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

2105 lines
58 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 uvtangent[1];
.global .align 8 .b8 texCoords[8];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 1 .b8 rnd_seeds[1];
.global .align 1 .b8 lightmapDirect[1];
.global .texref albedoTex;
.global .align 4 .u32 samples;
.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_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[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_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 0};
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[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_typeenum9texCoordsE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 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 16 .b8 _ZN21rti_internal_semantic9texCoordsE[20] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 101, 120, 67, 111, 111, 114, 100, 115, 0};
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[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_annotation9texCoordsE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[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<117>;
.reg .b16 %rs<144>;
.reg .f32 %f<961>;
.reg .b32 %r<413>;
.reg .b64 %rd<278>;
mov.u64 %rd277, __local_depot0;
cvta.local.u64 %SP, %rd277;
ld.global.u32 %r1, [samples];
ld.global.v2.u32 {%r102, %r103}, [pixelID];
cvt.u64.u32 %rd24, %r102;
cvt.u64.u32 %rd25, %r103;
mov.u64 %rd28, uvnormal;
cvta.global.u64 %rd23, %rd28;
mov.u32 %r100, 2;
mov.u32 %r101, 4;
mov.u64 %rd27, 0;
// inline asm
call (%rd22), _rt_buffer_get_64, (%rd23, %r100, %r101, %rd24, %rd25, %rd27, %rd27);
// inline asm
ld.u32 %r2, [%rd22];
shr.u32 %r106, %r2, 16;
cvt.u16.u32 %rs1, %r106;
and.b16 %rs5, %rs1, 255;
cvt.u16.u32 %rs6, %r2;
or.b16 %rs7, %rs6, %rs5;
setp.eq.s16 %p6, %rs7, 0;
mov.f32 %f892, 0f00000000;
mov.f32 %f893, %f892;
mov.f32 %f894, %f892;
@%p6 bra BB0_2;
ld.u8 %rs8, [%rd22+1];
and.b16 %rs10, %rs6, 255;
cvt.rn.f32.u16 %f213, %rs10;
div.rn.f32 %f214, %f213, 0f437F0000;
fma.rn.f32 %f215, %f214, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f216, %rs8;
div.rn.f32 %f217, %f216, 0f437F0000;
fma.rn.f32 %f218, %f217, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f219, %rs5;
div.rn.f32 %f220, %f219, 0f437F0000;
fma.rn.f32 %f221, %f220, 0f40000000, 0fBF800000;
mul.f32 %f222, %f218, %f218;
fma.rn.f32 %f223, %f215, %f215, %f222;
fma.rn.f32 %f224, %f221, %f221, %f223;
sqrt.rn.f32 %f225, %f224;
rcp.rn.f32 %f226, %f225;
mul.f32 %f892, %f215, %f226;
mul.f32 %f893, %f218, %f226;
mul.f32 %f894, %f221, %f226;
BB0_2:
ld.global.v2.u32 {%r107, %r108}, [pixelID];
ld.global.v2.u32 {%r110, %r111}, [tileInfo];
add.s32 %r3, %r107, %r110;
add.s32 %r4, %r108, %r111;
setp.eq.f32 %p7, %f893, 0f00000000;
setp.eq.f32 %p8, %f892, 0f00000000;
and.pred %p9, %p8, %p7;
setp.eq.f32 %p10, %f894, 0f00000000;
and.pred %p11, %p9, %p10;
@%p11 bra BB0_111;
bra.uni BB0_3;
BB0_111:
ld.global.u32 %r412, [imageEnabled];
and.b32 %r312, %r412, 1;
setp.eq.b32 %p109, %r312, 1;
@!%p109 bra BB0_113;
bra.uni BB0_112;
BB0_112:
cvt.u64.u32 %rd177, %r3;
cvt.u64.u32 %rd178, %r4;
mov.u64 %rd181, image;
cvta.global.u64 %rd176, %rd181;
mov.u64 %rd180, 0;
// inline asm
call (%rd175), _rt_buffer_get_64, (%rd176, %r100, %r101, %rd177, %rd178, %rd180, %rd180);
// inline asm
mov.u16 %rs100, 0;
st.v4.u8 [%rd175], {%rs100, %rs100, %rs100, %rs100};
ld.global.u32 %r412, [imageEnabled];
BB0_113:
and.b32 %r315, %r412, 4;
setp.eq.s32 %p110, %r315, 0;
@%p110 bra BB0_115;
cvt.u64.u32 %rd184, %r3;
cvt.u64.u32 %rd185, %r4;
mov.u64 %rd188, image_HDR;
cvta.global.u64 %rd183, %rd188;
mov.u32 %r317, 8;
mov.u64 %rd187, 0;
// inline asm
call (%rd182), _rt_buffer_get_64, (%rd183, %r100, %r317, %rd184, %rd185, %rd187, %rd187);
// inline asm
mov.f32 %f807, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs101, %f807;}
// inline asm
mov.u16 %rs102, 0;
st.v4.u16 [%rd182], {%rs101, %rs101, %rs101, %rs102};
ld.global.u32 %r412, [imageEnabled];
BB0_115:
and.b32 %r318, %r412, 16;
setp.eq.s32 %p111, %r318, 0;
@%p111 bra BB0_117;
cvt.u64.u32 %rd192, %r4;
cvt.u64.u32 %rd191, %r3;
mov.u64 %rd195, image_HDR2;
cvta.global.u64 %rd190, %rd195;
mov.u32 %r320, 8;
mov.u64 %rd194, 0;
// inline asm
call (%rd189), _rt_buffer_get_64, (%rd190, %r100, %r320, %rd191, %rd192, %rd194, %rd194);
// inline asm
mov.f32 %f808, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs103, %f808;}
// inline asm
mov.u16 %rs104, 0;
st.v4.u16 [%rd189], {%rs103, %rs103, %rs103, %rs104};
BB0_117:
cvt.u64.u32 %rd20, %r3;
cvt.u64.u32 %rd21, %r4;
ld.global.u32 %r321, [additive];
setp.eq.s32 %p112, %r321, 0;
@%p112 bra BB0_119;
mov.u64 %rd208, image_RNM0;
cvta.global.u64 %rd197, %rd208;
mov.u32 %r325, 8;
mov.u64 %rd207, 0;
// inline asm
call (%rd196), _rt_buffer_get_64, (%rd197, %r100, %r325, %rd20, %rd21, %rd207, %rd207);
// inline asm
ld.v4.u16 {%rs111, %rs112, %rs113, %rs114}, [%rd196];
// inline asm
{ cvt.f32.f16 %f809, %rs111;}
// inline asm
// inline asm
{ cvt.f32.f16 %f810, %rs112;}
// inline asm
// inline asm
{ cvt.f32.f16 %f811, %rs113;}
// inline asm
// inline asm
call (%rd202), _rt_buffer_get_64, (%rd197, %r100, %r325, %rd20, %rd21, %rd207, %rd207);
// inline asm
add.f32 %f812, %f809, 0f00000000;
add.f32 %f813, %f810, 0f00000000;
add.f32 %f814, %f811, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs110, %f814;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs109, %f813;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs108, %f812;}
// inline asm
mov.u16 %rs115, 0;
st.v4.u16 [%rd202], {%rs108, %rs109, %rs110, %rs115};
bra.uni BB0_120;
BB0_3:
ld.global.v2.u32 {%r123, %r124}, [pixelID];
cvt.u64.u32 %rd31, %r123;
cvt.u64.u32 %rd32, %r124;
mov.u64 %rd53, lightmapDirect;
cvta.global.u64 %rd30, %rd53;
mov.u32 %r116, 8;
// inline asm
call (%rd29), _rt_buffer_get_64, (%rd30, %r100, %r116, %rd31, %rd32, %rd27, %rd27);
// inline asm
ld.v4.u16 {%rs15, %rs16, %rs17, %rs18}, [%rd29];
// inline asm
{ cvt.f32.f16 %f227, %rs15;}
// inline asm
// inline asm
{ cvt.f32.f16 %f228, %rs16;}
// inline asm
// inline asm
{ cvt.f32.f16 %f229, %rs17;}
// inline asm
ld.global.v2.u32 {%r127, %r128}, [pixelID];
cvt.u64.u32 %rd37, %r127;
cvt.u64.u32 %rd38, %r128;
mov.u64 %rd54, uvpos;
cvta.global.u64 %rd36, %rd54;
mov.u32 %r118, 12;
// inline asm
call (%rd35), _rt_buffer_get_64, (%rd36, %r100, %r118, %rd37, %rd38, %rd27, %rd27);
// inline asm
ld.f32 %f233, [%rd35+8];
ld.f32 %f234, [%rd35+4];
ld.f32 %f235, [%rd35];
mul.f32 %f236, %f235, 0f3456BF95;
mul.f32 %f237, %f234, 0f3456BF95;
mul.f32 %f238, %f233, 0f3456BF95;
abs.f32 %f239, %f892;
div.rn.f32 %f240, %f236, %f239;
abs.f32 %f241, %f893;
div.rn.f32 %f242, %f237, %f241;
abs.f32 %f243, %f894;
div.rn.f32 %f244, %f238, %f243;
abs.f32 %f245, %f240;
abs.f32 %f246, %f242;
abs.f32 %f247, %f244;
mov.f32 %f248, 0f38D1B717;
max.f32 %f249, %f245, %f248;
max.f32 %f250, %f246, %f248;
max.f32 %f251, %f247, %f248;
fma.rn.f32 %f10, %f892, %f249, %f235;
fma.rn.f32 %f11, %f893, %f250, %f234;
fma.rn.f32 %f12, %f894, %f251, %f233;
setp.gt.f32 %p12, %f239, %f243;
neg.f32 %f252, %f893;
selp.f32 %f253, %f252, 0f00000000, %p12;
neg.f32 %f254, %f894;
selp.f32 %f255, %f892, %f254, %p12;
selp.f32 %f256, 0f00000000, %f893, %p12;
mul.f32 %f257, %f255, %f255;
fma.rn.f32 %f258, %f253, %f253, %f257;
fma.rn.f32 %f259, %f256, %f256, %f258;
sqrt.rn.f32 %f260, %f259;
rcp.rn.f32 %f261, %f260;
mul.f32 %f13, %f253, %f261;
mul.f32 %f14, %f255, %f261;
mul.f32 %f15, %f256, %f261;
ld.global.v2.u32 {%r131, %r132}, [pixelID];
cvt.u64.u32 %rd43, %r131;
cvt.u64.u32 %rd44, %r132;
mov.u64 %rd55, rnd_seeds;
cvta.global.u64 %rd42, %rd55;
// inline asm
call (%rd41), _rt_buffer_get_64, (%rd42, %r100, %r101, %rd43, %rd44, %rd27, %rd27);
// inline asm
ld.u32 %r388, [%rd41];
ld.global.v2.u32 {%r135, %r136}, [pixelID];
cvt.u64.u32 %rd49, %r135;
cvt.u64.u32 %rd50, %r136;
mov.u64 %rd56, uvtangent;
cvta.global.u64 %rd48, %rd56;
// inline asm
call (%rd47), _rt_buffer_get_64, (%rd48, %r100, %r101, %rd49, %rd50, %rd27, %rd27);
// inline asm
ld.u32 %r6, [%rd47];
shr.u32 %r7, %r6, 16;
cvt.u16.u32 %rs19, %r7;
and.b16 %rs20, %rs19, 255;
cvt.u16.u32 %rs21, %r6;
or.b16 %rs22, %rs21, %rs20;
setp.eq.s16 %p13, %rs22, 0;
mov.f32 %f898, 0f00000000;
mov.f32 %f895, %f898;
mov.f32 %f896, %f898;
mov.f32 %f897, %f898;
@%p13 bra BB0_5;
ld.u8 %rs23, [%rd47+1];
and.b16 %rs25, %rs21, 255;
cvt.rn.f32.u16 %f262, %rs25;
div.rn.f32 %f263, %f262, 0f437F0000;
fma.rn.f32 %f264, %f263, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f265, %rs23;
div.rn.f32 %f266, %f265, 0f437F0000;
fma.rn.f32 %f267, %f266, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f268, %rs20;
div.rn.f32 %f269, %f268, 0f437F0000;
fma.rn.f32 %f270, %f269, 0f40000000, 0fBF800000;
mul.f32 %f271, %f267, %f267;
fma.rn.f32 %f272, %f264, %f264, %f271;
fma.rn.f32 %f273, %f270, %f270, %f272;
sqrt.rn.f32 %f274, %f273;
rcp.rn.f32 %f275, %f274;
mul.f32 %f895, %f264, %f275;
mul.f32 %f896, %f267, %f275;
mul.f32 %f897, %f270, %f275;
BB0_5:
mul.f32 %f279, %f894, %f896;
mul.f32 %f280, %f893, %f897;
sub.f32 %f281, %f280, %f279;
mul.f32 %f282, %f892, %f897;
mul.f32 %f283, %f894, %f895;
sub.f32 %f284, %f283, %f282;
mul.f32 %f285, %f893, %f895;
mul.f32 %f286, %f892, %f896;
sub.f32 %f287, %f286, %f285;
setp.lt.u32 %p14, %r6, 16777216;
selp.f32 %f288, 0fBF800000, 0f3F800000, %p14;
mul.f32 %f289, %f281, %f288;
mul.f32 %f290, %f284, %f288;
mul.f32 %f291, %f287, %f288;
mul.f32 %f292, %f289, 0f00000000;
mul.f32 %f293, %f290, 0f00000000;
mul.f32 %f294, %f291, 0f00000000;
fma.rn.f32 %f295, %f895, 0f3F5105EC, %f292;
fma.rn.f32 %f296, %f896, 0f3F5105EC, %f293;
fma.rn.f32 %f297, %f897, 0f3F5105EC, %f294;
mul.f32 %f22, %f892, 0f3F13CD3A;
add.f32 %f23, %f22, %f295;
mul.f32 %f24, %f893, 0f3F13CD3A;
add.f32 %f25, %f24, %f296;
mul.f32 %f26, %f894, 0f3F13CD3A;
add.f32 %f27, %f26, %f297;
ld.global.v2.u32 {%r141, %r142}, [pixelID];
cvt.u64.u32 %rd59, %r141;
cvt.u64.u32 %rd60, %r142;
// inline asm
call (%rd57), _rt_buffer_get_64, (%rd48, %r100, %r101, %rd59, %rd60, %rd27, %rd27);
// inline asm
ld.u32 %r8, [%rd57];
shr.u32 %r9, %r8, 16;
cvt.u16.u32 %rs28, %r9;
and.b16 %rs29, %rs28, 255;
cvt.u16.u32 %rs30, %r8;
or.b16 %rs31, %rs30, %rs29;
setp.eq.s16 %p15, %rs31, 0;
mov.f32 %f899, %f898;
mov.f32 %f900, %f898;
@%p15 bra BB0_7;
ld.u8 %rs32, [%rd57+1];
and.b16 %rs34, %rs30, 255;
cvt.rn.f32.u16 %f298, %rs34;
div.rn.f32 %f299, %f298, 0f437F0000;
fma.rn.f32 %f300, %f299, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f301, %rs32;
div.rn.f32 %f302, %f301, 0f437F0000;
fma.rn.f32 %f303, %f302, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f304, %rs29;
div.rn.f32 %f305, %f304, 0f437F0000;
fma.rn.f32 %f306, %f305, 0f40000000, 0fBF800000;
mul.f32 %f307, %f303, %f303;
fma.rn.f32 %f308, %f300, %f300, %f307;
fma.rn.f32 %f309, %f306, %f306, %f308;
sqrt.rn.f32 %f310, %f309;
rcp.rn.f32 %f311, %f310;
mul.f32 %f898, %f300, %f311;
mul.f32 %f899, %f303, %f311;
mul.f32 %f900, %f306, %f311;
BB0_7:
mov.f32 %f901, 0f00000000;
mov.u32 %r384, 4;
mov.u64 %rd272, 0;
mov.u32 %r383, 2;
mul.f32 %f315, %f894, %f899;
mul.f32 %f316, %f893, %f900;
sub.f32 %f317, %f316, %f315;
mul.f32 %f318, %f892, %f900;
mul.f32 %f319, %f894, %f898;
sub.f32 %f320, %f319, %f318;
mul.f32 %f321, %f893, %f898;
mul.f32 %f322, %f892, %f899;
sub.f32 %f323, %f322, %f321;
setp.lt.u32 %p16, %r8, 16777216;
selp.f32 %f324, 0fBF800000, 0f3F800000, %p16;
mul.f32 %f325, %f317, %f324;
mul.f32 %f326, %f320, %f324;
mul.f32 %f327, %f323, %f324;
mul.f32 %f328, %f325, 0f3F3504F3;
mul.f32 %f329, %f326, 0f3F3504F3;
mul.f32 %f330, %f327, 0f3F3504F3;
fma.rn.f32 %f331, %f898, 0fBED105EC, %f328;
fma.rn.f32 %f332, %f899, 0fBED105EC, %f329;
fma.rn.f32 %f333, %f900, 0fBED105EC, %f330;
add.f32 %f34, %f22, %f331;
add.f32 %f35, %f24, %f332;
add.f32 %f36, %f26, %f333;
ld.global.v2.u32 {%r147, %r148}, [pixelID];
cvt.u64.u32 %rd66, %r147;
cvt.u64.u32 %rd67, %r148;
// inline asm
call (%rd64), _rt_buffer_get_64, (%rd48, %r383, %r384, %rd66, %rd67, %rd272, %rd272);
// inline asm
ld.u32 %r10, [%rd64];
shr.u32 %r11, %r10, 16;
cvt.u16.u32 %rs37, %r11;
and.b16 %rs38, %rs37, 255;
cvt.u16.u32 %rs39, %r10;
or.b16 %rs40, %rs39, %rs38;
setp.eq.s16 %p17, %rs40, 0;
mov.f32 %f902, %f901;
mov.f32 %f903, %f901;
@%p17 bra BB0_9;
ld.u8 %rs41, [%rd64+1];
and.b16 %rs43, %rs39, 255;
cvt.rn.f32.u16 %f334, %rs43;
div.rn.f32 %f335, %f334, 0f437F0000;
fma.rn.f32 %f336, %f335, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f337, %rs41;
div.rn.f32 %f338, %f337, 0f437F0000;
fma.rn.f32 %f339, %f338, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f340, %rs38;
div.rn.f32 %f341, %f340, 0f437F0000;
fma.rn.f32 %f342, %f341, 0f40000000, 0fBF800000;
mul.f32 %f343, %f339, %f339;
fma.rn.f32 %f344, %f336, %f336, %f343;
fma.rn.f32 %f345, %f342, %f342, %f344;
sqrt.rn.f32 %f346, %f345;
rcp.rn.f32 %f347, %f346;
mul.f32 %f901, %f336, %f347;
mul.f32 %f902, %f339, %f347;
mul.f32 %f903, %f342, %f347;
BB0_9:
mul.f32 %f887, %f893, 0f3F13CD3A;
mul.f32 %f886, %f892, 0f3F13CD3A;
mov.f32 %f916, 0f00000000;
mul.f32 %f360, %f894, %f902;
mul.f32 %f361, %f893, %f903;
sub.f32 %f362, %f361, %f360;
mul.f32 %f363, %f892, %f903;
mul.f32 %f364, %f894, %f901;
sub.f32 %f365, %f364, %f363;
mul.f32 %f366, %f893, %f901;
mul.f32 %f367, %f892, %f902;
sub.f32 %f368, %f367, %f366;
setp.lt.u32 %p18, %r10, 16777216;
selp.f32 %f369, 0fBF800000, 0f3F800000, %p18;
mul.f32 %f370, %f362, %f369;
mul.f32 %f371, %f365, %f369;
mul.f32 %f372, %f368, %f369;
mul.f32 %f373, %f370, 0fBF3504F3;
mul.f32 %f374, %f371, 0fBF3504F3;
mul.f32 %f375, %f372, 0fBF3504F3;
fma.rn.f32 %f376, %f901, 0fBED105EC, %f373;
fma.rn.f32 %f377, %f902, 0fBED105EC, %f374;
fma.rn.f32 %f378, %f903, 0fBED105EC, %f375;
add.f32 %f43, %f886, %f376;
add.f32 %f44, %f887, %f377;
add.f32 %f45, %f26, %f378;
setp.lt.s32 %p19, %r1, 1;
mov.f32 %f917, %f916;
mov.f32 %f918, %f916;
mov.f32 %f919, %f916;
mov.f32 %f920, %f916;
mov.f32 %f921, %f916;
mov.f32 %f922, %f916;
mov.f32 %f923, %f916;
mov.f32 %f924, %f916;
mov.f32 %f925, %f916;
mov.f32 %f926, %f916;
mov.f32 %f927, %f916;
@%p19 bra BB0_60;
cvt.rn.f32.s32 %f391, %r1;
rcp.rn.f32 %f46, %f391;
mul.f32 %f47, %f10, 0f3456BF95;
mul.f32 %f48, %f11, 0f3456BF95;
mul.f32 %f49, %f12, 0f3456BF95;
mul.f32 %f392, %f894, %f14;
mul.f32 %f393, %f893, %f15;
sub.f32 %f50, %f392, %f393;
mul.f32 %f394, %f892, %f15;
mul.f32 %f395, %f894, %f13;
sub.f32 %f51, %f394, %f395;
mul.f32 %f396, %f893, %f13;
mul.f32 %f397, %f892, %f14;
sub.f32 %f52, %f396, %f397;
mov.f32 %f916, 0f00000000;
mov.u32 %r151, 0;
abs.f32 %f398, %f48;
abs.f32 %f399, %f47;
max.f32 %f400, %f399, %f398;
abs.f32 %f401, %f49;
max.f32 %f402, %f400, %f401;
mov.u32 %r385, %r151;
mov.f32 %f917, %f916;
mov.f32 %f918, %f916;
mov.f32 %f919, %f916;
mov.f32 %f920, %f916;
mov.f32 %f921, %f916;
mov.f32 %f922, %f916;
mov.f32 %f923, %f916;
mov.f32 %f924, %f916;
mov.f32 %f925, %f916;
mov.f32 %f926, %f916;
mov.f32 %f927, %f916;
BB0_11:
mov.u32 %r387, %r151;
BB0_12:
mov.u32 %r15, %r388;
cvt.rn.f32.s32 %f870, %r385;
mad.lo.s32 %r153, %r15, 1664525, 1013904223;
and.b32 %r154, %r153, 16777215;
cvt.rn.f32.u32 %f404, %r154;
fma.rn.f32 %f405, %f404, 0f33800000, %f870;
mul.f32 %f79, %f46, %f405;
mad.lo.s32 %r16, %r153, 1664525, 1013904223;
and.b32 %r155, %r16, 16777215;
cvt.rn.f32.u32 %f406, %r155;
cvt.rn.f32.s32 %f407, %r387;
fma.rn.f32 %f408, %f406, 0f33800000, %f407;
mul.f32 %f409, %f46, %f408;
mul.f32 %f410, %f79, %f79;
mov.f32 %f411, 0f3F800000;
sub.f32 %f412, %f411, %f410;
mov.f32 %f413, 0f00000000;
max.f32 %f414, %f413, %f412;
sqrt.rn.f32 %f80, %f414;
mul.f32 %f934, %f409, 0f40C90FDB;
abs.f32 %f82, %f934;
setp.neu.f32 %p20, %f82, 0f7F800000;
mov.f32 %f928, %f934;
@%p20 bra BB0_14;
mov.f32 %f871, 0f00000000;
mul.rn.f32 %f928, %f934, %f871;
BB0_14:
mul.f32 %f416, %f928, 0f3F22F983;
cvt.rni.s32.f32 %r398, %f416;
cvt.rn.f32.s32 %f417, %r398;
neg.f32 %f418, %f417;
mov.f32 %f419, 0f3FC90FDA;
fma.rn.f32 %f420, %f418, %f419, %f928;
mov.f32 %f421, 0f33A22168;
fma.rn.f32 %f422, %f418, %f421, %f420;
mov.f32 %f423, 0f27C234C5;
fma.rn.f32 %f929, %f418, %f423, %f422;
abs.f32 %f424, %f928;
setp.leu.f32 %p21, %f424, 0f47CE4780;
@%p21 bra BB0_25;
add.u64 %rd72, %SP, 12;
cvta.to.local.u64 %rd273, %rd72;
mov.b32 %r18, %f928;
shl.b32 %r158, %r18, 8;
or.b32 %r20, %r158, -2147483648;
mov.u32 %r389, 0;
mov.u64 %rd274, 0;
mov.u32 %r390, %r389;
BB0_16:
.pragma "nounroll";
add.u64 %rd267, %SP, 12;
cvta.to.local.u64 %rd266, %rd267;
shl.b64 %rd73, %rd274, 2;
mov.u64 %rd74, __cudart_i2opi_f;
add.s64 %rd75, %rd74, %rd73;
ld.const.u32 %r161, [%rd75];
// inline asm
{
mad.lo.cc.u32 %r159, %r161, %r20, %r390;
madc.hi.u32 %r390, %r161, %r20, 0;
}
// inline asm
st.local.u32 [%rd273], %r159;
add.s32 %r389, %r389, 1;
cvt.s64.s32 %rd274, %r389;
mul.wide.s32 %rd78, %r389, 4;
add.s64 %rd273, %rd266, %rd78;
setp.ne.s32 %p22, %r389, 6;
@%p22 bra BB0_16;
mov.b32 %r365, %f928;
shr.u32 %r364, %r365, 23;
add.u64 %rd268, %SP, 12;
and.b32 %r164, %r364, 255;
add.s32 %r165, %r164, -128;
shr.u32 %r166, %r165, 5;
cvta.to.local.u64 %rd80, %rd268;
st.local.u32 [%rd80+24], %r390;
mov.u32 %r167, 6;
sub.s32 %r168, %r167, %r166;
mul.wide.s32 %rd81, %r168, 4;
add.s64 %rd10, %rd80, %rd81;
ld.local.u32 %r391, [%rd10];
ld.local.u32 %r392, [%rd10+-4];
and.b32 %r28, %r364, 31;
setp.eq.s32 %p23, %r28, 0;
@%p23 bra BB0_19;
mov.u32 %r169, 32;
sub.s32 %r170, %r169, %r28;
shr.u32 %r171, %r392, %r170;
shl.b32 %r172, %r391, %r28;
add.s32 %r391, %r171, %r172;
ld.local.u32 %r173, [%rd10+-8];
shr.u32 %r174, %r173, %r170;
shl.b32 %r175, %r392, %r28;
add.s32 %r392, %r174, %r175;
BB0_19:
mov.b32 %r368, %f928;
and.b32 %r394, %r368, -2147483648;
shr.u32 %r176, %r392, 30;
shl.b32 %r177, %r391, 2;
add.s32 %r393, %r176, %r177;
shl.b32 %r34, %r392, 2;
shr.u32 %r178, %r393, 31;
shr.u32 %r179, %r391, 30;
add.s32 %r35, %r178, %r179;
setp.eq.s32 %p24, %r178, 0;
@%p24 bra BB0_20;
bra.uni BB0_21;
BB0_20:
mov.u32 %r395, %r34;
bra.uni BB0_22;
BB0_21:
mov.b32 %r370, %f928;
and.b32 %r369, %r370, -2147483648;
not.b32 %r180, %r393;
neg.s32 %r395, %r34;
setp.eq.s32 %p25, %r34, 0;
selp.u32 %r181, 1, 0, %p25;
add.s32 %r393, %r181, %r180;
xor.b32 %r394, %r369, -2147483648;
BB0_22:
mov.b32 %r372, %f928;
and.b32 %r371, %r372, -2147483648;
clz.b32 %r397, %r393;
setp.eq.s32 %p26, %r397, 0;
shl.b32 %r182, %r393, %r397;
mov.u32 %r183, 32;
sub.s32 %r184, %r183, %r397;
shr.u32 %r185, %r395, %r184;
add.s32 %r186, %r185, %r182;
selp.b32 %r43, %r393, %r186, %p26;
mov.u32 %r187, -921707870;
mul.hi.u32 %r396, %r43, %r187;
setp.eq.s32 %p27, %r371, 0;
neg.s32 %r188, %r35;
selp.b32 %r398, %r35, %r188, %p27;
setp.lt.s32 %p28, %r396, 1;
@%p28 bra BB0_24;
mul.lo.s32 %r189, %r43, -921707870;
shr.u32 %r190, %r189, 31;
shl.b32 %r191, %r396, 1;
add.s32 %r396, %r190, %r191;
add.s32 %r397, %r397, 1;
BB0_24:
mov.u32 %r192, 126;
sub.s32 %r193, %r192, %r397;
shl.b32 %r194, %r193, 23;
add.s32 %r195, %r396, 1;
shr.u32 %r196, %r195, 7;
add.s32 %r197, %r196, 1;
shr.u32 %r198, %r197, 1;
add.s32 %r199, %r198, %r194;
or.b32 %r200, %r199, %r394;
mov.b32 %f929, %r200;
BB0_25:
add.s32 %r51, %r398, 1;
and.b32 %r52, %r51, 1;
setp.eq.s32 %p29, %r52, 0;
@%p29 bra BB0_27;
bra.uni BB0_26;
BB0_27:
mul.rn.f32 %f879, %f929, %f929;
mov.f32 %f427, 0f3C08839E;
mov.f32 %f428, 0fB94CA1F9;
fma.rn.f32 %f930, %f428, %f879, %f427;
bra.uni BB0_28;
BB0_26:
mul.rn.f32 %f875, %f929, %f929;
mov.f32 %f425, 0fBAB6061A;
mov.f32 %f426, 0f37CCF5CE;
fma.rn.f32 %f930, %f426, %f875, %f425;
BB0_28:
@%p29 bra BB0_30;
bra.uni BB0_29;
BB0_30:
mul.rn.f32 %f878, %f929, %f929;
mov.f32 %f874, 0f00000000;
mov.f32 %f432, 0fBE2AAAA3;
fma.rn.f32 %f433, %f930, %f878, %f432;
fma.rn.f32 %f931, %f433, %f878, %f874;
bra.uni BB0_31;
BB0_29:
mul.rn.f32 %f876, %f929, %f929;
mov.f32 %f429, 0f3D2AAAA5;
fma.rn.f32 %f430, %f930, %f876, %f429;
mov.f32 %f431, 0fBF000000;
fma.rn.f32 %f931, %f430, %f876, %f431;
BB0_31:
fma.rn.f32 %f932, %f931, %f929, %f929;
@%p29 bra BB0_33;
mul.rn.f32 %f877, %f929, %f929;
mov.f32 %f861, 0f3F800000;
fma.rn.f32 %f932, %f931, %f877, %f861;
BB0_33:
add.s32 %r373, %r398, 1;
and.b32 %r201, %r373, 2;
setp.eq.s32 %p32, %r201, 0;
@%p32 bra BB0_35;
mov.f32 %f862, 0f00000000;
mov.f32 %f437, 0fBF800000;
fma.rn.f32 %f932, %f932, %f437, %f862;
BB0_35:
abs.f32 %f863, %f934;
setp.neu.f32 %p115, %f863, 0f7F800000;
@%p115 bra BB0_37;
mov.f32 %f873, 0f00000000;
mul.rn.f32 %f934, %f934, %f873;
BB0_37:
mov.f32 %f866, 0f27C234C5;
mov.f32 %f865, 0f33A22168;
mov.f32 %f864, 0f3FC90FDA;
mul.f32 %f439, %f934, 0f3F22F983;
cvt.rni.s32.f32 %r408, %f439;
cvt.rn.f32.s32 %f440, %r408;
neg.f32 %f441, %f440;
fma.rn.f32 %f443, %f441, %f864, %f934;
fma.rn.f32 %f445, %f441, %f865, %f443;
fma.rn.f32 %f935, %f441, %f866, %f445;
abs.f32 %f447, %f934;
setp.leu.f32 %p34, %f447, 0f47CE4780;
@%p34 bra BB0_48;
mov.u64 %rd276, 0;
add.u64 %rd83, %SP, 12;
cvta.to.local.u64 %rd275, %rd83;
mov.b32 %r54, %f934;
shl.b32 %r204, %r54, 8;
or.b32 %r56, %r204, -2147483648;
mov.u32 %r399, 0;
mov.u32 %r400, %r399;
BB0_39:
.pragma "nounroll";
add.u64 %rd270, %SP, 12;
cvta.to.local.u64 %rd269, %rd270;
shl.b64 %rd84, %rd276, 2;
mov.u64 %rd85, __cudart_i2opi_f;
add.s64 %rd86, %rd85, %rd84;
ld.const.u32 %r207, [%rd86];
// inline asm
{
mad.lo.cc.u32 %r205, %r207, %r56, %r400;
madc.hi.u32 %r400, %r207, %r56, 0;
}
// inline asm
st.local.u32 [%rd275], %r205;
add.s32 %r399, %r399, 1;
cvt.s64.s32 %rd276, %r399;
mul.wide.s32 %rd87, %r399, 4;
add.s64 %rd275, %rd269, %rd87;
setp.ne.s32 %p35, %r399, 6;
@%p35 bra BB0_39;
mov.b32 %r375, %f934;
shr.u32 %r374, %r375, 23;
add.u64 %rd271, %SP, 12;
and.b32 %r210, %r374, 255;
add.s32 %r211, %r210, -128;
shr.u32 %r212, %r211, 5;
cvta.to.local.u64 %rd89, %rd271;
st.local.u32 [%rd89+24], %r400;
mov.u32 %r213, 6;
sub.s32 %r214, %r213, %r212;
mul.wide.s32 %rd90, %r214, 4;
add.s64 %rd17, %rd89, %rd90;
ld.local.u32 %r401, [%rd17];
ld.local.u32 %r402, [%rd17+-4];
and.b32 %r64, %r374, 31;
setp.eq.s32 %p36, %r64, 0;
@%p36 bra BB0_42;
mov.u32 %r215, 32;
sub.s32 %r216, %r215, %r64;
shr.u32 %r217, %r402, %r216;
shl.b32 %r218, %r401, %r64;
add.s32 %r401, %r217, %r218;
ld.local.u32 %r219, [%rd17+-8];
shr.u32 %r220, %r219, %r216;
shl.b32 %r221, %r402, %r64;
add.s32 %r402, %r220, %r221;
BB0_42:
mov.b32 %r378, %f934;
and.b32 %r404, %r378, -2147483648;
shr.u32 %r222, %r402, 30;
shl.b32 %r223, %r401, 2;
add.s32 %r403, %r222, %r223;
shl.b32 %r70, %r402, 2;
shr.u32 %r224, %r403, 31;
shr.u32 %r225, %r401, 30;
add.s32 %r71, %r224, %r225;
setp.eq.s32 %p37, %r224, 0;
@%p37 bra BB0_43;
bra.uni BB0_44;
BB0_43:
mov.u32 %r405, %r70;
bra.uni BB0_45;
BB0_44:
mov.b32 %r380, %f934;
and.b32 %r379, %r380, -2147483648;
not.b32 %r226, %r403;
neg.s32 %r405, %r70;
setp.eq.s32 %p38, %r70, 0;
selp.u32 %r227, 1, 0, %p38;
add.s32 %r403, %r227, %r226;
xor.b32 %r404, %r379, -2147483648;
BB0_45:
mov.b32 %r382, %f934;
and.b32 %r381, %r382, -2147483648;
clz.b32 %r407, %r403;
setp.eq.s32 %p39, %r407, 0;
shl.b32 %r228, %r403, %r407;
mov.u32 %r229, 32;
sub.s32 %r230, %r229, %r407;
shr.u32 %r231, %r405, %r230;
add.s32 %r232, %r231, %r228;
selp.b32 %r79, %r403, %r232, %p39;
mov.u32 %r233, -921707870;
mul.hi.u32 %r406, %r79, %r233;
setp.eq.s32 %p40, %r381, 0;
neg.s32 %r234, %r71;
selp.b32 %r408, %r71, %r234, %p40;
setp.lt.s32 %p41, %r406, 1;
@%p41 bra BB0_47;
mul.lo.s32 %r235, %r79, -921707870;
shr.u32 %r236, %r235, 31;
shl.b32 %r237, %r406, 1;
add.s32 %r406, %r236, %r237;
add.s32 %r407, %r407, 1;
BB0_47:
mov.u32 %r238, 126;
sub.s32 %r239, %r238, %r407;
shl.b32 %r240, %r239, 23;
add.s32 %r241, %r406, 1;
shr.u32 %r242, %r241, 7;
add.s32 %r243, %r242, 1;
shr.u32 %r244, %r243, 1;
add.s32 %r245, %r244, %r240;
or.b32 %r246, %r245, %r404;
mov.b32 %f935, %r246;
BB0_48:
and.b32 %r87, %r408, 1;
setp.eq.s32 %p42, %r87, 0;
@%p42 bra BB0_50;
bra.uni BB0_49;
BB0_50:
mul.rn.f32 %f891, %f935, %f935;
mov.f32 %f450, 0f3C08839E;
mov.f32 %f451, 0fB94CA1F9;
fma.rn.f32 %f936, %f451, %f891, %f450;
bra.uni BB0_51;
BB0_49:
mul.rn.f32 %f888, %f935, %f935;
mov.f32 %f448, 0fBAB6061A;
mov.f32 %f449, 0f37CCF5CE;
fma.rn.f32 %f936, %f449, %f888, %f448;
BB0_51:
@%p42 bra BB0_53;
bra.uni BB0_52;
BB0_53:
mul.rn.f32 %f890, %f935, %f935;
mov.f32 %f872, 0f00000000;
mov.f32 %f455, 0fBE2AAAA3;
fma.rn.f32 %f456, %f936, %f890, %f455;
fma.rn.f32 %f937, %f456, %f890, %f872;
bra.uni BB0_54;
BB0_52:
mul.rn.f32 %f889, %f935, %f935;
mov.f32 %f452, 0f3D2AAAA5;
fma.rn.f32 %f453, %f936, %f889, %f452;
mov.f32 %f454, 0fBF000000;
fma.rn.f32 %f937, %f453, %f889, %f454;
BB0_54:
fma.rn.f32 %f938, %f937, %f935, %f935;
@%p42 bra BB0_56;
mul.rn.f32 %f880, %f935, %f935;
mov.f32 %f867, 0f3F800000;
fma.rn.f32 %f938, %f937, %f880, %f867;
BB0_56:
and.b32 %r247, %r408, 2;
setp.eq.s32 %p45, %r247, 0;
@%p45 bra BB0_58;
mov.f32 %f868, 0f00000000;
mov.f32 %f460, 0fBF800000;
fma.rn.f32 %f938, %f938, %f460, %f868;
BB0_58:
mad.lo.s32 %r361, %r15, 1664525, 1013904223;
mad.lo.s32 %r388, %r361, 1664525, 1013904223;
max.f32 %f869, %f402, %f248;
mul.f32 %f469, %f80, %f932;
add.u64 %rd91, %SP, 0;
cvta.to.local.u64 %rd92, %rd91;
mul.f32 %f470, %f80, %f938;
mul.f32 %f471, %f13, %f470;
mul.f32 %f472, %f14, %f470;
mul.f32 %f473, %f15, %f470;
fma.rn.f32 %f474, %f50, %f469, %f471;
fma.rn.f32 %f475, %f51, %f469, %f472;
fma.rn.f32 %f476, %f52, %f469, %f473;
fma.rn.f32 %f464, %f892, %f79, %f474;
fma.rn.f32 %f465, %f893, %f79, %f475;
fma.rn.f32 %f466, %f894, %f79, %f476;
mov.u32 %r249, 0;
st.local.u32 [%rd92+8], %r249;
st.local.u32 [%rd92+4], %r249;
st.local.u32 [%rd92], %r249;
ld.global.u32 %r248, [root];
mov.f32 %f468, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r248, %f10, %f11, %f12, %f464, %f465, %f466, %r249, %f869, %f468, %rd91, %r118);
// inline asm
mul.f32 %f477, %f25, %f465;
fma.rn.f32 %f478, %f23, %f464, %f477;
fma.rn.f32 %f479, %f27, %f466, %f478;
cvt.sat.f32.f32 %f480, %f479;
ld.local.f32 %f481, [%rd92+8];
ld.local.f32 %f482, [%rd92+4];
ld.local.f32 %f483, [%rd92];
fma.rn.f32 %f921, %f483, %f480, %f921;
fma.rn.f32 %f920, %f482, %f480, %f920;
fma.rn.f32 %f919, %f481, %f480, %f919;
mul.f32 %f484, %f35, %f465;
fma.rn.f32 %f485, %f34, %f464, %f484;
fma.rn.f32 %f486, %f36, %f466, %f485;
cvt.sat.f32.f32 %f487, %f486;
fma.rn.f32 %f924, %f483, %f487, %f924;
fma.rn.f32 %f923, %f482, %f487, %f923;
fma.rn.f32 %f922, %f481, %f487, %f922;
mul.f32 %f488, %f44, %f465;
fma.rn.f32 %f489, %f43, %f464, %f488;
fma.rn.f32 %f490, %f45, %f466, %f489;
cvt.sat.f32.f32 %f491, %f490;
fma.rn.f32 %f927, %f483, %f491, %f927;
fma.rn.f32 %f926, %f482, %f491, %f926;
fma.rn.f32 %f925, %f481, %f491, %f925;
mul.f32 %f492, %f893, %f465;
fma.rn.f32 %f493, %f892, %f464, %f492;
fma.rn.f32 %f494, %f894, %f466, %f493;
cvt.sat.f32.f32 %f495, %f494;
fma.rn.f32 %f918, %f483, %f495, %f918;
fma.rn.f32 %f917, %f482, %f495, %f917;
fma.rn.f32 %f916, %f481, %f495, %f916;
add.s32 %r387, %r387, 1;
setp.lt.s32 %p46, %r387, %r1;
@%p46 bra BB0_12;
mad.lo.s32 %r363, %r15, 1664525, 1013904223;
mad.lo.s32 %r388, %r363, 1664525, 1013904223;
add.s32 %r385, %r385, 1;
setp.lt.s32 %p47, %r385, %r1;
@%p47 bra BB0_11;
BB0_60:
mul.lo.s32 %r251, %r1, %r1;
cvt.rn.f32.s32 %f496, %r251;
rcp.rn.f32 %f497, %f496;
cvt.rn.f32.u32 %f498, %r4;
cvt.rn.f32.u32 %f499, %r3;
tex.2d.v4.f32.f32 {%f500, %f501, %f502, %f503}, [albedoTex, {%f499, %f498}];
mul.f32 %f141, %f921, %f497;
mul.f32 %f142, %f920, %f497;
mul.f32 %f143, %f919, %f497;
mul.f32 %f144, %f924, %f497;
mul.f32 %f145, %f923, %f497;
mul.f32 %f146, %f922, %f497;
mul.f32 %f147, %f927, %f497;
mul.f32 %f148, %f926, %f497;
mul.f32 %f149, %f925, %f497;
mul.f32 %f504, %f918, %f497;
mul.f32 %f505, %f917, %f497;
mul.f32 %f506, %f916, %f497;
fma.rn.f32 %f150, %f918, %f497, %f504;
fma.rn.f32 %f151, %f917, %f497, %f505;
fma.rn.f32 %f152, %f916, %f497, %f506;
mul.f32 %f153, %f150, %f500;
mul.f32 %f154, %f151, %f501;
mul.f32 %f155, %f152, %f502;
add.f32 %f156, %f227, %f153;
add.f32 %f157, %f228, %f154;
add.f32 %f158, %f229, %f155;
abs.f32 %f507, %f156;
setp.gtu.f32 %p49, %f507, 0f7F800000;
mov.pred %p116, -1;
@%p49 bra BB0_63;
abs.f32 %f508, %f157;
setp.gtu.f32 %p51, %f508, 0f7F800000;
@%p51 bra BB0_63;
abs.f32 %f509, %f158;
setp.gtu.f32 %p116, %f509, 0f7F800000;
BB0_63:
selp.f32 %f159, 0f00000000, %f156, %p116;
selp.f32 %f160, 0f00000000, %f157, %p116;
selp.f32 %f161, 0f00000000, %f158, %p116;
ld.global.u32 %r410, [imageEnabled];
and.b32 %r252, %r410, 1;
setp.eq.b32 %p52, %r252, 1;
@!%p52 bra BB0_98;
bra.uni BB0_64;
BB0_64:
abs.f32 %f163, %f159;
setp.lt.f32 %p53, %f163, 0f00800000;
mul.f32 %f515, %f163, 0f4B800000;
selp.f32 %f516, 0fC3170000, 0fC2FE0000, %p53;
selp.f32 %f517, %f515, %f163, %p53;
mov.b32 %r253, %f517;
and.b32 %r254, %r253, 8388607;
or.b32 %r255, %r254, 1065353216;
mov.b32 %f518, %r255;
shr.u32 %r256, %r253, 23;
cvt.rn.f32.u32 %f519, %r256;
add.f32 %f520, %f516, %f519;
setp.gt.f32 %p54, %f518, 0f3FB504F3;
mul.f32 %f521, %f518, 0f3F000000;
add.f32 %f522, %f520, 0f3F800000;
selp.f32 %f523, %f521, %f518, %p54;
selp.f32 %f524, %f522, %f520, %p54;
add.f32 %f525, %f523, 0fBF800000;
add.f32 %f511, %f523, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f510,%f511;
// inline asm
add.f32 %f526, %f525, %f525;
mul.f32 %f527, %f510, %f526;
mul.f32 %f528, %f527, %f527;
mov.f32 %f529, 0f3C4CAF63;
mov.f32 %f530, 0f3B18F0FE;
fma.rn.f32 %f531, %f530, %f528, %f529;
mov.f32 %f532, 0f3DAAAABD;
fma.rn.f32 %f533, %f531, %f528, %f532;
mul.rn.f32 %f534, %f533, %f528;
mul.rn.f32 %f535, %f534, %f527;
sub.f32 %f536, %f525, %f527;
neg.f32 %f537, %f527;
add.f32 %f538, %f536, %f536;
fma.rn.f32 %f539, %f537, %f525, %f538;
mul.rn.f32 %f540, %f510, %f539;
add.f32 %f541, %f535, %f527;
sub.f32 %f542, %f527, %f541;
add.f32 %f543, %f535, %f542;
add.f32 %f544, %f540, %f543;
add.f32 %f545, %f541, %f544;
sub.f32 %f546, %f541, %f545;
add.f32 %f547, %f544, %f546;
mov.f32 %f548, 0f3F317200;
mul.rn.f32 %f549, %f524, %f548;
mov.f32 %f550, 0f35BFBE8E;
mul.rn.f32 %f551, %f524, %f550;
add.f32 %f552, %f549, %f545;
sub.f32 %f553, %f549, %f552;
add.f32 %f554, %f545, %f553;
add.f32 %f555, %f547, %f554;
add.f32 %f556, %f551, %f555;
add.f32 %f557, %f552, %f556;
sub.f32 %f558, %f552, %f557;
add.f32 %f559, %f556, %f558;
mov.f32 %f560, 0f3EE66666;
mul.rn.f32 %f561, %f560, %f557;
neg.f32 %f562, %f561;
fma.rn.f32 %f563, %f560, %f557, %f562;
fma.rn.f32 %f564, %f560, %f559, %f563;
mov.f32 %f565, 0f00000000;
fma.rn.f32 %f566, %f565, %f557, %f564;
add.rn.f32 %f567, %f561, %f566;
neg.f32 %f568, %f567;
add.rn.f32 %f569, %f561, %f568;
add.rn.f32 %f570, %f569, %f566;
mov.b32 %r257, %f567;
setp.eq.s32 %p55, %r257, 1118925336;
add.s32 %r258, %r257, -1;
mov.b32 %f571, %r258;
add.f32 %f572, %f570, 0f37000000;
selp.f32 %f573, %f571, %f567, %p55;
selp.f32 %f164, %f572, %f570, %p55;
mul.f32 %f574, %f573, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f575, %f574;
mov.f32 %f576, 0fBF317200;
fma.rn.f32 %f577, %f575, %f576, %f573;
mov.f32 %f578, 0fB5BFBE8E;
fma.rn.f32 %f579, %f575, %f578, %f577;
mul.f32 %f580, %f579, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f581, %f580;
add.f32 %f582, %f575, 0f00000000;
ex2.approx.f32 %f583, %f582;
mul.f32 %f584, %f581, %f583;
setp.lt.f32 %p56, %f573, 0fC2D20000;
selp.f32 %f585, 0f00000000, %f584, %p56;
setp.gt.f32 %p57, %f573, 0f42D20000;
selp.f32 %f952, 0f7F800000, %f585, %p57;
setp.eq.f32 %p58, %f952, 0f7F800000;
@%p58 bra BB0_66;
fma.rn.f32 %f952, %f952, %f164, %f952;
BB0_66:
mov.f32 %f833, 0f3E666666;
cvt.rzi.f32.f32 %f832, %f833;
fma.rn.f32 %f831, %f832, 0fC0000000, 0f3EE66666;
abs.f32 %f830, %f831;
setp.lt.f32 %p59, %f159, 0f00000000;
setp.eq.f32 %p60, %f830, 0f3F800000;
and.pred %p3, %p59, %p60;
mov.b32 %r259, %f952;
xor.b32 %r260, %r259, -2147483648;
mov.b32 %f586, %r260;
selp.f32 %f954, %f586, %f952, %p3;
setp.eq.f32 %p61, %f159, 0f00000000;
@%p61 bra BB0_69;
bra.uni BB0_67;
BB0_69:
add.f32 %f589, %f159, %f159;
selp.f32 %f954, %f589, 0f00000000, %p60;
bra.uni BB0_70;
BB0_119:
mov.u64 %rd215, image_RNM0;
cvta.global.u64 %rd210, %rd215;
mov.u32 %r327, 8;
mov.u64 %rd214, 0;
// inline asm
call (%rd209), _rt_buffer_get_64, (%rd210, %r100, %r327, %rd20, %rd21, %rd214, %rd214);
// inline asm
mov.f32 %f815, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs116, %f815;}
// inline asm
mov.u16 %rs117, 0;
st.v4.u16 [%rd209], {%rs116, %rs116, %rs116, %rs117};
BB0_120:
ld.global.u32 %r328, [additive];
setp.eq.s32 %p113, %r328, 0;
@%p113 bra BB0_122;
mov.u64 %rd228, image_RNM1;
cvta.global.u64 %rd217, %rd228;
mov.u32 %r332, 8;
mov.u64 %rd227, 0;
// inline asm
call (%rd216), _rt_buffer_get_64, (%rd217, %r100, %r332, %rd20, %rd21, %rd227, %rd227);
// inline asm
ld.v4.u16 {%rs124, %rs125, %rs126, %rs127}, [%rd216];
// inline asm
{ cvt.f32.f16 %f816, %rs124;}
// inline asm
// inline asm
{ cvt.f32.f16 %f817, %rs125;}
// inline asm
// inline asm
{ cvt.f32.f16 %f818, %rs126;}
// inline asm
// inline asm
call (%rd222), _rt_buffer_get_64, (%rd217, %r100, %r332, %rd20, %rd21, %rd227, %rd227);
// inline asm
add.f32 %f819, %f816, 0f00000000;
add.f32 %f820, %f817, 0f00000000;
add.f32 %f821, %f818, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs123, %f821;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs122, %f820;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs121, %f819;}
// inline asm
mov.u16 %rs128, 0;
st.v4.u16 [%rd222], {%rs121, %rs122, %rs123, %rs128};
bra.uni BB0_123;
BB0_122:
mov.u64 %rd235, image_RNM1;
cvta.global.u64 %rd230, %rd235;
mov.u32 %r334, 8;
mov.u64 %rd234, 0;
// inline asm
call (%rd229), _rt_buffer_get_64, (%rd230, %r100, %r334, %rd20, %rd21, %rd234, %rd234);
// inline asm
mov.f32 %f822, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs129, %f822;}
// inline asm
mov.u16 %rs130, 0;
st.v4.u16 [%rd229], {%rs129, %rs129, %rs129, %rs130};
BB0_123:
ld.global.u32 %r335, [additive];
setp.eq.s32 %p114, %r335, 0;
@%p114 bra BB0_125;
mov.u64 %rd248, image_RNM2;
cvta.global.u64 %rd237, %rd248;
mov.u32 %r339, 8;
mov.u64 %rd247, 0;
// inline asm
call (%rd236), _rt_buffer_get_64, (%rd237, %r100, %r339, %rd20, %rd21, %rd247, %rd247);
// inline asm
ld.v4.u16 {%rs137, %rs138, %rs139, %rs140}, [%rd236];
// inline asm
{ cvt.f32.f16 %f823, %rs137;}
// inline asm
// inline asm
{ cvt.f32.f16 %f824, %rs138;}
// inline asm
// inline asm
{ cvt.f32.f16 %f825, %rs139;}
// inline asm
// inline asm
call (%rd242), _rt_buffer_get_64, (%rd237, %r100, %r339, %rd20, %rd21, %rd247, %rd247);
// inline asm
add.f32 %f826, %f823, 0f00000000;
add.f32 %f827, %f824, 0f00000000;
add.f32 %f828, %f825, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs136, %f828;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs135, %f827;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs134, %f826;}
// inline asm
mov.u16 %rs141, 0;
st.v4.u16 [%rd242], {%rs134, %rs135, %rs136, %rs141};
bra.uni BB0_126;
BB0_125:
mov.u64 %rd255, image_RNM2;
cvta.global.u64 %rd250, %rd255;
mov.u32 %r341, 8;
mov.u64 %rd254, 0;
// inline asm
call (%rd249), _rt_buffer_get_64, (%rd250, %r100, %r341, %rd20, %rd21, %rd254, %rd254);
// inline asm
mov.f32 %f829, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs142, %f829;}
// inline asm
mov.u16 %rs143, 0;
st.v4.u16 [%rd249], {%rs142, %rs142, %rs142, %rs143};
bra.uni BB0_126;
BB0_67:
setp.geu.f32 %p62, %f159, 0f00000000;
@%p62 bra BB0_70;
mov.f32 %f857, 0f3EE66666;
cvt.rzi.f32.f32 %f588, %f857;
setp.neu.f32 %p63, %f588, 0f3EE66666;
selp.f32 %f954, 0f7FFFFFFF, %f954, %p63;
BB0_70:
abs.f32 %f834, %f159;
add.f32 %f590, %f834, 0f3EE66666;
mov.b32 %r261, %f590;
setp.lt.s32 %p65, %r261, 2139095040;
@%p65 bra BB0_75;
abs.f32 %f855, %f159;
setp.gtu.f32 %p66, %f855, 0f7F800000;
@%p66 bra BB0_74;
bra.uni BB0_72;
BB0_74:
add.f32 %f954, %f159, 0f3EE66666;
bra.uni BB0_75;
BB0_72:
abs.f32 %f856, %f159;
setp.neu.f32 %p67, %f856, 0f7F800000;
@%p67 bra BB0_75;
selp.f32 %f954, 0fFF800000, 0f7F800000, %p3;
BB0_75:
mov.f32 %f843, 0fB5BFBE8E;
mov.f32 %f842, 0fBF317200;
mov.f32 %f841, 0f00000000;
mov.f32 %f840, 0f35BFBE8E;
mov.f32 %f839, 0f3F317200;
mov.f32 %f838, 0f3DAAAABD;
mov.f32 %f837, 0f3C4CAF63;
mov.f32 %f836, 0f3B18F0FE;
mov.f32 %f835, 0f3EE66666;
setp.eq.f32 %p68, %f159, 0f3F800000;
selp.f32 %f175, 0f3F800000, %f954, %p68;
abs.f32 %f176, %f160;
setp.lt.f32 %p69, %f176, 0f00800000;
mul.f32 %f593, %f176, 0f4B800000;
selp.f32 %f594, 0fC3170000, 0fC2FE0000, %p69;
selp.f32 %f595, %f593, %f176, %p69;
mov.b32 %r262, %f595;
and.b32 %r263, %r262, 8388607;
or.b32 %r264, %r263, 1065353216;
mov.b32 %f596, %r264;
shr.u32 %r265, %r262, 23;
cvt.rn.f32.u32 %f597, %r265;
add.f32 %f598, %f594, %f597;
setp.gt.f32 %p70, %f596, 0f3FB504F3;
mul.f32 %f599, %f596, 0f3F000000;
add.f32 %f600, %f598, 0f3F800000;
selp.f32 %f601, %f599, %f596, %p70;
selp.f32 %f602, %f600, %f598, %p70;
add.f32 %f603, %f601, 0fBF800000;
add.f32 %f592, %f601, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f591,%f592;
// inline asm
add.f32 %f604, %f603, %f603;
mul.f32 %f605, %f591, %f604;
mul.f32 %f606, %f605, %f605;
fma.rn.f32 %f609, %f836, %f606, %f837;
fma.rn.f32 %f611, %f609, %f606, %f838;
mul.rn.f32 %f612, %f611, %f606;
mul.rn.f32 %f613, %f612, %f605;
sub.f32 %f614, %f603, %f605;
neg.f32 %f615, %f605;
add.f32 %f616, %f614, %f614;
fma.rn.f32 %f617, %f615, %f603, %f616;
mul.rn.f32 %f618, %f591, %f617;
add.f32 %f619, %f613, %f605;
sub.f32 %f620, %f605, %f619;
add.f32 %f621, %f613, %f620;
add.f32 %f622, %f618, %f621;
add.f32 %f623, %f619, %f622;
sub.f32 %f624, %f619, %f623;
add.f32 %f625, %f622, %f624;
mul.rn.f32 %f627, %f602, %f839;
mul.rn.f32 %f629, %f602, %f840;
add.f32 %f630, %f627, %f623;
sub.f32 %f631, %f627, %f630;
add.f32 %f632, %f623, %f631;
add.f32 %f633, %f625, %f632;
add.f32 %f634, %f629, %f633;
add.f32 %f635, %f630, %f634;
sub.f32 %f636, %f630, %f635;
add.f32 %f637, %f634, %f636;
mul.rn.f32 %f639, %f835, %f635;
neg.f32 %f640, %f639;
fma.rn.f32 %f641, %f835, %f635, %f640;
fma.rn.f32 %f642, %f835, %f637, %f641;
fma.rn.f32 %f644, %f841, %f635, %f642;
add.rn.f32 %f645, %f639, %f644;
neg.f32 %f646, %f645;
add.rn.f32 %f647, %f639, %f646;
add.rn.f32 %f648, %f647, %f644;
mov.b32 %r266, %f645;
setp.eq.s32 %p71, %r266, 1118925336;
add.s32 %r267, %r266, -1;
mov.b32 %f649, %r267;
add.f32 %f650, %f648, 0f37000000;
selp.f32 %f651, %f649, %f645, %p71;
selp.f32 %f177, %f650, %f648, %p71;
mul.f32 %f652, %f651, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f653, %f652;
fma.rn.f32 %f655, %f653, %f842, %f651;
fma.rn.f32 %f657, %f653, %f843, %f655;
mul.f32 %f658, %f657, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f659, %f658;
add.f32 %f660, %f653, 0f00000000;
ex2.approx.f32 %f661, %f660;
mul.f32 %f662, %f659, %f661;
setp.lt.f32 %p72, %f651, 0fC2D20000;
selp.f32 %f663, 0f00000000, %f662, %p72;
setp.gt.f32 %p73, %f651, 0f42D20000;
selp.f32 %f955, 0f7F800000, %f663, %p73;
setp.eq.f32 %p74, %f955, 0f7F800000;
@%p74 bra BB0_77;
fma.rn.f32 %f955, %f955, %f177, %f955;
BB0_77:
setp.lt.f32 %p75, %f160, 0f00000000;
and.pred %p4, %p75, %p60;
mov.b32 %r268, %f955;
xor.b32 %r269, %r268, -2147483648;
mov.b32 %f664, %r269;
selp.f32 %f957, %f664, %f955, %p4;
setp.eq.f32 %p77, %f160, 0f00000000;
@%p77 bra BB0_80;
bra.uni BB0_78;
BB0_80:
add.f32 %f667, %f160, %f160;
selp.f32 %f957, %f667, 0f00000000, %p60;
bra.uni BB0_81;
BB0_78:
setp.geu.f32 %p78, %f160, 0f00000000;
@%p78 bra BB0_81;
mov.f32 %f854, 0f3EE66666;
cvt.rzi.f32.f32 %f666, %f854;
setp.neu.f32 %p79, %f666, 0f3EE66666;
selp.f32 %f957, 0f7FFFFFFF, %f957, %p79;
BB0_81:
abs.f32 %f858, %f160;
add.f32 %f668, %f858, 0f3EE66666;
mov.b32 %r270, %f668;
setp.lt.s32 %p81, %r270, 2139095040;
@%p81 bra BB0_86;
abs.f32 %f859, %f160;
setp.gtu.f32 %p82, %f859, 0f7F800000;
@%p82 bra BB0_85;
bra.uni BB0_83;
BB0_85:
add.f32 %f957, %f160, 0f3EE66666;
bra.uni BB0_86;
BB0_83:
abs.f32 %f860, %f160;
setp.neu.f32 %p83, %f860, 0f7F800000;
@%p83 bra BB0_86;
selp.f32 %f957, 0fFF800000, 0f7F800000, %p4;
BB0_86:
mov.f32 %f852, 0fB5BFBE8E;
mov.f32 %f851, 0fBF317200;
mov.f32 %f850, 0f00000000;
mov.f32 %f849, 0f35BFBE8E;
mov.f32 %f848, 0f3F317200;
mov.f32 %f847, 0f3DAAAABD;
mov.f32 %f846, 0f3C4CAF63;
mov.f32 %f845, 0f3B18F0FE;
mov.f32 %f844, 0f3EE66666;
setp.eq.f32 %p84, %f160, 0f3F800000;
selp.f32 %f188, 0f3F800000, %f957, %p84;
abs.f32 %f189, %f161;
setp.lt.f32 %p85, %f189, 0f00800000;
mul.f32 %f671, %f189, 0f4B800000;
selp.f32 %f672, 0fC3170000, 0fC2FE0000, %p85;
selp.f32 %f673, %f671, %f189, %p85;
mov.b32 %r271, %f673;
and.b32 %r272, %r271, 8388607;
or.b32 %r273, %r272, 1065353216;
mov.b32 %f674, %r273;
shr.u32 %r274, %r271, 23;
cvt.rn.f32.u32 %f675, %r274;
add.f32 %f676, %f672, %f675;
setp.gt.f32 %p86, %f674, 0f3FB504F3;
mul.f32 %f677, %f674, 0f3F000000;
add.f32 %f678, %f676, 0f3F800000;
selp.f32 %f679, %f677, %f674, %p86;
selp.f32 %f680, %f678, %f676, %p86;
add.f32 %f681, %f679, 0fBF800000;
add.f32 %f670, %f679, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f669,%f670;
// inline asm
add.f32 %f682, %f681, %f681;
mul.f32 %f683, %f669, %f682;
mul.f32 %f684, %f683, %f683;
fma.rn.f32 %f687, %f845, %f684, %f846;
fma.rn.f32 %f689, %f687, %f684, %f847;
mul.rn.f32 %f690, %f689, %f684;
mul.rn.f32 %f691, %f690, %f683;
sub.f32 %f692, %f681, %f683;
neg.f32 %f693, %f683;
add.f32 %f694, %f692, %f692;
fma.rn.f32 %f695, %f693, %f681, %f694;
mul.rn.f32 %f696, %f669, %f695;
add.f32 %f697, %f691, %f683;
sub.f32 %f698, %f683, %f697;
add.f32 %f699, %f691, %f698;
add.f32 %f700, %f696, %f699;
add.f32 %f701, %f697, %f700;
sub.f32 %f702, %f697, %f701;
add.f32 %f703, %f700, %f702;
mul.rn.f32 %f705, %f680, %f848;
mul.rn.f32 %f707, %f680, %f849;
add.f32 %f708, %f705, %f701;
sub.f32 %f709, %f705, %f708;
add.f32 %f710, %f701, %f709;
add.f32 %f711, %f703, %f710;
add.f32 %f712, %f707, %f711;
add.f32 %f713, %f708, %f712;
sub.f32 %f714, %f708, %f713;
add.f32 %f715, %f712, %f714;
mul.rn.f32 %f717, %f844, %f713;
neg.f32 %f718, %f717;
fma.rn.f32 %f719, %f844, %f713, %f718;
fma.rn.f32 %f720, %f844, %f715, %f719;
fma.rn.f32 %f722, %f850, %f713, %f720;
add.rn.f32 %f723, %f717, %f722;
neg.f32 %f724, %f723;
add.rn.f32 %f725, %f717, %f724;
add.rn.f32 %f726, %f725, %f722;
mov.b32 %r275, %f723;
setp.eq.s32 %p87, %r275, 1118925336;
add.s32 %r276, %r275, -1;
mov.b32 %f727, %r276;
add.f32 %f728, %f726, 0f37000000;
selp.f32 %f729, %f727, %f723, %p87;
selp.f32 %f190, %f728, %f726, %p87;
mul.f32 %f730, %f729, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f731, %f730;
fma.rn.f32 %f733, %f731, %f851, %f729;
fma.rn.f32 %f735, %f731, %f852, %f733;
mul.f32 %f736, %f735, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f737, %f736;
add.f32 %f738, %f731, 0f00000000;
ex2.approx.f32 %f739, %f738;
mul.f32 %f740, %f737, %f739;
setp.lt.f32 %p88, %f729, 0fC2D20000;
selp.f32 %f741, 0f00000000, %f740, %p88;
setp.gt.f32 %p89, %f729, 0f42D20000;
selp.f32 %f958, 0f7F800000, %f741, %p89;
setp.eq.f32 %p90, %f958, 0f7F800000;
@%p90 bra BB0_88;
fma.rn.f32 %f958, %f958, %f190, %f958;
BB0_88:
setp.lt.f32 %p91, %f161, 0f00000000;
and.pred %p5, %p91, %p60;
mov.b32 %r277, %f958;
xor.b32 %r278, %r277, -2147483648;
mov.b32 %f742, %r278;
selp.f32 %f960, %f742, %f958, %p5;
setp.eq.f32 %p93, %f161, 0f00000000;
@%p93 bra BB0_91;
bra.uni BB0_89;
BB0_91:
add.f32 %f745, %f161, %f161;
selp.f32 %f960, %f745, 0f00000000, %p60;
bra.uni BB0_92;
BB0_89:
setp.geu.f32 %p94, %f161, 0f00000000;
@%p94 bra BB0_92;
mov.f32 %f853, 0f3EE66666;
cvt.rzi.f32.f32 %f744, %f853;
setp.neu.f32 %p95, %f744, 0f3EE66666;
selp.f32 %f960, 0f7FFFFFFF, %f960, %p95;
BB0_92:
abs.f32 %f881, %f161;
add.f32 %f746, %f881, 0f3EE66666;
mov.b32 %r279, %f746;
setp.lt.s32 %p97, %r279, 2139095040;
@%p97 bra BB0_97;
abs.f32 %f882, %f161;
setp.gtu.f32 %p98, %f882, 0f7F800000;
@%p98 bra BB0_96;
bra.uni BB0_94;
BB0_96:
add.f32 %f960, %f161, 0f3EE66666;
bra.uni BB0_97;
BB0_94:
abs.f32 %f883, %f161;
setp.neu.f32 %p99, %f883, 0f7F800000;
@%p99 bra BB0_97;
selp.f32 %f960, 0fFF800000, 0f7F800000, %p5;
BB0_97:
mov.u32 %r351, 2;
mov.u32 %r342, 4;
mov.u64 %rd256, 0;
setp.eq.f32 %p100, %f161, 0f3F800000;
selp.f32 %f747, 0f3F800000, %f960, %p100;
cvt.u64.u32 %rd97, %r4;
cvt.u64.u32 %rd96, %r3;
mov.u64 %rd100, image;
cvta.global.u64 %rd95, %rd100;
// inline asm
call (%rd94), _rt_buffer_get_64, (%rd95, %r351, %r342, %rd96, %rd97, %rd256, %rd256);
// inline asm
cvt.sat.f32.f32 %f748, %f747;
mul.f32 %f749, %f748, 0f437FFD71;
cvt.rzi.u32.f32 %r282, %f749;
cvt.sat.f32.f32 %f750, %f188;
mul.f32 %f751, %f750, 0f437FFD71;
cvt.rzi.u32.f32 %r283, %f751;
cvt.sat.f32.f32 %f752, %f175;
mul.f32 %f753, %f752, 0f437FFD71;
cvt.rzi.u32.f32 %r284, %f753;
cvt.u16.u32 %rs46, %r282;
cvt.u16.u32 %rs47, %r284;
cvt.u16.u32 %rs48, %r283;
mov.u16 %rs49, 255;
st.v4.u8 [%rd94], {%rs46, %rs48, %rs47, %rs49};
ld.global.u32 %r410, [imageEnabled];
BB0_98:
and.b32 %r285, %r410, 4;
setp.eq.s32 %p101, %r285, 0;
@%p101 bra BB0_100;
mov.u32 %r352, 2;
mov.u32 %r343, 8;
mov.u64 %rd257, 0;
cvt.u64.u32 %rd103, %r3;
cvt.u64.u32 %rd104, %r4;
mov.u64 %rd107, image_HDR;
cvta.global.u64 %rd102, %rd107;
// inline asm
call (%rd101), _rt_buffer_get_64, (%rd102, %r352, %r343, %rd103, %rd104, %rd257, %rd257);
// inline asm
mov.f32 %f757, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs53, %f757;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs52, %f161;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs51, %f160;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs50, %f159;}
// inline asm
st.v4.u16 [%rd101], {%rs50, %rs51, %rs52, %rs53};
ld.global.u32 %r410, [imageEnabled];
BB0_100:
and.b32 %r288, %r410, 16;
setp.eq.s32 %p102, %r288, 0;
@%p102 bra BB0_102;
mov.u32 %r353, 2;
mov.u32 %r344, 8;
mov.u64 %rd258, 0;
cvt.u64.u32 %rd111, %r4;
cvt.u64.u32 %rd110, %r3;
mov.u64 %rd114, image_HDR2;
cvta.global.u64 %rd109, %rd114;
// inline asm
call (%rd108), _rt_buffer_get_64, (%rd109, %r353, %r344, %rd110, %rd111, %rd258, %rd258);
// inline asm
mov.f32 %f761, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs57, %f761;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs56, %f155;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs55, %f154;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs54, %f153;}
// inline asm
st.v4.u16 [%rd108], {%rs54, %rs55, %rs56, %rs57};
BB0_102:
cvt.u64.u32 %rd18, %r3;
cvt.u64.u32 %rd19, %r4;
add.f32 %f763, %f141, %f144;
add.f32 %f764, %f147, %f763;
add.f32 %f765, %f142, %f145;
add.f32 %f766, %f148, %f765;
add.f32 %f767, %f143, %f146;
add.f32 %f768, %f149, %f767;
mul.f32 %f769, %f764, 0f3F13CD3A;
mul.f32 %f770, %f766, 0f3F13CD3A;
mul.f32 %f771, %f768, 0f3F13CD3A;
div.rn.f32 %f772, %f150, %f769;
div.rn.f32 %f773, %f151, %f770;
div.rn.f32 %f774, %f152, %f771;
setp.eq.f32 %p103, %f150, 0f00000000;
selp.f32 %f775, 0f00000000, %f772, %p103;
setp.eq.f32 %p104, %f151, 0f00000000;
selp.f32 %f776, 0f00000000, %f773, %p104;
setp.eq.f32 %p105, %f152, 0f00000000;
selp.f32 %f777, 0f00000000, %f774, %p105;
mul.f32 %f201, %f141, %f775;
mul.f32 %f202, %f142, %f776;
mul.f32 %f203, %f143, %f777;
mul.f32 %f204, %f144, %f775;
mul.f32 %f205, %f145, %f776;
mul.f32 %f206, %f146, %f777;
mul.f32 %f207, %f147, %f775;
mul.f32 %f208, %f148, %f776;
mul.f32 %f209, %f149, %f777;
ld.global.u32 %r291, [additive];
setp.eq.s32 %p106, %r291, 0;
mov.f32 %f762, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs58, %f762;}
// inline asm
@%p106 bra BB0_104;
mov.u32 %r354, 2;
mov.u32 %r345, 8;
mov.u64 %rd259, 0;
mov.u64 %rd127, image_RNM0;
cvta.global.u64 %rd116, %rd127;
// inline asm
call (%rd115), _rt_buffer_get_64, (%rd116, %r354, %r345, %rd18, %rd19, %rd259, %rd259);
// inline asm
ld.v4.u16 {%rs65, %rs66, %rs67, %rs68}, [%rd115];
// inline asm
{ cvt.f32.f16 %f778, %rs65;}
// inline asm
// inline asm
{ cvt.f32.f16 %f779, %rs66;}
// inline asm
// inline asm
{ cvt.f32.f16 %f780, %rs67;}
// inline asm
// inline asm
call (%rd121), _rt_buffer_get_64, (%rd116, %r354, %r345, %rd18, %rd19, %rd259, %rd259);
// inline asm
add.f32 %f781, %f201, %f778;
add.f32 %f782, %f202, %f779;
add.f32 %f783, %f203, %f780;
// inline asm
{ cvt.rn.f16.f32 %rs64, %f783;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs63, %f782;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs62, %f781;}
// inline asm
st.v4.u16 [%rd121], {%rs62, %rs63, %rs64, %rs58};
bra.uni BB0_105;
BB0_104:
mov.u32 %r359, 2;
mov.u32 %r350, 8;
mov.u64 %rd264, 0;
mov.u64 %rd134, image_RNM0;
cvta.global.u64 %rd129, %rd134;
// inline asm
call (%rd128), _rt_buffer_get_64, (%rd129, %r359, %r350, %rd18, %rd19, %rd264, %rd264);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs71, %f203;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs70, %f202;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs69, %f201;}
// inline asm
st.v4.u16 [%rd128], {%rs69, %rs70, %rs71, %rs58};
BB0_105:
ld.global.u32 %r298, [additive];
setp.eq.s32 %p107, %r298, 0;
// inline asm
{ cvt.rn.f16.f32 %rs72, %f762;}
// inline asm
@%p107 bra BB0_107;
mov.u32 %r355, 2;
mov.u32 %r346, 8;
mov.u64 %rd260, 0;
mov.u64 %rd147, image_RNM1;
cvta.global.u64 %rd136, %rd147;
// inline asm
call (%rd135), _rt_buffer_get_64, (%rd136, %r355, %r346, %rd18, %rd19, %rd260, %rd260);
// inline asm
ld.v4.u16 {%rs79, %rs80, %rs81, %rs82}, [%rd135];
// inline asm
{ cvt.f32.f16 %f788, %rs79;}
// inline asm
// inline asm
{ cvt.f32.f16 %f789, %rs80;}
// inline asm
// inline asm
{ cvt.f32.f16 %f790, %rs81;}
// inline asm
// inline asm
call (%rd141), _rt_buffer_get_64, (%rd136, %r355, %r346, %rd18, %rd19, %rd260, %rd260);
// inline asm
add.f32 %f791, %f204, %f788;
add.f32 %f792, %f205, %f789;
add.f32 %f793, %f206, %f790;
// inline asm
{ cvt.rn.f16.f32 %rs78, %f793;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs77, %f792;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs76, %f791;}
// inline asm
st.v4.u16 [%rd141], {%rs76, %rs77, %rs78, %rs72};
bra.uni BB0_108;
BB0_107:
mov.u32 %r358, 2;
mov.u32 %r349, 8;
mov.u64 %rd263, 0;
mov.u64 %rd154, image_RNM1;
cvta.global.u64 %rd149, %rd154;
// inline asm
call (%rd148), _rt_buffer_get_64, (%rd149, %r358, %r349, %rd18, %rd19, %rd263, %rd263);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs85, %f206;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs84, %f205;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs83, %f204;}
// inline asm
st.v4.u16 [%rd148], {%rs83, %rs84, %rs85, %rs72};
BB0_108:
ld.global.u32 %r305, [additive];
setp.eq.s32 %p108, %r305, 0;
// inline asm
{ cvt.rn.f16.f32 %rs86, %f762;}
// inline asm
@%p108 bra BB0_110;
mov.u32 %r356, 2;
mov.u32 %r347, 8;
mov.u64 %rd261, 0;
mov.u64 %rd167, image_RNM2;
cvta.global.u64 %rd156, %rd167;
// inline asm
call (%rd155), _rt_buffer_get_64, (%rd156, %r356, %r347, %rd18, %rd19, %rd261, %rd261);
// inline asm
ld.v4.u16 {%rs93, %rs94, %rs95, %rs96}, [%rd155];
// inline asm
{ cvt.f32.f16 %f798, %rs93;}
// inline asm
// inline asm
{ cvt.f32.f16 %f799, %rs94;}
// inline asm
// inline asm
{ cvt.f32.f16 %f800, %rs95;}
// inline asm
// inline asm
call (%rd161), _rt_buffer_get_64, (%rd156, %r356, %r347, %rd18, %rd19, %rd261, %rd261);
// inline asm
add.f32 %f801, %f207, %f798;
add.f32 %f802, %f208, %f799;
add.f32 %f803, %f209, %f800;
// inline asm
{ cvt.rn.f16.f32 %rs92, %f803;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs91, %f802;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs90, %f801;}
// inline asm
st.v4.u16 [%rd161], {%rs90, %rs91, %rs92, %rs86};
bra.uni BB0_126;
BB0_110:
mov.u32 %r357, 2;
mov.u32 %r348, 8;
mov.u64 %rd262, 0;
mov.u64 %rd174, image_RNM2;
cvta.global.u64 %rd169, %rd174;
// inline asm
call (%rd168), _rt_buffer_get_64, (%rd169, %r357, %r348, %rd18, %rd19, %rd262, %rd262);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs99, %f209;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs98, %f208;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs97, %f207;}
// inline asm
st.v4.u16 [%rd168], {%rs97, %rs98, %rs99, %rs86};
BB0_126:
ret;
}