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

2045 lines
56 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_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 .texref sky;
.global .align 4 .b8 skyColor[12];
.global .align 4 .u32 samples;
.global .align 4 .u32 hemispherical;
.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_typeinfo8skyColorE[8] = {82, 97, 121, 0, 12, 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_typeinfo13hemisphericalE[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_typename8skyColorE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename13hemisphericalE[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_typeenum8skyColorE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum13hemisphericalE = 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_semantic8skyColorE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic13hemisphericalE[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_annotation8skyColorE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation13hemisphericalE[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<106>;
.reg .b16 %rs<154>;
.reg .f32 %f<881>;
.reg .b32 %r<382>;
.reg .b64 %rd<288>;
mov.u64 %rd287, __local_depot0;
cvta.local.u64 %SP, %rd287;
ld.global.u32 %r1, [samples];
shl.b32 %r2, %r1, 1;
ld.global.v2.u32 {%r94, %r95}, [pixelID];
cvt.u64.u32 %rd22, %r94;
cvt.u64.u32 %rd23, %r95;
mov.u64 %rd26, uvnormal;
cvta.global.u64 %rd21, %rd26;
mov.u32 %r92, 2;
mov.u32 %r93, 4;
mov.u64 %rd25, 0;
// inline asm
call (%rd20), _rt_buffer_get_64, (%rd21, %r92, %r93, %rd22, %rd23, %rd25, %rd25);
// inline asm
ld.u32 %r3, [%rd20];
shr.u32 %r98, %r3, 16;
cvt.u16.u32 %rs1, %r98;
and.b16 %rs7, %rs1, 255;
cvt.u16.u32 %rs8, %r3;
or.b16 %rs9, %rs8, %rs7;
setp.eq.s16 %p4, %rs9, 0;
mov.f32 %f797, 0f00000000;
mov.f32 %f798, %f797;
mov.f32 %f799, %f797;
@%p4 bra BB0_2;
ld.u8 %rs10, [%rd20+1];
and.b16 %rs12, %rs8, 255;
cvt.rn.f32.u16 %f209, %rs12;
div.rn.f32 %f210, %f209, 0f437F0000;
fma.rn.f32 %f211, %f210, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f212, %rs10;
div.rn.f32 %f213, %f212, 0f437F0000;
fma.rn.f32 %f214, %f213, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f215, %rs7;
div.rn.f32 %f216, %f215, 0f437F0000;
fma.rn.f32 %f217, %f216, 0f40000000, 0fBF800000;
mul.f32 %f218, %f214, %f214;
fma.rn.f32 %f219, %f211, %f211, %f218;
fma.rn.f32 %f220, %f217, %f217, %f219;
sqrt.rn.f32 %f221, %f220;
rcp.rn.f32 %f222, %f221;
mul.f32 %f797, %f211, %f222;
mul.f32 %f798, %f214, %f222;
mul.f32 %f799, %f217, %f222;
BB0_2:
ld.global.v2.u32 {%r99, %r100}, [pixelID];
ld.global.v2.u32 {%r102, %r103}, [tileInfo];
add.s32 %r4, %r99, %r102;
add.s32 %r5, %r100, %r103;
setp.eq.f32 %p5, %f798, 0f00000000;
setp.eq.f32 %p6, %f797, 0f00000000;
and.pred %p7, %p6, %p5;
setp.eq.f32 %p8, %f799, 0f00000000;
and.pred %p9, %p7, %p8;
@%p9 bra BB0_107;
bra.uni BB0_3;
BB0_107:
ld.global.u32 %r381, [imageEnabled];
and.b32 %r290, %r381, 1;
setp.eq.b32 %p98, %r290, 1;
@!%p98 bra BB0_109;
bra.uni BB0_108;
BB0_108:
cvt.u64.u32 %rd173, %r5;
cvt.u64.u32 %rd172, %r4;
mov.u64 %rd176, image;
cvta.global.u64 %rd171, %rd176;
mov.u64 %rd175, 0;
// inline asm
call (%rd170), _rt_buffer_get_64, (%rd171, %r92, %r93, %rd172, %rd173, %rd175, %rd175);
// inline asm
mov.u16 %rs88, 0;
st.v4.u8 [%rd170], {%rs88, %rs88, %rs88, %rs88};
ld.global.u32 %r381, [imageEnabled];
BB0_109:
cvt.u64.u32 %rd18, %r4;
cvt.u64.u32 %rd19, %r5;
and.b32 %r293, %r381, 4;
setp.eq.s32 %p99, %r293, 0;
@%p99 bra BB0_113;
ld.global.u32 %r294, [additive];
setp.eq.s32 %p100, %r294, 0;
@%p100 bra BB0_112;
mov.u64 %rd189, image_HDR;
cvta.global.u64 %rd178, %rd189;
mov.u32 %r298, 8;
mov.u64 %rd188, 0;
// inline asm
call (%rd177), _rt_buffer_get_64, (%rd178, %r92, %r298, %rd18, %rd19, %rd188, %rd188);
// inline asm
ld.v4.u16 {%rs95, %rs96, %rs97, %rs98}, [%rd177];
// inline asm
{ cvt.f32.f16 %f706, %rs95;}
// inline asm
// inline asm
{ cvt.f32.f16 %f707, %rs96;}
// inline asm
// inline asm
{ cvt.f32.f16 %f708, %rs97;}
// inline asm
// inline asm
call (%rd183), _rt_buffer_get_64, (%rd178, %r92, %r298, %rd18, %rd19, %rd188, %rd188);
// inline asm
add.f32 %f709, %f706, 0f00000000;
add.f32 %f710, %f707, 0f00000000;
add.f32 %f711, %f708, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs94, %f711;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs93, %f710;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs92, %f709;}
// inline asm
mov.u16 %rs99, 0;
st.v4.u16 [%rd183], {%rs92, %rs93, %rs94, %rs99};
bra.uni BB0_113;
BB0_3:
ld.global.v2.u32 {%r111, %r112}, [pixelID];
cvt.u64.u32 %rd29, %r111;
cvt.u64.u32 %rd30, %r112;
mov.u64 %rd39, uvpos;
cvta.global.u64 %rd28, %rd39;
mov.u32 %r108, 12;
// inline asm
call (%rd27), _rt_buffer_get_64, (%rd28, %r92, %r108, %rd29, %rd30, %rd25, %rd25);
// inline asm
ld.f32 %f238, [%rd27+8];
ld.f32 %f239, [%rd27+4];
ld.f32 %f240, [%rd27];
mul.f32 %f241, %f240, 0f3456BF95;
mul.f32 %f242, %f239, 0f3456BF95;
mul.f32 %f243, %f238, 0f3456BF95;
abs.f32 %f244, %f797;
div.rn.f32 %f245, %f241, %f244;
abs.f32 %f246, %f798;
div.rn.f32 %f247, %f242, %f246;
abs.f32 %f248, %f799;
div.rn.f32 %f249, %f243, %f248;
abs.f32 %f250, %f245;
abs.f32 %f251, %f247;
abs.f32 %f252, %f249;
mov.f32 %f253, 0f38D1B717;
max.f32 %f254, %f250, %f253;
max.f32 %f255, %f251, %f253;
max.f32 %f256, %f252, %f253;
fma.rn.f32 %f7, %f797, %f254, %f240;
fma.rn.f32 %f8, %f798, %f255, %f239;
fma.rn.f32 %f9, %f799, %f256, %f238;
ld.global.u32 %r6, [hemispherical];
setp.gt.f32 %p10, %f244, %f248;
neg.f32 %f257, %f798;
selp.f32 %f258, %f257, 0f00000000, %p10;
neg.f32 %f259, %f799;
selp.f32 %f260, %f797, %f259, %p10;
selp.f32 %f261, 0f00000000, %f798, %p10;
mul.f32 %f262, %f260, %f260;
fma.rn.f32 %f263, %f258, %f258, %f262;
fma.rn.f32 %f264, %f261, %f261, %f263;
sqrt.rn.f32 %f265, %f264;
rcp.rn.f32 %f266, %f265;
mul.f32 %f10, %f258, %f266;
mul.f32 %f11, %f260, %f266;
mul.f32 %f12, %f261, %f266;
ld.global.v2.u32 {%r115, %r116}, [pixelID];
cvt.u64.u32 %rd35, %r115;
cvt.u64.u32 %rd36, %r116;
mov.u64 %rd40, rnd_seeds;
cvta.global.u64 %rd34, %rd40;
// inline asm
call (%rd33), _rt_buffer_get_64, (%rd34, %r92, %r93, %rd35, %rd36, %rd25, %rd25);
// inline asm
mov.f32 %f856, 0f00000000;
setp.lt.s32 %p11, %r1, 1;
mov.f32 %f855, %f856;
mov.f32 %f854, %f856;
mov.f32 %f853, %f856;
mov.f32 %f852, %f856;
mov.f32 %f851, %f856;
mov.f32 %f850, %f856;
mov.f32 %f849, %f856;
mov.f32 %f848, %f856;
mov.f32 %f847, %f856;
mov.f32 %f846, %f856;
mov.f32 %f845, %f856;
mov.f32 %f844, %f856;
mov.f32 %f843, %f856;
mov.f32 %f842, %f856;
@%p11 bra BB0_56;
cvt.rn.f32.s32 %f282, %r2;
rcp.rn.f32 %f13, %f282;
ld.u32 %r359, [%rd33];
mul.f32 %f14, %f7, 0f3456BF95;
mul.f32 %f15, %f8, 0f3456BF95;
mul.f32 %f16, %f9, 0f3456BF95;
mul.f32 %f283, %f797, %f11;
mul.f32 %f284, %f798, %f10;
sub.f32 %f17, %f284, %f283;
mul.f32 %f285, %f799, %f10;
mul.f32 %f286, %f797, %f12;
sub.f32 %f18, %f286, %f285;
mul.f32 %f287, %f798, %f12;
mul.f32 %f288, %f799, %f11;
sub.f32 %f19, %f288, %f287;
mov.f32 %f856, 0f00000000;
mov.u32 %r119, 0;
abs.f32 %f363, %f15;
abs.f32 %f364, %f14;
max.f32 %f365, %f364, %f363;
abs.f32 %f366, %f16;
max.f32 %f367, %f365, %f366;
mov.u32 %r356, %r119;
mov.f32 %f855, %f856;
mov.f32 %f854, %f856;
mov.f32 %f853, %f856;
mov.f32 %f852, %f856;
mov.f32 %f851, %f856;
mov.f32 %f850, %f856;
mov.f32 %f849, %f856;
mov.f32 %f848, %f856;
mov.f32 %f847, %f856;
mov.f32 %f846, %f856;
mov.f32 %f845, %f856;
mov.f32 %f844, %f856;
mov.f32 %f843, %f856;
mov.f32 %f842, %f856;
BB0_5:
mov.u32 %r358, %r119;
BB0_6:
mov.u32 %r11, %r359;
cvt.rn.f32.s32 %f747, %r356;
mad.lo.s32 %r121, %r11, 1664525, 1013904223;
and.b32 %r122, %r121, 16777215;
cvt.rn.f32.u32 %f289, %r122;
fma.rn.f32 %f290, %f289, 0f33800000, %f747;
mul.f32 %f291, %f13, %f290;
mad.lo.s32 %r12, %r121, 1664525, 1013904223;
and.b32 %r123, %r12, 16777215;
cvt.rn.f32.u32 %f292, %r123;
cvt.rn.f32.s32 %f293, %r358;
fma.rn.f32 %f294, %f292, 0f33800000, %f293;
mul.f32 %f295, %f13, %f294;
fma.rn.f32 %f51, %f291, 0fC0000000, 0f3F800000;
mul.f32 %f296, %f51, %f51;
mov.f32 %f297, 0f3F800000;
sub.f32 %f298, %f297, %f296;
mov.f32 %f299, 0f00000000;
max.f32 %f300, %f299, %f298;
sqrt.rn.f32 %f52, %f300;
mul.f32 %f836, %f295, 0f40C90FDB;
abs.f32 %f54, %f836;
setp.neu.f32 %p12, %f54, 0f7F800000;
mov.f32 %f830, %f836;
@%p12 bra BB0_8;
mul.rn.f32 %f830, %f836, %f299;
BB0_8:
mul.f32 %f302, %f830, 0f3F22F983;
cvt.rni.s32.f32 %r369, %f302;
cvt.rn.f32.s32 %f303, %r369;
neg.f32 %f304, %f303;
mov.f32 %f305, 0f3FC90FDA;
fma.rn.f32 %f306, %f304, %f305, %f830;
mov.f32 %f307, 0f33A22168;
fma.rn.f32 %f308, %f304, %f307, %f306;
mov.f32 %f309, 0f27C234C5;
fma.rn.f32 %f831, %f304, %f309, %f308;
abs.f32 %f310, %f830;
setp.leu.f32 %p13, %f310, 0f47CE4780;
@%p13 bra BB0_19;
add.u64 %rd42, %SP, 12;
cvta.to.local.u64 %rd283, %rd42;
mov.u32 %r360, 0;
mov.u64 %rd284, 0;
mov.u32 %r361, %r360;
BB0_10:
.pragma "nounroll";
mov.b32 %r331, %f830;
shl.b32 %r330, %r331, 8;
or.b32 %r329, %r330, -2147483648;
add.u64 %rd278, %SP, 12;
cvta.to.local.u64 %rd277, %rd278;
shl.b64 %rd43, %rd284, 2;
mov.u64 %rd44, __cudart_i2opi_f;
add.s64 %rd45, %rd44, %rd43;
ld.const.u32 %r129, [%rd45];
// inline asm
{
mad.lo.cc.u32 %r127, %r129, %r329, %r361;
madc.hi.u32 %r361, %r129, %r329, 0;
}
// inline asm
st.local.u32 [%rd283], %r127;
add.s32 %r360, %r360, 1;
cvt.s64.s32 %rd284, %r360;
mul.wide.s32 %rd48, %r360, 4;
add.s64 %rd283, %rd277, %rd48;
setp.ne.s32 %p14, %r360, 6;
@%p14 bra BB0_10;
mov.b32 %r333, %f830;
shr.u32 %r332, %r333, 23;
add.u64 %rd279, %SP, 12;
and.b32 %r132, %r332, 255;
add.s32 %r133, %r132, -128;
shr.u32 %r134, %r133, 5;
cvta.to.local.u64 %rd50, %rd279;
st.local.u32 [%rd50+24], %r361;
mov.u32 %r135, 6;
sub.s32 %r136, %r135, %r134;
mul.wide.s32 %rd51, %r136, 4;
add.s64 %rd8, %rd50, %rd51;
ld.local.u32 %r362, [%rd8];
ld.local.u32 %r363, [%rd8+-4];
and.b32 %r24, %r332, 31;
setp.eq.s32 %p15, %r24, 0;
@%p15 bra BB0_13;
mov.u32 %r137, 32;
sub.s32 %r138, %r137, %r24;
shr.u32 %r139, %r363, %r138;
shl.b32 %r140, %r362, %r24;
add.s32 %r362, %r139, %r140;
ld.local.u32 %r141, [%rd8+-8];
shr.u32 %r142, %r141, %r138;
shl.b32 %r143, %r363, %r24;
add.s32 %r363, %r142, %r143;
BB0_13:
mov.b32 %r341, %f830;
and.b32 %r365, %r341, -2147483648;
shr.u32 %r144, %r363, 30;
shl.b32 %r145, %r362, 2;
add.s32 %r364, %r144, %r145;
shl.b32 %r30, %r363, 2;
shr.u32 %r146, %r364, 31;
shr.u32 %r147, %r362, 30;
add.s32 %r31, %r146, %r147;
setp.eq.s32 %p16, %r146, 0;
@%p16 bra BB0_14;
bra.uni BB0_15;
BB0_14:
mov.u32 %r366, %r30;
bra.uni BB0_16;
BB0_15:
mov.b32 %r343, %f830;
and.b32 %r342, %r343, -2147483648;
not.b32 %r148, %r364;
neg.s32 %r366, %r30;
setp.eq.s32 %p17, %r30, 0;
selp.u32 %r149, 1, 0, %p17;
add.s32 %r364, %r149, %r148;
xor.b32 %r365, %r342, -2147483648;
BB0_16:
mov.b32 %r345, %f830;
and.b32 %r344, %r345, -2147483648;
clz.b32 %r368, %r364;
setp.eq.s32 %p18, %r368, 0;
shl.b32 %r150, %r364, %r368;
mov.u32 %r151, 32;
sub.s32 %r152, %r151, %r368;
shr.u32 %r153, %r366, %r152;
add.s32 %r154, %r153, %r150;
selp.b32 %r39, %r364, %r154, %p18;
mov.u32 %r155, -921707870;
mul.hi.u32 %r367, %r39, %r155;
setp.eq.s32 %p19, %r344, 0;
neg.s32 %r156, %r31;
selp.b32 %r369, %r31, %r156, %p19;
setp.lt.s32 %p20, %r367, 1;
@%p20 bra BB0_18;
mul.lo.s32 %r157, %r39, -921707870;
shr.u32 %r158, %r157, 31;
shl.b32 %r159, %r367, 1;
add.s32 %r367, %r158, %r159;
add.s32 %r368, %r368, 1;
BB0_18:
mov.u32 %r160, 126;
sub.s32 %r161, %r160, %r368;
shl.b32 %r162, %r161, 23;
add.s32 %r163, %r367, 1;
shr.u32 %r164, %r163, 7;
add.s32 %r165, %r164, 1;
shr.u32 %r166, %r165, 1;
add.s32 %r167, %r166, %r162;
or.b32 %r168, %r167, %r365;
mov.b32 %f831, %r168;
BB0_19:
add.s32 %r47, %r369, 1;
and.b32 %r48, %r47, 1;
setp.eq.s32 %p21, %r48, 0;
@%p21 bra BB0_21;
bra.uni BB0_20;
BB0_21:
mul.rn.f32 %f757, %f831, %f831;
mov.f32 %f313, 0f3C08839E;
mov.f32 %f314, 0fB94CA1F9;
fma.rn.f32 %f832, %f314, %f757, %f313;
bra.uni BB0_22;
BB0_20:
mul.rn.f32 %f753, %f831, %f831;
mov.f32 %f311, 0fBAB6061A;
mov.f32 %f312, 0f37CCF5CE;
fma.rn.f32 %f832, %f312, %f753, %f311;
BB0_22:
@%p21 bra BB0_24;
bra.uni BB0_23;
BB0_24:
mul.rn.f32 %f756, %f831, %f831;
mov.f32 %f752, 0f00000000;
mov.f32 %f318, 0fBE2AAAA3;
fma.rn.f32 %f319, %f832, %f756, %f318;
fma.rn.f32 %f833, %f319, %f756, %f752;
bra.uni BB0_25;
BB0_23:
mul.rn.f32 %f754, %f831, %f831;
mov.f32 %f315, 0f3D2AAAA5;
fma.rn.f32 %f316, %f832, %f754, %f315;
mov.f32 %f317, 0fBF000000;
fma.rn.f32 %f833, %f316, %f754, %f317;
BB0_25:
fma.rn.f32 %f834, %f833, %f831, %f831;
@%p21 bra BB0_27;
mul.rn.f32 %f755, %f831, %f831;
mov.f32 %f741, 0f3F800000;
fma.rn.f32 %f834, %f833, %f755, %f741;
BB0_27:
add.s32 %r346, %r369, 1;
and.b32 %r169, %r346, 2;
setp.eq.s32 %p24, %r169, 0;
@%p24 bra BB0_29;
mov.f32 %f748, 0f00000000;
mov.f32 %f323, 0fBF800000;
fma.rn.f32 %f834, %f834, %f323, %f748;
BB0_29:
abs.f32 %f742, %f836;
setp.neu.f32 %p105, %f742, 0f7F800000;
@%p105 bra BB0_31;
mov.f32 %f751, 0f00000000;
mul.rn.f32 %f836, %f836, %f751;
BB0_31:
mov.f32 %f745, 0f27C234C5;
mov.f32 %f744, 0f33A22168;
mov.f32 %f743, 0f3FC90FDA;
mul.f32 %f325, %f836, 0f3F22F983;
cvt.rni.s32.f32 %r379, %f325;
cvt.rn.f32.s32 %f326, %r379;
neg.f32 %f327, %f326;
fma.rn.f32 %f329, %f327, %f743, %f836;
fma.rn.f32 %f331, %f327, %f744, %f329;
fma.rn.f32 %f837, %f327, %f745, %f331;
abs.f32 %f333, %f836;
setp.leu.f32 %p26, %f333, 0f47CE4780;
@%p26 bra BB0_42;
add.u64 %rd53, %SP, 12;
cvta.to.local.u64 %rd285, %rd53;
mov.b32 %r50, %f836;
shl.b32 %r172, %r50, 8;
or.b32 %r52, %r172, -2147483648;
mov.u32 %r370, 0;
mov.u64 %rd286, %rd25;
mov.u32 %r371, %r370;
BB0_33:
.pragma "nounroll";
add.u64 %rd281, %SP, 12;
cvta.to.local.u64 %rd280, %rd281;
shl.b64 %rd54, %rd286, 2;
mov.u64 %rd55, __cudart_i2opi_f;
add.s64 %rd56, %rd55, %rd54;
ld.const.u32 %r175, [%rd56];
// inline asm
{
mad.lo.cc.u32 %r173, %r175, %r52, %r371;
madc.hi.u32 %r371, %r175, %r52, 0;
}
// inline asm
st.local.u32 [%rd285], %r173;
add.s32 %r370, %r370, 1;
cvt.s64.s32 %rd286, %r370;
mul.wide.s32 %rd57, %r370, 4;
add.s64 %rd285, %rd280, %rd57;
setp.ne.s32 %p27, %r370, 6;
@%p27 bra BB0_33;
mov.b32 %r348, %f836;
shr.u32 %r347, %r348, 23;
add.u64 %rd282, %SP, 12;
and.b32 %r178, %r347, 255;
add.s32 %r179, %r178, -128;
shr.u32 %r180, %r179, 5;
cvta.to.local.u64 %rd59, %rd282;
st.local.u32 [%rd59+24], %r371;
mov.u32 %r181, 6;
sub.s32 %r182, %r181, %r180;
mul.wide.s32 %rd60, %r182, 4;
add.s64 %rd15, %rd59, %rd60;
ld.local.u32 %r372, [%rd15];
ld.local.u32 %r373, [%rd15+-4];
and.b32 %r60, %r347, 31;
setp.eq.s32 %p28, %r60, 0;
@%p28 bra BB0_36;
mov.u32 %r183, 32;
sub.s32 %r184, %r183, %r60;
shr.u32 %r185, %r373, %r184;
shl.b32 %r186, %r372, %r60;
add.s32 %r372, %r185, %r186;
ld.local.u32 %r187, [%rd15+-8];
shr.u32 %r188, %r187, %r184;
shl.b32 %r189, %r373, %r60;
add.s32 %r373, %r188, %r189;
BB0_36:
mov.b32 %r351, %f836;
and.b32 %r375, %r351, -2147483648;
shr.u32 %r190, %r373, 30;
shl.b32 %r191, %r372, 2;
add.s32 %r374, %r190, %r191;
shl.b32 %r66, %r373, 2;
shr.u32 %r192, %r374, 31;
shr.u32 %r193, %r372, 30;
add.s32 %r67, %r192, %r193;
setp.eq.s32 %p29, %r192, 0;
@%p29 bra BB0_37;
bra.uni BB0_38;
BB0_37:
mov.u32 %r376, %r66;
bra.uni BB0_39;
BB0_38:
mov.b32 %r353, %f836;
and.b32 %r352, %r353, -2147483648;
not.b32 %r194, %r374;
neg.s32 %r376, %r66;
setp.eq.s32 %p30, %r66, 0;
selp.u32 %r195, 1, 0, %p30;
add.s32 %r374, %r195, %r194;
xor.b32 %r375, %r352, -2147483648;
BB0_39:
mov.b32 %r355, %f836;
and.b32 %r354, %r355, -2147483648;
clz.b32 %r378, %r374;
setp.eq.s32 %p31, %r378, 0;
shl.b32 %r196, %r374, %r378;
mov.u32 %r197, 32;
sub.s32 %r198, %r197, %r378;
shr.u32 %r199, %r376, %r198;
add.s32 %r200, %r199, %r196;
selp.b32 %r75, %r374, %r200, %p31;
mov.u32 %r201, -921707870;
mul.hi.u32 %r377, %r75, %r201;
setp.eq.s32 %p32, %r354, 0;
neg.s32 %r202, %r67;
selp.b32 %r379, %r67, %r202, %p32;
setp.lt.s32 %p33, %r377, 1;
@%p33 bra BB0_41;
mul.lo.s32 %r203, %r75, -921707870;
shr.u32 %r204, %r203, 31;
shl.b32 %r205, %r377, 1;
add.s32 %r377, %r204, %r205;
add.s32 %r378, %r378, 1;
BB0_41:
mov.u32 %r206, 126;
sub.s32 %r207, %r206, %r378;
shl.b32 %r208, %r207, 23;
add.s32 %r209, %r377, 1;
shr.u32 %r210, %r209, 7;
add.s32 %r211, %r210, 1;
shr.u32 %r212, %r211, 1;
add.s32 %r213, %r212, %r208;
or.b32 %r214, %r213, %r375;
mov.b32 %f837, %r214;
BB0_42:
and.b32 %r83, %r379, 1;
setp.eq.s32 %p34, %r83, 0;
@%p34 bra BB0_44;
bra.uni BB0_43;
BB0_44:
mul.rn.f32 %f762, %f837, %f837;
mov.f32 %f336, 0f3C08839E;
mov.f32 %f337, 0fB94CA1F9;
fma.rn.f32 %f838, %f337, %f762, %f336;
bra.uni BB0_45;
BB0_43:
mul.rn.f32 %f758, %f837, %f837;
mov.f32 %f334, 0fBAB6061A;
mov.f32 %f335, 0f37CCF5CE;
fma.rn.f32 %f838, %f335, %f758, %f334;
BB0_45:
@%p34 bra BB0_47;
bra.uni BB0_46;
BB0_47:
mul.rn.f32 %f761, %f837, %f837;
mov.f32 %f750, 0f00000000;
mov.f32 %f341, 0fBE2AAAA3;
fma.rn.f32 %f342, %f838, %f761, %f341;
fma.rn.f32 %f839, %f342, %f761, %f750;
bra.uni BB0_48;
BB0_46:
mul.rn.f32 %f759, %f837, %f837;
mov.f32 %f338, 0f3D2AAAA5;
fma.rn.f32 %f339, %f838, %f759, %f338;
mov.f32 %f340, 0fBF000000;
fma.rn.f32 %f839, %f339, %f759, %f340;
BB0_48:
fma.rn.f32 %f840, %f839, %f837, %f837;
@%p34 bra BB0_50;
mul.rn.f32 %f760, %f837, %f837;
mov.f32 %f746, 0f3F800000;
fma.rn.f32 %f840, %f839, %f760, %f746;
BB0_50:
and.b32 %r215, %r379, 2;
setp.eq.s32 %p37, %r215, 0;
@%p37 bra BB0_52;
mov.f32 %f749, 0f00000000;
mov.f32 %f346, 0fBF800000;
fma.rn.f32 %f840, %f840, %f346, %f749;
BB0_52:
mul.f32 %f347, %f52, %f834;
mul.f32 %f348, %f52, %f840;
mul.f32 %f349, %f10, %f348;
mul.f32 %f350, %f11, %f348;
mul.f32 %f351, %f12, %f348;
fma.rn.f32 %f352, %f19, %f347, %f349;
fma.rn.f32 %f353, %f18, %f347, %f350;
fma.rn.f32 %f354, %f17, %f347, %f351;
fma.rn.f32 %f89, %f797, %f51, %f352;
fma.rn.f32 %f90, %f798, %f51, %f353;
fma.rn.f32 %f91, %f799, %f51, %f354;
setp.gt.f32 %p38, %f90, 0f00000000;
setp.eq.s32 %p39, %r6, 0;
or.pred %p40, %p39, %p38;
@!%p40 bra BB0_54;
bra.uni BB0_53;
BB0_53:
add.u64 %rd61, %SP, 0;
cvta.to.local.u64 %rd62, %rd61;
max.f32 %f361, %f367, %f253;
mov.u32 %r219, 1065353216;
st.local.u32 [%rd62+8], %r219;
st.local.u32 [%rd62+4], %r219;
st.local.u32 [%rd62], %r219;
ld.global.u32 %r216, [root];
mov.u32 %r217, 1;
mov.f32 %f362, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r216, %f7, %f8, %f9, %f89, %f90, %f91, %r217, %f361, %f362, %rd61, %r108);
// inline asm
mul.f32 %f369, %f798, %f90;
fma.rn.f32 %f370, %f797, %f89, %f369;
fma.rn.f32 %f371, %f799, %f91, %f370;
ld.local.f32 %f372, [%rd62+8];
ld.local.f32 %f373, [%rd62+4];
ld.local.f32 %f374, [%rd62];
fma.rn.f32 %f848, %f89, %f374, %f848;
fma.rn.f32 %f849, %f89, %f373, %f849;
fma.rn.f32 %f850, %f89, %f372, %f850;
fma.rn.f32 %f845, %f90, %f374, %f845;
fma.rn.f32 %f846, %f90, %f373, %f846;
fma.rn.f32 %f847, %f90, %f372, %f847;
fma.rn.f32 %f842, %f91, %f374, %f842;
fma.rn.f32 %f843, %f91, %f373, %f843;
fma.rn.f32 %f844, %f91, %f372, %f844;
add.f32 %f851, %f851, %f374;
add.f32 %f852, %f852, %f373;
add.f32 %f853, %f853, %f372;
cvt.sat.f32.f32 %f375, %f371;
fma.rn.f32 %f854, %f375, %f374, %f854;
fma.rn.f32 %f855, %f375, %f373, %f855;
fma.rn.f32 %f856, %f375, %f372, %f856;
BB0_54:
mad.lo.s32 %r336, %r11, 1664525, 1013904223;
mad.lo.s32 %r359, %r336, 1664525, 1013904223;
add.s32 %r358, %r358, 1;
setp.lt.s32 %p41, %r358, %r2;
@%p41 bra BB0_6;
mad.lo.s32 %r338, %r11, 1664525, 1013904223;
mad.lo.s32 %r359, %r338, 1664525, 1013904223;
add.s32 %r356, %r356, 1;
setp.lt.s32 %p42, %r356, %r2;
@%p42 bra BB0_5;
BB0_56:
mul.lo.s32 %r220, %r2, %r2;
cvt.rn.f32.s32 %f376, %r220;
rcp.rn.f32 %f377, %f376;
mul.f32 %f378, %f854, %f377;
mul.f32 %f379, %f855, %f377;
mul.f32 %f380, %f856, %f377;
mul.f32 %f137, %f851, %f377;
mul.f32 %f138, %f852, %f377;
mul.f32 %f139, %f853, %f377;
mul.f32 %f140, %f848, %f377;
mul.f32 %f141, %f849, %f377;
mul.f32 %f142, %f850, %f377;
mul.f32 %f143, %f845, %f377;
mul.f32 %f144, %f846, %f377;
mul.f32 %f145, %f847, %f377;
mul.f32 %f146, %f842, %f377;
mul.f32 %f147, %f843, %f377;
mul.f32 %f148, %f844, %f377;
fma.rn.f32 %f381, %f854, %f377, %f378;
fma.rn.f32 %f382, %f855, %f377, %f379;
fma.rn.f32 %f383, %f856, %f377, %f380;
ld.global.f32 %f384, [skyColor];
mul.f32 %f149, %f384, %f381;
ld.global.f32 %f385, [skyColor+4];
mul.f32 %f150, %f385, %f382;
ld.global.f32 %f386, [skyColor+8];
mul.f32 %f151, %f383, %f386;
ld.global.u32 %r380, [imageEnabled];
and.b32 %r221, %r380, 1;
setp.eq.b32 %p43, %r221, 1;
@!%p43 bra BB0_91;
bra.uni BB0_57;
BB0_57:
abs.f32 %f153, %f149;
setp.lt.f32 %p44, %f153, 0f00800000;
mul.f32 %f392, %f153, 0f4B800000;
selp.f32 %f393, 0fC3170000, 0fC2FE0000, %p44;
selp.f32 %f394, %f392, %f153, %p44;
mov.b32 %r222, %f394;
and.b32 %r223, %r222, 8388607;
or.b32 %r224, %r223, 1065353216;
mov.b32 %f395, %r224;
shr.u32 %r225, %r222, 23;
cvt.rn.f32.u32 %f396, %r225;
add.f32 %f397, %f393, %f396;
setp.gt.f32 %p45, %f395, 0f3FB504F3;
mul.f32 %f398, %f395, 0f3F000000;
add.f32 %f399, %f397, 0f3F800000;
selp.f32 %f400, %f398, %f395, %p45;
selp.f32 %f401, %f399, %f397, %p45;
add.f32 %f402, %f400, 0fBF800000;
add.f32 %f388, %f400, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f387,%f388;
// inline asm
add.f32 %f403, %f402, %f402;
mul.f32 %f404, %f387, %f403;
mul.f32 %f405, %f404, %f404;
mov.f32 %f406, 0f3C4CAF63;
mov.f32 %f407, 0f3B18F0FE;
fma.rn.f32 %f408, %f407, %f405, %f406;
mov.f32 %f409, 0f3DAAAABD;
fma.rn.f32 %f410, %f408, %f405, %f409;
mul.rn.f32 %f411, %f410, %f405;
mul.rn.f32 %f412, %f411, %f404;
sub.f32 %f413, %f402, %f404;
neg.f32 %f414, %f404;
add.f32 %f415, %f413, %f413;
fma.rn.f32 %f416, %f414, %f402, %f415;
mul.rn.f32 %f417, %f387, %f416;
add.f32 %f418, %f412, %f404;
sub.f32 %f419, %f404, %f418;
add.f32 %f420, %f412, %f419;
add.f32 %f421, %f417, %f420;
add.f32 %f422, %f418, %f421;
sub.f32 %f423, %f418, %f422;
add.f32 %f424, %f421, %f423;
mov.f32 %f425, 0f3F317200;
mul.rn.f32 %f426, %f401, %f425;
mov.f32 %f427, 0f35BFBE8E;
mul.rn.f32 %f428, %f401, %f427;
add.f32 %f429, %f426, %f422;
sub.f32 %f430, %f426, %f429;
add.f32 %f431, %f422, %f430;
add.f32 %f432, %f424, %f431;
add.f32 %f433, %f428, %f432;
add.f32 %f434, %f429, %f433;
sub.f32 %f435, %f429, %f434;
add.f32 %f436, %f433, %f435;
mov.f32 %f437, 0f3EE66666;
mul.rn.f32 %f438, %f437, %f434;
neg.f32 %f439, %f438;
fma.rn.f32 %f440, %f437, %f434, %f439;
fma.rn.f32 %f441, %f437, %f436, %f440;
mov.f32 %f442, 0f00000000;
fma.rn.f32 %f443, %f442, %f434, %f441;
add.rn.f32 %f444, %f438, %f443;
neg.f32 %f445, %f444;
add.rn.f32 %f446, %f438, %f445;
add.rn.f32 %f447, %f446, %f443;
mov.b32 %r226, %f444;
setp.eq.s32 %p46, %r226, 1118925336;
add.s32 %r227, %r226, -1;
mov.b32 %f448, %r227;
add.f32 %f449, %f447, 0f37000000;
selp.f32 %f450, %f448, %f444, %p46;
selp.f32 %f154, %f449, %f447, %p46;
mul.f32 %f451, %f450, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f452, %f451;
mov.f32 %f453, 0fBF317200;
fma.rn.f32 %f454, %f452, %f453, %f450;
mov.f32 %f455, 0fB5BFBE8E;
fma.rn.f32 %f456, %f452, %f455, %f454;
mul.f32 %f457, %f456, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f458, %f457;
add.f32 %f459, %f452, 0f00000000;
ex2.approx.f32 %f460, %f459;
mul.f32 %f461, %f458, %f460;
setp.lt.f32 %p47, %f450, 0fC2D20000;
selp.f32 %f462, 0f00000000, %f461, %p47;
setp.gt.f32 %p48, %f450, 0f42D20000;
selp.f32 %f872, 0f7F800000, %f462, %p48;
setp.eq.f32 %p49, %f872, 0f7F800000;
@%p49 bra BB0_59;
fma.rn.f32 %f872, %f872, %f154, %f872;
BB0_59:
mov.f32 %f766, 0f3E666666;
cvt.rzi.f32.f32 %f765, %f766;
fma.rn.f32 %f764, %f765, 0fC0000000, 0f3EE66666;
abs.f32 %f763, %f764;
setp.lt.f32 %p50, %f149, 0f00000000;
setp.eq.f32 %p51, %f763, 0f3F800000;
and.pred %p1, %p50, %p51;
mov.b32 %r228, %f872;
xor.b32 %r229, %r228, -2147483648;
mov.b32 %f463, %r229;
selp.f32 %f874, %f463, %f872, %p1;
setp.eq.f32 %p52, %f149, 0f00000000;
@%p52 bra BB0_62;
bra.uni BB0_60;
BB0_62:
add.f32 %f466, %f149, %f149;
selp.f32 %f874, %f466, 0f00000000, %p51;
bra.uni BB0_63;
BB0_112:
mov.u64 %rd196, image_HDR;
cvta.global.u64 %rd191, %rd196;
mov.u32 %r300, 8;
mov.u64 %rd195, 0;
// inline asm
call (%rd190), _rt_buffer_get_64, (%rd191, %r92, %r300, %rd18, %rd19, %rd195, %rd195);
// inline asm
mov.f32 %f712, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs100, %f712;}
// inline asm
mov.u16 %rs101, 0;
st.v4.u16 [%rd190], {%rs100, %rs100, %rs100, %rs101};
BB0_113:
ld.global.u32 %r301, [additive];
setp.eq.s32 %p101, %r301, 0;
@%p101 bra BB0_115;
mov.u64 %rd209, image_RNM0;
cvta.global.u64 %rd198, %rd209;
mov.u32 %r305, 8;
mov.u64 %rd208, 0;
// inline asm
call (%rd197), _rt_buffer_get_64, (%rd198, %r92, %r305, %rd18, %rd19, %rd208, %rd208);
// inline asm
ld.v4.u16 {%rs108, %rs109, %rs110, %rs111}, [%rd197];
// inline asm
{ cvt.f32.f16 %f713, %rs108;}
// inline asm
// inline asm
{ cvt.f32.f16 %f714, %rs109;}
// inline asm
// inline asm
{ cvt.f32.f16 %f715, %rs110;}
// inline asm
// inline asm
call (%rd203), _rt_buffer_get_64, (%rd198, %r92, %r305, %rd18, %rd19, %rd208, %rd208);
// inline asm
add.f32 %f716, %f713, 0f00000000;
add.f32 %f717, %f714, 0f00000000;
add.f32 %f718, %f715, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs107, %f718;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs106, %f717;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs105, %f716;}
// inline asm
mov.u16 %rs112, 0;
st.v4.u16 [%rd203], {%rs105, %rs106, %rs107, %rs112};
bra.uni BB0_116;
BB0_115:
mov.u64 %rd216, image_RNM0;
cvta.global.u64 %rd211, %rd216;
mov.u32 %r307, 8;
mov.u64 %rd215, 0;
// inline asm
call (%rd210), _rt_buffer_get_64, (%rd211, %r92, %r307, %rd18, %rd19, %rd215, %rd215);
// inline asm
mov.f32 %f719, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs113, %f719;}
// inline asm
mov.u16 %rs114, 0;
st.v4.u16 [%rd210], {%rs113, %rs113, %rs113, %rs114};
BB0_116:
ld.global.u32 %r308, [additive];
setp.eq.s32 %p102, %r308, 0;
@%p102 bra BB0_118;
mov.u64 %rd229, image_RNM1;
cvta.global.u64 %rd218, %rd229;
mov.u32 %r312, 8;
mov.u64 %rd228, 0;
// inline asm
call (%rd217), _rt_buffer_get_64, (%rd218, %r92, %r312, %rd18, %rd19, %rd228, %rd228);
// inline asm
ld.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd217];
// inline asm
{ cvt.f32.f16 %f720, %rs121;}
// inline asm
// inline asm
{ cvt.f32.f16 %f721, %rs122;}
// inline asm
// inline asm
{ cvt.f32.f16 %f722, %rs123;}
// inline asm
// inline asm
call (%rd223), _rt_buffer_get_64, (%rd218, %r92, %r312, %rd18, %rd19, %rd228, %rd228);
// inline asm
add.f32 %f723, %f720, 0f00000000;
add.f32 %f724, %f721, 0f00000000;
add.f32 %f725, %f722, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs120, %f725;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs119, %f724;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs118, %f723;}
// inline asm
mov.u16 %rs125, 0;
st.v4.u16 [%rd223], {%rs118, %rs119, %rs120, %rs125};
bra.uni BB0_119;
BB0_118:
mov.u64 %rd236, image_RNM1;
cvta.global.u64 %rd231, %rd236;
mov.u32 %r314, 8;
mov.u64 %rd235, 0;
// inline asm
call (%rd230), _rt_buffer_get_64, (%rd231, %r92, %r314, %rd18, %rd19, %rd235, %rd235);
// inline asm
mov.f32 %f726, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs126, %f726;}
// inline asm
mov.u16 %rs127, 0;
st.v4.u16 [%rd230], {%rs126, %rs126, %rs126, %rs127};
BB0_119:
ld.global.u32 %r315, [additive];
setp.eq.s32 %p103, %r315, 0;
@%p103 bra BB0_121;
mov.u64 %rd249, image_RNM2;
cvta.global.u64 %rd238, %rd249;
mov.u32 %r319, 8;
mov.u64 %rd248, 0;
// inline asm
call (%rd237), _rt_buffer_get_64, (%rd238, %r92, %r319, %rd18, %rd19, %rd248, %rd248);
// inline asm
ld.v4.u16 {%rs134, %rs135, %rs136, %rs137}, [%rd237];
// inline asm
{ cvt.f32.f16 %f727, %rs134;}
// inline asm
// inline asm
{ cvt.f32.f16 %f728, %rs135;}
// inline asm
// inline asm
{ cvt.f32.f16 %f729, %rs136;}
// inline asm
// inline asm
call (%rd243), _rt_buffer_get_64, (%rd238, %r92, %r319, %rd18, %rd19, %rd248, %rd248);
// inline asm
add.f32 %f730, %f727, 0f00000000;
add.f32 %f731, %f728, 0f00000000;
add.f32 %f732, %f729, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs133, %f732;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs132, %f731;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs131, %f730;}
// inline asm
mov.u16 %rs138, 0;
st.v4.u16 [%rd243], {%rs131, %rs132, %rs133, %rs138};
bra.uni BB0_122;
BB0_121:
mov.u64 %rd256, image_RNM2;
cvta.global.u64 %rd251, %rd256;
mov.u32 %r321, 8;
mov.u64 %rd255, 0;
// inline asm
call (%rd250), _rt_buffer_get_64, (%rd251, %r92, %r321, %rd18, %rd19, %rd255, %rd255);
// inline asm
mov.f32 %f733, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs139, %f733;}
// inline asm
mov.u16 %rs140, 0;
st.v4.u16 [%rd250], {%rs139, %rs139, %rs139, %rs140};
BB0_122:
ld.global.u32 %r322, [additive];
setp.eq.s32 %p104, %r322, 0;
@%p104 bra BB0_124;
mov.u64 %rd269, image_RNM3;
cvta.global.u64 %rd258, %rd269;
mov.u32 %r326, 8;
mov.u64 %rd268, 0;
// inline asm
call (%rd257), _rt_buffer_get_64, (%rd258, %r92, %r326, %rd18, %rd19, %rd268, %rd268);
// inline asm
ld.v4.u16 {%rs147, %rs148, %rs149, %rs150}, [%rd257];
// inline asm
{ cvt.f32.f16 %f734, %rs147;}
// inline asm
// inline asm
{ cvt.f32.f16 %f735, %rs148;}
// inline asm
// inline asm
{ cvt.f32.f16 %f736, %rs149;}
// inline asm
// inline asm
call (%rd263), _rt_buffer_get_64, (%rd258, %r92, %r326, %rd18, %rd19, %rd268, %rd268);
// inline asm
add.f32 %f737, %f734, 0f00000000;
add.f32 %f738, %f735, 0f00000000;
add.f32 %f739, %f736, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs146, %f739;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs145, %f738;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs144, %f737;}
// inline asm
mov.u16 %rs151, 0;
st.v4.u16 [%rd263], {%rs144, %rs145, %rs146, %rs151};
bra.uni BB0_125;
BB0_124:
mov.u64 %rd276, image_RNM3;
cvta.global.u64 %rd271, %rd276;
mov.u32 %r328, 8;
mov.u64 %rd275, 0;
// inline asm
call (%rd270), _rt_buffer_get_64, (%rd271, %r92, %r328, %rd18, %rd19, %rd275, %rd275);
// inline asm
mov.f32 %f740, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs152, %f740;}
// inline asm
mov.u16 %rs153, 0;
st.v4.u16 [%rd270], {%rs152, %rs152, %rs152, %rs153};
bra.uni BB0_125;
BB0_60:
setp.geu.f32 %p53, %f149, 0f00000000;
@%p53 bra BB0_63;
mov.f32 %f790, 0f3EE66666;
cvt.rzi.f32.f32 %f465, %f790;
setp.neu.f32 %p54, %f465, 0f3EE66666;
selp.f32 %f874, 0f7FFFFFFF, %f874, %p54;
BB0_63:
abs.f32 %f767, %f149;
add.f32 %f467, %f767, 0f3EE66666;
mov.b32 %r230, %f467;
setp.lt.s32 %p56, %r230, 2139095040;
@%p56 bra BB0_68;
abs.f32 %f788, %f149;
setp.gtu.f32 %p57, %f788, 0f7F800000;
@%p57 bra BB0_67;
bra.uni BB0_65;
BB0_67:
add.f32 %f874, %f149, 0f3EE66666;
bra.uni BB0_68;
BB0_65:
abs.f32 %f789, %f149;
setp.neu.f32 %p58, %f789, 0f7F800000;
@%p58 bra BB0_68;
selp.f32 %f874, 0fFF800000, 0f7F800000, %p1;
BB0_68:
mov.f32 %f776, 0fB5BFBE8E;
mov.f32 %f775, 0fBF317200;
mov.f32 %f774, 0f00000000;
mov.f32 %f773, 0f35BFBE8E;
mov.f32 %f772, 0f3F317200;
mov.f32 %f771, 0f3DAAAABD;
mov.f32 %f770, 0f3C4CAF63;
mov.f32 %f769, 0f3B18F0FE;
mov.f32 %f768, 0f3EE66666;
setp.eq.f32 %p59, %f149, 0f3F800000;
selp.f32 %f165, 0f3F800000, %f874, %p59;
abs.f32 %f166, %f150;
setp.lt.f32 %p60, %f166, 0f00800000;
mul.f32 %f470, %f166, 0f4B800000;
selp.f32 %f471, 0fC3170000, 0fC2FE0000, %p60;
selp.f32 %f472, %f470, %f166, %p60;
mov.b32 %r231, %f472;
and.b32 %r232, %r231, 8388607;
or.b32 %r233, %r232, 1065353216;
mov.b32 %f473, %r233;
shr.u32 %r234, %r231, 23;
cvt.rn.f32.u32 %f474, %r234;
add.f32 %f475, %f471, %f474;
setp.gt.f32 %p61, %f473, 0f3FB504F3;
mul.f32 %f476, %f473, 0f3F000000;
add.f32 %f477, %f475, 0f3F800000;
selp.f32 %f478, %f476, %f473, %p61;
selp.f32 %f479, %f477, %f475, %p61;
add.f32 %f480, %f478, 0fBF800000;
add.f32 %f469, %f478, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f468,%f469;
// inline asm
add.f32 %f481, %f480, %f480;
mul.f32 %f482, %f468, %f481;
mul.f32 %f483, %f482, %f482;
fma.rn.f32 %f486, %f769, %f483, %f770;
fma.rn.f32 %f488, %f486, %f483, %f771;
mul.rn.f32 %f489, %f488, %f483;
mul.rn.f32 %f490, %f489, %f482;
sub.f32 %f491, %f480, %f482;
neg.f32 %f492, %f482;
add.f32 %f493, %f491, %f491;
fma.rn.f32 %f494, %f492, %f480, %f493;
mul.rn.f32 %f495, %f468, %f494;
add.f32 %f496, %f490, %f482;
sub.f32 %f497, %f482, %f496;
add.f32 %f498, %f490, %f497;
add.f32 %f499, %f495, %f498;
add.f32 %f500, %f496, %f499;
sub.f32 %f501, %f496, %f500;
add.f32 %f502, %f499, %f501;
mul.rn.f32 %f504, %f479, %f772;
mul.rn.f32 %f506, %f479, %f773;
add.f32 %f507, %f504, %f500;
sub.f32 %f508, %f504, %f507;
add.f32 %f509, %f500, %f508;
add.f32 %f510, %f502, %f509;
add.f32 %f511, %f506, %f510;
add.f32 %f512, %f507, %f511;
sub.f32 %f513, %f507, %f512;
add.f32 %f514, %f511, %f513;
mul.rn.f32 %f516, %f768, %f512;
neg.f32 %f517, %f516;
fma.rn.f32 %f518, %f768, %f512, %f517;
fma.rn.f32 %f519, %f768, %f514, %f518;
fma.rn.f32 %f521, %f774, %f512, %f519;
add.rn.f32 %f522, %f516, %f521;
neg.f32 %f523, %f522;
add.rn.f32 %f524, %f516, %f523;
add.rn.f32 %f525, %f524, %f521;
mov.b32 %r235, %f522;
setp.eq.s32 %p62, %r235, 1118925336;
add.s32 %r236, %r235, -1;
mov.b32 %f526, %r236;
add.f32 %f527, %f525, 0f37000000;
selp.f32 %f528, %f526, %f522, %p62;
selp.f32 %f167, %f527, %f525, %p62;
mul.f32 %f529, %f528, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f530, %f529;
fma.rn.f32 %f532, %f530, %f775, %f528;
fma.rn.f32 %f534, %f530, %f776, %f532;
mul.f32 %f535, %f534, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f536, %f535;
add.f32 %f537, %f530, 0f00000000;
ex2.approx.f32 %f538, %f537;
mul.f32 %f539, %f536, %f538;
setp.lt.f32 %p63, %f528, 0fC2D20000;
selp.f32 %f540, 0f00000000, %f539, %p63;
setp.gt.f32 %p64, %f528, 0f42D20000;
selp.f32 %f875, 0f7F800000, %f540, %p64;
setp.eq.f32 %p65, %f875, 0f7F800000;
@%p65 bra BB0_70;
fma.rn.f32 %f875, %f875, %f167, %f875;
BB0_70:
setp.lt.f32 %p66, %f150, 0f00000000;
and.pred %p2, %p66, %p51;
mov.b32 %r237, %f875;
xor.b32 %r238, %r237, -2147483648;
mov.b32 %f541, %r238;
selp.f32 %f877, %f541, %f875, %p2;
setp.eq.f32 %p68, %f150, 0f00000000;
@%p68 bra BB0_73;
bra.uni BB0_71;
BB0_73:
add.f32 %f544, %f150, %f150;
selp.f32 %f877, %f544, 0f00000000, %p51;
bra.uni BB0_74;
BB0_71:
setp.geu.f32 %p69, %f150, 0f00000000;
@%p69 bra BB0_74;
mov.f32 %f787, 0f3EE66666;
cvt.rzi.f32.f32 %f543, %f787;
setp.neu.f32 %p70, %f543, 0f3EE66666;
selp.f32 %f877, 0f7FFFFFFF, %f877, %p70;
BB0_74:
abs.f32 %f791, %f150;
add.f32 %f545, %f791, 0f3EE66666;
mov.b32 %r239, %f545;
setp.lt.s32 %p72, %r239, 2139095040;
@%p72 bra BB0_79;
abs.f32 %f792, %f150;
setp.gtu.f32 %p73, %f792, 0f7F800000;
@%p73 bra BB0_78;
bra.uni BB0_76;
BB0_78:
add.f32 %f877, %f150, 0f3EE66666;
bra.uni BB0_79;
BB0_76:
abs.f32 %f793, %f150;
setp.neu.f32 %p74, %f793, 0f7F800000;
@%p74 bra BB0_79;
selp.f32 %f877, 0fFF800000, 0f7F800000, %p2;
BB0_79:
mov.f32 %f785, 0fB5BFBE8E;
mov.f32 %f784, 0fBF317200;
mov.f32 %f783, 0f00000000;
mov.f32 %f782, 0f35BFBE8E;
mov.f32 %f781, 0f3F317200;
mov.f32 %f780, 0f3DAAAABD;
mov.f32 %f779, 0f3C4CAF63;
mov.f32 %f778, 0f3B18F0FE;
mov.f32 %f777, 0f3EE66666;
setp.eq.f32 %p75, %f150, 0f3F800000;
selp.f32 %f178, 0f3F800000, %f877, %p75;
abs.f32 %f179, %f151;
setp.lt.f32 %p76, %f179, 0f00800000;
mul.f32 %f548, %f179, 0f4B800000;
selp.f32 %f549, 0fC3170000, 0fC2FE0000, %p76;
selp.f32 %f550, %f548, %f179, %p76;
mov.b32 %r240, %f550;
and.b32 %r241, %r240, 8388607;
or.b32 %r242, %r241, 1065353216;
mov.b32 %f551, %r242;
shr.u32 %r243, %r240, 23;
cvt.rn.f32.u32 %f552, %r243;
add.f32 %f553, %f549, %f552;
setp.gt.f32 %p77, %f551, 0f3FB504F3;
mul.f32 %f554, %f551, 0f3F000000;
add.f32 %f555, %f553, 0f3F800000;
selp.f32 %f556, %f554, %f551, %p77;
selp.f32 %f557, %f555, %f553, %p77;
add.f32 %f558, %f556, 0fBF800000;
add.f32 %f547, %f556, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f546,%f547;
// inline asm
add.f32 %f559, %f558, %f558;
mul.f32 %f560, %f546, %f559;
mul.f32 %f561, %f560, %f560;
fma.rn.f32 %f564, %f778, %f561, %f779;
fma.rn.f32 %f566, %f564, %f561, %f780;
mul.rn.f32 %f567, %f566, %f561;
mul.rn.f32 %f568, %f567, %f560;
sub.f32 %f569, %f558, %f560;
neg.f32 %f570, %f560;
add.f32 %f571, %f569, %f569;
fma.rn.f32 %f572, %f570, %f558, %f571;
mul.rn.f32 %f573, %f546, %f572;
add.f32 %f574, %f568, %f560;
sub.f32 %f575, %f560, %f574;
add.f32 %f576, %f568, %f575;
add.f32 %f577, %f573, %f576;
add.f32 %f578, %f574, %f577;
sub.f32 %f579, %f574, %f578;
add.f32 %f580, %f577, %f579;
mul.rn.f32 %f582, %f557, %f781;
mul.rn.f32 %f584, %f557, %f782;
add.f32 %f585, %f582, %f578;
sub.f32 %f586, %f582, %f585;
add.f32 %f587, %f578, %f586;
add.f32 %f588, %f580, %f587;
add.f32 %f589, %f584, %f588;
add.f32 %f590, %f585, %f589;
sub.f32 %f591, %f585, %f590;
add.f32 %f592, %f589, %f591;
mul.rn.f32 %f594, %f777, %f590;
neg.f32 %f595, %f594;
fma.rn.f32 %f596, %f777, %f590, %f595;
fma.rn.f32 %f597, %f777, %f592, %f596;
fma.rn.f32 %f599, %f783, %f590, %f597;
add.rn.f32 %f600, %f594, %f599;
neg.f32 %f601, %f600;
add.rn.f32 %f602, %f594, %f601;
add.rn.f32 %f603, %f602, %f599;
mov.b32 %r244, %f600;
setp.eq.s32 %p78, %r244, 1118925336;
add.s32 %r245, %r244, -1;
mov.b32 %f604, %r245;
add.f32 %f605, %f603, 0f37000000;
selp.f32 %f606, %f604, %f600, %p78;
selp.f32 %f180, %f605, %f603, %p78;
mul.f32 %f607, %f606, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f608, %f607;
fma.rn.f32 %f610, %f608, %f784, %f606;
fma.rn.f32 %f612, %f608, %f785, %f610;
mul.f32 %f613, %f612, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f614, %f613;
add.f32 %f615, %f608, 0f00000000;
ex2.approx.f32 %f616, %f615;
mul.f32 %f617, %f614, %f616;
setp.lt.f32 %p79, %f606, 0fC2D20000;
selp.f32 %f618, 0f00000000, %f617, %p79;
setp.gt.f32 %p80, %f606, 0f42D20000;
selp.f32 %f878, 0f7F800000, %f618, %p80;
setp.eq.f32 %p81, %f878, 0f7F800000;
@%p81 bra BB0_81;
fma.rn.f32 %f878, %f878, %f180, %f878;
BB0_81:
setp.lt.f32 %p82, %f151, 0f00000000;
and.pred %p3, %p82, %p51;
mov.b32 %r246, %f878;
xor.b32 %r247, %r246, -2147483648;
mov.b32 %f619, %r247;
selp.f32 %f880, %f619, %f878, %p3;
setp.eq.f32 %p84, %f151, 0f00000000;
@%p84 bra BB0_84;
bra.uni BB0_82;
BB0_84:
add.f32 %f622, %f151, %f151;
selp.f32 %f880, %f622, 0f00000000, %p51;
bra.uni BB0_85;
BB0_82:
setp.geu.f32 %p85, %f151, 0f00000000;
@%p85 bra BB0_85;
mov.f32 %f786, 0f3EE66666;
cvt.rzi.f32.f32 %f621, %f786;
setp.neu.f32 %p86, %f621, 0f3EE66666;
selp.f32 %f880, 0f7FFFFFFF, %f880, %p86;
BB0_85:
abs.f32 %f794, %f151;
add.f32 %f623, %f794, 0f3EE66666;
mov.b32 %r248, %f623;
setp.lt.s32 %p88, %r248, 2139095040;
@%p88 bra BB0_90;
abs.f32 %f795, %f151;
setp.gtu.f32 %p89, %f795, 0f7F800000;
@%p89 bra BB0_89;
bra.uni BB0_87;
BB0_89:
add.f32 %f880, %f151, 0f3EE66666;
bra.uni BB0_90;
BB0_87:
abs.f32 %f796, %f151;
setp.neu.f32 %p90, %f796, 0f7F800000;
@%p90 bra BB0_90;
selp.f32 %f880, 0fFF800000, 0f7F800000, %p3;
BB0_90:
mov.u32 %r339, 4;
setp.eq.f32 %p91, %f151, 0f3F800000;
selp.f32 %f624, 0f3F800000, %f880, %p91;
cvt.u64.u32 %rd66, %r5;
cvt.u64.u32 %rd65, %r4;
mov.u64 %rd69, image;
cvta.global.u64 %rd64, %rd69;
// inline asm
call (%rd63), _rt_buffer_get_64, (%rd64, %r92, %r339, %rd65, %rd66, %rd25, %rd25);
// inline asm
cvt.sat.f32.f32 %f625, %f624;
mul.f32 %f626, %f625, 0f437FFD71;
cvt.rzi.u32.f32 %r251, %f626;
cvt.sat.f32.f32 %f627, %f178;
mul.f32 %f628, %f627, 0f437FFD71;
cvt.rzi.u32.f32 %r252, %f628;
cvt.sat.f32.f32 %f629, %f165;
mul.f32 %f630, %f629, 0f437FFD71;
cvt.rzi.u32.f32 %r253, %f630;
cvt.u16.u32 %rs14, %r251;
cvt.u16.u32 %rs15, %r253;
cvt.u16.u32 %rs16, %r252;
mov.u16 %rs17, 255;
st.v4.u8 [%rd63], {%rs14, %rs16, %rs15, %rs17};
ld.global.u32 %r380, [imageEnabled];
BB0_91:
cvt.u64.u32 %rd16, %r4;
cvt.u64.u32 %rd17, %r5;
and.b32 %r254, %r380, 4;
setp.eq.s32 %p92, %r254, 0;
@%p92 bra BB0_95;
ld.global.u32 %r255, [additive];
setp.eq.s32 %p93, %r255, 0;
mov.f32 %f631, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs18, %f631;}
// inline asm
@%p93 bra BB0_94;
mov.u64 %rd82, image_HDR;
cvta.global.u64 %rd71, %rd82;
mov.u32 %r259, 8;
// inline asm
call (%rd70), _rt_buffer_get_64, (%rd71, %r92, %r259, %rd16, %rd17, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs25, %rs26, %rs27, %rs28}, [%rd70];
// inline asm
{ cvt.f32.f16 %f632, %rs25;}
// inline asm
// inline asm
{ cvt.f32.f16 %f633, %rs26;}
// inline asm
// inline asm
{ cvt.f32.f16 %f634, %rs27;}
// inline asm
// inline asm
call (%rd76), _rt_buffer_get_64, (%rd71, %r92, %r259, %rd16, %rd17, %rd25, %rd25);
// inline asm
add.f32 %f635, %f149, %f632;
add.f32 %f636, %f150, %f633;
add.f32 %f637, %f151, %f634;
// inline asm
{ cvt.rn.f16.f32 %rs24, %f637;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs23, %f636;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs22, %f635;}
// inline asm
st.v4.u16 [%rd76], {%rs22, %rs23, %rs24, %rs18};
bra.uni BB0_95;
BB0_94:
mov.u64 %rd89, image_HDR;
cvta.global.u64 %rd84, %rd89;
mov.u32 %r261, 8;
// inline asm
call (%rd83), _rt_buffer_get_64, (%rd84, %r92, %r261, %rd16, %rd17, %rd25, %rd25);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs31, %f151;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs30, %f150;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs29, %f149;}
// inline asm
st.v4.u16 [%rd83], {%rs29, %rs30, %rs31, %rs18};
BB0_95:
mul.f32 %f191, %f137, 0f3F000000;
ld.global.f32 %f642, [skyColor];
mul.f32 %f192, %f191, %f642;
ld.global.f32 %f643, [skyColor+4];
mul.f32 %f193, %f138, 0f3F000000;
mul.f32 %f194, %f193, %f643;
ld.global.f32 %f644, [skyColor+8];
mul.f32 %f195, %f139, 0f3F000000;
mul.f32 %f196, %f195, %f644;
ld.global.u32 %r262, [additive];
setp.eq.s32 %p94, %r262, 0;
mov.f32 %f641, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs32, %f641;}
// inline asm
@%p94 bra BB0_97;
mov.u64 %rd102, image_RNM0;
cvta.global.u64 %rd91, %rd102;
mov.u32 %r266, 8;
// inline asm
call (%rd90), _rt_buffer_get_64, (%rd91, %r92, %r266, %rd16, %rd17, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs39, %rs40, %rs41, %rs42}, [%rd90];
// inline asm
{ cvt.f32.f16 %f645, %rs39;}
// inline asm
// inline asm
{ cvt.f32.f16 %f646, %rs40;}
// inline asm
// inline asm
{ cvt.f32.f16 %f647, %rs41;}
// inline asm
// inline asm
call (%rd96), _rt_buffer_get_64, (%rd91, %r92, %r266, %rd16, %rd17, %rd25, %rd25);
// inline asm
add.f32 %f648, %f192, %f645;
add.f32 %f649, %f194, %f646;
add.f32 %f650, %f196, %f647;
// inline asm
{ cvt.rn.f16.f32 %rs38, %f650;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs37, %f649;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs36, %f648;}
// inline asm
st.v4.u16 [%rd96], {%rs36, %rs37, %rs38, %rs32};
bra.uni BB0_98;
BB0_97:
mov.u64 %rd109, image_RNM0;
cvta.global.u64 %rd104, %rd109;
mov.u32 %r268, 8;
// inline asm
call (%rd103), _rt_buffer_get_64, (%rd104, %r92, %r268, %rd16, %rd17, %rd25, %rd25);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs45, %f196;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs44, %f194;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs43, %f192;}
// inline asm
st.v4.u16 [%rd103], {%rs43, %rs44, %rs45, %rs32};
BB0_98:
mov.f32 %f655, 0f34000000;
max.f32 %f656, %f191, %f655;
mul.f32 %f657, %f140, 0f3F000000;
div.rn.f32 %f658, %f657, %f656;
max.f32 %f659, %f193, %f655;
mul.f32 %f660, %f141, 0f3F000000;
div.rn.f32 %f661, %f660, %f659;
max.f32 %f662, %f195, %f655;
mul.f32 %f663, %f142, 0f3F000000;
div.rn.f32 %f664, %f663, %f662;
fma.rn.f32 %f197, %f658, 0f3F000000, 0f3F000000;
fma.rn.f32 %f198, %f661, 0f3F000000, 0f3F000000;
fma.rn.f32 %f199, %f664, 0f3F000000, 0f3F000000;
mul.f32 %f665, %f143, 0f3F000000;
div.rn.f32 %f666, %f665, %f656;
mul.f32 %f667, %f144, 0f3F000000;
div.rn.f32 %f668, %f667, %f659;
mul.f32 %f669, %f145, 0f3F000000;
div.rn.f32 %f670, %f669, %f662;
fma.rn.f32 %f200, %f666, 0f3F000000, 0f3F000000;
fma.rn.f32 %f201, %f668, 0f3F000000, 0f3F000000;
fma.rn.f32 %f202, %f670, 0f3F000000, 0f3F000000;
mul.f32 %f671, %f146, 0f3F000000;
div.rn.f32 %f672, %f671, %f656;
mul.f32 %f673, %f147, 0f3F000000;
div.rn.f32 %f674, %f673, %f659;
mul.f32 %f675, %f148, 0f3F000000;
div.rn.f32 %f676, %f675, %f662;
fma.rn.f32 %f203, %f672, 0f3F000000, 0f3F000000;
fma.rn.f32 %f204, %f674, 0f3F000000, 0f3F000000;
fma.rn.f32 %f205, %f676, 0f3F000000, 0f3F000000;
ld.global.u32 %r269, [additive];
setp.eq.s32 %p95, %r269, 0;
// inline asm
{ cvt.rn.f16.f32 %rs46, %f641;}
// inline asm
@%p95 bra BB0_100;
mov.u64 %rd122, image_RNM1;
cvta.global.u64 %rd111, %rd122;
mov.u32 %r273, 8;
// inline asm
call (%rd110), _rt_buffer_get_64, (%rd111, %r92, %r273, %rd16, %rd17, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs53, %rs54, %rs55, %rs56}, [%rd110];
// inline asm
{ cvt.f32.f16 %f677, %rs53;}
// inline asm
// inline asm
{ cvt.f32.f16 %f678, %rs54;}
// inline asm
// inline asm
{ cvt.f32.f16 %f679, %rs55;}
// inline asm
// inline asm
call (%rd116), _rt_buffer_get_64, (%rd111, %r92, %r273, %rd16, %rd17, %rd25, %rd25);
// inline asm
add.f32 %f680, %f197, %f677;
add.f32 %f681, %f198, %f678;
add.f32 %f682, %f199, %f679;
// inline asm
{ cvt.rn.f16.f32 %rs52, %f682;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs51, %f681;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs50, %f680;}
// inline asm
st.v4.u16 [%rd116], {%rs50, %rs51, %rs52, %rs46};
bra.uni BB0_101;
BB0_100:
mov.u64 %rd129, image_RNM1;
cvta.global.u64 %rd124, %rd129;
mov.u32 %r275, 8;
// inline asm
call (%rd123), _rt_buffer_get_64, (%rd124, %r92, %r275, %rd16, %rd17, %rd25, %rd25);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs59, %f199;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs58, %f198;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs57, %f197;}
// inline asm
st.v4.u16 [%rd123], {%rs57, %rs58, %rs59, %rs46};
BB0_101:
ld.global.u32 %r276, [additive];
setp.eq.s32 %p96, %r276, 0;
// inline asm
{ cvt.rn.f16.f32 %rs60, %f641;}
// inline asm
@%p96 bra BB0_103;
mov.u64 %rd142, image_RNM2;
cvta.global.u64 %rd131, %rd142;
mov.u32 %r280, 8;
// inline asm
call (%rd130), _rt_buffer_get_64, (%rd131, %r92, %r280, %rd16, %rd17, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs67, %rs68, %rs69, %rs70}, [%rd130];
// inline asm
{ cvt.f32.f16 %f687, %rs67;}
// inline asm
// inline asm
{ cvt.f32.f16 %f688, %rs68;}
// inline asm
// inline asm
{ cvt.f32.f16 %f689, %rs69;}
// inline asm
// inline asm
call (%rd136), _rt_buffer_get_64, (%rd131, %r92, %r280, %rd16, %rd17, %rd25, %rd25);
// inline asm
add.f32 %f690, %f200, %f687;
add.f32 %f691, %f201, %f688;
add.f32 %f692, %f202, %f689;
// inline asm
{ cvt.rn.f16.f32 %rs66, %f692;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs65, %f691;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs64, %f690;}
// inline asm
st.v4.u16 [%rd136], {%rs64, %rs65, %rs66, %rs60};
bra.uni BB0_104;
BB0_103:
mov.u64 %rd149, image_RNM2;
cvta.global.u64 %rd144, %rd149;
mov.u32 %r282, 8;
// inline asm
call (%rd143), _rt_buffer_get_64, (%rd144, %r92, %r282, %rd16, %rd17, %rd25, %rd25);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs73, %f202;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs72, %f201;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs71, %f200;}
// inline asm
st.v4.u16 [%rd143], {%rs71, %rs72, %rs73, %rs60};
BB0_104:
ld.global.u32 %r283, [additive];
setp.eq.s32 %p97, %r283, 0;
// inline asm
{ cvt.rn.f16.f32 %rs74, %f641;}
// inline asm
@%p97 bra BB0_106;
mov.u64 %rd162, image_RNM3;
cvta.global.u64 %rd151, %rd162;
mov.u32 %r287, 8;
// inline asm
call (%rd150), _rt_buffer_get_64, (%rd151, %r92, %r287, %rd16, %rd17, %rd25, %rd25);
// inline asm
ld.v4.u16 {%rs81, %rs82, %rs83, %rs84}, [%rd150];
// inline asm
{ cvt.f32.f16 %f697, %rs81;}
// inline asm
// inline asm
{ cvt.f32.f16 %f698, %rs82;}
// inline asm
// inline asm
{ cvt.f32.f16 %f699, %rs83;}
// inline asm
// inline asm
call (%rd156), _rt_buffer_get_64, (%rd151, %r92, %r287, %rd16, %rd17, %rd25, %rd25);
// inline asm
add.f32 %f700, %f203, %f697;
add.f32 %f701, %f204, %f698;
add.f32 %f702, %f205, %f699;
// inline asm
{ cvt.rn.f16.f32 %rs80, %f702;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs79, %f701;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs78, %f700;}
// inline asm
st.v4.u16 [%rd156], {%rs78, %rs79, %rs80, %rs74};
bra.uni BB0_125;
BB0_106:
mov.u64 %rd169, image_RNM3;
cvta.global.u64 %rd164, %rd169;
mov.u32 %r289, 8;
// inline asm
call (%rd163), _rt_buffer_get_64, (%rd164, %r92, %r289, %rd16, %rd17, %rd25, %rd25);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs87, %f205;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs86, %f204;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs85, %f203;}
// inline asm
st.v4.u16 [%rd163], {%rs85, %rs86, %rs87, %rs74};
BB0_125:
ret;
}