saad
/
ArabDesert
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
ArabDesert/Assets/Editor/x64/Bakery/lmSunSuperSampleRNM.ptx

3091 lines
92 KiB
Plaintext
Raw Blame History

//
// 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 uvtangent[1];
.global .align 1 .b8 uvpos[1];
.global .align 1 .b8 uvnormal[1];
.global .align 1 .b8 rnd_seeds[1];
.global .align 4 .b8 directDir[12];
.global .align 4 .b8 directColor[12];
.global .align 4 .f32 shadowSpread;
.global .align 4 .u32 samples;
.global .align 4 .u32 ignoreNormal;
.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_typeinfo9directDirE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo11directColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo12shadowSpreadE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
.global .align 4 .b8 _ZN21rti_internal_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 8 .b8 _ZN21rti_internal_typename9directDirE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename11directColorE[7] = {102, 108, 111, 97, 116, 51, 0};
.global .align 8 .b8 _ZN21rti_internal_typename12shadowSpreadE[6] = {102, 108, 111, 97, 116, 0};
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[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_typeenum9directDirE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum11directColorE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum12shadowSpreadE = 4919;
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 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_semantic9directDirE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic11directColorE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic12shadowSpreadE[1];
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[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_annotation9directDirE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation11directColorE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12shadowSpreadE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[1];
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
.visible .entry _Z6oxMainv(
)
{
.local .align 4 .b8 __local_depot0[36];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<179>;
.reg .b16 %rs<182>;
.reg .f32 %f<1689>;
.reg .b32 %r<521>;
.reg .b64 %rd<312>;
mov.u64 %rd311, __local_depot0;
cvta.local.u64 %SP, %rd311;
ld.global.v2.u32 {%r115, %r116}, [pixelID];
cvt.u64.u32 %rd29, %r115;
cvt.u64.u32 %rd30, %r116;
mov.u64 %rd33, uvnormal;
cvta.global.u64 %rd28, %rd33;
mov.u32 %r113, 2;
mov.u32 %r114, 4;
mov.u64 %rd32, 0;
// inline asm
call (%rd27), _rt_buffer_get_64, (%rd28, %r113, %r114, %rd29, %rd30, %rd32, %rd32);
// inline asm
ld.u32 %r1, [%rd27];
shr.u32 %r119, %r1, 16;
cvt.u16.u32 %rs1, %r119;
and.b16 %rs14, %rs1, 255;
cvt.u16.u32 %rs15, %r1;
or.b16 %rs16, %rs15, %rs14;
setp.eq.s16 %p6, %rs16, 0;
mov.f32 %f1623, 0f00000000;
mov.f32 %f1624, %f1623;
mov.f32 %f1625, %f1623;
@%p6 bra BB0_2;
ld.u8 %rs17, [%rd27+1];
and.b16 %rs19, %rs15, 255;
cvt.rn.f32.u16 %f258, %rs19;
div.rn.f32 %f259, %f258, 0f437F0000;
fma.rn.f32 %f260, %f259, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f261, %rs17;
div.rn.f32 %f262, %f261, 0f437F0000;
fma.rn.f32 %f263, %f262, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f264, %rs14;
div.rn.f32 %f265, %f264, 0f437F0000;
fma.rn.f32 %f266, %f265, 0f40000000, 0fBF800000;
mul.f32 %f267, %f263, %f263;
fma.rn.f32 %f268, %f260, %f260, %f267;
fma.rn.f32 %f269, %f266, %f266, %f268;
sqrt.rn.f32 %f270, %f269;
rcp.rn.f32 %f271, %f270;
mul.f32 %f1623, %f260, %f271;
mul.f32 %f1624, %f263, %f271;
mul.f32 %f1625, %f266, %f271;
BB0_2:
ld.global.v2.u32 {%r120, %r121}, [pixelID];
ld.global.v2.u32 {%r123, %r124}, [tileInfo];
add.s32 %r2, %r120, %r123;
add.s32 %r3, %r121, %r124;
setp.eq.f32 %p7, %f1624, 0f00000000;
setp.eq.f32 %p8, %f1623, 0f00000000;
and.pred %p9, %p8, %p7;
setp.eq.f32 %p10, %f1625, 0f00000000;
and.pred %p11, %p9, %p10;
@%p11 bra BB0_154;
bra.uni BB0_3;
BB0_154:
ld.global.u32 %r520, [imageEnabled];
and.b32 %r453, %r520, 1;
setp.eq.b32 %p172, %r453, 1;
@!%p172 bra BB0_156;
bra.uni BB0_155;
BB0_155:
cvt.u64.u32 %rd215, %r2;
cvt.u64.u32 %rd216, %r3;
mov.u64 %rd219, image;
cvta.global.u64 %rd214, %rd219;
mov.u64 %rd218, 0;
// inline asm
call (%rd213), _rt_buffer_get_64, (%rd214, %r113, %r114, %rd215, %rd216, %rd218, %rd218);
// inline asm
mov.u16 %rs119, 0;
st.v4.u8 [%rd213], {%rs119, %rs119, %rs119, %rs119};
ld.global.u32 %r520, [imageEnabled];
BB0_156:
and.b32 %r456, %r520, 8;
setp.eq.s32 %p173, %r456, 0;
@%p173 bra BB0_158;
cvt.u64.u32 %rd223, %r3;
cvt.u64.u32 %rd222, %r2;
mov.u64 %rd226, image_Mask;
cvta.global.u64 %rd221, %rd226;
mov.u64 %rd225, 0;
// inline asm
call (%rd220), _rt_buffer_get_64, (%rd221, %r113, %r113, %rd222, %rd223, %rd225, %rd225);
// inline asm
mov.f32 %f1592, 0f00000000;
cvt.rzi.u32.f32 %r459, %f1592;
cvt.u16.u32 %rs120, %r459;
mov.u16 %rs121, 0;
st.v2.u8 [%rd220], {%rs120, %rs121};
ld.global.u32 %r520, [imageEnabled];
BB0_158:
cvt.u64.u32 %rd25, %r2;
cvt.u64.u32 %rd26, %r3;
and.b32 %r460, %r520, 4;
setp.eq.s32 %p174, %r460, 0;
@%p174 bra BB0_162;
ld.global.u32 %r461, [additive];
setp.eq.s32 %p175, %r461, 0;
@%p175 bra BB0_161;
mov.u64 %rd239, image_HDR;
cvta.global.u64 %rd228, %rd239;
mov.u32 %r465, 8;
mov.u64 %rd238, 0;
// inline asm
call (%rd227), _rt_buffer_get_64, (%rd228, %r113, %r465, %rd25, %rd26, %rd238, %rd238);
// inline asm
ld.v4.u16 {%rs128, %rs129, %rs130, %rs131}, [%rd227];
// inline asm
{ cvt.f32.f16 %f1593, %rs128;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1594, %rs129;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1595, %rs130;}
// inline asm
// inline asm
call (%rd233), _rt_buffer_get_64, (%rd228, %r113, %r465, %rd25, %rd26, %rd238, %rd238);
// inline asm
add.f32 %f1596, %f1593, 0f00000000;
add.f32 %f1597, %f1594, 0f00000000;
add.f32 %f1598, %f1595, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs127, %f1598;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs126, %f1597;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs125, %f1596;}
// inline asm
mov.u16 %rs132, 0;
st.v4.u16 [%rd233], {%rs125, %rs126, %rs127, %rs132};
bra.uni BB0_162;
BB0_3:
ld.global.f32 %f8, [directDir+4];
ld.global.f32 %f9, [directDir+8];
ld.global.f32 %f7, [directDir];
ld.global.v2.u32 {%r136, %r137}, [pixelID];
cvt.u64.u32 %rd36, %r136;
cvt.u64.u32 %rd37, %r137;
mov.u64 %rd58, uvpos;
cvta.global.u64 %rd35, %rd58;
mov.u32 %r133, 12;
// inline asm
call (%rd34), _rt_buffer_get_64, (%rd35, %r113, %r133, %rd36, %rd37, %rd32, %rd32);
// inline asm
ld.f32 %f274, [%rd34+8];
ld.f32 %f275, [%rd34+4];
ld.f32 %f276, [%rd34];
mul.f32 %f277, %f276, 0f3456BF95;
mul.f32 %f278, %f275, 0f3456BF95;
mul.f32 %f279, %f274, 0f3456BF95;
abs.f32 %f280, %f1623;
div.rn.f32 %f281, %f277, %f280;
abs.f32 %f282, %f1624;
div.rn.f32 %f283, %f278, %f282;
abs.f32 %f284, %f1625;
div.rn.f32 %f285, %f279, %f284;
abs.f32 %f286, %f281;
abs.f32 %f287, %f283;
abs.f32 %f288, %f285;
mov.f32 %f289, 0f38D1B717;
max.f32 %f290, %f286, %f289;
max.f32 %f291, %f287, %f289;
max.f32 %f292, %f288, %f289;
fma.rn.f32 %f10, %f1623, %f290, %f276;
fma.rn.f32 %f11, %f1624, %f291, %f275;
fma.rn.f32 %f12, %f1625, %f292, %f274;
setp.gt.f32 %p12, %f280, %f284;
neg.f32 %f293, %f1624;
selp.f32 %f294, %f293, 0f00000000, %p12;
neg.f32 %f295, %f1625;
selp.f32 %f296, %f1623, %f295, %p12;
selp.f32 %f297, 0f00000000, %f1624, %p12;
mul.f32 %f298, %f296, %f296;
fma.rn.f32 %f299, %f294, %f294, %f298;
fma.rn.f32 %f300, %f297, %f297, %f299;
sqrt.rn.f32 %f301, %f300;
rcp.rn.f32 %f302, %f301;
mul.f32 %f13, %f294, %f302;
mul.f32 %f14, %f296, %f302;
mul.f32 %f15, %f297, %f302;
mul.f32 %f303, %f1625, %f14;
mul.f32 %f304, %f1624, %f15;
sub.f32 %f16, %f303, %f304;
mul.f32 %f305, %f1623, %f15;
mul.f32 %f306, %f1625, %f13;
sub.f32 %f17, %f305, %f306;
mul.f32 %f307, %f1624, %f13;
mul.f32 %f308, %f1623, %f14;
sub.f32 %f18, %f307, %f308;
ld.global.v2.u32 {%r140, %r141}, [pixelID];
cvt.u64.u32 %rd43, %r141;
add.s32 %r144, %r140, 1;
ld.global.u32 %r145, [resolution];
add.s32 %r146, %r145, -1;
min.u32 %r147, %r144, %r146;
cvt.u64.u32 %rd42, %r147;
// inline asm
call (%rd40), _rt_buffer_get_64, (%rd35, %r113, %r133, %rd42, %rd43, %rd32, %rd32);
// inline asm
ld.f32 %f309, [%rd40+8];
ld.f32 %f310, [%rd40+4];
ld.f32 %f311, [%rd40];
setp.eq.f32 %p13, %f311, 0f00000000;
setp.eq.f32 %p14, %f310, 0f00000000;
and.pred %p15, %p13, %p14;
setp.eq.f32 %p16, %f309, 0f00000000;
and.pred %p17, %p15, %p16;
selp.f32 %f312, %f276, %f311, %p17;
selp.f32 %f313, %f275, %f310, %p17;
selp.f32 %f314, %f274, %f309, %p17;
ld.global.v2.u32 {%r148, %r149}, [pixelID];
add.s32 %r152, %r149, 1;
ld.global.u32 %r153, [resolution+4];
add.s32 %r154, %r153, -1;
min.u32 %r155, %r152, %r154;
cvt.u64.u32 %rd48, %r148;
cvt.u64.u32 %rd49, %r155;
// inline asm
call (%rd46), _rt_buffer_get_64, (%rd35, %r113, %r133, %rd48, %rd49, %rd32, %rd32);
// inline asm
ld.f32 %f315, [%rd46+8];
ld.f32 %f316, [%rd46+4];
ld.f32 %f317, [%rd46];
setp.eq.f32 %p18, %f317, 0f00000000;
setp.eq.f32 %p19, %f316, 0f00000000;
and.pred %p20, %p18, %p19;
setp.eq.f32 %p21, %f315, 0f00000000;
and.pred %p22, %p20, %p21;
selp.f32 %f318, %f276, %f317, %p22;
selp.f32 %f319, %f275, %f316, %p22;
selp.f32 %f320, %f274, %f315, %p22;
sub.f32 %f321, %f312, %f276;
sub.f32 %f322, %f313, %f275;
sub.f32 %f323, %f314, %f274;
sub.f32 %f324, %f318, %f276;
sub.f32 %f325, %f319, %f275;
sub.f32 %f326, %f320, %f274;
abs.f32 %f327, %f321;
abs.f32 %f328, %f324;
max.f32 %f329, %f327, %f328;
abs.f32 %f330, %f322;
abs.f32 %f331, %f325;
max.f32 %f332, %f330, %f331;
abs.f32 %f333, %f323;
abs.f32 %f334, %f326;
max.f32 %f335, %f333, %f334;
max.f32 %f336, %f329, %f332;
max.f32 %f337, %f336, %f335;
setp.neu.f32 %p23, %f321, 0f00000000;
setp.neu.f32 %p24, %f322, 0f00000000;
or.pred %p25, %p23, %p24;
setp.neu.f32 %p26, %f323, 0f00000000;
setp.neu.f32 %p27, %f324, 0f00000000;
setp.neu.f32 %p28, %f325, 0f00000000;
or.pred %p29, %p27, %p28;
setp.neu.f32 %p30, %f326, 0f00000000;
or.pred %p31, %p29, %p30;
or.pred %p32, %p25, %p26;
and.pred %p33, %p32, %p31;
selp.f32 %f19, %f337, 0f00000000, %p33;
abs.f32 %f338, %f7;
abs.f32 %f339, %f9;
setp.gt.f32 %p34, %f338, %f339;
neg.f32 %f340, %f8;
neg.f32 %f341, %f9;
selp.f32 %f342, %f340, 0f00000000, %p34;
selp.f32 %f343, %f7, %f341, %p34;
selp.f32 %f344, 0f00000000, %f8, %p34;
mul.f32 %f345, %f343, %f343;
fma.rn.f32 %f346, %f342, %f342, %f345;
fma.rn.f32 %f347, %f344, %f344, %f346;
sqrt.rn.f32 %f348, %f347;
rcp.rn.f32 %f349, %f348;
mul.f32 %f20, %f342, %f349;
mul.f32 %f21, %f343, %f349;
mul.f32 %f22, %f344, %f349;
ld.global.v2.u32 {%r156, %r157}, [pixelID];
cvt.u64.u32 %rd54, %r156;
cvt.u64.u32 %rd55, %r157;
mov.u64 %rd59, rnd_seeds;
cvta.global.u64 %rd53, %rd59;
// inline asm
call (%rd52), _rt_buffer_get_64, (%rd53, %r113, %r114, %rd54, %rd55, %rd32, %rd32);
// inline asm
ld.global.u32 %r489, [samples];
mov.f32 %f1628, 0f00000000;
setp.lt.s32 %p35, %r489, 1;
@%p35 bra BB0_4;
cvt.rn.f32.s32 %f352, %r489;
rcp.rn.f32 %f23, %f352;
ld.u32 %r515, [%rd52];
mul.f32 %f24, %f10, 0f3456BF95;
mul.f32 %f25, %f11, 0f3456BF95;
mul.f32 %f26, %f12, 0f3456BF95;
mul.f32 %f353, %f7, %f21;
mul.f32 %f354, %f8, %f20;
sub.f32 %f27, %f354, %f353;
mul.f32 %f355, %f9, %f20;
mul.f32 %f356, %f7, %f22;
sub.f32 %f28, %f356, %f355;
mul.f32 %f357, %f8, %f22;
mul.f32 %f358, %f9, %f21;
sub.f32 %f29, %f358, %f357;
mov.f32 %f1628, 0f00000000;
mov.u32 %r490, 0;
abs.f32 %f359, %f25;
abs.f32 %f360, %f24;
max.f32 %f361, %f360, %f359;
abs.f32 %f362, %f26;
max.f32 %f363, %f361, %f362;
mov.f32 %f1629, %f1628;
BB0_6:
setp.lt.s32 %p36, %r489, 1;
@%p36 bra BB0_73;
mov.u32 %r492, 0;
BB0_8:
cvt.rn.f32.s32 %f1621, %r490;
mad.lo.s32 %r162, %r515, 1664525, 1013904223;
and.b32 %r163, %r162, 16777215;
cvt.rn.f32.u32 %f365, %r163;
fma.rn.f32 %f366, %f365, 0f33800000, %f1621;
mul.f32 %f367, %f23, %f366;
mad.lo.s32 %r11, %r162, 1664525, 1013904223;
and.b32 %r164, %r11, 16777215;
cvt.rn.f32.u32 %f368, %r164;
cvt.rn.f32.s32 %f369, %r492;
fma.rn.f32 %f370, %f368, 0f33800000, %f369;
mul.f32 %f371, %f23, %f370;
sqrt.rn.f32 %f36, %f367;
mul.f32 %f1636, %f371, 0f40C90FDB;
abs.f32 %f38, %f1636;
setp.neu.f32 %p37, %f38, 0f7F800000;
mov.f32 %f1630, %f1636;
@%p37 bra BB0_10;
mov.f32 %f372, 0f00000000;
mul.rn.f32 %f1630, %f1636, %f372;
BB0_10:
mul.f32 %f373, %f1630, 0f3F22F983;
cvt.rni.s32.f32 %r503, %f373;
cvt.rn.f32.s32 %f374, %r503;
neg.f32 %f375, %f374;
mov.f32 %f376, 0f3FC90FDA;
fma.rn.f32 %f377, %f375, %f376, %f1630;
mov.f32 %f378, 0f33A22168;
fma.rn.f32 %f379, %f375, %f378, %f377;
mov.f32 %f380, 0f27C234C5;
fma.rn.f32 %f1631, %f375, %f380, %f379;
abs.f32 %f381, %f1630;
setp.leu.f32 %p38, %f381, 0f47CE4780;
@%p38 bra BB0_21;
add.u64 %rd61, %SP, 8;
cvta.to.local.u64 %rd3, %rd61;
mov.b32 %r13, %f1630;
shr.u32 %r14, %r13, 23;
shl.b32 %r167, %r13, 8;
or.b32 %r15, %r167, -2147483648;
mov.u32 %r494, 0;
mov.u64 %rd308, 0;
mov.u64 %rd307, %rd3;
mov.u32 %r495, %r494;
BB0_12:
.pragma "nounroll";
shl.b64 %rd62, %rd308, 2;
mov.u64 %rd63, __cudart_i2opi_f;
add.s64 %rd64, %rd63, %rd62;
ld.const.u32 %r170, [%rd64];
// inline asm
{
mad.lo.cc.u32 %r168, %r170, %r15, %r495;
madc.hi.u32 %r495, %r170, %r15, 0;
}
// inline asm
st.local.u32 [%rd307], %r168;
add.s32 %r494, %r494, 1;
cvt.s64.s32 %rd308, %r494;
mul.wide.s32 %rd67, %r494, 4;
add.s64 %rd307, %rd3, %rd67;
setp.ne.s32 %p39, %r494, 6;
@%p39 bra BB0_12;
and.b32 %r173, %r14, 255;
add.s32 %r174, %r173, -128;
shr.u32 %r175, %r174, 5;
and.b32 %r20, %r13, -2147483648;
cvta.to.local.u64 %rd69, %rd61;
st.local.u32 [%rd69+24], %r495;
mov.u32 %r176, 6;
sub.s32 %r177, %r176, %r175;
mul.wide.s32 %rd70, %r177, 4;
add.s64 %rd8, %rd69, %rd70;
ld.local.u32 %r496, [%rd8];
ld.local.u32 %r497, [%rd8+-4];
and.b32 %r23, %r14, 31;
setp.eq.s32 %p40, %r23, 0;
@%p40 bra BB0_15;
mov.u32 %r178, 32;
sub.s32 %r179, %r178, %r23;
shr.u32 %r180, %r497, %r179;
shl.b32 %r181, %r496, %r23;
add.s32 %r496, %r180, %r181;
ld.local.u32 %r182, [%rd8+-8];
shr.u32 %r183, %r182, %r179;
shl.b32 %r184, %r497, %r23;
add.s32 %r497, %r183, %r184;
BB0_15:
shr.u32 %r185, %r497, 30;
shl.b32 %r186, %r496, 2;
add.s32 %r498, %r185, %r186;
shl.b32 %r29, %r497, 2;
shr.u32 %r187, %r498, 31;
shr.u32 %r188, %r496, 30;
add.s32 %r30, %r187, %r188;
setp.eq.s32 %p41, %r187, 0;
@%p41 bra BB0_16;
bra.uni BB0_17;
BB0_16:
mov.u32 %r499, %r20;
mov.u32 %r500, %r29;
bra.uni BB0_18;
BB0_17:
not.b32 %r189, %r498;
neg.s32 %r500, %r29;
setp.eq.s32 %p42, %r29, 0;
selp.u32 %r190, 1, 0, %p42;
add.s32 %r498, %r190, %r189;
xor.b32 %r499, %r20, -2147483648;
BB0_18:
clz.b32 %r502, %r498;
setp.eq.s32 %p43, %r502, 0;
shl.b32 %r191, %r498, %r502;
mov.u32 %r192, 32;
sub.s32 %r193, %r192, %r502;
shr.u32 %r194, %r500, %r193;
add.s32 %r195, %r194, %r191;
selp.b32 %r38, %r498, %r195, %p43;
mov.u32 %r196, -921707870;
mul.hi.u32 %r501, %r38, %r196;
setp.eq.s32 %p44, %r20, 0;
neg.s32 %r197, %r30;
selp.b32 %r503, %r30, %r197, %p44;
setp.lt.s32 %p45, %r501, 1;
@%p45 bra BB0_20;
mul.lo.s32 %r198, %r38, -921707870;
shr.u32 %r199, %r198, 31;
shl.b32 %r200, %r501, 1;
add.s32 %r501, %r199, %r200;
add.s32 %r502, %r502, 1;
BB0_20:
mov.u32 %r201, 126;
sub.s32 %r202, %r201, %r502;
shl.b32 %r203, %r202, 23;
add.s32 %r204, %r501, 1;
shr.u32 %r205, %r204, 7;
add.s32 %r206, %r205, 1;
shr.u32 %r207, %r206, 1;
add.s32 %r208, %r207, %r203;
or.b32 %r209, %r208, %r499;
mov.b32 %f1631, %r209;
BB0_21:
mul.rn.f32 %f44, %f1631, %f1631;
add.s32 %r46, %r503, 1;
and.b32 %r47, %r46, 1;
setp.eq.s32 %p46, %r47, 0;
@%p46 bra BB0_23;
bra.uni BB0_22;
BB0_23:
mov.f32 %f384, 0f3C08839E;
mov.f32 %f385, 0fB94CA1F9;
fma.rn.f32 %f1632, %f385, %f44, %f384;
bra.uni BB0_24;
BB0_22:
mov.f32 %f382, 0fBAB6061A;
mov.f32 %f383, 0f37CCF5CE;
fma.rn.f32 %f1632, %f383, %f44, %f382;
BB0_24:
@%p46 bra BB0_26;
bra.uni BB0_25;
BB0_26:
mov.f32 %f389, 0fBE2AAAA3;
fma.rn.f32 %f390, %f1632, %f44, %f389;
mov.f32 %f391, 0f00000000;
fma.rn.f32 %f1633, %f390, %f44, %f391;
bra.uni BB0_27;
BB0_25:
mov.f32 %f386, 0f3D2AAAA5;
fma.rn.f32 %f387, %f1632, %f44, %f386;
mov.f32 %f388, 0fBF000000;
fma.rn.f32 %f1633, %f387, %f44, %f388;
BB0_27:
fma.rn.f32 %f1634, %f1633, %f1631, %f1631;
@%p46 bra BB0_29;
mov.f32 %f392, 0f3F800000;
fma.rn.f32 %f1634, %f1633, %f44, %f392;
BB0_29:
and.b32 %r210, %r46, 2;
setp.eq.s32 %p49, %r210, 0;
@%p49 bra BB0_31;
mov.f32 %f393, 0f00000000;
mov.f32 %f394, 0fBF800000;
fma.rn.f32 %f1634, %f1634, %f394, %f393;
BB0_31:
@%p37 bra BB0_33;
mov.f32 %f395, 0f00000000;
mul.rn.f32 %f1636, %f1636, %f395;
BB0_33:
mul.f32 %f396, %f1636, 0f3F22F983;
cvt.rni.s32.f32 %r513, %f396;
cvt.rn.f32.s32 %f397, %r513;
neg.f32 %f398, %f397;
fma.rn.f32 %f400, %f398, %f376, %f1636;
fma.rn.f32 %f402, %f398, %f378, %f400;
fma.rn.f32 %f1637, %f398, %f380, %f402;
abs.f32 %f404, %f1636;
setp.leu.f32 %p51, %f404, 0f47CE4780;
@%p51 bra BB0_44;
add.u64 %rd72, %SP, 8;
cvta.to.local.u64 %rd9, %rd72;
mov.b32 %r49, %f1636;
shr.u32 %r50, %r49, 23;
shl.b32 %r213, %r49, 8;
or.b32 %r51, %r213, -2147483648;
mov.u32 %r504, 0;
mov.u64 %rd309, %rd9;
mov.u64 %rd310, %rd32;
mov.u32 %r505, %r504;
BB0_35:
.pragma "nounroll";
shl.b64 %rd73, %rd310, 2;
mov.u64 %rd74, __cudart_i2opi_f;
add.s64 %rd75, %rd74, %rd73;
ld.const.u32 %r216, [%rd75];
// inline asm
{
mad.lo.cc.u32 %r214, %r216, %r51, %r505;
madc.hi.u32 %r505, %r216, %r51, 0;
}
// inline asm
st.local.u32 [%rd309], %r214;
add.s32 %r504, %r504, 1;
cvt.s64.s32 %rd310, %r504;
mul.wide.s32 %rd76, %r504, 4;
add.s64 %rd309, %rd9, %rd76;
setp.ne.s32 %p52, %r504, 6;
@%p52 bra BB0_35;
and.b32 %r219, %r50, 255;
add.s32 %r220, %r219, -128;
shr.u32 %r221, %r220, 5;
and.b32 %r56, %r49, -2147483648;
cvta.to.local.u64 %rd78, %rd72;
st.local.u32 [%rd78+24], %r505;
mov.u32 %r222, 6;
sub.s32 %r223, %r222, %r221;
mul.wide.s32 %rd79, %r223, 4;
add.s64 %rd15, %rd78, %rd79;
ld.local.u32 %r506, [%rd15];
ld.local.u32 %r507, [%rd15+-4];
and.b32 %r59, %r50, 31;
setp.eq.s32 %p53, %r59, 0;
@%p53 bra BB0_38;
mov.u32 %r224, 32;
sub.s32 %r225, %r224, %r59;
shr.u32 %r226, %r507, %r225;
shl.b32 %r227, %r506, %r59;
add.s32 %r506, %r226, %r227;
ld.local.u32 %r228, [%rd15+-8];
shr.u32 %r229, %r228, %r225;
shl.b32 %r230, %r507, %r59;
add.s32 %r507, %r229, %r230;
BB0_38:
shr.u32 %r231, %r507, 30;
shl.b32 %r232, %r506, 2;
add.s32 %r508, %r231, %r232;
shl.b32 %r65, %r507, 2;
shr.u32 %r233, %r508, 31;
shr.u32 %r234, %r506, 30;
add.s32 %r66, %r233, %r234;
setp.eq.s32 %p54, %r233, 0;
@%p54 bra BB0_39;
bra.uni BB0_40;
BB0_39:
mov.u32 %r509, %r56;
mov.u32 %r510, %r65;
bra.uni BB0_41;
BB0_40:
not.b32 %r235, %r508;
neg.s32 %r510, %r65;
setp.eq.s32 %p55, %r65, 0;
selp.u32 %r236, 1, 0, %p55;
add.s32 %r508, %r236, %r235;
xor.b32 %r509, %r56, -2147483648;
BB0_41:
clz.b32 %r512, %r508;
setp.eq.s32 %p56, %r512, 0;
shl.b32 %r237, %r508, %r512;
mov.u32 %r238, 32;
sub.s32 %r239, %r238, %r512;
shr.u32 %r240, %r510, %r239;
add.s32 %r241, %r240, %r237;
selp.b32 %r74, %r508, %r241, %p56;
mov.u32 %r242, -921707870;
mul.hi.u32 %r511, %r74, %r242;
setp.eq.s32 %p57, %r56, 0;
neg.s32 %r243, %r66;
selp.b32 %r513, %r66, %r243, %p57;
setp.lt.s32 %p58, %r511, 1;
@%p58 bra BB0_43;
mul.lo.s32 %r244, %r74, -921707870;
shr.u32 %r245, %r244, 31;
shl.b32 %r246, %r511, 1;
add.s32 %r511, %r245, %r246;
add.s32 %r512, %r512, 1;
BB0_43:
mov.u32 %r247, 126;
sub.s32 %r248, %r247, %r512;
shl.b32 %r249, %r248, 23;
add.s32 %r250, %r511, 1;
shr.u32 %r251, %r250, 7;
add.s32 %r252, %r251, 1;
shr.u32 %r253, %r252, 1;
add.s32 %r254, %r253, %r249;
or.b32 %r255, %r254, %r509;
mov.b32 %f1637, %r255;
BB0_44:
mul.rn.f32 %f61, %f1637, %f1637;
and.b32 %r82, %r513, 1;
setp.eq.s32 %p59, %r82, 0;
@%p59 bra BB0_46;
bra.uni BB0_45;
BB0_46:
mov.f32 %f407, 0f3C08839E;
mov.f32 %f408, 0fB94CA1F9;
fma.rn.f32 %f1638, %f408, %f61, %f407;
bra.uni BB0_47;
BB0_45:
mov.f32 %f405, 0fBAB6061A;
mov.f32 %f406, 0f37CCF5CE;
fma.rn.f32 %f1638, %f406, %f61, %f405;
BB0_47:
@%p59 bra BB0_49;
bra.uni BB0_48;
BB0_49:
mov.f32 %f412, 0fBE2AAAA3;
fma.rn.f32 %f413, %f1638, %f61, %f412;
mov.f32 %f414, 0f00000000;
fma.rn.f32 %f1639, %f413, %f61, %f414;
bra.uni BB0_50;
BB0_48:
mov.f32 %f409, 0f3D2AAAA5;
fma.rn.f32 %f410, %f1638, %f61, %f409;
mov.f32 %f411, 0fBF000000;
fma.rn.f32 %f1639, %f410, %f61, %f411;
BB0_50:
fma.rn.f32 %f1640, %f1639, %f1637, %f1637;
@%p59 bra BB0_52;
mov.f32 %f415, 0f3F800000;
fma.rn.f32 %f1640, %f1639, %f61, %f415;
BB0_52:
and.b32 %r256, %r513, 2;
setp.eq.s32 %p62, %r256, 0;
@%p62 bra BB0_54;
mov.f32 %f416, 0f00000000;
mov.f32 %f417, 0fBF800000;
fma.rn.f32 %f1640, %f1640, %f417, %f416;
BB0_54:
max.f32 %f1622, %f363, %f289;
mul.f32 %f426, %f36, %f1634;
add.u64 %rd80, %SP, 4;
cvta.to.local.u64 %rd16, %rd80;
mul.f32 %f427, %f426, %f426;
mov.f32 %f428, 0f3F800000;
sub.f32 %f429, %f428, %f427;
mul.f32 %f430, %f36, %f1640;
mul.f32 %f431, %f430, %f430;
sub.f32 %f432, %f429, %f431;
mov.f32 %f433, 0f00000000;
max.f32 %f434, %f433, %f432;
sqrt.rn.f32 %f435, %f434;
mul.f32 %f436, %f20, %f430;
mul.f32 %f437, %f21, %f430;
mul.f32 %f438, %f22, %f430;
fma.rn.f32 %f439, %f29, %f426, %f436;
fma.rn.f32 %f440, %f28, %f426, %f437;
fma.rn.f32 %f441, %f27, %f426, %f438;
fma.rn.f32 %f442, %f7, %f435, %f439;
fma.rn.f32 %f443, %f8, %f435, %f440;
fma.rn.f32 %f444, %f9, %f435, %f441;
add.f32 %f445, %f7, %f442;
add.f32 %f446, %f8, %f443;
add.f32 %f447, %f9, %f444;
ld.global.f32 %f448, [shadowSpread];
mul.f32 %f449, %f448, %f445;
mul.f32 %f450, %f448, %f446;
mul.f32 %f451, %f448, %f447;
sub.f32 %f452, %f449, %f7;
sub.f32 %f453, %f450, %f8;
sub.f32 %f454, %f451, %f9;
mul.f32 %f455, %f453, %f453;
fma.rn.f32 %f456, %f452, %f452, %f455;
fma.rn.f32 %f457, %f454, %f454, %f456;
sqrt.rn.f32 %f458, %f457;
rcp.rn.f32 %f459, %f458;
mul.f32 %f73, %f459, %f452;
mul.f32 %f74, %f459, %f453;
mul.f32 %f75, %f459, %f454;
ld.global.u32 %r260, [imageEnabled];
and.b32 %r261, %r260, 32;
setp.eq.s32 %p63, %r261, 0;
selp.f32 %f460, 0f3F800000, 0f41200000, %p63;
mul.f32 %f424, %f460, %f1622;
mad.lo.s32 %r262, %r11, 1664525, 1013904223;
and.b32 %r263, %r262, 16777215;
cvt.rn.f32.u32 %f461, %r263;
mul.f32 %f462, %f461, 0f33800000;
fma.rn.f32 %f463, %f462, 0f40000000, 0fBF800000;
mad.lo.s32 %r83, %r262, 1664525, 1013904223;
and.b32 %r264, %r83, 16777215;
cvt.rn.f32.u32 %f464, %r264;
mul.f32 %f465, %f464, 0f33800000;
fma.rn.f32 %f466, %f465, 0f40000000, 0fBF800000;
mul.f32 %f467, %f13, %f466;
mul.f32 %f468, %f14, %f466;
mul.f32 %f469, %f15, %f466;
fma.rn.f32 %f470, %f16, %f463, %f467;
fma.rn.f32 %f471, %f17, %f463, %f468;
fma.rn.f32 %f472, %f18, %f463, %f469;
mul.f32 %f473, %f471, %f471;
fma.rn.f32 %f474, %f470, %f470, %f473;
fma.rn.f32 %f475, %f472, %f472, %f474;
sqrt.rn.f32 %f476, %f475;
rcp.rn.f32 %f477, %f476;
mul.f32 %f478, %f470, %f477;
mul.f32 %f479, %f471, %f477;
mul.f32 %f480, %f472, %f477;
mul.f32 %f481, %f1624, %f480;
mul.f32 %f482, %f1625, %f479;
sub.f32 %f483, %f481, %f482;
mul.f32 %f484, %f1625, %f478;
mul.f32 %f485, %f1623, %f480;
sub.f32 %f486, %f484, %f485;
mul.f32 %f487, %f1623, %f479;
mul.f32 %f488, %f1624, %f478;
sub.f32 %f489, %f487, %f488;
mul.f32 %f490, %f478, 0f3E24A490;
mul.f32 %f491, %f479, 0f3E24A490;
mul.f32 %f492, %f480, 0f3E24A490;
fma.rn.f32 %f493, %f19, %f490, %f10;
fma.rn.f32 %f494, %f19, %f491, %f11;
fma.rn.f32 %f495, %f19, %f492, %f12;
mul.f32 %f496, %f483, 0fBECECEDE;
mul.f32 %f497, %f486, 0fBECECEDE;
mul.f32 %f498, %f489, 0fBECECEDE;
fma.rn.f32 %f77, %f19, %f496, %f493;
fma.rn.f32 %f78, %f19, %f497, %f494;
fma.rn.f32 %f79, %f19, %f498, %f495;
sub.f32 %f499, %f77, %f10;
sub.f32 %f500, %f78, %f11;
sub.f32 %f501, %f79, %f12;
mul.f32 %f502, %f500, %f500;
fma.rn.f32 %f503, %f499, %f499, %f502;
fma.rn.f32 %f504, %f501, %f501, %f503;
sqrt.rn.f32 %f425, %f504;
rcp.rn.f32 %f505, %f425;
mul.f32 %f421, %f505, %f499;
mul.f32 %f422, %f505, %f500;
mul.f32 %f423, %f505, %f501;
mov.u32 %r265, 1065353216;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r257, [root];
mov.u32 %r258, 1;
// inline asm
call _rt_trace_64, (%r257, %f10, %f11, %f12, %f421, %f422, %f423, %r258, %f424, %f425, %rd80, %r114);
// inline asm
ld.local.f32 %f506, [%rd16];
mov.u16 %rs175, 0;
setp.neu.f32 %p64, %f506, 0f3F800000;
@%p64 bra BB0_56;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r266, [root];
mov.f32 %f514, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r266, %f77, %f78, %f79, %f73, %f74, %f75, %r258, %f424, %f514, %rd80, %r114);
// inline asm
ld.local.f32 %f515, [%rd16];
add.f32 %f1629, %f1629, %f515;
add.f32 %f1628, %f1628, 0f3F800000;
mov.u16 %rs175, 1;
BB0_56:
mad.lo.s32 %r273, %r83, 1664525, 1013904223;
and.b32 %r274, %r273, 16777215;
cvt.rn.f32.u32 %f524, %r274;
mul.f32 %f525, %f524, 0f33800000;
fma.rn.f32 %f526, %f525, 0f40000000, 0fBF800000;
mad.lo.s32 %r84, %r273, 1664525, 1013904223;
and.b32 %r275, %r84, 16777215;
cvt.rn.f32.u32 %f527, %r275;
mul.f32 %f528, %f527, 0f33800000;
fma.rn.f32 %f529, %f528, 0f40000000, 0fBF800000;
mul.f32 %f530, %f13, %f529;
mul.f32 %f531, %f14, %f529;
mul.f32 %f532, %f15, %f529;
fma.rn.f32 %f533, %f16, %f526, %f530;
fma.rn.f32 %f534, %f17, %f526, %f531;
fma.rn.f32 %f535, %f18, %f526, %f532;
mul.f32 %f536, %f534, %f534;
fma.rn.f32 %f537, %f533, %f533, %f536;
fma.rn.f32 %f538, %f535, %f535, %f537;
sqrt.rn.f32 %f539, %f538;
rcp.rn.f32 %f540, %f539;
mul.f32 %f541, %f533, %f540;
mul.f32 %f542, %f534, %f540;
mul.f32 %f543, %f535, %f540;
mul.f32 %f544, %f1624, %f543;
mul.f32 %f545, %f1625, %f542;
sub.f32 %f546, %f544, %f545;
mul.f32 %f547, %f1625, %f541;
mul.f32 %f548, %f1623, %f543;
sub.f32 %f549, %f547, %f548;
mul.f32 %f550, %f1623, %f542;
mul.f32 %f551, %f1624, %f541;
sub.f32 %f552, %f550, %f551;
mul.f32 %f553, %f541, 0fBE24A490;
mul.f32 %f554, %f542, 0fBE24A490;
mul.f32 %f555, %f543, 0fBE24A490;
fma.rn.f32 %f556, %f19, %f553, %f10;
fma.rn.f32 %f557, %f19, %f554, %f11;
fma.rn.f32 %f558, %f19, %f555, %f12;
mul.f32 %f559, %f546, 0f3ECACABC;
mul.f32 %f560, %f549, 0f3ECACABC;
mul.f32 %f561, %f552, 0f3ECACABC;
fma.rn.f32 %f84, %f19, %f559, %f556;
fma.rn.f32 %f85, %f19, %f560, %f557;
fma.rn.f32 %f86, %f19, %f561, %f558;
sub.f32 %f562, %f84, %f10;
sub.f32 %f563, %f85, %f11;
sub.f32 %f564, %f86, %f12;
mul.f32 %f565, %f563, %f563;
fma.rn.f32 %f566, %f562, %f562, %f565;
fma.rn.f32 %f567, %f564, %f564, %f566;
sqrt.rn.f32 %f523, %f567;
rcp.rn.f32 %f568, %f523;
mul.f32 %f519, %f568, %f562;
mul.f32 %f520, %f568, %f563;
mul.f32 %f521, %f568, %f564;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r270, [root];
// inline asm
call _rt_trace_64, (%r270, %f10, %f11, %f12, %f519, %f520, %f521, %r258, %f424, %f523, %rd80, %r114);
// inline asm
ld.local.f32 %f569, [%rd16];
setp.neu.f32 %p65, %f569, 0f3F800000;
@%p65 bra BB0_58;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r277, [root];
mov.f32 %f577, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r277, %f84, %f85, %f86, %f73, %f74, %f75, %r258, %f424, %f577, %rd80, %r114);
// inline asm
ld.local.f32 %f578, [%rd16];
add.f32 %f1629, %f1629, %f578;
add.f32 %f1628, %f1628, 0f3F800000;
mov.u16 %rs175, 1;
BB0_58:
mad.lo.s32 %r284, %r84, 1664525, 1013904223;
and.b32 %r285, %r284, 16777215;
cvt.rn.f32.u32 %f587, %r285;
mul.f32 %f588, %f587, 0f33800000;
fma.rn.f32 %f589, %f588, 0f40000000, 0fBF800000;
mad.lo.s32 %r85, %r284, 1664525, 1013904223;
and.b32 %r286, %r85, 16777215;
cvt.rn.f32.u32 %f590, %r286;
mul.f32 %f591, %f590, 0f33800000;
fma.rn.f32 %f592, %f591, 0f40000000, 0fBF800000;
mul.f32 %f593, %f13, %f592;
mul.f32 %f594, %f14, %f592;
mul.f32 %f595, %f15, %f592;
fma.rn.f32 %f596, %f16, %f589, %f593;
fma.rn.f32 %f597, %f17, %f589, %f594;
fma.rn.f32 %f598, %f18, %f589, %f595;
mul.f32 %f599, %f597, %f597;
fma.rn.f32 %f600, %f596, %f596, %f599;
fma.rn.f32 %f601, %f598, %f598, %f600;
sqrt.rn.f32 %f602, %f601;
rcp.rn.f32 %f603, %f602;
mul.f32 %f604, %f596, %f603;
mul.f32 %f605, %f597, %f603;
mul.f32 %f606, %f598, %f603;
mul.f32 %f607, %f1624, %f606;
mul.f32 %f608, %f1625, %f605;
sub.f32 %f609, %f607, %f608;
mul.f32 %f610, %f1625, %f604;
mul.f32 %f611, %f1623, %f606;
sub.f32 %f612, %f610, %f611;
mul.f32 %f613, %f1623, %f605;
mul.f32 %f614, %f1624, %f604;
sub.f32 %f615, %f613, %f614;
mul.f32 %f616, %f604, 0f3F23A3A8;
mul.f32 %f617, %f605, 0f3F23A3A8;
mul.f32 %f618, %f606, 0f3F23A3A8;
fma.rn.f32 %f619, %f19, %f616, %f10;
fma.rn.f32 %f620, %f19, %f617, %f11;
fma.rn.f32 %f621, %f19, %f618, %f12;
mul.f32 %f622, %f609, 0f3E24A490;
mul.f32 %f623, %f612, 0f3E24A490;
mul.f32 %f624, %f615, 0f3E24A490;
fma.rn.f32 %f91, %f19, %f622, %f619;
fma.rn.f32 %f92, %f19, %f623, %f620;
fma.rn.f32 %f93, %f19, %f624, %f621;
sub.f32 %f625, %f91, %f10;
sub.f32 %f626, %f92, %f11;
sub.f32 %f627, %f93, %f12;
mul.f32 %f628, %f626, %f626;
fma.rn.f32 %f629, %f625, %f625, %f628;
fma.rn.f32 %f630, %f627, %f627, %f629;
sqrt.rn.f32 %f586, %f630;
rcp.rn.f32 %f631, %f586;
mul.f32 %f582, %f631, %f625;
mul.f32 %f583, %f631, %f626;
mul.f32 %f584, %f631, %f627;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r281, [root];
// inline asm
call _rt_trace_64, (%r281, %f10, %f11, %f12, %f582, %f583, %f584, %r258, %f424, %f586, %rd80, %r114);
// inline asm
ld.local.f32 %f632, [%rd16];
setp.neu.f32 %p66, %f632, 0f3F800000;
@%p66 bra BB0_60;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r288, [root];
mov.f32 %f640, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r288, %f91, %f92, %f93, %f73, %f74, %f75, %r258, %f424, %f640, %rd80, %r114);
// inline asm
ld.local.f32 %f641, [%rd16];
add.f32 %f1629, %f1629, %f641;
add.f32 %f1628, %f1628, 0f3F800000;
mov.u16 %rs175, 1;
BB0_60:
mad.lo.s32 %r295, %r85, 1664525, 1013904223;
and.b32 %r296, %r295, 16777215;
cvt.rn.f32.u32 %f650, %r296;
mul.f32 %f651, %f650, 0f33800000;
fma.rn.f32 %f652, %f651, 0f40000000, 0fBF800000;
mad.lo.s32 %r86, %r295, 1664525, 1013904223;
and.b32 %r297, %r86, 16777215;
cvt.rn.f32.u32 %f653, %r297;
mul.f32 %f654, %f653, 0f33800000;
fma.rn.f32 %f655, %f654, 0f40000000, 0fBF800000;
mul.f32 %f656, %f13, %f655;
mul.f32 %f657, %f14, %f655;
mul.f32 %f658, %f15, %f655;
fma.rn.f32 %f659, %f16, %f652, %f656;
fma.rn.f32 %f660, %f17, %f652, %f657;
fma.rn.f32 %f661, %f18, %f652, %f658;
mul.f32 %f662, %f660, %f660;
fma.rn.f32 %f663, %f659, %f659, %f662;
fma.rn.f32 %f664, %f661, %f661, %f663;
sqrt.rn.f32 %f665, %f664;
rcp.rn.f32 %f666, %f665;
mul.f32 %f667, %f659, %f666;
mul.f32 %f668, %f660, %f666;
mul.f32 %f669, %f661, %f666;
mul.f32 %f670, %f1624, %f669;
mul.f32 %f671, %f1625, %f668;
sub.f32 %f672, %f670, %f671;
mul.f32 %f673, %f1625, %f667;
mul.f32 %f674, %f1623, %f669;
sub.f32 %f675, %f673, %f674;
mul.f32 %f676, %f1623, %f668;
mul.f32 %f677, %f1624, %f667;
sub.f32 %f678, %f676, %f677;
mul.f32 %f679, %f667, 0fBECECEDE;
mul.f32 %f680, %f668, 0fBECECEDE;
mul.f32 %f681, %f669, 0fBECECEDE;
fma.rn.f32 %f682, %f19, %f679, %f10;
fma.rn.f32 %f683, %f19, %f680, %f11;
fma.rn.f32 %f684, %f19, %f681, %f12;
mul.f32 %f685, %f672, 0fBF23A3A9;
mul.f32 %f686, %f675, 0fBF23A3A9;
mul.f32 %f687, %f678, 0fBF23A3A9;
fma.rn.f32 %f98, %f19, %f685, %f682;
fma.rn.f32 %f99, %f19, %f686, %f683;
fma.rn.f32 %f100, %f19, %f687, %f684;
sub.f32 %f688, %f98, %f10;
sub.f32 %f689, %f99, %f11;
sub.f32 %f690, %f100, %f12;
mul.f32 %f691, %f689, %f689;
fma.rn.f32 %f692, %f688, %f688, %f691;
fma.rn.f32 %f693, %f690, %f690, %f692;
sqrt.rn.f32 %f649, %f693;
rcp.rn.f32 %f694, %f649;
mul.f32 %f645, %f694, %f688;
mul.f32 %f646, %f694, %f689;
mul.f32 %f647, %f694, %f690;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r292, [root];
// inline asm
call _rt_trace_64, (%r292, %f10, %f11, %f12, %f645, %f646, %f647, %r258, %f424, %f649, %rd80, %r114);
// inline asm
ld.local.f32 %f695, [%rd16];
setp.neu.f32 %p67, %f695, 0f3F800000;
@%p67 bra BB0_62;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r299, [root];
mov.f32 %f703, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r299, %f98, %f99, %f100, %f73, %f74, %f75, %r258, %f424, %f703, %rd80, %r114);
// inline asm
ld.local.f32 %f704, [%rd16];
add.f32 %f1629, %f1629, %f704;
add.f32 %f1628, %f1628, 0f3F800000;
mov.u16 %rs175, 1;
BB0_62:
mad.lo.s32 %r306, %r86, 1664525, 1013904223;
and.b32 %r307, %r306, 16777215;
cvt.rn.f32.u32 %f713, %r307;
mul.f32 %f714, %f713, 0f33800000;
fma.rn.f32 %f715, %f714, 0f40000000, 0fBF800000;
mad.lo.s32 %r87, %r306, 1664525, 1013904223;
and.b32 %r308, %r87, 16777215;
cvt.rn.f32.u32 %f716, %r308;
mul.f32 %f717, %f716, 0f33800000;
fma.rn.f32 %f718, %f717, 0f40000000, 0fBF800000;
mul.f32 %f719, %f13, %f718;
mul.f32 %f720, %f14, %f718;
mul.f32 %f721, %f15, %f718;
fma.rn.f32 %f722, %f16, %f715, %f719;
fma.rn.f32 %f723, %f17, %f715, %f720;
fma.rn.f32 %f724, %f18, %f715, %f721;
mul.f32 %f725, %f723, %f723;
fma.rn.f32 %f726, %f722, %f722, %f725;
fma.rn.f32 %f727, %f724, %f724, %f726;
sqrt.rn.f32 %f728, %f727;
rcp.rn.f32 %f729, %f728;
mul.f32 %f730, %f722, %f729;
mul.f32 %f731, %f723, %f729;
mul.f32 %f732, %f724, %f729;
mul.f32 %f733, %f1624, %f732;
mul.f32 %f734, %f1625, %f731;
sub.f32 %f735, %f733, %f734;
mul.f32 %f736, %f1625, %f730;
mul.f32 %f737, %f1623, %f732;
sub.f32 %f738, %f736, %f737;
mul.f32 %f739, %f1623, %f731;
mul.f32 %f740, %f1624, %f730;
sub.f32 %f741, %f739, %f740;
mul.f32 %f742, %f730, 0fBF23A3A9;
mul.f32 %f743, %f731, 0fBF23A3A9;
mul.f32 %f744, %f732, 0fBF23A3A9;
fma.rn.f32 %f745, %f19, %f742, %f10;
fma.rn.f32 %f746, %f19, %f743, %f11;
fma.rn.f32 %f747, %f19, %f744, %f12;
mul.f32 %f748, %f735, 0f3F23A3A8;
mul.f32 %f749, %f738, 0f3F23A3A8;
mul.f32 %f750, %f741, 0f3F23A3A8;
fma.rn.f32 %f105, %f19, %f748, %f745;
fma.rn.f32 %f106, %f19, %f749, %f746;
fma.rn.f32 %f107, %f19, %f750, %f747;
sub.f32 %f751, %f105, %f10;
sub.f32 %f752, %f106, %f11;
sub.f32 %f753, %f107, %f12;
mul.f32 %f754, %f752, %f752;
fma.rn.f32 %f755, %f751, %f751, %f754;
fma.rn.f32 %f756, %f753, %f753, %f755;
sqrt.rn.f32 %f712, %f756;
rcp.rn.f32 %f757, %f712;
mul.f32 %f708, %f757, %f751;
mul.f32 %f709, %f757, %f752;
mul.f32 %f710, %f757, %f753;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r303, [root];
// inline asm
call _rt_trace_64, (%r303, %f10, %f11, %f12, %f708, %f709, %f710, %r258, %f424, %f712, %rd80, %r114);
// inline asm
ld.local.f32 %f758, [%rd16];
setp.neu.f32 %p68, %f758, 0f3F800000;
@%p68 bra BB0_64;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r310, [root];
mov.f32 %f766, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r310, %f105, %f106, %f107, %f73, %f74, %f75, %r258, %f424, %f766, %rd80, %r114);
// inline asm
ld.local.f32 %f767, [%rd16];
add.f32 %f1629, %f1629, %f767;
add.f32 %f1628, %f1628, 0f3F800000;
mov.u16 %rs175, 1;
BB0_64:
mad.lo.s32 %r317, %r87, 1664525, 1013904223;
and.b32 %r318, %r317, 16777215;
cvt.rn.f32.u32 %f776, %r318;
mul.f32 %f777, %f776, 0f33800000;
fma.rn.f32 %f778, %f777, 0f40000000, 0fBF800000;
mad.lo.s32 %r88, %r317, 1664525, 1013904223;
and.b32 %r319, %r88, 16777215;
cvt.rn.f32.u32 %f779, %r319;
mul.f32 %f780, %f779, 0f33800000;
fma.rn.f32 %f781, %f780, 0f40000000, 0fBF800000;
mul.f32 %f782, %f13, %f781;
mul.f32 %f783, %f14, %f781;
mul.f32 %f784, %f15, %f781;
fma.rn.f32 %f785, %f16, %f778, %f782;
fma.rn.f32 %f786, %f17, %f778, %f783;
fma.rn.f32 %f787, %f18, %f778, %f784;
mul.f32 %f788, %f786, %f786;
fma.rn.f32 %f789, %f785, %f785, %f788;
fma.rn.f32 %f790, %f787, %f787, %f789;
sqrt.rn.f32 %f791, %f790;
rcp.rn.f32 %f792, %f791;
mul.f32 %f793, %f785, %f792;
mul.f32 %f794, %f786, %f792;
mul.f32 %f795, %f787, %f792;
mul.f32 %f796, %f1624, %f795;
mul.f32 %f797, %f1625, %f794;
sub.f32 %f798, %f796, %f797;
mul.f32 %f799, %f1625, %f793;
mul.f32 %f800, %f1623, %f795;
sub.f32 %f801, %f799, %f800;
mul.f32 %f802, %f1623, %f794;
mul.f32 %f803, %f1624, %f793;
sub.f32 %f804, %f802, %f803;
mul.f32 %f805, %f793, 0fBF61E1E3;
mul.f32 %f806, %f794, 0fBF61E1E3;
mul.f32 %f807, %f795, 0fBF61E1E3;
fma.rn.f32 %f808, %f19, %f805, %f10;
fma.rn.f32 %f809, %f19, %f806, %f11;
fma.rn.f32 %f810, %f19, %f807, %f12;
mul.f32 %f811, %f798, 0fBE24A490;
mul.f32 %f812, %f801, 0fBE24A490;
mul.f32 %f813, %f804, 0fBE24A490;
fma.rn.f32 %f112, %f19, %f811, %f808;
fma.rn.f32 %f113, %f19, %f812, %f809;
fma.rn.f32 %f114, %f19, %f813, %f810;
sub.f32 %f814, %f112, %f10;
sub.f32 %f815, %f113, %f11;
sub.f32 %f816, %f114, %f12;
mul.f32 %f817, %f815, %f815;
fma.rn.f32 %f818, %f814, %f814, %f817;
fma.rn.f32 %f819, %f816, %f816, %f818;
sqrt.rn.f32 %f775, %f819;
rcp.rn.f32 %f820, %f775;
mul.f32 %f771, %f820, %f814;
mul.f32 %f772, %f820, %f815;
mul.f32 %f773, %f820, %f816;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r314, [root];
// inline asm
call _rt_trace_64, (%r314, %f10, %f11, %f12, %f771, %f772, %f773, %r258, %f424, %f775, %rd80, %r114);
// inline asm
ld.local.f32 %f821, [%rd16];
setp.neu.f32 %p69, %f821, 0f3F800000;
@%p69 bra BB0_66;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r321, [root];
mov.f32 %f829, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r321, %f112, %f113, %f114, %f73, %f74, %f75, %r258, %f424, %f829, %rd80, %r114);
// inline asm
ld.local.f32 %f830, [%rd16];
add.f32 %f1629, %f1629, %f830;
add.f32 %f1628, %f1628, 0f3F800000;
mov.u16 %rs175, 1;
BB0_66:
mad.lo.s32 %r328, %r88, 1664525, 1013904223;
and.b32 %r329, %r328, 16777215;
cvt.rn.f32.u32 %f839, %r329;
mul.f32 %f840, %f839, 0f33800000;
fma.rn.f32 %f841, %f840, 0f40000000, 0fBF800000;
mad.lo.s32 %r89, %r328, 1664525, 1013904223;
and.b32 %r330, %r89, 16777215;
cvt.rn.f32.u32 %f842, %r330;
mul.f32 %f843, %f842, 0f33800000;
fma.rn.f32 %f844, %f843, 0f40000000, 0fBF800000;
mul.f32 %f845, %f13, %f844;
mul.f32 %f846, %f14, %f844;
mul.f32 %f847, %f15, %f844;
fma.rn.f32 %f848, %f16, %f841, %f845;
fma.rn.f32 %f849, %f17, %f841, %f846;
fma.rn.f32 %f850, %f18, %f841, %f847;
mul.f32 %f851, %f849, %f849;
fma.rn.f32 %f852, %f848, %f848, %f851;
fma.rn.f32 %f853, %f850, %f850, %f852;
sqrt.rn.f32 %f854, %f853;
rcp.rn.f32 %f855, %f854;
mul.f32 %f856, %f848, %f855;
mul.f32 %f857, %f849, %f855;
mul.f32 %f858, %f850, %f855;
mul.f32 %f859, %f1624, %f858;
mul.f32 %f860, %f1625, %f857;
sub.f32 %f861, %f859, %f860;
mul.f32 %f862, %f1625, %f856;
mul.f32 %f863, %f1623, %f858;
sub.f32 %f864, %f862, %f863;
mul.f32 %f865, %f1623, %f857;
mul.f32 %f866, %f1624, %f856;
sub.f32 %f867, %f865, %f866;
mul.f32 %f868, %f856, 0f3ECACABC;
mul.f32 %f869, %f857, 0f3ECACABC;
mul.f32 %f870, %f858, 0f3ECACABC;
fma.rn.f32 %f871, %f19, %f868, %f10;
fma.rn.f32 %f872, %f19, %f869, %f11;
fma.rn.f32 %f873, %f19, %f870, %f12;
mul.f32 %f874, %f861, 0f3F61E1D2;
mul.f32 %f875, %f864, 0f3F61E1D2;
mul.f32 %f876, %f867, 0f3F61E1D2;
fma.rn.f32 %f119, %f19, %f874, %f871;
fma.rn.f32 %f120, %f19, %f875, %f872;
fma.rn.f32 %f121, %f19, %f876, %f873;
sub.f32 %f877, %f119, %f10;
sub.f32 %f878, %f120, %f11;
sub.f32 %f879, %f121, %f12;
mul.f32 %f880, %f878, %f878;
fma.rn.f32 %f881, %f877, %f877, %f880;
fma.rn.f32 %f882, %f879, %f879, %f881;
sqrt.rn.f32 %f838, %f882;
rcp.rn.f32 %f883, %f838;
mul.f32 %f834, %f883, %f877;
mul.f32 %f835, %f883, %f878;
mul.f32 %f836, %f883, %f879;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r325, [root];
// inline asm
call _rt_trace_64, (%r325, %f10, %f11, %f12, %f834, %f835, %f836, %r258, %f424, %f838, %rd80, %r114);
// inline asm
ld.local.f32 %f884, [%rd16];
setp.neu.f32 %p70, %f884, 0f3F800000;
@%p70 bra BB0_68;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r332, [root];
mov.f32 %f892, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r332, %f119, %f120, %f121, %f73, %f74, %f75, %r258, %f424, %f892, %rd80, %r114);
// inline asm
ld.local.f32 %f893, [%rd16];
add.f32 %f1629, %f1629, %f893;
add.f32 %f1628, %f1628, 0f3F800000;
mov.u16 %rs175, 1;
BB0_68:
mad.lo.s32 %r339, %r89, 1664525, 1013904223;
and.b32 %r340, %r339, 16777215;
cvt.rn.f32.u32 %f902, %r340;
mul.f32 %f903, %f902, 0f33800000;
fma.rn.f32 %f904, %f903, 0f40000000, 0fBF800000;
mad.lo.s32 %r515, %r339, 1664525, 1013904223;
and.b32 %r341, %r515, 16777215;
cvt.rn.f32.u32 %f905, %r341;
mul.f32 %f906, %f905, 0f33800000;
fma.rn.f32 %f907, %f906, 0f40000000, 0fBF800000;
mul.f32 %f908, %f13, %f907;
mul.f32 %f909, %f14, %f907;
mul.f32 %f910, %f15, %f907;
fma.rn.f32 %f911, %f16, %f904, %f908;
fma.rn.f32 %f912, %f17, %f904, %f909;
fma.rn.f32 %f913, %f18, %f904, %f910;
mul.f32 %f914, %f912, %f912;
fma.rn.f32 %f915, %f911, %f911, %f914;
fma.rn.f32 %f916, %f913, %f913, %f915;
sqrt.rn.f32 %f917, %f916;
rcp.rn.f32 %f918, %f917;
mul.f32 %f919, %f911, %f918;
mul.f32 %f920, %f912, %f918;
mul.f32 %f921, %f913, %f918;
mul.f32 %f922, %f1624, %f921;
mul.f32 %f923, %f1625, %f920;
sub.f32 %f924, %f922, %f923;
mul.f32 %f925, %f1625, %f919;
mul.f32 %f926, %f1623, %f921;
sub.f32 %f927, %f925, %f926;
mul.f32 %f928, %f1623, %f920;
mul.f32 %f929, %f1624, %f919;
sub.f32 %f930, %f928, %f929;
mul.f32 %f931, %f919, 0f3F61E1D2;
mul.f32 %f932, %f920, 0f3F61E1D2;
mul.f32 %f933, %f921, 0f3F61E1D2;
fma.rn.f32 %f934, %f19, %f931, %f10;
fma.rn.f32 %f935, %f19, %f932, %f11;
fma.rn.f32 %f936, %f19, %f933, %f12;
mul.f32 %f937, %f924, 0fBF61E1E3;
mul.f32 %f938, %f927, 0fBF61E1E3;
mul.f32 %f939, %f930, 0fBF61E1E3;
fma.rn.f32 %f126, %f19, %f937, %f934;
fma.rn.f32 %f127, %f19, %f938, %f935;
fma.rn.f32 %f128, %f19, %f939, %f936;
sub.f32 %f940, %f126, %f10;
sub.f32 %f941, %f127, %f11;
sub.f32 %f942, %f128, %f12;
mul.f32 %f943, %f941, %f941;
fma.rn.f32 %f944, %f940, %f940, %f943;
fma.rn.f32 %f945, %f942, %f942, %f944;
sqrt.rn.f32 %f901, %f945;
rcp.rn.f32 %f946, %f901;
mul.f32 %f897, %f946, %f940;
mul.f32 %f898, %f946, %f941;
mul.f32 %f899, %f946, %f942;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r336, [root];
// inline asm
call _rt_trace_64, (%r336, %f10, %f11, %f12, %f897, %f898, %f899, %r258, %f424, %f901, %rd80, %r114);
// inline asm
ld.local.f32 %f947, [%rd16];
setp.neu.f32 %p71, %f947, 0f3F800000;
@%p71 bra BB0_70;
st.local.u32 [%rd16], %r265;
ld.global.u32 %r343, [root];
mov.f32 %f955, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r343, %f126, %f127, %f128, %f73, %f74, %f75, %r258, %f424, %f955, %rd80, %r114);
// inline asm
ld.local.f32 %f956, [%rd16];
add.f32 %f1629, %f1629, %f956;
add.f32 %f1628, %f1628, 0f3F800000;
mov.u16 %rs175, 1;
BB0_70:
setp.ne.s16 %p72, %rs175, 0;
@%p72 bra BB0_72;
add.u64 %rd96, %SP, 0;
cvta.to.local.u64 %rd97, %rd96;
st.local.u32 [%rd97], %r265;
ld.global.u32 %r347, [root];
mov.f32 %f964, 0f6C4ECB8F;
// inline asm
call _rt_trace_64, (%r347, %f10, %f11, %f12, %f73, %f74, %f75, %r258, %f424, %f964, %rd96, %r114);
// inline asm
ld.local.f32 %f965, [%rd97];
add.f32 %f1629, %f1629, %f965;
add.f32 %f1628, %f1628, 0f3F800000;
BB0_72:
ld.global.u32 %r489, [samples];
add.s32 %r492, %r492, 1;
setp.lt.s32 %p73, %r492, %r489;
@%p73 bra BB0_8;
BB0_73:
add.s32 %r490, %r490, 1;
setp.lt.s32 %p74, %r490, %r489;
@%p74 bra BB0_6;
bra.uni BB0_74;
BB0_4:
mov.f32 %f1629, %f1628;
BB0_74:
mul.f32 %f967, %f1625, %f9;
mul.f32 %f968, %f1624, %f8;
neg.f32 %f969, %f968;
mul.f32 %f970, %f1623, %f7;
sub.f32 %f971, %f969, %f970;
sub.f32 %f141, %f971, %f967;
setp.eq.s32 %p75, %r489, 0;
mov.f32 %f1664, 0f3F800000;
@%p75 bra BB0_76;
mov.f32 %f972, 0f3F800000;
max.f32 %f973, %f1628, %f972;
div.rn.f32 %f1664, %f1629, %f973;
BB0_76:
ld.global.f32 %f976, [directColor];
mul.f32 %f144, %f1664, %f976;
ld.global.f32 %f977, [directColor+4];
mul.f32 %f145, %f1664, %f977;
ld.global.f32 %f978, [directColor+8];
mul.f32 %f146, %f1664, %f978;
cvt.sat.f32.f32 %f979, %f141;
mul.f32 %f147, %f144, %f979;
mul.f32 %f148, %f145, %f979;
mul.f32 %f149, %f146, %f979;
fma.rn.f32 %f980, %f141, 0f3F000000, 0f3F000000;
cvt.sat.f32.f32 %f981, %f980;
add.f32 %f150, %f981, %f981;
mov.f32 %f982, 0f41200000;
cvt.rzi.f32.f32 %f983, %f982;
add.f32 %f984, %f983, %f983;
mov.f32 %f985, 0f41A00000;
sub.f32 %f986, %f985, %f984;
abs.f32 %f151, %f986;
abs.f32 %f152, %f150;
setp.lt.f32 %p76, %f152, 0f00800000;
mul.f32 %f987, %f152, 0f4B800000;
selp.f32 %f988, 0fC3170000, 0fC2FE0000, %p76;
selp.f32 %f989, %f987, %f152, %p76;
mov.b32 %r351, %f989;
and.b32 %r352, %r351, 8388607;
or.b32 %r353, %r352, 1065353216;
mov.b32 %f990, %r353;
shr.u32 %r354, %r351, 23;
cvt.rn.f32.u32 %f991, %r354;
add.f32 %f992, %f988, %f991;
setp.gt.f32 %p77, %f990, 0f3FB504F3;
mul.f32 %f993, %f990, 0f3F000000;
add.f32 %f994, %f992, 0f3F800000;
selp.f32 %f995, %f993, %f990, %p77;
selp.f32 %f996, %f994, %f992, %p77;
add.f32 %f997, %f995, 0fBF800000;
add.f32 %f975, %f995, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f974,%f975;
// inline asm
add.f32 %f998, %f997, %f997;
mul.f32 %f999, %f974, %f998;
mul.f32 %f1000, %f999, %f999;
mov.f32 %f1001, 0f3C4CAF63;
mov.f32 %f1002, 0f3B18F0FE;
fma.rn.f32 %f1003, %f1002, %f1000, %f1001;
mov.f32 %f1004, 0f3DAAAABD;
fma.rn.f32 %f1005, %f1003, %f1000, %f1004;
mul.rn.f32 %f1006, %f1005, %f1000;
mul.rn.f32 %f1007, %f1006, %f999;
sub.f32 %f1008, %f997, %f999;
neg.f32 %f1009, %f999;
add.f32 %f1010, %f1008, %f1008;
fma.rn.f32 %f1011, %f1009, %f997, %f1010;
mul.rn.f32 %f1012, %f974, %f1011;
add.f32 %f1013, %f1007, %f999;
sub.f32 %f1014, %f999, %f1013;
add.f32 %f1015, %f1007, %f1014;
add.f32 %f1016, %f1012, %f1015;
add.f32 %f1017, %f1013, %f1016;
sub.f32 %f1018, %f1013, %f1017;
add.f32 %f1019, %f1016, %f1018;
mov.f32 %f1020, 0f3F317200;
mul.rn.f32 %f1021, %f996, %f1020;
mov.f32 %f1022, 0f35BFBE8E;
mul.rn.f32 %f1023, %f996, %f1022;
add.f32 %f1024, %f1021, %f1017;
sub.f32 %f1025, %f1021, %f1024;
add.f32 %f1026, %f1017, %f1025;
add.f32 %f1027, %f1019, %f1026;
add.f32 %f1028, %f1023, %f1027;
add.f32 %f1029, %f1024, %f1028;
sub.f32 %f1030, %f1024, %f1029;
add.f32 %f1031, %f1028, %f1030;
mul.rn.f32 %f1032, %f985, %f1029;
neg.f32 %f1033, %f1032;
fma.rn.f32 %f1034, %f985, %f1029, %f1033;
fma.rn.f32 %f1035, %f985, %f1031, %f1034;
mov.f32 %f1686, 0f00000000;
fma.rn.f32 %f1037, %f1686, %f1029, %f1035;
add.rn.f32 %f1038, %f1032, %f1037;
neg.f32 %f1039, %f1038;
add.rn.f32 %f1040, %f1032, %f1039;
add.rn.f32 %f1041, %f1040, %f1037;
mov.b32 %r355, %f1038;
setp.eq.s32 %p78, %r355, 1118925336;
add.s32 %r356, %r355, -1;
mov.b32 %f1042, %r356;
add.f32 %f1043, %f1041, 0f37000000;
selp.f32 %f1044, %f1042, %f1038, %p78;
selp.f32 %f153, %f1043, %f1041, %p78;
mul.f32 %f1045, %f1044, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1046, %f1045;
mov.f32 %f1047, 0fBF317200;
fma.rn.f32 %f1048, %f1046, %f1047, %f1044;
mov.f32 %f1049, 0fB5BFBE8E;
fma.rn.f32 %f1050, %f1046, %f1049, %f1048;
mul.f32 %f1051, %f1050, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1052, %f1051;
add.f32 %f1053, %f1046, 0f00000000;
ex2.approx.f32 %f1054, %f1053;
mul.f32 %f1055, %f1052, %f1054;
setp.lt.f32 %p79, %f1044, 0fC2D20000;
selp.f32 %f1056, 0f00000000, %f1055, %p79;
setp.gt.f32 %p80, %f1044, 0f42D20000;
selp.f32 %f1665, 0f7F800000, %f1056, %p80;
setp.eq.f32 %p81, %f1665, 0f7F800000;
@%p81 bra BB0_78;
fma.rn.f32 %f1665, %f1665, %f153, %f1665;
BB0_78:
setp.lt.f32 %p82, %f150, 0f00000000;
setp.eq.f32 %p83, %f151, 0f3F800000;
and.pred %p1, %p82, %p83;
mov.b32 %r357, %f1665;
xor.b32 %r358, %r357, -2147483648;
mov.b32 %f1057, %r358;
selp.f32 %f1667, %f1057, %f1665, %p1;
setp.eq.f32 %p84, %f150, 0f00000000;
@%p84 bra BB0_81;
bra.uni BB0_79;
BB0_81:
add.f32 %f1060, %f150, %f150;
selp.f32 %f1667, %f1060, 0f00000000, %p83;
bra.uni BB0_82;
BB0_79:
setp.geu.f32 %p85, %f150, 0f00000000;
@%p85 bra BB0_82;
cvt.rzi.f32.f32 %f1059, %f985;
setp.neu.f32 %p86, %f1059, 0f41A00000;
selp.f32 %f1667, 0f7FFFFFFF, %f1667, %p86;
BB0_82:
add.f32 %f1061, %f152, 0f41A00000;
mov.b32 %r359, %f1061;
setp.lt.s32 %p88, %r359, 2139095040;
@%p88 bra BB0_87;
setp.gtu.f32 %p89, %f152, 0f7F800000;
@%p89 bra BB0_86;
bra.uni BB0_84;
BB0_86:
add.f32 %f1667, %f150, 0f41A00000;
bra.uni BB0_87;
BB0_84:
setp.neu.f32 %p90, %f152, 0f7F800000;
@%p90 bra BB0_87;
selp.f32 %f1667, 0fFF800000, 0f7F800000, %p1;
BB0_87:
setp.eq.f32 %p91, %f150, 0f3F800000;
selp.f32 %f1062, 0f3F800000, %f1667, %p91;
cvt.sat.f32.f32 %f1063, %f1062;
mul.f32 %f164, %f144, %f1063;
mul.f32 %f165, %f145, %f1063;
mul.f32 %f166, %f146, %f1063;
ld.global.u32 %r518, [imageEnabled];
and.b32 %r360, %r518, 8;
setp.eq.s32 %p92, %r360, 0;
@%p92 bra BB0_100;
cvt.u64.u32 %rd100, %r2;
cvt.u64.u32 %rd101, %r3;
mov.u64 %rd104, image_Mask;
cvta.global.u64 %rd99, %rd104;
// inline asm
call (%rd98), _rt_buffer_get_64, (%rd99, %r113, %r113, %rd100, %rd101, %rd32, %rd32);
// inline asm
mov.f32 %f1066, 0f3E68BA2E;
cvt.rzi.f32.f32 %f1067, %f1066;
fma.rn.f32 %f1068, %f1067, 0fC0000000, 0f3EE8BA2E;
abs.f32 %f167, %f1068;
abs.f32 %f168, %f1664;
setp.lt.f32 %p93, %f168, 0f00800000;
mul.f32 %f1069, %f168, 0f4B800000;
selp.f32 %f1070, 0fC3170000, 0fC2FE0000, %p93;
selp.f32 %f1071, %f1069, %f168, %p93;
mov.b32 %r363, %f1071;
and.b32 %r364, %r363, 8388607;
or.b32 %r365, %r364, 1065353216;
mov.b32 %f1072, %r365;
shr.u32 %r366, %r363, 23;
cvt.rn.f32.u32 %f1073, %r366;
add.f32 %f1074, %f1070, %f1073;
setp.gt.f32 %p94, %f1072, 0f3FB504F3;
mul.f32 %f1075, %f1072, 0f3F000000;
add.f32 %f1076, %f1074, 0f3F800000;
selp.f32 %f1077, %f1075, %f1072, %p94;
selp.f32 %f1078, %f1076, %f1074, %p94;
add.f32 %f1079, %f1077, 0fBF800000;
add.f32 %f1065, %f1077, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1064,%f1065;
// inline asm
add.f32 %f1080, %f1079, %f1079;
mul.f32 %f1081, %f1064, %f1080;
mul.f32 %f1082, %f1081, %f1081;
fma.rn.f32 %f1085, %f1002, %f1082, %f1001;
fma.rn.f32 %f1087, %f1085, %f1082, %f1004;
mul.rn.f32 %f1088, %f1087, %f1082;
mul.rn.f32 %f1089, %f1088, %f1081;
sub.f32 %f1090, %f1079, %f1081;
neg.f32 %f1091, %f1081;
add.f32 %f1092, %f1090, %f1090;
fma.rn.f32 %f1093, %f1091, %f1079, %f1092;
mul.rn.f32 %f1094, %f1064, %f1093;
add.f32 %f1095, %f1089, %f1081;
sub.f32 %f1096, %f1081, %f1095;
add.f32 %f1097, %f1089, %f1096;
add.f32 %f1098, %f1094, %f1097;
add.f32 %f1099, %f1095, %f1098;
sub.f32 %f1100, %f1095, %f1099;
add.f32 %f1101, %f1098, %f1100;
mul.rn.f32 %f1103, %f1078, %f1020;
mul.rn.f32 %f1105, %f1078, %f1022;
add.f32 %f1106, %f1103, %f1099;
sub.f32 %f1107, %f1103, %f1106;
add.f32 %f1108, %f1099, %f1107;
add.f32 %f1109, %f1101, %f1108;
add.f32 %f1110, %f1105, %f1109;
add.f32 %f1111, %f1106, %f1110;
sub.f32 %f1112, %f1106, %f1111;
add.f32 %f1113, %f1110, %f1112;
mov.f32 %f1114, 0f3EE8BA2E;
mul.rn.f32 %f1115, %f1114, %f1111;
neg.f32 %f1116, %f1115;
fma.rn.f32 %f1117, %f1114, %f1111, %f1116;
fma.rn.f32 %f1118, %f1114, %f1113, %f1117;
fma.rn.f32 %f1120, %f1686, %f1111, %f1118;
add.rn.f32 %f1121, %f1115, %f1120;
neg.f32 %f1122, %f1121;
add.rn.f32 %f1123, %f1115, %f1122;
add.rn.f32 %f1124, %f1123, %f1120;
mov.b32 %r367, %f1121;
setp.eq.s32 %p95, %r367, 1118925336;
add.s32 %r368, %r367, -1;
mov.b32 %f1125, %r368;
add.f32 %f1126, %f1124, 0f37000000;
selp.f32 %f1127, %f1125, %f1121, %p95;
selp.f32 %f169, %f1126, %f1124, %p95;
mul.f32 %f1128, %f1127, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1129, %f1128;
fma.rn.f32 %f1131, %f1129, %f1047, %f1127;
fma.rn.f32 %f1133, %f1129, %f1049, %f1131;
mul.f32 %f1134, %f1133, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1135, %f1134;
add.f32 %f1136, %f1129, 0f00000000;
ex2.approx.f32 %f1137, %f1136;
mul.f32 %f1138, %f1135, %f1137;
setp.lt.f32 %p96, %f1127, 0fC2D20000;
selp.f32 %f1139, 0f00000000, %f1138, %p96;
setp.gt.f32 %p97, %f1127, 0f42D20000;
selp.f32 %f1668, 0f7F800000, %f1139, %p97;
setp.eq.f32 %p98, %f1668, 0f7F800000;
@%p98 bra BB0_90;
fma.rn.f32 %f1668, %f1668, %f169, %f1668;
BB0_90:
setp.lt.f32 %p99, %f1664, 0f00000000;
setp.eq.f32 %p100, %f167, 0f3F800000;
and.pred %p2, %p99, %p100;
mov.b32 %r369, %f1668;
xor.b32 %r370, %r369, -2147483648;
mov.b32 %f1140, %r370;
selp.f32 %f1670, %f1140, %f1668, %p2;
setp.eq.f32 %p101, %f1664, 0f00000000;
@%p101 bra BB0_93;
bra.uni BB0_91;
BB0_93:
add.f32 %f1143, %f1664, %f1664;
selp.f32 %f1670, %f1143, 0f00000000, %p100;
bra.uni BB0_94;
BB0_161:
mov.u64 %rd246, image_HDR;
cvta.global.u64 %rd241, %rd246;
mov.u32 %r467, 8;
mov.u64 %rd245, 0;
// inline asm
call (%rd240), _rt_buffer_get_64, (%rd241, %r113, %r467, %rd25, %rd26, %rd245, %rd245);
// inline asm
mov.f32 %f1599, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs133, %f1599;}
// inline asm
mov.u16 %rs134, 0;
st.v4.u16 [%rd240], {%rs133, %rs133, %rs133, %rs134};
BB0_162:
ld.global.u32 %r468, [additive];
setp.eq.s32 %p176, %r468, 0;
@%p176 bra BB0_164;
mov.u64 %rd259, image_RNM0;
cvta.global.u64 %rd248, %rd259;
mov.u32 %r472, 8;
mov.u64 %rd258, 0;
// inline asm
call (%rd247), _rt_buffer_get_64, (%rd248, %r113, %r472, %rd25, %rd26, %rd258, %rd258);
// inline asm
ld.v4.u16 {%rs141, %rs142, %rs143, %rs144}, [%rd247];
// inline asm
{ cvt.f32.f16 %f1600, %rs141;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1601, %rs142;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1602, %rs143;}
// inline asm
// inline asm
call (%rd253), _rt_buffer_get_64, (%rd248, %r113, %r472, %rd25, %rd26, %rd258, %rd258);
// inline asm
add.f32 %f1603, %f1600, 0f00000000;
add.f32 %f1604, %f1601, 0f00000000;
add.f32 %f1605, %f1602, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs140, %f1605;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs139, %f1604;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs138, %f1603;}
// inline asm
mov.u16 %rs145, 0;
st.v4.u16 [%rd253], {%rs138, %rs139, %rs140, %rs145};
bra.uni BB0_165;
BB0_164:
mov.u64 %rd266, image_RNM0;
cvta.global.u64 %rd261, %rd266;
mov.u32 %r474, 8;
mov.u64 %rd265, 0;
// inline asm
call (%rd260), _rt_buffer_get_64, (%rd261, %r113, %r474, %rd25, %rd26, %rd265, %rd265);
// inline asm
mov.f32 %f1606, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs146, %f1606;}
// inline asm
mov.u16 %rs147, 0;
st.v4.u16 [%rd260], {%rs146, %rs146, %rs146, %rs147};
BB0_165:
ld.global.u32 %r475, [additive];
setp.eq.s32 %p177, %r475, 0;
@%p177 bra BB0_167;
mov.u64 %rd279, image_RNM1;
cvta.global.u64 %rd268, %rd279;
mov.u32 %r479, 8;
mov.u64 %rd278, 0;
// inline asm
call (%rd267), _rt_buffer_get_64, (%rd268, %r113, %r479, %rd25, %rd26, %rd278, %rd278);
// inline asm
ld.v4.u16 {%rs154, %rs155, %rs156, %rs157}, [%rd267];
// inline asm
{ cvt.f32.f16 %f1607, %rs154;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1608, %rs155;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1609, %rs156;}
// inline asm
// inline asm
call (%rd273), _rt_buffer_get_64, (%rd268, %r113, %r479, %rd25, %rd26, %rd278, %rd278);
// inline asm
add.f32 %f1610, %f1607, 0f00000000;
add.f32 %f1611, %f1608, 0f00000000;
add.f32 %f1612, %f1609, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs153, %f1612;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs152, %f1611;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs151, %f1610;}
// inline asm
mov.u16 %rs158, 0;
st.v4.u16 [%rd273], {%rs151, %rs152, %rs153, %rs158};
bra.uni BB0_168;
BB0_167:
mov.u64 %rd286, image_RNM1;
cvta.global.u64 %rd281, %rd286;
mov.u32 %r481, 8;
mov.u64 %rd285, 0;
// inline asm
call (%rd280), _rt_buffer_get_64, (%rd281, %r113, %r481, %rd25, %rd26, %rd285, %rd285);
// inline asm
mov.f32 %f1613, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs159, %f1613;}
// inline asm
mov.u16 %rs160, 0;
st.v4.u16 [%rd280], {%rs159, %rs159, %rs159, %rs160};
BB0_168:
ld.global.u32 %r482, [additive];
setp.eq.s32 %p178, %r482, 0;
@%p178 bra BB0_170;
mov.u64 %rd299, image_RNM2;
cvta.global.u64 %rd288, %rd299;
mov.u32 %r486, 8;
mov.u64 %rd298, 0;
// inline asm
call (%rd287), _rt_buffer_get_64, (%rd288, %r113, %r486, %rd25, %rd26, %rd298, %rd298);
// inline asm
ld.v4.u16 {%rs167, %rs168, %rs169, %rs170}, [%rd287];
// inline asm
{ cvt.f32.f16 %f1614, %rs167;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1615, %rs168;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1616, %rs169;}
// inline asm
// inline asm
call (%rd293), _rt_buffer_get_64, (%rd288, %r113, %r486, %rd25, %rd26, %rd298, %rd298);
// inline asm
add.f32 %f1617, %f1614, 0f00000000;
add.f32 %f1618, %f1615, 0f00000000;
add.f32 %f1619, %f1616, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs166, %f1619;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs165, %f1618;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs164, %f1617;}
// inline asm
mov.u16 %rs171, 0;
st.v4.u16 [%rd293], {%rs164, %rs165, %rs166, %rs171};
bra.uni BB0_171;
BB0_170:
mov.u64 %rd306, image_RNM2;
cvta.global.u64 %rd301, %rd306;
mov.u32 %r488, 8;
mov.u64 %rd305, 0;
// inline asm
call (%rd300), _rt_buffer_get_64, (%rd301, %r113, %r488, %rd25, %rd26, %rd305, %rd305);
// inline asm
mov.f32 %f1620, 0f00000000;
// inline asm
{ cvt.rn.f16.f32 %rs172, %f1620;}
// inline asm
mov.u16 %rs173, 0;
st.v4.u16 [%rd300], {%rs172, %rs172, %rs172, %rs173};
bra.uni BB0_171;
BB0_91:
setp.geu.f32 %p102, %f1664, 0f00000000;
@%p102 bra BB0_94;
cvt.rzi.f32.f32 %f1142, %f1114;
setp.neu.f32 %p103, %f1142, 0f3EE8BA2E;
selp.f32 %f1670, 0f7FFFFFFF, %f1670, %p103;
BB0_94:
add.f32 %f1144, %f168, 0f3EE8BA2E;
mov.b32 %r371, %f1144;
setp.lt.s32 %p105, %r371, 2139095040;
@%p105 bra BB0_99;
setp.gtu.f32 %p106, %f168, 0f7F800000;
@%p106 bra BB0_98;
bra.uni BB0_96;
BB0_98:
add.f32 %f1670, %f1664, 0f3EE8BA2E;
bra.uni BB0_99;
BB0_96:
setp.neu.f32 %p107, %f168, 0f7F800000;
@%p107 bra BB0_99;
selp.f32 %f1670, 0fFF800000, 0f7F800000, %p2;
BB0_99:
mul.f32 %f1145, %f1670, 0f437F0000;
setp.eq.f32 %p108, %f1664, 0f3F800000;
selp.f32 %f1146, 0f437F0000, %f1145, %p108;
cvt.rzi.u32.f32 %r372, %f1146;
cvt.u16.u32 %rs30, %r372;
mov.u16 %rs31, 255;
st.v2.u8 [%rd98], {%rs30, %rs31};
ld.global.u32 %r518, [imageEnabled];
BB0_100:
and.b32 %r373, %r518, 1;
setp.eq.b32 %p109, %r373, 1;
@!%p109 bra BB0_135;
bra.uni BB0_101;
BB0_101:
mov.f32 %f1149, 0f3E666666;
cvt.rzi.f32.f32 %f1150, %f1149;
fma.rn.f32 %f1151, %f1150, 0fC0000000, 0f3EE66666;
abs.f32 %f180, %f1151;
abs.f32 %f181, %f147;
setp.lt.f32 %p110, %f181, 0f00800000;
mul.f32 %f1152, %f181, 0f4B800000;
selp.f32 %f1153, 0fC3170000, 0fC2FE0000, %p110;
selp.f32 %f1154, %f1152, %f181, %p110;
mov.b32 %r374, %f1154;
and.b32 %r375, %r374, 8388607;
or.b32 %r376, %r375, 1065353216;
mov.b32 %f1155, %r376;
shr.u32 %r377, %r374, 23;
cvt.rn.f32.u32 %f1156, %r377;
add.f32 %f1157, %f1153, %f1156;
setp.gt.f32 %p111, %f1155, 0f3FB504F3;
mul.f32 %f1158, %f1155, 0f3F000000;
add.f32 %f1159, %f1157, 0f3F800000;
selp.f32 %f1160, %f1158, %f1155, %p111;
selp.f32 %f1161, %f1159, %f1157, %p111;
add.f32 %f1162, %f1160, 0fBF800000;
add.f32 %f1148, %f1160, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1147,%f1148;
// inline asm
add.f32 %f1163, %f1162, %f1162;
mul.f32 %f1164, %f1147, %f1163;
mul.f32 %f1165, %f1164, %f1164;
fma.rn.f32 %f1168, %f1002, %f1165, %f1001;
fma.rn.f32 %f1170, %f1168, %f1165, %f1004;
mul.rn.f32 %f1171, %f1170, %f1165;
mul.rn.f32 %f1172, %f1171, %f1164;
sub.f32 %f1173, %f1162, %f1164;
neg.f32 %f1174, %f1164;
add.f32 %f1175, %f1173, %f1173;
fma.rn.f32 %f1176, %f1174, %f1162, %f1175;
mul.rn.f32 %f1177, %f1147, %f1176;
add.f32 %f1178, %f1172, %f1164;
sub.f32 %f1179, %f1164, %f1178;
add.f32 %f1180, %f1172, %f1179;
add.f32 %f1181, %f1177, %f1180;
add.f32 %f1182, %f1178, %f1181;
sub.f32 %f1183, %f1178, %f1182;
add.f32 %f1184, %f1181, %f1183;
mul.rn.f32 %f1186, %f1161, %f1020;
mul.rn.f32 %f1188, %f1161, %f1022;
add.f32 %f1189, %f1186, %f1182;
sub.f32 %f1190, %f1186, %f1189;
add.f32 %f1191, %f1182, %f1190;
add.f32 %f1192, %f1184, %f1191;
add.f32 %f1193, %f1188, %f1192;
add.f32 %f1194, %f1189, %f1193;
sub.f32 %f1195, %f1189, %f1194;
add.f32 %f1196, %f1193, %f1195;
mov.f32 %f1197, 0f3EE66666;
mul.rn.f32 %f1198, %f1197, %f1194;
neg.f32 %f1199, %f1198;
fma.rn.f32 %f1200, %f1197, %f1194, %f1199;
fma.rn.f32 %f1201, %f1197, %f1196, %f1200;
fma.rn.f32 %f1203, %f1686, %f1194, %f1201;
add.rn.f32 %f1204, %f1198, %f1203;
neg.f32 %f1205, %f1204;
add.rn.f32 %f1206, %f1198, %f1205;
add.rn.f32 %f1207, %f1206, %f1203;
mov.b32 %r378, %f1204;
setp.eq.s32 %p112, %r378, 1118925336;
add.s32 %r379, %r378, -1;
mov.b32 %f1208, %r379;
add.f32 %f1209, %f1207, 0f37000000;
selp.f32 %f1210, %f1208, %f1204, %p112;
selp.f32 %f182, %f1209, %f1207, %p112;
mul.f32 %f1211, %f1210, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1212, %f1211;
fma.rn.f32 %f1214, %f1212, %f1047, %f1210;
fma.rn.f32 %f1216, %f1212, %f1049, %f1214;
mul.f32 %f1217, %f1216, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1218, %f1217;
add.f32 %f1219, %f1212, 0f00000000;
ex2.approx.f32 %f1220, %f1219;
mul.f32 %f1221, %f1218, %f1220;
setp.lt.f32 %p113, %f1210, 0fC2D20000;
selp.f32 %f1222, 0f00000000, %f1221, %p113;
setp.gt.f32 %p114, %f1210, 0f42D20000;
selp.f32 %f1671, 0f7F800000, %f1222, %p114;
setp.eq.f32 %p115, %f1671, 0f7F800000;
@%p115 bra BB0_103;
fma.rn.f32 %f1671, %f1671, %f182, %f1671;
BB0_103:
setp.lt.f32 %p116, %f147, 0f00000000;
setp.eq.f32 %p117, %f180, 0f3F800000;
and.pred %p3, %p116, %p117;
mov.b32 %r380, %f1671;
xor.b32 %r381, %r380, -2147483648;
mov.b32 %f1223, %r381;
selp.f32 %f1673, %f1223, %f1671, %p3;
setp.eq.f32 %p118, %f147, 0f00000000;
@%p118 bra BB0_106;
bra.uni BB0_104;
BB0_106:
add.f32 %f1226, %f147, %f147;
selp.f32 %f1673, %f1226, 0f00000000, %p117;
bra.uni BB0_107;
BB0_104:
setp.geu.f32 %p119, %f147, 0f00000000;
@%p119 bra BB0_107;
cvt.rzi.f32.f32 %f1225, %f1197;
setp.neu.f32 %p120, %f1225, 0f3EE66666;
selp.f32 %f1673, 0f7FFFFFFF, %f1673, %p120;
BB0_107:
add.f32 %f1227, %f181, 0f3EE66666;
mov.b32 %r382, %f1227;
setp.lt.s32 %p122, %r382, 2139095040;
@%p122 bra BB0_112;
setp.gtu.f32 %p123, %f181, 0f7F800000;
@%p123 bra BB0_111;
bra.uni BB0_109;
BB0_111:
add.f32 %f1673, %f147, 0f3EE66666;
bra.uni BB0_112;
BB0_109:
setp.neu.f32 %p124, %f181, 0f7F800000;
@%p124 bra BB0_112;
selp.f32 %f1673, 0fFF800000, 0f7F800000, %p3;
BB0_112:
setp.eq.f32 %p125, %f147, 0f3F800000;
selp.f32 %f193, 0f3F800000, %f1673, %p125;
abs.f32 %f194, %f148;
setp.lt.f32 %p126, %f194, 0f00800000;
mul.f32 %f1230, %f194, 0f4B800000;
selp.f32 %f1231, 0fC3170000, 0fC2FE0000, %p126;
selp.f32 %f1232, %f1230, %f194, %p126;
mov.b32 %r383, %f1232;
and.b32 %r384, %r383, 8388607;
or.b32 %r385, %r384, 1065353216;
mov.b32 %f1233, %r385;
shr.u32 %r386, %r383, 23;
cvt.rn.f32.u32 %f1234, %r386;
add.f32 %f1235, %f1231, %f1234;
setp.gt.f32 %p127, %f1233, 0f3FB504F3;
mul.f32 %f1236, %f1233, 0f3F000000;
add.f32 %f1237, %f1235, 0f3F800000;
selp.f32 %f1238, %f1236, %f1233, %p127;
selp.f32 %f1239, %f1237, %f1235, %p127;
add.f32 %f1240, %f1238, 0fBF800000;
add.f32 %f1229, %f1238, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1228,%f1229;
// inline asm
add.f32 %f1241, %f1240, %f1240;
mul.f32 %f1242, %f1228, %f1241;
mul.f32 %f1243, %f1242, %f1242;
fma.rn.f32 %f1246, %f1002, %f1243, %f1001;
fma.rn.f32 %f1248, %f1246, %f1243, %f1004;
mul.rn.f32 %f1249, %f1248, %f1243;
mul.rn.f32 %f1250, %f1249, %f1242;
sub.f32 %f1251, %f1240, %f1242;
neg.f32 %f1252, %f1242;
add.f32 %f1253, %f1251, %f1251;
fma.rn.f32 %f1254, %f1252, %f1240, %f1253;
mul.rn.f32 %f1255, %f1228, %f1254;
add.f32 %f1256, %f1250, %f1242;
sub.f32 %f1257, %f1242, %f1256;
add.f32 %f1258, %f1250, %f1257;
add.f32 %f1259, %f1255, %f1258;
add.f32 %f1260, %f1256, %f1259;
sub.f32 %f1261, %f1256, %f1260;
add.f32 %f1262, %f1259, %f1261;
mul.rn.f32 %f1264, %f1239, %f1020;
mul.rn.f32 %f1266, %f1239, %f1022;
add.f32 %f1267, %f1264, %f1260;
sub.f32 %f1268, %f1264, %f1267;
add.f32 %f1269, %f1260, %f1268;
add.f32 %f1270, %f1262, %f1269;
add.f32 %f1271, %f1266, %f1270;
add.f32 %f1272, %f1267, %f1271;
sub.f32 %f1273, %f1267, %f1272;
add.f32 %f1274, %f1271, %f1273;
mul.rn.f32 %f1276, %f1197, %f1272;
neg.f32 %f1277, %f1276;
fma.rn.f32 %f1278, %f1197, %f1272, %f1277;
fma.rn.f32 %f1279, %f1197, %f1274, %f1278;
fma.rn.f32 %f1281, %f1686, %f1272, %f1279;
add.rn.f32 %f1282, %f1276, %f1281;
neg.f32 %f1283, %f1282;
add.rn.f32 %f1284, %f1276, %f1283;
add.rn.f32 %f1285, %f1284, %f1281;
mov.b32 %r387, %f1282;
setp.eq.s32 %p128, %r387, 1118925336;
add.s32 %r388, %r387, -1;
mov.b32 %f1286, %r388;
add.f32 %f1287, %f1285, 0f37000000;
selp.f32 %f1288, %f1286, %f1282, %p128;
selp.f32 %f195, %f1287, %f1285, %p128;
mul.f32 %f1289, %f1288, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1290, %f1289;
fma.rn.f32 %f1292, %f1290, %f1047, %f1288;
fma.rn.f32 %f1294, %f1290, %f1049, %f1292;
mul.f32 %f1295, %f1294, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1296, %f1295;
add.f32 %f1297, %f1290, 0f00000000;
ex2.approx.f32 %f1298, %f1297;
mul.f32 %f1299, %f1296, %f1298;
setp.lt.f32 %p129, %f1288, 0fC2D20000;
selp.f32 %f1300, 0f00000000, %f1299, %p129;
setp.gt.f32 %p130, %f1288, 0f42D20000;
selp.f32 %f1674, 0f7F800000, %f1300, %p130;
setp.eq.f32 %p131, %f1674, 0f7F800000;
@%p131 bra BB0_114;
fma.rn.f32 %f1674, %f1674, %f195, %f1674;
BB0_114:
setp.lt.f32 %p132, %f148, 0f00000000;
and.pred %p4, %p132, %p117;
mov.b32 %r389, %f1674;
xor.b32 %r390, %r389, -2147483648;
mov.b32 %f1301, %r390;
selp.f32 %f1676, %f1301, %f1674, %p4;
setp.eq.f32 %p134, %f148, 0f00000000;
@%p134 bra BB0_117;
bra.uni BB0_115;
BB0_117:
add.f32 %f1304, %f148, %f148;
selp.f32 %f1676, %f1304, 0f00000000, %p117;
bra.uni BB0_118;
BB0_115:
setp.geu.f32 %p135, %f148, 0f00000000;
@%p135 bra BB0_118;
cvt.rzi.f32.f32 %f1303, %f1197;
setp.neu.f32 %p136, %f1303, 0f3EE66666;
selp.f32 %f1676, 0f7FFFFFFF, %f1676, %p136;
BB0_118:
add.f32 %f1305, %f194, 0f3EE66666;
mov.b32 %r391, %f1305;
setp.lt.s32 %p138, %r391, 2139095040;
@%p138 bra BB0_123;
setp.gtu.f32 %p139, %f194, 0f7F800000;
@%p139 bra BB0_122;
bra.uni BB0_120;
BB0_122:
add.f32 %f1676, %f148, 0f3EE66666;
bra.uni BB0_123;
BB0_120:
setp.neu.f32 %p140, %f194, 0f7F800000;
@%p140 bra BB0_123;
selp.f32 %f1676, 0fFF800000, 0f7F800000, %p4;
BB0_123:
setp.eq.f32 %p141, %f148, 0f3F800000;
selp.f32 %f206, 0f3F800000, %f1676, %p141;
abs.f32 %f207, %f149;
setp.lt.f32 %p142, %f207, 0f00800000;
mul.f32 %f1308, %f207, 0f4B800000;
selp.f32 %f1309, 0fC3170000, 0fC2FE0000, %p142;
selp.f32 %f1310, %f1308, %f207, %p142;
mov.b32 %r392, %f1310;
and.b32 %r393, %r392, 8388607;
or.b32 %r394, %r393, 1065353216;
mov.b32 %f1311, %r394;
shr.u32 %r395, %r392, 23;
cvt.rn.f32.u32 %f1312, %r395;
add.f32 %f1313, %f1309, %f1312;
setp.gt.f32 %p143, %f1311, 0f3FB504F3;
mul.f32 %f1314, %f1311, 0f3F000000;
add.f32 %f1315, %f1313, 0f3F800000;
selp.f32 %f1316, %f1314, %f1311, %p143;
selp.f32 %f1317, %f1315, %f1313, %p143;
add.f32 %f1318, %f1316, 0fBF800000;
add.f32 %f1307, %f1316, 0f3F800000;
// inline asm
rcp.approx.ftz.f32 %f1306,%f1307;
// inline asm
add.f32 %f1319, %f1318, %f1318;
mul.f32 %f1320, %f1306, %f1319;
mul.f32 %f1321, %f1320, %f1320;
fma.rn.f32 %f1324, %f1002, %f1321, %f1001;
fma.rn.f32 %f1326, %f1324, %f1321, %f1004;
mul.rn.f32 %f1327, %f1326, %f1321;
mul.rn.f32 %f1328, %f1327, %f1320;
sub.f32 %f1329, %f1318, %f1320;
neg.f32 %f1330, %f1320;
add.f32 %f1331, %f1329, %f1329;
fma.rn.f32 %f1332, %f1330, %f1318, %f1331;
mul.rn.f32 %f1333, %f1306, %f1332;
add.f32 %f1334, %f1328, %f1320;
sub.f32 %f1335, %f1320, %f1334;
add.f32 %f1336, %f1328, %f1335;
add.f32 %f1337, %f1333, %f1336;
add.f32 %f1338, %f1334, %f1337;
sub.f32 %f1339, %f1334, %f1338;
add.f32 %f1340, %f1337, %f1339;
mul.rn.f32 %f1342, %f1317, %f1020;
mul.rn.f32 %f1344, %f1317, %f1022;
add.f32 %f1345, %f1342, %f1338;
sub.f32 %f1346, %f1342, %f1345;
add.f32 %f1347, %f1338, %f1346;
add.f32 %f1348, %f1340, %f1347;
add.f32 %f1349, %f1344, %f1348;
add.f32 %f1350, %f1345, %f1349;
sub.f32 %f1351, %f1345, %f1350;
add.f32 %f1352, %f1349, %f1351;
mul.rn.f32 %f1354, %f1197, %f1350;
neg.f32 %f1355, %f1354;
fma.rn.f32 %f1356, %f1197, %f1350, %f1355;
fma.rn.f32 %f1357, %f1197, %f1352, %f1356;
fma.rn.f32 %f1359, %f1686, %f1350, %f1357;
add.rn.f32 %f1360, %f1354, %f1359;
neg.f32 %f1361, %f1360;
add.rn.f32 %f1362, %f1354, %f1361;
add.rn.f32 %f1363, %f1362, %f1359;
mov.b32 %r396, %f1360;
setp.eq.s32 %p144, %r396, 1118925336;
add.s32 %r397, %r396, -1;
mov.b32 %f1364, %r397;
add.f32 %f1365, %f1363, 0f37000000;
selp.f32 %f1366, %f1364, %f1360, %p144;
selp.f32 %f208, %f1365, %f1363, %p144;
mul.f32 %f1367, %f1366, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f1368, %f1367;
fma.rn.f32 %f1370, %f1368, %f1047, %f1366;
fma.rn.f32 %f1372, %f1368, %f1049, %f1370;
mul.f32 %f1373, %f1372, 0f3FB8AA3B;
ex2.approx.ftz.f32 %f1374, %f1373;
add.f32 %f1375, %f1368, 0f00000000;
ex2.approx.f32 %f1376, %f1375;
mul.f32 %f1377, %f1374, %f1376;
setp.lt.f32 %p145, %f1366, 0fC2D20000;
selp.f32 %f1378, 0f00000000, %f1377, %p145;
setp.gt.f32 %p146, %f1366, 0f42D20000;
selp.f32 %f1677, 0f7F800000, %f1378, %p146;
setp.eq.f32 %p147, %f1677, 0f7F800000;
@%p147 bra BB0_125;
fma.rn.f32 %f1677, %f1677, %f208, %f1677;
BB0_125:
setp.lt.f32 %p148, %f149, 0f00000000;
and.pred %p5, %p148, %p117;
mov.b32 %r398, %f1677;
xor.b32 %r399, %r398, -2147483648;
mov.b32 %f1379, %r399;
selp.f32 %f1679, %f1379, %f1677, %p5;
setp.eq.f32 %p150, %f149, 0f00000000;
@%p150 bra BB0_128;
bra.uni BB0_126;
BB0_128:
add.f32 %f1382, %f149, %f149;
selp.f32 %f1679, %f1382, 0f00000000, %p117;
bra.uni BB0_129;
BB0_126:
setp.geu.f32 %p151, %f149, 0f00000000;
@%p151 bra BB0_129;
cvt.rzi.f32.f32 %f1381, %f1197;
setp.neu.f32 %p152, %f1381, 0f3EE66666;
selp.f32 %f1679, 0f7FFFFFFF, %f1679, %p152;
BB0_129:
add.f32 %f1383, %f207, 0f3EE66666;
mov.b32 %r400, %f1383;
setp.lt.s32 %p154, %r400, 2139095040;
@%p154 bra BB0_134;
setp.gtu.f32 %p155, %f207, 0f7F800000;
@%p155 bra BB0_133;
bra.uni BB0_131;
BB0_133:
add.f32 %f1679, %f149, 0f3EE66666;
bra.uni BB0_134;
BB0_131:
setp.neu.f32 %p156, %f207, 0f7F800000;
@%p156 bra BB0_134;
selp.f32 %f1679, 0fFF800000, 0f7F800000, %p5;
BB0_134:
setp.eq.f32 %p157, %f149, 0f3F800000;
selp.f32 %f1384, 0f3F800000, %f1679, %p157;
cvt.u64.u32 %rd108, %r3;
cvt.u64.u32 %rd107, %r2;
mov.u64 %rd111, image;
cvta.global.u64 %rd106, %rd111;
// inline asm
call (%rd105), _rt_buffer_get_64, (%rd106, %r113, %r114, %rd107, %rd108, %rd32, %rd32);
// inline asm
cvt.sat.f32.f32 %f1385, %f1384;
mul.f32 %f1386, %f1385, 0f437FFD71;
cvt.rzi.u32.f32 %r403, %f1386;
cvt.sat.f32.f32 %f1387, %f206;
mul.f32 %f1388, %f1387, 0f437FFD71;
cvt.rzi.u32.f32 %r404, %f1388;
cvt.sat.f32.f32 %f1389, %f193;
mul.f32 %f1390, %f1389, 0f437FFD71;
cvt.rzi.u32.f32 %r405, %f1390;
cvt.u16.u32 %rs32, %r403;
cvt.u16.u32 %rs33, %r405;
cvt.u16.u32 %rs34, %r404;
mov.u16 %rs35, 255;
st.v4.u8 [%rd105], {%rs32, %rs34, %rs33, %rs35};
ld.global.u32 %r518, [imageEnabled];
BB0_135:
and.b32 %r406, %r518, 4;
setp.eq.s32 %p158, %r406, 0;
@%p158 bra BB0_139;
ld.global.u32 %r407, [additive];
setp.eq.s32 %p159, %r407, 0;
cvt.u64.u32 %rd18, %r2;
cvt.u64.u32 %rd19, %r3;
mov.f32 %f1391, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs36, %f1391;}
// inline asm
@%p159 bra BB0_138;
mov.u64 %rd124, image_HDR;
cvta.global.u64 %rd113, %rd124;
mov.u32 %r411, 8;
// inline asm
call (%rd112), _rt_buffer_get_64, (%rd113, %r113, %r411, %rd18, %rd19, %rd32, %rd32);
// inline asm
ld.v4.u16 {%rs43, %rs44, %rs45, %rs46}, [%rd112];
// inline asm
{ cvt.f32.f16 %f1392, %rs43;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1393, %rs44;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1394, %rs45;}
// inline asm
// inline asm
call (%rd118), _rt_buffer_get_64, (%rd113, %r113, %r411, %rd18, %rd19, %rd32, %rd32);
// inline asm
add.f32 %f1395, %f147, %f1392;
add.f32 %f1396, %f148, %f1393;
add.f32 %f1397, %f149, %f1394;
// inline asm
{ cvt.rn.f16.f32 %rs42, %f1397;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs41, %f1396;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs40, %f1395;}
// inline asm
st.v4.u16 [%rd118], {%rs40, %rs41, %rs42, %rs36};
bra.uni BB0_139;
BB0_138:
mov.u64 %rd131, image_HDR;
cvta.global.u64 %rd126, %rd131;
mov.u32 %r413, 8;
// inline asm
call (%rd125), _rt_buffer_get_64, (%rd126, %r113, %r413, %rd18, %rd19, %rd32, %rd32);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs49, %f149;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs48, %f148;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs47, %f147;}
// inline asm
st.v4.u16 [%rd125], {%rs47, %rs48, %rs49, %rs36};
BB0_139:
ld.global.v2.u32 {%r416, %r417}, [pixelID];
cvt.u64.u32 %rd134, %r416;
cvt.u64.u32 %rd135, %r417;
mov.u64 %rd138, uvtangent;
cvta.global.u64 %rd133, %rd138;
// inline asm
call (%rd132), _rt_buffer_get_64, (%rd133, %r113, %r114, %rd134, %rd135, %rd32, %rd32);
// inline asm
ld.u32 %r102, [%rd132];
shr.u32 %r103, %r102, 16;
cvt.u16.u32 %rs50, %r103;
and.b16 %rs51, %rs50, 255;
cvt.u16.u32 %rs52, %r102;
or.b16 %rs53, %rs52, %rs51;
setp.eq.s16 %p160, %rs53, 0;
mov.f32 %f1680, %f1686;
mov.f32 %f1681, %f1686;
mov.f32 %f1682, %f1686;
@%p160 bra BB0_141;
ld.u8 %rs54, [%rd132+1];
and.b16 %rs56, %rs52, 255;
cvt.rn.f32.u16 %f1404, %rs56;
div.rn.f32 %f1405, %f1404, 0f437F0000;
fma.rn.f32 %f1406, %f1405, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f1407, %rs54;
div.rn.f32 %f1408, %f1407, 0f437F0000;
fma.rn.f32 %f1409, %f1408, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f1410, %rs51;
div.rn.f32 %f1411, %f1410, 0f437F0000;
fma.rn.f32 %f1412, %f1411, 0f40000000, 0fBF800000;
mul.f32 %f1413, %f1409, %f1409;
fma.rn.f32 %f1414, %f1406, %f1406, %f1413;
fma.rn.f32 %f1415, %f1412, %f1412, %f1414;
sqrt.rn.f32 %f1416, %f1415;
rcp.rn.f32 %f1417, %f1416;
mul.f32 %f1680, %f1406, %f1417;
mul.f32 %f1681, %f1409, %f1417;
mul.f32 %f1682, %f1412, %f1417;
BB0_141:
mul.f32 %f1421, %f1625, %f1681;
mul.f32 %f1422, %f1624, %f1682;
sub.f32 %f1423, %f1422, %f1421;
mul.f32 %f1424, %f1623, %f1682;
mul.f32 %f1425, %f1625, %f1680;
sub.f32 %f1426, %f1425, %f1424;
mul.f32 %f1427, %f1624, %f1680;
mul.f32 %f1428, %f1623, %f1681;
sub.f32 %f1429, %f1428, %f1427;
setp.lt.u32 %p161, %r102, 16777216;
selp.f32 %f1430, 0fBF800000, 0f3F800000, %p161;
mul.f32 %f1431, %f1423, %f1430;
mul.f32 %f1432, %f1426, %f1430;
mul.f32 %f1433, %f1429, %f1430;
mul.f32 %f1434, %f1431, 0f00000000;
mul.f32 %f1435, %f1432, 0f00000000;
mul.f32 %f1436, %f1433, 0f00000000;
fma.rn.f32 %f1437, %f1680, 0f3F5105EC, %f1434;
fma.rn.f32 %f1438, %f1681, 0f3F5105EC, %f1435;
fma.rn.f32 %f1439, %f1682, 0f3F5105EC, %f1436;
mul.f32 %f225, %f1623, 0f3F13CD3A;
add.f32 %f226, %f225, %f1437;
mul.f32 %f227, %f1624, 0f3F13CD3A;
add.f32 %f228, %f227, %f1438;
mul.f32 %f229, %f1625, 0f3F13CD3A;
add.f32 %f230, %f229, %f1439;
ld.global.v2.u32 {%r422, %r423}, [pixelID];
cvt.u64.u32 %rd141, %r422;
cvt.u64.u32 %rd142, %r423;
// inline asm
call (%rd139), _rt_buffer_get_64, (%rd133, %r113, %r114, %rd141, %rd142, %rd32, %rd32);
// inline asm
ld.u32 %r104, [%rd139];
shr.u32 %r105, %r104, 16;
cvt.u16.u32 %rs59, %r105;
and.b16 %rs60, %rs59, 255;
cvt.u16.u32 %rs61, %r104;
or.b16 %rs62, %rs61, %rs60;
setp.eq.s16 %p162, %rs62, 0;
mov.f32 %f1683, %f1686;
mov.f32 %f1684, %f1686;
mov.f32 %f1685, %f1686;
@%p162 bra BB0_143;
ld.u8 %rs63, [%rd139+1];
and.b16 %rs65, %rs61, 255;
cvt.rn.f32.u16 %f1440, %rs65;
div.rn.f32 %f1441, %f1440, 0f437F0000;
fma.rn.f32 %f1442, %f1441, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f1443, %rs63;
div.rn.f32 %f1444, %f1443, 0f437F0000;
fma.rn.f32 %f1445, %f1444, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f1446, %rs60;
div.rn.f32 %f1447, %f1446, 0f437F0000;
fma.rn.f32 %f1448, %f1447, 0f40000000, 0fBF800000;
mul.f32 %f1449, %f1445, %f1445;
fma.rn.f32 %f1450, %f1442, %f1442, %f1449;
fma.rn.f32 %f1451, %f1448, %f1448, %f1450;
sqrt.rn.f32 %f1452, %f1451;
rcp.rn.f32 %f1453, %f1452;
mul.f32 %f1683, %f1442, %f1453;
mul.f32 %f1684, %f1445, %f1453;
mul.f32 %f1685, %f1448, %f1453;
BB0_143:
mul.f32 %f1457, %f1625, %f1684;
mul.f32 %f1458, %f1624, %f1685;
sub.f32 %f1459, %f1458, %f1457;
mul.f32 %f1460, %f1623, %f1685;
mul.f32 %f1461, %f1625, %f1683;
sub.f32 %f1462, %f1461, %f1460;
mul.f32 %f1463, %f1624, %f1683;
mul.f32 %f1464, %f1623, %f1684;
sub.f32 %f1465, %f1464, %f1463;
setp.lt.u32 %p163, %r104, 16777216;
selp.f32 %f1466, 0fBF800000, 0f3F800000, %p163;
mul.f32 %f1467, %f1459, %f1466;
mul.f32 %f1468, %f1462, %f1466;
mul.f32 %f1469, %f1465, %f1466;
mul.f32 %f1470, %f1467, 0f3F3504F3;
mul.f32 %f1471, %f1468, 0f3F3504F3;
mul.f32 %f1472, %f1469, 0f3F3504F3;
fma.rn.f32 %f1473, %f1683, 0fBED105EC, %f1470;
fma.rn.f32 %f1474, %f1684, 0fBED105EC, %f1471;
fma.rn.f32 %f1475, %f1685, 0fBED105EC, %f1472;
add.f32 %f237, %f225, %f1473;
add.f32 %f238, %f227, %f1474;
add.f32 %f239, %f229, %f1475;
ld.global.v2.u32 {%r428, %r429}, [pixelID];
cvt.u64.u32 %rd148, %r428;
cvt.u64.u32 %rd149, %r429;
// inline asm
call (%rd146), _rt_buffer_get_64, (%rd133, %r113, %r114, %rd148, %rd149, %rd32, %rd32);
// inline asm
ld.u32 %r106, [%rd146];
shr.u32 %r107, %r106, 16;
cvt.u16.u32 %rs68, %r107;
and.b16 %rs69, %rs68, 255;
cvt.u16.u32 %rs70, %r106;
or.b16 %rs71, %rs70, %rs69;
setp.eq.s16 %p164, %rs71, 0;
mov.f32 %f1687, %f1686;
mov.f32 %f1688, %f1686;
@%p164 bra BB0_145;
ld.u8 %rs72, [%rd146+1];
and.b16 %rs74, %rs70, 255;
cvt.rn.f32.u16 %f1476, %rs74;
div.rn.f32 %f1477, %f1476, 0f437F0000;
fma.rn.f32 %f1478, %f1477, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f1479, %rs72;
div.rn.f32 %f1480, %f1479, 0f437F0000;
fma.rn.f32 %f1481, %f1480, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f1482, %rs69;
div.rn.f32 %f1483, %f1482, 0f437F0000;
fma.rn.f32 %f1484, %f1483, 0f40000000, 0fBF800000;
mul.f32 %f1485, %f1481, %f1481;
fma.rn.f32 %f1486, %f1478, %f1478, %f1485;
fma.rn.f32 %f1487, %f1484, %f1484, %f1486;
sqrt.rn.f32 %f1488, %f1487;
rcp.rn.f32 %f1489, %f1488;
mul.f32 %f1686, %f1478, %f1489;
mul.f32 %f1687, %f1481, %f1489;
mul.f32 %f1688, %f1484, %f1489;
BB0_145:
mul.f32 %f1491, %f1625, %f1687;
mul.f32 %f1492, %f1624, %f1688;
sub.f32 %f1493, %f1492, %f1491;
mul.f32 %f1494, %f1623, %f1688;
mul.f32 %f1495, %f1625, %f1686;
sub.f32 %f1496, %f1495, %f1494;
mul.f32 %f1497, %f1624, %f1686;
mul.f32 %f1498, %f1623, %f1687;
sub.f32 %f1499, %f1498, %f1497;
setp.lt.u32 %p165, %r106, 16777216;
selp.f32 %f1500, 0fBF800000, 0f3F800000, %p165;
mul.f32 %f1501, %f1493, %f1500;
mul.f32 %f1502, %f1496, %f1500;
mul.f32 %f1503, %f1499, %f1500;
mul.f32 %f1504, %f1501, 0fBF3504F3;
mul.f32 %f1505, %f1502, 0fBF3504F3;
mul.f32 %f1506, %f1503, 0fBF3504F3;
fma.rn.f32 %f1507, %f1686, 0fBED105EC, %f1504;
fma.rn.f32 %f1508, %f1687, 0fBED105EC, %f1505;
fma.rn.f32 %f1509, %f1688, 0fBED105EC, %f1506;
add.f32 %f1510, %f225, %f1507;
add.f32 %f1511, %f227, %f1508;
add.f32 %f1512, %f229, %f1509;
ld.global.f32 %f1513, [directDir];
mul.f32 %f1514, %f226, %f1513;
ld.global.f32 %f1515, [directDir+4];
mul.f32 %f1516, %f228, %f1515;
neg.f32 %f1517, %f1516;
sub.f32 %f1518, %f1517, %f1514;
ld.global.f32 %f1519, [directDir+8];
mul.f32 %f1520, %f230, %f1519;
sub.f32 %f1521, %f1518, %f1520;
cvt.sat.f32.f32 %f1522, %f1521;
mul.f32 %f1523, %f164, %f1522;
mul.f32 %f1524, %f165, %f1522;
mul.f32 %f1525, %f166, %f1522;
mul.f32 %f1526, %f237, %f1513;
mul.f32 %f1527, %f238, %f1515;
neg.f32 %f1528, %f1527;
sub.f32 %f1529, %f1528, %f1526;
mul.f32 %f1530, %f239, %f1519;
sub.f32 %f1531, %f1529, %f1530;
cvt.sat.f32.f32 %f1532, %f1531;
mul.f32 %f1533, %f164, %f1532;
mul.f32 %f1534, %f165, %f1532;
mul.f32 %f1535, %f166, %f1532;
mul.f32 %f1536, %f1510, %f1513;
mul.f32 %f1537, %f1511, %f1515;
neg.f32 %f1538, %f1537;
sub.f32 %f1539, %f1538, %f1536;
mul.f32 %f1540, %f1512, %f1519;
sub.f32 %f1541, %f1539, %f1540;
cvt.sat.f32.f32 %f1542, %f1541;
mul.f32 %f1543, %f164, %f1542;
mul.f32 %f1544, %f165, %f1542;
mul.f32 %f1545, %f166, %f1542;
add.f32 %f1546, %f1523, %f1533;
add.f32 %f1547, %f1524, %f1534;
add.f32 %f1548, %f1525, %f1535;
add.f32 %f1549, %f1546, %f1543;
add.f32 %f1550, %f1547, %f1544;
add.f32 %f1551, %f1548, %f1545;
mul.f32 %f1552, %f1549, 0f3F13CD3A;
mul.f32 %f1553, %f1550, 0f3F13CD3A;
mul.f32 %f1554, %f1551, 0f3F13CD3A;
div.rn.f32 %f1555, %f147, %f1552;
div.rn.f32 %f1556, %f148, %f1553;
div.rn.f32 %f1557, %f149, %f1554;
setp.eq.f32 %p166, %f147, 0f00000000;
selp.f32 %f1558, 0f00000000, %f1555, %p166;
setp.eq.f32 %p167, %f148, 0f00000000;
selp.f32 %f1559, 0f00000000, %f1556, %p167;
setp.eq.f32 %p168, %f149, 0f00000000;
selp.f32 %f1560, 0f00000000, %f1557, %p168;
mul.f32 %f246, %f1523, %f1558;
mul.f32 %f247, %f1524, %f1559;
mul.f32 %f248, %f1525, %f1560;
mul.f32 %f249, %f1533, %f1558;
mul.f32 %f250, %f1534, %f1559;
mul.f32 %f251, %f1535, %f1560;
mul.f32 %f252, %f1543, %f1558;
mul.f32 %f253, %f1544, %f1559;
mul.f32 %f254, %f1545, %f1560;
ld.global.u32 %r432, [additive];
setp.eq.s32 %p169, %r432, 0;
cvt.u64.u32 %rd23, %r2;
cvt.u64.u32 %rd24, %r3;
mov.f32 %f1490, 0f3F800000;
// inline asm
{ cvt.rn.f16.f32 %rs77, %f1490;}
// inline asm
@%p169 bra BB0_147;
mov.u64 %rd165, image_RNM0;
cvta.global.u64 %rd154, %rd165;
mov.u32 %r436, 8;
// inline asm
call (%rd153), _rt_buffer_get_64, (%rd154, %r113, %r436, %rd23, %rd24, %rd32, %rd32);
// inline asm
ld.v4.u16 {%rs84, %rs85, %rs86, %rs87}, [%rd153];
// inline asm
{ cvt.f32.f16 %f1561, %rs84;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1562, %rs85;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1563, %rs86;}
// inline asm
// inline asm
call (%rd159), _rt_buffer_get_64, (%rd154, %r113, %r436, %rd23, %rd24, %rd32, %rd32);
// inline asm
add.f32 %f1564, %f246, %f1561;
add.f32 %f1565, %f247, %f1562;
add.f32 %f1566, %f248, %f1563;
// inline asm
{ cvt.rn.f16.f32 %rs83, %f1566;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs82, %f1565;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs81, %f1564;}
// inline asm
st.v4.u16 [%rd159], {%rs81, %rs82, %rs83, %rs77};
bra.uni BB0_148;
BB0_147:
mov.u64 %rd172, image_RNM0;
cvta.global.u64 %rd167, %rd172;
mov.u32 %r438, 8;
// inline asm
call (%rd166), _rt_buffer_get_64, (%rd167, %r113, %r438, %rd23, %rd24, %rd32, %rd32);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs90, %f248;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs89, %f247;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs88, %f246;}
// inline asm
st.v4.u16 [%rd166], {%rs88, %rs89, %rs90, %rs77};
BB0_148:
ld.global.u32 %r439, [additive];
setp.eq.s32 %p170, %r439, 0;
// inline asm
{ cvt.rn.f16.f32 %rs91, %f1490;}
// inline asm
@%p170 bra BB0_150;
mov.u64 %rd185, image_RNM1;
cvta.global.u64 %rd174, %rd185;
mov.u32 %r443, 8;
// inline asm
call (%rd173), _rt_buffer_get_64, (%rd174, %r113, %r443, %rd23, %rd24, %rd32, %rd32);
// inline asm
ld.v4.u16 {%rs98, %rs99, %rs100, %rs101}, [%rd173];
// inline asm
{ cvt.f32.f16 %f1571, %rs98;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1572, %rs99;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1573, %rs100;}
// inline asm
// inline asm
call (%rd179), _rt_buffer_get_64, (%rd174, %r113, %r443, %rd23, %rd24, %rd32, %rd32);
// inline asm
add.f32 %f1574, %f249, %f1571;
add.f32 %f1575, %f250, %f1572;
add.f32 %f1576, %f251, %f1573;
// inline asm
{ cvt.rn.f16.f32 %rs97, %f1576;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs96, %f1575;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs95, %f1574;}
// inline asm
st.v4.u16 [%rd179], {%rs95, %rs96, %rs97, %rs91};
bra.uni BB0_151;
BB0_150:
mov.u64 %rd192, image_RNM1;
cvta.global.u64 %rd187, %rd192;
mov.u32 %r445, 8;
// inline asm
call (%rd186), _rt_buffer_get_64, (%rd187, %r113, %r445, %rd23, %rd24, %rd32, %rd32);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs104, %f251;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs103, %f250;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs102, %f249;}
// inline asm
st.v4.u16 [%rd186], {%rs102, %rs103, %rs104, %rs91};
BB0_151:
ld.global.u32 %r446, [additive];
setp.eq.s32 %p171, %r446, 0;
// inline asm
{ cvt.rn.f16.f32 %rs105, %f1490;}
// inline asm
@%p171 bra BB0_153;
mov.u64 %rd205, image_RNM2;
cvta.global.u64 %rd194, %rd205;
mov.u32 %r450, 8;
// inline asm
call (%rd193), _rt_buffer_get_64, (%rd194, %r113, %r450, %rd23, %rd24, %rd32, %rd32);
// inline asm
ld.v4.u16 {%rs112, %rs113, %rs114, %rs115}, [%rd193];
// inline asm
{ cvt.f32.f16 %f1581, %rs112;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1582, %rs113;}
// inline asm
// inline asm
{ cvt.f32.f16 %f1583, %rs114;}
// inline asm
// inline asm
call (%rd199), _rt_buffer_get_64, (%rd194, %r113, %r450, %rd23, %rd24, %rd32, %rd32);
// inline asm
add.f32 %f1584, %f252, %f1581;
add.f32 %f1585, %f253, %f1582;
add.f32 %f1586, %f254, %f1583;
// inline asm
{ cvt.rn.f16.f32 %rs111, %f1586;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs110, %f1585;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs109, %f1584;}
// inline asm
st.v4.u16 [%rd199], {%rs109, %rs110, %rs111, %rs105};
bra.uni BB0_171;
BB0_153:
mov.u64 %rd212, image_RNM2;
cvta.global.u64 %rd207, %rd212;
mov.u32 %r452, 8;
// inline asm
call (%rd206), _rt_buffer_get_64, (%rd207, %r113, %r452, %rd23, %rd24, %rd32, %rd32);
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs118, %f254;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs117, %f253;}
// inline asm
// inline asm
{ cvt.rn.f16.f32 %rs116, %f252;}
// inline asm
st.v4.u16 [%rd206], {%rs116, %rs117, %rs118, %rs105};
BB0_171:
ret;
}
Reference in New Issue View Git Blame Copy Permalink