Files
FORT-BO/Assets/Editor/x64/Bakery/optix9/lmBatchPointLightLegacy.ptx
2026-08-06 18:56:32 +03:00

2092 lines
61 KiB
Plaintext

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-31968024
// Cuda compilation tools, release 12.0, V12.0.76
// Based on NVVM 7.0.1
//
.version 8.0
.target sm_52
.address_size 64
// .globl __raygen__oxMain
.const .align 16 .b8 params[1184];
.visible .entry __raygen__oxMain()
{
.reg .pred %p<179>;
.reg .b16 %rs<65>;
.reg .f32 %f<1369>;
.reg .b32 %r<418>;
.reg .b64 %rd<54>;
// begin inline asm
call (%r23), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r24), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u64 %rd13, [params+400];
cvta.to.global.u64 %rd14, %rd13;
ld.const.u32 %r29, [params+392];
mad.lo.s32 %r30, %r29, %r24, %r23;
mul.wide.u32 %rd15, %r30, 4;
add.s64 %rd1, %rd14, %rd15;
ld.global.v2.u8 {%rs9, %rs63}, [%rd1];
or.b16 %rs11, %rs9, %rs63;
and.b16 %rs12, %rs11, 255;
setp.eq.s16 %p10, %rs12, 0;
@%p10 bra $L__BB0_2;
ld.global.u8 %rs62, [%rd1+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs62, [%rd1+2];
setp.eq.s16 %p11, %rs62, 0;
mov.f32 %f1298, 0f00000000;
mov.u16 %rs63, 0;
mov.f32 %f1299, %f1298;
mov.f32 %f1300, %f1298;
@%p11 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f221, %rs9;
div.rn.f32 %f222, %f221, 0f437F0000;
fma.rn.f32 %f223, %f222, 0f40000000, 0fBF800000;
and.b16 %rs15, %rs63, 255;
cvt.rn.f32.u16 %f224, %rs15;
div.rn.f32 %f225, %f224, 0f437F0000;
fma.rn.f32 %f226, %f225, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f227, %rs62;
div.rn.f32 %f228, %f227, 0f437F0000;
fma.rn.f32 %f229, %f228, 0f40000000, 0fBF800000;
mul.f32 %f230, %f226, %f226;
fma.rn.f32 %f231, %f223, %f223, %f230;
fma.rn.f32 %f232, %f229, %f229, %f231;
sqrt.rn.f32 %f233, %f232;
rcp.rn.f32 %f234, %f233;
mul.f32 %f1300, %f234, %f229;
mul.f32 %f1299, %f234, %f226;
mul.f32 %f1298, %f223, %f234;
$L__BB0_4:
ld.const.v2.u32 {%r31, %r32}, [params];
add.s32 %r3, %r31, %r23;
add.s32 %r4, %r32, %r24;
setp.eq.f32 %p12, %f1298, 0f00000000;
setp.eq.f32 %p13, %f1299, 0f00000000;
and.pred %p14, %p12, %p13;
setp.eq.f32 %p15, %f1300, 0f00000000;
and.pred %p16, %p15, %p14;
@%p16 bra $L__BB0_141;
bra.uni $L__BB0_5;
$L__BB0_141:
ld.const.u32 %r22, [params+104];
and.b32 %r399, %r22, 1;
setp.eq.b32 %p170, %r399, 1;
mov.pred %p171, 0;
xor.pred %p172, %p170, %p171;
not.pred %p173, %p172;
@%p173 bra $L__BB0_143;
ld.const.u64 %rd38, [params+144];
cvta.to.global.u64 %rd39, %rd38;
ld.const.u32 %r400, [params+136];
mad.lo.s32 %r401, %r400, %r4, %r3;
mul.wide.u32 %rd40, %r401, 4;
add.s64 %rd41, %rd39, %rd40;
mov.u16 %rs43, 0;
st.global.v4.u8 [%rd41], {%rs43, %rs43, %rs43, %rs43};
$L__BB0_143:
and.b32 %r402, %r22, 8;
setp.eq.s32 %p174, %r402, 0;
@%p174 bra $L__BB0_145;
ld.const.u64 %rd42, [params+192];
cvta.to.global.u64 %rd43, %rd42;
ld.const.u32 %r403, [params+184];
mad.lo.s32 %r404, %r403, %r4, %r3;
mov.f32 %f1285, 0f00000000;
cvt.rzi.u32.f32 %r405, %f1285;
mul.wide.u32 %rd44, %r404, 2;
add.s64 %rd45, %rd43, %rd44;
mov.u16 %rs44, 0;
cvt.u16.u32 %rs45, %r405;
st.global.v2.u8 [%rd45], {%rs45, %rs44};
$L__BB0_145:
and.b32 %r406, %r22, 4;
setp.eq.s32 %p175, %r406, 0;
@%p175 bra $L__BB0_149;
ld.const.u32 %r407, [params+108];
setp.eq.s32 %p176, %r407, 0;
ld.const.u64 %rd46, [params+224];
cvta.to.global.u64 %rd47, %rd46;
ld.const.u32 %r408, [params+216];
mad.lo.s32 %r409, %r408, %r4, %r3;
mul.wide.u32 %rd48, %r409, 8;
add.s64 %rd12, %rd47, %rd48;
@%p176 bra $L__BB0_148;
ld.global.v4.u16 {%rs52, %rs53, %rs54, %rs55}, [%rd12];
// begin inline asm
{ cvt.f32.f16 %f1286, %rs52;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1287, %rs53;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1288, %rs54;}
// end inline asm
add.f32 %f1289, %f1286, 0f00000000;
add.f32 %f1290, %f1287, 0f00000000;
add.f32 %f1291, %f1288, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f1291;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f1290;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs49, %f1289;}
// end inline asm
mov.u16 %rs56, 0;
st.global.v4.u16 [%rd12], {%rs49, %rs50, %rs51, %rs56};
bra.uni $L__BB0_149;
$L__BB0_5:
ld.const.u64 %rd16, [params+432];
cvta.to.global.u64 %rd17, %rd16;
ld.const.u32 %r35, [params+424];
mad.lo.s32 %r36, %r35, %r24, %r23;
mul.wide.u32 %rd18, %r36, 12;
add.s64 %rd19, %rd17, %rd18;
ld.global.f32 %f10, [%rd19];
mul.f32 %f235, %f10, 0f3456BF95;
ld.global.f32 %f11, [%rd19+4];
mul.f32 %f236, %f11, 0f3456BF95;
ld.global.f32 %f12, [%rd19+8];
mul.f32 %f237, %f12, 0f3456BF95;
abs.f32 %f238, %f1298;
div.rn.f32 %f239, %f235, %f238;
abs.f32 %f240, %f1299;
div.rn.f32 %f241, %f236, %f240;
abs.f32 %f242, %f1300;
div.rn.f32 %f243, %f237, %f242;
abs.f32 %f244, %f239;
abs.f32 %f245, %f241;
abs.f32 %f246, %f243;
mov.f32 %f247, 0f38D1B717;
max.f32 %f248, %f244, %f247;
max.f32 %f249, %f245, %f247;
max.f32 %f250, %f246, %f247;
fma.rn.f32 %f13, %f1298, %f248, %f10;
fma.rn.f32 %f14, %f1299, %f249, %f11;
fma.rn.f32 %f15, %f1300, %f250, %f12;
ld.const.u32 %r5, [params+600];
setp.eq.s32 %p17, %r5, 0;
@%p17 bra $L__BB0_46;
ld.const.u64 %rd20, [params+608];
cvta.to.global.u64 %rd2, %rd20;
mov.f32 %f258, 0f40000000;
cvt.rzi.f32.f32 %f259, %f258;
add.f32 %f260, %f259, %f259;
mov.f32 %f261, 0f40800000;
sub.f32 %f262, %f261, %f260;
abs.f32 %f16, %f262;
ld.const.u32 %r6, [params+544];
ld.const.u32 %r417, [params+104];
and.b32 %r8, %r417, 64;
and.b32 %r38, %r417, 32;
setp.eq.s32 %p18, %r38, 0;
mov.u32 %r413, 0;
selp.f32 %f17, 0f3F800000, 0f41200000, %p18;
mul.f32 %f18, %f13, 0f3456BF95;
mul.f32 %f19, %f14, 0f3456BF95;
mul.f32 %f20, %f15, 0f3456BF95;
mov.f32 %f257, 0f00000000;
ld.const.u64 %rd3, [params+96];
abs.f32 %f447, %f18;
abs.f32 %f448, %f19;
max.f32 %f449, %f447, %f448;
abs.f32 %f450, %f20;
max.f32 %f451, %f449, %f450;
mov.f32 %f1340, %f257;
mov.f32 %f1339, %f257;
mov.f32 %f1338, %f257;
mov.f32 %f1337, %f257;
mov.f32 %f1336, %f257;
mov.f32 %f1335, %f257;
mov.f32 %f1334, %f257;
$L__BB0_7:
mul.wide.u32 %rd21, %r413, 112;
add.s64 %rd5, %rd2, %rd21;
ld.global.v4.f32 {%f264, %f265, %f266, %f267}, [%rd5];
ld.global.v4.f32 {%f271, %f272, %f273, %f274}, [%rd5+16];
ld.global.v4.f32 {%f277, %f1316, %f1317, %f280}, [%rd5+32];
ld.global.v4.f32 {%f281, %f282, %f283, %f1314}, [%rd5+48];
ld.global.v4.f32 {%f285, %f286, %f287, %f288}, [%rd5+64];
ld.global.v4.u32 {%r39, %r40, %r41, %r42}, [%rd5+80];
mov.b32 %f41, %r39;
mov.b32 %f42, %r40;
mov.b32 %f43, %r41;
ld.global.u64 %rd6, [%rd5+96];
sub.f32 %f44, %f265, %f10;
sub.f32 %f45, %f266, %f11;
sub.f32 %f46, %f267, %f12;
mul.f32 %f289, %f45, %f45;
fma.rn.f32 %f290, %f44, %f44, %f289;
fma.rn.f32 %f291, %f46, %f46, %f290;
sqrt.rn.f32 %f292, %f291;
rcp.rn.f32 %f293, %f292;
mul.f32 %f47, %f44, %f293;
mul.f32 %f48, %f45, %f293;
mul.f32 %f49, %f46, %f293;
mul.f32 %f50, %f292, %f272;
mul.f32 %f51, %f292, %f274;
abs.f32 %f52, %f51;
setp.lt.f32 %p19, %f52, 0f00800000;
mul.f32 %f294, %f52, 0f4B800000;
selp.f32 %f295, %f294, %f52, %p19;
selp.f32 %f296, 0fC3170000, 0fC2FE0000, %p19;
mov.b32 %r46, %f295;
and.b32 %r47, %r46, 8388607;
or.b32 %r48, %r47, 1065353216;
mov.b32 %f297, %r48;
shr.u32 %r49, %r46, 23;
cvt.rn.f32.u32 %f298, %r49;
add.f32 %f299, %f296, %f298;
setp.gt.f32 %p20, %f297, 0f3FB504F3;
mul.f32 %f300, %f297, 0f3F000000;
add.f32 %f301, %f299, 0f3F800000;
selp.f32 %f302, %f301, %f299, %p20;
selp.f32 %f303, %f300, %f297, %p20;
add.f32 %f304, %f303, 0fBF800000;
add.f32 %f305, %f303, 0f3F800000;
rcp.approx.ftz.f32 %f306, %f305;
add.f32 %f307, %f304, %f304;
mul.f32 %f308, %f307, %f306;
mul.f32 %f309, %f308, %f308;
mov.f32 %f310, 0f3C4CAF63;
mov.f32 %f311, 0f3B18F0FE;
fma.rn.f32 %f312, %f311, %f309, %f310;
mov.f32 %f313, 0f3DAAAABD;
fma.rn.f32 %f314, %f312, %f309, %f313;
mul.rn.f32 %f315, %f314, %f309;
mul.rn.f32 %f316, %f315, %f308;
sub.f32 %f317, %f304, %f308;
add.f32 %f318, %f317, %f317;
neg.f32 %f319, %f308;
fma.rn.f32 %f320, %f319, %f304, %f318;
mul.rn.f32 %f321, %f306, %f320;
add.f32 %f322, %f316, %f308;
sub.f32 %f323, %f308, %f322;
add.f32 %f324, %f316, %f323;
add.f32 %f325, %f321, %f324;
add.f32 %f326, %f322, %f325;
sub.f32 %f327, %f322, %f326;
add.f32 %f328, %f325, %f327;
mov.f32 %f329, 0f3F317200;
mul.rn.f32 %f330, %f302, %f329;
mov.f32 %f331, 0f35BFBE8E;
mul.rn.f32 %f332, %f302, %f331;
add.f32 %f333, %f330, %f326;
sub.f32 %f334, %f330, %f333;
add.f32 %f335, %f326, %f334;
add.f32 %f336, %f328, %f335;
add.f32 %f337, %f332, %f336;
add.f32 %f338, %f333, %f337;
sub.f32 %f339, %f333, %f338;
add.f32 %f340, %f337, %f339;
mul.rn.f32 %f342, %f261, %f338;
neg.f32 %f343, %f342;
fma.rn.f32 %f344, %f261, %f338, %f343;
fma.rn.f32 %f345, %f261, %f340, %f344;
fma.rn.f32 %f347, %f257, %f338, %f345;
add.rn.f32 %f348, %f342, %f347;
neg.f32 %f349, %f348;
add.rn.f32 %f350, %f342, %f349;
add.rn.f32 %f351, %f350, %f347;
mov.b32 %r50, %f348;
setp.eq.s32 %p21, %r50, 1118925336;
add.s32 %r51, %r50, -1;
mov.b32 %f352, %r51;
add.f32 %f353, %f351, 0f37000000;
selp.f32 %f53, %f353, %f351, %p21;
selp.f32 %f354, %f352, %f348, %p21;
mov.f32 %f355, 0f3FB8AA3B;
mul.rn.f32 %f356, %f354, %f355;
cvt.rzi.f32.f32 %f357, %f356;
abs.f32 %f358, %f357;
setp.gt.f32 %p22, %f358, 0f42FC0000;
mov.b32 %r52, %f357;
and.b32 %r53, %r52, -2147483648;
or.b32 %r54, %r53, 1123811328;
mov.b32 %f359, %r54;
selp.f32 %f360, %f359, %f357, %p22;
mov.f32 %f361, 0fBF317218;
fma.rn.f32 %f362, %f360, %f361, %f354;
mov.f32 %f363, 0f3102E308;
fma.rn.f32 %f364, %f360, %f363, %f362;
mul.f32 %f365, %f364, 0f3FB8AA3B;
add.f32 %f366, %f360, 0f4B40007F;
mov.b32 %r55, %f366;
shl.b32 %r56, %r55, 23;
mov.b32 %f367, %r56;
ex2.approx.ftz.f32 %f368, %f365;
mul.f32 %f54, %f368, %f367;
setp.eq.f32 %p23, %f54, 0f7F800000;
mov.f32 %f1308, 0f7F800000;
@%p23 bra $L__BB0_9;
fma.rn.f32 %f1308, %f54, %f53, %f54;
$L__BB0_9:
setp.lt.f32 %p24, %f51, 0f00000000;
setp.eq.f32 %p25, %f16, 0f3F800000;
and.pred %p1, %p24, %p25;
setp.eq.f32 %p26, %f51, 0f00000000;
@%p26 bra $L__BB0_13;
bra.uni $L__BB0_10;
$L__BB0_13:
add.f32 %f373, %f51, %f51;
selp.f32 %f1310, %f373, 0f00000000, %p25;
bra.uni $L__BB0_14;
$L__BB0_10:
mov.b32 %r57, %f1308;
xor.b32 %r58, %r57, -2147483648;
mov.b32 %f369, %r58;
selp.f32 %f1310, %f369, %f1308, %p1;
setp.geu.f32 %p27, %f51, 0f00000000;
@%p27 bra $L__BB0_14;
mov.f32 %f370, 0f40800000;
cvt.rzi.f32.f32 %f371, %f370;
setp.eq.f32 %p28, %f371, 0f40800000;
@%p28 bra $L__BB0_14;
mov.f32 %f1310, 0f7FFFFFFF;
$L__BB0_14:
abs.f32 %f1295, %f51;
add.f32 %f374, %f1295, 0f40800000;
mov.b32 %r59, %f374;
setp.lt.s32 %p30, %r59, 2139095040;
@%p30 bra $L__BB0_19;
abs.f32 %f1296, %f51;
setp.gtu.f32 %p31, %f1296, 0f7F800000;
@%p31 bra $L__BB0_18;
bra.uni $L__BB0_16;
$L__BB0_18:
add.f32 %f1310, %f51, 0f40800000;
bra.uni $L__BB0_19;
$L__BB0_16:
abs.f32 %f1297, %f51;
setp.neu.f32 %p32, %f1297, 0f7F800000;
@%p32 bra $L__BB0_19;
selp.f32 %f1310, 0fFF800000, 0f7F800000, %p1;
$L__BB0_19:
mov.f32 %f375, 0f3F800000;
sub.f32 %f376, %f375, %f1310;
setp.eq.f32 %p33, %f51, 0f3F800000;
selp.f32 %f377, 0f00000000, %f376, %p33;
cvt.sat.f32.f32 %f378, %f377;
fma.rn.f32 %f379, %f50, %f50, %f273;
div.rn.f32 %f63, %f378, %f379;
mul.f32 %f380, %f1299, %f48;
fma.rn.f32 %f381, %f1298, %f47, %f380;
fma.rn.f32 %f64, %f1300, %f49, %f381;
setp.eq.f32 %p34, %f264, 0f3F800000;
@%p34 bra $L__BB0_25;
bra.uni $L__BB0_20;
$L__BB0_25:
setp.eq.s64 %p38, %rd6, 0;
mov.f32 %f1315, %f277;
mov.f32 %f1318, %f63;
@%p38 bra $L__BB0_29;
setp.geu.f32 %p39, %f1314, 0f00000000;
mov.f32 %f1311, %f47;
mov.f32 %f1312, %f48;
mov.f32 %f1313, %f49;
@%p39 bra $L__BB0_28;
mul.f32 %f404, %f42, %f45;
fma.rn.f32 %f405, %f41, %f44, %f404;
fma.rn.f32 %f406, %f43, %f46, %f405;
rcp.rn.f32 %f407, %f406;
mul.f32 %f1311, %f44, %f407;
mul.f32 %f1312, %f45, %f407;
mul.f32 %f1313, %f46, %f407;
neg.f32 %f1314, %f1314;
$L__BB0_28:
mul.f32 %f408, %f43, %f49;
mul.f32 %f409, %f42, %f48;
neg.f32 %f410, %f409;
mul.f32 %f411, %f41, %f47;
sub.f32 %f412, %f410, %f411;
sub.f32 %f413, %f412, %f408;
setp.gt.f32 %p40, %f413, 0f00000000;
selp.f32 %f414, 0f3F800000, 0f00000000, %p40;
mov.f32 %f415, 0f3F800000;
mul.f32 %f416, %f282, %f1312;
fma.rn.f32 %f417, %f281, %f1311, %f416;
mul.f32 %f418, %f286, %f1312;
fma.rn.f32 %f419, %f285, %f1311, %f418;
fma.rn.f32 %f420, %f283, %f1313, %f417;
fma.rn.f32 %f421, %f287, %f1313, %f419;
fma.rn.f32 %f422, %f1314, %f420, 0f3F000000;
sub.f32 %f423, %f415, %f422;
fma.rn.f32 %f424, %f1314, %f421, 0f3F000000;
tex.2d.v4.f32.f32 {%f425, %f426, %f427, %f428}, [%rd6, {%f423, %f424}];
mul.f32 %f429, %f414, %f425;
mul.f32 %f430, %f414, %f426;
mul.f32 %f431, %f414, %f427;
mul.f32 %f1315, %f277, %f429;
mul.f32 %f1316, %f1316, %f430;
mul.f32 %f1317, %f1317, %f431;
mov.f32 %f1318, %f63;
bra.uni $L__BB0_29;
$L__BB0_20:
setp.eq.f32 %p35, %f264, 0f40000000;
@%p35 bra $L__BB0_23;
bra.uni $L__BB0_21;
$L__BB0_23:
setp.eq.s64 %p37, %rd6, 0;
mov.f32 %f1315, %f277;
mov.f32 %f1318, %f63;
@%p37 bra $L__BB0_29;
mul.f32 %f391, %f282, %f48;
fma.rn.f32 %f392, %f281, %f47, %f391;
mul.f32 %f393, %f286, %f48;
fma.rn.f32 %f394, %f285, %f47, %f393;
mul.f32 %f395, %f42, %f48;
fma.rn.f32 %f396, %f41, %f47, %f395;
fma.rn.f32 %f397, %f283, %f49, %f392;
fma.rn.f32 %f398, %f287, %f49, %f394;
fma.rn.f32 %f399, %f43, %f49, %f396;
tex.cube.v4.f32.f32 {%f400, %f401, %f402, %f403}, [%rd6, {%f397, %f398, %f399, %f399}];
mul.f32 %f1315, %f277, %f400;
mul.f32 %f1316, %f1316, %f401;
mul.f32 %f1317, %f1317, %f402;
mov.f32 %f1318, %f63;
bra.uni $L__BB0_29;
$L__BB0_21:
setp.neu.f32 %p36, %f264, 0f40800000;
mov.f32 %f1315, %f277;
mov.f32 %f1318, %f63;
@%p36 bra $L__BB0_29;
mul.f32 %f382, %f41, %f47;
mul.f32 %f383, %f42, %f48;
neg.f32 %f384, %f383;
sub.f32 %f385, %f384, %f382;
mul.f32 %f386, %f43, %f49;
sub.f32 %f387, %f385, %f386;
fma.rn.f32 %f388, %f280, %f387, %f1314;
cvt.sat.f32.f32 %f389, %f388;
mul.f32 %f390, %f389, %f389;
mul.f32 %f1318, %f63, %f390;
mov.f32 %f1315, %f277;
$L__BB0_29:
max.f32 %f438, %f1315, %f1316;
max.f32 %f439, %f438, %f1317;
setp.eq.s32 %p42, %r6, 0;
selp.f32 %f440, %f64, 0f3F800000, %p42;
cvt.sat.f32.f32 %f441, %f440;
mul.f32 %f84, %f441, %f1318;
mul.f32 %f442, %f84, %f439;
setp.lt.f32 %p43, %f442, 0f3727C5AC;
mov.pred %p178, -1;
mov.f32 %f1319, 0f00000000;
mov.f32 %f1320, %f1319;
mov.f32 %f1321, %f1319;
mov.f32 %f1322, %f1319;
mov.f32 %f1323, %f1319;
mov.f32 %f1324, %f1319;
@%p43 bra $L__BB0_31;
setp.eq.s32 %p45, %r8, 0;
mul.f32 %f1319, %f1315, %f84;
mul.f32 %f1320, %f1316, %f84;
mul.f32 %f1321, %f1317, %f84;
selp.f32 %f1322, 0f00000000, %f47, %p45;
selp.f32 %f1323, 0f00000000, %f48, %p45;
selp.f32 %f1324, 0f00000000, %f49, %p45;
mov.pred %p178, 0;
$L__BB0_31:
@%p178 bra $L__BB0_45;
setp.eq.s32 %p46, %r42, 0;
mov.u16 %rs64, 0;
mov.f32 %f1333, 0f3F800000;
@%p46 bra $L__BB0_40;
abs.s32 %r11, %r42;
setp.lt.s32 %p47, %r11, 1;
mov.f32 %f1326, 0f00000000;
@%p47 bra $L__BB0_39;
mov.f32 %f452, 0f38D1B717;
max.f32 %f453, %f451, %f452;
mul.f32 %f97, %f17, %f453;
and.b32 %r12, %r11, 1;
setp.eq.s32 %p48, %r11, 1;
mov.f32 %f1326, 0f00000000;
mov.u32 %r416, 0;
@%p48 bra $L__BB0_37;
sub.s32 %r415, %r11, %r12;
$L__BB0_36:
cvt.rn.f32.s32 %f473, %r416;
mul.f32 %f474, %f473, 0f3DD32618;
cvt.rmi.f32.f32 %f475, %f474;
sub.f32 %f476, %f474, %f475;
mul.f32 %f477, %f473, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f478, %f477;
sub.f32 %f479, %f477, %f478;
mul.f32 %f480, %f473, 0f3DC74539;
cvt.rmi.f32.f32 %f481, %f480;
sub.f32 %f482, %f480, %f481;
add.f32 %f483, %f479, 0f4199851F;
add.f32 %f484, %f482, 0f4199851F;
add.f32 %f485, %f476, 0f4199851F;
mul.f32 %f486, %f479, %f484;
fma.rn.f32 %f487, %f476, %f483, %f486;
fma.rn.f32 %f488, %f485, %f482, %f487;
add.f32 %f489, %f476, %f488;
add.f32 %f490, %f479, %f488;
add.f32 %f491, %f482, %f488;
add.f32 %f492, %f489, %f490;
mul.f32 %f493, %f491, %f492;
cvt.rmi.f32.f32 %f494, %f493;
sub.f32 %f495, %f493, %f494;
add.f32 %f496, %f489, %f491;
mul.f32 %f497, %f490, %f496;
cvt.rmi.f32.f32 %f498, %f497;
sub.f32 %f499, %f497, %f498;
add.f32 %f500, %f490, %f491;
mul.f32 %f501, %f489, %f500;
cvt.rmi.f32.f32 %f502, %f501;
sub.f32 %f503, %f501, %f502;
fma.rn.f32 %f504, %f495, 0f40000000, 0fBF800000;
fma.rn.f32 %f505, %f499, 0f40000000, 0fBF800000;
fma.rn.f32 %f506, %f503, 0f40000000, 0fBF800000;
ld.global.f32 %f507, [%rd5+16];
ld.global.f32 %f508, [%rd5+4];
fma.rn.f32 %f509, %f507, %f504, %f508;
ld.global.f32 %f510, [%rd5+8];
fma.rn.f32 %f511, %f507, %f505, %f510;
ld.global.f32 %f512, [%rd5+12];
fma.rn.f32 %f513, %f507, %f506, %f512;
sub.f32 %f514, %f509, %f10;
sub.f32 %f515, %f511, %f11;
sub.f32 %f516, %f513, %f12;
mul.f32 %f517, %f515, %f515;
fma.rn.f32 %f518, %f514, %f514, %f517;
fma.rn.f32 %f519, %f516, %f516, %f518;
sqrt.rn.f32 %f462, %f519;
rcp.rn.f32 %f520, %f462;
mul.f32 %f458, %f520, %f514;
mul.f32 %f459, %f520, %f515;
mul.f32 %f460, %f520, %f516;
mov.f32 %f472, 0f00000000;
mov.u32 %r169, 2;
mov.u32 %r171, 1;
mov.u32 %r172, 1065353216;
mov.u32 %r203, 0;
// begin inline asm
call(%r62,%r63,%r64,%r65,%r66,%r67,%r68,%r69,%r70,%r71,%r72,%r73,%r74,%r75,%r76,%r77,%r78,%r79,%r80,%r81,%r82,%r83,%r84,%r85,%r86,%r87,%r88,%r89,%r90,%r91,%r92,%r93),_optix_trace_typed_32,(%r203,%rd3,%f13,%f14,%f15,%f458,%f459,%f460,%f97,%f462,%f472,%r171,%r203,%r171,%r169,%r171,%r171,%r172,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203);
// end inline asm
mov.b32 %f521, %r62;
add.f32 %f522, %f1326, %f521;
add.s32 %r204, %r416, 1;
cvt.rn.f32.s32 %f523, %r204;
mul.f32 %f524, %f523, 0f3DD32618;
cvt.rmi.f32.f32 %f525, %f524;
sub.f32 %f526, %f524, %f525;
mul.f32 %f527, %f523, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f528, %f527;
sub.f32 %f529, %f527, %f528;
mul.f32 %f530, %f523, 0f3DC74539;
cvt.rmi.f32.f32 %f531, %f530;
sub.f32 %f532, %f530, %f531;
add.f32 %f533, %f529, 0f4199851F;
add.f32 %f534, %f532, 0f4199851F;
add.f32 %f535, %f526, 0f4199851F;
mul.f32 %f536, %f529, %f534;
fma.rn.f32 %f537, %f526, %f533, %f536;
fma.rn.f32 %f538, %f535, %f532, %f537;
add.f32 %f539, %f526, %f538;
add.f32 %f540, %f529, %f538;
add.f32 %f541, %f532, %f538;
add.f32 %f542, %f539, %f540;
mul.f32 %f543, %f541, %f542;
cvt.rmi.f32.f32 %f544, %f543;
sub.f32 %f545, %f543, %f544;
add.f32 %f546, %f539, %f541;
mul.f32 %f547, %f540, %f546;
cvt.rmi.f32.f32 %f548, %f547;
sub.f32 %f549, %f547, %f548;
add.f32 %f550, %f540, %f541;
mul.f32 %f551, %f539, %f550;
cvt.rmi.f32.f32 %f552, %f551;
sub.f32 %f553, %f551, %f552;
fma.rn.f32 %f554, %f545, 0f40000000, 0fBF800000;
fma.rn.f32 %f555, %f549, 0f40000000, 0fBF800000;
fma.rn.f32 %f556, %f553, 0f40000000, 0fBF800000;
ld.global.f32 %f557, [%rd5+16];
ld.global.f32 %f558, [%rd5+4];
fma.rn.f32 %f559, %f557, %f554, %f558;
ld.global.f32 %f560, [%rd5+8];
fma.rn.f32 %f561, %f557, %f555, %f560;
ld.global.f32 %f562, [%rd5+12];
fma.rn.f32 %f563, %f557, %f556, %f562;
sub.f32 %f564, %f559, %f10;
sub.f32 %f565, %f561, %f11;
sub.f32 %f566, %f563, %f12;
mul.f32 %f567, %f565, %f565;
fma.rn.f32 %f568, %f564, %f564, %f567;
fma.rn.f32 %f569, %f566, %f566, %f568;
sqrt.rn.f32 %f471, %f569;
rcp.rn.f32 %f570, %f471;
mul.f32 %f467, %f570, %f564;
mul.f32 %f468, %f570, %f565;
mul.f32 %f469, %f570, %f566;
// begin inline asm
call(%r133,%r134,%r135,%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164),_optix_trace_typed_32,(%r203,%rd3,%f13,%f14,%f15,%f467,%f468,%f469,%f97,%f471,%f472,%r171,%r203,%r171,%r169,%r171,%r171,%r172,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203);
// end inline asm
mov.b32 %f571, %r133;
add.f32 %f1326, %f522, %f571;
add.s32 %r416, %r416, 2;
add.s32 %r415, %r415, -2;
setp.ne.s32 %p49, %r415, 0;
@%p49 bra $L__BB0_36;
$L__BB0_37:
setp.eq.s32 %p50, %r12, 0;
@%p50 bra $L__BB0_39;
cvt.rn.f32.s32 %f581, %r416;
mul.f32 %f582, %f581, 0f3DD32618;
cvt.rmi.f32.f32 %f583, %f582;
sub.f32 %f584, %f582, %f583;
mul.f32 %f585, %f581, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f586, %f585;
sub.f32 %f587, %f585, %f586;
mul.f32 %f588, %f581, 0f3DC74539;
cvt.rmi.f32.f32 %f589, %f588;
sub.f32 %f590, %f588, %f589;
add.f32 %f591, %f587, 0f4199851F;
add.f32 %f592, %f590, 0f4199851F;
add.f32 %f593, %f584, 0f4199851F;
mul.f32 %f594, %f587, %f592;
fma.rn.f32 %f595, %f584, %f591, %f594;
fma.rn.f32 %f596, %f593, %f590, %f595;
add.f32 %f597, %f584, %f596;
add.f32 %f598, %f587, %f596;
add.f32 %f599, %f590, %f596;
add.f32 %f600, %f597, %f598;
mul.f32 %f601, %f599, %f600;
cvt.rmi.f32.f32 %f602, %f601;
sub.f32 %f603, %f601, %f602;
add.f32 %f604, %f597, %f599;
mul.f32 %f605, %f598, %f604;
cvt.rmi.f32.f32 %f606, %f605;
sub.f32 %f607, %f605, %f606;
add.f32 %f608, %f598, %f599;
mul.f32 %f609, %f597, %f608;
cvt.rmi.f32.f32 %f610, %f609;
sub.f32 %f611, %f609, %f610;
fma.rn.f32 %f612, %f603, 0f40000000, 0fBF800000;
fma.rn.f32 %f613, %f607, 0f40000000, 0fBF800000;
fma.rn.f32 %f614, %f611, 0f40000000, 0fBF800000;
ld.global.f32 %f615, [%rd5+16];
ld.global.f32 %f616, [%rd5+4];
fma.rn.f32 %f617, %f615, %f612, %f616;
ld.global.f32 %f618, [%rd5+8];
fma.rn.f32 %f619, %f615, %f613, %f618;
ld.global.f32 %f620, [%rd5+12];
fma.rn.f32 %f621, %f615, %f614, %f620;
sub.f32 %f622, %f617, %f10;
sub.f32 %f623, %f619, %f11;
sub.f32 %f624, %f621, %f12;
mul.f32 %f625, %f623, %f623;
fma.rn.f32 %f626, %f622, %f622, %f625;
fma.rn.f32 %f627, %f624, %f624, %f626;
sqrt.rn.f32 %f579, %f627;
rcp.rn.f32 %f628, %f579;
mul.f32 %f575, %f628, %f622;
mul.f32 %f576, %f628, %f623;
mul.f32 %f577, %f628, %f624;
mov.f32 %f580, 0f00000000;
mov.u32 %r241, 2;
mov.u32 %r243, 1;
mov.u32 %r244, 1065353216;
mov.u32 %r275, 0;
// begin inline asm
call(%r205,%r206,%r207,%r208,%r209,%r210,%r211,%r212,%r213,%r214,%r215,%r216,%r217,%r218,%r219,%r220,%r221,%r222,%r223,%r224,%r225,%r226,%r227,%r228,%r229,%r230,%r231,%r232,%r233,%r234,%r235,%r236),_optix_trace_typed_32,(%r275,%rd3,%f13,%f14,%f15,%f575,%f576,%f577,%f97,%f579,%f580,%r243,%r275,%r243,%r241,%r243,%r243,%r244,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275);
// end inline asm
mov.b32 %f629, %r205;
add.f32 %f1326, %f1326, %f629;
$L__BB0_39:
cvt.rn.f32.s32 %f630, %r11;
div.rn.f32 %f1333, %f1326, %f630;
shr.u32 %r276, %r42, 31;
cvt.u16.u32 %rs64, %r276;
$L__BB0_40:
fma.rn.f32 %f1334, %f1319, %f1333, %f1334;
fma.rn.f32 %f1335, %f1320, %f1333, %f1335;
fma.rn.f32 %f1336, %f1321, %f1333, %f1336;
setp.eq.s32 %p51, %r8, 0;
@%p51 bra $L__BB0_42;
mul.f32 %f631, %f1320, 0f3F372474;
fma.rn.f32 %f632, %f1319, 0f3E59999A, %f631;
fma.rn.f32 %f633, %f1321, 0f3D93A92A, %f632;
fma.rn.f32 %f1337, %f1322, %f633, %f1337;
fma.rn.f32 %f1338, %f1323, %f633, %f1338;
fma.rn.f32 %f1339, %f633, %f1324, %f1339;
$L__BB0_42:
setp.eq.s16 %p52, %rs64, 0;
@%p52 bra $L__BB0_44;
div.rn.f32 %f634, %f1319, %f277;
div.rn.f32 %f635, %f634, %f63;
cvt.sat.f32.f32 %f636, %f635;
mul.f32 %f1333, %f1333, %f636;
$L__BB0_44:
add.f32 %f1340, %f1340, %f1333;
$L__BB0_45:
cvt.u64.u32 %rd53, %r413;
cvt.u32.u64 %r277, %rd53;
add.s32 %r413, %r277, 1;
setp.lt.u32 %p53, %r413, %r5;
@%p53 bra $L__BB0_7;
bra.uni $L__BB0_47;
$L__BB0_46:
mov.f32 %f1334, 0f00000000;
ld.const.u32 %r417, [params+104];
mov.f32 %f1335, %f1334;
mov.f32 %f1336, %f1334;
mov.f32 %f1337, %f1334;
mov.f32 %f1338, %f1334;
mov.f32 %f1339, %f1334;
mov.f32 %f1340, %f1334;
$L__BB0_47:
and.b32 %r278, %r417, 8;
setp.eq.s32 %p54, %r278, 0;
@%p54 bra $L__BB0_61;
cvt.sat.f32.f32 %f132, %f1340;
ld.const.u64 %rd25, [params+192];
cvta.to.global.u64 %rd7, %rd25;
ld.const.u32 %r279, [params+184];
mad.lo.s32 %r280, %r279, %r4, %r3;
cvt.u64.u32 %rd8, %r280;
mov.f32 %f645, 0f3E68BA2E;
cvt.rzi.f32.f32 %f646, %f645;
add.f32 %f647, %f646, %f646;
mov.f32 %f648, 0f3EE8BA2E;
sub.f32 %f649, %f648, %f647;
abs.f32 %f133, %f649;
abs.f32 %f134, %f132;
setp.lt.f32 %p55, %f134, 0f00800000;
mul.f32 %f650, %f134, 0f4B800000;
selp.f32 %f651, %f650, %f134, %p55;
selp.f32 %f652, 0fC3170000, 0fC2FE0000, %p55;
mov.b32 %r281, %f651;
and.b32 %r282, %r281, 8388607;
or.b32 %r283, %r282, 1065353216;
mov.b32 %f653, %r283;
shr.u32 %r284, %r281, 23;
cvt.rn.f32.u32 %f654, %r284;
add.f32 %f655, %f652, %f654;
setp.gt.f32 %p56, %f653, 0f3FB504F3;
mul.f32 %f656, %f653, 0f3F000000;
add.f32 %f657, %f655, 0f3F800000;
selp.f32 %f658, %f657, %f655, %p56;
selp.f32 %f659, %f656, %f653, %p56;
add.f32 %f660, %f659, 0fBF800000;
add.f32 %f661, %f659, 0f3F800000;
rcp.approx.ftz.f32 %f662, %f661;
add.f32 %f663, %f660, %f660;
mul.f32 %f664, %f663, %f662;
mul.f32 %f665, %f664, %f664;
mov.f32 %f666, 0f3C4CAF63;
mov.f32 %f667, 0f3B18F0FE;
fma.rn.f32 %f668, %f667, %f665, %f666;
mov.f32 %f669, 0f3DAAAABD;
fma.rn.f32 %f670, %f668, %f665, %f669;
mul.rn.f32 %f671, %f670, %f665;
mul.rn.f32 %f672, %f671, %f664;
sub.f32 %f673, %f660, %f664;
add.f32 %f674, %f673, %f673;
neg.f32 %f675, %f664;
fma.rn.f32 %f676, %f675, %f660, %f674;
mul.rn.f32 %f677, %f662, %f676;
add.f32 %f678, %f672, %f664;
sub.f32 %f679, %f664, %f678;
add.f32 %f680, %f672, %f679;
add.f32 %f681, %f677, %f680;
add.f32 %f682, %f678, %f681;
sub.f32 %f683, %f678, %f682;
add.f32 %f684, %f681, %f683;
mov.f32 %f685, 0f3F317200;
mul.rn.f32 %f686, %f658, %f685;
mov.f32 %f687, 0f35BFBE8E;
mul.rn.f32 %f688, %f658, %f687;
add.f32 %f689, %f686, %f682;
sub.f32 %f690, %f686, %f689;
add.f32 %f691, %f682, %f690;
add.f32 %f692, %f684, %f691;
add.f32 %f693, %f688, %f692;
add.f32 %f694, %f689, %f693;
sub.f32 %f695, %f689, %f694;
add.f32 %f696, %f693, %f695;
mul.rn.f32 %f697, %f648, %f694;
neg.f32 %f698, %f697;
fma.rn.f32 %f699, %f648, %f694, %f698;
fma.rn.f32 %f700, %f648, %f696, %f699;
mov.f32 %f701, 0f00000000;
fma.rn.f32 %f702, %f701, %f694, %f700;
add.rn.f32 %f703, %f697, %f702;
neg.f32 %f704, %f703;
add.rn.f32 %f705, %f697, %f704;
add.rn.f32 %f706, %f705, %f702;
mov.b32 %r285, %f703;
setp.eq.s32 %p57, %r285, 1118925336;
add.s32 %r286, %r285, -1;
mov.b32 %f707, %r286;
add.f32 %f708, %f706, 0f37000000;
selp.f32 %f135, %f708, %f706, %p57;
selp.f32 %f709, %f707, %f703, %p57;
mov.f32 %f710, 0f3FB8AA3B;
mul.rn.f32 %f711, %f709, %f710;
cvt.rzi.f32.f32 %f712, %f711;
abs.f32 %f713, %f712;
setp.gt.f32 %p58, %f713, 0f42FC0000;
mov.b32 %r287, %f712;
and.b32 %r288, %r287, -2147483648;
or.b32 %r289, %r288, 1123811328;
mov.b32 %f714, %r289;
selp.f32 %f715, %f714, %f712, %p58;
mov.f32 %f716, 0fBF317218;
fma.rn.f32 %f717, %f715, %f716, %f709;
mov.f32 %f718, 0f3102E308;
fma.rn.f32 %f719, %f715, %f718, %f717;
mul.f32 %f720, %f719, 0f3FB8AA3B;
add.f32 %f721, %f715, 0f4B40007F;
mov.b32 %r290, %f721;
shl.b32 %r291, %r290, 23;
mov.b32 %f722, %r291;
ex2.approx.ftz.f32 %f723, %f720;
mul.f32 %f136, %f723, %f722;
setp.eq.f32 %p59, %f136, 0f7F800000;
mov.f32 %f1348, 0f7F800000;
@%p59 bra $L__BB0_50;
fma.rn.f32 %f1348, %f136, %f135, %f136;
$L__BB0_50:
setp.lt.f32 %p60, %f132, 0f00000000;
setp.eq.f32 %p61, %f133, 0f3F800000;
and.pred %p3, %p60, %p61;
setp.eq.f32 %p62, %f132, 0f00000000;
@%p62 bra $L__BB0_54;
bra.uni $L__BB0_51;
$L__BB0_54:
add.f32 %f728, %f132, %f132;
selp.f32 %f1350, %f728, 0f00000000, %p61;
bra.uni $L__BB0_55;
$L__BB0_148:
mov.f32 %f1294, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs59, %f1294;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs58, %f1294;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs57, %f1294;}
// end inline asm
mov.u16 %rs60, 0;
st.global.v4.u16 [%rd12], {%rs57, %rs58, %rs59, %rs60};
$L__BB0_149:
and.b32 %r410, %r22, 64;
setp.eq.s32 %p177, %r410, 0;
@%p177 bra $L__BB0_151;
ld.const.u64 %rd49, [params+208];
cvta.to.global.u64 %rd50, %rd49;
ld.const.u32 %r411, [params+200];
mad.lo.s32 %r412, %r411, %r4, %r3;
mul.wide.u32 %rd51, %r412, 4;
add.s64 %rd52, %rd50, %rd51;
mov.u16 %rs61, 0;
st.global.v4.u8 [%rd52], {%rs61, %rs61, %rs61, %rs61};
bra.uni $L__BB0_151;
$L__BB0_51:
mov.b32 %r292, %f1348;
xor.b32 %r293, %r292, -2147483648;
mov.b32 %f724, %r293;
selp.f32 %f1350, %f724, %f1348, %p3;
setp.geu.f32 %p63, %f132, 0f00000000;
@%p63 bra $L__BB0_55;
mov.f32 %f725, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f726, %f725;
setp.eq.f32 %p64, %f726, 0f3EE8BA2E;
@%p64 bra $L__BB0_55;
mov.f32 %f1350, 0f7FFFFFFF;
$L__BB0_55:
add.f32 %f729, %f134, 0f3EE8BA2E;
mov.b32 %r294, %f729;
setp.lt.s32 %p66, %r294, 2139095040;
@%p66 bra $L__BB0_60;
setp.gtu.f32 %p67, %f134, 0f7F800000;
@%p67 bra $L__BB0_59;
bra.uni $L__BB0_57;
$L__BB0_59:
add.f32 %f1350, %f132, 0f3EE8BA2E;
bra.uni $L__BB0_60;
$L__BB0_57:
setp.neu.f32 %p68, %f134, 0f7F800000;
@%p68 bra $L__BB0_60;
selp.f32 %f1350, 0fFF800000, 0f7F800000, %p3;
$L__BB0_60:
mul.f32 %f730, %f1350, 0f437F0000;
setp.eq.f32 %p69, %f132, 0f3F800000;
selp.f32 %f731, 0f437F0000, %f730, %p69;
cvt.rzi.u32.f32 %r295, %f731;
shl.b64 %rd26, %rd8, 1;
add.s64 %rd27, %rd7, %rd26;
cvt.u16.u32 %rs18, %r295;
mov.u16 %rs19, 255;
st.global.v2.u8 [%rd27], {%rs18, %rs19};
$L__BB0_61:
and.b32 %r296, %r417, 1;
setp.eq.b32 %p70, %r296, 1;
mov.pred %p71, 0;
xor.pred %p72, %p70, %p71;
not.pred %p73, %p72;
@%p73 bra $L__BB0_135;
mov.f32 %f733, 0f3E666666;
cvt.rzi.f32.f32 %f734, %f733;
add.f32 %f735, %f734, %f734;
mov.f32 %f736, 0f3EE66666;
sub.f32 %f737, %f736, %f735;
abs.f32 %f145, %f737;
abs.f32 %f146, %f1334;
setp.lt.f32 %p74, %f146, 0f00800000;
mul.f32 %f738, %f146, 0f4B800000;
selp.f32 %f739, %f738, %f146, %p74;
selp.f32 %f740, 0fC3170000, 0fC2FE0000, %p74;
mov.b32 %r297, %f739;
and.b32 %r298, %r297, 8388607;
or.b32 %r299, %r298, 1065353216;
mov.b32 %f741, %r299;
shr.u32 %r300, %r297, 23;
cvt.rn.f32.u32 %f742, %r300;
add.f32 %f743, %f740, %f742;
setp.gt.f32 %p75, %f741, 0f3FB504F3;
mul.f32 %f744, %f741, 0f3F000000;
add.f32 %f745, %f743, 0f3F800000;
selp.f32 %f746, %f745, %f743, %p75;
selp.f32 %f747, %f744, %f741, %p75;
add.f32 %f748, %f747, 0fBF800000;
add.f32 %f749, %f747, 0f3F800000;
rcp.approx.ftz.f32 %f750, %f749;
add.f32 %f751, %f748, %f748;
mul.f32 %f752, %f751, %f750;
mul.f32 %f753, %f752, %f752;
mov.f32 %f754, 0f3C4CAF63;
mov.f32 %f755, 0f3B18F0FE;
fma.rn.f32 %f756, %f755, %f753, %f754;
mov.f32 %f757, 0f3DAAAABD;
fma.rn.f32 %f758, %f756, %f753, %f757;
mul.rn.f32 %f759, %f758, %f753;
mul.rn.f32 %f760, %f759, %f752;
sub.f32 %f761, %f748, %f752;
add.f32 %f762, %f761, %f761;
neg.f32 %f763, %f752;
fma.rn.f32 %f764, %f763, %f748, %f762;
mul.rn.f32 %f765, %f750, %f764;
add.f32 %f766, %f760, %f752;
sub.f32 %f767, %f752, %f766;
add.f32 %f768, %f760, %f767;
add.f32 %f769, %f765, %f768;
add.f32 %f770, %f766, %f769;
sub.f32 %f771, %f766, %f770;
add.f32 %f772, %f769, %f771;
mov.f32 %f773, 0f3F317200;
mul.rn.f32 %f774, %f746, %f773;
mov.f32 %f775, 0f35BFBE8E;
mul.rn.f32 %f776, %f746, %f775;
add.f32 %f777, %f774, %f770;
sub.f32 %f778, %f774, %f777;
add.f32 %f779, %f770, %f778;
add.f32 %f780, %f772, %f779;
add.f32 %f781, %f776, %f780;
add.f32 %f782, %f777, %f781;
sub.f32 %f783, %f777, %f782;
add.f32 %f784, %f781, %f783;
mul.rn.f32 %f785, %f736, %f782;
neg.f32 %f786, %f785;
fma.rn.f32 %f787, %f736, %f782, %f786;
fma.rn.f32 %f788, %f736, %f784, %f787;
mov.f32 %f789, 0f00000000;
fma.rn.f32 %f790, %f789, %f782, %f788;
add.rn.f32 %f791, %f785, %f790;
neg.f32 %f792, %f791;
add.rn.f32 %f793, %f785, %f792;
add.rn.f32 %f794, %f793, %f790;
mov.b32 %r301, %f791;
setp.eq.s32 %p76, %r301, 1118925336;
add.s32 %r302, %r301, -1;
mov.b32 %f795, %r302;
add.f32 %f796, %f794, 0f37000000;
selp.f32 %f147, %f796, %f794, %p76;
selp.f32 %f797, %f795, %f791, %p76;
mov.f32 %f798, 0f3FB8AA3B;
mul.rn.f32 %f799, %f797, %f798;
cvt.rzi.f32.f32 %f800, %f799;
abs.f32 %f801, %f800;
setp.gt.f32 %p77, %f801, 0f42FC0000;
mov.b32 %r303, %f800;
and.b32 %r304, %r303, -2147483648;
or.b32 %r305, %r304, 1123811328;
mov.b32 %f802, %r305;
selp.f32 %f803, %f802, %f800, %p77;
mov.f32 %f804, 0fBF317218;
fma.rn.f32 %f805, %f803, %f804, %f797;
mov.f32 %f806, 0f3102E308;
fma.rn.f32 %f807, %f803, %f806, %f805;
mul.f32 %f808, %f807, 0f3FB8AA3B;
add.f32 %f809, %f803, 0f4B40007F;
mov.b32 %r306, %f809;
shl.b32 %r307, %r306, 23;
mov.b32 %f810, %r307;
ex2.approx.ftz.f32 %f811, %f808;
mul.f32 %f148, %f811, %f810;
setp.eq.f32 %p78, %f148, 0f7F800000;
mov.f32 %f1351, 0f7F800000;
@%p78 bra $L__BB0_64;
fma.rn.f32 %f1351, %f148, %f147, %f148;
$L__BB0_64:
setp.lt.f32 %p79, %f1334, 0f00000000;
setp.eq.f32 %p80, %f145, 0f3F800000;
and.pred %p4, %p79, %p80;
setp.eq.f32 %p81, %f1334, 0f00000000;
@%p81 bra $L__BB0_68;
bra.uni $L__BB0_65;
$L__BB0_68:
add.f32 %f816, %f1334, %f1334;
selp.f32 %f1353, %f816, 0f00000000, %p80;
bra.uni $L__BB0_69;
$L__BB0_65:
mov.b32 %r308, %f1351;
xor.b32 %r309, %r308, -2147483648;
mov.b32 %f812, %r309;
selp.f32 %f1353, %f812, %f1351, %p4;
setp.geu.f32 %p82, %f1334, 0f00000000;
@%p82 bra $L__BB0_69;
mov.f32 %f813, 0f3EE66666;
cvt.rzi.f32.f32 %f814, %f813;
setp.eq.f32 %p83, %f814, 0f3EE66666;
@%p83 bra $L__BB0_69;
mov.f32 %f1353, 0f7FFFFFFF;
$L__BB0_69:
add.f32 %f817, %f146, 0f3EE66666;
mov.b32 %r310, %f817;
setp.lt.s32 %p85, %r310, 2139095040;
@%p85 bra $L__BB0_74;
setp.gtu.f32 %p86, %f146, 0f7F800000;
@%p86 bra $L__BB0_73;
bra.uni $L__BB0_71;
$L__BB0_73:
add.f32 %f1353, %f1334, 0f3EE66666;
bra.uni $L__BB0_74;
$L__BB0_71:
setp.neu.f32 %p87, %f146, 0f7F800000;
@%p87 bra $L__BB0_74;
selp.f32 %f1353, 0fFF800000, 0f7F800000, %p4;
$L__BB0_74:
setp.eq.f32 %p88, %f1334, 0f3F800000;
selp.f32 %f157, 0f3F800000, %f1353, %p88;
abs.f32 %f158, %f1335;
setp.lt.f32 %p89, %f158, 0f00800000;
mul.f32 %f819, %f158, 0f4B800000;
selp.f32 %f820, %f819, %f158, %p89;
selp.f32 %f821, 0fC3170000, 0fC2FE0000, %p89;
mov.b32 %r311, %f820;
and.b32 %r312, %r311, 8388607;
or.b32 %r313, %r312, 1065353216;
mov.b32 %f822, %r313;
shr.u32 %r314, %r311, 23;
cvt.rn.f32.u32 %f823, %r314;
add.f32 %f824, %f821, %f823;
setp.gt.f32 %p90, %f822, 0f3FB504F3;
mul.f32 %f825, %f822, 0f3F000000;
add.f32 %f826, %f824, 0f3F800000;
selp.f32 %f827, %f826, %f824, %p90;
selp.f32 %f828, %f825, %f822, %p90;
add.f32 %f829, %f828, 0fBF800000;
add.f32 %f830, %f828, 0f3F800000;
rcp.approx.ftz.f32 %f831, %f830;
add.f32 %f832, %f829, %f829;
mul.f32 %f833, %f832, %f831;
mul.f32 %f834, %f833, %f833;
mov.f32 %f835, 0f3C4CAF63;
mov.f32 %f836, 0f3B18F0FE;
fma.rn.f32 %f837, %f836, %f834, %f835;
mov.f32 %f838, 0f3DAAAABD;
fma.rn.f32 %f839, %f837, %f834, %f838;
mul.rn.f32 %f840, %f839, %f834;
mul.rn.f32 %f841, %f840, %f833;
sub.f32 %f842, %f829, %f833;
add.f32 %f843, %f842, %f842;
neg.f32 %f844, %f833;
fma.rn.f32 %f845, %f844, %f829, %f843;
mul.rn.f32 %f846, %f831, %f845;
add.f32 %f847, %f841, %f833;
sub.f32 %f848, %f833, %f847;
add.f32 %f849, %f841, %f848;
add.f32 %f850, %f846, %f849;
add.f32 %f851, %f847, %f850;
sub.f32 %f852, %f847, %f851;
add.f32 %f853, %f850, %f852;
mov.f32 %f854, 0f3F317200;
mul.rn.f32 %f855, %f827, %f854;
mov.f32 %f856, 0f35BFBE8E;
mul.rn.f32 %f857, %f827, %f856;
add.f32 %f858, %f855, %f851;
sub.f32 %f859, %f855, %f858;
add.f32 %f860, %f851, %f859;
add.f32 %f861, %f853, %f860;
add.f32 %f862, %f857, %f861;
add.f32 %f863, %f858, %f862;
sub.f32 %f864, %f858, %f863;
add.f32 %f865, %f862, %f864;
mov.f32 %f866, 0f3EE66666;
mul.rn.f32 %f867, %f866, %f863;
neg.f32 %f868, %f867;
fma.rn.f32 %f869, %f866, %f863, %f868;
fma.rn.f32 %f870, %f866, %f865, %f869;
mov.f32 %f871, 0f00000000;
fma.rn.f32 %f872, %f871, %f863, %f870;
add.rn.f32 %f873, %f867, %f872;
neg.f32 %f874, %f873;
add.rn.f32 %f875, %f867, %f874;
add.rn.f32 %f876, %f875, %f872;
mov.b32 %r315, %f873;
setp.eq.s32 %p91, %r315, 1118925336;
add.s32 %r316, %r315, -1;
mov.b32 %f877, %r316;
add.f32 %f878, %f876, 0f37000000;
selp.f32 %f159, %f878, %f876, %p91;
selp.f32 %f879, %f877, %f873, %p91;
mov.f32 %f880, 0f3FB8AA3B;
mul.rn.f32 %f881, %f879, %f880;
cvt.rzi.f32.f32 %f882, %f881;
abs.f32 %f883, %f882;
setp.gt.f32 %p92, %f883, 0f42FC0000;
mov.b32 %r317, %f882;
and.b32 %r318, %r317, -2147483648;
or.b32 %r319, %r318, 1123811328;
mov.b32 %f884, %r319;
selp.f32 %f885, %f884, %f882, %p92;
mov.f32 %f886, 0fBF317218;
fma.rn.f32 %f887, %f885, %f886, %f879;
mov.f32 %f888, 0f3102E308;
fma.rn.f32 %f889, %f885, %f888, %f887;
mul.f32 %f890, %f889, 0f3FB8AA3B;
add.f32 %f891, %f885, 0f4B40007F;
mov.b32 %r320, %f891;
shl.b32 %r321, %r320, 23;
mov.b32 %f892, %r321;
ex2.approx.ftz.f32 %f893, %f890;
mul.f32 %f160, %f893, %f892;
setp.eq.f32 %p93, %f160, 0f7F800000;
mov.f32 %f1354, 0f7F800000;
@%p93 bra $L__BB0_76;
fma.rn.f32 %f1354, %f160, %f159, %f160;
$L__BB0_76:
setp.lt.f32 %p94, %f1335, 0f00000000;
and.pred %p5, %p94, %p80;
setp.eq.f32 %p96, %f1335, 0f00000000;
@%p96 bra $L__BB0_80;
bra.uni $L__BB0_77;
$L__BB0_80:
add.f32 %f898, %f1335, %f1335;
selp.f32 %f1356, %f898, 0f00000000, %p80;
bra.uni $L__BB0_81;
$L__BB0_77:
mov.b32 %r322, %f1354;
xor.b32 %r323, %r322, -2147483648;
mov.b32 %f894, %r323;
selp.f32 %f1356, %f894, %f1354, %p5;
setp.geu.f32 %p97, %f1335, 0f00000000;
@%p97 bra $L__BB0_81;
mov.f32 %f895, 0f3EE66666;
cvt.rzi.f32.f32 %f896, %f895;
setp.eq.f32 %p98, %f896, 0f3EE66666;
@%p98 bra $L__BB0_81;
mov.f32 %f1356, 0f7FFFFFFF;
$L__BB0_81:
add.f32 %f899, %f158, 0f3EE66666;
mov.b32 %r324, %f899;
setp.lt.s32 %p100, %r324, 2139095040;
@%p100 bra $L__BB0_86;
setp.gtu.f32 %p101, %f158, 0f7F800000;
@%p101 bra $L__BB0_85;
bra.uni $L__BB0_83;
$L__BB0_85:
add.f32 %f1356, %f1335, 0f3EE66666;
bra.uni $L__BB0_86;
$L__BB0_83:
setp.neu.f32 %p102, %f158, 0f7F800000;
@%p102 bra $L__BB0_86;
selp.f32 %f1356, 0fFF800000, 0f7F800000, %p5;
$L__BB0_86:
setp.eq.f32 %p103, %f1335, 0f3F800000;
selp.f32 %f169, 0f3F800000, %f1356, %p103;
abs.f32 %f170, %f1336;
setp.lt.f32 %p104, %f170, 0f00800000;
mul.f32 %f901, %f170, 0f4B800000;
selp.f32 %f902, %f901, %f170, %p104;
selp.f32 %f903, 0fC3170000, 0fC2FE0000, %p104;
mov.b32 %r325, %f902;
and.b32 %r326, %r325, 8388607;
or.b32 %r327, %r326, 1065353216;
mov.b32 %f904, %r327;
shr.u32 %r328, %r325, 23;
cvt.rn.f32.u32 %f905, %r328;
add.f32 %f906, %f903, %f905;
setp.gt.f32 %p105, %f904, 0f3FB504F3;
mul.f32 %f907, %f904, 0f3F000000;
add.f32 %f908, %f906, 0f3F800000;
selp.f32 %f909, %f908, %f906, %p105;
selp.f32 %f910, %f907, %f904, %p105;
add.f32 %f911, %f910, 0fBF800000;
add.f32 %f912, %f910, 0f3F800000;
rcp.approx.ftz.f32 %f913, %f912;
add.f32 %f914, %f911, %f911;
mul.f32 %f915, %f914, %f913;
mul.f32 %f916, %f915, %f915;
mov.f32 %f917, 0f3C4CAF63;
mov.f32 %f918, 0f3B18F0FE;
fma.rn.f32 %f919, %f918, %f916, %f917;
mov.f32 %f920, 0f3DAAAABD;
fma.rn.f32 %f921, %f919, %f916, %f920;
mul.rn.f32 %f922, %f921, %f916;
mul.rn.f32 %f923, %f922, %f915;
sub.f32 %f924, %f911, %f915;
add.f32 %f925, %f924, %f924;
neg.f32 %f926, %f915;
fma.rn.f32 %f927, %f926, %f911, %f925;
mul.rn.f32 %f928, %f913, %f927;
add.f32 %f929, %f923, %f915;
sub.f32 %f930, %f915, %f929;
add.f32 %f931, %f923, %f930;
add.f32 %f932, %f928, %f931;
add.f32 %f933, %f929, %f932;
sub.f32 %f934, %f929, %f933;
add.f32 %f935, %f932, %f934;
mov.f32 %f936, 0f3F317200;
mul.rn.f32 %f937, %f909, %f936;
mov.f32 %f938, 0f35BFBE8E;
mul.rn.f32 %f939, %f909, %f938;
add.f32 %f940, %f937, %f933;
sub.f32 %f941, %f937, %f940;
add.f32 %f942, %f933, %f941;
add.f32 %f943, %f935, %f942;
add.f32 %f944, %f939, %f943;
add.f32 %f945, %f940, %f944;
sub.f32 %f946, %f940, %f945;
add.f32 %f947, %f944, %f946;
mov.f32 %f948, 0f3EE66666;
mul.rn.f32 %f949, %f948, %f945;
neg.f32 %f950, %f949;
fma.rn.f32 %f951, %f948, %f945, %f950;
fma.rn.f32 %f952, %f948, %f947, %f951;
mov.f32 %f953, 0f00000000;
fma.rn.f32 %f954, %f953, %f945, %f952;
add.rn.f32 %f955, %f949, %f954;
neg.f32 %f956, %f955;
add.rn.f32 %f957, %f949, %f956;
add.rn.f32 %f958, %f957, %f954;
mov.b32 %r329, %f955;
setp.eq.s32 %p106, %r329, 1118925336;
add.s32 %r330, %r329, -1;
mov.b32 %f959, %r330;
add.f32 %f960, %f958, 0f37000000;
selp.f32 %f171, %f960, %f958, %p106;
selp.f32 %f961, %f959, %f955, %p106;
mov.f32 %f962, 0f3FB8AA3B;
mul.rn.f32 %f963, %f961, %f962;
cvt.rzi.f32.f32 %f964, %f963;
abs.f32 %f965, %f964;
setp.gt.f32 %p107, %f965, 0f42FC0000;
mov.b32 %r331, %f964;
and.b32 %r332, %r331, -2147483648;
or.b32 %r333, %r332, 1123811328;
mov.b32 %f966, %r333;
selp.f32 %f967, %f966, %f964, %p107;
mov.f32 %f968, 0fBF317218;
fma.rn.f32 %f969, %f967, %f968, %f961;
mov.f32 %f970, 0f3102E308;
fma.rn.f32 %f971, %f967, %f970, %f969;
mul.f32 %f972, %f971, 0f3FB8AA3B;
add.f32 %f973, %f967, 0f4B40007F;
mov.b32 %r334, %f973;
shl.b32 %r335, %r334, 23;
mov.b32 %f974, %r335;
ex2.approx.ftz.f32 %f975, %f972;
mul.f32 %f172, %f975, %f974;
setp.eq.f32 %p108, %f172, 0f7F800000;
mov.f32 %f1357, 0f7F800000;
@%p108 bra $L__BB0_88;
fma.rn.f32 %f1357, %f172, %f171, %f172;
$L__BB0_88:
setp.lt.f32 %p109, %f1336, 0f00000000;
and.pred %p6, %p109, %p80;
setp.eq.f32 %p111, %f1336, 0f00000000;
@%p111 bra $L__BB0_92;
bra.uni $L__BB0_89;
$L__BB0_92:
add.f32 %f980, %f1336, %f1336;
selp.f32 %f1359, %f980, 0f00000000, %p80;
bra.uni $L__BB0_93;
$L__BB0_89:
mov.b32 %r336, %f1357;
xor.b32 %r337, %r336, -2147483648;
mov.b32 %f976, %r337;
selp.f32 %f1359, %f976, %f1357, %p6;
setp.geu.f32 %p112, %f1336, 0f00000000;
@%p112 bra $L__BB0_93;
mov.f32 %f977, 0f3EE66666;
cvt.rzi.f32.f32 %f978, %f977;
setp.eq.f32 %p113, %f978, 0f3EE66666;
@%p113 bra $L__BB0_93;
mov.f32 %f1359, 0f7FFFFFFF;
$L__BB0_93:
add.f32 %f981, %f170, 0f3EE66666;
mov.b32 %r338, %f981;
setp.lt.s32 %p115, %r338, 2139095040;
@%p115 bra $L__BB0_98;
setp.gtu.f32 %p116, %f170, 0f7F800000;
@%p116 bra $L__BB0_97;
bra.uni $L__BB0_95;
$L__BB0_97:
add.f32 %f1359, %f1336, 0f3EE66666;
bra.uni $L__BB0_98;
$L__BB0_95:
setp.neu.f32 %p117, %f170, 0f7F800000;
@%p117 bra $L__BB0_98;
selp.f32 %f1359, 0fFF800000, 0f7F800000, %p6;
$L__BB0_98:
setp.eq.f32 %p118, %f1336, 0f3F800000;
mov.f32 %f983, 0f3F800000;
selp.f32 %f984, 0f3F800000, %f1359, %p118;
ld.const.u64 %rd28, [params+144];
cvta.to.global.u64 %rd9, %rd28;
ld.const.u32 %r339, [params+136];
mad.lo.s32 %r340, %r339, %r4, %r3;
cvt.u64.u32 %rd10, %r340;
min.f32 %f985, %f157, %f983;
mov.f32 %f986, 0f00000000;
max.f32 %f181, %f986, %f985;
min.f32 %f987, %f169, %f983;
max.f32 %f182, %f986, %f987;
min.f32 %f988, %f984, %f983;
max.f32 %f183, %f986, %f988;
mov.f32 %f989, 0f3E555555;
cvt.rzi.f32.f32 %f990, %f989;
add.f32 %f991, %f990, %f990;
mov.f32 %f992, 0f3ED55555;
sub.f32 %f993, %f992, %f991;
abs.f32 %f184, %f993;
abs.f32 %f185, %f181;
setp.lt.f32 %p119, %f185, 0f00800000;
mul.f32 %f994, %f185, 0f4B800000;
selp.f32 %f995, %f994, %f185, %p119;
selp.f32 %f996, 0fC3170000, 0fC2FE0000, %p119;
mov.b32 %r341, %f995;
and.b32 %r342, %r341, 8388607;
or.b32 %r343, %r342, 1065353216;
mov.b32 %f997, %r343;
shr.u32 %r344, %r341, 23;
cvt.rn.f32.u32 %f998, %r344;
add.f32 %f999, %f996, %f998;
setp.gt.f32 %p120, %f997, 0f3FB504F3;
mul.f32 %f1000, %f997, 0f3F000000;
add.f32 %f1001, %f999, 0f3F800000;
selp.f32 %f1002, %f1001, %f999, %p120;
selp.f32 %f1003, %f1000, %f997, %p120;
add.f32 %f1004, %f1003, 0fBF800000;
add.f32 %f1005, %f1003, 0f3F800000;
rcp.approx.ftz.f32 %f1006, %f1005;
add.f32 %f1007, %f1004, %f1004;
mul.f32 %f1008, %f1007, %f1006;
mul.f32 %f1009, %f1008, %f1008;
mov.f32 %f1010, 0f3C4CAF63;
mov.f32 %f1011, 0f3B18F0FE;
fma.rn.f32 %f1012, %f1011, %f1009, %f1010;
mov.f32 %f1013, 0f3DAAAABD;
fma.rn.f32 %f1014, %f1012, %f1009, %f1013;
mul.rn.f32 %f1015, %f1014, %f1009;
mul.rn.f32 %f1016, %f1015, %f1008;
sub.f32 %f1017, %f1004, %f1008;
add.f32 %f1018, %f1017, %f1017;
neg.f32 %f1019, %f1008;
fma.rn.f32 %f1020, %f1019, %f1004, %f1018;
mul.rn.f32 %f1021, %f1006, %f1020;
add.f32 %f1022, %f1016, %f1008;
sub.f32 %f1023, %f1008, %f1022;
add.f32 %f1024, %f1016, %f1023;
add.f32 %f1025, %f1021, %f1024;
add.f32 %f1026, %f1022, %f1025;
sub.f32 %f1027, %f1022, %f1026;
add.f32 %f1028, %f1025, %f1027;
mov.f32 %f1029, 0f3F317200;
mul.rn.f32 %f1030, %f1002, %f1029;
mov.f32 %f1031, 0f35BFBE8E;
mul.rn.f32 %f1032, %f1002, %f1031;
add.f32 %f1033, %f1030, %f1026;
sub.f32 %f1034, %f1030, %f1033;
add.f32 %f1035, %f1026, %f1034;
add.f32 %f1036, %f1028, %f1035;
add.f32 %f1037, %f1032, %f1036;
add.f32 %f1038, %f1033, %f1037;
sub.f32 %f1039, %f1033, %f1038;
add.f32 %f1040, %f1037, %f1039;
mul.rn.f32 %f1041, %f992, %f1038;
neg.f32 %f1042, %f1041;
fma.rn.f32 %f1043, %f992, %f1038, %f1042;
fma.rn.f32 %f1044, %f992, %f1040, %f1043;
fma.rn.f32 %f1045, %f986, %f1038, %f1044;
add.rn.f32 %f1046, %f1041, %f1045;
neg.f32 %f1047, %f1046;
add.rn.f32 %f1048, %f1041, %f1047;
add.rn.f32 %f1049, %f1048, %f1045;
mov.b32 %r345, %f1046;
setp.eq.s32 %p121, %r345, 1118925336;
add.s32 %r346, %r345, -1;
mov.b32 %f1050, %r346;
add.f32 %f1051, %f1049, 0f37000000;
selp.f32 %f186, %f1051, %f1049, %p121;
selp.f32 %f1052, %f1050, %f1046, %p121;
mov.f32 %f1053, 0f3FB8AA3B;
mul.rn.f32 %f1054, %f1052, %f1053;
cvt.rzi.f32.f32 %f1055, %f1054;
abs.f32 %f1056, %f1055;
setp.gt.f32 %p122, %f1056, 0f42FC0000;
mov.b32 %r347, %f1055;
and.b32 %r348, %r347, -2147483648;
or.b32 %r349, %r348, 1123811328;
mov.b32 %f1057, %r349;
selp.f32 %f1058, %f1057, %f1055, %p122;
mov.f32 %f1059, 0fBF317218;
fma.rn.f32 %f1060, %f1058, %f1059, %f1052;
mov.f32 %f1061, 0f3102E308;
fma.rn.f32 %f1062, %f1058, %f1061, %f1060;
mul.f32 %f1063, %f1062, 0f3FB8AA3B;
add.f32 %f1064, %f1058, 0f4B40007F;
mov.b32 %r350, %f1064;
shl.b32 %r351, %r350, 23;
mov.b32 %f1065, %r351;
ex2.approx.ftz.f32 %f1066, %f1063;
mul.f32 %f187, %f1066, %f1065;
setp.eq.f32 %p123, %f187, 0f7F800000;
mov.f32 %f1360, 0f7F800000;
@%p123 bra $L__BB0_100;
fma.rn.f32 %f1360, %f187, %f186, %f187;
$L__BB0_100:
setp.lt.f32 %p124, %f181, 0f00000000;
setp.eq.f32 %p125, %f184, 0f3F800000;
and.pred %p7, %p124, %p125;
setp.eq.f32 %p126, %f181, 0f00000000;
@%p126 bra $L__BB0_104;
bra.uni $L__BB0_101;
$L__BB0_104:
add.f32 %f1071, %f181, %f181;
selp.f32 %f1362, %f1071, 0f00000000, %p125;
bra.uni $L__BB0_105;
$L__BB0_101:
mov.b32 %r352, %f1360;
xor.b32 %r353, %r352, -2147483648;
mov.b32 %f1067, %r353;
selp.f32 %f1362, %f1067, %f1360, %p7;
setp.geu.f32 %p127, %f181, 0f00000000;
@%p127 bra $L__BB0_105;
mov.f32 %f1068, 0f3ED55555;
cvt.rzi.f32.f32 %f1069, %f1068;
setp.eq.f32 %p128, %f1069, 0f3ED55555;
@%p128 bra $L__BB0_105;
mov.f32 %f1362, 0f7FFFFFFF;
$L__BB0_105:
add.f32 %f1072, %f185, 0f3ED55555;
mov.b32 %r354, %f1072;
setp.lt.s32 %p130, %r354, 2139095040;
@%p130 bra $L__BB0_110;
setp.gtu.f32 %p131, %f185, 0f7F800000;
@%p131 bra $L__BB0_109;
bra.uni $L__BB0_107;
$L__BB0_109:
add.f32 %f1362, %f181, 0f3ED55555;
bra.uni $L__BB0_110;
$L__BB0_107:
setp.neu.f32 %p132, %f185, 0f7F800000;
@%p132 bra $L__BB0_110;
selp.f32 %f1362, 0fFF800000, 0f7F800000, %p7;
$L__BB0_110:
abs.f32 %f196, %f182;
setp.lt.f32 %p133, %f196, 0f00800000;
mul.f32 %f1074, %f196, 0f4B800000;
selp.f32 %f1075, %f1074, %f196, %p133;
selp.f32 %f1076, 0fC3170000, 0fC2FE0000, %p133;
mov.b32 %r355, %f1075;
and.b32 %r356, %r355, 8388607;
or.b32 %r357, %r356, 1065353216;
mov.b32 %f1077, %r357;
shr.u32 %r358, %r355, 23;
cvt.rn.f32.u32 %f1078, %r358;
add.f32 %f1079, %f1076, %f1078;
setp.gt.f32 %p134, %f1077, 0f3FB504F3;
mul.f32 %f1080, %f1077, 0f3F000000;
add.f32 %f1081, %f1079, 0f3F800000;
selp.f32 %f1082, %f1081, %f1079, %p134;
selp.f32 %f1083, %f1080, %f1077, %p134;
add.f32 %f1084, %f1083, 0fBF800000;
add.f32 %f1085, %f1083, 0f3F800000;
rcp.approx.ftz.f32 %f1086, %f1085;
add.f32 %f1087, %f1084, %f1084;
mul.f32 %f1088, %f1087, %f1086;
mul.f32 %f1089, %f1088, %f1088;
mov.f32 %f1090, 0f3C4CAF63;
mov.f32 %f1091, 0f3B18F0FE;
fma.rn.f32 %f1092, %f1091, %f1089, %f1090;
mov.f32 %f1093, 0f3DAAAABD;
fma.rn.f32 %f1094, %f1092, %f1089, %f1093;
mul.rn.f32 %f1095, %f1094, %f1089;
mul.rn.f32 %f1096, %f1095, %f1088;
sub.f32 %f1097, %f1084, %f1088;
add.f32 %f1098, %f1097, %f1097;
neg.f32 %f1099, %f1088;
fma.rn.f32 %f1100, %f1099, %f1084, %f1098;
mul.rn.f32 %f1101, %f1086, %f1100;
add.f32 %f1102, %f1096, %f1088;
sub.f32 %f1103, %f1088, %f1102;
add.f32 %f1104, %f1096, %f1103;
add.f32 %f1105, %f1101, %f1104;
add.f32 %f1106, %f1102, %f1105;
sub.f32 %f1107, %f1102, %f1106;
add.f32 %f1108, %f1105, %f1107;
mov.f32 %f1109, 0f3F317200;
mul.rn.f32 %f1110, %f1082, %f1109;
mov.f32 %f1111, 0f35BFBE8E;
mul.rn.f32 %f1112, %f1082, %f1111;
add.f32 %f1113, %f1110, %f1106;
sub.f32 %f1114, %f1110, %f1113;
add.f32 %f1115, %f1106, %f1114;
add.f32 %f1116, %f1108, %f1115;
add.f32 %f1117, %f1112, %f1116;
add.f32 %f1118, %f1113, %f1117;
sub.f32 %f1119, %f1113, %f1118;
add.f32 %f1120, %f1117, %f1119;
mov.f32 %f1121, 0f3ED55555;
mul.rn.f32 %f1122, %f1121, %f1118;
neg.f32 %f1123, %f1122;
fma.rn.f32 %f1124, %f1121, %f1118, %f1123;
fma.rn.f32 %f1125, %f1121, %f1120, %f1124;
mov.f32 %f1126, 0f00000000;
fma.rn.f32 %f1127, %f1126, %f1118, %f1125;
add.rn.f32 %f1128, %f1122, %f1127;
neg.f32 %f1129, %f1128;
add.rn.f32 %f1130, %f1122, %f1129;
add.rn.f32 %f1131, %f1130, %f1127;
mov.b32 %r359, %f1128;
setp.eq.s32 %p135, %r359, 1118925336;
add.s32 %r360, %r359, -1;
mov.b32 %f1132, %r360;
add.f32 %f1133, %f1131, 0f37000000;
selp.f32 %f197, %f1133, %f1131, %p135;
selp.f32 %f1134, %f1132, %f1128, %p135;
mov.f32 %f1135, 0f3FB8AA3B;
mul.rn.f32 %f1136, %f1134, %f1135;
cvt.rzi.f32.f32 %f1137, %f1136;
abs.f32 %f1138, %f1137;
setp.gt.f32 %p136, %f1138, 0f42FC0000;
mov.b32 %r361, %f1137;
and.b32 %r362, %r361, -2147483648;
or.b32 %r363, %r362, 1123811328;
mov.b32 %f1139, %r363;
selp.f32 %f1140, %f1139, %f1137, %p136;
mov.f32 %f1141, 0fBF317218;
fma.rn.f32 %f1142, %f1140, %f1141, %f1134;
mov.f32 %f1143, 0f3102E308;
fma.rn.f32 %f1144, %f1140, %f1143, %f1142;
mul.f32 %f1145, %f1144, 0f3FB8AA3B;
add.f32 %f1146, %f1140, 0f4B40007F;
mov.b32 %r364, %f1146;
shl.b32 %r365, %r364, 23;
mov.b32 %f1147, %r365;
ex2.approx.ftz.f32 %f1148, %f1145;
mul.f32 %f198, %f1148, %f1147;
setp.eq.f32 %p137, %f198, 0f7F800000;
mov.f32 %f1363, 0f7F800000;
@%p137 bra $L__BB0_112;
fma.rn.f32 %f1363, %f198, %f197, %f198;
$L__BB0_112:
setp.lt.f32 %p138, %f182, 0f00000000;
and.pred %p8, %p138, %p125;
setp.eq.f32 %p140, %f182, 0f00000000;
@%p140 bra $L__BB0_116;
bra.uni $L__BB0_113;
$L__BB0_116:
add.f32 %f1153, %f182, %f182;
selp.f32 %f1365, %f1153, 0f00000000, %p125;
bra.uni $L__BB0_117;
$L__BB0_113:
mov.b32 %r366, %f1363;
xor.b32 %r367, %r366, -2147483648;
mov.b32 %f1149, %r367;
selp.f32 %f1365, %f1149, %f1363, %p8;
setp.geu.f32 %p141, %f182, 0f00000000;
@%p141 bra $L__BB0_117;
mov.f32 %f1150, 0f3ED55555;
cvt.rzi.f32.f32 %f1151, %f1150;
setp.eq.f32 %p142, %f1151, 0f3ED55555;
@%p142 bra $L__BB0_117;
mov.f32 %f1365, 0f7FFFFFFF;
$L__BB0_117:
add.f32 %f1154, %f196, 0f3ED55555;
mov.b32 %r368, %f1154;
setp.lt.s32 %p144, %r368, 2139095040;
@%p144 bra $L__BB0_122;
setp.gtu.f32 %p145, %f196, 0f7F800000;
@%p145 bra $L__BB0_121;
bra.uni $L__BB0_119;
$L__BB0_121:
add.f32 %f1365, %f182, 0f3ED55555;
bra.uni $L__BB0_122;
$L__BB0_119:
setp.neu.f32 %p146, %f196, 0f7F800000;
@%p146 bra $L__BB0_122;
selp.f32 %f1365, 0fFF800000, 0f7F800000, %p8;
$L__BB0_122:
abs.f32 %f207, %f183;
setp.lt.f32 %p147, %f207, 0f00800000;
mul.f32 %f1156, %f207, 0f4B800000;
selp.f32 %f1157, %f1156, %f207, %p147;
selp.f32 %f1158, 0fC3170000, 0fC2FE0000, %p147;
mov.b32 %r369, %f1157;
and.b32 %r370, %r369, 8388607;
or.b32 %r371, %r370, 1065353216;
mov.b32 %f1159, %r371;
shr.u32 %r372, %r369, 23;
cvt.rn.f32.u32 %f1160, %r372;
add.f32 %f1161, %f1158, %f1160;
setp.gt.f32 %p148, %f1159, 0f3FB504F3;
mul.f32 %f1162, %f1159, 0f3F000000;
add.f32 %f1163, %f1161, 0f3F800000;
selp.f32 %f1164, %f1163, %f1161, %p148;
selp.f32 %f1165, %f1162, %f1159, %p148;
add.f32 %f1166, %f1165, 0fBF800000;
add.f32 %f1167, %f1165, 0f3F800000;
rcp.approx.ftz.f32 %f1168, %f1167;
add.f32 %f1169, %f1166, %f1166;
mul.f32 %f1170, %f1169, %f1168;
mul.f32 %f1171, %f1170, %f1170;
mov.f32 %f1172, 0f3C4CAF63;
mov.f32 %f1173, 0f3B18F0FE;
fma.rn.f32 %f1174, %f1173, %f1171, %f1172;
mov.f32 %f1175, 0f3DAAAABD;
fma.rn.f32 %f1176, %f1174, %f1171, %f1175;
mul.rn.f32 %f1177, %f1176, %f1171;
mul.rn.f32 %f1178, %f1177, %f1170;
sub.f32 %f1179, %f1166, %f1170;
add.f32 %f1180, %f1179, %f1179;
neg.f32 %f1181, %f1170;
fma.rn.f32 %f1182, %f1181, %f1166, %f1180;
mul.rn.f32 %f1183, %f1168, %f1182;
add.f32 %f1184, %f1178, %f1170;
sub.f32 %f1185, %f1170, %f1184;
add.f32 %f1186, %f1178, %f1185;
add.f32 %f1187, %f1183, %f1186;
add.f32 %f1188, %f1184, %f1187;
sub.f32 %f1189, %f1184, %f1188;
add.f32 %f1190, %f1187, %f1189;
mov.f32 %f1191, 0f3F317200;
mul.rn.f32 %f1192, %f1164, %f1191;
mov.f32 %f1193, 0f35BFBE8E;
mul.rn.f32 %f1194, %f1164, %f1193;
add.f32 %f1195, %f1192, %f1188;
sub.f32 %f1196, %f1192, %f1195;
add.f32 %f1197, %f1188, %f1196;
add.f32 %f1198, %f1190, %f1197;
add.f32 %f1199, %f1194, %f1198;
add.f32 %f1200, %f1195, %f1199;
sub.f32 %f1201, %f1195, %f1200;
add.f32 %f1202, %f1199, %f1201;
mov.f32 %f1203, 0f3ED55555;
mul.rn.f32 %f1204, %f1203, %f1200;
neg.f32 %f1205, %f1204;
fma.rn.f32 %f1206, %f1203, %f1200, %f1205;
fma.rn.f32 %f1207, %f1203, %f1202, %f1206;
mov.f32 %f1208, 0f00000000;
fma.rn.f32 %f1209, %f1208, %f1200, %f1207;
add.rn.f32 %f1210, %f1204, %f1209;
neg.f32 %f1211, %f1210;
add.rn.f32 %f1212, %f1204, %f1211;
add.rn.f32 %f1213, %f1212, %f1209;
mov.b32 %r373, %f1210;
setp.eq.s32 %p149, %r373, 1118925336;
add.s32 %r374, %r373, -1;
mov.b32 %f1214, %r374;
add.f32 %f1215, %f1213, 0f37000000;
selp.f32 %f208, %f1215, %f1213, %p149;
selp.f32 %f1216, %f1214, %f1210, %p149;
mov.f32 %f1217, 0f3FB8AA3B;
mul.rn.f32 %f1218, %f1216, %f1217;
cvt.rzi.f32.f32 %f1219, %f1218;
abs.f32 %f1220, %f1219;
setp.gt.f32 %p150, %f1220, 0f42FC0000;
mov.b32 %r375, %f1219;
and.b32 %r376, %r375, -2147483648;
or.b32 %r377, %r376, 1123811328;
mov.b32 %f1221, %r377;
selp.f32 %f1222, %f1221, %f1219, %p150;
mov.f32 %f1223, 0fBF317218;
fma.rn.f32 %f1224, %f1222, %f1223, %f1216;
mov.f32 %f1225, 0f3102E308;
fma.rn.f32 %f1226, %f1222, %f1225, %f1224;
mul.f32 %f1227, %f1226, 0f3FB8AA3B;
add.f32 %f1228, %f1222, 0f4B40007F;
mov.b32 %r378, %f1228;
shl.b32 %r379, %r378, 23;
mov.b32 %f1229, %r379;
ex2.approx.ftz.f32 %f1230, %f1227;
mul.f32 %f209, %f1230, %f1229;
setp.eq.f32 %p151, %f209, 0f7F800000;
mov.f32 %f1366, 0f7F800000;
@%p151 bra $L__BB0_124;
fma.rn.f32 %f1366, %f209, %f208, %f209;
$L__BB0_124:
setp.lt.f32 %p152, %f183, 0f00000000;
and.pred %p9, %p152, %p125;
setp.eq.f32 %p154, %f183, 0f00000000;
@%p154 bra $L__BB0_128;
bra.uni $L__BB0_125;
$L__BB0_128:
add.f32 %f1235, %f183, %f183;
selp.f32 %f1368, %f1235, 0f00000000, %p125;
bra.uni $L__BB0_129;
$L__BB0_125:
mov.b32 %r380, %f1366;
xor.b32 %r381, %r380, -2147483648;
mov.b32 %f1231, %r381;
selp.f32 %f1368, %f1231, %f1366, %p9;
setp.geu.f32 %p155, %f183, 0f00000000;
@%p155 bra $L__BB0_129;
mov.f32 %f1232, 0f3ED55555;
cvt.rzi.f32.f32 %f1233, %f1232;
setp.eq.f32 %p156, %f1233, 0f3ED55555;
@%p156 bra $L__BB0_129;
mov.f32 %f1368, 0f7FFFFFFF;
$L__BB0_129:
add.f32 %f1236, %f207, 0f3ED55555;
mov.b32 %r382, %f1236;
setp.lt.s32 %p158, %r382, 2139095040;
@%p158 bra $L__BB0_134;
setp.gtu.f32 %p159, %f207, 0f7F800000;
@%p159 bra $L__BB0_133;
bra.uni $L__BB0_131;
$L__BB0_133:
add.f32 %f1368, %f183, 0f3ED55555;
bra.uni $L__BB0_134;
$L__BB0_131:
setp.neu.f32 %p160, %f207, 0f7F800000;
@%p160 bra $L__BB0_134;
selp.f32 %f1368, 0fFF800000, 0f7F800000, %p9;
$L__BB0_134:
fma.rn.f32 %f1237, %f1362, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p161, %f181, 0f3F800000;
mov.f32 %f1238, 0f3F800000;
selp.f32 %f1239, 0f3F7FFFFF, %f1237, %p161;
mul.f32 %f1240, %f181, 0f414EB852;
setp.lt.f32 %p162, %f181, 0f3B4D2E1C;
selp.f32 %f1241, %f1240, %f1239, %p162;
fma.rn.f32 %f1242, %f1365, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p163, %f182, 0f3F800000;
selp.f32 %f1243, 0f3F7FFFFF, %f1242, %p163;
mul.f32 %f1244, %f182, 0f414EB852;
setp.lt.f32 %p164, %f182, 0f3B4D2E1C;
selp.f32 %f1245, %f1244, %f1243, %p164;
fma.rn.f32 %f1246, %f1368, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p165, %f183, 0f3F800000;
selp.f32 %f1247, 0f3F7FFFFF, %f1246, %p165;
mul.f32 %f1248, %f183, 0f414EB852;
setp.lt.f32 %p166, %f183, 0f3B4D2E1C;
selp.f32 %f1249, %f1248, %f1247, %p166;
min.f32 %f1250, %f1241, %f1238;
mov.f32 %f1251, 0f00000000;
max.f32 %f1252, %f1251, %f1250;
mul.f32 %f1253, %f1252, 0f43800000;
cvt.rzi.u32.f32 %r383, %f1253;
min.u32 %r384, %r383, 255;
min.f32 %f1254, %f1245, %f1238;
max.f32 %f1255, %f1251, %f1254;
mul.f32 %f1256, %f1255, 0f43800000;
cvt.rzi.u32.f32 %r385, %f1256;
min.u32 %r386, %r385, 255;
min.f32 %f1257, %f1249, %f1238;
max.f32 %f1258, %f1251, %f1257;
mul.f32 %f1259, %f1258, 0f43800000;
cvt.rzi.u32.f32 %r387, %f1259;
min.u32 %r388, %r387, 255;
shl.b64 %rd29, %rd10, 2;
add.s64 %rd30, %rd9, %rd29;
cvt.u16.u32 %rs20, %r388;
cvt.u16.u32 %rs21, %r386;
cvt.u16.u32 %rs22, %r384;
mov.u16 %rs23, 255;
st.global.v4.u8 [%rd30], {%rs22, %rs21, %rs20, %rs23};
$L__BB0_135:
and.b32 %r389, %r417, 4;
setp.eq.s32 %p167, %r389, 0;
@%p167 bra $L__BB0_139;
ld.const.u32 %r390, [params+108];
setp.eq.s32 %p168, %r390, 0;
ld.const.u64 %rd31, [params+224];
cvta.to.global.u64 %rd32, %rd31;
ld.const.u32 %r391, [params+216];
mad.lo.s32 %r392, %r391, %r4, %r3;
mul.wide.u32 %rd33, %r392, 8;
add.s64 %rd11, %rd32, %rd33;
@%p168 bra $L__BB0_138;
ld.global.v4.u16 {%rs31, %rs32, %rs33, %rs34}, [%rd11];
// begin inline asm
{ cvt.f32.f16 %f1260, %rs31;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1261, %rs32;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1262, %rs33;}
// end inline asm
add.f32 %f1263, %f1334, %f1260;
add.f32 %f1264, %f1335, %f1261;
add.f32 %f1265, %f1336, %f1262;
mov.f32 %f1266, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f1265;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1264;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1263;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f1266;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs27, %rs28, %rs29, %rs30};
bra.uni $L__BB0_139;
$L__BB0_138:
mov.f32 %f1270, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f1270;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f1336;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f1335;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f1334;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs35, %rs36, %rs37, %rs38};
$L__BB0_139:
and.b32 %r393, %r417, 64;
setp.eq.s32 %p169, %r393, 0;
@%p169 bra $L__BB0_151;
mul.f32 %f1271, %f1338, %f1338;
fma.rn.f32 %f1272, %f1337, %f1337, %f1271;
fma.rn.f32 %f1273, %f1339, %f1339, %f1272;
sqrt.rn.f32 %f1274, %f1273;
rcp.rn.f32 %f1275, %f1274;
mul.f32 %f1276, %f1337, %f1275;
mul.f32 %f1277, %f1338, %f1275;
mul.f32 %f1278, %f1339, %f1275;
ld.const.u64 %rd34, [params+208];
cvta.to.global.u64 %rd35, %rd34;
ld.const.u32 %r394, [params+200];
mad.lo.s32 %r395, %r394, %r4, %r3;
fma.rn.f32 %f1279, %f1276, 0f3F000000, 0f3F000000;
mul.f32 %f1280, %f1279, 0f437F0000;
cvt.rzi.u32.f32 %r396, %f1280;
fma.rn.f32 %f1281, %f1277, 0f3F000000, 0f3F000000;
mul.f32 %f1282, %f1281, 0f437F0000;
cvt.rzi.u32.f32 %r397, %f1282;
fma.rn.f32 %f1283, %f1278, 0f3F000000, 0f3F000000;
mul.f32 %f1284, %f1283, 0f437F0000;
cvt.rzi.u32.f32 %r398, %f1284;
mul.wide.u32 %rd36, %r395, 4;
add.s64 %rd37, %rd35, %rd36;
cvt.u16.u32 %rs39, %r398;
cvt.u16.u32 %rs40, %r397;
cvt.u16.u32 %rs41, %r396;
mov.u16 %rs42, 255;
st.global.v4.u8 [%rd37], {%rs41, %rs40, %rs39, %rs42};
$L__BB0_151:
ret;
}