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

1950 lines
59 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[12];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<134>;
.reg .b16 %rs<63>;
.reg .f32 %f<1298>;
.reg .b32 %r<203>;
.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 %f1223, 0f00000000;
mov.f32 %f1224, %f1223;
mov.f32 %f1225, %f1223;
@%p8 bra BB0_2;
ld.u8 %rs8, [%rd8+1];
and.b16 %rs10, %rs6, 255;
cvt.rn.f32.u16 %f238, %rs10;
div.rn.f32 %f239, %f238, 0f437F0000;
fma.rn.f32 %f240, %f239, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f241, %rs8;
div.rn.f32 %f242, %f241, 0f437F0000;
fma.rn.f32 %f243, %f242, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f244, %rs5;
div.rn.f32 %f245, %f244, 0f437F0000;
fma.rn.f32 %f246, %f245, 0f40000000, 0fBF800000;
mul.f32 %f247, %f243, %f243;
fma.rn.f32 %f248, %f240, %f240, %f247;
fma.rn.f32 %f249, %f246, %f246, %f248;
sqrt.rn.f32 %f250, %f249;
rcp.rn.f32 %f251, %f250;
mul.f32 %f1223, %f240, %f251;
mul.f32 %f1224, %f243, %f251;
mul.f32 %f1225, %f246, %f251;
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, %f1224, 0f00000000;
setp.eq.f32 %p10, %f1223, 0f00000000;
and.pred %p11, %p10, %p9;
setp.eq.f32 %p12, %f1225, 0f00000000;
and.pred %p13, %p11, %p12;
@%p13 bra BB0_103;
bra.uni BB0_3;
BB0_103:
ld.global.u32 %r202, [imageEnabled];
and.b32 %r175, %r202, 1;
setp.eq.b32 %p128, %r175, 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 %r202, [imageEnabled];
BB0_105:
and.b32 %r178, %r202, 8;
setp.eq.s32 %p129, %r178, 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 %f1212, 0f00000000;
cvt.rzi.u32.f32 %r181, %f1212;
cvt.u16.u32 %rs44, %r181;
mov.u16 %rs45, 0;
st.v2.u8 [%rd91], {%rs44, %rs45};
ld.global.u32 %r202, [imageEnabled];
BB0_107:
and.b32 %r182, %r202, 4;
setp.eq.s32 %p130, %r182, 0;
@%p130 bra BB0_111;
ld.global.u32 %r183, [additive];
setp.eq.s32 %p131, %r183, 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 %r187, 8;
// inline asm
call (%rd98), _rt_buffer_get_64, (%rd99, %r28, %r187, %rd6, %rd7, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs52, %rs53, %rs54, %rs55}, [%rd98];
// inline asm
{ cvt.f32.f16 %f1213, %rs52;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1214, %rs53;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1215, %rs54;}
// inline asm
// inline asm
call (%rd104), _rt_buffer_get_64, (%rd99, %r28, %r187, %rd6, %rd7, %rd13, %rd13);
// inline asm
add.f32 %f1216, %f1213, 0f00000000;
add.f32 %f1217, %f1214, 0f00000000;
add.f32 %f1218, %f1215, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs51, %f1218;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs50, %f1217;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs49, %f1216;}
// 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 %f259, %f7, 0f3456BF95;
mul.f32 %f260, %f8, 0f3456BF95;
mul.f32 %f261, %f9, 0f3456BF95;
abs.f32 %f262, %f1223;
div.rn.f32 %f263, %f259, %f262;
abs.f32 %f264, %f1224;
div.rn.f32 %f265, %f260, %f264;
abs.f32 %f266, %f1225;
div.rn.f32 %f267, %f261, %f266;
abs.f32 %f268, %f263;
abs.f32 %f269, %f265;
abs.f32 %f270, %f267;
mov.f32 %f271, 0f38D1B717;
max.f32 %f272, %f268, %f271;
max.f32 %f273, %f269, %f271;
max.f32 %f274, %f270, %f271;
fma.rn.f32 %f10, %f1223, %f272, %f7;
fma.rn.f32 %f11, %f1224, %f273, %f8;
fma.rn.f32 %f12, %f1225, %f274, %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 %p14, %r4, 0;
mov.f32 %f1226, 0f00000000;
mov.f32 %f18, %f1226;
mov.f32 %f19, %f1226;
mov.f32 %f20, %f1226;
mov.f32 %f1230, %f1226;
mov.f32 %f1231, %f1226;
mov.f32 %f1232, %f1226;
@%p14 bra BB0_49;
mov.f32 %f282, 0f40000000;
cvt.rzi.f32.f32 %f283, %f282;
add.f32 %f284, %f283, %f283;
mov.f32 %f285, 0f40800000;
sub.f32 %f286, %f285, %f284;
abs.f32 %f13, %f286;
mul.f32 %f14, %f10, 0f3456BF95;
mul.f32 %f15, %f11, 0f3456BF95;
mul.f32 %f16, %f12, 0f3456BF95;
mov.f32 %f281, 0f00000000;
mov.u32 %r194, 0;
abs.f32 %f472, %f14;
abs.f32 %f473, %f15;
max.f32 %f474, %f472, %f473;
abs.f32 %f475, %f16;
max.f32 %f476, %f474, %f475;
mov.f32 %f1226, %f281;
mov.f32 %f18, %f281;
mov.f32 %f19, %f281;
mov.f32 %f20, %f281;
mov.f32 %f1230, %f281;
mov.f32 %f1231, %f281;
mov.f32 %f1232, %f281;
BB0_5:
cvt.u64.u32 %rd30, %r194;
// inline asm
call (%rd28), _rt_buffer_get_64, (%rd25, %r45, %r46, %rd30, %rd13, %rd13, %rd13);
// inline asm
ld.v4.f32 {%f289, %f290, %f291, %f292}, [%rd28+80];
ld.v4.f32 {%f293, %f294, %f295, %f296}, [%rd28+64];
ld.v4.f32 {%f297, %f298, %f299, %f300}, [%rd28+48];
ld.v4.f32 {%f301, %f1241, %f1242, %f304}, [%rd28+32];
ld.v4.f32 {%f305, %f306, %f307, %f308}, [%rd28+16];
ld.v4.f32 {%f309, %f310, %f311, %f312}, [%rd28];
mov.b32 %r6, %f292;
sub.f32 %f48, %f310, %f7;
sub.f32 %f50, %f311, %f8;
sub.f32 %f52, %f312, %f9;
mul.f32 %f314, %f50, %f50;
fma.rn.f32 %f315, %f48, %f48, %f314;
fma.rn.f32 %f316, %f52, %f52, %f315;
sqrt.rn.f32 %f53, %f316;
rcp.rn.f32 %f317, %f53;
mul.f32 %f54, %f48, %f317;
mul.f32 %f55, %f50, %f317;
mul.f32 %f56, %f52, %f317;
mul.f32 %f57, %f53, %f308;
abs.f32 %f58, %f57;
setp.lt.f32 %p15, %f58, 0f00800000;
mul.f32 %f318, %f58, 0f4B800000;
selp.f32 %f319, 0fC3170000, 0fC2FE0000, %p15;
selp.f32 %f320, %f318, %f58, %p15;
mov.b32 %r54, %f320;
and.b32 %r55, %r54, 8388607;
or.b32 %r56, %r55, 1065353216;
mov.b32 %f321, %r56;
shr.u32 %r57, %r54, 23;
cvt.rn.f32.u32 %f322, %r57;
add.f32 %f323, %f319, %f322;
setp.gt.f32 %p16, %f321, 0f3FB504F3;
mul.f32 %f324, %f321, 0f3F000000;
add.f32 %f325, %f323, 0f3F800000;
selp.f32 %f326, %f324, %f321, %p16;
selp.f32 %f327, %f325, %f323, %p16;
add.f32 %f328, %f326, 0fBF800000;
add.f32 %f288, %f326, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f287,%f288;
// inline asm
add.f32 %f329, %f328, %f328;
mul.f32 %f330, %f287, %f329;
mul.f32 %f331, %f330, %f330;
mov.f32 %f332, 0f3C4CAF63;
mov.f32 %f333, 0f3B18F0FE;
fma.rn.f32 %f334, %f333, %f331, %f332;
mov.f32 %f335, 0f3DAAAABD;
fma.rn.f32 %f336, %f334, %f331, %f335;
mul.rn.f32 %f337, %f336, %f331;
mul.rn.f32 %f338, %f337, %f330;
sub.f32 %f339, %f328, %f330;
neg.f32 %f340, %f330;
add.f32 %f341, %f339, %f339;
fma.rn.f32 %f342, %f340, %f328, %f341;
mul.rn.f32 %f343, %f287, %f342;
add.f32 %f344, %f338, %f330;
sub.f32 %f345, %f330, %f344;
add.f32 %f346, %f338, %f345;
add.f32 %f347, %f343, %f346;
add.f32 %f348, %f344, %f347;
sub.f32 %f349, %f344, %f348;
add.f32 %f350, %f347, %f349;
mov.f32 %f351, 0f3F317200;
mul.rn.f32 %f352, %f327, %f351;
mov.f32 %f353, 0f35BFBE8E;
mul.rn.f32 %f354, %f327, %f353;
add.f32 %f355, %f352, %f348;
sub.f32 %f356, %f352, %f355;
add.f32 %f357, %f348, %f356;
add.f32 %f358, %f350, %f357;
add.f32 %f359, %f354, %f358;
add.f32 %f360, %f355, %f359;
sub.f32 %f361, %f355, %f360;
add.f32 %f362, %f359, %f361;
mul.rn.f32 %f59, %f285, %f360;
neg.f32 %f364, %f59;
fma.rn.f32 %f365, %f285, %f360, %f364;
fma.rn.f32 %f366, %f285, %f362, %f365;
fma.rn.f32 %f60, %f281, %f360, %f366;
add.rn.f32 %f61, %f59, %f60;
mov.b32 %r58, %f61;
setp.eq.s32 %p1, %r58, 1118925336;
add.s32 %r59, %r58, -1;
mov.b32 %f368, %r59;
selp.f32 %f369, %f368, %f61, %p1;
mul.f32 %f370, %f369, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f371, %f370;
mov.f32 %f372, 0fBF317200;
fma.rn.f32 %f373, %f371, %f372, %f369;
mov.f32 %f374, 0fB5BFBE8E;
fma.rn.f32 %f375, %f371, %f374, %f373;
mul.f32 %f376, %f375, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f377, %f376;
add.f32 %f378, %f371, 0f00000000;
ex2.approx.f32 %f379, %f378;
mul.f32 %f380, %f377, %f379;
setp.lt.f32 %p17, %f369, 0fC2D20000;
selp.f32 %f381, 0f00000000, %f380, %p17;
setp.gt.f32 %p18, %f369, 0f42D20000;
selp.f32 %f1233, 0f7F800000, %f381, %p18;
setp.eq.f32 %p19, %f1233, 0f7F800000;
@%p19 bra BB0_7;
neg.f32 %f382, %f61;
add.rn.f32 %f383, %f59, %f382;
add.rn.f32 %f384, %f383, %f60;
add.f32 %f385, %f384, 0f37000000;
selp.f32 %f386, %f385, %f384, %p1;
fma.rn.f32 %f1233, %f1233, %f386, %f1233;
BB0_7:
setp.lt.f32 %p20, %f57, 0f00000000;
setp.eq.f32 %p21, %f13, 0f3F800000;
and.pred %p2, %p20, %p21;
mov.b32 %r60, %f1233;
xor.b32 %r61, %r60, -2147483648;
mov.b32 %f387, %r61;
selp.f32 %f1235, %f387, %f1233, %p2;
setp.eq.f32 %p22, %f57, 0f00000000;
@%p22 bra BB0_10;
bra.uni BB0_8;
BB0_10:
add.f32 %f390, %f57, %f57;
selp.f32 %f1235, %f390, 0f00000000, %p21;
bra.uni BB0_11;
BB0_8:
setp.geu.f32 %p23, %f57, 0f00000000;
@%p23 bra BB0_11;
cvt.rzi.f32.f32 %f389, %f285;
setp.neu.f32 %p24, %f389, 0f40800000;
selp.f32 %f1235, 0f7FFFFFFF, %f1235, %p24;
BB0_11:
add.f32 %f391, %f58, 0f40800000;
mov.b32 %r62, %f391;
setp.lt.s32 %p26, %r62, 2139095040;
@%p26 bra BB0_16;
setp.gtu.f32 %p27, %f58, 0f7F800000;
@%p27 bra BB0_15;
bra.uni BB0_13;
BB0_15:
add.f32 %f1235, %f57, 0f40800000;
bra.uni BB0_16;
BB0_13:
setp.neu.f32 %p28, %f58, 0f7F800000;
@%p28 bra BB0_16;
selp.f32 %f1235, 0fFF800000, 0f7F800000, %p2;
BB0_16:
mul.f32 %f392, %f53, %f306;
mov.f32 %f1265, 0f3F800000;
sub.f32 %f394, %f1265, %f1235;
setp.eq.f32 %p29, %f57, 0f3F800000;
selp.f32 %f395, 0f00000000, %f394, %p29;
cvt.sat.f32.f32 %f396, %f395;
fma.rn.f32 %f397, %f392, %f392, %f307;
div.rn.f32 %f87, %f396, %f397;
mul.f32 %f398, %f1224, %f55;
fma.rn.f32 %f399, %f1223, %f54, %f398;
fma.rn.f32 %f400, %f1225, %f56, %f399;
ld.global.u32 %r63, [ignoreNormal];
setp.eq.s32 %p30, %r63, 0;
selp.f32 %f401, %f400, 0f3F800000, %p30;
cvt.sat.f32.f32 %f88, %f401;
setp.eq.f32 %p31, %f309, 0f3F800000;
@%p31 bra BB0_23;
bra.uni BB0_17;
BB0_23:
setp.leu.f32 %p35, %f304, 0f00000000;
@%p35 bra BB0_19;
setp.geu.f32 %p36, %f300, 0f00000000;
@%p36 bra BB0_25;
sub.f32 %f1222, %f312, %f9;
sub.f32 %f1221, %f310, %f7;
sub.f32 %f1220, %f311, %f8;
mul.f32 %f424, %f290, %f1220;
fma.rn.f32 %f425, %f289, %f1221, %f424;
fma.rn.f32 %f426, %f291, %f1222, %f425;
rcp.rn.f32 %f427, %f426;
mul.f32 %f1236, %f1221, %f427;
mul.f32 %f1237, %f1220, %f427;
mul.f32 %f1238, %f1222, %f427;
neg.f32 %f300, %f300;
bra.uni BB0_27;
BB0_17:
setp.eq.f32 %p32, %f309, 0f40000000;
@%p32 bra BB0_21;
bra.uni BB0_18;
BB0_21:
setp.leu.f32 %p34, %f304, 0f00000000;
@%p34 bra BB0_19;
mul.f32 %f418, %f298, %f55;
fma.rn.f32 %f419, %f297, %f54, %f418;
mul.f32 %f420, %f294, %f55;
fma.rn.f32 %f421, %f293, %f54, %f420;
mul.f32 %f422, %f290, %f55;
fma.rn.f32 %f423, %f289, %f54, %f422;
fma.rn.f32 %f415, %f299, %f56, %f419;
fma.rn.f32 %f416, %f295, %f56, %f421;
fma.rn.f32 %f417, %f291, %f56, %f423;
cvt.rzi.s32.f32 %r64, %f304;
mov.u32 %r65, 6;
mov.u32 %r66, 0;
// inline asm
call (%f411, %f412, %f413, %f414), _rt_texture_get_base_id, (%r64, %r65, %f415, %f416, %f417, %r66);
// inline asm
mul.f32 %f1240, %f301, %f411;
mul.f32 %f1241, %f1241, %f412;
mul.f32 %f1242, %f1242, %f413;
bra.uni BB0_28;
BB0_18:
setp.neu.f32 %p33, %f309, 0f40800000;
@%p33 bra BB0_19;
mul.f32 %f402, %f289, %f54;
mul.f32 %f403, %f290, %f55;
neg.f32 %f404, %f403;
sub.f32 %f405, %f404, %f402;
mul.f32 %f406, %f291, %f56;
sub.f32 %f407, %f405, %f406;
fma.rn.f32 %f408, %f304, %f407, %f300;
cvt.sat.f32.f32 %f409, %f408;
mul.f32 %f410, %f409, %f409;
mul.f32 %f1243, %f87, %f410;
mov.f32 %f1240, %f301;
bra.uni BB0_29;
BB0_19:
mov.f32 %f1240, %f301;
bra.uni BB0_28;
BB0_25:
mov.f32 %f1236, %f54;
mov.f32 %f1237, %f55;
mov.f32 %f1238, %f56;
BB0_27:
mul.f32 %f436, %f291, %f56;
mul.f32 %f437, %f290, %f55;
neg.f32 %f438, %f437;
mul.f32 %f439, %f289, %f54;
sub.f32 %f440, %f438, %f439;
sub.f32 %f441, %f440, %f436;
setp.gt.f32 %p37, %f441, 0f00000000;
selp.f32 %f442, 0f3F800000, 0f00000000, %p37;
mul.f32 %f443, %f298, %f1237;
fma.rn.f32 %f444, %f297, %f1236, %f443;
mul.f32 %f445, %f294, %f1237;
fma.rn.f32 %f446, %f293, %f1236, %f445;
fma.rn.f32 %f447, %f299, %f1238, %f444;
fma.rn.f32 %f448, %f295, %f1238, %f446;
fma.rn.f32 %f449, %f300, %f447, 0f3F000000;
mov.f32 %f450, 0f3F800000;
sub.f32 %f432, %f450, %f449;
fma.rn.f32 %f433, %f300, %f448, 0f3F000000;
cvt.rzi.s32.f32 %r67, %f304;
mov.f32 %f435, 0f00000000;
// inline asm
call (%f428, %f429, %f430, %f431), _rt_texture_get_f_id, (%r67, %r28, %f432, %f433, %f435, %f435);
// inline asm
mul.f32 %f451, %f442, %f428;
mul.f32 %f452, %f442, %f429;
mul.f32 %f453, %f442, %f430;
mul.f32 %f1240, %f301, %f451;
mul.f32 %f1241, %f1241, %f452;
mul.f32 %f1242, %f1242, %f453;
BB0_28:
mov.f32 %f1243, %f87;
BB0_29:
max.f32 %f460, %f1240, %f1241;
max.f32 %f461, %f460, %f1242;
mul.f32 %f108, %f88, %f1243;
mul.f32 %f462, %f108, %f461;
setp.lt.f32 %p39, %f462, 0f3727C5AC;
mov.pred %p133, -1;
mov.f32 %f115, 0f00000000;
mov.f32 %f116, %f115;
mov.f32 %f117, %f115;
mov.f32 %f118, %f115;
mov.f32 %f119, %f115;
mov.f32 %f120, %f115;
@%p39 bra BB0_31;
mul.f32 %f115, %f1240, %f108;
mul.f32 %f116, %f1241, %f108;
mul.f32 %f117, %f1242, %f108;
ld.global.u8 %rs12, [imageEnabled];
and.b16 %rs13, %rs12, 64;
setp.eq.s16 %p41, %rs13, 0;
selp.f32 %f118, 0f00000000, %f54, %p41;
selp.f32 %f119, 0f00000000, %f55, %p41;
selp.f32 %f120, 0f00000000, %f56, %p41;
mov.pred %p133, 0;
BB0_31:
@%p133 bra BB0_48;
setp.eq.s32 %p42, %r6, 0;
mov.u16 %rs62, 0;
mov.f32 %f1266, %f1265;
mov.f32 %f1267, %f1265;
@%p42 bra BB0_43;
abs.s32 %r8, %r6;
mov.f32 %f1262, 0f00000000;
setp.lt.s32 %p43, %r8, 1;
mov.f32 %f1263, %f1262;
mov.f32 %f1264, %f1262;
@%p43 bra BB0_42;
max.f32 %f122, %f476, %f271;
and.b32 %r9, %r8, 3;
setp.eq.s32 %p44, %r9, 0;
add.u64 %rd35, %SP, 0;
cvta.to.local.u64 %rd2, %rd35;
mov.f32 %f1262, 0f00000000;
mov.u32 %r198, 0;
mov.f32 %f1263, %f1262;
mov.f32 %f1264, %f1262;
@%p44 bra BB0_40;
setp.eq.s32 %p45, %r9, 1;
mov.f32 %f1253, 0f00000000;
mov.u32 %r196, 0;
mov.f32 %f1254, %f1253;
mov.f32 %f1255, %f1253;
@%p45 bra BB0_39;
setp.eq.s32 %p46, %r9, 2;
mov.f32 %f1250, 0f00000000;
mov.u32 %r195, 0;
mov.f32 %f1251, %f1250;
mov.f32 %f1252, %f1250;
@%p46 bra BB0_38;
sub.f32 %f492, %f310, %f305;
sub.f32 %f493, %f311, %f305;
sub.f32 %f494, %f312, %f305;
sub.f32 %f495, %f492, %f7;
sub.f32 %f496, %f493, %f8;
sub.f32 %f497, %f494, %f9;
mul.f32 %f498, %f496, %f496;
fma.rn.f32 %f499, %f495, %f495, %f498;
fma.rn.f32 %f500, %f497, %f497, %f499;
sqrt.rn.f32 %f491, %f500;
rcp.rn.f32 %f501, %f491;
mul.f32 %f487, %f501, %f495;
mul.f32 %f488, %f501, %f496;
mul.f32 %f489, %f501, %f497;
ld.global.u32 %r76, [imageEnabled];
and.b32 %r77, %r76, 32;
setp.eq.s32 %p47, %r77, 0;
selp.f32 %f502, 0f3F800000, 0f41200000, %p47;
mul.f32 %f490, %f502, %f122;
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, %f487, %f488, %f489, %r45, %f490, %f491, %rd35, %r44);
// inline asm
ld.local.f32 %f503, [%rd2];
add.f32 %f1252, %f503, 0f00000000;
ld.local.f32 %f504, [%rd2+4];
add.f32 %f1251, %f504, 0f00000000;
ld.local.f32 %f505, [%rd2+8];
add.f32 %f1250, %f505, 0f00000000;
mov.u32 %r195, %r45;
BB0_38:
cvt.rn.f32.s32 %f514, %r195;
mul.f32 %f515, %f514, 0f3DD32618;
cvt.rmi.f32.f32 %f516, %f515;
sub.f32 %f517, %f515, %f516;
mul.f32 %f518, %f514, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f519, %f518;
sub.f32 %f520, %f518, %f519;
mul.f32 %f521, %f514, 0f3DC74539;
cvt.rmi.f32.f32 %f522, %f521;
sub.f32 %f523, %f521, %f522;
add.f32 %f524, %f520, 0f4199851F;
add.f32 %f525, %f523, 0f4199851F;
add.f32 %f526, %f517, 0f4199851F;
mul.f32 %f527, %f520, %f525;
fma.rn.f32 %f528, %f517, %f524, %f527;
fma.rn.f32 %f529, %f526, %f523, %f528;
add.f32 %f530, %f517, %f529;
add.f32 %f531, %f520, %f529;
add.f32 %f532, %f523, %f529;
add.f32 %f533, %f530, %f531;
mul.f32 %f534, %f532, %f533;
cvt.rmi.f32.f32 %f535, %f534;
sub.f32 %f536, %f534, %f535;
add.f32 %f537, %f530, %f532;
mul.f32 %f538, %f531, %f537;
cvt.rmi.f32.f32 %f539, %f538;
sub.f32 %f540, %f538, %f539;
add.f32 %f541, %f531, %f532;
mul.f32 %f542, %f530, %f541;
cvt.rmi.f32.f32 %f543, %f542;
sub.f32 %f544, %f542, %f543;
fma.rn.f32 %f545, %f536, 0f40000000, 0fBF800000;
fma.rn.f32 %f546, %f540, 0f40000000, 0fBF800000;
fma.rn.f32 %f547, %f544, 0f40000000, 0fBF800000;
fma.rn.f32 %f548, %f305, %f545, %f310;
fma.rn.f32 %f549, %f305, %f546, %f311;
fma.rn.f32 %f550, %f305, %f547, %f312;
sub.f32 %f551, %f548, %f7;
sub.f32 %f552, %f549, %f8;
sub.f32 %f553, %f550, %f9;
mul.f32 %f554, %f552, %f552;
fma.rn.f32 %f555, %f551, %f551, %f554;
fma.rn.f32 %f556, %f553, %f553, %f555;
sqrt.rn.f32 %f513, %f556;
rcp.rn.f32 %f557, %f513;
mul.f32 %f509, %f557, %f551;
mul.f32 %f510, %f557, %f552;
mul.f32 %f511, %f557, %f553;
ld.global.u32 %r82, [imageEnabled];
and.b32 %r83, %r82, 32;
setp.eq.s32 %p48, %r83, 0;
selp.f32 %f558, 0f3F800000, 0f41200000, %p48;
mul.f32 %f512, %f558, %f122;
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, %f509, %f510, %f511, %r45, %f512, %f513, %rd35, %r44);
// inline asm
ld.local.f32 %f559, [%rd2];
add.f32 %f1255, %f1252, %f559;
ld.local.f32 %f560, [%rd2+4];
add.f32 %f1254, %f1251, %f560;
ld.local.f32 %f561, [%rd2+8];
add.f32 %f1253, %f1250, %f561;
add.s32 %r196, %r195, 1;
BB0_39:
cvt.rn.f32.s32 %f570, %r196;
mul.f32 %f571, %f570, 0f3DD32618;
cvt.rmi.f32.f32 %f572, %f571;
sub.f32 %f573, %f571, %f572;
mul.f32 %f574, %f570, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f575, %f574;
sub.f32 %f576, %f574, %f575;
mul.f32 %f577, %f570, 0f3DC74539;
cvt.rmi.f32.f32 %f578, %f577;
sub.f32 %f579, %f577, %f578;
add.f32 %f580, %f576, 0f4199851F;
add.f32 %f581, %f579, 0f4199851F;
add.f32 %f582, %f573, 0f4199851F;
mul.f32 %f583, %f576, %f581;
fma.rn.f32 %f584, %f573, %f580, %f583;
fma.rn.f32 %f585, %f582, %f579, %f584;
add.f32 %f586, %f573, %f585;
add.f32 %f587, %f576, %f585;
add.f32 %f588, %f579, %f585;
add.f32 %f589, %f586, %f587;
mul.f32 %f590, %f588, %f589;
cvt.rmi.f32.f32 %f591, %f590;
sub.f32 %f592, %f590, %f591;
add.f32 %f593, %f586, %f588;
mul.f32 %f594, %f587, %f593;
cvt.rmi.f32.f32 %f595, %f594;
sub.f32 %f596, %f594, %f595;
add.f32 %f597, %f587, %f588;
mul.f32 %f598, %f586, %f597;
cvt.rmi.f32.f32 %f599, %f598;
sub.f32 %f600, %f598, %f599;
fma.rn.f32 %f601, %f592, 0f40000000, 0fBF800000;
fma.rn.f32 %f602, %f596, 0f40000000, 0fBF800000;
fma.rn.f32 %f603, %f600, 0f40000000, 0fBF800000;
fma.rn.f32 %f604, %f305, %f601, %f310;
fma.rn.f32 %f605, %f305, %f602, %f311;
fma.rn.f32 %f606, %f305, %f603, %f312;
sub.f32 %f607, %f604, %f7;
sub.f32 %f608, %f605, %f8;
sub.f32 %f609, %f606, %f9;
mul.f32 %f610, %f608, %f608;
fma.rn.f32 %f611, %f607, %f607, %f610;
fma.rn.f32 %f612, %f609, %f609, %f611;
sqrt.rn.f32 %f569, %f612;
rcp.rn.f32 %f613, %f569;
mul.f32 %f565, %f613, %f607;
mul.f32 %f566, %f613, %f608;
mul.f32 %f567, %f613, %f609;
ld.global.u32 %r88, [imageEnabled];
and.b32 %r89, %r88, 32;
setp.eq.s32 %p49, %r89, 0;
selp.f32 %f614, 0f3F800000, 0f41200000, %p49;
mul.f32 %f568, %f614, %f122;
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, %f565, %f566, %f567, %r86, %f568, %f569, %rd35, %r44);
// inline asm
ld.local.f32 %f615, [%rd2];
add.f32 %f1264, %f1255, %f615;
ld.local.f32 %f616, [%rd2+4];
add.f32 %f1263, %f1254, %f616;
ld.local.f32 %f617, [%rd2+8];
add.f32 %f1262, %f1253, %f617;
add.s32 %r198, %r196, 1;
BB0_40:
setp.lt.u32 %p50, %r8, 4;
@%p50 bra BB0_42;
BB0_41:
cvt.rn.f32.s32 %f650, %r198;
mul.f32 %f651, %f650, 0f3DD32618;
cvt.rmi.f32.f32 %f652, %f651;
sub.f32 %f653, %f651, %f652;
mul.f32 %f654, %f650, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f655, %f654;
sub.f32 %f656, %f654, %f655;
mul.f32 %f657, %f650, 0f3DC74539;
cvt.rmi.f32.f32 %f658, %f657;
sub.f32 %f659, %f657, %f658;
add.f32 %f660, %f656, 0f4199851F;
add.f32 %f661, %f659, 0f4199851F;
add.f32 %f662, %f653, 0f4199851F;
mul.f32 %f663, %f656, %f661;
fma.rn.f32 %f664, %f653, %f660, %f663;
fma.rn.f32 %f665, %f662, %f659, %f664;
add.f32 %f666, %f653, %f665;
add.f32 %f667, %f656, %f665;
add.f32 %f668, %f659, %f665;
add.f32 %f669, %f666, %f667;
mul.f32 %f670, %f668, %f669;
cvt.rmi.f32.f32 %f671, %f670;
sub.f32 %f672, %f670, %f671;
add.f32 %f673, %f666, %f668;
mul.f32 %f674, %f667, %f673;
cvt.rmi.f32.f32 %f675, %f674;
sub.f32 %f676, %f674, %f675;
add.f32 %f677, %f667, %f668;
mul.f32 %f678, %f666, %f677;
cvt.rmi.f32.f32 %f679, %f678;
sub.f32 %f680, %f678, %f679;
fma.rn.f32 %f681, %f672, 0f40000000, 0fBF800000;
fma.rn.f32 %f682, %f676, 0f40000000, 0fBF800000;
fma.rn.f32 %f683, %f680, 0f40000000, 0fBF800000;
fma.rn.f32 %f684, %f305, %f681, %f310;
fma.rn.f32 %f685, %f305, %f682, %f311;
fma.rn.f32 %f686, %f305, %f683, %f312;
sub.f32 %f687, %f684, %f7;
sub.f32 %f688, %f685, %f8;
sub.f32 %f689, %f686, %f9;
mul.f32 %f690, %f688, %f688;
fma.rn.f32 %f691, %f687, %f687, %f690;
fma.rn.f32 %f692, %f689, %f689, %f691;
sqrt.rn.f32 %f625, %f692;
rcp.rn.f32 %f693, %f625;
mul.f32 %f621, %f693, %f687;
mul.f32 %f622, %f693, %f688;
mul.f32 %f623, %f693, %f689;
ld.global.u32 %r103, [imageEnabled];
and.b32 %r104, %r103, 32;
setp.eq.s32 %p51, %r104, 0;
selp.f32 %f694, 0f3F800000, 0f41200000, %p51;
mul.f32 %f624, %f694, %f122;
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, %f621, %f622, %f623, %r101, %f624, %f625, %rd35, %r44);
// inline asm
ld.local.f32 %f695, [%rd2];
add.f32 %f696, %f1264, %f695;
ld.local.f32 %f697, [%rd2+4];
add.f32 %f698, %f1263, %f697;
ld.local.f32 %f699, [%rd2+8];
add.f32 %f700, %f1262, %f699;
add.s32 %r106, %r198, 1;
cvt.rn.f32.s32 %f701, %r106;
mul.f32 %f702, %f701, 0f3DD32618;
cvt.rmi.f32.f32 %f703, %f702;
sub.f32 %f704, %f702, %f703;
mul.f32 %f705, %f701, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f706, %f705;
sub.f32 %f707, %f705, %f706;
mul.f32 %f708, %f701, 0f3DC74539;
cvt.rmi.f32.f32 %f709, %f708;
sub.f32 %f710, %f708, %f709;
add.f32 %f711, %f707, 0f4199851F;
add.f32 %f712, %f710, 0f4199851F;
add.f32 %f713, %f704, 0f4199851F;
mul.f32 %f714, %f707, %f712;
fma.rn.f32 %f715, %f704, %f711, %f714;
fma.rn.f32 %f716, %f713, %f710, %f715;
add.f32 %f717, %f704, %f716;
add.f32 %f718, %f707, %f716;
add.f32 %f719, %f710, %f716;
add.f32 %f720, %f717, %f718;
mul.f32 %f721, %f719, %f720;
cvt.rmi.f32.f32 %f722, %f721;
sub.f32 %f723, %f721, %f722;
add.f32 %f724, %f717, %f719;
mul.f32 %f725, %f718, %f724;
cvt.rmi.f32.f32 %f726, %f725;
sub.f32 %f727, %f725, %f726;
add.f32 %f728, %f718, %f719;
mul.f32 %f729, %f717, %f728;
cvt.rmi.f32.f32 %f730, %f729;
sub.f32 %f731, %f729, %f730;
fma.rn.f32 %f732, %f723, 0f40000000, 0fBF800000;
fma.rn.f32 %f733, %f727, 0f40000000, 0fBF800000;
fma.rn.f32 %f734, %f731, 0f40000000, 0fBF800000;
fma.rn.f32 %f735, %f305, %f732, %f310;
fma.rn.f32 %f736, %f305, %f733, %f311;
fma.rn.f32 %f737, %f305, %f734, %f312;
sub.f32 %f738, %f735, %f7;
sub.f32 %f739, %f736, %f8;
sub.f32 %f740, %f737, %f9;
mul.f32 %f741, %f739, %f739;
fma.rn.f32 %f742, %f738, %f738, %f741;
fma.rn.f32 %f743, %f740, %f740, %f742;
sqrt.rn.f32 %f633, %f743;
rcp.rn.f32 %f744, %f633;
mul.f32 %f629, %f744, %f738;
mul.f32 %f630, %f744, %f739;
mul.f32 %f631, %f744, %f740;
ld.global.u32 %r107, [imageEnabled];
and.b32 %r108, %r107, 32;
setp.eq.s32 %p52, %r108, 0;
selp.f32 %f745, 0f3F800000, 0f41200000, %p52;
mul.f32 %f632, %f745, %f122;
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, %f629, %f630, %f631, %r101, %f632, %f633, %rd35, %r44);
// inline asm
ld.local.f32 %f746, [%rd2];
add.f32 %f747, %f696, %f746;
ld.local.f32 %f748, [%rd2+4];
add.f32 %f749, %f698, %f748;
ld.local.f32 %f750, [%rd2+8];
add.f32 %f751, %f700, %f750;
add.s32 %r109, %r198, 2;
cvt.rn.f32.s32 %f752, %r109;
mul.f32 %f753, %f752, 0f3DD32618;
cvt.rmi.f32.f32 %f754, %f753;
sub.f32 %f755, %f753, %f754;
mul.f32 %f756, %f752, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f757, %f756;
sub.f32 %f758, %f756, %f757;
mul.f32 %f759, %f752, 0f3DC74539;
cvt.rmi.f32.f32 %f760, %f759;
sub.f32 %f761, %f759, %f760;
add.f32 %f762, %f758, 0f4199851F;
add.f32 %f763, %f761, 0f4199851F;
add.f32 %f764, %f755, 0f4199851F;
mul.f32 %f765, %f758, %f763;
fma.rn.f32 %f766, %f755, %f762, %f765;
fma.rn.f32 %f767, %f764, %f761, %f766;
add.f32 %f768, %f755, %f767;
add.f32 %f769, %f758, %f767;
add.f32 %f770, %f761, %f767;
add.f32 %f771, %f768, %f769;
mul.f32 %f772, %f770, %f771;
cvt.rmi.f32.f32 %f773, %f772;
sub.f32 %f774, %f772, %f773;
add.f32 %f775, %f768, %f770;
mul.f32 %f776, %f769, %f775;
cvt.rmi.f32.f32 %f777, %f776;
sub.f32 %f778, %f776, %f777;
add.f32 %f779, %f769, %f770;
mul.f32 %f780, %f768, %f779;
cvt.rmi.f32.f32 %f781, %f780;
sub.f32 %f782, %f780, %f781;
fma.rn.f32 %f783, %f774, 0f40000000, 0fBF800000;
fma.rn.f32 %f784, %f778, 0f40000000, 0fBF800000;
fma.rn.f32 %f785, %f782, 0f40000000, 0fBF800000;
fma.rn.f32 %f786, %f305, %f783, %f310;
fma.rn.f32 %f787, %f305, %f784, %f311;
fma.rn.f32 %f788, %f305, %f785, %f312;
sub.f32 %f789, %f786, %f7;
sub.f32 %f790, %f787, %f8;
sub.f32 %f791, %f788, %f9;
mul.f32 %f792, %f790, %f790;
fma.rn.f32 %f793, %f789, %f789, %f792;
fma.rn.f32 %f794, %f791, %f791, %f793;
sqrt.rn.f32 %f641, %f794;
rcp.rn.f32 %f795, %f641;
mul.f32 %f637, %f795, %f789;
mul.f32 %f638, %f795, %f790;
mul.f32 %f639, %f795, %f791;
ld.global.u32 %r110, [imageEnabled];
and.b32 %r111, %r110, 32;
setp.eq.s32 %p53, %r111, 0;
selp.f32 %f796, 0f3F800000, 0f41200000, %p53;
mul.f32 %f640, %f796, %f122;
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, %f637, %f638, %f639, %r101, %f640, %f641, %rd35, %r44);
// inline asm
ld.local.f32 %f797, [%rd2];
add.f32 %f798, %f747, %f797;
ld.local.f32 %f799, [%rd2+4];
add.f32 %f800, %f749, %f799;
ld.local.f32 %f801, [%rd2+8];
add.f32 %f802, %f751, %f801;
add.s32 %r112, %r198, 3;
cvt.rn.f32.s32 %f803, %r112;
mul.f32 %f804, %f803, 0f3DD32618;
cvt.rmi.f32.f32 %f805, %f804;
sub.f32 %f806, %f804, %f805;
mul.f32 %f807, %f803, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f808, %f807;
sub.f32 %f809, %f807, %f808;
mul.f32 %f810, %f803, 0f3DC74539;
cvt.rmi.f32.f32 %f811, %f810;
sub.f32 %f812, %f810, %f811;
add.f32 %f813, %f809, 0f4199851F;
add.f32 %f814, %f812, 0f4199851F;
add.f32 %f815, %f806, 0f4199851F;
mul.f32 %f816, %f809, %f814;
fma.rn.f32 %f817, %f806, %f813, %f816;
fma.rn.f32 %f818, %f815, %f812, %f817;
add.f32 %f819, %f806, %f818;
add.f32 %f820, %f809, %f818;
add.f32 %f821, %f812, %f818;
add.f32 %f822, %f819, %f820;
mul.f32 %f823, %f821, %f822;
cvt.rmi.f32.f32 %f824, %f823;
sub.f32 %f825, %f823, %f824;
add.f32 %f826, %f819, %f821;
mul.f32 %f827, %f820, %f826;
cvt.rmi.f32.f32 %f828, %f827;
sub.f32 %f829, %f827, %f828;
add.f32 %f830, %f820, %f821;
mul.f32 %f831, %f819, %f830;
cvt.rmi.f32.f32 %f832, %f831;
sub.f32 %f833, %f831, %f832;
fma.rn.f32 %f834, %f825, 0f40000000, 0fBF800000;
fma.rn.f32 %f835, %f829, 0f40000000, 0fBF800000;
fma.rn.f32 %f836, %f833, 0f40000000, 0fBF800000;
fma.rn.f32 %f837, %f305, %f834, %f310;
fma.rn.f32 %f838, %f305, %f835, %f311;
fma.rn.f32 %f839, %f305, %f836, %f312;
sub.f32 %f840, %f837, %f7;
sub.f32 %f841, %f838, %f8;
sub.f32 %f842, %f839, %f9;
mul.f32 %f843, %f841, %f841;
fma.rn.f32 %f844, %f840, %f840, %f843;
fma.rn.f32 %f845, %f842, %f842, %f844;
sqrt.rn.f32 %f649, %f845;
rcp.rn.f32 %f846, %f649;
mul.f32 %f645, %f846, %f840;
mul.f32 %f646, %f846, %f841;
mul.f32 %f647, %f846, %f842;
ld.global.u32 %r113, [imageEnabled];
and.b32 %r114, %r113, 32;
setp.eq.s32 %p54, %r114, 0;
selp.f32 %f847, 0f3F800000, 0f41200000, %p54;
mul.f32 %f648, %f847, %f122;
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, %f645, %f646, %f647, %r101, %f648, %f649, %rd35, %r44);
// inline asm
ld.local.f32 %f848, [%rd2];
add.f32 %f1264, %f798, %f848;
ld.local.f32 %f849, [%rd2+4];
add.f32 %f1263, %f800, %f849;
ld.local.f32 %f850, [%rd2+8];
add.f32 %f1262, %f802, %f850;
add.s32 %r198, %r198, 4;
setp.lt.s32 %p55, %r198, %r8;
@%p55 bra BB0_41;
BB0_42:
cvt.rn.f32.s32 %f851, %r8;
rcp.rn.f32 %f852, %f851;
mul.f32 %f1265, %f1264, %f852;
mul.f32 %f1266, %f1263, %f852;
mul.f32 %f1267, %f1262, %f852;
shr.u32 %r115, %r6, 31;
cvt.u16.u32 %rs62, %r115;
BB0_43:
fma.rn.f32 %f1232, %f115, %f1265, %f1232;
fma.rn.f32 %f1231, %f116, %f1266, %f1231;
fma.rn.f32 %f1230, %f117, %f1267, %f1230;
ld.global.u8 %rs15, [imageEnabled];
and.b16 %rs16, %rs15, 64;
setp.eq.s16 %p56, %rs16, 0;
@%p56 bra BB0_45;
mul.f32 %f853, %f116, 0f3F372474;
fma.rn.f32 %f854, %f115, 0f3E59999A, %f853;
fma.rn.f32 %f855, %f117, 0f3D93A92A, %f854;
fma.rn.f32 %f20, %f118, %f855, %f20;
fma.rn.f32 %f19, %f119, %f855, %f19;
fma.rn.f32 %f18, %f855, %f120, %f18;
BB0_45:
setp.eq.s16 %p57, %rs62, 0;
@%p57 bra BB0_47;
div.rn.f32 %f856, %f115, %f301;
div.rn.f32 %f857, %f856, %f87;
cvt.sat.f32.f32 %f858, %f857;
mul.f32 %f1265, %f1265, %f858;
BB0_47:
add.f32 %f1226, %f1226, %f1265;
BB0_48:
add.s32 %r194, %r194, 1;
setp.lt.u32 %p58, %r194, %r4;
@%p58 bra BB0_5;
BB0_49:
ld.global.u32 %r200, [imageEnabled];
and.b32 %r116, %r200, 8;
setp.eq.s32 %p59, %r116, 0;
@%p59 bra BB0_62;
cvt.sat.f32.f32 %f182, %f1226;
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 %f861, 0f3E68BA2E;
cvt.rzi.f32.f32 %f862, %f861;
fma.rn.f32 %f863, %f862, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f183, %f863;
abs.f32 %f184, %f182;
setp.lt.f32 %p60, %f184, 0f00800000;
mul.f32 %f864, %f184, 0f4B800000;
selp.f32 %f865, 0fC3170000, 0fC2FE0000, %p60;
selp.f32 %f866, %f864, %f184, %p60;
mov.b32 %r119, %f866;
and.b32 %r120, %r119, 8388607;
or.b32 %r121, %r120, 1065353216;
mov.b32 %f867, %r121;
shr.u32 %r122, %r119, 23;
cvt.rn.f32.u32 %f868, %r122;
add.f32 %f869, %f865, %f868;
setp.gt.f32 %p61, %f867, 0f3FB504F3;
mul.f32 %f870, %f867, 0f3F000000;
add.f32 %f871, %f869, 0f3F800000;
selp.f32 %f872, %f870, %f867, %p61;
selp.f32 %f873, %f871, %f869, %p61;
add.f32 %f874, %f872, 0fBF800000;
add.f32 %f860, %f872, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f859,%f860;
// inline asm
add.f32 %f875, %f874, %f874;
mul.f32 %f876, %f859, %f875;
mul.f32 %f877, %f876, %f876;
mov.f32 %f878, 0f3C4CAF63;
mov.f32 %f879, 0f3B18F0FE;
fma.rn.f32 %f880, %f879, %f877, %f878;
mov.f32 %f881, 0f3DAAAABD;
fma.rn.f32 %f882, %f880, %f877, %f881;
mul.rn.f32 %f883, %f882, %f877;
mul.rn.f32 %f884, %f883, %f876;
sub.f32 %f885, %f874, %f876;
neg.f32 %f886, %f876;
add.f32 %f887, %f885, %f885;
fma.rn.f32 %f888, %f886, %f874, %f887;
mul.rn.f32 %f889, %f859, %f888;
add.f32 %f890, %f884, %f876;
sub.f32 %f891, %f876, %f890;
add.f32 %f892, %f884, %f891;
add.f32 %f893, %f889, %f892;
add.f32 %f894, %f890, %f893;
sub.f32 %f895, %f890, %f894;
add.f32 %f896, %f893, %f895;
mov.f32 %f897, 0f3F317200;
mul.rn.f32 %f898, %f873, %f897;
mov.f32 %f899, 0f35BFBE8E;
mul.rn.f32 %f900, %f873, %f899;
add.f32 %f901, %f898, %f894;
sub.f32 %f902, %f898, %f901;
add.f32 %f903, %f894, %f902;
add.f32 %f904, %f896, %f903;
add.f32 %f905, %f900, %f904;
add.f32 %f906, %f901, %f905;
sub.f32 %f907, %f901, %f906;
add.f32 %f908, %f905, %f907;
mov.f32 %f909, 0f3EE8BA2E;
mul.rn.f32 %f910, %f909, %f906;
neg.f32 %f911, %f910;
fma.rn.f32 %f912, %f909, %f906, %f911;
fma.rn.f32 %f913, %f909, %f908, %f912;
mov.f32 %f914, 0f00000000;
fma.rn.f32 %f915, %f914, %f906, %f913;
add.rn.f32 %f916, %f910, %f915;
neg.f32 %f917, %f916;
add.rn.f32 %f918, %f910, %f917;
add.rn.f32 %f919, %f918, %f915;
mov.b32 %r123, %f916;
setp.eq.s32 %p62, %r123, 1118925336;
add.s32 %r124, %r123, -1;
mov.b32 %f920, %r124;
add.f32 %f921, %f919, 0f37000000;
selp.f32 %f922, %f920, %f916, %p62;
selp.f32 %f185, %f921, %f919, %p62;
mul.f32 %f923, %f922, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f924, %f923;
mov.f32 %f925, 0fBF317200;
fma.rn.f32 %f926, %f924, %f925, %f922;
mov.f32 %f927, 0fB5BFBE8E;
fma.rn.f32 %f928, %f924, %f927, %f926;
mul.f32 %f929, %f928, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f930, %f929;
add.f32 %f931, %f924, 0f00000000;
ex2.approx.f32 %f932, %f931;
mul.f32 %f933, %f930, %f932;
setp.lt.f32 %p63, %f922, 0fC2D20000;
selp.f32 %f934, 0f00000000, %f933, %p63;
setp.gt.f32 %p64, %f922, 0f42D20000;
selp.f32 %f1286, 0f7F800000, %f934, %p64;
setp.eq.f32 %p65, %f1286, 0f7F800000;
@%p65 bra BB0_52;
fma.rn.f32 %f1286, %f1286, %f185, %f1286;
BB0_52:
setp.lt.f32 %p66, %f182, 0f00000000;
setp.eq.f32 %p67, %f183, 0f3F800000;
and.pred %p4, %p66, %p67;
mov.b32 %r125, %f1286;
xor.b32 %r126, %r125, -2147483648;
mov.b32 %f935, %r126;
selp.f32 %f1288, %f935, %f1286, %p4;
setp.eq.f32 %p68, %f182, 0f00000000;
@%p68 bra BB0_55;
bra.uni BB0_53;
BB0_55:
add.f32 %f938, %f182, %f182;
selp.f32 %f1288, %f938, 0f00000000, %p67;
bra.uni BB0_56;
BB0_110:
mov.u64 %rd117, image_HDR;
cvta.global.u64 %rd112, %rd117;
mov.u32 %r189, 8;
// inline asm
call (%rd111), _rt_buffer_get_64, (%rd112, %r28, %r189, %rd6, %rd7, %rd13, %rd13);
// inline asm
mov.f32 %f1219, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs57, %f1219;}
// 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, %f182, 0f00000000;
@%p69 bra BB0_56;
cvt.rzi.f32.f32 %f937, %f909;
setp.neu.f32 %p70, %f937, 0f3EE8BA2E;
selp.f32 %f1288, 0f7FFFFFFF, %f1288, %p70;
BB0_56:
add.f32 %f939, %f184, 0f3EE8BA2E;
mov.b32 %r127, %f939;
setp.lt.s32 %p72, %r127, 2139095040;
@%p72 bra BB0_61;
setp.gtu.f32 %p73, %f184, 0f7F800000;
@%p73 bra BB0_60;
bra.uni BB0_58;
BB0_60:
add.f32 %f1288, %f182, 0f3EE8BA2E;
bra.uni BB0_61;
BB0_58:
setp.neu.f32 %p74, %f184, 0f7F800000;
@%p74 bra BB0_61;
selp.f32 %f1288, 0fFF800000, 0f7F800000, %p4;
BB0_61:
mul.f32 %f940, %f1288, 0f437F0000;
setp.eq.f32 %p75, %f182, 0f3F800000;
selp.f32 %f941, 0f437F0000, %f940, %p75;
cvt.rzi.u32.f32 %r128, %f941;
cvt.u16.u32 %rs17, %r128;
mov.u16 %rs18, 255;
st.v2.u8 [%rd43], {%rs17, %rs18};
ld.global.u32 %r200, [imageEnabled];
BB0_62:
and.b32 %r129, %r200, 1;
setp.eq.b32 %p76, %r129, 1;
@!%p76 bra BB0_97;
bra.uni BB0_63;
BB0_63:
mov.f32 %f944, 0f3E666666;
cvt.rzi.f32.f32 %f945, %f944;
fma.rn.f32 %f946, %f945, 0fC0000000, 0f3EE66666;
abs.f32 %f196, %f946;
abs.f32 %f197, %f1232;
setp.lt.f32 %p77, %f197, 0f00800000;
mul.f32 %f947, %f197, 0f4B800000;
selp.f32 %f948, 0fC3170000, 0fC2FE0000, %p77;
selp.f32 %f949, %f947, %f197, %p77;
mov.b32 %r130, %f949;
and.b32 %r131, %r130, 8388607;
or.b32 %r132, %r131, 1065353216;
mov.b32 %f950, %r132;
shr.u32 %r133, %r130, 23;
cvt.rn.f32.u32 %f951, %r133;
add.f32 %f952, %f948, %f951;
setp.gt.f32 %p78, %f950, 0f3FB504F3;
mul.f32 %f953, %f950, 0f3F000000;
add.f32 %f954, %f952, 0f3F800000;
selp.f32 %f955, %f953, %f950, %p78;
selp.f32 %f956, %f954, %f952, %p78;
add.f32 %f957, %f955, 0fBF800000;
add.f32 %f943, %f955, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f942,%f943;
// inline asm
add.f32 %f958, %f957, %f957;
mul.f32 %f959, %f942, %f958;
mul.f32 %f960, %f959, %f959;
mov.f32 %f961, 0f3C4CAF63;
mov.f32 %f962, 0f3B18F0FE;
fma.rn.f32 %f963, %f962, %f960, %f961;
mov.f32 %f964, 0f3DAAAABD;
fma.rn.f32 %f965, %f963, %f960, %f964;
mul.rn.f32 %f966, %f965, %f960;
mul.rn.f32 %f967, %f966, %f959;
sub.f32 %f968, %f957, %f959;
neg.f32 %f969, %f959;
add.f32 %f970, %f968, %f968;
fma.rn.f32 %f971, %f969, %f957, %f970;
mul.rn.f32 %f972, %f942, %f971;
add.f32 %f973, %f967, %f959;
sub.f32 %f974, %f959, %f973;
add.f32 %f975, %f967, %f974;
add.f32 %f976, %f972, %f975;
add.f32 %f977, %f973, %f976;
sub.f32 %f978, %f973, %f977;
add.f32 %f979, %f976, %f978;
mov.f32 %f980, 0f3F317200;
mul.rn.f32 %f981, %f956, %f980;
mov.f32 %f982, 0f35BFBE8E;
mul.rn.f32 %f983, %f956, %f982;
add.f32 %f984, %f981, %f977;
sub.f32 %f985, %f981, %f984;
add.f32 %f986, %f977, %f985;
add.f32 %f987, %f979, %f986;
add.f32 %f988, %f983, %f987;
add.f32 %f989, %f984, %f988;
sub.f32 %f990, %f984, %f989;
add.f32 %f991, %f988, %f990;
mov.f32 %f992, 0f3EE66666;
mul.rn.f32 %f993, %f992, %f989;
neg.f32 %f994, %f993;
fma.rn.f32 %f995, %f992, %f989, %f994;
fma.rn.f32 %f996, %f992, %f991, %f995;
mov.f32 %f997, 0f00000000;
fma.rn.f32 %f998, %f997, %f989, %f996;
add.rn.f32 %f999, %f993, %f998;
neg.f32 %f1000, %f999;
add.rn.f32 %f1001, %f993, %f1000;
add.rn.f32 %f1002, %f1001, %f998;
mov.b32 %r134, %f999;
setp.eq.s32 %p79, %r134, 1118925336;
add.s32 %r135, %r134, -1;
mov.b32 %f1003, %r135;
add.f32 %f1004, %f1002, 0f37000000;
selp.f32 %f1005, %f1003, %f999, %p79;
selp.f32 %f198, %f1004, %f1002, %p79;
mul.f32 %f1006, %f1005, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1007, %f1006;
mov.f32 %f1008, 0fBF317200;
fma.rn.f32 %f1009, %f1007, %f1008, %f1005;
mov.f32 %f1010, 0fB5BFBE8E;
fma.rn.f32 %f1011, %f1007, %f1010, %f1009;
mul.f32 %f1012, %f1011, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1013, %f1012;
add.f32 %f1014, %f1007, 0f00000000;
ex2.approx.f32 %f1015, %f1014;
mul.f32 %f1016, %f1013, %f1015;
setp.lt.f32 %p80, %f1005, 0fC2D20000;
selp.f32 %f1017, 0f00000000, %f1016, %p80;
setp.gt.f32 %p81, %f1005, 0f42D20000;
selp.f32 %f1289, 0f7F800000, %f1017, %p81;
setp.eq.f32 %p82, %f1289, 0f7F800000;
@%p82 bra BB0_65;
fma.rn.f32 %f1289, %f1289, %f198, %f1289;
BB0_65:
setp.lt.f32 %p83, %f1232, 0f00000000;
setp.eq.f32 %p84, %f196, 0f3F800000;
and.pred %p5, %p83, %p84;
mov.b32 %r136, %f1289;
xor.b32 %r137, %r136, -2147483648;
mov.b32 %f1018, %r137;
selp.f32 %f1291, %f1018, %f1289, %p5;
setp.eq.f32 %p85, %f1232, 0f00000000;
@%p85 bra BB0_68;
bra.uni BB0_66;
BB0_68:
add.f32 %f1021, %f1232, %f1232;
selp.f32 %f1291, %f1021, 0f00000000, %p84;
bra.uni BB0_69;
BB0_66:
setp.geu.f32 %p86, %f1232, 0f00000000;
@%p86 bra BB0_69;
cvt.rzi.f32.f32 %f1020, %f992;
setp.neu.f32 %p87, %f1020, 0f3EE66666;
selp.f32 %f1291, 0f7FFFFFFF, %f1291, %p87;
BB0_69:
add.f32 %f1022, %f197, 0f3EE66666;
mov.b32 %r138, %f1022;
setp.lt.s32 %p89, %r138, 2139095040;
@%p89 bra BB0_74;
setp.gtu.f32 %p90, %f197, 0f7F800000;
@%p90 bra BB0_73;
bra.uni BB0_71;
BB0_73:
add.f32 %f1291, %f1232, 0f3EE66666;
bra.uni BB0_74;
BB0_71:
setp.neu.f32 %p91, %f197, 0f7F800000;
@%p91 bra BB0_74;
selp.f32 %f1291, 0fFF800000, 0f7F800000, %p5;
BB0_74:
setp.eq.f32 %p92, %f1232, 0f3F800000;
selp.f32 %f209, 0f3F800000, %f1291, %p92;
abs.f32 %f210, %f1231;
setp.lt.f32 %p93, %f210, 0f00800000;
mul.f32 %f1025, %f210, 0f4B800000;
selp.f32 %f1026, 0fC3170000, 0fC2FE0000, %p93;
selp.f32 %f1027, %f1025, %f210, %p93;
mov.b32 %r139, %f1027;
and.b32 %r140, %r139, 8388607;
or.b32 %r141, %r140, 1065353216;
mov.b32 %f1028, %r141;
shr.u32 %r142, %r139, 23;
cvt.rn.f32.u32 %f1029, %r142;
add.f32 %f1030, %f1026, %f1029;
setp.gt.f32 %p94, %f1028, 0f3FB504F3;
mul.f32 %f1031, %f1028, 0f3F000000;
add.f32 %f1032, %f1030, 0f3F800000;
selp.f32 %f1033, %f1031, %f1028, %p94;
selp.f32 %f1034, %f1032, %f1030, %p94;
add.f32 %f1035, %f1033, 0fBF800000;
add.f32 %f1024, %f1033, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1023,%f1024;
// inline asm
add.f32 %f1036, %f1035, %f1035;
mul.f32 %f1037, %f1023, %f1036;
mul.f32 %f1038, %f1037, %f1037;
fma.rn.f32 %f1041, %f962, %f1038, %f961;
fma.rn.f32 %f1043, %f1041, %f1038, %f964;
mul.rn.f32 %f1044, %f1043, %f1038;
mul.rn.f32 %f1045, %f1044, %f1037;
sub.f32 %f1046, %f1035, %f1037;
neg.f32 %f1047, %f1037;
add.f32 %f1048, %f1046, %f1046;
fma.rn.f32 %f1049, %f1047, %f1035, %f1048;
mul.rn.f32 %f1050, %f1023, %f1049;
add.f32 %f1051, %f1045, %f1037;
sub.f32 %f1052, %f1037, %f1051;
add.f32 %f1053, %f1045, %f1052;
add.f32 %f1054, %f1050, %f1053;
add.f32 %f1055, %f1051, %f1054;
sub.f32 %f1056, %f1051, %f1055;
add.f32 %f1057, %f1054, %f1056;
mul.rn.f32 %f1059, %f1034, %f980;
mul.rn.f32 %f1061, %f1034, %f982;
add.f32 %f1062, %f1059, %f1055;
sub.f32 %f1063, %f1059, %f1062;
add.f32 %f1064, %f1055, %f1063;
add.f32 %f1065, %f1057, %f1064;
add.f32 %f1066, %f1061, %f1065;
add.f32 %f1067, %f1062, %f1066;
sub.f32 %f1068, %f1062, %f1067;
add.f32 %f1069, %f1066, %f1068;
mul.rn.f32 %f1071, %f992, %f1067;
neg.f32 %f1072, %f1071;
fma.rn.f32 %f1073, %f992, %f1067, %f1072;
fma.rn.f32 %f1074, %f992, %f1069, %f1073;
fma.rn.f32 %f1076, %f997, %f1067, %f1074;
add.rn.f32 %f1077, %f1071, %f1076;
neg.f32 %f1078, %f1077;
add.rn.f32 %f1079, %f1071, %f1078;
add.rn.f32 %f1080, %f1079, %f1076;
mov.b32 %r143, %f1077;
setp.eq.s32 %p95, %r143, 1118925336;
add.s32 %r144, %r143, -1;
mov.b32 %f1081, %r144;
add.f32 %f1082, %f1080, 0f37000000;
selp.f32 %f1083, %f1081, %f1077, %p95;
selp.f32 %f211, %f1082, %f1080, %p95;
mul.f32 %f1084, %f1083, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1085, %f1084;
fma.rn.f32 %f1087, %f1085, %f1008, %f1083;
fma.rn.f32 %f1089, %f1085, %f1010, %f1087;
mul.f32 %f1090, %f1089, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1091, %f1090;
add.f32 %f1092, %f1085, 0f00000000;
ex2.approx.f32 %f1093, %f1092;
mul.f32 %f1094, %f1091, %f1093;
setp.lt.f32 %p96, %f1083, 0fC2D20000;
selp.f32 %f1095, 0f00000000, %f1094, %p96;
setp.gt.f32 %p97, %f1083, 0f42D20000;
selp.f32 %f1292, 0f7F800000, %f1095, %p97;
setp.eq.f32 %p98, %f1292, 0f7F800000;
@%p98 bra BB0_76;
fma.rn.f32 %f1292, %f1292, %f211, %f1292;
BB0_76:
setp.lt.f32 %p99, %f1231, 0f00000000;
and.pred %p6, %p99, %p84;
mov.b32 %r145, %f1292;
xor.b32 %r146, %r145, -2147483648;
mov.b32 %f1096, %r146;
selp.f32 %f1294, %f1096, %f1292, %p6;
setp.eq.f32 %p101, %f1231, 0f00000000;
@%p101 bra BB0_79;
bra.uni BB0_77;
BB0_79:
add.f32 %f1099, %f1231, %f1231;
selp.f32 %f1294, %f1099, 0f00000000, %p84;
bra.uni BB0_80;
BB0_77:
setp.geu.f32 %p102, %f1231, 0f00000000;
@%p102 bra BB0_80;
cvt.rzi.f32.f32 %f1098, %f992;
setp.neu.f32 %p103, %f1098, 0f3EE66666;
selp.f32 %f1294, 0f7FFFFFFF, %f1294, %p103;
BB0_80:
add.f32 %f1100, %f210, 0f3EE66666;
mov.b32 %r147, %f1100;
setp.lt.s32 %p105, %r147, 2139095040;
@%p105 bra BB0_85;
setp.gtu.f32 %p106, %f210, 0f7F800000;
@%p106 bra BB0_84;
bra.uni BB0_82;
BB0_84:
add.f32 %f1294, %f1231, 0f3EE66666;
bra.uni BB0_85;
BB0_82:
setp.neu.f32 %p107, %f210, 0f7F800000;
@%p107 bra BB0_85;
selp.f32 %f1294, 0fFF800000, 0f7F800000, %p6;
BB0_85:
setp.eq.f32 %p108, %f1231, 0f3F800000;
selp.f32 %f222, 0f3F800000, %f1294, %p108;
abs.f32 %f223, %f1230;
setp.lt.f32 %p109, %f223, 0f00800000;
mul.f32 %f1103, %f223, 0f4B800000;
selp.f32 %f1104, 0fC3170000, 0fC2FE0000, %p109;
selp.f32 %f1105, %f1103, %f223, %p109;
mov.b32 %r148, %f1105;
and.b32 %r149, %r148, 8388607;
or.b32 %r150, %r149, 1065353216;
mov.b32 %f1106, %r150;
shr.u32 %r151, %r148, 23;
cvt.rn.f32.u32 %f1107, %r151;
add.f32 %f1108, %f1104, %f1107;
setp.gt.f32 %p110, %f1106, 0f3FB504F3;
mul.f32 %f1109, %f1106, 0f3F000000;
add.f32 %f1110, %f1108, 0f3F800000;
selp.f32 %f1111, %f1109, %f1106, %p110;
selp.f32 %f1112, %f1110, %f1108, %p110;
add.f32 %f1113, %f1111, 0fBF800000;
add.f32 %f1102, %f1111, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1101,%f1102;
// inline asm
add.f32 %f1114, %f1113, %f1113;
mul.f32 %f1115, %f1101, %f1114;
mul.f32 %f1116, %f1115, %f1115;
fma.rn.f32 %f1119, %f962, %f1116, %f961;
fma.rn.f32 %f1121, %f1119, %f1116, %f964;
mul.rn.f32 %f1122, %f1121, %f1116;
mul.rn.f32 %f1123, %f1122, %f1115;
sub.f32 %f1124, %f1113, %f1115;
neg.f32 %f1125, %f1115;
add.f32 %f1126, %f1124, %f1124;
fma.rn.f32 %f1127, %f1125, %f1113, %f1126;
mul.rn.f32 %f1128, %f1101, %f1127;
add.f32 %f1129, %f1123, %f1115;
sub.f32 %f1130, %f1115, %f1129;
add.f32 %f1131, %f1123, %f1130;
add.f32 %f1132, %f1128, %f1131;
add.f32 %f1133, %f1129, %f1132;
sub.f32 %f1134, %f1129, %f1133;
add.f32 %f1135, %f1132, %f1134;
mul.rn.f32 %f1137, %f1112, %f980;
mul.rn.f32 %f1139, %f1112, %f982;
add.f32 %f1140, %f1137, %f1133;
sub.f32 %f1141, %f1137, %f1140;
add.f32 %f1142, %f1133, %f1141;
add.f32 %f1143, %f1135, %f1142;
add.f32 %f1144, %f1139, %f1143;
add.f32 %f1145, %f1140, %f1144;
sub.f32 %f1146, %f1140, %f1145;
add.f32 %f1147, %f1144, %f1146;
mul.rn.f32 %f1149, %f992, %f1145;
neg.f32 %f1150, %f1149;
fma.rn.f32 %f1151, %f992, %f1145, %f1150;
fma.rn.f32 %f1152, %f992, %f1147, %f1151;
fma.rn.f32 %f1154, %f997, %f1145, %f1152;
add.rn.f32 %f1155, %f1149, %f1154;
neg.f32 %f1156, %f1155;
add.rn.f32 %f1157, %f1149, %f1156;
add.rn.f32 %f1158, %f1157, %f1154;
mov.b32 %r152, %f1155;
setp.eq.s32 %p111, %r152, 1118925336;
add.s32 %r153, %r152, -1;
mov.b32 %f1159, %r153;
add.f32 %f1160, %f1158, 0f37000000;
selp.f32 %f1161, %f1159, %f1155, %p111;
selp.f32 %f224, %f1160, %f1158, %p111;
mul.f32 %f1162, %f1161, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1163, %f1162;
fma.rn.f32 %f1165, %f1163, %f1008, %f1161;
fma.rn.f32 %f1167, %f1163, %f1010, %f1165;
mul.f32 %f1168, %f1167, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1169, %f1168;
add.f32 %f1170, %f1163, 0f00000000;
ex2.approx.f32 %f1171, %f1170;
mul.f32 %f1172, %f1169, %f1171;
setp.lt.f32 %p112, %f1161, 0fC2D20000;
selp.f32 %f1173, 0f00000000, %f1172, %p112;
setp.gt.f32 %p113, %f1161, 0f42D20000;
selp.f32 %f1295, 0f7F800000, %f1173, %p113;
setp.eq.f32 %p114, %f1295, 0f7F800000;
@%p114 bra BB0_87;
fma.rn.f32 %f1295, %f1295, %f224, %f1295;
BB0_87:
setp.lt.f32 %p115, %f1230, 0f00000000;
and.pred %p7, %p115, %p84;
mov.b32 %r154, %f1295;
xor.b32 %r155, %r154, -2147483648;
mov.b32 %f1174, %r155;
selp.f32 %f1297, %f1174, %f1295, %p7;
setp.eq.f32 %p117, %f1230, 0f00000000;
@%p117 bra BB0_90;
bra.uni BB0_88;
BB0_90:
add.f32 %f1177, %f1230, %f1230;
selp.f32 %f1297, %f1177, 0f00000000, %p84;
bra.uni BB0_91;
BB0_88:
setp.geu.f32 %p118, %f1230, 0f00000000;
@%p118 bra BB0_91;
cvt.rzi.f32.f32 %f1176, %f992;
setp.neu.f32 %p119, %f1176, 0f3EE66666;
selp.f32 %f1297, 0f7FFFFFFF, %f1297, %p119;
BB0_91:
add.f32 %f1178, %f223, 0f3EE66666;
mov.b32 %r156, %f1178;
setp.lt.s32 %p121, %r156, 2139095040;
@%p121 bra BB0_96;
setp.gtu.f32 %p122, %f223, 0f7F800000;
@%p122 bra BB0_95;
bra.uni BB0_93;
BB0_95:
add.f32 %f1297, %f1230, 0f3EE66666;
bra.uni BB0_96;
BB0_93:
setp.neu.f32 %p123, %f223, 0f7F800000;
@%p123 bra BB0_96;
selp.f32 %f1297, 0fFF800000, 0f7F800000, %p7;
BB0_96:
mov.u32 %r192, 4;
setp.eq.f32 %p124, %f1230, 0f3F800000;
selp.f32 %f1179, 0f3F800000, %f1297, %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, %r192, %rd52, %rd53, %rd13, %rd13);
// inline asm
cvt.sat.f32.f32 %f1180, %f1179;
mul.f32 %f1181, %f1180, 0f437FFD71;
cvt.rzi.u32.f32 %r159, %f1181;
cvt.sat.f32.f32 %f1182, %f222;
mul.f32 %f1183, %f1182, 0f437FFD71;
cvt.rzi.u32.f32 %r160, %f1183;
cvt.sat.f32.f32 %f1184, %f209;
mul.f32 %f1185, %f1184, 0f437FFD71;
cvt.rzi.u32.f32 %r161, %f1185;
cvt.u16.u32 %rs19, %r159;
cvt.u16.u32 %rs20, %r161;
cvt.u16.u32 %rs21, %r160;
mov.u16 %rs22, 255;
st.v4.u8 [%rd50], {%rs19, %rs21, %rs20, %rs22};
ld.global.u32 %r200, [imageEnabled];
BB0_97:
and.b32 %r162, %r200, 4;
setp.eq.s32 %p125, %r162, 0;
@%p125 bra BB0_101;
ld.global.u32 %r163, [additive];
setp.eq.s32 %p126, %r163, 0;
cvt.u64.u32 %rd4, %r2;
cvt.u64.u32 %rd5, %r3;
mov.f32 %f1186, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs23, %f1186;}
// inline asm
@%p126 bra BB0_100;
mov.u64 %rd69, image_HDR;
cvta.global.u64 %rd58, %rd69;
mov.u32 %r167, 8;
// inline asm
call (%rd57), _rt_buffer_get_64, (%rd58, %r28, %r167, %rd4, %rd5, %rd13, %rd13);
// inline asm
ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd57];
// inline asm
{ cvt.f32.f16 %f1187, %rs30;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1188, %rs31;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1189, %rs32;}
// inline asm
// inline asm
call (%rd63), _rt_buffer_get_64, (%rd58, %r28, %r167, %rd4, %rd5, %rd13, %rd13);
// inline asm
add.f32 %f1190, %f1232, %f1187;
add.f32 %f1191, %f1231, %f1188;
add.f32 %f1192, %f1230, %f1189;
// inline asm
{ cvt.rn.f16.f32 %rs29, %f1192;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs28, %f1191;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs27, %f1190;}
// 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 %r169, 8;
// inline asm
call (%rd70), _rt_buffer_get_64, (%rd71, %r28, %r169, %rd4, %rd5, %rd13, %rd13);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs36, %f1230;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs35, %f1231;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs34, %f1232;}
// 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;
mov.u32 %r193, 4;
mul.f32 %f1196, %f19, %f19;
fma.rn.f32 %f1197, %f20, %f20, %f1196;
fma.rn.f32 %f1198, %f18, %f18, %f1197;
sqrt.rn.f32 %f1199, %f1198;
rcp.rn.f32 %f1200, %f1199;
mul.f32 %f1201, %f20, %f1200;
mul.f32 %f1202, %f19, %f1200;
mul.f32 %f1203, %f18, %f1200;
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, %r193, %rd79, %rd80, %rd13, %rd13);
// inline asm
fma.rn.f32 %f1204, %f1201, 0f3F000000, 0f3F000000;
mul.f32 %f1205, %f1204, 0f437F0000;
cvt.rzi.u32.f32 %r172, %f1205;
fma.rn.f32 %f1206, %f1202, 0f3F000000, 0f3F000000;
mul.f32 %f1207, %f1206, 0f437F0000;
cvt.rzi.u32.f32 %r173, %f1207;
fma.rn.f32 %f1208, %f1203, 0f3F000000, 0f3F000000;
mul.f32 %f1209, %f1208, 0f437F0000;
cvt.rzi.u32.f32 %r174, %f1209;
cvt.u16.u32 %rs39, %r174;
cvt.u16.u32 %rs40, %r173;
cvt.u16.u32 %rs41, %r172;
mov.u16 %rs42, 255;
st.v4.u8 [%rd77], {%rs41, %rs40, %rs39, %rs42};
BB0_113:
ret;
}