Files
2026-08-06 18:56:32 +03:00

1929 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_Mask[1];
.global .align 1 .b8 image_Dir[1];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 4 .u32 ignoreNormal;
.global .align 1 .b8 localLights[1];
.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_typeinfo12ignoreNormalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[4] = {105, 110, 116, 0};
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 4919;
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[1];
.visible .entry _Z6oxMainv(
)
{
.local .align 4 .b8 __local_depot0[4];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<134>;
.reg .b16 %rs<63>;
.reg .f32 %f<1268>;
.reg .b32 %r<205>;
.reg .b64 %rd<126>;
mov.u64 %rd125, __local_depot0;
cvta.local.u64 %SP, %rd125;
ld.global.v2.u32 {%r30, %r31}, [pixelID];
cvt.u64.u32 %rd10, %r30;
cvt.u64.u32 %rd11, %r31;
mov.u64 %rd14, uvnormal;
cvta.global.u64 %rd9, %rd14;
mov.u32 %r28, 2;
mov.u32 %r29, 4;
mov.u64 %rd13, 0;
// inline asm
call (%rd8), _rt_buffer_get_64, (%rd9, %r28, %r29, %rd10, %rd11, %rd13, %rd13);
// inline asm
ld.u32 %r1, [%rd8];
shr.u32 %r34, %r1, 16;
cvt.u16.u32 %rs1, %r34;
and.b16 %rs5, %rs1, 255;
cvt.u16.u32 %rs6, %r1;
or.b16 %rs7, %rs6, %rs5;
setp.eq.s16 %p8, %rs7, 0;
mov.f32 %f1205, 0f00000000;
mov.f32 %f1206, %f1205;
mov.f32 %f1207, %f1205;
@%p8 bra BB0_2;
ld.u8 %rs8, [%rd8+1];
and.b16 %rs10, %rs6, 255;
cvt.rn.f32.u16 %f220, %rs10;
div.rn.f32 %f221, %f220, 0f437F0000;
fma.rn.f32 %f222, %f221, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f223, %rs8;
div.rn.f32 %f224, %f223, 0f437F0000;
fma.rn.f32 %f225, %f224, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f226, %rs5;
div.rn.f32 %f227, %f226, 0f437F0000;
fma.rn.f32 %f228, %f227, 0f40000000, 0fBF800000;
mul.f32 %f229, %f225, %f225;
fma.rn.f32 %f230, %f222, %f222, %f229;
fma.rn.f32 %f231, %f228, %f228, %f230;
sqrt.rn.f32 %f232, %f231;
rcp.rn.f32 %f233, %f232;
mul.f32 %f1205, %f222, %f233;
mul.f32 %f1206, %f225, %f233;
mul.f32 %f1207, %f228, %f233;
BB0_2:
ld.global.v2.u32 {%r35, %r36}, [pixelID];
ld.global.v2.u32 {%r38, %r39}, [tileInfo];
add.s32 %r2, %r35, %r38;
add.s32 %r3, %r36, %r39;
setp.eq.f32 %p9, %f1206, 0f00000000;
setp.eq.f32 %p10, %f1205, 0f00000000;
and.pred %p11, %p10, %p9;
setp.eq.f32 %p12, %f1207, 0f00000000;
and.pred %p13, %p11, %p12;
@%p13 bra BB0_103;
bra.uni BB0_3;
BB0_103:
ld.global.u32 %r204, [imageEnabled];
and.b32 %r179, %r204, 1;
setp.eq.b32 %p128, %r179, 1;
@!%p128 bra BB0_105;
bra.uni BB0_104;
BB0_104:
cvt.u64.u32 %rd86, %r2;
cvt.u64.u32 %rd87, %r3;
mov.u64 %rd90, image;
cvta.global.u64 %rd85, %rd90;
// inline asm
call (%rd84), _rt_buffer_get_64, (%rd85, %r28, %r29, %rd86, %rd87, %rd13, %rd13);
// inline asm
mov.u16 %rs43, 0;
st.v4.u8 [%rd84], {%rs43, %rs43, %rs43, %rs43};
ld.global.u32 %r204, [imageEnabled];
BB0_105:
and.b32 %r182, %r204, 8;
setp.eq.s32 %p129, %r182, 0;
@%p129 bra BB0_107;
cvt.u64.u32 %rd93, %r2;
cvt.u64.u32 %rd94, %r3;
mov.u64 %rd97, image_Mask;
cvta.global.u64 %rd92, %rd97;
// inline asm
call (%rd91), _rt_buffer_get_64, (%rd92, %r28, %r28, %rd93, %rd94, %rd13, %rd13);
// inline asm
mov.f32 %f1194, 0f00000000;
cvt.rzi.u32.f32 %r185, %f1194;
cvt.u16.u32 %rs44, %r185;
mov.u16 %rs45, 0;
st.v2.u8 [%rd91], {%rs44, %rs45};
ld.global.u32 %r204, [imageEnabled];
BB0_107:
and.b32 %r186, %r204, 4;
setp.eq.s32 %p130, %r186, 0;
@%p130 bra BB0_111;
ld.global.u32 %r187, [additive];
setp.eq.s32 %p131, %r187, 0;
cvt.u64.u32 %rd6, %r2;
cvt.u64.u32 %rd7, %r3;
@%p131 bra BB0_110;
mov.u64 %rd110, image_HDR;
cvta.global.u64 %rd99, %rd110;
mov.u32 %r191, 8;
// inline asm
call (%rd98), _rt_buffer_get_64, (%rd99, %r28, %r191, %rd6, %rd7, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs52, %rs53, %rs54, %rs55}, [%rd98];
// inline asm
{ cvt.f32.f16 %f1195, %rs52;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1196, %rs53;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1197, %rs54;}
// inline asm
// inline asm
call (%rd104), _rt_buffer_get_64, (%rd99, %r28, %r191, %rd6, %rd7, %rd13, %rd13);
// inline asm
add.f32 %f1198, %f1195, 0f00000000;
add.f32 %f1199, %f1196, 0f00000000;
add.f32 %f1200, %f1197, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs51, %f1200;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs50, %f1199;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs49, %f1198;}
// inline asm
mov.u16 %rs56, 0;
st.v4.u16 [%rd104], {%rs49, %rs50, %rs51, %rs56};
bra.uni BB0_111;
BB0_3:
ld.global.v2.u32 {%r47, %r48}, [pixelID];
cvt.u64.u32 %rd17, %r47;
cvt.u64.u32 %rd18, %r48;
mov.u64 %rd26, uvpos;
cvta.global.u64 %rd16, %rd26;
mov.u32 %r44, 12;
// inline asm
call (%rd15), _rt_buffer_get_64, (%rd16, %r28, %r44, %rd17, %rd18, %rd13, %rd13);
// inline asm
ld.f32 %f9, [%rd15+8];
ld.f32 %f8, [%rd15+4];
ld.f32 %f7, [%rd15];
mul.f32 %f241, %f7, 0f3456BF95;
mul.f32 %f242, %f8, 0f3456BF95;
mul.f32 %f243, %f9, 0f3456BF95;
abs.f32 %f244, %f1205;
div.rn.f32 %f245, %f241, %f244;
abs.f32 %f246, %f1206;
div.rn.f32 %f247, %f242, %f246;
abs.f32 %f248, %f1207;
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 %f10, %f1205, %f254, %f7;
fma.rn.f32 %f11, %f1206, %f255, %f8;
fma.rn.f32 %f12, %f1207, %f256, %f9;
ld.global.v2.u32 {%r51, %r52}, [pixelID];
cvt.rn.f32.u32 %f13, %r51;
cvt.rn.f32.u32 %f14, %r52;
mov.u64 %rd27, localLights;
cvta.global.u64 %rd25, %rd27;
mov.u32 %r45, 1;
mov.u32 %r46, 96;
// inline asm
call (%rd21, %rd22, %rd23, %rd24), _rt_buffer_get_size_64, (%rd25, %r45, %r46);
// inline asm
cvt.u32.u64 %r4, %rd21;
setp.eq.s32 %p14, %r4, 0;
mov.f32 %f1208, 0f00000000;
mov.f32 %f22, %f1208;
mov.f32 %f23, %f1208;
mov.f32 %f24, %f1208;
mov.f32 %f1212, %f1208;
mov.f32 %f1213, %f1208;
mov.f32 %f1214, %f1208;
@%p14 bra BB0_49;
mov.f32 %f264, 0f40000000;
cvt.rzi.f32.f32 %f265, %f264;
add.f32 %f266, %f265, %f265;
mov.f32 %f267, 0f40800000;
sub.f32 %f268, %f267, %f266;
abs.f32 %f15, %f268;
mul.f32 %f16, %f10, 0f3456BF95;
mul.f32 %f17, %f11, 0f3456BF95;
mul.f32 %f18, %f12, 0f3456BF95;
mul.f32 %f19, %f13, 0f3DD32618;
mul.f32 %f20, %f14, 0f3DD2F1AA;
mov.f32 %f263, 0f00000000;
mov.u32 %r196, 0;
abs.f32 %f448, %f16;
abs.f32 %f449, %f17;
max.f32 %f450, %f448, %f449;
abs.f32 %f451, %f18;
max.f32 %f452, %f450, %f451;
mov.f32 %f1208, %f263;
mov.f32 %f22, %f263;
mov.f32 %f23, %f263;
mov.f32 %f24, %f263;
mov.f32 %f1212, %f263;
mov.f32 %f1213, %f263;
mov.f32 %f1214, %f263;
BB0_5:
cvt.u64.u32 %rd30, %r196;
// inline asm
call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
// inline asm
ld.v4.f32 {%f271, %f272, %f273, %f274}, [%rd28+80];
ld.v4.f32 {%f275, %f276, %f277, %f278}, [%rd28+64];
ld.v4.f32 {%f279, %f280, %f281, %f282}, [%rd28+48];
ld.v4.f32 {%f283, %f1223, %f1224, %f286}, [%rd28+32];
ld.v4.f32 {%f287, %f288, %f289, %f290}, [%rd28+16];
ld.v4.f32 {%f291, %f292, %f293, %f294}, [%rd28];
mov.b32 %r6, %f274;
sub.f32 %f52, %f292, %f7;
sub.f32 %f54, %f293, %f8;
sub.f32 %f56, %f294, %f9;
mul.f32 %f296, %f54, %f54;
fma.rn.f32 %f297, %f52, %f52, %f296;
fma.rn.f32 %f298, %f56, %f56, %f297;
sqrt.rn.f32 %f57, %f298;
rcp.rn.f32 %f299, %f57;
mul.f32 %f58, %f52, %f299;
mul.f32 %f59, %f54, %f299;
mul.f32 %f60, %f56, %f299;
mul.f32 %f61, %f57, %f290;
abs.f32 %f62, %f61;
setp.lt.f32 %p15, %f62, 0f00800000;
mul.f32 %f300, %f62, 0f4B800000;
selp.f32 %f301, 0fC3170000, 0fC2FE0000, %p15;
selp.f32 %f302, %f300, %f62, %p15;
mov.b32 %r58, %f302;
and.b32 %r59, %r58, 8388607;
or.b32 %r60, %r59, 1065353216;
mov.b32 %f303, %r60;
shr.u32 %r61, %r58, 23;
cvt.rn.f32.u32 %f304, %r61;
add.f32 %f305, %f301, %f304;
setp.gt.f32 %p16, %f303, 0f3FB504F3;
mul.f32 %f306, %f303, 0f3F000000;
add.f32 %f307, %f305, 0f3F800000;
selp.f32 %f308, %f306, %f303, %p16;
selp.f32 %f309, %f307, %f305, %p16;
add.f32 %f310, %f308, 0fBF800000;
add.f32 %f270, %f308, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f269,%f270;
// inline asm
add.f32 %f311, %f310, %f310;
mul.f32 %f312, %f269, %f311;
mul.f32 %f313, %f312, %f312;
mov.f32 %f314, 0f3C4CAF63;
mov.f32 %f315, 0f3B18F0FE;
fma.rn.f32 %f316, %f315, %f313, %f314;
mov.f32 %f317, 0f3DAAAABD;
fma.rn.f32 %f318, %f316, %f313, %f317;
mul.rn.f32 %f319, %f318, %f313;
mul.rn.f32 %f320, %f319, %f312;
sub.f32 %f321, %f310, %f312;
neg.f32 %f322, %f312;
add.f32 %f323, %f321, %f321;
fma.rn.f32 %f324, %f322, %f310, %f323;
mul.rn.f32 %f325, %f269, %f324;
add.f32 %f326, %f320, %f312;
sub.f32 %f327, %f312, %f326;
add.f32 %f328, %f320, %f327;
add.f32 %f329, %f325, %f328;
add.f32 %f330, %f326, %f329;
sub.f32 %f331, %f326, %f330;
add.f32 %f332, %f329, %f331;
mov.f32 %f333, 0f3F317200;
mul.rn.f32 %f334, %f309, %f333;
mov.f32 %f335, 0f35BFBE8E;
mul.rn.f32 %f336, %f309, %f335;
add.f32 %f337, %f334, %f330;
sub.f32 %f338, %f334, %f337;
add.f32 %f339, %f330, %f338;
add.f32 %f340, %f332, %f339;
add.f32 %f341, %f336, %f340;
add.f32 %f342, %f337, %f341;
sub.f32 %f343, %f337, %f342;
add.f32 %f344, %f341, %f343;
mul.rn.f32 %f63, %f267, %f342;
neg.f32 %f346, %f63;
fma.rn.f32 %f347, %f267, %f342, %f346;
fma.rn.f32 %f348, %f267, %f344, %f347;
fma.rn.f32 %f64, %f263, %f342, %f348;
add.rn.f32 %f65, %f63, %f64;
mov.b32 %r62, %f65;
setp.eq.s32 %p1, %r62, 1118925336;
add.s32 %r63, %r62, -1;
mov.b32 %f350, %r63;
selp.f32 %f351, %f350, %f65, %p1;
mul.f32 %f352, %f351, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f353, %f352;
mov.f32 %f354, 0fBF317200;
fma.rn.f32 %f355, %f353, %f354, %f351;
mov.f32 %f356, 0fB5BFBE8E;
fma.rn.f32 %f357, %f353, %f356, %f355;
mul.f32 %f358, %f357, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f359, %f358;
add.f32 %f360, %f353, 0f00000000;
ex2.approx.f32 %f361, %f360;
mul.f32 %f362, %f359, %f361;
setp.lt.f32 %p17, %f351, 0fC2D20000;
selp.f32 %f363, 0f00000000, %f362, %p17;
setp.gt.f32 %p18, %f351, 0f42D20000;
selp.f32 %f1215, 0f7F800000, %f363, %p18;
setp.eq.f32 %p19, %f1215, 0f7F800000;
@%p19 bra BB0_7;
neg.f32 %f364, %f65;
add.rn.f32 %f365, %f63, %f364;
add.rn.f32 %f366, %f365, %f64;
add.f32 %f367, %f366, 0f37000000;
selp.f32 %f368, %f367, %f366, %p1;
fma.rn.f32 %f1215, %f1215, %f368, %f1215;
BB0_7:
setp.lt.f32 %p20, %f61, 0f00000000;
setp.eq.f32 %p21, %f15, 0f3F800000;
and.pred %p2, %p20, %p21;
mov.b32 %r64, %f1215;
xor.b32 %r65, %r64, -2147483648;
mov.b32 %f369, %r65;
selp.f32 %f1217, %f369, %f1215, %p2;
setp.eq.f32 %p22, %f61, 0f00000000;
@%p22 bra BB0_10;
bra.uni BB0_8;
BB0_10:
add.f32 %f372, %f61, %f61;
selp.f32 %f1217, %f372, 0f00000000, %p21;
bra.uni BB0_11;
BB0_8:
setp.geu.f32 %p23, %f61, 0f00000000;
@%p23 bra BB0_11;
cvt.rzi.f32.f32 %f371, %f267;
setp.neu.f32 %p24, %f371, 0f40800000;
selp.f32 %f1217, 0f7FFFFFFF, %f1217, %p24;
BB0_11:
add.f32 %f373, %f62, 0f40800000;
mov.b32 %r66, %f373;
setp.lt.s32 %p26, %r66, 2139095040;
@%p26 bra BB0_16;
setp.gtu.f32 %p27, %f62, 0f7F800000;
@%p27 bra BB0_15;
bra.uni BB0_13;
BB0_15:
add.f32 %f1217, %f61, 0f40800000;
bra.uni BB0_16;
BB0_13:
setp.neu.f32 %p28, %f62, 0f7F800000;
@%p28 bra BB0_16;
selp.f32 %f1217, 0fFF800000, 0f7F800000, %p2;
BB0_16:
mul.f32 %f374, %f57, %f288;
mov.f32 %f1237, 0f3F800000;
sub.f32 %f376, %f1237, %f1217;
setp.eq.f32 %p29, %f61, 0f3F800000;
selp.f32 %f377, 0f00000000, %f376, %p29;
cvt.sat.f32.f32 %f378, %f377;
fma.rn.f32 %f379, %f374, %f374, %f289;
div.rn.f32 %f91, %f378, %f379;
mul.f32 %f380, %f1206, %f59;
fma.rn.f32 %f381, %f1205, %f58, %f380;
fma.rn.f32 %f382, %f1207, %f60, %f381;
ld.global.u32 %r67, [ignoreNormal];
setp.eq.s32 %p30, %r67, 0;
selp.f32 %f383, %f382, 0f3F800000, %p30;
cvt.sat.f32.f32 %f92, %f383;
setp.eq.f32 %p31, %f291, 0f3F800000;
@%p31 bra BB0_23;
bra.uni BB0_17;
BB0_23:
setp.leu.f32 %p35, %f286, 0f00000000;
@%p35 bra BB0_19;
setp.geu.f32 %p36, %f282, 0f00000000;
@%p36 bra BB0_25;
sub.f32 %f1204, %f294, %f9;
sub.f32 %f1203, %f292, %f7;
sub.f32 %f1202, %f293, %f8;
mul.f32 %f406, %f272, %f1202;
fma.rn.f32 %f407, %f271, %f1203, %f406;
fma.rn.f32 %f408, %f273, %f1204, %f407;
rcp.rn.f32 %f409, %f408;
mul.f32 %f1218, %f1203, %f409;
mul.f32 %f1219, %f1202, %f409;
mul.f32 %f1220, %f1204, %f409;
neg.f32 %f282, %f282;
bra.uni BB0_27;
BB0_17:
setp.eq.f32 %p32, %f291, 0f40000000;
@%p32 bra BB0_21;
bra.uni BB0_18;
BB0_21:
setp.leu.f32 %p34, %f286, 0f00000000;
@%p34 bra BB0_19;
mul.f32 %f400, %f280, %f59;
fma.rn.f32 %f401, %f279, %f58, %f400;
mul.f32 %f402, %f276, %f59;
fma.rn.f32 %f403, %f275, %f58, %f402;
mul.f32 %f404, %f272, %f59;
fma.rn.f32 %f405, %f271, %f58, %f404;
fma.rn.f32 %f397, %f281, %f60, %f401;
fma.rn.f32 %f398, %f277, %f60, %f403;
fma.rn.f32 %f399, %f273, %f60, %f405;
cvt.rzi.s32.f32 %r68, %f286;
mov.u32 %r69, 6;
mov.u32 %r70, 0;
// inline asm
call (%f393, %f394, %f395, %f396), _rt_texture_get_base_id, (%r68, %r69, %f397, %f398, %f399, %r70);
// inline asm
mul.f32 %f1222, %f283, %f393;
mul.f32 %f1223, %f1223, %f394;
mul.f32 %f1224, %f1224, %f395;
bra.uni BB0_28;
BB0_18:
setp.neu.f32 %p33, %f291, 0f40800000;
@%p33 bra BB0_19;
mul.f32 %f384, %f271, %f58;
mul.f32 %f385, %f272, %f59;
neg.f32 %f386, %f385;
sub.f32 %f387, %f386, %f384;
mul.f32 %f388, %f273, %f60;
sub.f32 %f389, %f387, %f388;
fma.rn.f32 %f390, %f286, %f389, %f282;
cvt.sat.f32.f32 %f391, %f390;
mul.f32 %f392, %f391, %f391;
mul.f32 %f1225, %f91, %f392;
mov.f32 %f1222, %f283;
bra.uni BB0_29;
BB0_19:
mov.f32 %f1222, %f283;
bra.uni BB0_28;
BB0_25:
mov.f32 %f1218, %f58;
mov.f32 %f1219, %f59;
mov.f32 %f1220, %f60;
BB0_27:
mul.f32 %f418, %f273, %f60;
mul.f32 %f419, %f272, %f59;
neg.f32 %f420, %f419;
mul.f32 %f421, %f271, %f58;
sub.f32 %f422, %f420, %f421;
sub.f32 %f423, %f422, %f418;
setp.gt.f32 %p37, %f423, 0f00000000;
selp.f32 %f424, 0f3F800000, 0f00000000, %p37;
mul.f32 %f425, %f280, %f1219;
fma.rn.f32 %f426, %f279, %f1218, %f425;
mul.f32 %f427, %f276, %f1219;
fma.rn.f32 %f428, %f275, %f1218, %f427;
fma.rn.f32 %f429, %f281, %f1220, %f426;
fma.rn.f32 %f430, %f277, %f1220, %f428;
fma.rn.f32 %f431, %f282, %f429, 0f3F000000;
mov.f32 %f432, 0f3F800000;
sub.f32 %f414, %f432, %f431;
fma.rn.f32 %f415, %f282, %f430, 0f3F000000;
cvt.rzi.s32.f32 %r71, %f286;
mov.f32 %f417, 0f00000000;
// inline asm
call (%f410, %f411, %f412, %f413), _rt_texture_get_f_id, (%r71, %r28, %f414, %f415, %f417, %f417);
// inline asm
mul.f32 %f433, %f424, %f410;
mul.f32 %f434, %f424, %f411;
mul.f32 %f435, %f424, %f412;
mul.f32 %f1222, %f283, %f433;
mul.f32 %f1223, %f1223, %f434;
mul.f32 %f1224, %f1224, %f435;
BB0_28:
mov.f32 %f1225, %f91;
BB0_29:
max.f32 %f442, %f1222, %f1223;
max.f32 %f443, %f442, %f1224;
mul.f32 %f112, %f92, %f1225;
mul.f32 %f444, %f112, %f443;
setp.lt.f32 %p39, %f444, 0f3727C5AC;
mov.pred %p133, -1;
mov.f32 %f119, 0f00000000;
mov.f32 %f120, %f119;
mov.f32 %f121, %f119;
mov.f32 %f122, %f119;
mov.f32 %f123, %f119;
mov.f32 %f124, %f119;
@%p39 bra BB0_31;
mul.f32 %f119, %f1222, %f112;
mul.f32 %f120, %f1223, %f112;
mul.f32 %f121, %f1224, %f112;
ld.global.u8 %rs12, [imageEnabled];
and.b16 %rs13, %rs12, 64;
setp.eq.s16 %p41, %rs13, 0;
selp.f32 %f122, 0f00000000, %f58, %p41;
selp.f32 %f123, 0f00000000, %f59, %p41;
selp.f32 %f124, 0f00000000, %f60, %p41;
mov.pred %p133, 0;
BB0_31:
@%p133 bra BB0_48;
setp.eq.s32 %p42, %r6, 0;
mov.u16 %rs62, 0;
@%p42 bra BB0_43;
abs.s32 %r8, %r6;
mov.f32 %f1236, 0f00000000;
setp.lt.s32 %p43, %r8, 1;
@%p43 bra BB0_42;
max.f32 %f126, %f452, %f253;
and.b32 %r9, %r8, 3;
setp.eq.s32 %p44, %r9, 0;
add.u64 %rd35, %SP, 0;
cvta.to.local.u64 %rd2, %rd35;
mov.f32 %f1236, 0f00000000;
mov.u32 %r200, 0;
@%p44 bra BB0_40;
setp.eq.s32 %p45, %r9, 1;
mov.f32 %f1233, 0f00000000;
mov.u32 %r198, 0;
@%p45 bra BB0_39;
setp.eq.s32 %p46, %r9, 2;
mov.f32 %f1232, 0f00000000;
mov.u32 %r197, 0;
@%p46 bra BB0_38;
cvt.rmi.f32.f32 %f464, %f19;
sub.f32 %f465, %f19, %f464;
cvt.rmi.f32.f32 %f466, %f20;
sub.f32 %f467, %f20, %f466;
add.f32 %f468, %f467, 0f420551EC;
add.f32 %f469, %f465, 0f420551EC;
mul.f32 %f470, %f465, %f468;
fma.rn.f32 %f471, %f467, %f469, %f470;
add.f32 %f472, %f471, 0f00000000;
add.f32 %f473, %f465, %f472;
add.f32 %f474, %f467, %f472;
add.f32 %f475, %f473, %f474;
mul.f32 %f476, %f472, %f475;
cvt.rmi.f32.f32 %f477, %f476;
sub.f32 %f478, %f476, %f477;
add.f32 %f479, %f473, %f473;
mul.f32 %f480, %f474, %f479;
cvt.rmi.f32.f32 %f481, %f480;
sub.f32 %f482, %f480, %f481;
mul.f32 %f483, %f473, %f475;
cvt.rmi.f32.f32 %f484, %f483;
sub.f32 %f485, %f483, %f484;
fma.rn.f32 %f486, %f478, 0f40000000, 0fBF800000;
fma.rn.f32 %f487, %f482, 0f40000000, 0fBF800000;
fma.rn.f32 %f488, %f485, 0f40000000, 0fBF800000;
fma.rn.f32 %f489, %f287, %f486, %f292;
fma.rn.f32 %f490, %f287, %f487, %f293;
fma.rn.f32 %f491, %f287, %f488, %f294;
sub.f32 %f492, %f489, %f7;
sub.f32 %f493, %f490, %f8;
sub.f32 %f494, %f491, %f9;
mul.f32 %f495, %f493, %f493;
fma.rn.f32 %f496, %f492, %f492, %f495;
fma.rn.f32 %f497, %f494, %f494, %f496;
sqrt.rn.f32 %f463, %f497;
rcp.rn.f32 %f498, %f463;
mul.f32 %f459, %f498, %f492;
mul.f32 %f460, %f498, %f493;
mul.f32 %f461, %f498, %f494;
ld.global.u32 %r80, [imageEnabled];
and.b32 %r81, %r80, 32;
setp.eq.s32 %p47, %r81, 0;
selp.f32 %f499, 0f3F800000, 0f41200000, %p47;
mul.f32 %f462, %f499, %f126;
mov.u32 %r82, 1065353216;
st.local.u32 [%rd2], %r82;
ld.global.u32 %r76, [root];
// inline asm
call _rt_trace_64, (%r76, %f10, %f11, %f12, %f459, %f460, %f461, %r45, %f462, %f463, %rd35, %r29);
// inline asm
ld.local.f32 %f500, [%rd2];
add.f32 %f1232, %f500, 0f00000000;
mov.u32 %r197, %r45;
BB0_38:
cvt.rn.f32.s32 %f509, %r197;
add.f32 %f510, %f13, %f509;
sub.f32 %f511, %f14, %f509;
mul.f32 %f512, %f510, 0f3DD32618;
cvt.rmi.f32.f32 %f513, %f512;
sub.f32 %f514, %f512, %f513;
mul.f32 %f515, %f511, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f516, %f515;
sub.f32 %f517, %f515, %f516;
mul.f32 %f518, %f509, 0f3DC74539;
cvt.rmi.f32.f32 %f519, %f518;
sub.f32 %f520, %f518, %f519;
add.f32 %f521, %f517, 0f420551EC;
add.f32 %f522, %f514, 0f420551EC;
add.f32 %f523, %f520, 0f420551EC;
mul.f32 %f524, %f514, %f521;
fma.rn.f32 %f525, %f517, %f522, %f524;
fma.rn.f32 %f526, %f520, %f523, %f525;
add.f32 %f527, %f514, %f526;
add.f32 %f528, %f517, %f526;
add.f32 %f529, %f520, %f526;
add.f32 %f530, %f527, %f528;
mul.f32 %f531, %f529, %f530;
cvt.rmi.f32.f32 %f532, %f531;
sub.f32 %f533, %f531, %f532;
add.f32 %f534, %f527, %f527;
mul.f32 %f535, %f528, %f534;
cvt.rmi.f32.f32 %f536, %f535;
sub.f32 %f537, %f535, %f536;
mul.f32 %f538, %f527, %f530;
cvt.rmi.f32.f32 %f539, %f538;
sub.f32 %f540, %f538, %f539;
fma.rn.f32 %f541, %f533, 0f40000000, 0fBF800000;
fma.rn.f32 %f542, %f537, 0f40000000, 0fBF800000;
fma.rn.f32 %f543, %f540, 0f40000000, 0fBF800000;
fma.rn.f32 %f544, %f287, %f541, %f292;
fma.rn.f32 %f545, %f287, %f542, %f293;
fma.rn.f32 %f546, %f287, %f543, %f294;
sub.f32 %f547, %f544, %f7;
sub.f32 %f548, %f545, %f8;
sub.f32 %f549, %f546, %f9;
mul.f32 %f550, %f548, %f548;
fma.rn.f32 %f551, %f547, %f547, %f550;
fma.rn.f32 %f552, %f549, %f549, %f551;
sqrt.rn.f32 %f508, %f552;
rcp.rn.f32 %f553, %f508;
mul.f32 %f504, %f553, %f547;
mul.f32 %f505, %f553, %f548;
mul.f32 %f506, %f553, %f549;
ld.global.u32 %r86, [imageEnabled];
and.b32 %r87, %r86, 32;
setp.eq.s32 %p48, %r87, 0;
selp.f32 %f554, 0f3F800000, 0f41200000, %p48;
mul.f32 %f507, %f554, %f126;
mov.u32 %r88, 1065353216;
st.local.u32 [%rd2], %r88;
ld.global.u32 %r83, [root];
// inline asm
call _rt_trace_64, (%r83, %f10, %f11, %f12, %f504, %f505, %f506, %r45, %f507, %f508, %rd35, %r29);
// inline asm
ld.local.f32 %f555, [%rd2];
add.f32 %f1233, %f1232, %f555;
add.s32 %r198, %r197, 1;
BB0_39:
cvt.rn.f32.s32 %f564, %r198;
add.f32 %f565, %f13, %f564;
sub.f32 %f566, %f14, %f564;
mul.f32 %f567, %f565, 0f3DD32618;
cvt.rmi.f32.f32 %f568, %f567;
sub.f32 %f569, %f567, %f568;
mul.f32 %f570, %f566, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f571, %f570;
sub.f32 %f572, %f570, %f571;
mul.f32 %f573, %f564, 0f3DC74539;
cvt.rmi.f32.f32 %f574, %f573;
sub.f32 %f575, %f573, %f574;
add.f32 %f576, %f572, 0f420551EC;
add.f32 %f577, %f569, 0f420551EC;
add.f32 %f578, %f575, 0f420551EC;
mul.f32 %f579, %f569, %f576;
fma.rn.f32 %f580, %f572, %f577, %f579;
fma.rn.f32 %f581, %f575, %f578, %f580;
add.f32 %f582, %f569, %f581;
add.f32 %f583, %f572, %f581;
add.f32 %f584, %f575, %f581;
add.f32 %f585, %f582, %f583;
mul.f32 %f586, %f584, %f585;
cvt.rmi.f32.f32 %f587, %f586;
sub.f32 %f588, %f586, %f587;
add.f32 %f589, %f582, %f582;
mul.f32 %f590, %f583, %f589;
cvt.rmi.f32.f32 %f591, %f590;
sub.f32 %f592, %f590, %f591;
mul.f32 %f593, %f582, %f585;
cvt.rmi.f32.f32 %f594, %f593;
sub.f32 %f595, %f593, %f594;
fma.rn.f32 %f596, %f588, 0f40000000, 0fBF800000;
fma.rn.f32 %f597, %f592, 0f40000000, 0fBF800000;
fma.rn.f32 %f598, %f595, 0f40000000, 0fBF800000;
fma.rn.f32 %f599, %f287, %f596, %f292;
fma.rn.f32 %f600, %f287, %f597, %f293;
fma.rn.f32 %f601, %f287, %f598, %f294;
sub.f32 %f602, %f599, %f7;
sub.f32 %f603, %f600, %f8;
sub.f32 %f604, %f601, %f9;
mul.f32 %f605, %f603, %f603;
fma.rn.f32 %f606, %f602, %f602, %f605;
fma.rn.f32 %f607, %f604, %f604, %f606;
sqrt.rn.f32 %f563, %f607;
rcp.rn.f32 %f608, %f563;
mul.f32 %f559, %f608, %f602;
mul.f32 %f560, %f608, %f603;
mul.f32 %f561, %f608, %f604;
ld.global.u32 %r92, [imageEnabled];
and.b32 %r93, %r92, 32;
setp.eq.s32 %p49, %r93, 0;
selp.f32 %f609, 0f3F800000, 0f41200000, %p49;
mul.f32 %f562, %f609, %f126;
mov.u32 %r94, 1065353216;
st.local.u32 [%rd2], %r94;
ld.global.u32 %r89, [root];
mov.u32 %r90, 1;
// inline asm
call _rt_trace_64, (%r89, %f10, %f11, %f12, %f559, %f560, %f561, %r90, %f562, %f563, %rd35, %r29);
// inline asm
ld.local.f32 %f610, [%rd2];
add.f32 %f1236, %f1233, %f610;
add.s32 %r200, %r198, 1;
BB0_40:
setp.lt.u32 %p50, %r8, 4;
@%p50 bra BB0_42;
BB0_41:
cvt.rn.f32.s32 %f643, %r200;
add.f32 %f644, %f13, %f643;
sub.f32 %f645, %f14, %f643;
mul.f32 %f646, %f644, 0f3DD32618;
cvt.rmi.f32.f32 %f647, %f646;
sub.f32 %f648, %f646, %f647;
mul.f32 %f649, %f645, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f650, %f649;
sub.f32 %f651, %f649, %f650;
mul.f32 %f652, %f643, 0f3DC74539;
cvt.rmi.f32.f32 %f653, %f652;
sub.f32 %f654, %f652, %f653;
add.f32 %f655, %f651, 0f420551EC;
add.f32 %f656, %f648, 0f420551EC;
add.f32 %f657, %f654, 0f420551EC;
mul.f32 %f658, %f648, %f655;
fma.rn.f32 %f659, %f651, %f656, %f658;
fma.rn.f32 %f660, %f654, %f657, %f659;
add.f32 %f661, %f648, %f660;
add.f32 %f662, %f651, %f660;
add.f32 %f663, %f654, %f660;
add.f32 %f664, %f661, %f662;
mul.f32 %f665, %f663, %f664;
cvt.rmi.f32.f32 %f666, %f665;
sub.f32 %f667, %f665, %f666;
add.f32 %f668, %f661, %f661;
mul.f32 %f669, %f662, %f668;
cvt.rmi.f32.f32 %f670, %f669;
sub.f32 %f671, %f669, %f670;
mul.f32 %f672, %f661, %f664;
cvt.rmi.f32.f32 %f673, %f672;
sub.f32 %f674, %f672, %f673;
fma.rn.f32 %f675, %f667, 0f40000000, 0fBF800000;
fma.rn.f32 %f676, %f671, 0f40000000, 0fBF800000;
fma.rn.f32 %f677, %f674, 0f40000000, 0fBF800000;
fma.rn.f32 %f678, %f287, %f675, %f292;
fma.rn.f32 %f679, %f287, %f676, %f293;
fma.rn.f32 %f680, %f287, %f677, %f294;
sub.f32 %f681, %f678, %f7;
sub.f32 %f682, %f679, %f8;
sub.f32 %f683, %f680, %f9;
mul.f32 %f684, %f682, %f682;
fma.rn.f32 %f685, %f681, %f681, %f684;
fma.rn.f32 %f686, %f683, %f683, %f685;
sqrt.rn.f32 %f618, %f686;
rcp.rn.f32 %f687, %f618;
mul.f32 %f614, %f687, %f681;
mul.f32 %f615, %f687, %f682;
mul.f32 %f616, %f687, %f683;
ld.global.u32 %r107, [imageEnabled];
and.b32 %r108, %r107, 32;
setp.eq.s32 %p51, %r108, 0;
selp.f32 %f688, 0f3F800000, 0f41200000, %p51;
mul.f32 %f617, %f688, %f126;
mov.u32 %r109, 1065353216;
st.local.u32 [%rd2], %r109;
ld.global.u32 %r95, [root];
mov.u32 %r105, 1;
// inline asm
call _rt_trace_64, (%r95, %f10, %f11, %f12, %f614, %f615, %f616, %r105, %f617, %f618, %rd35, %r29);
// inline asm
ld.local.f32 %f689, [%rd2];
add.f32 %f690, %f1236, %f689;
add.s32 %r110, %r200, 1;
cvt.rn.f32.s32 %f691, %r110;
add.f32 %f692, %f13, %f691;
sub.f32 %f693, %f14, %f691;
mul.f32 %f694, %f692, 0f3DD32618;
cvt.rmi.f32.f32 %f695, %f694;
sub.f32 %f696, %f694, %f695;
mul.f32 %f697, %f693, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f698, %f697;
sub.f32 %f699, %f697, %f698;
mul.f32 %f700, %f691, 0f3DC74539;
cvt.rmi.f32.f32 %f701, %f700;
sub.f32 %f702, %f700, %f701;
add.f32 %f703, %f699, 0f420551EC;
add.f32 %f704, %f696, 0f420551EC;
add.f32 %f705, %f702, 0f420551EC;
mul.f32 %f706, %f696, %f703;
fma.rn.f32 %f707, %f699, %f704, %f706;
fma.rn.f32 %f708, %f702, %f705, %f707;
add.f32 %f709, %f696, %f708;
add.f32 %f710, %f699, %f708;
add.f32 %f711, %f702, %f708;
add.f32 %f712, %f709, %f710;
mul.f32 %f713, %f711, %f712;
cvt.rmi.f32.f32 %f714, %f713;
sub.f32 %f715, %f713, %f714;
add.f32 %f716, %f709, %f709;
mul.f32 %f717, %f710, %f716;
cvt.rmi.f32.f32 %f718, %f717;
sub.f32 %f719, %f717, %f718;
mul.f32 %f720, %f709, %f712;
cvt.rmi.f32.f32 %f721, %f720;
sub.f32 %f722, %f720, %f721;
fma.rn.f32 %f723, %f715, 0f40000000, 0fBF800000;
fma.rn.f32 %f724, %f719, 0f40000000, 0fBF800000;
fma.rn.f32 %f725, %f722, 0f40000000, 0fBF800000;
fma.rn.f32 %f726, %f287, %f723, %f292;
fma.rn.f32 %f727, %f287, %f724, %f293;
fma.rn.f32 %f728, %f287, %f725, %f294;
sub.f32 %f729, %f726, %f7;
sub.f32 %f730, %f727, %f8;
sub.f32 %f731, %f728, %f9;
mul.f32 %f732, %f730, %f730;
fma.rn.f32 %f733, %f729, %f729, %f732;
fma.rn.f32 %f734, %f731, %f731, %f733;
sqrt.rn.f32 %f626, %f734;
rcp.rn.f32 %f735, %f626;
mul.f32 %f622, %f735, %f729;
mul.f32 %f623, %f735, %f730;
mul.f32 %f624, %f735, %f731;
ld.global.u32 %r111, [imageEnabled];
and.b32 %r112, %r111, 32;
setp.eq.s32 %p52, %r112, 0;
selp.f32 %f736, 0f3F800000, 0f41200000, %p52;
mul.f32 %f625, %f736, %f126;
st.local.u32 [%rd2], %r109;
ld.global.u32 %r98, [root];
// inline asm
call _rt_trace_64, (%r98, %f10, %f11, %f12, %f622, %f623, %f624, %r105, %f625, %f626, %rd35, %r29);
// inline asm
ld.local.f32 %f737, [%rd2];
add.f32 %f738, %f690, %f737;
add.s32 %r113, %r200, 2;
cvt.rn.f32.s32 %f739, %r113;
add.f32 %f740, %f13, %f739;
sub.f32 %f741, %f14, %f739;
mul.f32 %f742, %f740, 0f3DD32618;
cvt.rmi.f32.f32 %f743, %f742;
sub.f32 %f744, %f742, %f743;
mul.f32 %f745, %f741, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f746, %f745;
sub.f32 %f747, %f745, %f746;
mul.f32 %f748, %f739, 0f3DC74539;
cvt.rmi.f32.f32 %f749, %f748;
sub.f32 %f750, %f748, %f749;
add.f32 %f751, %f747, 0f420551EC;
add.f32 %f752, %f744, 0f420551EC;
add.f32 %f753, %f750, 0f420551EC;
mul.f32 %f754, %f744, %f751;
fma.rn.f32 %f755, %f747, %f752, %f754;
fma.rn.f32 %f756, %f750, %f753, %f755;
add.f32 %f757, %f744, %f756;
add.f32 %f758, %f747, %f756;
add.f32 %f759, %f750, %f756;
add.f32 %f760, %f757, %f758;
mul.f32 %f761, %f759, %f760;
cvt.rmi.f32.f32 %f762, %f761;
sub.f32 %f763, %f761, %f762;
add.f32 %f764, %f757, %f757;
mul.f32 %f765, %f758, %f764;
cvt.rmi.f32.f32 %f766, %f765;
sub.f32 %f767, %f765, %f766;
mul.f32 %f768, %f757, %f760;
cvt.rmi.f32.f32 %f769, %f768;
sub.f32 %f770, %f768, %f769;
fma.rn.f32 %f771, %f763, 0f40000000, 0fBF800000;
fma.rn.f32 %f772, %f767, 0f40000000, 0fBF800000;
fma.rn.f32 %f773, %f770, 0f40000000, 0fBF800000;
fma.rn.f32 %f774, %f287, %f771, %f292;
fma.rn.f32 %f775, %f287, %f772, %f293;
fma.rn.f32 %f776, %f287, %f773, %f294;
sub.f32 %f777, %f774, %f7;
sub.f32 %f778, %f775, %f8;
sub.f32 %f779, %f776, %f9;
mul.f32 %f780, %f778, %f778;
fma.rn.f32 %f781, %f777, %f777, %f780;
fma.rn.f32 %f782, %f779, %f779, %f781;
sqrt.rn.f32 %f634, %f782;
rcp.rn.f32 %f783, %f634;
mul.f32 %f630, %f783, %f777;
mul.f32 %f631, %f783, %f778;
mul.f32 %f632, %f783, %f779;
ld.global.u32 %r114, [imageEnabled];
and.b32 %r115, %r114, 32;
setp.eq.s32 %p53, %r115, 0;
selp.f32 %f784, 0f3F800000, 0f41200000, %p53;
mul.f32 %f633, %f784, %f126;
st.local.u32 [%rd2], %r109;
ld.global.u32 %r101, [root];
// inline asm
call _rt_trace_64, (%r101, %f10, %f11, %f12, %f630, %f631, %f632, %r105, %f633, %f634, %rd35, %r29);
// inline asm
ld.local.f32 %f785, [%rd2];
add.f32 %f786, %f738, %f785;
add.s32 %r116, %r200, 3;
cvt.rn.f32.s32 %f787, %r116;
add.f32 %f788, %f13, %f787;
sub.f32 %f789, %f14, %f787;
mul.f32 %f790, %f788, 0f3DD32618;
cvt.rmi.f32.f32 %f791, %f790;
sub.f32 %f792, %f790, %f791;
mul.f32 %f793, %f789, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f794, %f793;
sub.f32 %f795, %f793, %f794;
mul.f32 %f796, %f787, 0f3DC74539;
cvt.rmi.f32.f32 %f797, %f796;
sub.f32 %f798, %f796, %f797;
add.f32 %f799, %f795, 0f420551EC;
add.f32 %f800, %f792, 0f420551EC;
add.f32 %f801, %f798, 0f420551EC;
mul.f32 %f802, %f792, %f799;
fma.rn.f32 %f803, %f795, %f800, %f802;
fma.rn.f32 %f804, %f798, %f801, %f803;
add.f32 %f805, %f792, %f804;
add.f32 %f806, %f795, %f804;
add.f32 %f807, %f798, %f804;
add.f32 %f808, %f805, %f806;
mul.f32 %f809, %f807, %f808;
cvt.rmi.f32.f32 %f810, %f809;
sub.f32 %f811, %f809, %f810;
add.f32 %f812, %f805, %f805;
mul.f32 %f813, %f806, %f812;
cvt.rmi.f32.f32 %f814, %f813;
sub.f32 %f815, %f813, %f814;
mul.f32 %f816, %f805, %f808;
cvt.rmi.f32.f32 %f817, %f816;
sub.f32 %f818, %f816, %f817;
fma.rn.f32 %f819, %f811, 0f40000000, 0fBF800000;
fma.rn.f32 %f820, %f815, 0f40000000, 0fBF800000;
fma.rn.f32 %f821, %f818, 0f40000000, 0fBF800000;
fma.rn.f32 %f822, %f287, %f819, %f292;
fma.rn.f32 %f823, %f287, %f820, %f293;
fma.rn.f32 %f824, %f287, %f821, %f294;
sub.f32 %f825, %f822, %f7;
sub.f32 %f826, %f823, %f8;
sub.f32 %f827, %f824, %f9;
mul.f32 %f828, %f826, %f826;
fma.rn.f32 %f829, %f825, %f825, %f828;
fma.rn.f32 %f830, %f827, %f827, %f829;
sqrt.rn.f32 %f642, %f830;
rcp.rn.f32 %f831, %f642;
mul.f32 %f638, %f831, %f825;
mul.f32 %f639, %f831, %f826;
mul.f32 %f640, %f831, %f827;
ld.global.u32 %r117, [imageEnabled];
and.b32 %r118, %r117, 32;
setp.eq.s32 %p54, %r118, 0;
selp.f32 %f832, 0f3F800000, 0f41200000, %p54;
mul.f32 %f641, %f832, %f126;
st.local.u32 [%rd2], %r109;
ld.global.u32 %r104, [root];
// inline asm
call _rt_trace_64, (%r104, %f10, %f11, %f12, %f638, %f639, %f640, %r105, %f641, %f642, %rd35, %r29);
// inline asm
ld.local.f32 %f833, [%rd2];
add.f32 %f1236, %f786, %f833;
add.s32 %r200, %r200, 4;
setp.lt.s32 %p55, %r200, %r8;
@%p55 bra BB0_41;
BB0_42:
cvt.rn.f32.s32 %f834, %r8;
div.rn.f32 %f1237, %f1236, %f834;
shr.u32 %r119, %r6, 31;
cvt.u16.u32 %rs62, %r119;
BB0_43:
fma.rn.f32 %f1214, %f119, %f1237, %f1214;
fma.rn.f32 %f1213, %f120, %f1237, %f1213;
fma.rn.f32 %f1212, %f121, %f1237, %f1212;
ld.global.u8 %rs15, [imageEnabled];
and.b16 %rs16, %rs15, 64;
setp.eq.s16 %p56, %rs16, 0;
@%p56 bra BB0_45;
mul.f32 %f835, %f120, 0f3F372474;
fma.rn.f32 %f836, %f119, 0f3E59999A, %f835;
fma.rn.f32 %f837, %f121, 0f3D93A92A, %f836;
fma.rn.f32 %f24, %f122, %f837, %f24;
fma.rn.f32 %f23, %f123, %f837, %f23;
fma.rn.f32 %f22, %f837, %f124, %f22;
BB0_45:
setp.eq.s16 %p57, %rs62, 0;
@%p57 bra BB0_47;
div.rn.f32 %f838, %f119, %f283;
div.rn.f32 %f839, %f838, %f91;
cvt.sat.f32.f32 %f840, %f839;
mul.f32 %f1237, %f1237, %f840;
BB0_47:
add.f32 %f1208, %f1208, %f1237;
BB0_48:
add.s32 %r196, %r196, 1;
setp.lt.u32 %p58, %r196, %r4;
@%p58 bra BB0_5;
BB0_49:
ld.global.u32 %r202, [imageEnabled];
and.b32 %r120, %r202, 8;
setp.eq.s32 %p59, %r120, 0;
@%p59 bra BB0_62;
cvt.sat.f32.f32 %f164, %f1208;
cvt.u64.u32 %rd46, %r3;
cvt.u64.u32 %rd45, %r2;
mov.u64 %rd49, image_Mask;
cvta.global.u64 %rd44, %rd49;
// inline asm
call (%rd43), _rt_buffer_get_64, (%rd44, %r28, %r28, %rd45, %rd46, %rd13, %rd13);
// inline asm
mov.f32 %f843, 0f3E68BA2E;
cvt.rzi.f32.f32 %f844, %f843;
fma.rn.f32 %f845, %f844, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f165, %f845;
abs.f32 %f166, %f164;
setp.lt.f32 %p60, %f166, 0f00800000;
mul.f32 %f846, %f166, 0f4B800000;
selp.f32 %f847, 0fC3170000, 0fC2FE0000, %p60;
selp.f32 %f848, %f846, %f166, %p60;
mov.b32 %r123, %f848;
and.b32 %r124, %r123, 8388607;
or.b32 %r125, %r124, 1065353216;
mov.b32 %f849, %r125;
shr.u32 %r126, %r123, 23;
cvt.rn.f32.u32 %f850, %r126;
add.f32 %f851, %f847, %f850;
setp.gt.f32 %p61, %f849, 0f3FB504F3;
mul.f32 %f852, %f849, 0f3F000000;
add.f32 %f853, %f851, 0f3F800000;
selp.f32 %f854, %f852, %f849, %p61;
selp.f32 %f855, %f853, %f851, %p61;
add.f32 %f856, %f854, 0fBF800000;
add.f32 %f842, %f854, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f841,%f842;
// inline asm
add.f32 %f857, %f856, %f856;
mul.f32 %f858, %f841, %f857;
mul.f32 %f859, %f858, %f858;
mov.f32 %f860, 0f3C4CAF63;
mov.f32 %f861, 0f3B18F0FE;
fma.rn.f32 %f862, %f861, %f859, %f860;
mov.f32 %f863, 0f3DAAAABD;
fma.rn.f32 %f864, %f862, %f859, %f863;
mul.rn.f32 %f865, %f864, %f859;
mul.rn.f32 %f866, %f865, %f858;
sub.f32 %f867, %f856, %f858;
neg.f32 %f868, %f858;
add.f32 %f869, %f867, %f867;
fma.rn.f32 %f870, %f868, %f856, %f869;
mul.rn.f32 %f871, %f841, %f870;
add.f32 %f872, %f866, %f858;
sub.f32 %f873, %f858, %f872;
add.f32 %f874, %f866, %f873;
add.f32 %f875, %f871, %f874;
add.f32 %f876, %f872, %f875;
sub.f32 %f877, %f872, %f876;
add.f32 %f878, %f875, %f877;
mov.f32 %f879, 0f3F317200;
mul.rn.f32 %f880, %f855, %f879;
mov.f32 %f881, 0f35BFBE8E;
mul.rn.f32 %f882, %f855, %f881;
add.f32 %f883, %f880, %f876;
sub.f32 %f884, %f880, %f883;
add.f32 %f885, %f876, %f884;
add.f32 %f886, %f878, %f885;
add.f32 %f887, %f882, %f886;
add.f32 %f888, %f883, %f887;
sub.f32 %f889, %f883, %f888;
add.f32 %f890, %f887, %f889;
mov.f32 %f891, 0f3EE8BA2E;
mul.rn.f32 %f892, %f891, %f888;
neg.f32 %f893, %f892;
fma.rn.f32 %f894, %f891, %f888, %f893;
fma.rn.f32 %f895, %f891, %f890, %f894;
mov.f32 %f896, 0f00000000;
fma.rn.f32 %f897, %f896, %f888, %f895;
add.rn.f32 %f898, %f892, %f897;
neg.f32 %f899, %f898;
add.rn.f32 %f900, %f892, %f899;
add.rn.f32 %f901, %f900, %f897;
mov.b32 %r127, %f898;
setp.eq.s32 %p62, %r127, 1118925336;
add.s32 %r128, %r127, -1;
mov.b32 %f902, %r128;
add.f32 %f903, %f901, 0f37000000;
selp.f32 %f904, %f902, %f898, %p62;
selp.f32 %f167, %f903, %f901, %p62;
mul.f32 %f905, %f904, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f906, %f905;
mov.f32 %f907, 0fBF317200;
fma.rn.f32 %f908, %f906, %f907, %f904;
mov.f32 %f909, 0fB5BFBE8E;
fma.rn.f32 %f910, %f906, %f909, %f908;
mul.f32 %f911, %f910, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f912, %f911;
add.f32 %f913, %f906, 0f00000000;
ex2.approx.f32 %f914, %f913;
mul.f32 %f915, %f912, %f914;
setp.lt.f32 %p63, %f904, 0fC2D20000;
selp.f32 %f916, 0f00000000, %f915, %p63;
setp.gt.f32 %p64, %f904, 0f42D20000;
selp.f32 %f1256, 0f7F800000, %f916, %p64;
setp.eq.f32 %p65, %f1256, 0f7F800000;
@%p65 bra BB0_52;
fma.rn.f32 %f1256, %f1256, %f167, %f1256;
BB0_52:
setp.lt.f32 %p66, %f164, 0f00000000;
setp.eq.f32 %p67, %f165, 0f3F800000;
and.pred %p4, %p66, %p67;
mov.b32 %r129, %f1256;
xor.b32 %r130, %r129, -2147483648;
mov.b32 %f917, %r130;
selp.f32 %f1258, %f917, %f1256, %p4;
setp.eq.f32 %p68, %f164, 0f00000000;
@%p68 bra BB0_55;
bra.uni BB0_53;
BB0_55:
add.f32 %f920, %f164, %f164;
selp.f32 %f1258, %f920, 0f00000000, %p67;
bra.uni BB0_56;
BB0_110:
mov.u64 %rd117, image_HDR;
cvta.global.u64 %rd112, %rd117;
mov.u32 %r193, 8;
// inline asm
call (%rd111), _rt_buffer_get_64, (%rd112, %r28, %r193, %rd6, %rd7, %rd13, %rd13);
// inline asm
mov.f32 %f1201, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs57, %f1201;}
// inline asm
mov.u16 %rs58, 0;
st.v4.u16 [%rd111], {%rs57, %rs57, %rs57, %rs58};
BB0_111:
ld.global.u8 %rs59, [imageEnabled];
and.b16 %rs60, %rs59, 64;
setp.eq.s16 %p132, %rs60, 0;
@%p132 bra BB0_113;
cvt.u64.u32 %rd120, %r2;
cvt.u64.u32 %rd121, %r3;
mov.u64 %rd124, image_Dir;
cvta.global.u64 %rd119, %rd124;
// inline asm
call (%rd118), _rt_buffer_get_64, (%rd119, %r28, %r29, %rd120, %rd121, %rd13, %rd13);
// inline asm
mov.u16 %rs61, 0;
st.v4.u8 [%rd118], {%rs61, %rs61, %rs61, %rs61};
bra.uni BB0_113;
BB0_53:
setp.geu.f32 %p69, %f164, 0f00000000;
@%p69 bra BB0_56;
cvt.rzi.f32.f32 %f919, %f891;
setp.neu.f32 %p70, %f919, 0f3EE8BA2E;
selp.f32 %f1258, 0f7FFFFFFF, %f1258, %p70;
BB0_56:
add.f32 %f921, %f166, 0f3EE8BA2E;
mov.b32 %r131, %f921;
setp.lt.s32 %p72, %r131, 2139095040;
@%p72 bra BB0_61;
setp.gtu.f32 %p73, %f166, 0f7F800000;
@%p73 bra BB0_60;
bra.uni BB0_58;
BB0_60:
add.f32 %f1258, %f164, 0f3EE8BA2E;
bra.uni BB0_61;
BB0_58:
setp.neu.f32 %p74, %f166, 0f7F800000;
@%p74 bra BB0_61;
selp.f32 %f1258, 0fFF800000, 0f7F800000, %p4;
BB0_61:
mul.f32 %f922, %f1258, 0f437F0000;
setp.eq.f32 %p75, %f164, 0f3F800000;
selp.f32 %f923, 0f437F0000, %f922, %p75;
cvt.rzi.u32.f32 %r132, %f923;
cvt.u16.u32 %rs17, %r132;
mov.u16 %rs18, 255;
st.v2.u8 [%rd43], {%rs17, %rs18};
ld.global.u32 %r202, [imageEnabled];
BB0_62:
and.b32 %r133, %r202, 1;
setp.eq.b32 %p76, %r133, 1;
@!%p76 bra BB0_97;
bra.uni BB0_63;
BB0_63:
mov.f32 %f926, 0f3E666666;
cvt.rzi.f32.f32 %f927, %f926;
fma.rn.f32 %f928, %f927, 0fC0000000, 0f3EE66666;
abs.f32 %f178, %f928;
abs.f32 %f179, %f1214;
setp.lt.f32 %p77, %f179, 0f00800000;
mul.f32 %f929, %f179, 0f4B800000;
selp.f32 %f930, 0fC3170000, 0fC2FE0000, %p77;
selp.f32 %f931, %f929, %f179, %p77;
mov.b32 %r134, %f931;
and.b32 %r135, %r134, 8388607;
or.b32 %r136, %r135, 1065353216;
mov.b32 %f932, %r136;
shr.u32 %r137, %r134, 23;
cvt.rn.f32.u32 %f933, %r137;
add.f32 %f934, %f930, %f933;
setp.gt.f32 %p78, %f932, 0f3FB504F3;
mul.f32 %f935, %f932, 0f3F000000;
add.f32 %f936, %f934, 0f3F800000;
selp.f32 %f937, %f935, %f932, %p78;
selp.f32 %f938, %f936, %f934, %p78;
add.f32 %f939, %f937, 0fBF800000;
add.f32 %f925, %f937, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f924,%f925;
// inline asm
add.f32 %f940, %f939, %f939;
mul.f32 %f941, %f924, %f940;
mul.f32 %f942, %f941, %f941;
mov.f32 %f943, 0f3C4CAF63;
mov.f32 %f944, 0f3B18F0FE;
fma.rn.f32 %f945, %f944, %f942, %f943;
mov.f32 %f946, 0f3DAAAABD;
fma.rn.f32 %f947, %f945, %f942, %f946;
mul.rn.f32 %f948, %f947, %f942;
mul.rn.f32 %f949, %f948, %f941;
sub.f32 %f950, %f939, %f941;
neg.f32 %f951, %f941;
add.f32 %f952, %f950, %f950;
fma.rn.f32 %f953, %f951, %f939, %f952;
mul.rn.f32 %f954, %f924, %f953;
add.f32 %f955, %f949, %f941;
sub.f32 %f956, %f941, %f955;
add.f32 %f957, %f949, %f956;
add.f32 %f958, %f954, %f957;
add.f32 %f959, %f955, %f958;
sub.f32 %f960, %f955, %f959;
add.f32 %f961, %f958, %f960;
mov.f32 %f962, 0f3F317200;
mul.rn.f32 %f963, %f938, %f962;
mov.f32 %f964, 0f35BFBE8E;
mul.rn.f32 %f965, %f938, %f964;
add.f32 %f966, %f963, %f959;
sub.f32 %f967, %f963, %f966;
add.f32 %f968, %f959, %f967;
add.f32 %f969, %f961, %f968;
add.f32 %f970, %f965, %f969;
add.f32 %f971, %f966, %f970;
sub.f32 %f972, %f966, %f971;
add.f32 %f973, %f970, %f972;
mov.f32 %f974, 0f3EE66666;
mul.rn.f32 %f975, %f974, %f971;
neg.f32 %f976, %f975;
fma.rn.f32 %f977, %f974, %f971, %f976;
fma.rn.f32 %f978, %f974, %f973, %f977;
mov.f32 %f979, 0f00000000;
fma.rn.f32 %f980, %f979, %f971, %f978;
add.rn.f32 %f981, %f975, %f980;
neg.f32 %f982, %f981;
add.rn.f32 %f983, %f975, %f982;
add.rn.f32 %f984, %f983, %f980;
mov.b32 %r138, %f981;
setp.eq.s32 %p79, %r138, 1118925336;
add.s32 %r139, %r138, -1;
mov.b32 %f985, %r139;
add.f32 %f986, %f984, 0f37000000;
selp.f32 %f987, %f985, %f981, %p79;
selp.f32 %f180, %f986, %f984, %p79;
mul.f32 %f988, %f987, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f989, %f988;
mov.f32 %f990, 0fBF317200;
fma.rn.f32 %f991, %f989, %f990, %f987;
mov.f32 %f992, 0fB5BFBE8E;
fma.rn.f32 %f993, %f989, %f992, %f991;
mul.f32 %f994, %f993, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f995, %f994;
add.f32 %f996, %f989, 0f00000000;
ex2.approx.f32 %f997, %f996;
mul.f32 %f998, %f995, %f997;
setp.lt.f32 %p80, %f987, 0fC2D20000;
selp.f32 %f999, 0f00000000, %f998, %p80;
setp.gt.f32 %p81, %f987, 0f42D20000;
selp.f32 %f1259, 0f7F800000, %f999, %p81;
setp.eq.f32 %p82, %f1259, 0f7F800000;
@%p82 bra BB0_65;
fma.rn.f32 %f1259, %f1259, %f180, %f1259;
BB0_65:
setp.lt.f32 %p83, %f1214, 0f00000000;
setp.eq.f32 %p84, %f178, 0f3F800000;
and.pred %p5, %p83, %p84;
mov.b32 %r140, %f1259;
xor.b32 %r141, %r140, -2147483648;
mov.b32 %f1000, %r141;
selp.f32 %f1261, %f1000, %f1259, %p5;
setp.eq.f32 %p85, %f1214, 0f00000000;
@%p85 bra BB0_68;
bra.uni BB0_66;
BB0_68:
add.f32 %f1003, %f1214, %f1214;
selp.f32 %f1261, %f1003, 0f00000000, %p84;
bra.uni BB0_69;
BB0_66:
setp.geu.f32 %p86, %f1214, 0f00000000;
@%p86 bra BB0_69;
cvt.rzi.f32.f32 %f1002, %f974;
setp.neu.f32 %p87, %f1002, 0f3EE66666;
selp.f32 %f1261, 0f7FFFFFFF, %f1261, %p87;
BB0_69:
add.f32 %f1004, %f179, 0f3EE66666;
mov.b32 %r142, %f1004;
setp.lt.s32 %p89, %r142, 2139095040;
@%p89 bra BB0_74;
setp.gtu.f32 %p90, %f179, 0f7F800000;
@%p90 bra BB0_73;
bra.uni BB0_71;
BB0_73:
add.f32 %f1261, %f1214, 0f3EE66666;
bra.uni BB0_74;
BB0_71:
setp.neu.f32 %p91, %f179, 0f7F800000;
@%p91 bra BB0_74;
selp.f32 %f1261, 0fFF800000, 0f7F800000, %p5;
BB0_74:
setp.eq.f32 %p92, %f1214, 0f3F800000;
selp.f32 %f191, 0f3F800000, %f1261, %p92;
abs.f32 %f192, %f1213;
setp.lt.f32 %p93, %f192, 0f00800000;
mul.f32 %f1007, %f192, 0f4B800000;
selp.f32 %f1008, 0fC3170000, 0fC2FE0000, %p93;
selp.f32 %f1009, %f1007, %f192, %p93;
mov.b32 %r143, %f1009;
and.b32 %r144, %r143, 8388607;
or.b32 %r145, %r144, 1065353216;
mov.b32 %f1010, %r145;
shr.u32 %r146, %r143, 23;
cvt.rn.f32.u32 %f1011, %r146;
add.f32 %f1012, %f1008, %f1011;
setp.gt.f32 %p94, %f1010, 0f3FB504F3;
mul.f32 %f1013, %f1010, 0f3F000000;
add.f32 %f1014, %f1012, 0f3F800000;
selp.f32 %f1015, %f1013, %f1010, %p94;
selp.f32 %f1016, %f1014, %f1012, %p94;
add.f32 %f1017, %f1015, 0fBF800000;
add.f32 %f1006, %f1015, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1005,%f1006;
// inline asm
add.f32 %f1018, %f1017, %f1017;
mul.f32 %f1019, %f1005, %f1018;
mul.f32 %f1020, %f1019, %f1019;
fma.rn.f32 %f1023, %f944, %f1020, %f943;
fma.rn.f32 %f1025, %f1023, %f1020, %f946;
mul.rn.f32 %f1026, %f1025, %f1020;
mul.rn.f32 %f1027, %f1026, %f1019;
sub.f32 %f1028, %f1017, %f1019;
neg.f32 %f1029, %f1019;
add.f32 %f1030, %f1028, %f1028;
fma.rn.f32 %f1031, %f1029, %f1017, %f1030;
mul.rn.f32 %f1032, %f1005, %f1031;
add.f32 %f1033, %f1027, %f1019;
sub.f32 %f1034, %f1019, %f1033;
add.f32 %f1035, %f1027, %f1034;
add.f32 %f1036, %f1032, %f1035;
add.f32 %f1037, %f1033, %f1036;
sub.f32 %f1038, %f1033, %f1037;
add.f32 %f1039, %f1036, %f1038;
mul.rn.f32 %f1041, %f1016, %f962;
mul.rn.f32 %f1043, %f1016, %f964;
add.f32 %f1044, %f1041, %f1037;
sub.f32 %f1045, %f1041, %f1044;
add.f32 %f1046, %f1037, %f1045;
add.f32 %f1047, %f1039, %f1046;
add.f32 %f1048, %f1043, %f1047;
add.f32 %f1049, %f1044, %f1048;
sub.f32 %f1050, %f1044, %f1049;
add.f32 %f1051, %f1048, %f1050;
mul.rn.f32 %f1053, %f974, %f1049;
neg.f32 %f1054, %f1053;
fma.rn.f32 %f1055, %f974, %f1049, %f1054;
fma.rn.f32 %f1056, %f974, %f1051, %f1055;
fma.rn.f32 %f1058, %f979, %f1049, %f1056;
add.rn.f32 %f1059, %f1053, %f1058;
neg.f32 %f1060, %f1059;
add.rn.f32 %f1061, %f1053, %f1060;
add.rn.f32 %f1062, %f1061, %f1058;
mov.b32 %r147, %f1059;
setp.eq.s32 %p95, %r147, 1118925336;
add.s32 %r148, %r147, -1;
mov.b32 %f1063, %r148;
add.f32 %f1064, %f1062, 0f37000000;
selp.f32 %f1065, %f1063, %f1059, %p95;
selp.f32 %f193, %f1064, %f1062, %p95;
mul.f32 %f1066, %f1065, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1067, %f1066;
fma.rn.f32 %f1069, %f1067, %f990, %f1065;
fma.rn.f32 %f1071, %f1067, %f992, %f1069;
mul.f32 %f1072, %f1071, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1073, %f1072;
add.f32 %f1074, %f1067, 0f00000000;
ex2.approx.f32 %f1075, %f1074;
mul.f32 %f1076, %f1073, %f1075;
setp.lt.f32 %p96, %f1065, 0fC2D20000;
selp.f32 %f1077, 0f00000000, %f1076, %p96;
setp.gt.f32 %p97, %f1065, 0f42D20000;
selp.f32 %f1262, 0f7F800000, %f1077, %p97;
setp.eq.f32 %p98, %f1262, 0f7F800000;
@%p98 bra BB0_76;
fma.rn.f32 %f1262, %f1262, %f193, %f1262;
BB0_76:
setp.lt.f32 %p99, %f1213, 0f00000000;
and.pred %p6, %p99, %p84;
mov.b32 %r149, %f1262;
xor.b32 %r150, %r149, -2147483648;
mov.b32 %f1078, %r150;
selp.f32 %f1264, %f1078, %f1262, %p6;
setp.eq.f32 %p101, %f1213, 0f00000000;
@%p101 bra BB0_79;
bra.uni BB0_77;
BB0_79:
add.f32 %f1081, %f1213, %f1213;
selp.f32 %f1264, %f1081, 0f00000000, %p84;
bra.uni BB0_80;
BB0_77:
setp.geu.f32 %p102, %f1213, 0f00000000;
@%p102 bra BB0_80;
cvt.rzi.f32.f32 %f1080, %f974;
setp.neu.f32 %p103, %f1080, 0f3EE66666;
selp.f32 %f1264, 0f7FFFFFFF, %f1264, %p103;
BB0_80:
add.f32 %f1082, %f192, 0f3EE66666;
mov.b32 %r151, %f1082;
setp.lt.s32 %p105, %r151, 2139095040;
@%p105 bra BB0_85;
setp.gtu.f32 %p106, %f192, 0f7F800000;
@%p106 bra BB0_84;
bra.uni BB0_82;
BB0_84:
add.f32 %f1264, %f1213, 0f3EE66666;
bra.uni BB0_85;
BB0_82:
setp.neu.f32 %p107, %f192, 0f7F800000;
@%p107 bra BB0_85;
selp.f32 %f1264, 0fFF800000, 0f7F800000, %p6;
BB0_85:
setp.eq.f32 %p108, %f1213, 0f3F800000;
selp.f32 %f204, 0f3F800000, %f1264, %p108;
abs.f32 %f205, %f1212;
setp.lt.f32 %p109, %f205, 0f00800000;
mul.f32 %f1085, %f205, 0f4B800000;
selp.f32 %f1086, 0fC3170000, 0fC2FE0000, %p109;
selp.f32 %f1087, %f1085, %f205, %p109;
mov.b32 %r152, %f1087;
and.b32 %r153, %r152, 8388607;
or.b32 %r154, %r153, 1065353216;
mov.b32 %f1088, %r154;
shr.u32 %r155, %r152, 23;
cvt.rn.f32.u32 %f1089, %r155;
add.f32 %f1090, %f1086, %f1089;
setp.gt.f32 %p110, %f1088, 0f3FB504F3;
mul.f32 %f1091, %f1088, 0f3F000000;
add.f32 %f1092, %f1090, 0f3F800000;
selp.f32 %f1093, %f1091, %f1088, %p110;
selp.f32 %f1094, %f1092, %f1090, %p110;
add.f32 %f1095, %f1093, 0fBF800000;
add.f32 %f1084, %f1093, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1083,%f1084;
// inline asm
add.f32 %f1096, %f1095, %f1095;
mul.f32 %f1097, %f1083, %f1096;
mul.f32 %f1098, %f1097, %f1097;
fma.rn.f32 %f1101, %f944, %f1098, %f943;
fma.rn.f32 %f1103, %f1101, %f1098, %f946;
mul.rn.f32 %f1104, %f1103, %f1098;
mul.rn.f32 %f1105, %f1104, %f1097;
sub.f32 %f1106, %f1095, %f1097;
neg.f32 %f1107, %f1097;
add.f32 %f1108, %f1106, %f1106;
fma.rn.f32 %f1109, %f1107, %f1095, %f1108;
mul.rn.f32 %f1110, %f1083, %f1109;
add.f32 %f1111, %f1105, %f1097;
sub.f32 %f1112, %f1097, %f1111;
add.f32 %f1113, %f1105, %f1112;
add.f32 %f1114, %f1110, %f1113;
add.f32 %f1115, %f1111, %f1114;
sub.f32 %f1116, %f1111, %f1115;
add.f32 %f1117, %f1114, %f1116;
mul.rn.f32 %f1119, %f1094, %f962;
mul.rn.f32 %f1121, %f1094, %f964;
add.f32 %f1122, %f1119, %f1115;
sub.f32 %f1123, %f1119, %f1122;
add.f32 %f1124, %f1115, %f1123;
add.f32 %f1125, %f1117, %f1124;
add.f32 %f1126, %f1121, %f1125;
add.f32 %f1127, %f1122, %f1126;
sub.f32 %f1128, %f1122, %f1127;
add.f32 %f1129, %f1126, %f1128;
mul.rn.f32 %f1131, %f974, %f1127;
neg.f32 %f1132, %f1131;
fma.rn.f32 %f1133, %f974, %f1127, %f1132;
fma.rn.f32 %f1134, %f974, %f1129, %f1133;
fma.rn.f32 %f1136, %f979, %f1127, %f1134;
add.rn.f32 %f1137, %f1131, %f1136;
neg.f32 %f1138, %f1137;
add.rn.f32 %f1139, %f1131, %f1138;
add.rn.f32 %f1140, %f1139, %f1136;
mov.b32 %r156, %f1137;
setp.eq.s32 %p111, %r156, 1118925336;
add.s32 %r157, %r156, -1;
mov.b32 %f1141, %r157;
add.f32 %f1142, %f1140, 0f37000000;
selp.f32 %f1143, %f1141, %f1137, %p111;
selp.f32 %f206, %f1142, %f1140, %p111;
mul.f32 %f1144, %f1143, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1145, %f1144;
fma.rn.f32 %f1147, %f1145, %f990, %f1143;
fma.rn.f32 %f1149, %f1145, %f992, %f1147;
mul.f32 %f1150, %f1149, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1151, %f1150;
add.f32 %f1152, %f1145, 0f00000000;
ex2.approx.f32 %f1153, %f1152;
mul.f32 %f1154, %f1151, %f1153;
setp.lt.f32 %p112, %f1143, 0fC2D20000;
selp.f32 %f1155, 0f00000000, %f1154, %p112;
setp.gt.f32 %p113, %f1143, 0f42D20000;
selp.f32 %f1265, 0f7F800000, %f1155, %p113;
setp.eq.f32 %p114, %f1265, 0f7F800000;
@%p114 bra BB0_87;
fma.rn.f32 %f1265, %f1265, %f206, %f1265;
BB0_87:
setp.lt.f32 %p115, %f1212, 0f00000000;
and.pred %p7, %p115, %p84;
mov.b32 %r158, %f1265;
xor.b32 %r159, %r158, -2147483648;
mov.b32 %f1156, %r159;
selp.f32 %f1267, %f1156, %f1265, %p7;
setp.eq.f32 %p117, %f1212, 0f00000000;
@%p117 bra BB0_90;
bra.uni BB0_88;
BB0_90:
add.f32 %f1159, %f1212, %f1212;
selp.f32 %f1267, %f1159, 0f00000000, %p84;
bra.uni BB0_91;
BB0_88:
setp.geu.f32 %p118, %f1212, 0f00000000;
@%p118 bra BB0_91;
cvt.rzi.f32.f32 %f1158, %f974;
setp.neu.f32 %p119, %f1158, 0f3EE66666;
selp.f32 %f1267, 0f7FFFFFFF, %f1267, %p119;
BB0_91:
add.f32 %f1160, %f205, 0f3EE66666;
mov.b32 %r160, %f1160;
setp.lt.s32 %p121, %r160, 2139095040;
@%p121 bra BB0_96;
setp.gtu.f32 %p122, %f205, 0f7F800000;
@%p122 bra BB0_95;
bra.uni BB0_93;
BB0_95:
add.f32 %f1267, %f1212, 0f3EE66666;
bra.uni BB0_96;
BB0_93:
setp.neu.f32 %p123, %f205, 0f7F800000;
@%p123 bra BB0_96;
selp.f32 %f1267, 0fFF800000, 0f7F800000, %p7;
BB0_96:
setp.eq.f32 %p124, %f1212, 0f3F800000;
selp.f32 %f1161, 0f3F800000, %f1267, %p124;
cvt.u64.u32 %rd53, %r3;
cvt.u64.u32 %rd52, %r2;
mov.u64 %rd56, image;
cvta.global.u64 %rd51, %rd56;
// inline asm
call (%rd50), _rt_buffer_get_64, (%rd51, %r28, %r29, %rd52, %rd53, %rd13, %rd13);
// inline asm
cvt.sat.f32.f32 %f1162, %f1161;
mul.f32 %f1163, %f1162, 0f437FFD71;
cvt.rzi.u32.f32 %r163, %f1163;
cvt.sat.f32.f32 %f1164, %f204;
mul.f32 %f1165, %f1164, 0f437FFD71;
cvt.rzi.u32.f32 %r164, %f1165;
cvt.sat.f32.f32 %f1166, %f191;
mul.f32 %f1167, %f1166, 0f437FFD71;
cvt.rzi.u32.f32 %r165, %f1167;
cvt.u16.u32 %rs19, %r163;
cvt.u16.u32 %rs20, %r165;
cvt.u16.u32 %rs21, %r164;
mov.u16 %rs22, 255;
st.v4.u8 [%rd50], {%rs19, %rs21, %rs20, %rs22};
ld.global.u32 %r202, [imageEnabled];
BB0_97:
and.b32 %r166, %r202, 4;
setp.eq.s32 %p125, %r166, 0;
@%p125 bra BB0_101;
ld.global.u32 %r167, [additive];
setp.eq.s32 %p126, %r167, 0;
cvt.u64.u32 %rd4, %r2;
cvt.u64.u32 %rd5, %r3;
mov.f32 %f1168, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs23, %f1168;}
// inline asm
@%p126 bra BB0_100;
mov.u64 %rd69, image_HDR;
cvta.global.u64 %rd58, %rd69;
mov.u32 %r171, 8;
// inline asm
call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r171, %rd4, %rd5, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd57];
// inline asm
{ cvt.f32.f16 %f1169, %rs30;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1170, %rs31;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1171, %rs32;}
// inline asm
// inline asm
call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r171, %rd4, %rd5, %rd13, %rd13);
// inline asm
add.f32 %f1172, %f1214, %f1169;
add.f32 %f1173, %f1213, %f1170;
add.f32 %f1174, %f1212, %f1171;
// inline asm
{ cvt.rn.f16.f32 %rs29, %f1174;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs28, %f1173;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs27, %f1172;}
// inline asm
st.v4.u16 [%rd63], {%rs27, %rs28, %rs29, %rs23};
bra.uni BB0_101;
BB0_100:
mov.u64 %rd76, image_HDR;
cvta.global.u64 %rd71, %rd76;
mov.u32 %r173, 8;
// inline asm
call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r173, %rd4, %rd5, %rd13, %rd13);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs36, %f1212;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs35, %f1213;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs34, %f1214;}
// inline asm
st.v4.u16 [%rd70], {%rs34, %rs35, %rs36, %rs23};
BB0_101:
ld.global.u8 %rs37, [imageEnabled];
and.b16 %rs38, %rs37, 64;
setp.eq.s16 %p127, %rs38, 0;
@%p127 bra BB0_113;
mul.f32 %f1178, %f23, %f23;
fma.rn.f32 %f1179, %f24, %f24, %f1178;
fma.rn.f32 %f1180, %f22, %f22, %f1179;
sqrt.rn.f32 %f1181, %f1180;
rcp.rn.f32 %f1182, %f1181;
mul.f32 %f1183, %f24, %f1182;
mul.f32 %f1184, %f23, %f1182;
mul.f32 %f1185, %f22, %f1182;
cvt.u64.u32 %rd80, %r3;
cvt.u64.u32 %rd79, %r2;
mov.u64 %rd83, image_Dir;
cvta.global.u64 %rd78, %rd83;
// inline asm
call (%rd77), _rt_buffer_get_64, (%rd78, %r28, %r29, %rd79, %rd80, %rd13, %rd13);
// inline asm
fma.rn.f32 %f1186, %f1183, 0f3F000000, 0f3F000000;
mul.f32 %f1187, %f1186, 0f437F0000;
cvt.rzi.u32.f32 %r176, %f1187;
fma.rn.f32 %f1188, %f1184, 0f3F000000, 0f3F000000;
mul.f32 %f1189, %f1188, 0f437F0000;
cvt.rzi.u32.f32 %r177, %f1189;
fma.rn.f32 %f1190, %f1185, 0f3F000000, 0f3F000000;
mul.f32 %f1191, %f1190, 0f437F0000;
cvt.rzi.u32.f32 %r178, %f1191;
cvt.u16.u32 %rs39, %r178;
cvt.u16.u32 %rs40, %r177;
cvt.u16.u32 %rs41, %r176;
mov.u16 %rs42, 255;
st.v4.u8 [%rd77], {%rs41, %rs40, %rs39, %rs42};
BB0_113:
ret;
}