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

2691 lines
78 KiB
Plaintext

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-23083092
// Cuda compilation tools, release 9.1, V9.1.85
// Based on LLVM 3.4svn
//
.version 6.1
.target sm_30
.address_size 64
// .globl _Z6oxMainv
.global .align 8 .b8 pixelID[8];
.global .align 8 .b8 resolution[8];
.global .align 4 .b8 normal[12];
.global .align 4 .b8 camPos[12];
.global .align 4 .b8 root[4];
.global .align 4 .u32 imageEnabled;
.global .texref lightmap;
.global .align 16 .b8 tileInfo[16];
.global .align 4 .u32 additive;
.global .align 1 .b8 image[1];
.global .align 1 .b8 image_HDR[1];
.global .align 1 .b8 image_HDR2[1];
.global .align 1 .b8 image_Mask[1];
.global .align 1 .b8 image_RNM0[1];
.global .align 1 .b8 image_RNM1[1];
.global .align 1 .b8 image_RNM2[1];
.global .align 1 .b8 image_RNM3[1];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 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[12];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<166>;
.reg .b16 %rs<162>;
.reg .f32 %f<1624>;
.reg .b32 %r<270>;
.reg .b64 %rd<272>;
mov.u64 %rd271, __local_depot0;
cvta.local.u64 %SP, %rd271;
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 %rs9, %rs1, 255;
cvt.u16.u32 %rs10, %r1;
or.b16 %rs11, %rs10, %rs9;
setp.eq.s16 %p9, %rs11, 0;
mov.f32 %f1513, 0f00000000;
mov.f32 %f1514, %f1513;
mov.f32 %f1515, %f1513;
@%p9 bra BB0_2;
ld.u8 %rs12, [%rd8+1];
and.b16 %rs14, %rs10, 255;
cvt.rn.f32.u16 %f317, %rs14;
div.rn.f32 %f318, %f317, 0f437F0000;
fma.rn.f32 %f319, %f318, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f320, %rs12;
div.rn.f32 %f321, %f320, 0f437F0000;
fma.rn.f32 %f322, %f321, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f323, %rs9;
div.rn.f32 %f324, %f323, 0f437F0000;
fma.rn.f32 %f325, %f324, 0f40000000, 0fBF800000;
mul.f32 %f326, %f322, %f322;
fma.rn.f32 %f327, %f319, %f319, %f326;
fma.rn.f32 %f328, %f325, %f325, %f327;
sqrt.rn.f32 %f329, %f328;
rcp.rn.f32 %f330, %f329;
mul.f32 %f1513, %f319, %f330;
mul.f32 %f1514, %f322, %f330;
mul.f32 %f1515, %f325, %f330;
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 %p10, %f1514, 0f00000000;
setp.eq.f32 %p11, %f1513, 0f00000000;
and.pred %p12, %p11, %p10;
setp.eq.f32 %p13, %f1515, 0f00000000;
and.pred %p14, %p12, %p13;
@%p14 bra BB0_124;
bra.uni BB0_3;
BB0_124:
ld.global.u32 %r269, [imageEnabled];
and.b32 %r217, %r269, 1;
setp.eq.b32 %p157, %r217, 1;
@!%p157 bra BB0_126;
bra.uni BB0_125;
BB0_125:
cvt.u64.u32 %rd159, %r2;
cvt.u64.u32 %rd160, %r3;
mov.u64 %rd163, image;
cvta.global.u64 %rd158, %rd163;
// inline asm
call (%rd157), _rt_buffer_get_64, (%rd158, %r28, %r29, %rd159, %rd160, %rd13, %rd13);
// inline asm
mov.u16 %rs93, 0;
st.v4.u8 [%rd157], {%rs93, %rs93, %rs93, %rs93};
ld.global.u32 %r269, [imageEnabled];
BB0_126:
and.b32 %r220, %r269, 8;
setp.eq.s32 %p158, %r220, 0;
@%p158 bra BB0_128;
cvt.u64.u32 %rd167, %r3;
cvt.u64.u32 %rd166, %r2;
mov.u64 %rd170, image_Mask;
cvta.global.u64 %rd165, %rd170;
// inline asm
call (%rd164), _rt_buffer_get_64, (%rd165, %r28, %r28, %rd166, %rd167, %rd13, %rd13);
// inline asm
mov.f32 %f1467, 0f00000000;
cvt.rzi.u32.f32 %r223, %f1467;
cvt.u16.u32 %rs94, %r223;
mov.u16 %rs95, 0;
st.v2.u8 [%rd164], {%rs94, %rs95};
ld.global.u32 %r269, [imageEnabled];
BB0_128:
cvt.u64.u32 %rd6, %r2;
cvt.u64.u32 %rd7, %r3;
and.b32 %r224, %r269, 4;
setp.eq.s32 %p159, %r224, 0;
@%p159 bra BB0_132;
ld.global.u32 %r225, [additive];
setp.eq.s32 %p160, %r225, 0;
@%p160 bra BB0_131;
mov.u64 %rd183, image_HDR;
cvta.global.u64 %rd172, %rd183;
mov.u32 %r229, 8;
// inline asm
call (%rd171), _rt_buffer_get_64, (%rd172, %r28, %r229, %rd6, %rd7, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd171];
// inline asm
{ cvt.f32.f16 %f1468, %rs102;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1469, %rs103;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1470, %rs104;}
// inline asm
// inline asm
call (%rd177), _rt_buffer_get_64, (%rd172, %r28, %r229, %rd6, %rd7, %rd13, %rd13);
// inline asm
add.f32 %f1471, %f1468, 0f00000000;
add.f32 %f1472, %f1469, 0f00000000;
add.f32 %f1473, %f1470, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs101, %f1473;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs100, %f1472;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs99, %f1471;}
// inline asm
mov.u16 %rs106, 0;
st.v4.u16 [%rd177], {%rs99, %rs100, %rs101, %rs106};
bra.uni BB0_132;
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 %f347, %f7, 0f3456BF95;
mul.f32 %f348, %f8, 0f3456BF95;
mul.f32 %f349, %f9, 0f3456BF95;
abs.f32 %f350, %f1513;
div.rn.f32 %f351, %f347, %f350;
abs.f32 %f352, %f1514;
div.rn.f32 %f353, %f348, %f352;
abs.f32 %f354, %f1515;
div.rn.f32 %f355, %f349, %f354;
abs.f32 %f356, %f351;
abs.f32 %f357, %f353;
abs.f32 %f358, %f355;
mov.f32 %f359, 0f38D1B717;
max.f32 %f360, %f356, %f359;
max.f32 %f361, %f357, %f359;
max.f32 %f362, %f358, %f359;
fma.rn.f32 %f10, %f1513, %f360, %f7;
fma.rn.f32 %f11, %f1514, %f361, %f8;
fma.rn.f32 %f12, %f1515, %f362, %f9;
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 %p15, %r4, 0;
mov.f32 %f1516, 0f00000000;
mov.f32 %f1517, %f1516;
mov.f32 %f1518, %f1516;
mov.f32 %f1519, %f1516;
mov.f32 %f1520, %f1516;
mov.f32 %f1521, %f1516;
mov.f32 %f1522, %f1516;
mov.f32 %f1523, %f1516;
mov.f32 %f1524, %f1516;
mov.f32 %f1525, %f1516;
mov.f32 %f1526, %f1516;
mov.f32 %f1527, %f1516;
mov.f32 %f1528, %f1516;
mov.f32 %f1529, %f1516;
mov.f32 %f1530, %f1516;
mov.f32 %f1531, %f1516;
@%p15 bra BB0_60;
mov.f32 %f379, 0f40000000;
cvt.rzi.f32.f32 %f380, %f379;
add.f32 %f381, %f380, %f380;
mov.f32 %f382, 0f40800000;
sub.f32 %f383, %f382, %f381;
abs.f32 %f13, %f383;
mul.f32 %f14, %f10, 0f3456BF95;
mul.f32 %f15, %f11, 0f3456BF95;
mul.f32 %f16, %f12, 0f3456BF95;
mov.f32 %f378, 0f00000000;
mov.u32 %r261, 0;
abs.f32 %f583, %f14;
abs.f32 %f584, %f15;
max.f32 %f585, %f583, %f584;
abs.f32 %f586, %f16;
max.f32 %f587, %f585, %f586;
mov.f32 %f1516, %f378;
mov.f32 %f1517, %f378;
mov.f32 %f1518, %f378;
mov.f32 %f1519, %f378;
mov.f32 %f1520, %f378;
mov.f32 %f1521, %f378;
mov.f32 %f1522, %f378;
mov.f32 %f1523, %f378;
mov.f32 %f1524, %f378;
mov.f32 %f1525, %f378;
mov.f32 %f1526, %f378;
mov.f32 %f1527, %f378;
mov.f32 %f1528, %f378;
mov.f32 %f1529, %f378;
mov.f32 %f1530, %f378;
mov.f32 %f1531, %f378;
BB0_5:
cvt.u64.u32 %rd30, %r261;
// inline asm
call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
// inline asm
ld.v4.f32 {%f386, %f387, %f388, %f389}, [%rd28+80];
ld.v4.f32 {%f390, %f391, %f392, %f393}, [%rd28+64];
ld.v4.f32 {%f394, %f395, %f396, %f397}, [%rd28+48];
ld.v4.f32 {%f398, %f1540, %f1541, %f401}, [%rd28+32];
ld.v4.f32 {%f402, %f403, %f404, %f405}, [%rd28+16];
ld.v4.f32 {%f406, %f407, %f408, %f409}, [%rd28];
mov.b32 %r6, %f389;
sub.f32 %f57, %f407, %f7;
sub.f32 %f59, %f408, %f8;
sub.f32 %f61, %f409, %f9;
mul.f32 %f411, %f59, %f59;
fma.rn.f32 %f412, %f57, %f57, %f411;
fma.rn.f32 %f413, %f61, %f61, %f412;
sqrt.rn.f32 %f62, %f413;
rcp.rn.f32 %f414, %f62;
mul.f32 %f63, %f57, %f414;
mul.f32 %f64, %f59, %f414;
mul.f32 %f65, %f61, %f414;
mul.f32 %f66, %f62, %f405;
abs.f32 %f67, %f66;
setp.lt.f32 %p16, %f67, 0f00800000;
mul.f32 %f415, %f67, 0f4B800000;
selp.f32 %f416, 0fC3170000, 0fC2FE0000, %p16;
selp.f32 %f417, %f415, %f67, %p16;
mov.b32 %r54, %f417;
and.b32 %r55, %r54, 8388607;
or.b32 %r56, %r55, 1065353216;
mov.b32 %f418, %r56;
shr.u32 %r57, %r54, 23;
cvt.rn.f32.u32 %f419, %r57;
add.f32 %f420, %f416, %f419;
setp.gt.f32 %p17, %f418, 0f3FB504F3;
mul.f32 %f421, %f418, 0f3F000000;
add.f32 %f422, %f420, 0f3F800000;
selp.f32 %f423, %f421, %f418, %p17;
selp.f32 %f424, %f422, %f420, %p17;
add.f32 %f425, %f423, 0fBF800000;
add.f32 %f385, %f423, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f384,%f385;
// inline asm
add.f32 %f426, %f425, %f425;
mul.f32 %f427, %f384, %f426;
mul.f32 %f428, %f427, %f427;
mov.f32 %f429, 0f3C4CAF63;
mov.f32 %f430, 0f3B18F0FE;
fma.rn.f32 %f431, %f430, %f428, %f429;
mov.f32 %f432, 0f3DAAAABD;
fma.rn.f32 %f433, %f431, %f428, %f432;
mul.rn.f32 %f434, %f433, %f428;
mul.rn.f32 %f435, %f434, %f427;
sub.f32 %f436, %f425, %f427;
neg.f32 %f437, %f427;
add.f32 %f438, %f436, %f436;
fma.rn.f32 %f439, %f437, %f425, %f438;
mul.rn.f32 %f440, %f384, %f439;
add.f32 %f441, %f435, %f427;
sub.f32 %f442, %f427, %f441;
add.f32 %f443, %f435, %f442;
add.f32 %f444, %f440, %f443;
add.f32 %f445, %f441, %f444;
sub.f32 %f446, %f441, %f445;
add.f32 %f447, %f444, %f446;
mov.f32 %f448, 0f3F317200;
mul.rn.f32 %f449, %f424, %f448;
mov.f32 %f450, 0f35BFBE8E;
mul.rn.f32 %f451, %f424, %f450;
add.f32 %f452, %f449, %f445;
sub.f32 %f453, %f449, %f452;
add.f32 %f454, %f445, %f453;
add.f32 %f455, %f447, %f454;
add.f32 %f456, %f451, %f455;
add.f32 %f457, %f452, %f456;
sub.f32 %f458, %f452, %f457;
add.f32 %f459, %f456, %f458;
mul.rn.f32 %f68, %f382, %f457;
neg.f32 %f461, %f68;
fma.rn.f32 %f462, %f382, %f457, %f461;
fma.rn.f32 %f463, %f382, %f459, %f462;
fma.rn.f32 %f69, %f378, %f457, %f463;
add.rn.f32 %f70, %f68, %f69;
mov.b32 %r58, %f70;
setp.eq.s32 %p1, %r58, 1118925336;
add.s32 %r59, %r58, -1;
mov.b32 %f465, %r59;
selp.f32 %f466, %f465, %f70, %p1;
mul.f32 %f467, %f466, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f468, %f467;
mov.f32 %f469, 0fBF317200;
fma.rn.f32 %f470, %f468, %f469, %f466;
mov.f32 %f471, 0fB5BFBE8E;
fma.rn.f32 %f472, %f468, %f471, %f470;
mul.f32 %f473, %f472, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f474, %f473;
add.f32 %f475, %f468, 0f00000000;
ex2.approx.f32 %f476, %f475;
mul.f32 %f477, %f474, %f476;
setp.lt.f32 %p18, %f466, 0fC2D20000;
selp.f32 %f478, 0f00000000, %f477, %p18;
setp.gt.f32 %p19, %f466, 0f42D20000;
selp.f32 %f1532, 0f7F800000, %f478, %p19;
setp.eq.f32 %p20, %f1532, 0f7F800000;
@%p20 bra BB0_7;
neg.f32 %f479, %f70;
add.rn.f32 %f480, %f68, %f479;
add.rn.f32 %f481, %f480, %f69;
add.f32 %f482, %f481, 0f37000000;
selp.f32 %f483, %f482, %f481, %p1;
fma.rn.f32 %f1532, %f1532, %f483, %f1532;
BB0_7:
setp.lt.f32 %p21, %f66, 0f00000000;
setp.eq.f32 %p22, %f13, 0f3F800000;
and.pred %p2, %p21, %p22;
mov.b32 %r60, %f1532;
xor.b32 %r61, %r60, -2147483648;
mov.b32 %f484, %r61;
selp.f32 %f1534, %f484, %f1532, %p2;
setp.eq.f32 %p23, %f66, 0f00000000;
@%p23 bra BB0_10;
bra.uni BB0_8;
BB0_10:
add.f32 %f487, %f66, %f66;
selp.f32 %f1534, %f487, 0f00000000, %p22;
bra.uni BB0_11;
BB0_8:
setp.geu.f32 %p24, %f66, 0f00000000;
@%p24 bra BB0_11;
cvt.rzi.f32.f32 %f486, %f382;
setp.neu.f32 %p25, %f486, 0f40800000;
selp.f32 %f1534, 0f7FFFFFFF, %f1534, %p25;
BB0_11:
add.f32 %f488, %f67, 0f40800000;
mov.b32 %r62, %f488;
setp.lt.s32 %p27, %r62, 2139095040;
@%p27 bra BB0_16;
setp.gtu.f32 %p28, %f67, 0f7F800000;
@%p28 bra BB0_15;
bra.uni BB0_13;
BB0_15:
add.f32 %f1534, %f66, 0f40800000;
bra.uni BB0_16;
BB0_13:
setp.neu.f32 %p29, %f67, 0f7F800000;
@%p29 bra BB0_16;
selp.f32 %f1534, 0fFF800000, 0f7F800000, %p2;
BB0_16:
mul.f32 %f489, %f62, %f403;
mov.f32 %f1573, 0f3F800000;
sub.f32 %f491, %f1573, %f1534;
setp.eq.f32 %p30, %f66, 0f3F800000;
selp.f32 %f492, 0f00000000, %f491, %p30;
cvt.sat.f32.f32 %f493, %f492;
fma.rn.f32 %f494, %f489, %f489, %f404;
div.rn.f32 %f96, %f493, %f494;
mul.f32 %f495, %f1514, %f64;
fma.rn.f32 %f496, %f1513, %f63, %f495;
fma.rn.f32 %f497, %f1515, %f65, %f496;
ld.global.u32 %r63, [ignoreNormal];
setp.eq.s32 %p31, %r63, 0;
selp.f32 %f97, %f497, 0f3F800000, %p31;
mul.f32 %f498, %f97, 0f40800000;
cvt.sat.f32.f32 %f98, %f498;
setp.eq.f32 %p32, %f406, 0f3F800000;
@%p32 bra BB0_23;
bra.uni BB0_17;
BB0_23:
setp.leu.f32 %p36, %f401, 0f00000000;
@%p36 bra BB0_19;
setp.geu.f32 %p37, %f397, 0f00000000;
@%p37 bra BB0_25;
sub.f32 %f1512, %f409, %f9;
sub.f32 %f1511, %f407, %f7;
sub.f32 %f1510, %f408, %f8;
mul.f32 %f521, %f387, %f1510;
fma.rn.f32 %f522, %f386, %f1511, %f521;
fma.rn.f32 %f523, %f388, %f1512, %f522;
rcp.rn.f32 %f524, %f523;
mul.f32 %f1535, %f1511, %f524;
mul.f32 %f1536, %f1510, %f524;
mul.f32 %f1537, %f1512, %f524;
neg.f32 %f397, %f397;
bra.uni BB0_27;
BB0_17:
setp.eq.f32 %p33, %f406, 0f40000000;
@%p33 bra BB0_21;
bra.uni BB0_18;
BB0_21:
setp.leu.f32 %p35, %f401, 0f00000000;
@%p35 bra BB0_19;
mul.f32 %f515, %f395, %f64;
fma.rn.f32 %f516, %f394, %f63, %f515;
mul.f32 %f517, %f391, %f64;
fma.rn.f32 %f518, %f390, %f63, %f517;
mul.f32 %f519, %f387, %f64;
fma.rn.f32 %f520, %f386, %f63, %f519;
fma.rn.f32 %f512, %f396, %f65, %f516;
fma.rn.f32 %f513, %f392, %f65, %f518;
fma.rn.f32 %f514, %f388, %f65, %f520;
cvt.rzi.s32.f32 %r64, %f401;
mov.u32 %r65, 6;
mov.u32 %r66, 0;
// inline asm
call (%f508, %f509, %f510, %f511), _rt_texture_get_base_id, (%r64, %r65, %f512, %f513, %f514, %r66);
// inline asm
mul.f32 %f1539, %f398, %f508;
mul.f32 %f1540, %f1540, %f509;
mul.f32 %f1541, %f1541, %f510;
bra.uni BB0_28;
BB0_18:
setp.neu.f32 %p34, %f406, 0f40800000;
@%p34 bra BB0_19;
mul.f32 %f499, %f386, %f63;
mul.f32 %f500, %f387, %f64;
neg.f32 %f501, %f500;
sub.f32 %f502, %f501, %f499;
mul.f32 %f503, %f388, %f65;
sub.f32 %f504, %f502, %f503;
fma.rn.f32 %f505, %f401, %f504, %f397;
cvt.sat.f32.f32 %f506, %f505;
mul.f32 %f507, %f506, %f506;
mul.f32 %f1542, %f96, %f507;
mov.f32 %f1539, %f398;
bra.uni BB0_29;
BB0_19:
mov.f32 %f1539, %f398;
bra.uni BB0_28;
BB0_25:
mov.f32 %f1535, %f63;
mov.f32 %f1536, %f64;
mov.f32 %f1537, %f65;
BB0_27:
mul.f32 %f533, %f388, %f65;
mul.f32 %f534, %f387, %f64;
neg.f32 %f535, %f534;
mul.f32 %f536, %f386, %f63;
sub.f32 %f537, %f535, %f536;
sub.f32 %f538, %f537, %f533;
setp.gt.f32 %p38, %f538, 0f00000000;
selp.f32 %f539, 0f3F800000, 0f00000000, %p38;
mul.f32 %f540, %f395, %f1536;
fma.rn.f32 %f541, %f394, %f1535, %f540;
mul.f32 %f542, %f391, %f1536;
fma.rn.f32 %f543, %f390, %f1535, %f542;
fma.rn.f32 %f544, %f396, %f1537, %f541;
fma.rn.f32 %f545, %f392, %f1537, %f543;
fma.rn.f32 %f546, %f397, %f544, 0f3F000000;
mov.f32 %f547, 0f3F800000;
sub.f32 %f529, %f547, %f546;
fma.rn.f32 %f530, %f397, %f545, 0f3F000000;
cvt.rzi.s32.f32 %r67, %f401;
mov.f32 %f532, 0f00000000;
// inline asm
call (%f525, %f526, %f527, %f528), _rt_texture_get_f_id, (%r67, %r28, %f529, %f530, %f532, %f532);
// inline asm
mul.f32 %f548, %f539, %f525;
mul.f32 %f549, %f539, %f526;
mul.f32 %f550, %f539, %f527;
mul.f32 %f1539, %f398, %f548;
mul.f32 %f1540, %f1540, %f549;
mul.f32 %f1541, %f1541, %f550;
BB0_28:
mov.f32 %f1542, %f96;
BB0_29:
max.f32 %f566, %f1539, %f1540;
max.f32 %f567, %f566, %f1541;
mul.f32 %f568, %f98, %f1542;
mul.f32 %f569, %f568, %f567;
setp.lt.f32 %p40, %f569, 0f3727C5AC;
mov.pred %p165, -1;
mov.f32 %f1543, 0f00000000;
mov.f32 %f1544, %f1543;
mov.f32 %f1545, %f1543;
mov.f32 %f136, %f1543;
mov.f32 %f1547, %f1543;
mov.f32 %f1548, %f1543;
mov.f32 %f139, %f1543;
mov.f32 %f1550, %f1543;
mov.f32 %f1551, %f1543;
mov.f32 %f142, %f1543;
mov.f32 %f1553, %f1543;
mov.f32 %f1554, %f1543;
mov.f32 %f145, %f1543;
mov.f32 %f1556, %f1543;
mov.f32 %f1557, %f1543;
@%p40 bra BB0_31;
cvt.sat.f32.f32 %f570, %f97;
mul.f32 %f571, %f1542, %f570;
mul.f32 %f1543, %f1539, %f571;
mul.f32 %f1544, %f1540, %f571;
mul.f32 %f1545, %f1541, %f571;
mul.f32 %f572, %f1542, 0f3E800000;
mul.f32 %f573, %f98, %f572;
mul.f32 %f136, %f1539, %f573;
mul.f32 %f1547, %f1540, %f573;
mul.f32 %f1548, %f1541, %f573;
mul.f32 %f139, %f63, %f136;
mul.f32 %f1550, %f63, %f1547;
mul.f32 %f1551, %f63, %f1548;
mul.f32 %f142, %f64, %f136;
mul.f32 %f1553, %f64, %f1547;
mul.f32 %f1554, %f64, %f1548;
mul.f32 %f145, %f65, %f136;
mul.f32 %f1556, %f65, %f1547;
mul.f32 %f1557, %f65, %f1548;
mov.pred %p165, 0;
BB0_31:
@%p165 bra BB0_59;
setp.eq.s32 %p42, %r6, 0;
mov.u16 %rs161, 0;
mov.f32 %f1574, %f1573;
mov.f32 %f1575, %f1573;
@%p42 bra BB0_43;
abs.s32 %r8, %r6;
mov.f32 %f1570, 0f00000000;
setp.lt.s32 %p43, %r8, 1;
mov.f32 %f1571, %f1570;
mov.f32 %f1572, %f1570;
@%p43 bra BB0_42;
max.f32 %f149, %f587, %f359;
and.b32 %r9, %r8, 3;
setp.eq.s32 %p44, %r9, 0;
add.u64 %rd35, %SP, 0;
cvta.to.local.u64 %rd2, %rd35;
mov.f32 %f1570, 0f00000000;
mov.u32 %r265, 0;
mov.f32 %f1571, %f1570;
mov.f32 %f1572, %f1570;
@%p44 bra BB0_40;
setp.eq.s32 %p45, %r9, 1;
mov.f32 %f1561, 0f00000000;
mov.u32 %r263, 0;
mov.f32 %f1562, %f1561;
mov.f32 %f1563, %f1561;
@%p45 bra BB0_39;
setp.eq.s32 %p46, %r9, 2;
mov.f32 %f1558, 0f00000000;
mov.u32 %r262, 0;
mov.f32 %f1559, %f1558;
mov.f32 %f1560, %f1558;
@%p46 bra BB0_38;
sub.f32 %f603, %f407, %f402;
sub.f32 %f604, %f408, %f402;
sub.f32 %f605, %f409, %f402;
sub.f32 %f606, %f603, %f7;
sub.f32 %f607, %f604, %f8;
sub.f32 %f608, %f605, %f9;
mul.f32 %f609, %f607, %f607;
fma.rn.f32 %f610, %f606, %f606, %f609;
fma.rn.f32 %f611, %f608, %f608, %f610;
sqrt.rn.f32 %f602, %f611;
rcp.rn.f32 %f612, %f602;
mul.f32 %f598, %f612, %f606;
mul.f32 %f599, %f612, %f607;
mul.f32 %f600, %f612, %f608;
ld.global.u32 %r76, [imageEnabled];
and.b32 %r77, %r76, 32;
setp.eq.s32 %p47, %r77, 0;
selp.f32 %f613, 0f3F800000, 0f41200000, %p47;
mul.f32 %f601, %f613, %f149;
mov.u32 %r78, 1065353216;
st.local.u32 [%rd2+8], %r78;
st.local.u32 [%rd2+4], %r78;
st.local.u32 [%rd2], %r78;
ld.global.u32 %r72, [root];
// inline asm
call _rt_trace_64, (%r72, %f10, %f11, %f12, %f598, %f599, %f600, %r45, %f601, %f602, %rd35, %r44);
// inline asm
ld.local.f32 %f614, [%rd2];
add.f32 %f1560, %f614, 0f00000000;
ld.local.f32 %f615, [%rd2+4];
add.f32 %f1559, %f615, 0f00000000;
ld.local.f32 %f616, [%rd2+8];
add.f32 %f1558, %f616, 0f00000000;
mov.u32 %r262, %r45;
BB0_38:
cvt.rn.f32.s32 %f625, %r262;
mul.f32 %f626, %f625, 0f3DD32618;
cvt.rmi.f32.f32 %f627, %f626;
sub.f32 %f628, %f626, %f627;
mul.f32 %f629, %f625, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f630, %f629;
sub.f32 %f631, %f629, %f630;
mul.f32 %f632, %f625, 0f3DC74539;
cvt.rmi.f32.f32 %f633, %f632;
sub.f32 %f634, %f632, %f633;
add.f32 %f635, %f631, 0f4199851F;
add.f32 %f636, %f634, 0f4199851F;
add.f32 %f637, %f628, 0f4199851F;
mul.f32 %f638, %f631, %f636;
fma.rn.f32 %f639, %f628, %f635, %f638;
fma.rn.f32 %f640, %f637, %f634, %f639;
add.f32 %f641, %f628, %f640;
add.f32 %f642, %f631, %f640;
add.f32 %f643, %f634, %f640;
add.f32 %f644, %f641, %f642;
mul.f32 %f645, %f643, %f644;
cvt.rmi.f32.f32 %f646, %f645;
sub.f32 %f647, %f645, %f646;
add.f32 %f648, %f641, %f643;
mul.f32 %f649, %f642, %f648;
cvt.rmi.f32.f32 %f650, %f649;
sub.f32 %f651, %f649, %f650;
add.f32 %f652, %f642, %f643;
mul.f32 %f653, %f641, %f652;
cvt.rmi.f32.f32 %f654, %f653;
sub.f32 %f655, %f653, %f654;
fma.rn.f32 %f656, %f647, 0f40000000, 0fBF800000;
fma.rn.f32 %f657, %f651, 0f40000000, 0fBF800000;
fma.rn.f32 %f658, %f655, 0f40000000, 0fBF800000;
fma.rn.f32 %f659, %f402, %f656, %f407;
fma.rn.f32 %f660, %f402, %f657, %f408;
fma.rn.f32 %f661, %f402, %f658, %f409;
sub.f32 %f662, %f659, %f7;
sub.f32 %f663, %f660, %f8;
sub.f32 %f664, %f661, %f9;
mul.f32 %f665, %f663, %f663;
fma.rn.f32 %f666, %f662, %f662, %f665;
fma.rn.f32 %f667, %f664, %f664, %f666;
sqrt.rn.f32 %f624, %f667;
rcp.rn.f32 %f668, %f624;
mul.f32 %f620, %f668, %f662;
mul.f32 %f621, %f668, %f663;
mul.f32 %f622, %f668, %f664;
ld.global.u32 %r82, [imageEnabled];
and.b32 %r83, %r82, 32;
setp.eq.s32 %p48, %r83, 0;
selp.f32 %f669, 0f3F800000, 0f41200000, %p48;
mul.f32 %f623, %f669, %f149;
mov.u32 %r84, 1065353216;
st.local.u32 [%rd2+8], %r84;
st.local.u32 [%rd2+4], %r84;
st.local.u32 [%rd2], %r84;
ld.global.u32 %r79, [root];
// inline asm
call _rt_trace_64, (%r79, %f10, %f11, %f12, %f620, %f621, %f622, %r45, %f623, %f624, %rd35, %r44);
// inline asm
ld.local.f32 %f670, [%rd2];
add.f32 %f1563, %f1560, %f670;
ld.local.f32 %f671, [%rd2+4];
add.f32 %f1562, %f1559, %f671;
ld.local.f32 %f672, [%rd2+8];
add.f32 %f1561, %f1558, %f672;
add.s32 %r263, %r262, 1;
BB0_39:
cvt.rn.f32.s32 %f681, %r263;
mul.f32 %f682, %f681, 0f3DD32618;
cvt.rmi.f32.f32 %f683, %f682;
sub.f32 %f684, %f682, %f683;
mul.f32 %f685, %f681, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f686, %f685;
sub.f32 %f687, %f685, %f686;
mul.f32 %f688, %f681, 0f3DC74539;
cvt.rmi.f32.f32 %f689, %f688;
sub.f32 %f690, %f688, %f689;
add.f32 %f691, %f687, 0f4199851F;
add.f32 %f692, %f690, 0f4199851F;
add.f32 %f693, %f684, 0f4199851F;
mul.f32 %f694, %f687, %f692;
fma.rn.f32 %f695, %f684, %f691, %f694;
fma.rn.f32 %f696, %f693, %f690, %f695;
add.f32 %f697, %f684, %f696;
add.f32 %f698, %f687, %f696;
add.f32 %f699, %f690, %f696;
add.f32 %f700, %f697, %f698;
mul.f32 %f701, %f699, %f700;
cvt.rmi.f32.f32 %f702, %f701;
sub.f32 %f703, %f701, %f702;
add.f32 %f704, %f697, %f699;
mul.f32 %f705, %f698, %f704;
cvt.rmi.f32.f32 %f706, %f705;
sub.f32 %f707, %f705, %f706;
add.f32 %f708, %f698, %f699;
mul.f32 %f709, %f697, %f708;
cvt.rmi.f32.f32 %f710, %f709;
sub.f32 %f711, %f709, %f710;
fma.rn.f32 %f712, %f703, 0f40000000, 0fBF800000;
fma.rn.f32 %f713, %f707, 0f40000000, 0fBF800000;
fma.rn.f32 %f714, %f711, 0f40000000, 0fBF800000;
fma.rn.f32 %f715, %f402, %f712, %f407;
fma.rn.f32 %f716, %f402, %f713, %f408;
fma.rn.f32 %f717, %f402, %f714, %f409;
sub.f32 %f718, %f715, %f7;
sub.f32 %f719, %f716, %f8;
sub.f32 %f720, %f717, %f9;
mul.f32 %f721, %f719, %f719;
fma.rn.f32 %f722, %f718, %f718, %f721;
fma.rn.f32 %f723, %f720, %f720, %f722;
sqrt.rn.f32 %f680, %f723;
rcp.rn.f32 %f724, %f680;
mul.f32 %f676, %f724, %f718;
mul.f32 %f677, %f724, %f719;
mul.f32 %f678, %f724, %f720;
ld.global.u32 %r88, [imageEnabled];
and.b32 %r89, %r88, 32;
setp.eq.s32 %p49, %r89, 0;
selp.f32 %f725, 0f3F800000, 0f41200000, %p49;
mul.f32 %f679, %f725, %f149;
mov.u32 %r90, 1065353216;
st.local.u32 [%rd2+8], %r90;
st.local.u32 [%rd2+4], %r90;
st.local.u32 [%rd2], %r90;
ld.global.u32 %r85, [root];
mov.u32 %r86, 1;
// inline asm
call _rt_trace_64, (%r85, %f10, %f11, %f12, %f676, %f677, %f678, %r86, %f679, %f680, %rd35, %r44);
// inline asm
ld.local.f32 %f726, [%rd2];
add.f32 %f1572, %f1563, %f726;
ld.local.f32 %f727, [%rd2+4];
add.f32 %f1571, %f1562, %f727;
ld.local.f32 %f728, [%rd2+8];
add.f32 %f1570, %f1561, %f728;
add.s32 %r265, %r263, 1;
BB0_40:
setp.lt.u32 %p50, %r8, 4;
@%p50 bra BB0_42;
BB0_41:
cvt.rn.f32.s32 %f761, %r265;
mul.f32 %f762, %f761, 0f3DD32618;
cvt.rmi.f32.f32 %f763, %f762;
sub.f32 %f764, %f762, %f763;
mul.f32 %f765, %f761, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f766, %f765;
sub.f32 %f767, %f765, %f766;
mul.f32 %f768, %f761, 0f3DC74539;
cvt.rmi.f32.f32 %f769, %f768;
sub.f32 %f770, %f768, %f769;
add.f32 %f771, %f767, 0f4199851F;
add.f32 %f772, %f770, 0f4199851F;
add.f32 %f773, %f764, 0f4199851F;
mul.f32 %f774, %f767, %f772;
fma.rn.f32 %f775, %f764, %f771, %f774;
fma.rn.f32 %f776, %f773, %f770, %f775;
add.f32 %f777, %f764, %f776;
add.f32 %f778, %f767, %f776;
add.f32 %f779, %f770, %f776;
add.f32 %f780, %f777, %f778;
mul.f32 %f781, %f779, %f780;
cvt.rmi.f32.f32 %f782, %f781;
sub.f32 %f783, %f781, %f782;
add.f32 %f784, %f777, %f779;
mul.f32 %f785, %f778, %f784;
cvt.rmi.f32.f32 %f786, %f785;
sub.f32 %f787, %f785, %f786;
add.f32 %f788, %f778, %f779;
mul.f32 %f789, %f777, %f788;
cvt.rmi.f32.f32 %f790, %f789;
sub.f32 %f791, %f789, %f790;
fma.rn.f32 %f792, %f783, 0f40000000, 0fBF800000;
fma.rn.f32 %f793, %f787, 0f40000000, 0fBF800000;
fma.rn.f32 %f794, %f791, 0f40000000, 0fBF800000;
fma.rn.f32 %f795, %f402, %f792, %f407;
fma.rn.f32 %f796, %f402, %f793, %f408;
fma.rn.f32 %f797, %f402, %f794, %f409;
sub.f32 %f798, %f795, %f7;
sub.f32 %f799, %f796, %f8;
sub.f32 %f800, %f797, %f9;
mul.f32 %f801, %f799, %f799;
fma.rn.f32 %f802, %f798, %f798, %f801;
fma.rn.f32 %f803, %f800, %f800, %f802;
sqrt.rn.f32 %f736, %f803;
rcp.rn.f32 %f804, %f736;
mul.f32 %f732, %f804, %f798;
mul.f32 %f733, %f804, %f799;
mul.f32 %f734, %f804, %f800;
ld.global.u32 %r103, [imageEnabled];
and.b32 %r104, %r103, 32;
setp.eq.s32 %p51, %r104, 0;
selp.f32 %f805, 0f3F800000, 0f41200000, %p51;
mul.f32 %f735, %f805, %f149;
mov.u32 %r105, 1065353216;
st.local.u32 [%rd2+8], %r105;
st.local.u32 [%rd2+4], %r105;
st.local.u32 [%rd2], %r105;
ld.global.u32 %r91, [root];
mov.u32 %r101, 1;
// inline asm
call _rt_trace_64, (%r91, %f10, %f11, %f12, %f732, %f733, %f734, %r101, %f735, %f736, %rd35, %r44);
// inline asm
ld.local.f32 %f806, [%rd2];
add.f32 %f807, %f1572, %f806;
ld.local.f32 %f808, [%rd2+4];
add.f32 %f809, %f1571, %f808;
ld.local.f32 %f810, [%rd2+8];
add.f32 %f811, %f1570, %f810;
add.s32 %r106, %r265, 1;
cvt.rn.f32.s32 %f812, %r106;
mul.f32 %f813, %f812, 0f3DD32618;
cvt.rmi.f32.f32 %f814, %f813;
sub.f32 %f815, %f813, %f814;
mul.f32 %f816, %f812, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f817, %f816;
sub.f32 %f818, %f816, %f817;
mul.f32 %f819, %f812, 0f3DC74539;
cvt.rmi.f32.f32 %f820, %f819;
sub.f32 %f821, %f819, %f820;
add.f32 %f822, %f818, 0f4199851F;
add.f32 %f823, %f821, 0f4199851F;
add.f32 %f824, %f815, 0f4199851F;
mul.f32 %f825, %f818, %f823;
fma.rn.f32 %f826, %f815, %f822, %f825;
fma.rn.f32 %f827, %f824, %f821, %f826;
add.f32 %f828, %f815, %f827;
add.f32 %f829, %f818, %f827;
add.f32 %f830, %f821, %f827;
add.f32 %f831, %f828, %f829;
mul.f32 %f832, %f830, %f831;
cvt.rmi.f32.f32 %f833, %f832;
sub.f32 %f834, %f832, %f833;
add.f32 %f835, %f828, %f830;
mul.f32 %f836, %f829, %f835;
cvt.rmi.f32.f32 %f837, %f836;
sub.f32 %f838, %f836, %f837;
add.f32 %f839, %f829, %f830;
mul.f32 %f840, %f828, %f839;
cvt.rmi.f32.f32 %f841, %f840;
sub.f32 %f842, %f840, %f841;
fma.rn.f32 %f843, %f834, 0f40000000, 0fBF800000;
fma.rn.f32 %f844, %f838, 0f40000000, 0fBF800000;
fma.rn.f32 %f845, %f842, 0f40000000, 0fBF800000;
fma.rn.f32 %f846, %f402, %f843, %f407;
fma.rn.f32 %f847, %f402, %f844, %f408;
fma.rn.f32 %f848, %f402, %f845, %f409;
sub.f32 %f849, %f846, %f7;
sub.f32 %f850, %f847, %f8;
sub.f32 %f851, %f848, %f9;
mul.f32 %f852, %f850, %f850;
fma.rn.f32 %f853, %f849, %f849, %f852;
fma.rn.f32 %f854, %f851, %f851, %f853;
sqrt.rn.f32 %f744, %f854;
rcp.rn.f32 %f855, %f744;
mul.f32 %f740, %f855, %f849;
mul.f32 %f741, %f855, %f850;
mul.f32 %f742, %f855, %f851;
ld.global.u32 %r107, [imageEnabled];
and.b32 %r108, %r107, 32;
setp.eq.s32 %p52, %r108, 0;
selp.f32 %f856, 0f3F800000, 0f41200000, %p52;
mul.f32 %f743, %f856, %f149;
st.local.u32 [%rd2+8], %r105;
st.local.u32 [%rd2+4], %r105;
st.local.u32 [%rd2], %r105;
ld.global.u32 %r94, [root];
// inline asm
call _rt_trace_64, (%r94, %f10, %f11, %f12, %f740, %f741, %f742, %r101, %f743, %f744, %rd35, %r44);
// inline asm
ld.local.f32 %f857, [%rd2];
add.f32 %f858, %f807, %f857;
ld.local.f32 %f859, [%rd2+4];
add.f32 %f860, %f809, %f859;
ld.local.f32 %f861, [%rd2+8];
add.f32 %f862, %f811, %f861;
add.s32 %r109, %r265, 2;
cvt.rn.f32.s32 %f863, %r109;
mul.f32 %f864, %f863, 0f3DD32618;
cvt.rmi.f32.f32 %f865, %f864;
sub.f32 %f866, %f864, %f865;
mul.f32 %f867, %f863, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f868, %f867;
sub.f32 %f869, %f867, %f868;
mul.f32 %f870, %f863, 0f3DC74539;
cvt.rmi.f32.f32 %f871, %f870;
sub.f32 %f872, %f870, %f871;
add.f32 %f873, %f869, 0f4199851F;
add.f32 %f874, %f872, 0f4199851F;
add.f32 %f875, %f866, 0f4199851F;
mul.f32 %f876, %f869, %f874;
fma.rn.f32 %f877, %f866, %f873, %f876;
fma.rn.f32 %f878, %f875, %f872, %f877;
add.f32 %f879, %f866, %f878;
add.f32 %f880, %f869, %f878;
add.f32 %f881, %f872, %f878;
add.f32 %f882, %f879, %f880;
mul.f32 %f883, %f881, %f882;
cvt.rmi.f32.f32 %f884, %f883;
sub.f32 %f885, %f883, %f884;
add.f32 %f886, %f879, %f881;
mul.f32 %f887, %f880, %f886;
cvt.rmi.f32.f32 %f888, %f887;
sub.f32 %f889, %f887, %f888;
add.f32 %f890, %f880, %f881;
mul.f32 %f891, %f879, %f890;
cvt.rmi.f32.f32 %f892, %f891;
sub.f32 %f893, %f891, %f892;
fma.rn.f32 %f894, %f885, 0f40000000, 0fBF800000;
fma.rn.f32 %f895, %f889, 0f40000000, 0fBF800000;
fma.rn.f32 %f896, %f893, 0f40000000, 0fBF800000;
fma.rn.f32 %f897, %f402, %f894, %f407;
fma.rn.f32 %f898, %f402, %f895, %f408;
fma.rn.f32 %f899, %f402, %f896, %f409;
sub.f32 %f900, %f897, %f7;
sub.f32 %f901, %f898, %f8;
sub.f32 %f902, %f899, %f9;
mul.f32 %f903, %f901, %f901;
fma.rn.f32 %f904, %f900, %f900, %f903;
fma.rn.f32 %f905, %f902, %f902, %f904;
sqrt.rn.f32 %f752, %f905;
rcp.rn.f32 %f906, %f752;
mul.f32 %f748, %f906, %f900;
mul.f32 %f749, %f906, %f901;
mul.f32 %f750, %f906, %f902;
ld.global.u32 %r110, [imageEnabled];
and.b32 %r111, %r110, 32;
setp.eq.s32 %p53, %r111, 0;
selp.f32 %f907, 0f3F800000, 0f41200000, %p53;
mul.f32 %f751, %f907, %f149;
st.local.u32 [%rd2+8], %r105;
st.local.u32 [%rd2+4], %r105;
st.local.u32 [%rd2], %r105;
ld.global.u32 %r97, [root];
// inline asm
call _rt_trace_64, (%r97, %f10, %f11, %f12, %f748, %f749, %f750, %r101, %f751, %f752, %rd35, %r44);
// inline asm
ld.local.f32 %f908, [%rd2];
add.f32 %f909, %f858, %f908;
ld.local.f32 %f910, [%rd2+4];
add.f32 %f911, %f860, %f910;
ld.local.f32 %f912, [%rd2+8];
add.f32 %f913, %f862, %f912;
add.s32 %r112, %r265, 3;
cvt.rn.f32.s32 %f914, %r112;
mul.f32 %f915, %f914, 0f3DD32618;
cvt.rmi.f32.f32 %f916, %f915;
sub.f32 %f917, %f915, %f916;
mul.f32 %f918, %f914, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f919, %f918;
sub.f32 %f920, %f918, %f919;
mul.f32 %f921, %f914, 0f3DC74539;
cvt.rmi.f32.f32 %f922, %f921;
sub.f32 %f923, %f921, %f922;
add.f32 %f924, %f920, 0f4199851F;
add.f32 %f925, %f923, 0f4199851F;
add.f32 %f926, %f917, 0f4199851F;
mul.f32 %f927, %f920, %f925;
fma.rn.f32 %f928, %f917, %f924, %f927;
fma.rn.f32 %f929, %f926, %f923, %f928;
add.f32 %f930, %f917, %f929;
add.f32 %f931, %f920, %f929;
add.f32 %f932, %f923, %f929;
add.f32 %f933, %f930, %f931;
mul.f32 %f934, %f932, %f933;
cvt.rmi.f32.f32 %f935, %f934;
sub.f32 %f936, %f934, %f935;
add.f32 %f937, %f930, %f932;
mul.f32 %f938, %f931, %f937;
cvt.rmi.f32.f32 %f939, %f938;
sub.f32 %f940, %f938, %f939;
add.f32 %f941, %f931, %f932;
mul.f32 %f942, %f930, %f941;
cvt.rmi.f32.f32 %f943, %f942;
sub.f32 %f944, %f942, %f943;
fma.rn.f32 %f945, %f936, 0f40000000, 0fBF800000;
fma.rn.f32 %f946, %f940, 0f40000000, 0fBF800000;
fma.rn.f32 %f947, %f944, 0f40000000, 0fBF800000;
fma.rn.f32 %f948, %f402, %f945, %f407;
fma.rn.f32 %f949, %f402, %f946, %f408;
fma.rn.f32 %f950, %f402, %f947, %f409;
sub.f32 %f951, %f948, %f7;
sub.f32 %f952, %f949, %f8;
sub.f32 %f953, %f950, %f9;
mul.f32 %f954, %f952, %f952;
fma.rn.f32 %f955, %f951, %f951, %f954;
fma.rn.f32 %f956, %f953, %f953, %f955;
sqrt.rn.f32 %f760, %f956;
rcp.rn.f32 %f957, %f760;
mul.f32 %f756, %f957, %f951;
mul.f32 %f757, %f957, %f952;
mul.f32 %f758, %f957, %f953;
ld.global.u32 %r113, [imageEnabled];
and.b32 %r114, %r113, 32;
setp.eq.s32 %p54, %r114, 0;
selp.f32 %f958, 0f3F800000, 0f41200000, %p54;
mul.f32 %f759, %f958, %f149;
st.local.u32 [%rd2+8], %r105;
st.local.u32 [%rd2+4], %r105;
st.local.u32 [%rd2], %r105;
ld.global.u32 %r100, [root];
// inline asm
call _rt_trace_64, (%r100, %f10, %f11, %f12, %f756, %f757, %f758, %r101, %f759, %f760, %rd35, %r44);
// inline asm
ld.local.f32 %f959, [%rd2];
add.f32 %f1572, %f909, %f959;
ld.local.f32 %f960, [%rd2+4];
add.f32 %f1571, %f911, %f960;
ld.local.f32 %f961, [%rd2+8];
add.f32 %f1570, %f913, %f961;
add.s32 %r265, %r265, 4;
setp.lt.s32 %p55, %r265, %r8;
@%p55 bra BB0_41;
BB0_42:
cvt.rn.f32.s32 %f962, %r8;
rcp.rn.f32 %f963, %f962;
mul.f32 %f1573, %f1572, %f963;
mul.f32 %f1574, %f1571, %f963;
mul.f32 %f1575, %f1570, %f963;
shr.u32 %r115, %r6, 31;
cvt.u16.u32 %rs161, %r115;
BB0_43:
fma.rn.f32 %f1531, %f1543, %f1573, %f1531;
fma.rn.f32 %f1530, %f1544, %f1574, %f1530;
fma.rn.f32 %f1529, %f1545, %f1575, %f1529;
fma.rn.f32 %f1528, %f136, %f1573, %f1528;
fma.rn.f32 %f1527, %f1547, %f1574, %f1527;
fma.rn.f32 %f1526, %f1548, %f1575, %f1526;
fma.rn.f32 %f1525, %f139, %f1573, %f1525;
fma.rn.f32 %f1524, %f1550, %f1574, %f1524;
fma.rn.f32 %f1523, %f1551, %f1575, %f1523;
fma.rn.f32 %f1522, %f142, %f1573, %f1522;
fma.rn.f32 %f1521, %f1553, %f1574, %f1521;
fma.rn.f32 %f1520, %f1554, %f1575, %f1520;
fma.rn.f32 %f1519, %f145, %f1573, %f1519;
fma.rn.f32 %f1518, %f1556, %f1574, %f1518;
fma.rn.f32 %f1517, %f1557, %f1575, %f1517;
setp.eq.s16 %p56, %rs161, 0;
@%p56 bra BB0_58;
mov.f32 %f1509, 0fB5BFBE8E;
mov.f32 %f1508, 0fBF317200;
mov.f32 %f1507, 0f35BFBE8E;
mov.f32 %f1506, 0f3F317200;
mov.f32 %f1505, 0f3DAAAABD;
mov.f32 %f1504, 0f3C4CAF63;
mov.f32 %f1503, 0f3B18F0FE;
mul.f32 %f966, %f139, 0f3F000000;
mul.f32 %f967, %f142, 0f3F000000;
mul.f32 %f968, %f967, %f967;
fma.rn.f32 %f969, %f966, %f966, %f968;
mul.f32 %f970, %f145, 0f3F000000;
fma.rn.f32 %f971, %f970, %f970, %f969;
sqrt.rn.f32 %f972, %f971;
rcp.rn.f32 %f973, %f972;
mul.f32 %f974, %f966, %f973;
mul.f32 %f975, %f967, %f973;
mul.f32 %f976, %f970, %f973;
mul.f32 %f977, %f1514, %f975;
fma.rn.f32 %f978, %f1513, %f974, %f977;
fma.rn.f32 %f979, %f1515, %f976, %f978;
fma.rn.f32 %f198, %f979, 0f3F000000, 0f3F000000;
add.f32 %f980, %f972, %f972;
div.rn.f32 %f981, %f980, %f136;
add.f32 %f199, %f981, 0f3F800000;
div.rn.f32 %f982, %f972, %f136;
mov.f32 %f983, 0f3F800000;
sub.f32 %f984, %f983, %f982;
add.f32 %f985, %f982, 0f3F800000;
div.rn.f32 %f200, %f984, %f985;
sub.f32 %f986, %f983, %f200;
add.f32 %f987, %f199, 0f3F800000;
mul.f32 %f201, %f987, %f986;
mul.f32 %f988, %f199, 0f3F000000;
cvt.rzi.f32.f32 %f989, %f988;
add.f32 %f990, %f989, %f989;
sub.f32 %f991, %f199, %f990;
abs.f32 %f202, %f991;
abs.f32 %f203, %f198;
setp.lt.f32 %p57, %f203, 0f00800000;
mul.f32 %f992, %f203, 0f4B800000;
selp.f32 %f993, 0fC3170000, 0fC2FE0000, %p57;
selp.f32 %f994, %f992, %f203, %p57;
mov.b32 %r116, %f994;
and.b32 %r117, %r116, 8388607;
or.b32 %r118, %r117, 1065353216;
mov.b32 %f995, %r118;
shr.u32 %r119, %r116, 23;
cvt.rn.f32.u32 %f996, %r119;
add.f32 %f997, %f993, %f996;
setp.gt.f32 %p58, %f995, 0f3FB504F3;
mul.f32 %f998, %f995, 0f3F000000;
add.f32 %f999, %f997, 0f3F800000;
selp.f32 %f1000, %f998, %f995, %p58;
selp.f32 %f1001, %f999, %f997, %p58;
add.f32 %f1002, %f1000, 0fBF800000;
add.f32 %f965, %f1000, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f964,%f965;
// inline asm
add.f32 %f1003, %f1002, %f1002;
mul.f32 %f1004, %f964, %f1003;
mul.f32 %f1005, %f1004, %f1004;
fma.rn.f32 %f1008, %f1503, %f1005, %f1504;
fma.rn.f32 %f1010, %f1008, %f1005, %f1505;
mul.rn.f32 %f1011, %f1010, %f1005;
mul.rn.f32 %f1012, %f1011, %f1004;
sub.f32 %f1013, %f1002, %f1004;
neg.f32 %f1014, %f1004;
add.f32 %f1015, %f1013, %f1013;
fma.rn.f32 %f1016, %f1014, %f1002, %f1015;
mul.rn.f32 %f1017, %f964, %f1016;
add.f32 %f1018, %f1012, %f1004;
sub.f32 %f1019, %f1004, %f1018;
add.f32 %f1020, %f1012, %f1019;
add.f32 %f1021, %f1017, %f1020;
add.f32 %f1022, %f1018, %f1021;
sub.f32 %f1023, %f1018, %f1022;
add.f32 %f1024, %f1021, %f1023;
mul.rn.f32 %f1026, %f1001, %f1506;
mul.rn.f32 %f1028, %f1001, %f1507;
add.f32 %f1029, %f1026, %f1022;
sub.f32 %f1030, %f1026, %f1029;
add.f32 %f1031, %f1022, %f1030;
add.f32 %f1032, %f1024, %f1031;
add.f32 %f1033, %f1028, %f1032;
add.f32 %f1034, %f1029, %f1033;
sub.f32 %f1035, %f1029, %f1034;
add.f32 %f1036, %f1033, %f1035;
abs.f32 %f204, %f199;
setp.gt.f32 %p59, %f204, 0f77F684DF;
mul.f32 %f1037, %f199, 0f39000000;
selp.f32 %f1038, %f1037, %f199, %p59;
mul.rn.f32 %f1039, %f1038, %f1034;
neg.f32 %f1040, %f1039;
fma.rn.f32 %f1041, %f1038, %f1034, %f1040;
fma.rn.f32 %f1042, %f1038, %f1036, %f1041;
mov.f32 %f1043, 0f00000000;
fma.rn.f32 %f1044, %f1043, %f1034, %f1042;
add.rn.f32 %f1045, %f1039, %f1044;
neg.f32 %f1046, %f1045;
add.rn.f32 %f1047, %f1039, %f1046;
add.rn.f32 %f1048, %f1047, %f1044;
mov.b32 %r120, %f1045;
setp.eq.s32 %p60, %r120, 1118925336;
add.s32 %r121, %r120, -1;
mov.b32 %f1049, %r121;
add.f32 %f1050, %f1048, 0f37000000;
selp.f32 %f1051, %f1049, %f1045, %p60;
selp.f32 %f205, %f1050, %f1048, %p60;
mul.f32 %f1052, %f1051, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1053, %f1052;
fma.rn.f32 %f1055, %f1053, %f1508, %f1051;
fma.rn.f32 %f1057, %f1053, %f1509, %f1055;
mul.f32 %f1058, %f1057, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1059, %f1058;
add.f32 %f1060, %f1053, 0f00000000;
ex2.approx.f32 %f1061, %f1060;
mul.f32 %f1062, %f1059, %f1061;
setp.lt.f32 %p61, %f1051, 0fC2D20000;
selp.f32 %f1063, 0f00000000, %f1062, %p61;
setp.gt.f32 %p62, %f1051, 0f42D20000;
selp.f32 %f1576, 0f7F800000, %f1063, %p62;
setp.eq.f32 %p63, %f1576, 0f7F800000;
@%p63 bra BB0_46;
fma.rn.f32 %f1576, %f1576, %f205, %f1576;
BB0_46:
setp.lt.f32 %p64, %f198, 0f00000000;
setp.eq.f32 %p65, %f202, 0f3F800000;
and.pred %p4, %p64, %p65;
mov.b32 %r122, %f1576;
xor.b32 %r123, %r122, -2147483648;
mov.b32 %f1064, %r123;
selp.f32 %f1578, %f1064, %f1576, %p4;
setp.eq.f32 %p66, %f198, 0f00000000;
@%p66 bra BB0_49;
bra.uni BB0_47;
BB0_49:
add.f32 %f1066, %f198, %f198;
mov.b32 %r124, %f1066;
selp.b32 %r125, %r124, 0, %p65;
or.b32 %r126, %r125, 2139095040;
setp.lt.f32 %p70, %f199, 0f00000000;
selp.b32 %r127, %r126, %r125, %p70;
mov.b32 %f1578, %r127;
bra.uni BB0_50;
BB0_47:
setp.geu.f32 %p67, %f198, 0f00000000;
@%p67 bra BB0_50;
cvt.rzi.f32.f32 %f1065, %f199;
setp.neu.f32 %p68, %f1065, %f199;
selp.f32 %f1578, 0f7FFFFFFF, %f1578, %p68;
BB0_50:
add.f32 %f1067, %f203, %f204;
mov.b32 %r128, %f1067;
setp.lt.s32 %p71, %r128, 2139095040;
@%p71 bra BB0_57;
setp.gtu.f32 %p72, %f203, 0f7F800000;
setp.gtu.f32 %p73, %f204, 0f7F800000;
or.pred %p74, %p72, %p73;
@%p74 bra BB0_56;
bra.uni BB0_52;
BB0_56:
add.f32 %f1578, %f199, %f198;
bra.uni BB0_57;
BB0_52:
setp.eq.f32 %p75, %f204, 0f7F800000;
@%p75 bra BB0_55;
bra.uni BB0_53;
BB0_55:
setp.gt.f32 %p78, %f203, 0f3F800000;
selp.b32 %r132, 2139095040, 0, %p78;
xor.b32 %r133, %r132, 2139095040;
setp.lt.f32 %p79, %f199, 0f00000000;
selp.b32 %r134, %r133, %r132, %p79;
mov.b32 %f1068, %r134;
setp.eq.f32 %p80, %f198, 0fBF800000;
selp.f32 %f1578, 0f3F800000, %f1068, %p80;
bra.uni BB0_57;
BB0_53:
setp.neu.f32 %p76, %f203, 0f7F800000;
@%p76 bra BB0_57;
setp.ltu.f32 %p77, %f199, 0f00000000;
selp.b32 %r129, 0, 2139095040, %p77;
or.b32 %r130, %r129, -2147483648;
selp.b32 %r131, %r130, %r129, %p4;
mov.b32 %f1578, %r131;
BB0_57:
setp.eq.f32 %p81, %f199, 0f00000000;
setp.eq.f32 %p82, %f198, 0f3F800000;
or.pred %p83, %p82, %p81;
selp.f32 %f1069, 0f3F800000, %f1578, %p83;
fma.rn.f32 %f1070, %f201, %f1069, %f200;
mul.f32 %f1071, %f136, %f1070;
div.rn.f32 %f1072, %f1071, %f398;
div.rn.f32 %f1073, %f1072, %f96;
cvt.sat.f32.f32 %f1074, %f1073;
mul.f32 %f1573, %f1573, %f1074;
BB0_58:
add.f32 %f1516, %f1516, %f1573;
BB0_59:
add.s32 %r261, %r261, 1;
setp.lt.u32 %p84, %r261, %r4;
@%p84 bra BB0_5;
BB0_60:
ld.global.u32 %r267, [imageEnabled];
and.b32 %r135, %r267, 8;
setp.eq.s32 %p85, %r135, 0;
@%p85 bra BB0_73;
cvt.sat.f32.f32 %f252, %f1516;
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 %f1077, 0f3E68BA2E;
cvt.rzi.f32.f32 %f1078, %f1077;
fma.rn.f32 %f1079, %f1078, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f253, %f1079;
abs.f32 %f254, %f252;
setp.lt.f32 %p86, %f254, 0f00800000;
mul.f32 %f1080, %f254, 0f4B800000;
selp.f32 %f1081, 0fC3170000, 0fC2FE0000, %p86;
selp.f32 %f1082, %f1080, %f254, %p86;
mov.b32 %r138, %f1082;
and.b32 %r139, %r138, 8388607;
or.b32 %r140, %r139, 1065353216;
mov.b32 %f1083, %r140;
shr.u32 %r141, %r138, 23;
cvt.rn.f32.u32 %f1084, %r141;
add.f32 %f1085, %f1081, %f1084;
setp.gt.f32 %p87, %f1083, 0f3FB504F3;
mul.f32 %f1086, %f1083, 0f3F000000;
add.f32 %f1087, %f1085, 0f3F800000;
selp.f32 %f1088, %f1086, %f1083, %p87;
selp.f32 %f1089, %f1087, %f1085, %p87;
add.f32 %f1090, %f1088, 0fBF800000;
add.f32 %f1076, %f1088, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1075,%f1076;
// inline asm
add.f32 %f1091, %f1090, %f1090;
mul.f32 %f1092, %f1075, %f1091;
mul.f32 %f1093, %f1092, %f1092;
mov.f32 %f1094, 0f3C4CAF63;
mov.f32 %f1095, 0f3B18F0FE;
fma.rn.f32 %f1096, %f1095, %f1093, %f1094;
mov.f32 %f1097, 0f3DAAAABD;
fma.rn.f32 %f1098, %f1096, %f1093, %f1097;
mul.rn.f32 %f1099, %f1098, %f1093;
mul.rn.f32 %f1100, %f1099, %f1092;
sub.f32 %f1101, %f1090, %f1092;
neg.f32 %f1102, %f1092;
add.f32 %f1103, %f1101, %f1101;
fma.rn.f32 %f1104, %f1102, %f1090, %f1103;
mul.rn.f32 %f1105, %f1075, %f1104;
add.f32 %f1106, %f1100, %f1092;
sub.f32 %f1107, %f1092, %f1106;
add.f32 %f1108, %f1100, %f1107;
add.f32 %f1109, %f1105, %f1108;
add.f32 %f1110, %f1106, %f1109;
sub.f32 %f1111, %f1106, %f1110;
add.f32 %f1112, %f1109, %f1111;
mov.f32 %f1113, 0f3F317200;
mul.rn.f32 %f1114, %f1089, %f1113;
mov.f32 %f1115, 0f35BFBE8E;
mul.rn.f32 %f1116, %f1089, %f1115;
add.f32 %f1117, %f1114, %f1110;
sub.f32 %f1118, %f1114, %f1117;
add.f32 %f1119, %f1110, %f1118;
add.f32 %f1120, %f1112, %f1119;
add.f32 %f1121, %f1116, %f1120;
add.f32 %f1122, %f1117, %f1121;
sub.f32 %f1123, %f1117, %f1122;
add.f32 %f1124, %f1121, %f1123;
mov.f32 %f1125, 0f3EE8BA2E;
mul.rn.f32 %f1126, %f1125, %f1122;
neg.f32 %f1127, %f1126;
fma.rn.f32 %f1128, %f1125, %f1122, %f1127;
fma.rn.f32 %f1129, %f1125, %f1124, %f1128;
mov.f32 %f1130, 0f00000000;
fma.rn.f32 %f1131, %f1130, %f1122, %f1129;
add.rn.f32 %f1132, %f1126, %f1131;
neg.f32 %f1133, %f1132;
add.rn.f32 %f1134, %f1126, %f1133;
add.rn.f32 %f1135, %f1134, %f1131;
mov.b32 %r142, %f1132;
setp.eq.s32 %p88, %r142, 1118925336;
add.s32 %r143, %r142, -1;
mov.b32 %f1136, %r143;
add.f32 %f1137, %f1135, 0f37000000;
selp.f32 %f1138, %f1136, %f1132, %p88;
selp.f32 %f255, %f1137, %f1135, %p88;
mul.f32 %f1139, %f1138, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1140, %f1139;
mov.f32 %f1141, 0fBF317200;
fma.rn.f32 %f1142, %f1140, %f1141, %f1138;
mov.f32 %f1143, 0fB5BFBE8E;
fma.rn.f32 %f1144, %f1140, %f1143, %f1142;
mul.f32 %f1145, %f1144, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1146, %f1145;
add.f32 %f1147, %f1140, 0f00000000;
ex2.approx.f32 %f1148, %f1147;
mul.f32 %f1149, %f1146, %f1148;
setp.lt.f32 %p89, %f1138, 0fC2D20000;
selp.f32 %f1150, 0f00000000, %f1149, %p89;
setp.gt.f32 %p90, %f1138, 0f42D20000;
selp.f32 %f1612, 0f7F800000, %f1150, %p90;
setp.eq.f32 %p91, %f1612, 0f7F800000;
@%p91 bra BB0_63;
fma.rn.f32 %f1612, %f1612, %f255, %f1612;
BB0_63:
setp.lt.f32 %p92, %f252, 0f00000000;
setp.eq.f32 %p93, %f253, 0f3F800000;
and.pred %p5, %p92, %p93;
mov.b32 %r144, %f1612;
xor.b32 %r145, %r144, -2147483648;
mov.b32 %f1151, %r145;
selp.f32 %f1614, %f1151, %f1612, %p5;
setp.eq.f32 %p94, %f252, 0f00000000;
@%p94 bra BB0_66;
bra.uni BB0_64;
BB0_66:
add.f32 %f1154, %f252, %f252;
selp.f32 %f1614, %f1154, 0f00000000, %p93;
bra.uni BB0_67;
BB0_131:
mov.u64 %rd190, image_HDR;
cvta.global.u64 %rd185, %rd190;
mov.u32 %r231, 8;
// inline asm
call (%rd184), _rt_buffer_get_64, (%rd185, %r28, %r231, %rd6, %rd7, %rd13, %rd13);
// inline asm
mov.f32 %f1474, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs107, %f1474;}
// inline asm
mov.u16 %rs108, 0;
st.v4.u16 [%rd184], {%rs107, %rs107, %rs107, %rs108};
BB0_132:
ld.global.u32 %r232, [additive];
setp.eq.s32 %p161, %r232, 0;
@%p161 bra BB0_134;
mov.u64 %rd203, image_RNM0;
cvta.global.u64 %rd192, %rd203;
mov.u32 %r236, 8;
// inline asm
call (%rd191), _rt_buffer_get_64, (%rd192, %r28, %r236, %rd6, %rd7, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs115, %rs116, %rs117, %rs118}, [%rd191];
// inline asm
{ cvt.f32.f16 %f1475, %rs115;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1476, %rs116;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1477, %rs117;}
// inline asm
// inline asm
call (%rd197), _rt_buffer_get_64, (%rd192, %r28, %r236, %rd6, %rd7, %rd13, %rd13);
// inline asm
add.f32 %f1478, %f1475, 0f00000000;
add.f32 %f1479, %f1476, 0f00000000;
add.f32 %f1480, %f1477, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs114, %f1480;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs113, %f1479;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs112, %f1478;}
// inline asm
mov.u16 %rs119, 0;
st.v4.u16 [%rd197], {%rs112, %rs113, %rs114, %rs119};
bra.uni BB0_135;
BB0_134:
mov.u64 %rd210, image_RNM0;
cvta.global.u64 %rd205, %rd210;
mov.u32 %r238, 8;
// inline asm
call (%rd204), _rt_buffer_get_64, (%rd205, %r28, %r238, %rd6, %rd7, %rd13, %rd13);
// inline asm
mov.f32 %f1481, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs120, %f1481;}
// inline asm
mov.u16 %rs121, 0;
st.v4.u16 [%rd204], {%rs120, %rs120, %rs120, %rs121};
BB0_135:
ld.global.u32 %r239, [additive];
setp.eq.s32 %p162, %r239, 0;
@%p162 bra BB0_137;
mov.u64 %rd223, image_RNM1;
cvta.global.u64 %rd212, %rd223;
mov.u32 %r243, 8;
// inline asm
call (%rd211), _rt_buffer_get_64, (%rd212, %r28, %r243, %rd6, %rd7, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs128, %rs129, %rs130, %rs131}, [%rd211];
// inline asm
{ cvt.f32.f16 %f1482, %rs128;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1483, %rs129;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1484, %rs130;}
// inline asm
// inline asm
call (%rd217), _rt_buffer_get_64, (%rd212, %r28, %r243, %rd6, %rd7, %rd13, %rd13);
// inline asm
add.f32 %f1485, %f1482, 0f00000000;
add.f32 %f1486, %f1483, 0f00000000;
add.f32 %f1487, %f1484, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs127, %f1487;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs126, %f1486;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs125, %f1485;}
// inline asm
mov.u16 %rs132, 0;
st.v4.u16 [%rd217], {%rs125, %rs126, %rs127, %rs132};
bra.uni BB0_138;
BB0_137:
mov.u64 %rd230, image_RNM1;
cvta.global.u64 %rd225, %rd230;
mov.u32 %r245, 8;
// inline asm
call (%rd224), _rt_buffer_get_64, (%rd225, %r28, %r245, %rd6, %rd7, %rd13, %rd13);
// inline asm
mov.f32 %f1488, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs133, %f1488;}
// inline asm
mov.u16 %rs134, 0;
st.v4.u16 [%rd224], {%rs133, %rs133, %rs133, %rs134};
BB0_138:
ld.global.u32 %r246, [additive];
setp.eq.s32 %p163, %r246, 0;
@%p163 bra BB0_140;
mov.u64 %rd243, image_RNM2;
cvta.global.u64 %rd232, %rd243;
mov.u32 %r250, 8;
// inline asm
call (%rd231), _rt_buffer_get_64, (%rd232, %r28, %r250, %rd6, %rd7, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs141, %rs142, %rs143, %rs144}, [%rd231];
// inline asm
{ cvt.f32.f16 %f1489, %rs141;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1490, %rs142;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1491, %rs143;}
// inline asm
// inline asm
call (%rd237), _rt_buffer_get_64, (%rd232, %r28, %r250, %rd6, %rd7, %rd13, %rd13);
// inline asm
add.f32 %f1492, %f1489, 0f00000000;
add.f32 %f1493, %f1490, 0f00000000;
add.f32 %f1494, %f1491, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs140, %f1494;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs139, %f1493;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs138, %f1492;}
// inline asm
mov.u16 %rs145, 0;
st.v4.u16 [%rd237], {%rs138, %rs139, %rs140, %rs145};
bra.uni BB0_141;
BB0_140:
mov.u64 %rd250, image_RNM2;
cvta.global.u64 %rd245, %rd250;
mov.u32 %r252, 8;
// inline asm
call (%rd244), _rt_buffer_get_64, (%rd245, %r28, %r252, %rd6, %rd7, %rd13, %rd13);
// inline asm
mov.f32 %f1495, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs146, %f1495;}
// inline asm
mov.u16 %rs147, 0;
st.v4.u16 [%rd244], {%rs146, %rs146, %rs146, %rs147};
BB0_141:
ld.global.u32 %r253, [additive];
setp.eq.s32 %p164, %r253, 0;
@%p164 bra BB0_143;
mov.u64 %rd263, image_RNM3;
cvta.global.u64 %rd252, %rd263;
mov.u32 %r257, 8;
// inline asm
call (%rd251), _rt_buffer_get_64, (%rd252, %r28, %r257, %rd6, %rd7, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs154, %rs155, %rs156, %rs157}, [%rd251];
// inline asm
{ cvt.f32.f16 %f1496, %rs154;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1497, %rs155;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1498, %rs156;}
// inline asm
// inline asm
call (%rd257), _rt_buffer_get_64, (%rd252, %r28, %r257, %rd6, %rd7, %rd13, %rd13);
// inline asm
add.f32 %f1499, %f1496, 0f00000000;
add.f32 %f1500, %f1497, 0f00000000;
add.f32 %f1501, %f1498, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs153, %f1501;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs152, %f1500;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs151, %f1499;}
// inline asm
mov.u16 %rs158, 0;
st.v4.u16 [%rd257], {%rs151, %rs152, %rs153, %rs158};
bra.uni BB0_144;
BB0_143:
mov.u64 %rd270, image_RNM3;
cvta.global.u64 %rd265, %rd270;
mov.u32 %r259, 8;
// inline asm
call (%rd264), _rt_buffer_get_64, (%rd265, %r28, %r259, %rd6, %rd7, %rd13, %rd13);
// inline asm
mov.f32 %f1502, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs159, %f1502;}
// inline asm
mov.u16 %rs160, 0;
st.v4.u16 [%rd264], {%rs159, %rs159, %rs159, %rs160};
bra.uni BB0_144;
BB0_64:
setp.geu.f32 %p95, %f252, 0f00000000;
@%p95 bra BB0_67;
cvt.rzi.f32.f32 %f1153, %f1125;
setp.neu.f32 %p96, %f1153, 0f3EE8BA2E;
selp.f32 %f1614, 0f7FFFFFFF, %f1614, %p96;
BB0_67:
add.f32 %f1155, %f254, 0f3EE8BA2E;
mov.b32 %r146, %f1155;
setp.lt.s32 %p98, %r146, 2139095040;
@%p98 bra BB0_72;
setp.gtu.f32 %p99, %f254, 0f7F800000;
@%p99 bra BB0_71;
bra.uni BB0_69;
BB0_71:
add.f32 %f1614, %f252, 0f3EE8BA2E;
bra.uni BB0_72;
BB0_69:
setp.neu.f32 %p100, %f254, 0f7F800000;
@%p100 bra BB0_72;
selp.f32 %f1614, 0fFF800000, 0f7F800000, %p5;
BB0_72:
mul.f32 %f1156, %f1614, 0f437F0000;
setp.eq.f32 %p101, %f252, 0f3F800000;
selp.f32 %f1157, 0f437F0000, %f1156, %p101;
cvt.rzi.u32.f32 %r147, %f1157;
cvt.u16.u32 %rs17, %r147;
mov.u16 %rs18, 255;
st.v2.u8 [%rd43], {%rs17, %rs18};
ld.global.u32 %r267, [imageEnabled];
BB0_73:
and.b32 %r148, %r267, 1;
setp.eq.b32 %p102, %r148, 1;
@!%p102 bra BB0_108;
bra.uni BB0_74;
BB0_74:
mov.f32 %f1160, 0f3E666666;
cvt.rzi.f32.f32 %f1161, %f1160;
fma.rn.f32 %f1162, %f1161, 0fC0000000, 0f3EE66666;
abs.f32 %f266, %f1162;
abs.f32 %f267, %f1531;
setp.lt.f32 %p103, %f267, 0f00800000;
mul.f32 %f1163, %f267, 0f4B800000;
selp.f32 %f1164, 0fC3170000, 0fC2FE0000, %p103;
selp.f32 %f1165, %f1163, %f267, %p103;
mov.b32 %r149, %f1165;
and.b32 %r150, %r149, 8388607;
or.b32 %r151, %r150, 1065353216;
mov.b32 %f1166, %r151;
shr.u32 %r152, %r149, 23;
cvt.rn.f32.u32 %f1167, %r152;
add.f32 %f1168, %f1164, %f1167;
setp.gt.f32 %p104, %f1166, 0f3FB504F3;
mul.f32 %f1169, %f1166, 0f3F000000;
add.f32 %f1170, %f1168, 0f3F800000;
selp.f32 %f1171, %f1169, %f1166, %p104;
selp.f32 %f1172, %f1170, %f1168, %p104;
add.f32 %f1173, %f1171, 0fBF800000;
add.f32 %f1159, %f1171, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1158,%f1159;
// inline asm
add.f32 %f1174, %f1173, %f1173;
mul.f32 %f1175, %f1158, %f1174;
mul.f32 %f1176, %f1175, %f1175;
mov.f32 %f1177, 0f3C4CAF63;
mov.f32 %f1178, 0f3B18F0FE;
fma.rn.f32 %f1179, %f1178, %f1176, %f1177;
mov.f32 %f1180, 0f3DAAAABD;
fma.rn.f32 %f1181, %f1179, %f1176, %f1180;
mul.rn.f32 %f1182, %f1181, %f1176;
mul.rn.f32 %f1183, %f1182, %f1175;
sub.f32 %f1184, %f1173, %f1175;
neg.f32 %f1185, %f1175;
add.f32 %f1186, %f1184, %f1184;
fma.rn.f32 %f1187, %f1185, %f1173, %f1186;
mul.rn.f32 %f1188, %f1158, %f1187;
add.f32 %f1189, %f1183, %f1175;
sub.f32 %f1190, %f1175, %f1189;
add.f32 %f1191, %f1183, %f1190;
add.f32 %f1192, %f1188, %f1191;
add.f32 %f1193, %f1189, %f1192;
sub.f32 %f1194, %f1189, %f1193;
add.f32 %f1195, %f1192, %f1194;
mov.f32 %f1196, 0f3F317200;
mul.rn.f32 %f1197, %f1172, %f1196;
mov.f32 %f1198, 0f35BFBE8E;
mul.rn.f32 %f1199, %f1172, %f1198;
add.f32 %f1200, %f1197, %f1193;
sub.f32 %f1201, %f1197, %f1200;
add.f32 %f1202, %f1193, %f1201;
add.f32 %f1203, %f1195, %f1202;
add.f32 %f1204, %f1199, %f1203;
add.f32 %f1205, %f1200, %f1204;
sub.f32 %f1206, %f1200, %f1205;
add.f32 %f1207, %f1204, %f1206;
mov.f32 %f1208, 0f3EE66666;
mul.rn.f32 %f1209, %f1208, %f1205;
neg.f32 %f1210, %f1209;
fma.rn.f32 %f1211, %f1208, %f1205, %f1210;
fma.rn.f32 %f1212, %f1208, %f1207, %f1211;
mov.f32 %f1213, 0f00000000;
fma.rn.f32 %f1214, %f1213, %f1205, %f1212;
add.rn.f32 %f1215, %f1209, %f1214;
neg.f32 %f1216, %f1215;
add.rn.f32 %f1217, %f1209, %f1216;
add.rn.f32 %f1218, %f1217, %f1214;
mov.b32 %r153, %f1215;
setp.eq.s32 %p105, %r153, 1118925336;
add.s32 %r154, %r153, -1;
mov.b32 %f1219, %r154;
add.f32 %f1220, %f1218, 0f37000000;
selp.f32 %f1221, %f1219, %f1215, %p105;
selp.f32 %f268, %f1220, %f1218, %p105;
mul.f32 %f1222, %f1221, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1223, %f1222;
mov.f32 %f1224, 0fBF317200;
fma.rn.f32 %f1225, %f1223, %f1224, %f1221;
mov.f32 %f1226, 0fB5BFBE8E;
fma.rn.f32 %f1227, %f1223, %f1226, %f1225;
mul.f32 %f1228, %f1227, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1229, %f1228;
add.f32 %f1230, %f1223, 0f00000000;
ex2.approx.f32 %f1231, %f1230;
mul.f32 %f1232, %f1229, %f1231;
setp.lt.f32 %p106, %f1221, 0fC2D20000;
selp.f32 %f1233, 0f00000000, %f1232, %p106;
setp.gt.f32 %p107, %f1221, 0f42D20000;
selp.f32 %f1615, 0f7F800000, %f1233, %p107;
setp.eq.f32 %p108, %f1615, 0f7F800000;
@%p108 bra BB0_76;
fma.rn.f32 %f1615, %f1615, %f268, %f1615;
BB0_76:
setp.lt.f32 %p109, %f1531, 0f00000000;
setp.eq.f32 %p110, %f266, 0f3F800000;
and.pred %p6, %p109, %p110;
mov.b32 %r155, %f1615;
xor.b32 %r156, %r155, -2147483648;
mov.b32 %f1234, %r156;
selp.f32 %f1617, %f1234, %f1615, %p6;
setp.eq.f32 %p111, %f1531, 0f00000000;
@%p111 bra BB0_79;
bra.uni BB0_77;
BB0_79:
add.f32 %f1237, %f1531, %f1531;
selp.f32 %f1617, %f1237, 0f00000000, %p110;
bra.uni BB0_80;
BB0_77:
setp.geu.f32 %p112, %f1531, 0f00000000;
@%p112 bra BB0_80;
cvt.rzi.f32.f32 %f1236, %f1208;
setp.neu.f32 %p113, %f1236, 0f3EE66666;
selp.f32 %f1617, 0f7FFFFFFF, %f1617, %p113;
BB0_80:
add.f32 %f1238, %f267, 0f3EE66666;
mov.b32 %r157, %f1238;
setp.lt.s32 %p115, %r157, 2139095040;
@%p115 bra BB0_85;
setp.gtu.f32 %p116, %f267, 0f7F800000;
@%p116 bra BB0_84;
bra.uni BB0_82;
BB0_84:
add.f32 %f1617, %f1531, 0f3EE66666;
bra.uni BB0_85;
BB0_82:
setp.neu.f32 %p117, %f267, 0f7F800000;
@%p117 bra BB0_85;
selp.f32 %f1617, 0fFF800000, 0f7F800000, %p6;
BB0_85:
setp.eq.f32 %p118, %f1531, 0f3F800000;
selp.f32 %f279, 0f3F800000, %f1617, %p118;
abs.f32 %f280, %f1530;
setp.lt.f32 %p119, %f280, 0f00800000;
mul.f32 %f1241, %f280, 0f4B800000;
selp.f32 %f1242, 0fC3170000, 0fC2FE0000, %p119;
selp.f32 %f1243, %f1241, %f280, %p119;
mov.b32 %r158, %f1243;
and.b32 %r159, %r158, 8388607;
or.b32 %r160, %r159, 1065353216;
mov.b32 %f1244, %r160;
shr.u32 %r161, %r158, 23;
cvt.rn.f32.u32 %f1245, %r161;
add.f32 %f1246, %f1242, %f1245;
setp.gt.f32 %p120, %f1244, 0f3FB504F3;
mul.f32 %f1247, %f1244, 0f3F000000;
add.f32 %f1248, %f1246, 0f3F800000;
selp.f32 %f1249, %f1247, %f1244, %p120;
selp.f32 %f1250, %f1248, %f1246, %p120;
add.f32 %f1251, %f1249, 0fBF800000;
add.f32 %f1240, %f1249, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1239,%f1240;
// inline asm
add.f32 %f1252, %f1251, %f1251;
mul.f32 %f1253, %f1239, %f1252;
mul.f32 %f1254, %f1253, %f1253;
fma.rn.f32 %f1257, %f1178, %f1254, %f1177;
fma.rn.f32 %f1259, %f1257, %f1254, %f1180;
mul.rn.f32 %f1260, %f1259, %f1254;
mul.rn.f32 %f1261, %f1260, %f1253;
sub.f32 %f1262, %f1251, %f1253;
neg.f32 %f1263, %f1253;
add.f32 %f1264, %f1262, %f1262;
fma.rn.f32 %f1265, %f1263, %f1251, %f1264;
mul.rn.f32 %f1266, %f1239, %f1265;
add.f32 %f1267, %f1261, %f1253;
sub.f32 %f1268, %f1253, %f1267;
add.f32 %f1269, %f1261, %f1268;
add.f32 %f1270, %f1266, %f1269;
add.f32 %f1271, %f1267, %f1270;
sub.f32 %f1272, %f1267, %f1271;
add.f32 %f1273, %f1270, %f1272;
mul.rn.f32 %f1275, %f1250, %f1196;
mul.rn.f32 %f1277, %f1250, %f1198;
add.f32 %f1278, %f1275, %f1271;
sub.f32 %f1279, %f1275, %f1278;
add.f32 %f1280, %f1271, %f1279;
add.f32 %f1281, %f1273, %f1280;
add.f32 %f1282, %f1277, %f1281;
add.f32 %f1283, %f1278, %f1282;
sub.f32 %f1284, %f1278, %f1283;
add.f32 %f1285, %f1282, %f1284;
mul.rn.f32 %f1287, %f1208, %f1283;
neg.f32 %f1288, %f1287;
fma.rn.f32 %f1289, %f1208, %f1283, %f1288;
fma.rn.f32 %f1290, %f1208, %f1285, %f1289;
fma.rn.f32 %f1292, %f1213, %f1283, %f1290;
add.rn.f32 %f1293, %f1287, %f1292;
neg.f32 %f1294, %f1293;
add.rn.f32 %f1295, %f1287, %f1294;
add.rn.f32 %f1296, %f1295, %f1292;
mov.b32 %r162, %f1293;
setp.eq.s32 %p121, %r162, 1118925336;
add.s32 %r163, %r162, -1;
mov.b32 %f1297, %r163;
add.f32 %f1298, %f1296, 0f37000000;
selp.f32 %f1299, %f1297, %f1293, %p121;
selp.f32 %f281, %f1298, %f1296, %p121;
mul.f32 %f1300, %f1299, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1301, %f1300;
fma.rn.f32 %f1303, %f1301, %f1224, %f1299;
fma.rn.f32 %f1305, %f1301, %f1226, %f1303;
mul.f32 %f1306, %f1305, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1307, %f1306;
add.f32 %f1308, %f1301, 0f00000000;
ex2.approx.f32 %f1309, %f1308;
mul.f32 %f1310, %f1307, %f1309;
setp.lt.f32 %p122, %f1299, 0fC2D20000;
selp.f32 %f1311, 0f00000000, %f1310, %p122;
setp.gt.f32 %p123, %f1299, 0f42D20000;
selp.f32 %f1618, 0f7F800000, %f1311, %p123;
setp.eq.f32 %p124, %f1618, 0f7F800000;
@%p124 bra BB0_87;
fma.rn.f32 %f1618, %f1618, %f281, %f1618;
BB0_87:
setp.lt.f32 %p125, %f1530, 0f00000000;
and.pred %p7, %p125, %p110;
mov.b32 %r164, %f1618;
xor.b32 %r165, %r164, -2147483648;
mov.b32 %f1312, %r165;
selp.f32 %f1620, %f1312, %f1618, %p7;
setp.eq.f32 %p127, %f1530, 0f00000000;
@%p127 bra BB0_90;
bra.uni BB0_88;
BB0_90:
add.f32 %f1315, %f1530, %f1530;
selp.f32 %f1620, %f1315, 0f00000000, %p110;
bra.uni BB0_91;
BB0_88:
setp.geu.f32 %p128, %f1530, 0f00000000;
@%p128 bra BB0_91;
cvt.rzi.f32.f32 %f1314, %f1208;
setp.neu.f32 %p129, %f1314, 0f3EE66666;
selp.f32 %f1620, 0f7FFFFFFF, %f1620, %p129;
BB0_91:
add.f32 %f1316, %f280, 0f3EE66666;
mov.b32 %r166, %f1316;
setp.lt.s32 %p131, %r166, 2139095040;
@%p131 bra BB0_96;
setp.gtu.f32 %p132, %f280, 0f7F800000;
@%p132 bra BB0_95;
bra.uni BB0_93;
BB0_95:
add.f32 %f1620, %f1530, 0f3EE66666;
bra.uni BB0_96;
BB0_93:
setp.neu.f32 %p133, %f280, 0f7F800000;
@%p133 bra BB0_96;
selp.f32 %f1620, 0fFF800000, 0f7F800000, %p7;
BB0_96:
setp.eq.f32 %p134, %f1530, 0f3F800000;
selp.f32 %f292, 0f3F800000, %f1620, %p134;
abs.f32 %f293, %f1529;
setp.lt.f32 %p135, %f293, 0f00800000;
mul.f32 %f1319, %f293, 0f4B800000;
selp.f32 %f1320, 0fC3170000, 0fC2FE0000, %p135;
selp.f32 %f1321, %f1319, %f293, %p135;
mov.b32 %r167, %f1321;
and.b32 %r168, %r167, 8388607;
or.b32 %r169, %r168, 1065353216;
mov.b32 %f1322, %r169;
shr.u32 %r170, %r167, 23;
cvt.rn.f32.u32 %f1323, %r170;
add.f32 %f1324, %f1320, %f1323;
setp.gt.f32 %p136, %f1322, 0f3FB504F3;
mul.f32 %f1325, %f1322, 0f3F000000;
add.f32 %f1326, %f1324, 0f3F800000;
selp.f32 %f1327, %f1325, %f1322, %p136;
selp.f32 %f1328, %f1326, %f1324, %p136;
add.f32 %f1329, %f1327, 0fBF800000;
add.f32 %f1318, %f1327, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1317,%f1318;
// inline asm
add.f32 %f1330, %f1329, %f1329;
mul.f32 %f1331, %f1317, %f1330;
mul.f32 %f1332, %f1331, %f1331;
fma.rn.f32 %f1335, %f1178, %f1332, %f1177;
fma.rn.f32 %f1337, %f1335, %f1332, %f1180;
mul.rn.f32 %f1338, %f1337, %f1332;
mul.rn.f32 %f1339, %f1338, %f1331;
sub.f32 %f1340, %f1329, %f1331;
neg.f32 %f1341, %f1331;
add.f32 %f1342, %f1340, %f1340;
fma.rn.f32 %f1343, %f1341, %f1329, %f1342;
mul.rn.f32 %f1344, %f1317, %f1343;
add.f32 %f1345, %f1339, %f1331;
sub.f32 %f1346, %f1331, %f1345;
add.f32 %f1347, %f1339, %f1346;
add.f32 %f1348, %f1344, %f1347;
add.f32 %f1349, %f1345, %f1348;
sub.f32 %f1350, %f1345, %f1349;
add.f32 %f1351, %f1348, %f1350;
mul.rn.f32 %f1353, %f1328, %f1196;
mul.rn.f32 %f1355, %f1328, %f1198;
add.f32 %f1356, %f1353, %f1349;
sub.f32 %f1357, %f1353, %f1356;
add.f32 %f1358, %f1349, %f1357;
add.f32 %f1359, %f1351, %f1358;
add.f32 %f1360, %f1355, %f1359;
add.f32 %f1361, %f1356, %f1360;
sub.f32 %f1362, %f1356, %f1361;
add.f32 %f1363, %f1360, %f1362;
mul.rn.f32 %f1365, %f1208, %f1361;
neg.f32 %f1366, %f1365;
fma.rn.f32 %f1367, %f1208, %f1361, %f1366;
fma.rn.f32 %f1368, %f1208, %f1363, %f1367;
fma.rn.f32 %f1370, %f1213, %f1361, %f1368;
add.rn.f32 %f1371, %f1365, %f1370;
neg.f32 %f1372, %f1371;
add.rn.f32 %f1373, %f1365, %f1372;
add.rn.f32 %f1374, %f1373, %f1370;
mov.b32 %r171, %f1371;
setp.eq.s32 %p137, %r171, 1118925336;
add.s32 %r172, %r171, -1;
mov.b32 %f1375, %r172;
add.f32 %f1376, %f1374, 0f37000000;
selp.f32 %f1377, %f1375, %f1371, %p137;
selp.f32 %f294, %f1376, %f1374, %p137;
mul.f32 %f1378, %f1377, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1379, %f1378;
fma.rn.f32 %f1381, %f1379, %f1224, %f1377;
fma.rn.f32 %f1383, %f1379, %f1226, %f1381;
mul.f32 %f1384, %f1383, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1385, %f1384;
add.f32 %f1386, %f1379, 0f00000000;
ex2.approx.f32 %f1387, %f1386;
mul.f32 %f1388, %f1385, %f1387;
setp.lt.f32 %p138, %f1377, 0fC2D20000;
selp.f32 %f1389, 0f00000000, %f1388, %p138;
setp.gt.f32 %p139, %f1377, 0f42D20000;
selp.f32 %f1621, 0f7F800000, %f1389, %p139;
setp.eq.f32 %p140, %f1621, 0f7F800000;
@%p140 bra BB0_98;
fma.rn.f32 %f1621, %f1621, %f294, %f1621;
BB0_98:
setp.lt.f32 %p141, %f1529, 0f00000000;
and.pred %p8, %p141, %p110;
mov.b32 %r173, %f1621;
xor.b32 %r174, %r173, -2147483648;
mov.b32 %f1390, %r174;
selp.f32 %f1623, %f1390, %f1621, %p8;
setp.eq.f32 %p143, %f1529, 0f00000000;
@%p143 bra BB0_101;
bra.uni BB0_99;
BB0_101:
add.f32 %f1393, %f1529, %f1529;
selp.f32 %f1623, %f1393, 0f00000000, %p110;
bra.uni BB0_102;
BB0_99:
setp.geu.f32 %p144, %f1529, 0f00000000;
@%p144 bra BB0_102;
cvt.rzi.f32.f32 %f1392, %f1208;
setp.neu.f32 %p145, %f1392, 0f3EE66666;
selp.f32 %f1623, 0f7FFFFFFF, %f1623, %p145;
BB0_102:
add.f32 %f1394, %f293, 0f3EE66666;
mov.b32 %r175, %f1394;
setp.lt.s32 %p147, %r175, 2139095040;
@%p147 bra BB0_107;
setp.gtu.f32 %p148, %f293, 0f7F800000;
@%p148 bra BB0_106;
bra.uni BB0_104;
BB0_106:
add.f32 %f1623, %f1529, 0f3EE66666;
bra.uni BB0_107;
BB0_104:
setp.neu.f32 %p149, %f293, 0f7F800000;
@%p149 bra BB0_107;
selp.f32 %f1623, 0fFF800000, 0f7F800000, %p8;
BB0_107:
mov.u32 %r260, 4;
setp.eq.f32 %p150, %f1529, 0f3F800000;
selp.f32 %f1395, 0f3F800000, %f1623, %p150;
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, %r260, %rd52, %rd53, %rd13, %rd13);
// inline asm
cvt.sat.f32.f32 %f1396, %f1395;
mul.f32 %f1397, %f1396, 0f437FFD71;
cvt.rzi.u32.f32 %r178, %f1397;
cvt.sat.f32.f32 %f1398, %f292;
mul.f32 %f1399, %f1398, 0f437FFD71;
cvt.rzi.u32.f32 %r179, %f1399;
cvt.sat.f32.f32 %f1400, %f279;
mul.f32 %f1401, %f1400, 0f437FFD71;
cvt.rzi.u32.f32 %r180, %f1401;
cvt.u16.u32 %rs19, %r178;
cvt.u16.u32 %rs20, %r180;
cvt.u16.u32 %rs21, %r179;
mov.u16 %rs22, 255;
st.v4.u8 [%rd50], {%rs19, %rs21, %rs20, %rs22};
ld.global.u32 %r267, [imageEnabled];
BB0_108:
cvt.u64.u32 %rd4, %r2;
cvt.u64.u32 %rd5, %r3;
and.b32 %r181, %r267, 4;
setp.eq.s32 %p151, %r181, 0;
@%p151 bra BB0_112;
ld.global.u32 %r182, [additive];
setp.eq.s32 %p152, %r182, 0;
mov.f32 %f1402, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs23, %f1402;}
// inline asm
@%p152 bra BB0_111;
mov.u64 %rd69, image_HDR;
cvta.global.u64 %rd58, %rd69;
mov.u32 %r186, 8;
// inline asm
call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r186, %rd4, %rd5, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd57];
// inline asm
{ cvt.f32.f16 %f1403, %rs30;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1404, %rs31;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1405, %rs32;}
// inline asm
// inline asm
call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r186, %rd4, %rd5, %rd13, %rd13);
// inline asm
add.f32 %f1406, %f1531, %f1403;
add.f32 %f1407, %f1530, %f1404;
add.f32 %f1408, %f1529, %f1405;
// inline asm
{ cvt.rn.f16.f32 %rs29, %f1408;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs28, %f1407;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs27, %f1406;}
// inline asm
st.v4.u16 [%rd63], {%rs27, %rs28, %rs29, %rs23};
bra.uni BB0_112;
BB0_111:
mov.u64 %rd76, image_HDR;
cvta.global.u64 %rd71, %rd76;
mov.u32 %r188, 8;
// inline asm
call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r188, %rd4, %rd5, %rd13, %rd13);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs36, %f1529;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs35, %f1530;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs34, %f1531;}
// inline asm
st.v4.u16 [%rd70], {%rs34, %rs35, %rs36, %rs23};
BB0_112:
mov.f32 %f1413, 0f34000000;
max.f32 %f1414, %f1528, %f1413;
div.rn.f32 %f1415, %f1525, %f1414;
max.f32 %f1416, %f1527, %f1413;
div.rn.f32 %f1417, %f1524, %f1416;
max.f32 %f1418, %f1526, %f1413;
div.rn.f32 %f1419, %f1523, %f1418;
fma.rn.f32 %f305, %f1415, 0f3F000000, 0f3F000000;
fma.rn.f32 %f306, %f1417, 0f3F000000, 0f3F000000;
fma.rn.f32 %f307, %f1419, 0f3F000000, 0f3F000000;
div.rn.f32 %f1420, %f1522, %f1414;
div.rn.f32 %f1421, %f1521, %f1416;
div.rn.f32 %f1422, %f1520, %f1418;
fma.rn.f32 %f308, %f1420, 0f3F000000, 0f3F000000;
fma.rn.f32 %f309, %f1421, 0f3F000000, 0f3F000000;
fma.rn.f32 %f310, %f1422, 0f3F000000, 0f3F000000;
div.rn.f32 %f1423, %f1519, %f1414;
div.rn.f32 %f1424, %f1518, %f1416;
div.rn.f32 %f1425, %f1517, %f1418;
fma.rn.f32 %f311, %f1423, 0f3F000000, 0f3F000000;
fma.rn.f32 %f312, %f1424, 0f3F000000, 0f3F000000;
fma.rn.f32 %f313, %f1425, 0f3F000000, 0f3F000000;
ld.global.u32 %r189, [additive];
setp.eq.s32 %p153, %r189, 0;
mov.f32 %f1412, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs37, %f1412;}
// inline asm
@%p153 bra BB0_114;
mov.u64 %rd89, image_RNM0;
cvta.global.u64 %rd78, %rd89;
mov.u32 %r193, 8;
// inline asm
call (%rd77), _rt_buffer_get_64, (%rd78, %r28, %r193, %rd4, %rd5, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd77];
// inline asm
{ cvt.f32.f16 %f1426, %rs44;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1427, %rs45;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1428, %rs46;}
// inline asm
// inline asm
call (%rd83), _rt_buffer_get_64, (%rd78, %r28, %r193, %rd4, %rd5, %rd13, %rd13);
// inline asm
add.f32 %f1429, %f1528, %f1426;
add.f32 %f1430, %f1527, %f1427;
add.f32 %f1431, %f1526, %f1428;
// inline asm
{ cvt.rn.f16.f32 %rs43, %f1431;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs42, %f1430;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs41, %f1429;}
// inline asm
st.v4.u16 [%rd83], {%rs41, %rs42, %rs43, %rs37};
bra.uni BB0_115;
BB0_114:
mov.u64 %rd96, image_RNM0;
cvta.global.u64 %rd91, %rd96;
mov.u32 %r195, 8;
// inline asm
call (%rd90), _rt_buffer_get_64, (%rd91, %r28, %r195, %rd4, %rd5, %rd13, %rd13);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs50, %f1526;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs49, %f1527;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs48, %f1528;}
// inline asm
st.v4.u16 [%rd90], {%rs48, %rs49, %rs50, %rs37};
BB0_115:
ld.global.u32 %r196, [additive];
setp.eq.s32 %p154, %r196, 0;
// inline asm
{ cvt.rn.f16.f32 %rs51, %f1412;}
// inline asm
@%p154 bra BB0_117;
mov.u64 %rd109, image_RNM1;
cvta.global.u64 %rd98, %rd109;
mov.u32 %r200, 8;
// inline asm
call (%rd97), _rt_buffer_get_64, (%rd98, %r28, %r200, %rd4, %rd5, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs58, %rs59, %rs60, %rs61}, [%rd97];
// inline asm
{ cvt.f32.f16 %f1436, %rs58;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1437, %rs59;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1438, %rs60;}
// inline asm
// inline asm
call (%rd103), _rt_buffer_get_64, (%rd98, %r28, %r200, %rd4, %rd5, %rd13, %rd13);
// inline asm
add.f32 %f1439, %f305, %f1436;
add.f32 %f1440, %f306, %f1437;
add.f32 %f1441, %f307, %f1438;
// inline asm
{ cvt.rn.f16.f32 %rs57, %f1441;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs56, %f1440;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs55, %f1439;}
// inline asm
st.v4.u16 [%rd103], {%rs55, %rs56, %rs57, %rs51};
bra.uni BB0_118;
BB0_117:
mov.u64 %rd116, image_RNM1;
cvta.global.u64 %rd111, %rd116;
mov.u32 %r202, 8;
// inline asm
call (%rd110), _rt_buffer_get_64, (%rd111, %r28, %r202, %rd4, %rd5, %rd13, %rd13);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs64, %f307;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs63, %f306;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs62, %f305;}
// inline asm
st.v4.u16 [%rd110], {%rs62, %rs63, %rs64, %rs51};
BB0_118:
ld.global.u32 %r203, [additive];
setp.eq.s32 %p155, %r203, 0;
// inline asm
{ cvt.rn.f16.f32 %rs65, %f1412;}
// inline asm
@%p155 bra BB0_120;
mov.u64 %rd129, image_RNM2;
cvta.global.u64 %rd118, %rd129;
mov.u32 %r207, 8;
// inline asm
call (%rd117), _rt_buffer_get_64, (%rd118, %r28, %r207, %rd4, %rd5, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd117];
// inline asm
{ cvt.f32.f16 %f1446, %rs72;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1447, %rs73;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1448, %rs74;}
// inline asm
// inline asm
call (%rd123), _rt_buffer_get_64, (%rd118, %r28, %r207, %rd4, %rd5, %rd13, %rd13);
// inline asm
add.f32 %f1449, %f308, %f1446;
add.f32 %f1450, %f309, %f1447;
add.f32 %f1451, %f310, %f1448;
// inline asm
{ cvt.rn.f16.f32 %rs71, %f1451;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs70, %f1450;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs69, %f1449;}
// inline asm
st.v4.u16 [%rd123], {%rs69, %rs70, %rs71, %rs65};
bra.uni BB0_121;
BB0_120:
mov.u64 %rd136, image_RNM2;
cvta.global.u64 %rd131, %rd136;
mov.u32 %r209, 8;
// inline asm
call (%rd130), _rt_buffer_get_64, (%rd131, %r28, %r209, %rd4, %rd5, %rd13, %rd13);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs78, %f310;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs77, %f309;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs76, %f308;}
// inline asm
st.v4.u16 [%rd130], {%rs76, %rs77, %rs78, %rs65};
BB0_121:
ld.global.u32 %r210, [additive];
setp.eq.s32 %p156, %r210, 0;
// inline asm
{ cvt.rn.f16.f32 %rs79, %f1412;}
// inline asm
@%p156 bra BB0_123;
mov.u64 %rd149, image_RNM3;
cvta.global.u64 %rd138, %rd149;
mov.u32 %r214, 8;
// inline asm
call (%rd137), _rt_buffer_get_64, (%rd138, %r28, %r214, %rd4, %rd5, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs86, %rs87, %rs88, %rs89}, [%rd137];
// inline asm
{ cvt.f32.f16 %f1456, %rs86;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1457, %rs87;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1458, %rs88;}
// inline asm
// inline asm
call (%rd143), _rt_buffer_get_64, (%rd138, %r28, %r214, %rd4, %rd5, %rd13, %rd13);
// inline asm
add.f32 %f1459, %f311, %f1456;
add.f32 %f1460, %f312, %f1457;
add.f32 %f1461, %f313, %f1458;
// inline asm
{ cvt.rn.f16.f32 %rs85, %f1461;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs84, %f1460;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs83, %f1459;}
// inline asm
st.v4.u16 [%rd143], {%rs83, %rs84, %rs85, %rs79};
bra.uni BB0_144;
BB0_123:
mov.u64 %rd156, image_RNM3;
cvta.global.u64 %rd151, %rd156;
mov.u32 %r216, 8;
// inline asm
call (%rd150), _rt_buffer_get_64, (%rd151, %r28, %r216, %rd4, %rd5, %rd13, %rd13);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs92, %f313;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs91, %f312;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs90, %f311;}
// inline asm
st.v4.u16 [%rd150], {%rs90, %rs91, %rs92, %rs79};
BB0_144:
ret;
}