;
add.f32 %f1220, %f1216, %f1219;
add.f32 %f1221, %f1217, %f1220;
sub.f32 %f1222, %f1217, %f1221;
add.f32 %f1223, %f1220, %f1222;
mul.rn.f32 %f1225, %f1200, %f1146;
mul.rn.f32 %f1227, %f1200, %f1148;
add.f32 %f1228, %f1225, %f1221;
sub.f32 %f1229, %f1225, %f1228;
add.f32 %f1230, %f1221, %f1229;
add.f32 %f1231, %f1223, %f1230;
add.f32 %f1232, %f1227, %f1231;
add.f32 %f1233, %f1228, %f1232;
sub.f32 %f1234, %f1228, %f1233;
add.f32 %f1235, %f1232, %f1234;
mul.rn.f32 %f1237, %f1158, %f1233;
neg.f32 %f1238, %f1237;
fma.rn.f32 %f1239, %f1158, %f1233, %f1238;
fma.rn.f32 %f1240, %f1158, %f1235, %f1239;
fma.rn.f32 %f1242, %f1163, %f1233, %f1240;
add.rn.f32 %f1243, %f1237, %f1242;
neg.f32 %f1244, %f1243;
add.rn.f32 %f1245, %f1237, %f1244;
add.rn.f32 %f1246, %f1245, %f1242;
mov.b32 %r372, %f1243;
setp.eq.s32 %p127, %r372, 1118925336;
add.s32 %r373, %r372, -1;
mov.b32 %f1247, %r373;
add.f32 %f1248, %f1246, 0f37000000;
selp.f32 %f1249, %f1247, %f1243, %p127;
selp.f32 %f226, %f1248, %f1246, %p127;
mul.f32 %f1250, %f1249, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1251, %f1250;
fma.rn.f32 %f1253, %f1251, %f1174, %f1249;
fma.rn.f32 %f1255, %f1251, %f1176, %f1253;
mul.f32 %f1256, %f1255, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1257, %f1256;
add.f32 %f1258, %f1251, 0f00000000;
ex2.approx.f32 %f1259, %f1258;
mul.f32 %f1260, %f1257, %f1259;
setp.lt.f32 %p128, %f1249, 0fC2D20000;
selp.f32 %f1261, 0f00000000, %f1260, %p128;
setp.gt.f32 %p129, %f1249, 0f42D20000;
selp.f32 %f1515, 0f7F800000, %f1261, %p129;
setp.eq.f32 %p130, %f1515, 0f7F800000;
@%p130 bra BB0_119;
fma.rn.f32 %f1515, %f1515, %f226, %f1515;
BB0_119:
setp.lt.f32 %p131, %f193, 0f00000000;
and.pred %p3, %p131, %p116;
mov.b32 %r374, %f1515;
xor.b32 %r375, %r374, -2147483648;
mov.b32 %f1262, %r375;
selp.f32 %f1517, %f1262, %f1515, %p3;
setp.eq.f32 %p133, %f193, 0f00000000;
@%p133 bra BB0_122;
bra.uni BB0_120;
BB0_122:
add.f32 %f1265, %f193, %f193;
selp.f32 %f1517, %f1265, 0f00000000, %p116;
bra.uni BB0_123;
BB0_120:
setp.geu.f32 %p134, %f193, 0f00000000;
@%p134 bra BB0_123;
cvt.rzi.f32.f32 %f1264, %f1158;
setp.neu.f32 %p135, %f1264, 0f3EE66666;
selp.f32 %f1517, 0f7FFFFFFF, %f1517, %p135;
BB0_123:
add.f32 %f1266, %f225, 0f3EE66666;
mov.b32 %r376, %f1266;
setp.lt.s32 %p137, %r376, 2139095040;
@%p137 bra BB0_128;
setp.gtu.f32 %p138, %f225, 0f7F800000;
@%p138 bra BB0_127;
bra.uni BB0_125;
BB0_127:
add.f32 %f1517, %f193, 0f3EE66666;
bra.uni BB0_128;
BB0_125:
setp.neu.f32 %p139, %f225, 0f7F800000;
@%p139 bra BB0_128;
selp.f32 %f1517, 0fFF800000, 0f7F800000, %p3;
BB0_128:
setp.eq.f32 %p140, %f193, 0f3F800000;
selp.f32 %f237, 0f3F800000, %f1517, %p140;
abs.f32 %f238, %f194;
setp.lt.f32 %p141, %f238, 0f00800000;
mul.f32 %f1269, %f238, 0f4B800000;
selp.f32 %f1270, 0fC3170000, 0fC2FE0000, %p141;
selp.f32 %f1271, %f1269, %f238, %p141;
mov.b32 %r377, %f1271;
and.b32 %r378, %r377, 8388607;
or.b32 %r379, %r378, 1065353216;
mov.b32 %f1272, %r379;
shr.u32 %r380, %r377, 23;
cvt.rn.f32.u32 %f1273, %r380;
add.f32 %f1274, %f1270, %f1273;
setp.gt.f32 %p142, %f1272, 0f3FB504F3;
mul.f32 %f1275, %f1272, 0f3F000000;
add.f32 %f1276, %f1274, 0f3F800000;
selp.f32 %f1277, %f1275, %f1272, %p142;
selp.f32 %f1278, %f1276, %f1274, %p142;
add.f32 %f1279, %f1277, 0fBF800000;
add.f32 %f1268, %f1277, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1267,%f1268;
// inline asm
add.f32 %f1280, %f1279, %f1279;
mul.f32 %f1281, %f1267, %f1280;
mul.f32 %f1282, %f1281, %f1281;
fma.rn.f32 %f1285, %f1128, %f1282, %f1127;
fma.rn.f32 %f1287, %f1285, %f1282, %f1130;
mul.rn.f32 %f1288, %f1287, %f1282;
mul.rn.f32 %f1289, %f1288, %f1281;
sub.f32 %f1290, %f1279, %f1281;
neg.f32 %f1291, %f1281;
add.f32 %f1292, %f1290, %f1290;
fma.rn.f32 %f1293, %f1291, %f1279, %f1292;
mul.rn.f32 %f1294, %f1267, %f1293;
add.f32 %f1295, %f1289, %f1281;
sub.f32 %f1296, %f1281, %f1295;
add.f32 %f1297, %f1289, %f1296;
add.f32 %f1298, %f1294, %f1297;
add.f32 %f1299, %f1295, %f1298;
sub.f32 %f1300, %f1295, %f1299;
add.f32 %f1301, %f1298, %f1300;
mul.rn.f32 %f1303, %f1278, %f1146;
mul.rn.f32 %f1305, %f1278, %f1148;
add.f32 %f1306, %f1303, %f1299;
sub.f32 %f1307, %f1303, %f1306;
add.f32 %f1308, %f1299, %f1307;
add.f32 %f1309, %f1301, %f1308;
add.f32 %f1310, %f1305, %f1309;
add.f32 %f1311, %f1306, %f1310;
sub.f32 %f1312, %f1306, %f1311;
add.f32 %f1313, %f1310, %f1312;
mul.rn.f32 %f1315, %f1158, %f1311;
neg.f32 %f1316, %f1315;
fma.rn.f32 %f1317, %f1158, %f1311, %f1316;
fma.rn.f32 %f1318, %f1158, %f1313, %f1317;
fma.rn.f32 %f1320, %f1163, %f1311, %f1318;
add.rn.f32 %f1321, %f1315, %f1320;
neg.f32 %f1322, %f1321;
add.rn.f32 %f1323, %f1315, %f1322;
add.rn.f32 %f1324, %f1323, %f1320;
mov.b32 %r381, %f1321;
setp.eq.s32 %p143, %r381, 1118925336;
add.s32 %r382, %r381, -1;
mov.b32 %f1325, %r382;
add.f32 %f1326, %f1324, 0f37000000;
selp.f32 %f1327, %f1325, %f1321, %p143;
selp.f32 %f239, %f1326, %f1324, %p143;
mul.f32 %f1328, %f1327, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1329, %f1328;
fma.rn.f32 %f1331, %f1329, %f1174, %f1327;
fma.rn.f32 %f1333, %f1329, %f1176, %f1331;
mul.f32 %f1334, %f1333, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1335, %f1334;
add.f32 %f1336, %f1329, 0f00000000;
ex2.approx.f32 %f1337, %f1336;
mul.f32 %f1338, %f1335, %f1337;
setp.lt.f32 %p144, %f1327, 0fC2D20000;
selp.f32 %f1339, 0f00000000, %f1338, %p144;
setp.gt.f32 %p145, %f1327, 0f42D20000;
selp.f32 %f1518, 0f7F800000, %f1339, %p145;
setp.eq.f32 %p146, %f1518, 0f7F800000;
@%p146 bra BB0_130;
fma.rn.f32 %f1518, %f1518, %f239, %f1518;
BB0_130:
setp.lt.f32 %p147, %f194, 0f00000000;
and.pred %p4, %p147, %p116;
mov.b32 %r383, %f1518;
xor.b32 %r384, %r383, -2147483648;
mov.b32 %f1340, %r384;
selp.f32 %f1520, %f1340, %f1518, %p4;
setp.eq.f32 %p149, %f194, 0f00000000;
@%p149 bra BB0_133;
bra.uni BB0_131;
BB0_133:
add.f32 %f1343, %f194, %f194;
selp.f32 %f1520, %f1343, 0f00000000, %p116;
bra.uni BB0_134;
BB0_131:
setp.geu.f32 %p150, %f194, 0f00000000;
@%p150 bra BB0_134;
cvt.rzi.f32.f32 %f1342, %f1158;
setp.neu.f32 %p151, %f1342, 0f3EE66666;
selp.f32 %f1520, 0f7FFFFFFF, %f1520, %p151;
BB0_134:
add.f32 %f1344, %f238, 0f3EE66666;
mov.b32 %r385, %f1344;
setp.lt.s32 %p153, %r385, 2139095040;
@%p153 bra BB0_139;
setp.gtu.f32 %p154, %f238, 0f7F800000;
@%p154 bra BB0_138;
bra.uni BB0_136;
BB0_138:
add.f32 %f1520, %f194, 0f3EE66666;
bra.uni BB0_139;
BB0_136:
setp.neu.f32 %p155, %f238, 0f7F800000;
@%p155 bra BB0_139;
selp.f32 %f1520, 0fFF800000, 0f7F800000, %p4;
BB0_139:
mov.u32 %r470, 4;
setp.eq.f32 %p156, %f194, 0f3F800000;
selp.f32 %f1345, 0f3F800000, %f1520, %p156;
cvt.u64.u32 %rd103, %r3;
cvt.u64.u32 %rd102, %r2;
mov.u64 %rd106, image;
cvta.global.u64 %rd101, %rd106;
// inline asm
call (%rd100), _rt_buffer_get_64, (%rd101, %r107, %r470, %rd102, %rd103, %rd27, %rd27);
// inline asm
cvt.sat.f32.f32 %f1346, %f1345;
mul.f32 %f1347, %f1346, 0f437FFD71;
cvt.rzi.u32.f32 %r388, %f1347;
cvt.sat.f32.f32 %f1348, %f237;
mul.f32 %f1349, %f1348, 0f437FFD71;
cvt.rzi.u32.f32 %r389, %f1349;
cvt.sat.f32.f32 %f1350, %f224;
mul.f32 %f1351, %f1350, 0f437FFD71;
cvt.rzi.u32.f32 %r390, %f1351;
cvt.u16.u32 %rs35, %r388;
cvt.u16.u32 %rs36, %r390;
cvt.u16.u32 %rs37, %r389;
mov.u16 %rs38, 255;
st.v4.u8 [%rd100], {%rs35, %rs37, %rs36, %rs38};
ld.global.u32 %r500, [imageEnabled];
BB0_140:
cvt.u64.u32 %rd18, %r2;
cvt.u64.u32 %rd19, %r3;
and.b32 %r391, %r500, 4;
setp.eq.s32 %p157, %r391, 0;
@%p157 bra BB0_144;
ld.global.u32 %r392, [additive];
setp.eq.s32 %p158, %r392, 0;
mov.f32 %f1352, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs39, %f1352;}
// inline asm
@%p158 bra BB0_143;
mov.u64 %rd119, image_HDR;
cvta.global.u64 %rd108, %rd119;
mov.u32 %r396, 8;
// inline asm
call (%rd107), _rt_buffer_get_64, (%rd108, %r107, %r396, %rd18, %rd19, %rd27, %rd27);
// inline asm
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd107];
// inline asm
{ cvt.f32.f16 %f1353, %rs46;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1354, %rs47;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1355, %rs48;}
// inline asm
// inline asm
call (%rd113), _rt_buffer_get_64, (%rd108, %r107, %r396, %rd18, %rd19, %rd27, %rd27);
// inline asm
add.f32 %f1356, %f192, %f1353;
add.f32 %f1357, %f193, %f1354;
add.f32 %f1358, %f194, %f1355;
// inline asm
{ cvt.rn.f16.f32 %rs45, %f1358;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs44, %f1357;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs43, %f1356;}
// inline asm
st.v4.u16 [%rd113], {%rs43, %rs44, %rs45, %rs39};
bra.uni BB0_144;
BB0_143:
mov.u64 %rd126, image_HDR;
cvta.global.u64 %rd121, %rd126;
mov.u32 %r398, 8;
// inline asm
call (%rd120), _rt_buffer_get_64, (%rd121, %r107, %r398, %rd18, %rd19, %rd27, %rd27);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs52, %f194;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs51, %f193;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs50, %f192;}
// inline asm
st.v4.u16 [%rd120], {%rs50, %rs51, %rs52, %rs39};
BB0_144:
ld.global.u32 %r399, [additive];
setp.eq.s32 %p159, %r399, 0;
mov.f32 %f1362, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs53, %f1362;}
// inline asm
@%p159 bra BB0_146;
mov.u64 %rd139, image_RNM0;
cvta.global.u64 %rd128, %rd139;
mov.u32 %r403, 8;
// inline asm
call (%rd127), _rt_buffer_get_64, (%rd128, %r107, %r403, %rd18, %rd19, %rd27, %rd27);
// inline asm
ld.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd127];
// inline asm
{ cvt.f32.f16 %f1363, %rs60;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1364, %rs61;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1365, %rs62;}
// inline asm
// inline asm
call (%rd133), _rt_buffer_get_64, (%rd128, %r107, %r403, %rd18, %rd19, %rd27, %rd27);
// inline asm
add.f32 %f1366, %f195, %f1363;
add.f32 %f1367, %f196, %f1364;
add.f32 %f1368, %f197, %f1365;
// inline asm
{ cvt.rn.f16.f32 %rs59, %f1368;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs58, %f1367;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs57, %f1366;}
// inline asm
st.v4.u16 [%rd133], {%rs57, %rs58, %rs59, %rs53};
bra.uni BB0_147;
BB0_146:
mov.u64 %rd146, image_RNM0;
cvta.global.u64 %rd141, %rd146;
mov.u32 %r405, 8;
// inline asm
call (%rd140), _rt_buffer_get_64, (%rd141, %r107, %r405, %rd18, %rd19, %rd27, %rd27);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs66, %f197;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs65, %f196;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs64, %f195;}
// inline asm
st.v4.u16 [%rd140], {%rs64, %rs65, %rs66, %rs53};
BB0_147:
ld.global.f32 %f1373, [directDir];
fma.rn.f32 %f250, %f1373, 0fBF000000, 0f3F000000;
ld.global.f32 %f1374, [directDir+4];
fma.rn.f32 %f251, %f1374, 0fBF000000, 0f3F000000;
ld.global.f32 %f1375, [directDir+8];
fma.rn.f32 %f252, %f1375, 0fBF000000, 0f3F000000;
ld.global.u32 %r406, [additive];
setp.eq.s32 %p160, %r406, 0;
// inline asm
{ cvt.rn.f16.f32 %rs67, %f1362;}
// inline asm
@%p160 bra BB0_149;
mov.u64 %rd159, image_RNM1;
cvta.global.u64 %rd148, %rd159;
mov.u32 %r410, 8;
// inline asm
call (%rd147), _rt_buffer_get_64, (%rd148, %r107, %r410, %rd18, %rd19, %rd27, %rd27);
// inline asm
ld.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd147];
// inline asm
{ cvt.f32.f16 %f1376, %rs74;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1377, %rs75;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1378, %rs76;}
// inline asm
// inline asm
call (%rd153), _rt_buffer_get_64, (%rd148, %r107, %r410, %rd18, %rd19, %rd27, %rd27);
// inline asm
add.f32 %f1379, %f250, %f1376;
add.f32 %f1380, %f250, %f1377;
add.f32 %f1381, %f250, %f1378;
// inline asm
{ cvt.rn.f16.f32 %rs73, %f1381;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs72, %f1380;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs71, %f1379;}
// inline asm
st.v4.u16 [%rd153], {%rs71, %rs72, %rs73, %rs67};
bra.uni BB0_150;
BB0_149:
mov.u64 %rd166, image_RNM1;
cvta.global.u64 %rd161, %rd166;
mov.u32 %r412, 8;
// inline asm
call (%rd160), _rt_buffer_get_64, (%rd161, %r107, %r412, %rd18, %rd19, %rd27, %rd27);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs78, %f250;}
// inline asm
st.v4.u16 [%rd160], {%rs78, %rs78, %rs78, %rs67};
BB0_150:
ld.global.u32 %r413, [additive];
setp.eq.s32 %p161, %r413, 0;
// inline asm
{ cvt.rn.f16.f32 %rs79, %f1362;}
// inline asm
@%p161 bra BB0_152;
mov.u64 %rd179, image_RNM2;
cvta.global.u64 %rd168, %rd179;
mov.u32 %r417, 8;
// inline asm
call (%rd167), _rt_buffer_get_64, (%rd168, %r107, %r417, %rd18, %rd19, %rd27, %rd27);
// inline asm
ld.v4.u16 {%rs86, %rs87, %rs88, %rs89}, [%rd167];
// inline asm
{ cvt.f32.f16 %f1384, %rs86;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1385, %rs87;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1386, %rs88;}
// inline asm
// inline asm
call (%rd173), _rt_buffer_get_64, (%rd168, %r107, %r417, %rd18, %rd19, %rd27, %rd27);
// inline asm
add.f32 %f1387, %f251, %f1384;
add.f32 %f1388, %f251, %f1385;
add.f32 %f1389, %f251, %f1386;
// inline asm
{ cvt.rn.f16.f32 %rs85, %f1389;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs84, %f1388;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs83, %f1387;}
// inline asm
st.v4.u16 [%rd173], {%rs83, %rs84, %rs85, %rs79};
bra.uni BB0_153;
BB0_152:
mov.u64 %rd186, image_RNM2;
cvta.global.u64 %rd181, %rd186;
mov.u32 %r419, 8;
// inline asm
call (%rd180), _rt_buffer_get_64, (%rd181, %r107, %r419, %rd18, %rd19, %rd27, %rd27);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs90, %f251;}
// inline asm
st.v4.u16 [%rd180], {%rs90, %rs90, %rs90, %rs79};
BB0_153:
ld.global.u32 %r420, [additive];
setp.eq.s32 %p162, %r420, 0;
// inline asm
{ cvt.rn.f16.f32 %rs91, %f1362;}
// inline asm
@%p162 bra BB0_155;
mov.u64 %rd199, image_RNM3;
cvta.global.u64 %rd188, %rd199;
mov.u32 %r424, 8;
// inline asm
call (%rd187), _rt_buffer_get_64, (%rd188, %r107, %r424, %rd18, %rd19, %rd27, %rd27);
// inline asm
ld.v4.u16 {%rs98, %rs99, %rs100, %rs101}, [%rd187];
// inline asm
{ cvt.f32.f16 %f1392, %rs98;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1393, %rs99;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1394, %rs100;}
// inline asm
// inline asm
call (%rd193), _rt_buffer_get_64, (%rd188, %r107, %r424, %rd18, %rd19, %rd27, %rd27);
// inline asm
add.f32 %f1395, %f252, %f1392;
add.f32 %f1396, %f252, %f1393;
add.f32 %f1397, %f252, %f1394;
// inline asm
{ cvt.rn.f16.f32 %rs97, %f1397;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs96, %f1396;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs95, %f1395;}
// inline asm
st.v4.u16 [%rd193], {%rs95, %rs96, %rs97, %rs91};
bra.uni BB0_176;
BB0_155:
mov.u64 %rd206, image_RNM3;
cvta.global.u64 %rd201, %rd206;
mov.u32 %r426, 8;
// inline asm
call (%rd200), _rt_buffer_get_64, (%rd201, %r107, %r426, %rd18, %rd19, %rd27, %rd27);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs102, %f252;}
// inline asm
st.v4.u16 [%rd200], {%rs102, %rs102, %rs102, %rs91};
BB0_176:
ret;
}
Reference in New Issue View Git Blame Copy Permalink