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

2598 lines
72 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<183>;
.reg .b16 %rs<180>;
.reg .f32 %f<1584>;
.reg .b32 %r<430>;
.reg .b64 %rd<78>;
// begin inline asm
call (%r26), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r27), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u64 %rd21, [params+400];
cvta.to.global.u64 %rd22, %rd21;
ld.const.u32 %r32, [params+392];
mad.lo.s32 %r33, %r32, %r27, %r26;
mul.wide.u32 %rd23, %r33, 4;
add.s64 %rd1, %rd22, %rd23;
ld.global.v2.u8 {%rs9, %rs178}, [%rd1];
or.b16 %rs11, %rs9, %rs178;
and.b16 %rs12, %rs11, 255;
setp.eq.s16 %p10, %rs12, 0;
@%p10 bra $L__BB0_2;
ld.global.u8 %rs177, [%rd1+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs177, [%rd1+2];
setp.eq.s16 %p11, %rs177, 0;
mov.f32 %f1480, 0f00000000;
mov.u16 %rs178, 0;
mov.f32 %f1481, %f1480;
mov.f32 %f1482, %f1480;
@%p11 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f294, %rs9;
div.rn.f32 %f295, %f294, 0f437F0000;
fma.rn.f32 %f296, %f295, 0f40000000, 0fBF800000;
and.b16 %rs15, %rs178, 255;
cvt.rn.f32.u16 %f297, %rs15;
div.rn.f32 %f298, %f297, 0f437F0000;
fma.rn.f32 %f299, %f298, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f300, %rs177;
div.rn.f32 %f301, %f300, 0f437F0000;
fma.rn.f32 %f302, %f301, 0f40000000, 0fBF800000;
mul.f32 %f303, %f299, %f299;
fma.rn.f32 %f304, %f296, %f296, %f303;
fma.rn.f32 %f305, %f302, %f302, %f304;
sqrt.rn.f32 %f306, %f305;
rcp.rn.f32 %f307, %f306;
mul.f32 %f1482, %f307, %f302;
mul.f32 %f1481, %f307, %f299;
mul.f32 %f1480, %f296, %f307;
$L__BB0_4:
ld.const.v2.u32 {%r34, %r35}, [params];
add.s32 %r3, %r34, %r26;
add.s32 %r4, %r35, %r27;
setp.eq.f32 %p12, %f1480, 0f00000000;
setp.eq.f32 %p13, %f1481, 0f00000000;
and.pred %p14, %p12, %p13;
setp.eq.f32 %p15, %f1482, 0f00000000;
and.pred %p16, %p15, %p14;
@%p16 bra $L__BB0_149;
bra.uni $L__BB0_5;
$L__BB0_149:
ld.const.u32 %r23, [params+104];
and.b32 %r404, %r23, 1;
setp.eq.b32 %p171, %r404, 1;
mov.pred %p172, 0;
xor.pred %p173, %p171, %p172;
not.pred %p174, %p173;
@%p174 bra $L__BB0_151;
ld.const.u64 %rd54, [params+144];
cvta.to.global.u64 %rd55, %rd54;
ld.const.u32 %r405, [params+136];
mad.lo.s32 %r406, %r405, %r4, %r3;
mul.wide.u32 %rd56, %r406, 4;
add.s64 %rd57, %rd55, %rd56;
mov.u16 %rs99, 0;
st.global.v4.u8 [%rd57], {%rs99, %rs99, %rs99, %rs99};
$L__BB0_151:
and.b32 %r407, %r23, 8;
setp.eq.s32 %p175, %r407, 0;
@%p175 bra $L__BB0_153;
ld.const.u64 %rd58, [params+192];
cvta.to.global.u64 %rd59, %rd58;
ld.const.u32 %r408, [params+184];
mad.lo.s32 %r409, %r408, %r4, %r3;
mov.f32 %f1431, 0f00000000;
cvt.rzi.u32.f32 %r410, %f1431;
mul.wide.u32 %rd60, %r409, 2;
add.s64 %rd61, %rd59, %rd60;
mov.u16 %rs100, 0;
cvt.u16.u32 %rs101, %r410;
st.global.v2.u8 [%rd61], {%rs101, %rs100};
$L__BB0_153:
and.b32 %r411, %r23, 4;
setp.eq.s32 %p176, %r411, 0;
ld.const.u32 %r429, [params+108];
@%p176 bra $L__BB0_157;
setp.eq.s32 %p177, %r429, 0;
ld.const.u64 %rd62, [params+224];
cvta.to.global.u64 %rd63, %rd62;
ld.const.u32 %r412, [params+216];
mad.lo.s32 %r413, %r412, %r4, %r3;
mul.wide.u32 %rd64, %r413, 8;
add.s64 %rd16, %rd63, %rd64;
@%p177 bra $L__BB0_156;
ld.global.v4.u16 {%rs108, %rs109, %rs110, %rs111}, [%rd16];
// begin inline asm
{ cvt.f32.f16 %f1432, %rs108;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1433, %rs109;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1434, %rs110;}
// end inline asm
add.f32 %f1435, %f1432, 0f00000000;
add.f32 %f1436, %f1433, 0f00000000;
add.f32 %f1437, %f1434, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs107, %f1437;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs106, %f1436;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs105, %f1435;}
// end inline asm
mov.u16 %rs112, 0;
st.global.v4.u16 [%rd16], {%rs105, %rs106, %rs107, %rs112};
bra.uni $L__BB0_157;
$L__BB0_5:
ld.const.u64 %rd24, [params+432];
cvta.to.global.u64 %rd25, %rd24;
ld.const.u32 %r38, [params+424];
mad.lo.s32 %r39, %r38, %r27, %r26;
mul.wide.u32 %rd26, %r39, 12;
add.s64 %rd27, %rd25, %rd26;
ld.global.f32 %f10, [%rd27];
mul.f32 %f308, %f10, 0f3456BF95;
ld.global.f32 %f11, [%rd27+4];
mul.f32 %f309, %f11, 0f3456BF95;
ld.global.f32 %f12, [%rd27+8];
mul.f32 %f310, %f12, 0f3456BF95;
abs.f32 %f311, %f1480;
div.rn.f32 %f312, %f308, %f311;
abs.f32 %f313, %f1481;
div.rn.f32 %f314, %f309, %f313;
abs.f32 %f315, %f1482;
div.rn.f32 %f316, %f310, %f315;
abs.f32 %f317, %f312;
abs.f32 %f318, %f314;
abs.f32 %f319, %f316;
mov.f32 %f320, 0f38D1B717;
max.f32 %f321, %f317, %f320;
max.f32 %f322, %f318, %f320;
max.f32 %f323, %f319, %f320;
fma.rn.f32 %f13, %f1480, %f321, %f10;
fma.rn.f32 %f14, %f1481, %f322, %f11;
fma.rn.f32 %f15, %f1482, %f323, %f12;
ld.const.u32 %r5, [params+600];
setp.eq.s32 %p17, %r5, 0;
@%p17 bra $L__BB0_44;
ld.const.u64 %rd28, [params+608];
cvta.to.global.u64 %rd2, %rd28;
mov.f32 %f340, 0f40000000;
cvt.rzi.f32.f32 %f341, %f340;
add.f32 %f342, %f341, %f341;
mov.f32 %f343, 0f40800000;
sub.f32 %f344, %f343, %f342;
abs.f32 %f16, %f344;
ld.const.u32 %r6, [params+544];
ld.const.u32 %r427, [params+104];
mul.f32 %f17, %f13, 0f3456BF95;
mul.f32 %f18, %f14, 0f3456BF95;
mul.f32 %f19, %f15, 0f3456BF95;
mov.f32 %f339, 0f00000000;
mov.u32 %r423, 0;
ld.const.u64 %rd3, [params+96];
abs.f32 %f543, %f17;
abs.f32 %f544, %f18;
max.f32 %f545, %f543, %f544;
abs.f32 %f546, %f19;
max.f32 %f547, %f545, %f546;
mov.f32 %f1546, %f339;
mov.f32 %f1545, %f339;
mov.f32 %f1544, %f339;
mov.f32 %f1543, %f339;
mov.f32 %f1542, %f339;
mov.f32 %f1541, %f339;
mov.f32 %f1540, %f339;
mov.f32 %f1539, %f339;
mov.f32 %f1538, %f339;
mov.f32 %f1537, %f339;
mov.f32 %f1536, %f339;
mov.f32 %f1535, %f339;
mov.f32 %f1534, %f339;
mov.f32 %f1533, %f339;
mov.f32 %f1532, %f339;
mov.f32 %f1531, %f339;
$L__BB0_7:
mul.wide.u32 %rd29, %r423, 112;
add.s64 %rd5, %rd2, %rd29;
ld.global.v4.f32 {%f346, %f347, %f348, %f349}, [%rd5];
ld.global.v4.f32 {%f350, %f351, %f352, %f353}, [%rd5+16];
ld.global.v4.f32 {%f356, %f1507, %f1508, %f359}, [%rd5+32];
ld.global.v4.f32 {%f360, %f361, %f362, %f1505}, [%rd5+48];
ld.global.v4.f32 {%f364, %f365, %f366, %f367}, [%rd5+64];
ld.global.v4.u32 {%r41, %r42, %r43, %r44}, [%rd5+80];
mov.b32 %f55, %r43;
mov.b32 %f54, %r42;
mov.b32 %f53, %r41;
ld.global.u64 %rd6, [%rd5+96];
sub.f32 %f368, %f347, %f10;
sub.f32 %f369, %f348, %f11;
sub.f32 %f370, %f349, %f12;
mul.f32 %f371, %f369, %f369;
fma.rn.f32 %f372, %f368, %f368, %f371;
fma.rn.f32 %f373, %f370, %f370, %f372;
sqrt.rn.f32 %f374, %f373;
rcp.rn.f32 %f375, %f374;
mul.f32 %f57, %f368, %f375;
mul.f32 %f58, %f369, %f375;
mul.f32 %f59, %f370, %f375;
mul.f32 %f60, %f374, %f351;
mul.f32 %f61, %f374, %f353;
abs.f32 %f62, %f61;
setp.lt.f32 %p18, %f62, 0f00800000;
mul.f32 %f376, %f62, 0f4B800000;
selp.f32 %f377, %f376, %f62, %p18;
selp.f32 %f378, 0fC3170000, 0fC2FE0000, %p18;
mov.b32 %r48, %f377;
and.b32 %r49, %r48, 8388607;
or.b32 %r50, %r49, 1065353216;
mov.b32 %f379, %r50;
shr.u32 %r51, %r48, 23;
cvt.rn.f32.u32 %f380, %r51;
add.f32 %f381, %f378, %f380;
setp.gt.f32 %p19, %f379, 0f3FB504F3;
mul.f32 %f382, %f379, 0f3F000000;
add.f32 %f383, %f381, 0f3F800000;
selp.f32 %f384, %f383, %f381, %p19;
selp.f32 %f385, %f382, %f379, %p19;
add.f32 %f386, %f385, 0fBF800000;
add.f32 %f387, %f385, 0f3F800000;
rcp.approx.ftz.f32 %f388, %f387;
add.f32 %f389, %f386, %f386;
mul.f32 %f390, %f389, %f388;
mul.f32 %f391, %f390, %f390;
mov.f32 %f392, 0f3C4CAF63;
mov.f32 %f393, 0f3B18F0FE;
fma.rn.f32 %f394, %f393, %f391, %f392;
mov.f32 %f395, 0f3DAAAABD;
fma.rn.f32 %f396, %f394, %f391, %f395;
mul.rn.f32 %f397, %f396, %f391;
mul.rn.f32 %f398, %f397, %f390;
sub.f32 %f399, %f386, %f390;
add.f32 %f400, %f399, %f399;
neg.f32 %f401, %f390;
fma.rn.f32 %f402, %f401, %f386, %f400;
mul.rn.f32 %f403, %f388, %f402;
add.f32 %f404, %f398, %f390;
sub.f32 %f405, %f390, %f404;
add.f32 %f406, %f398, %f405;
add.f32 %f407, %f403, %f406;
add.f32 %f408, %f404, %f407;
sub.f32 %f409, %f404, %f408;
add.f32 %f410, %f407, %f409;
mov.f32 %f411, 0f3F317200;
mul.rn.f32 %f412, %f384, %f411;
mov.f32 %f413, 0f35BFBE8E;
mul.rn.f32 %f414, %f384, %f413;
add.f32 %f415, %f412, %f408;
sub.f32 %f416, %f412, %f415;
add.f32 %f417, %f408, %f416;
add.f32 %f418, %f410, %f417;
add.f32 %f419, %f414, %f418;
add.f32 %f420, %f415, %f419;
sub.f32 %f421, %f415, %f420;
add.f32 %f422, %f419, %f421;
mul.rn.f32 %f424, %f343, %f420;
neg.f32 %f425, %f424;
fma.rn.f32 %f426, %f343, %f420, %f425;
fma.rn.f32 %f427, %f343, %f422, %f426;
fma.rn.f32 %f429, %f339, %f420, %f427;
add.rn.f32 %f430, %f424, %f429;
neg.f32 %f431, %f430;
add.rn.f32 %f432, %f424, %f431;
add.rn.f32 %f433, %f432, %f429;
mov.b32 %r52, %f430;
setp.eq.s32 %p20, %r52, 1118925336;
add.s32 %r53, %r52, -1;
mov.b32 %f434, %r53;
add.f32 %f435, %f433, 0f37000000;
selp.f32 %f63, %f435, %f433, %p20;
selp.f32 %f436, %f434, %f430, %p20;
mov.f32 %f437, 0f3FB8AA3B;
mul.rn.f32 %f438, %f436, %f437;
cvt.rzi.f32.f32 %f439, %f438;
abs.f32 %f440, %f439;
setp.gt.f32 %p21, %f440, 0f42FC0000;
mov.b32 %r54, %f439;
and.b32 %r55, %r54, -2147483648;
or.b32 %r56, %r55, 1123811328;
mov.b32 %f441, %r56;
selp.f32 %f442, %f441, %f439, %p21;
mov.f32 %f443, 0fBF317218;
fma.rn.f32 %f444, %f442, %f443, %f436;
mov.f32 %f445, 0f3102E308;
fma.rn.f32 %f446, %f442, %f445, %f444;
mul.f32 %f447, %f446, 0f3FB8AA3B;
add.f32 %f448, %f442, 0f4B40007F;
mov.b32 %r57, %f448;
shl.b32 %r58, %r57, 23;
mov.b32 %f449, %r58;
ex2.approx.ftz.f32 %f450, %f447;
mul.f32 %f64, %f450, %f449;
setp.eq.f32 %p22, %f64, 0f7F800000;
mov.f32 %f1499, 0f7F800000;
@%p22 bra $L__BB0_9;
fma.rn.f32 %f1499, %f64, %f63, %f64;
$L__BB0_9:
setp.lt.f32 %p23, %f61, 0f00000000;
setp.eq.f32 %p24, %f16, 0f3F800000;
and.pred %p1, %p23, %p24;
setp.eq.f32 %p25, %f61, 0f00000000;
@%p25 bra $L__BB0_13;
bra.uni $L__BB0_10;
$L__BB0_13:
add.f32 %f455, %f61, %f61;
selp.f32 %f1501, %f455, 0f00000000, %p24;
bra.uni $L__BB0_14;
$L__BB0_10:
mov.b32 %r59, %f1499;
xor.b32 %r60, %r59, -2147483648;
mov.b32 %f451, %r60;
selp.f32 %f1501, %f451, %f1499, %p1;
setp.geu.f32 %p26, %f61, 0f00000000;
@%p26 bra $L__BB0_14;
mov.f32 %f452, 0f40800000;
cvt.rzi.f32.f32 %f453, %f452;
setp.eq.f32 %p27, %f453, 0f40800000;
@%p27 bra $L__BB0_14;
mov.f32 %f1501, 0f7FFFFFFF;
$L__BB0_14:
abs.f32 %f1477, %f61;
add.f32 %f456, %f1477, 0f40800000;
mov.b32 %r61, %f456;
setp.lt.s32 %p29, %r61, 2139095040;
@%p29 bra $L__BB0_19;
abs.f32 %f1478, %f61;
setp.gtu.f32 %p30, %f1478, 0f7F800000;
@%p30 bra $L__BB0_18;
bra.uni $L__BB0_16;
$L__BB0_18:
add.f32 %f1501, %f61, 0f40800000;
bra.uni $L__BB0_19;
$L__BB0_16:
abs.f32 %f1479, %f61;
setp.neu.f32 %p31, %f1479, 0f7F800000;
@%p31 bra $L__BB0_19;
selp.f32 %f1501, 0fFF800000, 0f7F800000, %p1;
$L__BB0_19:
mov.f32 %f457, 0f3F800000;
sub.f32 %f458, %f457, %f1501;
setp.eq.f32 %p32, %f61, 0f3F800000;
selp.f32 %f459, 0f00000000, %f458, %p32;
cvt.sat.f32.f32 %f460, %f459;
fma.rn.f32 %f461, %f60, %f60, %f352;
div.rn.f32 %f73, %f460, %f461;
mul.f32 %f462, %f1481, %f58;
fma.rn.f32 %f463, %f1480, %f57, %f462;
fma.rn.f32 %f74, %f1482, %f59, %f463;
setp.eq.f32 %p33, %f346, 0f3F800000;
@%p33 bra $L__BB0_25;
bra.uni $L__BB0_20;
$L__BB0_25:
setp.eq.s64 %p37, %rd6, 0;
mov.f32 %f1506, %f356;
mov.f32 %f1509, %f73;
@%p37 bra $L__BB0_29;
setp.geu.f32 %p38, %f1505, 0f00000000;
mov.f32 %f1502, %f57;
mov.f32 %f1503, %f58;
mov.f32 %f1504, %f59;
@%p38 bra $L__BB0_28;
mul.f32 %f488, %f54, %f369;
fma.rn.f32 %f489, %f53, %f368, %f488;
fma.rn.f32 %f491, %f55, %f370, %f489;
rcp.rn.f32 %f492, %f491;
mul.f32 %f1502, %f368, %f492;
mul.f32 %f1503, %f369, %f492;
mul.f32 %f1504, %f370, %f492;
neg.f32 %f1505, %f1505;
$L__BB0_28:
mul.f32 %f493, %f55, %f59;
mul.f32 %f494, %f54, %f58;
neg.f32 %f495, %f494;
mul.f32 %f496, %f53, %f57;
sub.f32 %f497, %f495, %f496;
sub.f32 %f498, %f497, %f493;
setp.gt.f32 %p39, %f498, 0f00000000;
selp.f32 %f499, 0f3F800000, 0f00000000, %p39;
mov.f32 %f500, 0f3F800000;
mul.f32 %f501, %f361, %f1503;
mul.f32 %f502, %f365, %f1503;
fma.rn.f32 %f503, %f360, %f1502, %f501;
fma.rn.f32 %f504, %f364, %f1502, %f502;
fma.rn.f32 %f505, %f362, %f1504, %f503;
fma.rn.f32 %f506, %f366, %f1504, %f504;
fma.rn.f32 %f507, %f1505, %f505, 0f3F000000;
sub.f32 %f508, %f500, %f507;
fma.rn.f32 %f509, %f1505, %f506, 0f3F000000;
tex.2d.v4.f32.f32 {%f510, %f511, %f512, %f513}, [%rd6, {%f508, %f509}];
mul.f32 %f514, %f499, %f510;
mul.f32 %f515, %f499, %f511;
mul.f32 %f516, %f499, %f512;
mul.f32 %f1506, %f356, %f514;
mul.f32 %f1507, %f1507, %f515;
mul.f32 %f1508, %f1508, %f516;
mov.f32 %f1509, %f73;
bra.uni $L__BB0_29;
$L__BB0_20:
setp.eq.f32 %p34, %f346, 0f40000000;
@%p34 bra $L__BB0_23;
bra.uni $L__BB0_21;
$L__BB0_23:
setp.eq.s64 %p36, %rd6, 0;
mov.f32 %f1506, %f356;
mov.f32 %f1509, %f73;
@%p36 bra $L__BB0_29;
mul.f32 %f473, %f361, %f58;
mul.f32 %f474, %f365, %f58;
mul.f32 %f475, %f54, %f58;
fma.rn.f32 %f476, %f360, %f57, %f473;
fma.rn.f32 %f477, %f364, %f57, %f474;
fma.rn.f32 %f478, %f53, %f57, %f475;
fma.rn.f32 %f479, %f362, %f59, %f476;
fma.rn.f32 %f480, %f366, %f59, %f477;
fma.rn.f32 %f481, %f55, %f59, %f478;
tex.cube.v4.f32.f32 {%f482, %f483, %f484, %f485}, [%rd6, {%f479, %f480, %f481, %f481}];
mul.f32 %f1506, %f356, %f482;
mul.f32 %f1507, %f1507, %f483;
mul.f32 %f1508, %f1508, %f484;
mov.f32 %f1509, %f73;
bra.uni $L__BB0_29;
$L__BB0_21:
setp.neu.f32 %p35, %f346, 0f40800000;
mov.f32 %f1506, %f356;
mov.f32 %f1509, %f73;
@%p35 bra $L__BB0_29;
mul.f32 %f464, %f53, %f57;
mul.f32 %f465, %f54, %f58;
neg.f32 %f466, %f465;
sub.f32 %f467, %f466, %f464;
mul.f32 %f468, %f55, %f59;
sub.f32 %f469, %f467, %f468;
fma.rn.f32 %f470, %f359, %f469, %f1505;
cvt.sat.f32.f32 %f471, %f470;
mul.f32 %f472, %f471, %f471;
mul.f32 %f1509, %f73, %f472;
mov.f32 %f1506, %f356;
$L__BB0_29:
max.f32 %f532, %f1506, %f1507;
max.f32 %f533, %f532, %f1508;
mul.f32 %f534, %f1509, %f533;
setp.lt.f32 %p41, %f534, 0f3727C5AC;
mov.pred %p182, -1;
mov.f32 %f1510, 0f00000000;
mov.f32 %f1511, %f1510;
mov.f32 %f1512, %f1510;
mov.f32 %f1513, %f1510;
mov.f32 %f1514, %f1510;
mov.f32 %f1515, %f1510;
mov.f32 %f1516, %f1510;
mov.f32 %f1517, %f1510;
mov.f32 %f1518, %f1510;
mov.f32 %f1519, %f1510;
mov.f32 %f1520, %f1510;
mov.f32 %f1521, %f1510;
mov.f32 %f1522, %f1510;
mov.f32 %f1523, %f1510;
mov.f32 %f1524, %f1510;
@%p41 bra $L__BB0_31;
setp.eq.s32 %p43, %r6, 0;
selp.f32 %f535, %f74, 0f3F800000, %p43;
cvt.sat.f32.f32 %f536, %f535;
mul.f32 %f537, %f1509, %f536;
mul.f32 %f1511, %f1506, %f537;
mul.f32 %f1510, %f1507, %f537;
mul.f32 %f1512, %f1508, %f537;
mul.f32 %f538, %f1509, 0f3E800000;
mul.f32 %f1513, %f1506, %f538;
mul.f32 %f1514, %f1507, %f538;
mul.f32 %f1515, %f1508, %f538;
mul.f32 %f1516, %f57, %f1513;
mul.f32 %f1517, %f57, %f1514;
mul.f32 %f1518, %f57, %f1515;
mul.f32 %f1519, %f58, %f1513;
mul.f32 %f1520, %f58, %f1514;
mul.f32 %f1521, %f58, %f1515;
mul.f32 %f1522, %f59, %f1513;
mul.f32 %f1523, %f59, %f1514;
mul.f32 %f1524, %f59, %f1515;
mov.pred %p182, 0;
$L__BB0_31:
@%p182 bra $L__BB0_43;
setp.eq.s32 %p44, %r44, 0;
mov.u16 %rs179, 0;
mov.f32 %f1530, 0f3F800000;
@%p44 bra $L__BB0_40;
abs.s32 %r10, %r44;
setp.lt.s32 %p45, %r10, 1;
mov.f32 %f1526, 0f00000000;
@%p45 bra $L__BB0_39;
mov.f32 %f548, 0f38D1B717;
max.f32 %f549, %f547, %f548;
and.b32 %r63, %r427, 32;
setp.eq.s32 %p46, %r63, 0;
mov.u32 %r426, 0;
selp.f32 %f550, 0f3F800000, 0f41200000, %p46;
mul.f32 %f137, %f550, %f549;
and.b32 %r11, %r10, 1;
setp.eq.s32 %p47, %r10, 1;
mov.f32 %f1526, 0f00000000;
@%p47 bra $L__BB0_37;
sub.s32 %r425, %r10, %r11;
$L__BB0_36:
cvt.rn.f32.s32 %f570, %r426;
mul.f32 %f571, %f570, 0f3DD32618;
cvt.rmi.f32.f32 %f572, %f571;
sub.f32 %f573, %f571, %f572;
mul.f32 %f574, %f570, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f575, %f574;
sub.f32 %f576, %f574, %f575;
mul.f32 %f577, %f570, 0f3DC74539;
cvt.rmi.f32.f32 %f578, %f577;
sub.f32 %f579, %f577, %f578;
add.f32 %f580, %f576, 0f4199851F;
add.f32 %f581, %f579, 0f4199851F;
add.f32 %f582, %f573, 0f4199851F;
mul.f32 %f583, %f576, %f581;
fma.rn.f32 %f584, %f573, %f580, %f583;
fma.rn.f32 %f585, %f582, %f579, %f584;
add.f32 %f586, %f573, %f585;
add.f32 %f587, %f576, %f585;
add.f32 %f588, %f579, %f585;
add.f32 %f589, %f586, %f587;
mul.f32 %f590, %f588, %f589;
cvt.rmi.f32.f32 %f591, %f590;
sub.f32 %f592, %f590, %f591;
add.f32 %f593, %f586, %f588;
mul.f32 %f594, %f587, %f593;
cvt.rmi.f32.f32 %f595, %f594;
sub.f32 %f596, %f594, %f595;
add.f32 %f597, %f587, %f588;
mul.f32 %f598, %f586, %f597;
cvt.rmi.f32.f32 %f599, %f598;
sub.f32 %f600, %f598, %f599;
fma.rn.f32 %f601, %f592, 0f40000000, 0fBF800000;
fma.rn.f32 %f602, %f596, 0f40000000, 0fBF800000;
fma.rn.f32 %f603, %f600, 0f40000000, 0fBF800000;
ld.global.f32 %f604, [%rd5+16];
ld.global.f32 %f605, [%rd5+4];
fma.rn.f32 %f606, %f604, %f601, %f605;
ld.global.f32 %f607, [%rd5+8];
fma.rn.f32 %f608, %f604, %f602, %f607;
ld.global.f32 %f609, [%rd5+12];
fma.rn.f32 %f610, %f604, %f603, %f609;
sub.f32 %f611, %f606, %f10;
sub.f32 %f612, %f608, %f11;
sub.f32 %f613, %f610, %f12;
mul.f32 %f614, %f612, %f612;
fma.rn.f32 %f615, %f611, %f611, %f614;
fma.rn.f32 %f616, %f613, %f613, %f615;
sqrt.rn.f32 %f559, %f616;
rcp.rn.f32 %f617, %f559;
mul.f32 %f555, %f617, %f611;
mul.f32 %f556, %f617, %f612;
mul.f32 %f557, %f617, %f613;
mov.f32 %f569, 0f00000000;
mov.u32 %r172, 2;
mov.u32 %r174, 1;
mov.u32 %r175, 1065353216;
mov.u32 %r206, 0;
// begin inline asm
call(%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,%r94,%r95,%r96),_optix_trace_typed_32,(%r206,%rd3,%f13,%f14,%f15,%f555,%f556,%f557,%f137,%f559,%f569,%r174,%r206,%r174,%r172,%r174,%r174,%r175,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206);
// end inline asm
mov.b32 %f618, %r65;
add.f32 %f619, %f1526, %f618;
add.s32 %r207, %r426, 1;
cvt.rn.f32.s32 %f620, %r207;
mul.f32 %f621, %f620, 0f3DD32618;
cvt.rmi.f32.f32 %f622, %f621;
sub.f32 %f623, %f621, %f622;
mul.f32 %f624, %f620, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f625, %f624;
sub.f32 %f626, %f624, %f625;
mul.f32 %f627, %f620, 0f3DC74539;
cvt.rmi.f32.f32 %f628, %f627;
sub.f32 %f629, %f627, %f628;
add.f32 %f630, %f626, 0f4199851F;
add.f32 %f631, %f629, 0f4199851F;
add.f32 %f632, %f623, 0f4199851F;
mul.f32 %f633, %f626, %f631;
fma.rn.f32 %f634, %f623, %f630, %f633;
fma.rn.f32 %f635, %f632, %f629, %f634;
add.f32 %f636, %f623, %f635;
add.f32 %f637, %f626, %f635;
add.f32 %f638, %f629, %f635;
add.f32 %f639, %f636, %f637;
mul.f32 %f640, %f638, %f639;
cvt.rmi.f32.f32 %f641, %f640;
sub.f32 %f642, %f640, %f641;
add.f32 %f643, %f636, %f638;
mul.f32 %f644, %f637, %f643;
cvt.rmi.f32.f32 %f645, %f644;
sub.f32 %f646, %f644, %f645;
add.f32 %f647, %f637, %f638;
mul.f32 %f648, %f636, %f647;
cvt.rmi.f32.f32 %f649, %f648;
sub.f32 %f650, %f648, %f649;
fma.rn.f32 %f651, %f642, 0f40000000, 0fBF800000;
fma.rn.f32 %f652, %f646, 0f40000000, 0fBF800000;
fma.rn.f32 %f653, %f650, 0f40000000, 0fBF800000;
ld.global.f32 %f654, [%rd5+16];
ld.global.f32 %f655, [%rd5+4];
fma.rn.f32 %f656, %f654, %f651, %f655;
ld.global.f32 %f657, [%rd5+8];
fma.rn.f32 %f658, %f654, %f652, %f657;
ld.global.f32 %f659, [%rd5+12];
fma.rn.f32 %f660, %f654, %f653, %f659;
sub.f32 %f661, %f656, %f10;
sub.f32 %f662, %f658, %f11;
sub.f32 %f663, %f660, %f12;
mul.f32 %f664, %f662, %f662;
fma.rn.f32 %f665, %f661, %f661, %f664;
fma.rn.f32 %f666, %f663, %f663, %f665;
sqrt.rn.f32 %f568, %f666;
rcp.rn.f32 %f667, %f568;
mul.f32 %f564, %f667, %f661;
mul.f32 %f565, %f667, %f662;
mul.f32 %f566, %f667, %f663;
// begin inline asm
call(%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,%r165,%r166,%r167),_optix_trace_typed_32,(%r206,%rd3,%f13,%f14,%f15,%f564,%f565,%f566,%f137,%f568,%f569,%r174,%r206,%r174,%r172,%r174,%r174,%r175,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206);
// end inline asm
mov.b32 %f668, %r136;
add.f32 %f1526, %f619, %f668;
add.s32 %r426, %r426, 2;
add.s32 %r425, %r425, -2;
setp.ne.s32 %p48, %r425, 0;
@%p48 bra $L__BB0_36;
$L__BB0_37:
setp.eq.s32 %p49, %r11, 0;
@%p49 bra $L__BB0_39;
cvt.rn.f32.s32 %f678, %r426;
mul.f32 %f679, %f678, 0f3DD32618;
cvt.rmi.f32.f32 %f680, %f679;
sub.f32 %f681, %f679, %f680;
mul.f32 %f682, %f678, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f683, %f682;
sub.f32 %f684, %f682, %f683;
mul.f32 %f685, %f678, 0f3DC74539;
cvt.rmi.f32.f32 %f686, %f685;
sub.f32 %f687, %f685, %f686;
add.f32 %f688, %f684, 0f4199851F;
add.f32 %f689, %f687, 0f4199851F;
add.f32 %f690, %f681, 0f4199851F;
mul.f32 %f691, %f684, %f689;
fma.rn.f32 %f692, %f681, %f688, %f691;
fma.rn.f32 %f693, %f690, %f687, %f692;
add.f32 %f694, %f681, %f693;
add.f32 %f695, %f684, %f693;
add.f32 %f696, %f687, %f693;
add.f32 %f697, %f694, %f695;
mul.f32 %f698, %f696, %f697;
cvt.rmi.f32.f32 %f699, %f698;
sub.f32 %f700, %f698, %f699;
add.f32 %f701, %f694, %f696;
mul.f32 %f702, %f695, %f701;
cvt.rmi.f32.f32 %f703, %f702;
sub.f32 %f704, %f702, %f703;
add.f32 %f705, %f695, %f696;
mul.f32 %f706, %f694, %f705;
cvt.rmi.f32.f32 %f707, %f706;
sub.f32 %f708, %f706, %f707;
fma.rn.f32 %f709, %f700, 0f40000000, 0fBF800000;
fma.rn.f32 %f710, %f704, 0f40000000, 0fBF800000;
fma.rn.f32 %f711, %f708, 0f40000000, 0fBF800000;
ld.global.f32 %f712, [%rd5+16];
ld.global.f32 %f713, [%rd5+4];
fma.rn.f32 %f714, %f712, %f709, %f713;
ld.global.f32 %f715, [%rd5+8];
fma.rn.f32 %f716, %f712, %f710, %f715;
ld.global.f32 %f717, [%rd5+12];
fma.rn.f32 %f718, %f712, %f711, %f717;
sub.f32 %f719, %f714, %f10;
sub.f32 %f720, %f716, %f11;
sub.f32 %f721, %f718, %f12;
mul.f32 %f722, %f720, %f720;
fma.rn.f32 %f723, %f719, %f719, %f722;
fma.rn.f32 %f724, %f721, %f721, %f723;
sqrt.rn.f32 %f676, %f724;
rcp.rn.f32 %f725, %f676;
mul.f32 %f672, %f725, %f719;
mul.f32 %f673, %f725, %f720;
mul.f32 %f674, %f725, %f721;
mov.f32 %f677, 0f00000000;
mov.u32 %r244, 2;
mov.u32 %r246, 1;
mov.u32 %r247, 1065353216;
mov.u32 %r278, 0;
// begin inline asm
call(%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,%r237,%r238,%r239),_optix_trace_typed_32,(%r278,%rd3,%f13,%f14,%f15,%f672,%f673,%f674,%f137,%f676,%f677,%r246,%r278,%r246,%r244,%r246,%r246,%r247,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278);
// end inline asm
mov.b32 %f726, %r208;
add.f32 %f1526, %f1526, %f726;
$L__BB0_39:
cvt.rn.f32.s32 %f727, %r10;
div.rn.f32 %f1530, %f1526, %f727;
shr.u32 %r279, %r44, 31;
cvt.u16.u32 %rs179, %r279;
$L__BB0_40:
fma.rn.f32 %f1531, %f1511, %f1530, %f1531;
fma.rn.f32 %f1532, %f1510, %f1530, %f1532;
fma.rn.f32 %f1533, %f1512, %f1530, %f1533;
fma.rn.f32 %f1534, %f1513, %f1530, %f1534;
fma.rn.f32 %f1535, %f1514, %f1530, %f1535;
fma.rn.f32 %f1536, %f1515, %f1530, %f1536;
fma.rn.f32 %f1537, %f1516, %f1530, %f1537;
fma.rn.f32 %f1538, %f1517, %f1530, %f1538;
fma.rn.f32 %f1539, %f1518, %f1530, %f1539;
fma.rn.f32 %f1540, %f1519, %f1530, %f1540;
fma.rn.f32 %f1541, %f1520, %f1530, %f1541;
fma.rn.f32 %f1542, %f1521, %f1530, %f1542;
fma.rn.f32 %f1543, %f1522, %f1530, %f1543;
fma.rn.f32 %f1544, %f1523, %f1530, %f1544;
fma.rn.f32 %f1545, %f1524, %f1530, %f1545;
setp.eq.s16 %p50, %rs179, 0;
@%p50 bra $L__BB0_42;
div.rn.f32 %f728, %f1511, %f356;
div.rn.f32 %f729, %f728, %f73;
cvt.sat.f32.f32 %f730, %f729;
mul.f32 %f1530, %f1530, %f730;
$L__BB0_42:
add.f32 %f1546, %f1546, %f1530;
$L__BB0_43:
cvt.u64.u32 %rd77, %r423;
cvt.u32.u64 %r280, %rd77;
add.s32 %r423, %r280, 1;
setp.lt.u32 %p51, %r423, %r5;
@%p51 bra $L__BB0_7;
bra.uni $L__BB0_45;
$L__BB0_44:
mov.f32 %f1531, 0f00000000;
ld.const.u32 %r427, [params+104];
mov.f32 %f1532, %f1531;
mov.f32 %f1533, %f1531;
mov.f32 %f1534, %f1531;
mov.f32 %f1535, %f1531;
mov.f32 %f1536, %f1531;
mov.f32 %f1537, %f1531;
mov.f32 %f1538, %f1531;
mov.f32 %f1539, %f1531;
mov.f32 %f1540, %f1531;
mov.f32 %f1541, %f1531;
mov.f32 %f1542, %f1531;
mov.f32 %f1543, %f1531;
mov.f32 %f1544, %f1531;
mov.f32 %f1545, %f1531;
mov.f32 %f1546, %f1531;
$L__BB0_45:
and.b32 %r281, %r427, 8;
setp.eq.s32 %p52, %r281, 0;
@%p52 bra $L__BB0_59;
cvt.sat.f32.f32 %f196, %f1546;
ld.const.u64 %rd33, [params+192];
cvta.to.global.u64 %rd7, %rd33;
ld.const.u32 %r282, [params+184];
mad.lo.s32 %r283, %r282, %r4, %r3;
cvt.u64.u32 %rd8, %r283;
mov.f32 %f748, 0f3E68BA2E;
cvt.rzi.f32.f32 %f749, %f748;
add.f32 %f750, %f749, %f749;
mov.f32 %f751, 0f3EE8BA2E;
sub.f32 %f752, %f751, %f750;
abs.f32 %f197, %f752;
abs.f32 %f198, %f196;
setp.lt.f32 %p53, %f198, 0f00800000;
mul.f32 %f753, %f198, 0f4B800000;
selp.f32 %f754, %f753, %f198, %p53;
selp.f32 %f755, 0fC3170000, 0fC2FE0000, %p53;
mov.b32 %r284, %f754;
and.b32 %r285, %r284, 8388607;
or.b32 %r286, %r285, 1065353216;
mov.b32 %f756, %r286;
shr.u32 %r287, %r284, 23;
cvt.rn.f32.u32 %f757, %r287;
add.f32 %f758, %f755, %f757;
setp.gt.f32 %p54, %f756, 0f3FB504F3;
mul.f32 %f759, %f756, 0f3F000000;
add.f32 %f760, %f758, 0f3F800000;
selp.f32 %f761, %f760, %f758, %p54;
selp.f32 %f762, %f759, %f756, %p54;
add.f32 %f763, %f762, 0fBF800000;
add.f32 %f764, %f762, 0f3F800000;
rcp.approx.ftz.f32 %f765, %f764;
add.f32 %f766, %f763, %f763;
mul.f32 %f767, %f766, %f765;
mul.f32 %f768, %f767, %f767;
mov.f32 %f769, 0f3C4CAF63;
mov.f32 %f770, 0f3B18F0FE;
fma.rn.f32 %f771, %f770, %f768, %f769;
mov.f32 %f772, 0f3DAAAABD;
fma.rn.f32 %f773, %f771, %f768, %f772;
mul.rn.f32 %f774, %f773, %f768;
mul.rn.f32 %f775, %f774, %f767;
sub.f32 %f776, %f763, %f767;
add.f32 %f777, %f776, %f776;
neg.f32 %f778, %f767;
fma.rn.f32 %f779, %f778, %f763, %f777;
mul.rn.f32 %f780, %f765, %f779;
add.f32 %f781, %f775, %f767;
sub.f32 %f782, %f767, %f781;
add.f32 %f783, %f775, %f782;
add.f32 %f784, %f780, %f783;
add.f32 %f785, %f781, %f784;
sub.f32 %f786, %f781, %f785;
add.f32 %f787, %f784, %f786;
mov.f32 %f788, 0f3F317200;
mul.rn.f32 %f789, %f761, %f788;
mov.f32 %f790, 0f35BFBE8E;
mul.rn.f32 %f791, %f761, %f790;
add.f32 %f792, %f789, %f785;
sub.f32 %f793, %f789, %f792;
add.f32 %f794, %f785, %f793;
add.f32 %f795, %f787, %f794;
add.f32 %f796, %f791, %f795;
add.f32 %f797, %f792, %f796;
sub.f32 %f798, %f792, %f797;
add.f32 %f799, %f796, %f798;
mul.rn.f32 %f800, %f751, %f797;
neg.f32 %f801, %f800;
fma.rn.f32 %f802, %f751, %f797, %f801;
fma.rn.f32 %f803, %f751, %f799, %f802;
mov.f32 %f804, 0f00000000;
fma.rn.f32 %f805, %f804, %f797, %f803;
add.rn.f32 %f806, %f800, %f805;
neg.f32 %f807, %f806;
add.rn.f32 %f808, %f800, %f807;
add.rn.f32 %f809, %f808, %f805;
mov.b32 %r288, %f806;
setp.eq.s32 %p55, %r288, 1118925336;
add.s32 %r289, %r288, -1;
mov.b32 %f810, %r289;
add.f32 %f811, %f809, 0f37000000;
selp.f32 %f199, %f811, %f809, %p55;
selp.f32 %f812, %f810, %f806, %p55;
mov.f32 %f813, 0f3FB8AA3B;
mul.rn.f32 %f814, %f812, %f813;
cvt.rzi.f32.f32 %f815, %f814;
abs.f32 %f816, %f815;
setp.gt.f32 %p56, %f816, 0f42FC0000;
mov.b32 %r290, %f815;
and.b32 %r291, %r290, -2147483648;
or.b32 %r292, %r291, 1123811328;
mov.b32 %f817, %r292;
selp.f32 %f818, %f817, %f815, %p56;
mov.f32 %f819, 0fBF317218;
fma.rn.f32 %f820, %f818, %f819, %f812;
mov.f32 %f821, 0f3102E308;
fma.rn.f32 %f822, %f818, %f821, %f820;
mul.f32 %f823, %f822, 0f3FB8AA3B;
add.f32 %f824, %f818, 0f4B40007F;
mov.b32 %r293, %f824;
shl.b32 %r294, %r293, 23;
mov.b32 %f825, %r294;
ex2.approx.ftz.f32 %f826, %f823;
mul.f32 %f200, %f826, %f825;
setp.eq.f32 %p57, %f200, 0f7F800000;
mov.f32 %f1563, 0f7F800000;
@%p57 bra $L__BB0_48;
fma.rn.f32 %f1563, %f200, %f199, %f200;
$L__BB0_48:
setp.lt.f32 %p58, %f196, 0f00000000;
setp.eq.f32 %p59, %f197, 0f3F800000;
and.pred %p3, %p58, %p59;
setp.eq.f32 %p60, %f196, 0f00000000;
@%p60 bra $L__BB0_52;
bra.uni $L__BB0_49;
$L__BB0_52:
add.f32 %f831, %f196, %f196;
selp.f32 %f1565, %f831, 0f00000000, %p59;
bra.uni $L__BB0_53;
$L__BB0_156:
mov.f32 %f1440, 0f00000000;
mov.u32 %r429, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs115, %f1440;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs114, %f1440;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f1440;}
// end inline asm
mov.u16 %rs116, 0;
st.global.v4.u16 [%rd16], {%rs113, %rs114, %rs115, %rs116};
$L__BB0_157:
ld.const.u64 %rd65, [params+256];
cvta.to.global.u64 %rd66, %rd65;
ld.const.u32 %r415, [params+248];
mad.lo.s32 %r416, %r415, %r4, %r3;
mul.wide.u32 %rd67, %r416, 8;
add.s64 %rd17, %rd66, %rd67;
setp.eq.s32 %p178, %r429, 0;
@%p178 bra $L__BB0_159;
ld.global.v4.u16 {%rs123, %rs124, %rs125, %rs126}, [%rd17];
// begin inline asm
{ cvt.f32.f16 %f1441, %rs123;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1442, %rs124;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1443, %rs125;}
// end inline asm
add.f32 %f1444, %f1441, 0f00000000;
add.f32 %f1445, %f1442, 0f00000000;
add.f32 %f1446, %f1443, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs122, %f1446;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs121, %f1445;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs120, %f1444;}
// end inline asm
mov.u16 %rs127, 0;
st.global.v4.u16 [%rd17], {%rs120, %rs121, %rs122, %rs127};
bra.uni $L__BB0_160;
$L__BB0_159:
mov.f32 %f1449, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs130, %f1449;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs129, %f1449;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1449;}
// end inline asm
mov.u16 %rs131, 0;
st.global.v4.u16 [%rd17], {%rs128, %rs129, %rs130, %rs131};
$L__BB0_160:
ld.const.u64 %rd68, [params+272];
cvta.to.global.u64 %rd69, %rd68;
ld.const.u32 %r417, [params+264];
mad.lo.s32 %r418, %r417, %r4, %r3;
mul.wide.u32 %rd70, %r418, 8;
add.s64 %rd18, %rd69, %rd70;
@%p178 bra $L__BB0_162;
ld.global.v4.u16 {%rs138, %rs139, %rs140, %rs141}, [%rd18];
// begin inline asm
{ cvt.f32.f16 %f1450, %rs138;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1451, %rs139;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1452, %rs140;}
// end inline asm
add.f32 %f1453, %f1450, 0f00000000;
add.f32 %f1454, %f1451, 0f00000000;
add.f32 %f1455, %f1452, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs137, %f1455;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs136, %f1454;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs135, %f1453;}
// end inline asm
mov.u16 %rs142, 0;
st.global.v4.u16 [%rd18], {%rs135, %rs136, %rs137, %rs142};
bra.uni $L__BB0_163;
$L__BB0_162:
mov.f32 %f1458, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs145, %f1458;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs144, %f1458;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1458;}
// end inline asm
mov.u16 %rs146, 0;
st.global.v4.u16 [%rd18], {%rs143, %rs144, %rs145, %rs146};
$L__BB0_163:
ld.const.u64 %rd71, [params+288];
cvta.to.global.u64 %rd72, %rd71;
ld.const.u32 %r419, [params+280];
mad.lo.s32 %r420, %r419, %r4, %r3;
mul.wide.u32 %rd73, %r420, 8;
add.s64 %rd19, %rd72, %rd73;
@%p178 bra $L__BB0_165;
ld.global.v4.u16 {%rs153, %rs154, %rs155, %rs156}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f1459, %rs153;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1460, %rs154;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1461, %rs155;}
// end inline asm
add.f32 %f1462, %f1459, 0f00000000;
add.f32 %f1463, %f1460, 0f00000000;
add.f32 %f1464, %f1461, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs152, %f1464;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs151, %f1463;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs150, %f1462;}
// end inline asm
mov.u16 %rs157, 0;
st.global.v4.u16 [%rd19], {%rs150, %rs151, %rs152, %rs157};
bra.uni $L__BB0_166;
$L__BB0_165:
mov.f32 %f1467, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs160, %f1467;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs159, %f1467;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1467;}
// end inline asm
mov.u16 %rs161, 0;
st.global.v4.u16 [%rd19], {%rs158, %rs159, %rs160, %rs161};
$L__BB0_166:
ld.const.u64 %rd74, [params+304];
cvta.to.global.u64 %rd75, %rd74;
ld.const.u32 %r421, [params+296];
mad.lo.s32 %r422, %r421, %r4, %r3;
mul.wide.u32 %rd76, %r422, 8;
add.s64 %rd20, %rd75, %rd76;
@%p178 bra $L__BB0_168;
ld.global.v4.u16 {%rs168, %rs169, %rs170, %rs171}, [%rd20];
// begin inline asm
{ cvt.f32.f16 %f1468, %rs168;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1469, %rs169;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1470, %rs170;}
// end inline asm
add.f32 %f1471, %f1468, 0f00000000;
add.f32 %f1472, %f1469, 0f00000000;
add.f32 %f1473, %f1470, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs167, %f1473;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs166, %f1472;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs165, %f1471;}
// end inline asm
mov.u16 %rs172, 0;
st.global.v4.u16 [%rd20], {%rs165, %rs166, %rs167, %rs172};
bra.uni $L__BB0_169;
$L__BB0_168:
mov.f32 %f1476, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs175, %f1476;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs174, %f1476;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs173, %f1476;}
// end inline asm
mov.u16 %rs176, 0;
st.global.v4.u16 [%rd20], {%rs173, %rs174, %rs175, %rs176};
bra.uni $L__BB0_169;
$L__BB0_49:
mov.b32 %r295, %f1563;
xor.b32 %r296, %r295, -2147483648;
mov.b32 %f827, %r296;
selp.f32 %f1565, %f827, %f1563, %p3;
setp.geu.f32 %p61, %f196, 0f00000000;
@%p61 bra $L__BB0_53;
mov.f32 %f828, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f829, %f828;
setp.eq.f32 %p62, %f829, 0f3EE8BA2E;
@%p62 bra $L__BB0_53;
mov.f32 %f1565, 0f7FFFFFFF;
$L__BB0_53:
add.f32 %f832, %f198, 0f3EE8BA2E;
mov.b32 %r297, %f832;
setp.lt.s32 %p64, %r297, 2139095040;
@%p64 bra $L__BB0_58;
setp.gtu.f32 %p65, %f198, 0f7F800000;
@%p65 bra $L__BB0_57;
bra.uni $L__BB0_55;
$L__BB0_57:
add.f32 %f1565, %f196, 0f3EE8BA2E;
bra.uni $L__BB0_58;
$L__BB0_55:
setp.neu.f32 %p66, %f198, 0f7F800000;
@%p66 bra $L__BB0_58;
selp.f32 %f1565, 0fFF800000, 0f7F800000, %p3;
$L__BB0_58:
mul.f32 %f833, %f1565, 0f437F0000;
setp.eq.f32 %p67, %f196, 0f3F800000;
selp.f32 %f834, 0f437F0000, %f833, %p67;
cvt.rzi.u32.f32 %r298, %f834;
shl.b64 %rd34, %rd8, 1;
add.s64 %rd35, %rd7, %rd34;
cvt.u16.u32 %rs18, %r298;
mov.u16 %rs19, 255;
st.global.v2.u8 [%rd35], {%rs18, %rs19};
$L__BB0_59:
and.b32 %r299, %r427, 1;
setp.eq.b32 %p68, %r299, 1;
mov.pred %p69, 0;
xor.pred %p70, %p68, %p69;
not.pred %p71, %p70;
@%p71 bra $L__BB0_133;
mov.f32 %f836, 0f3E666666;
cvt.rzi.f32.f32 %f837, %f836;
add.f32 %f838, %f837, %f837;
mov.f32 %f839, 0f3EE66666;
sub.f32 %f840, %f839, %f838;
abs.f32 %f209, %f840;
abs.f32 %f210, %f1531;
setp.lt.f32 %p72, %f210, 0f00800000;
mul.f32 %f841, %f210, 0f4B800000;
selp.f32 %f842, %f841, %f210, %p72;
selp.f32 %f843, 0fC3170000, 0fC2FE0000, %p72;
mov.b32 %r300, %f842;
and.b32 %r301, %r300, 8388607;
or.b32 %r302, %r301, 1065353216;
mov.b32 %f844, %r302;
shr.u32 %r303, %r300, 23;
cvt.rn.f32.u32 %f845, %r303;
add.f32 %f846, %f843, %f845;
setp.gt.f32 %p73, %f844, 0f3FB504F3;
mul.f32 %f847, %f844, 0f3F000000;
add.f32 %f848, %f846, 0f3F800000;
selp.f32 %f849, %f848, %f846, %p73;
selp.f32 %f850, %f847, %f844, %p73;
add.f32 %f851, %f850, 0fBF800000;
add.f32 %f852, %f850, 0f3F800000;
rcp.approx.ftz.f32 %f853, %f852;
add.f32 %f854, %f851, %f851;
mul.f32 %f855, %f854, %f853;
mul.f32 %f856, %f855, %f855;
mov.f32 %f857, 0f3C4CAF63;
mov.f32 %f858, 0f3B18F0FE;
fma.rn.f32 %f859, %f858, %f856, %f857;
mov.f32 %f860, 0f3DAAAABD;
fma.rn.f32 %f861, %f859, %f856, %f860;
mul.rn.f32 %f862, %f861, %f856;
mul.rn.f32 %f863, %f862, %f855;
sub.f32 %f864, %f851, %f855;
add.f32 %f865, %f864, %f864;
neg.f32 %f866, %f855;
fma.rn.f32 %f867, %f866, %f851, %f865;
mul.rn.f32 %f868, %f853, %f867;
add.f32 %f869, %f863, %f855;
sub.f32 %f870, %f855, %f869;
add.f32 %f871, %f863, %f870;
add.f32 %f872, %f868, %f871;
add.f32 %f873, %f869, %f872;
sub.f32 %f874, %f869, %f873;
add.f32 %f875, %f872, %f874;
mov.f32 %f876, 0f3F317200;
mul.rn.f32 %f877, %f849, %f876;
mov.f32 %f878, 0f35BFBE8E;
mul.rn.f32 %f879, %f849, %f878;
add.f32 %f880, %f877, %f873;
sub.f32 %f881, %f877, %f880;
add.f32 %f882, %f873, %f881;
add.f32 %f883, %f875, %f882;
add.f32 %f884, %f879, %f883;
add.f32 %f885, %f880, %f884;
sub.f32 %f886, %f880, %f885;
add.f32 %f887, %f884, %f886;
mul.rn.f32 %f888, %f839, %f885;
neg.f32 %f889, %f888;
fma.rn.f32 %f890, %f839, %f885, %f889;
fma.rn.f32 %f891, %f839, %f887, %f890;
mov.f32 %f892, 0f00000000;
fma.rn.f32 %f893, %f892, %f885, %f891;
add.rn.f32 %f894, %f888, %f893;
neg.f32 %f895, %f894;
add.rn.f32 %f896, %f888, %f895;
add.rn.f32 %f897, %f896, %f893;
mov.b32 %r304, %f894;
setp.eq.s32 %p74, %r304, 1118925336;
add.s32 %r305, %r304, -1;
mov.b32 %f898, %r305;
add.f32 %f899, %f897, 0f37000000;
selp.f32 %f211, %f899, %f897, %p74;
selp.f32 %f900, %f898, %f894, %p74;
mov.f32 %f901, 0f3FB8AA3B;
mul.rn.f32 %f902, %f900, %f901;
cvt.rzi.f32.f32 %f903, %f902;
abs.f32 %f904, %f903;
setp.gt.f32 %p75, %f904, 0f42FC0000;
mov.b32 %r306, %f903;
and.b32 %r307, %r306, -2147483648;
or.b32 %r308, %r307, 1123811328;
mov.b32 %f905, %r308;
selp.f32 %f906, %f905, %f903, %p75;
mov.f32 %f907, 0fBF317218;
fma.rn.f32 %f908, %f906, %f907, %f900;
mov.f32 %f909, 0f3102E308;
fma.rn.f32 %f910, %f906, %f909, %f908;
mul.f32 %f911, %f910, 0f3FB8AA3B;
add.f32 %f912, %f906, 0f4B40007F;
mov.b32 %r309, %f912;
shl.b32 %r310, %r309, 23;
mov.b32 %f913, %r310;
ex2.approx.ftz.f32 %f914, %f911;
mul.f32 %f212, %f914, %f913;
setp.eq.f32 %p76, %f212, 0f7F800000;
mov.f32 %f1566, 0f7F800000;
@%p76 bra $L__BB0_62;
fma.rn.f32 %f1566, %f212, %f211, %f212;
$L__BB0_62:
setp.lt.f32 %p77, %f1531, 0f00000000;
setp.eq.f32 %p78, %f209, 0f3F800000;
and.pred %p4, %p77, %p78;
setp.eq.f32 %p79, %f1531, 0f00000000;
@%p79 bra $L__BB0_66;
bra.uni $L__BB0_63;
$L__BB0_66:
add.f32 %f919, %f1531, %f1531;
selp.f32 %f1568, %f919, 0f00000000, %p78;
bra.uni $L__BB0_67;
$L__BB0_63:
mov.b32 %r311, %f1566;
xor.b32 %r312, %r311, -2147483648;
mov.b32 %f915, %r312;
selp.f32 %f1568, %f915, %f1566, %p4;
setp.geu.f32 %p80, %f1531, 0f00000000;
@%p80 bra $L__BB0_67;
mov.f32 %f916, 0f3EE66666;
cvt.rzi.f32.f32 %f917, %f916;
setp.eq.f32 %p81, %f917, 0f3EE66666;
@%p81 bra $L__BB0_67;
mov.f32 %f1568, 0f7FFFFFFF;
$L__BB0_67:
add.f32 %f920, %f210, 0f3EE66666;
mov.b32 %r313, %f920;
setp.lt.s32 %p83, %r313, 2139095040;
@%p83 bra $L__BB0_72;
setp.gtu.f32 %p84, %f210, 0f7F800000;
@%p84 bra $L__BB0_71;
bra.uni $L__BB0_69;
$L__BB0_71:
add.f32 %f1568, %f1531, 0f3EE66666;
bra.uni $L__BB0_72;
$L__BB0_69:
setp.neu.f32 %p85, %f210, 0f7F800000;
@%p85 bra $L__BB0_72;
selp.f32 %f1568, 0fFF800000, 0f7F800000, %p4;
$L__BB0_72:
setp.eq.f32 %p86, %f1531, 0f3F800000;
selp.f32 %f221, 0f3F800000, %f1568, %p86;
abs.f32 %f222, %f1532;
setp.lt.f32 %p87, %f222, 0f00800000;
mul.f32 %f922, %f222, 0f4B800000;
selp.f32 %f923, %f922, %f222, %p87;
selp.f32 %f924, 0fC3170000, 0fC2FE0000, %p87;
mov.b32 %r314, %f923;
and.b32 %r315, %r314, 8388607;
or.b32 %r316, %r315, 1065353216;
mov.b32 %f925, %r316;
shr.u32 %r317, %r314, 23;
cvt.rn.f32.u32 %f926, %r317;
add.f32 %f927, %f924, %f926;
setp.gt.f32 %p88, %f925, 0f3FB504F3;
mul.f32 %f928, %f925, 0f3F000000;
add.f32 %f929, %f927, 0f3F800000;
selp.f32 %f930, %f929, %f927, %p88;
selp.f32 %f931, %f928, %f925, %p88;
add.f32 %f932, %f931, 0fBF800000;
add.f32 %f933, %f931, 0f3F800000;
rcp.approx.ftz.f32 %f934, %f933;
add.f32 %f935, %f932, %f932;
mul.f32 %f936, %f935, %f934;
mul.f32 %f937, %f936, %f936;
mov.f32 %f938, 0f3C4CAF63;
mov.f32 %f939, 0f3B18F0FE;
fma.rn.f32 %f940, %f939, %f937, %f938;
mov.f32 %f941, 0f3DAAAABD;
fma.rn.f32 %f942, %f940, %f937, %f941;
mul.rn.f32 %f943, %f942, %f937;
mul.rn.f32 %f944, %f943, %f936;
sub.f32 %f945, %f932, %f936;
add.f32 %f946, %f945, %f945;
neg.f32 %f947, %f936;
fma.rn.f32 %f948, %f947, %f932, %f946;
mul.rn.f32 %f949, %f934, %f948;
add.f32 %f950, %f944, %f936;
sub.f32 %f951, %f936, %f950;
add.f32 %f952, %f944, %f951;
add.f32 %f953, %f949, %f952;
add.f32 %f954, %f950, %f953;
sub.f32 %f955, %f950, %f954;
add.f32 %f956, %f953, %f955;
mov.f32 %f957, 0f3F317200;
mul.rn.f32 %f958, %f930, %f957;
mov.f32 %f959, 0f35BFBE8E;
mul.rn.f32 %f960, %f930, %f959;
add.f32 %f961, %f958, %f954;
sub.f32 %f962, %f958, %f961;
add.f32 %f963, %f954, %f962;
add.f32 %f964, %f956, %f963;
add.f32 %f965, %f960, %f964;
add.f32 %f966, %f961, %f965;
sub.f32 %f967, %f961, %f966;
add.f32 %f968, %f965, %f967;
mov.f32 %f969, 0f3EE66666;
mul.rn.f32 %f970, %f969, %f966;
neg.f32 %f971, %f970;
fma.rn.f32 %f972, %f969, %f966, %f971;
fma.rn.f32 %f973, %f969, %f968, %f972;
mov.f32 %f974, 0f00000000;
fma.rn.f32 %f975, %f974, %f966, %f973;
add.rn.f32 %f976, %f970, %f975;
neg.f32 %f977, %f976;
add.rn.f32 %f978, %f970, %f977;
add.rn.f32 %f979, %f978, %f975;
mov.b32 %r318, %f976;
setp.eq.s32 %p89, %r318, 1118925336;
add.s32 %r319, %r318, -1;
mov.b32 %f980, %r319;
add.f32 %f981, %f979, 0f37000000;
selp.f32 %f223, %f981, %f979, %p89;
selp.f32 %f982, %f980, %f976, %p89;
mov.f32 %f983, 0f3FB8AA3B;
mul.rn.f32 %f984, %f982, %f983;
cvt.rzi.f32.f32 %f985, %f984;
abs.f32 %f986, %f985;
setp.gt.f32 %p90, %f986, 0f42FC0000;
mov.b32 %r320, %f985;
and.b32 %r321, %r320, -2147483648;
or.b32 %r322, %r321, 1123811328;
mov.b32 %f987, %r322;
selp.f32 %f988, %f987, %f985, %p90;
mov.f32 %f989, 0fBF317218;
fma.rn.f32 %f990, %f988, %f989, %f982;
mov.f32 %f991, 0f3102E308;
fma.rn.f32 %f992, %f988, %f991, %f990;
mul.f32 %f993, %f992, 0f3FB8AA3B;
add.f32 %f994, %f988, 0f4B40007F;
mov.b32 %r323, %f994;
shl.b32 %r324, %r323, 23;
mov.b32 %f995, %r324;
ex2.approx.ftz.f32 %f996, %f993;
mul.f32 %f224, %f996, %f995;
setp.eq.f32 %p91, %f224, 0f7F800000;
mov.f32 %f1569, 0f7F800000;
@%p91 bra $L__BB0_74;
fma.rn.f32 %f1569, %f224, %f223, %f224;
$L__BB0_74:
setp.lt.f32 %p92, %f1532, 0f00000000;
and.pred %p5, %p92, %p78;
setp.eq.f32 %p94, %f1532, 0f00000000;
@%p94 bra $L__BB0_78;
bra.uni $L__BB0_75;
$L__BB0_78:
add.f32 %f1001, %f1532, %f1532;
selp.f32 %f1571, %f1001, 0f00000000, %p78;
bra.uni $L__BB0_79;
$L__BB0_75:
mov.b32 %r325, %f1569;
xor.b32 %r326, %r325, -2147483648;
mov.b32 %f997, %r326;
selp.f32 %f1571, %f997, %f1569, %p5;
setp.geu.f32 %p95, %f1532, 0f00000000;
@%p95 bra $L__BB0_79;
mov.f32 %f998, 0f3EE66666;
cvt.rzi.f32.f32 %f999, %f998;
setp.eq.f32 %p96, %f999, 0f3EE66666;
@%p96 bra $L__BB0_79;
mov.f32 %f1571, 0f7FFFFFFF;
$L__BB0_79:
add.f32 %f1002, %f222, 0f3EE66666;
mov.b32 %r327, %f1002;
setp.lt.s32 %p98, %r327, 2139095040;
@%p98 bra $L__BB0_84;
setp.gtu.f32 %p99, %f222, 0f7F800000;
@%p99 bra $L__BB0_83;
bra.uni $L__BB0_81;
$L__BB0_83:
add.f32 %f1571, %f1532, 0f3EE66666;
bra.uni $L__BB0_84;
$L__BB0_81:
setp.neu.f32 %p100, %f222, 0f7F800000;
@%p100 bra $L__BB0_84;
selp.f32 %f1571, 0fFF800000, 0f7F800000, %p5;
$L__BB0_84:
setp.eq.f32 %p101, %f1532, 0f3F800000;
selp.f32 %f233, 0f3F800000, %f1571, %p101;
abs.f32 %f234, %f1533;
setp.lt.f32 %p102, %f234, 0f00800000;
mul.f32 %f1004, %f234, 0f4B800000;
selp.f32 %f1005, %f1004, %f234, %p102;
selp.f32 %f1006, 0fC3170000, 0fC2FE0000, %p102;
mov.b32 %r328, %f1005;
and.b32 %r329, %r328, 8388607;
or.b32 %r330, %r329, 1065353216;
mov.b32 %f1007, %r330;
shr.u32 %r331, %r328, 23;
cvt.rn.f32.u32 %f1008, %r331;
add.f32 %f1009, %f1006, %f1008;
setp.gt.f32 %p103, %f1007, 0f3FB504F3;
mul.f32 %f1010, %f1007, 0f3F000000;
add.f32 %f1011, %f1009, 0f3F800000;
selp.f32 %f1012, %f1011, %f1009, %p103;
selp.f32 %f1013, %f1010, %f1007, %p103;
add.f32 %f1014, %f1013, 0fBF800000;
add.f32 %f1015, %f1013, 0f3F800000;
rcp.approx.ftz.f32 %f1016, %f1015;
add.f32 %f1017, %f1014, %f1014;
mul.f32 %f1018, %f1017, %f1016;
mul.f32 %f1019, %f1018, %f1018;
mov.f32 %f1020, 0f3C4CAF63;
mov.f32 %f1021, 0f3B18F0FE;
fma.rn.f32 %f1022, %f1021, %f1019, %f1020;
mov.f32 %f1023, 0f3DAAAABD;
fma.rn.f32 %f1024, %f1022, %f1019, %f1023;
mul.rn.f32 %f1025, %f1024, %f1019;
mul.rn.f32 %f1026, %f1025, %f1018;
sub.f32 %f1027, %f1014, %f1018;
add.f32 %f1028, %f1027, %f1027;
neg.f32 %f1029, %f1018;
fma.rn.f32 %f1030, %f1029, %f1014, %f1028;
mul.rn.f32 %f1031, %f1016, %f1030;
add.f32 %f1032, %f1026, %f1018;
sub.f32 %f1033, %f1018, %f1032;
add.f32 %f1034, %f1026, %f1033;
add.f32 %f1035, %f1031, %f1034;
add.f32 %f1036, %f1032, %f1035;
sub.f32 %f1037, %f1032, %f1036;
add.f32 %f1038, %f1035, %f1037;
mov.f32 %f1039, 0f3F317200;
mul.rn.f32 %f1040, %f1012, %f1039;
mov.f32 %f1041, 0f35BFBE8E;
mul.rn.f32 %f1042, %f1012, %f1041;
add.f32 %f1043, %f1040, %f1036;
sub.f32 %f1044, %f1040, %f1043;
add.f32 %f1045, %f1036, %f1044;
add.f32 %f1046, %f1038, %f1045;
add.f32 %f1047, %f1042, %f1046;
add.f32 %f1048, %f1043, %f1047;
sub.f32 %f1049, %f1043, %f1048;
add.f32 %f1050, %f1047, %f1049;
mov.f32 %f1051, 0f3EE66666;
mul.rn.f32 %f1052, %f1051, %f1048;
neg.f32 %f1053, %f1052;
fma.rn.f32 %f1054, %f1051, %f1048, %f1053;
fma.rn.f32 %f1055, %f1051, %f1050, %f1054;
mov.f32 %f1056, 0f00000000;
fma.rn.f32 %f1057, %f1056, %f1048, %f1055;
add.rn.f32 %f1058, %f1052, %f1057;
neg.f32 %f1059, %f1058;
add.rn.f32 %f1060, %f1052, %f1059;
add.rn.f32 %f1061, %f1060, %f1057;
mov.b32 %r332, %f1058;
setp.eq.s32 %p104, %r332, 1118925336;
add.s32 %r333, %r332, -1;
mov.b32 %f1062, %r333;
add.f32 %f1063, %f1061, 0f37000000;
selp.f32 %f235, %f1063, %f1061, %p104;
selp.f32 %f1064, %f1062, %f1058, %p104;
mov.f32 %f1065, 0f3FB8AA3B;
mul.rn.f32 %f1066, %f1064, %f1065;
cvt.rzi.f32.f32 %f1067, %f1066;
abs.f32 %f1068, %f1067;
setp.gt.f32 %p105, %f1068, 0f42FC0000;
mov.b32 %r334, %f1067;
and.b32 %r335, %r334, -2147483648;
or.b32 %r336, %r335, 1123811328;
mov.b32 %f1069, %r336;
selp.f32 %f1070, %f1069, %f1067, %p105;
mov.f32 %f1071, 0fBF317218;
fma.rn.f32 %f1072, %f1070, %f1071, %f1064;
mov.f32 %f1073, 0f3102E308;
fma.rn.f32 %f1074, %f1070, %f1073, %f1072;
mul.f32 %f1075, %f1074, 0f3FB8AA3B;
add.f32 %f1076, %f1070, 0f4B40007F;
mov.b32 %r337, %f1076;
shl.b32 %r338, %r337, 23;
mov.b32 %f1077, %r338;
ex2.approx.ftz.f32 %f1078, %f1075;
mul.f32 %f236, %f1078, %f1077;
setp.eq.f32 %p106, %f236, 0f7F800000;
mov.f32 %f1572, 0f7F800000;
@%p106 bra $L__BB0_86;
fma.rn.f32 %f1572, %f236, %f235, %f236;
$L__BB0_86:
setp.lt.f32 %p107, %f1533, 0f00000000;
and.pred %p6, %p107, %p78;
setp.eq.f32 %p109, %f1533, 0f00000000;
@%p109 bra $L__BB0_90;
bra.uni $L__BB0_87;
$L__BB0_90:
add.f32 %f1083, %f1533, %f1533;
selp.f32 %f1574, %f1083, 0f00000000, %p78;
bra.uni $L__BB0_91;
$L__BB0_87:
mov.b32 %r339, %f1572;
xor.b32 %r340, %r339, -2147483648;
mov.b32 %f1079, %r340;
selp.f32 %f1574, %f1079, %f1572, %p6;
setp.geu.f32 %p110, %f1533, 0f00000000;
@%p110 bra $L__BB0_91;
mov.f32 %f1080, 0f3EE66666;
cvt.rzi.f32.f32 %f1081, %f1080;
setp.eq.f32 %p111, %f1081, 0f3EE66666;
@%p111 bra $L__BB0_91;
mov.f32 %f1574, 0f7FFFFFFF;
$L__BB0_91:
add.f32 %f1084, %f234, 0f3EE66666;
mov.b32 %r341, %f1084;
setp.lt.s32 %p113, %r341, 2139095040;
@%p113 bra $L__BB0_96;
setp.gtu.f32 %p114, %f234, 0f7F800000;
@%p114 bra $L__BB0_95;
bra.uni $L__BB0_93;
$L__BB0_95:
add.f32 %f1574, %f1533, 0f3EE66666;
bra.uni $L__BB0_96;
$L__BB0_93:
setp.neu.f32 %p115, %f234, 0f7F800000;
@%p115 bra $L__BB0_96;
selp.f32 %f1574, 0fFF800000, 0f7F800000, %p6;
$L__BB0_96:
setp.eq.f32 %p116, %f1533, 0f3F800000;
mov.f32 %f1086, 0f3F800000;
selp.f32 %f1087, 0f3F800000, %f1574, %p116;
ld.const.u64 %rd36, [params+144];
cvta.to.global.u64 %rd9, %rd36;
ld.const.u32 %r342, [params+136];
mad.lo.s32 %r343, %r342, %r4, %r3;
cvt.u64.u32 %rd10, %r343;
min.f32 %f1088, %f221, %f1086;
mov.f32 %f1089, 0f00000000;
max.f32 %f245, %f1089, %f1088;
min.f32 %f1090, %f233, %f1086;
max.f32 %f246, %f1089, %f1090;
min.f32 %f1091, %f1087, %f1086;
max.f32 %f247, %f1089, %f1091;
mov.f32 %f1092, 0f3E555555;
cvt.rzi.f32.f32 %f1093, %f1092;
add.f32 %f1094, %f1093, %f1093;
mov.f32 %f1095, 0f3ED55555;
sub.f32 %f1096, %f1095, %f1094;
abs.f32 %f248, %f1096;
abs.f32 %f249, %f245;
setp.lt.f32 %p117, %f249, 0f00800000;
mul.f32 %f1097, %f249, 0f4B800000;
selp.f32 %f1098, %f1097, %f249, %p117;
selp.f32 %f1099, 0fC3170000, 0fC2FE0000, %p117;
mov.b32 %r344, %f1098;
and.b32 %r345, %r344, 8388607;
or.b32 %r346, %r345, 1065353216;
mov.b32 %f1100, %r346;
shr.u32 %r347, %r344, 23;
cvt.rn.f32.u32 %f1101, %r347;
add.f32 %f1102, %f1099, %f1101;
setp.gt.f32 %p118, %f1100, 0f3FB504F3;
mul.f32 %f1103, %f1100, 0f3F000000;
add.f32 %f1104, %f1102, 0f3F800000;
selp.f32 %f1105, %f1104, %f1102, %p118;
selp.f32 %f1106, %f1103, %f1100, %p118;
add.f32 %f1107, %f1106, 0fBF800000;
add.f32 %f1108, %f1106, 0f3F800000;
rcp.approx.ftz.f32 %f1109, %f1108;
add.f32 %f1110, %f1107, %f1107;
mul.f32 %f1111, %f1110, %f1109;
mul.f32 %f1112, %f1111, %f1111;
mov.f32 %f1113, 0f3C4CAF63;
mov.f32 %f1114, 0f3B18F0FE;
fma.rn.f32 %f1115, %f1114, %f1112, %f1113;
mov.f32 %f1116, 0f3DAAAABD;
fma.rn.f32 %f1117, %f1115, %f1112, %f1116;
mul.rn.f32 %f1118, %f1117, %f1112;
mul.rn.f32 %f1119, %f1118, %f1111;
sub.f32 %f1120, %f1107, %f1111;
add.f32 %f1121, %f1120, %f1120;
neg.f32 %f1122, %f1111;
fma.rn.f32 %f1123, %f1122, %f1107, %f1121;
mul.rn.f32 %f1124, %f1109, %f1123;
add.f32 %f1125, %f1119, %f1111;
sub.f32 %f1126, %f1111, %f1125;
add.f32 %f1127, %f1119, %f1126;
add.f32 %f1128, %f1124, %f1127;
add.f32 %f1129, %f1125, %f1128;
sub.f32 %f1130, %f1125, %f1129;
add.f32 %f1131, %f1128, %f1130;
mov.f32 %f1132, 0f3F317200;
mul.rn.f32 %f1133, %f1105, %f1132;
mov.f32 %f1134, 0f35BFBE8E;
mul.rn.f32 %f1135, %f1105, %f1134;
add.f32 %f1136, %f1133, %f1129;
sub.f32 %f1137, %f1133, %f1136;
add.f32 %f1138, %f1129, %f1137;
add.f32 %f1139, %f1131, %f1138;
add.f32 %f1140, %f1135, %f1139;
add.f32 %f1141, %f1136, %f1140;
sub.f32 %f1142, %f1136, %f1141;
add.f32 %f1143, %f1140, %f1142;
mul.rn.f32 %f1144, %f1095, %f1141;
neg.f32 %f1145, %f1144;
fma.rn.f32 %f1146, %f1095, %f1141, %f1145;
fma.rn.f32 %f1147, %f1095, %f1143, %f1146;
fma.rn.f32 %f1148, %f1089, %f1141, %f1147;
add.rn.f32 %f1149, %f1144, %f1148;
neg.f32 %f1150, %f1149;
add.rn.f32 %f1151, %f1144, %f1150;
add.rn.f32 %f1152, %f1151, %f1148;
mov.b32 %r348, %f1149;
setp.eq.s32 %p119, %r348, 1118925336;
add.s32 %r349, %r348, -1;
mov.b32 %f1153, %r349;
add.f32 %f1154, %f1152, 0f37000000;
selp.f32 %f250, %f1154, %f1152, %p119;
selp.f32 %f1155, %f1153, %f1149, %p119;
mov.f32 %f1156, 0f3FB8AA3B;
mul.rn.f32 %f1157, %f1155, %f1156;
cvt.rzi.f32.f32 %f1158, %f1157;
abs.f32 %f1159, %f1158;
setp.gt.f32 %p120, %f1159, 0f42FC0000;
mov.b32 %r350, %f1158;
and.b32 %r351, %r350, -2147483648;
or.b32 %r352, %r351, 1123811328;
mov.b32 %f1160, %r352;
selp.f32 %f1161, %f1160, %f1158, %p120;
mov.f32 %f1162, 0fBF317218;
fma.rn.f32 %f1163, %f1161, %f1162, %f1155;
mov.f32 %f1164, 0f3102E308;
fma.rn.f32 %f1165, %f1161, %f1164, %f1163;
mul.f32 %f1166, %f1165, 0f3FB8AA3B;
add.f32 %f1167, %f1161, 0f4B40007F;
mov.b32 %r353, %f1167;
shl.b32 %r354, %r353, 23;
mov.b32 %f1168, %r354;
ex2.approx.ftz.f32 %f1169, %f1166;
mul.f32 %f251, %f1169, %f1168;
setp.eq.f32 %p121, %f251, 0f7F800000;
mov.f32 %f1575, 0f7F800000;
@%p121 bra $L__BB0_98;
fma.rn.f32 %f1575, %f251, %f250, %f251;
$L__BB0_98:
setp.lt.f32 %p122, %f245, 0f00000000;
setp.eq.f32 %p123, %f248, 0f3F800000;
and.pred %p7, %p122, %p123;
setp.eq.f32 %p124, %f245, 0f00000000;
@%p124 bra $L__BB0_102;
bra.uni $L__BB0_99;
$L__BB0_102:
add.f32 %f1174, %f245, %f245;
selp.f32 %f1577, %f1174, 0f00000000, %p123;
bra.uni $L__BB0_103;
$L__BB0_99:
mov.b32 %r355, %f1575;
xor.b32 %r356, %r355, -2147483648;
mov.b32 %f1170, %r356;
selp.f32 %f1577, %f1170, %f1575, %p7;
setp.geu.f32 %p125, %f245, 0f00000000;
@%p125 bra $L__BB0_103;
mov.f32 %f1171, 0f3ED55555;
cvt.rzi.f32.f32 %f1172, %f1171;
setp.eq.f32 %p126, %f1172, 0f3ED55555;
@%p126 bra $L__BB0_103;
mov.f32 %f1577, 0f7FFFFFFF;
$L__BB0_103:
add.f32 %f1175, %f249, 0f3ED55555;
mov.b32 %r357, %f1175;
setp.lt.s32 %p128, %r357, 2139095040;
@%p128 bra $L__BB0_108;
setp.gtu.f32 %p129, %f249, 0f7F800000;
@%p129 bra $L__BB0_107;
bra.uni $L__BB0_105;
$L__BB0_107:
add.f32 %f1577, %f245, 0f3ED55555;
bra.uni $L__BB0_108;
$L__BB0_105:
setp.neu.f32 %p130, %f249, 0f7F800000;
@%p130 bra $L__BB0_108;
selp.f32 %f1577, 0fFF800000, 0f7F800000, %p7;
$L__BB0_108:
abs.f32 %f260, %f246;
setp.lt.f32 %p131, %f260, 0f00800000;
mul.f32 %f1177, %f260, 0f4B800000;
selp.f32 %f1178, %f1177, %f260, %p131;
selp.f32 %f1179, 0fC3170000, 0fC2FE0000, %p131;
mov.b32 %r358, %f1178;
and.b32 %r359, %r358, 8388607;
or.b32 %r360, %r359, 1065353216;
mov.b32 %f1180, %r360;
shr.u32 %r361, %r358, 23;
cvt.rn.f32.u32 %f1181, %r361;
add.f32 %f1182, %f1179, %f1181;
setp.gt.f32 %p132, %f1180, 0f3FB504F3;
mul.f32 %f1183, %f1180, 0f3F000000;
add.f32 %f1184, %f1182, 0f3F800000;
selp.f32 %f1185, %f1184, %f1182, %p132;
selp.f32 %f1186, %f1183, %f1180, %p132;
add.f32 %f1187, %f1186, 0fBF800000;
add.f32 %f1188, %f1186, 0f3F800000;
rcp.approx.ftz.f32 %f1189, %f1188;
add.f32 %f1190, %f1187, %f1187;
mul.f32 %f1191, %f1190, %f1189;
mul.f32 %f1192, %f1191, %f1191;
mov.f32 %f1193, 0f3C4CAF63;
mov.f32 %f1194, 0f3B18F0FE;
fma.rn.f32 %f1195, %f1194, %f1192, %f1193;
mov.f32 %f1196, 0f3DAAAABD;
fma.rn.f32 %f1197, %f1195, %f1192, %f1196;
mul.rn.f32 %f1198, %f1197, %f1192;
mul.rn.f32 %f1199, %f1198, %f1191;
sub.f32 %f1200, %f1187, %f1191;
add.f32 %f1201, %f1200, %f1200;
neg.f32 %f1202, %f1191;
fma.rn.f32 %f1203, %f1202, %f1187, %f1201;
mul.rn.f32 %f1204, %f1189, %f1203;
add.f32 %f1205, %f1199, %f1191;
sub.f32 %f1206, %f1191, %f1205;
add.f32 %f1207, %f1199, %f1206;
add.f32 %f1208, %f1204, %f1207;
add.f32 %f1209, %f1205, %f1208;
sub.f32 %f1210, %f1205, %f1209;
add.f32 %f1211, %f1208, %f1210;
mov.f32 %f1212, 0f3F317200;
mul.rn.f32 %f1213, %f1185, %f1212;
mov.f32 %f1214, 0f35BFBE8E;
mul.rn.f32 %f1215, %f1185, %f1214;
add.f32 %f1216, %f1213, %f1209;
sub.f32 %f1217, %f1213, %f1216;
add.f32 %f1218, %f1209, %f1217;
add.f32 %f1219, %f1211, %f1218;
add.f32 %f1220, %f1215, %f1219;
add.f32 %f1221, %f1216, %f1220;
sub.f32 %f1222, %f1216, %f1221;
add.f32 %f1223, %f1220, %f1222;
mov.f32 %f1224, 0f3ED55555;
mul.rn.f32 %f1225, %f1224, %f1221;
neg.f32 %f1226, %f1225;
fma.rn.f32 %f1227, %f1224, %f1221, %f1226;
fma.rn.f32 %f1228, %f1224, %f1223, %f1227;
mov.f32 %f1229, 0f00000000;
fma.rn.f32 %f1230, %f1229, %f1221, %f1228;
add.rn.f32 %f1231, %f1225, %f1230;
neg.f32 %f1232, %f1231;
add.rn.f32 %f1233, %f1225, %f1232;
add.rn.f32 %f1234, %f1233, %f1230;
mov.b32 %r362, %f1231;
setp.eq.s32 %p133, %r362, 1118925336;
add.s32 %r363, %r362, -1;
mov.b32 %f1235, %r363;
add.f32 %f1236, %f1234, 0f37000000;
selp.f32 %f261, %f1236, %f1234, %p133;
selp.f32 %f1237, %f1235, %f1231, %p133;
mov.f32 %f1238, 0f3FB8AA3B;
mul.rn.f32 %f1239, %f1237, %f1238;
cvt.rzi.f32.f32 %f1240, %f1239;
abs.f32 %f1241, %f1240;
setp.gt.f32 %p134, %f1241, 0f42FC0000;
mov.b32 %r364, %f1240;
and.b32 %r365, %r364, -2147483648;
or.b32 %r366, %r365, 1123811328;
mov.b32 %f1242, %r366;
selp.f32 %f1243, %f1242, %f1240, %p134;
mov.f32 %f1244, 0fBF317218;
fma.rn.f32 %f1245, %f1243, %f1244, %f1237;
mov.f32 %f1246, 0f3102E308;
fma.rn.f32 %f1247, %f1243, %f1246, %f1245;
mul.f32 %f1248, %f1247, 0f3FB8AA3B;
add.f32 %f1249, %f1243, 0f4B40007F;
mov.b32 %r367, %f1249;
shl.b32 %r368, %r367, 23;
mov.b32 %f1250, %r368;
ex2.approx.ftz.f32 %f1251, %f1248;
mul.f32 %f262, %f1251, %f1250;
setp.eq.f32 %p135, %f262, 0f7F800000;
mov.f32 %f1578, 0f7F800000;
@%p135 bra $L__BB0_110;
fma.rn.f32 %f1578, %f262, %f261, %f262;
$L__BB0_110:
setp.lt.f32 %p136, %f246, 0f00000000;
and.pred %p8, %p136, %p123;
setp.eq.f32 %p138, %f246, 0f00000000;
@%p138 bra $L__BB0_114;
bra.uni $L__BB0_111;
$L__BB0_114:
add.f32 %f1256, %f246, %f246;
selp.f32 %f1580, %f1256, 0f00000000, %p123;
bra.uni $L__BB0_115;
$L__BB0_111:
mov.b32 %r369, %f1578;
xor.b32 %r370, %r369, -2147483648;
mov.b32 %f1252, %r370;
selp.f32 %f1580, %f1252, %f1578, %p8;
setp.geu.f32 %p139, %f246, 0f00000000;
@%p139 bra $L__BB0_115;
mov.f32 %f1253, 0f3ED55555;
cvt.rzi.f32.f32 %f1254, %f1253;
setp.eq.f32 %p140, %f1254, 0f3ED55555;
@%p140 bra $L__BB0_115;
mov.f32 %f1580, 0f7FFFFFFF;
$L__BB0_115:
add.f32 %f1257, %f260, 0f3ED55555;
mov.b32 %r371, %f1257;
setp.lt.s32 %p142, %r371, 2139095040;
@%p142 bra $L__BB0_120;
setp.gtu.f32 %p143, %f260, 0f7F800000;
@%p143 bra $L__BB0_119;
bra.uni $L__BB0_117;
$L__BB0_119:
add.f32 %f1580, %f246, 0f3ED55555;
bra.uni $L__BB0_120;
$L__BB0_117:
setp.neu.f32 %p144, %f260, 0f7F800000;
@%p144 bra $L__BB0_120;
selp.f32 %f1580, 0fFF800000, 0f7F800000, %p8;
$L__BB0_120:
abs.f32 %f271, %f247;
setp.lt.f32 %p145, %f271, 0f00800000;
mul.f32 %f1259, %f271, 0f4B800000;
selp.f32 %f1260, %f1259, %f271, %p145;
selp.f32 %f1261, 0fC3170000, 0fC2FE0000, %p145;
mov.b32 %r372, %f1260;
and.b32 %r373, %r372, 8388607;
or.b32 %r374, %r373, 1065353216;
mov.b32 %f1262, %r374;
shr.u32 %r375, %r372, 23;
cvt.rn.f32.u32 %f1263, %r375;
add.f32 %f1264, %f1261, %f1263;
setp.gt.f32 %p146, %f1262, 0f3FB504F3;
mul.f32 %f1265, %f1262, 0f3F000000;
add.f32 %f1266, %f1264, 0f3F800000;
selp.f32 %f1267, %f1266, %f1264, %p146;
selp.f32 %f1268, %f1265, %f1262, %p146;
add.f32 %f1269, %f1268, 0fBF800000;
add.f32 %f1270, %f1268, 0f3F800000;
rcp.approx.ftz.f32 %f1271, %f1270;
add.f32 %f1272, %f1269, %f1269;
mul.f32 %f1273, %f1272, %f1271;
mul.f32 %f1274, %f1273, %f1273;
mov.f32 %f1275, 0f3C4CAF63;
mov.f32 %f1276, 0f3B18F0FE;
fma.rn.f32 %f1277, %f1276, %f1274, %f1275;
mov.f32 %f1278, 0f3DAAAABD;
fma.rn.f32 %f1279, %f1277, %f1274, %f1278;
mul.rn.f32 %f1280, %f1279, %f1274;
mul.rn.f32 %f1281, %f1280, %f1273;
sub.f32 %f1282, %f1269, %f1273;
add.f32 %f1283, %f1282, %f1282;
neg.f32 %f1284, %f1273;
fma.rn.f32 %f1285, %f1284, %f1269, %f1283;
mul.rn.f32 %f1286, %f1271, %f1285;
add.f32 %f1287, %f1281, %f1273;
sub.f32 %f1288, %f1273, %f1287;
add.f32 %f1289, %f1281, %f1288;
add.f32 %f1290, %f1286, %f1289;
add.f32 %f1291, %f1287, %f1290;
sub.f32 %f1292, %f1287, %f1291;
add.f32 %f1293, %f1290, %f1292;
mov.f32 %f1294, 0f3F317200;
mul.rn.f32 %f1295, %f1267, %f1294;
mov.f32 %f1296, 0f35BFBE8E;
mul.rn.f32 %f1297, %f1267, %f1296;
add.f32 %f1298, %f1295, %f1291;
sub.f32 %f1299, %f1295, %f1298;
add.f32 %f1300, %f1291, %f1299;
add.f32 %f1301, %f1293, %f1300;
add.f32 %f1302, %f1297, %f1301;
add.f32 %f1303, %f1298, %f1302;
sub.f32 %f1304, %f1298, %f1303;
add.f32 %f1305, %f1302, %f1304;
mov.f32 %f1306, 0f3ED55555;
mul.rn.f32 %f1307, %f1306, %f1303;
neg.f32 %f1308, %f1307;
fma.rn.f32 %f1309, %f1306, %f1303, %f1308;
fma.rn.f32 %f1310, %f1306, %f1305, %f1309;
mov.f32 %f1311, 0f00000000;
fma.rn.f32 %f1312, %f1311, %f1303, %f1310;
add.rn.f32 %f1313, %f1307, %f1312;
neg.f32 %f1314, %f1313;
add.rn.f32 %f1315, %f1307, %f1314;
add.rn.f32 %f1316, %f1315, %f1312;
mov.b32 %r376, %f1313;
setp.eq.s32 %p147, %r376, 1118925336;
add.s32 %r377, %r376, -1;
mov.b32 %f1317, %r377;
add.f32 %f1318, %f1316, 0f37000000;
selp.f32 %f272, %f1318, %f1316, %p147;
selp.f32 %f1319, %f1317, %f1313, %p147;
mov.f32 %f1320, 0f3FB8AA3B;
mul.rn.f32 %f1321, %f1319, %f1320;
cvt.rzi.f32.f32 %f1322, %f1321;
abs.f32 %f1323, %f1322;
setp.gt.f32 %p148, %f1323, 0f42FC0000;
mov.b32 %r378, %f1322;
and.b32 %r379, %r378, -2147483648;
or.b32 %r380, %r379, 1123811328;
mov.b32 %f1324, %r380;
selp.f32 %f1325, %f1324, %f1322, %p148;
mov.f32 %f1326, 0fBF317218;
fma.rn.f32 %f1327, %f1325, %f1326, %f1319;
mov.f32 %f1328, 0f3102E308;
fma.rn.f32 %f1329, %f1325, %f1328, %f1327;
mul.f32 %f1330, %f1329, 0f3FB8AA3B;
add.f32 %f1331, %f1325, 0f4B40007F;
mov.b32 %r381, %f1331;
shl.b32 %r382, %r381, 23;
mov.b32 %f1332, %r382;
ex2.approx.ftz.f32 %f1333, %f1330;
mul.f32 %f273, %f1333, %f1332;
setp.eq.f32 %p149, %f273, 0f7F800000;
mov.f32 %f1581, 0f7F800000;
@%p149 bra $L__BB0_122;
fma.rn.f32 %f1581, %f273, %f272, %f273;
$L__BB0_122:
setp.lt.f32 %p150, %f247, 0f00000000;
and.pred %p9, %p150, %p123;
setp.eq.f32 %p152, %f247, 0f00000000;
@%p152 bra $L__BB0_126;
bra.uni $L__BB0_123;
$L__BB0_126:
add.f32 %f1338, %f247, %f247;
selp.f32 %f1583, %f1338, 0f00000000, %p123;
bra.uni $L__BB0_127;
$L__BB0_123:
mov.b32 %r383, %f1581;
xor.b32 %r384, %r383, -2147483648;
mov.b32 %f1334, %r384;
selp.f32 %f1583, %f1334, %f1581, %p9;
setp.geu.f32 %p153, %f247, 0f00000000;
@%p153 bra $L__BB0_127;
mov.f32 %f1335, 0f3ED55555;
cvt.rzi.f32.f32 %f1336, %f1335;
setp.eq.f32 %p154, %f1336, 0f3ED55555;
@%p154 bra $L__BB0_127;
mov.f32 %f1583, 0f7FFFFFFF;
$L__BB0_127:
add.f32 %f1339, %f271, 0f3ED55555;
mov.b32 %r385, %f1339;
setp.lt.s32 %p156, %r385, 2139095040;
@%p156 bra $L__BB0_132;
setp.gtu.f32 %p157, %f271, 0f7F800000;
@%p157 bra $L__BB0_131;
bra.uni $L__BB0_129;
$L__BB0_131:
add.f32 %f1583, %f247, 0f3ED55555;
bra.uni $L__BB0_132;
$L__BB0_129:
setp.neu.f32 %p158, %f271, 0f7F800000;
@%p158 bra $L__BB0_132;
selp.f32 %f1583, 0fFF800000, 0f7F800000, %p9;
$L__BB0_132:
fma.rn.f32 %f1340, %f1577, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p159, %f245, 0f3F800000;
mov.f32 %f1341, 0f3F800000;
selp.f32 %f1342, 0f3F7FFFFF, %f1340, %p159;
mul.f32 %f1343, %f245, 0f414EB852;
setp.lt.f32 %p160, %f245, 0f3B4D2E1C;
selp.f32 %f1344, %f1343, %f1342, %p160;
fma.rn.f32 %f1345, %f1580, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p161, %f246, 0f3F800000;
selp.f32 %f1346, 0f3F7FFFFF, %f1345, %p161;
mul.f32 %f1347, %f246, 0f414EB852;
setp.lt.f32 %p162, %f246, 0f3B4D2E1C;
selp.f32 %f1348, %f1347, %f1346, %p162;
fma.rn.f32 %f1349, %f1583, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p163, %f247, 0f3F800000;
selp.f32 %f1350, 0f3F7FFFFF, %f1349, %p163;
mul.f32 %f1351, %f247, 0f414EB852;
setp.lt.f32 %p164, %f247, 0f3B4D2E1C;
selp.f32 %f1352, %f1351, %f1350, %p164;
min.f32 %f1353, %f1344, %f1341;
mov.f32 %f1354, 0f00000000;
max.f32 %f1355, %f1354, %f1353;
mul.f32 %f1356, %f1355, 0f43800000;
cvt.rzi.u32.f32 %r386, %f1356;
min.u32 %r387, %r386, 255;
min.f32 %f1357, %f1348, %f1341;
max.f32 %f1358, %f1354, %f1357;
mul.f32 %f1359, %f1358, 0f43800000;
cvt.rzi.u32.f32 %r388, %f1359;
min.u32 %r389, %r388, 255;
min.f32 %f1360, %f1352, %f1341;
max.f32 %f1361, %f1354, %f1360;
mul.f32 %f1362, %f1361, 0f43800000;
cvt.rzi.u32.f32 %r390, %f1362;
min.u32 %r391, %r390, 255;
shl.b64 %rd37, %rd10, 2;
add.s64 %rd38, %rd9, %rd37;
cvt.u16.u32 %rs20, %r391;
cvt.u16.u32 %rs21, %r389;
cvt.u16.u32 %rs22, %r387;
mov.u16 %rs23, 255;
st.global.v4.u8 [%rd38], {%rs22, %rs21, %rs20, %rs23};
$L__BB0_133:
and.b32 %r392, %r427, 4;
setp.eq.s32 %p165, %r392, 0;
ld.const.u32 %r428, [params+108];
@%p165 bra $L__BB0_137;
setp.eq.s32 %p166, %r428, 0;
ld.const.u64 %rd39, [params+224];
cvta.to.global.u64 %rd40, %rd39;
ld.const.u32 %r393, [params+216];
mad.lo.s32 %r394, %r393, %r4, %r3;
mul.wide.u32 %rd41, %r394, 8;
add.s64 %rd11, %rd40, %rd41;
@%p166 bra $L__BB0_136;
ld.global.v4.u16 {%rs31, %rs32, %rs33, %rs34}, [%rd11];
// begin inline asm
{ cvt.f32.f16 %f1363, %rs31;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1364, %rs32;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1365, %rs33;}
// end inline asm
add.f32 %f1366, %f1531, %f1363;
add.f32 %f1367, %f1532, %f1364;
add.f32 %f1368, %f1533, %f1365;
mov.f32 %f1369, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f1368;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1367;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1366;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f1369;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs27, %rs28, %rs29, %rs30};
bra.uni $L__BB0_137;
$L__BB0_136:
mov.f32 %f1373, 0f3F800000;
mov.u32 %r428, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f1373;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f1533;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f1532;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f1531;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs35, %rs36, %rs37, %rs38};
$L__BB0_137:
mov.f32 %f1374, 0f34000000;
max.f32 %f1375, %f1534, %f1374;
div.rn.f32 %f1376, %f1537, %f1375;
max.f32 %f1377, %f1535, %f1374;
div.rn.f32 %f1378, %f1538, %f1377;
max.f32 %f1379, %f1536, %f1374;
div.rn.f32 %f1380, %f1539, %f1379;
fma.rn.f32 %f282, %f1376, 0f3F000000, 0f3F000000;
fma.rn.f32 %f283, %f1378, 0f3F000000, 0f3F000000;
fma.rn.f32 %f284, %f1380, 0f3F000000, 0f3F000000;
div.rn.f32 %f1381, %f1540, %f1375;
div.rn.f32 %f1382, %f1541, %f1377;
div.rn.f32 %f1383, %f1542, %f1379;
fma.rn.f32 %f285, %f1381, 0f3F000000, 0f3F000000;
fma.rn.f32 %f286, %f1382, 0f3F000000, 0f3F000000;
fma.rn.f32 %f287, %f1383, 0f3F000000, 0f3F000000;
div.rn.f32 %f1384, %f1543, %f1375;
div.rn.f32 %f1385, %f1544, %f1377;
div.rn.f32 %f1386, %f1545, %f1379;
fma.rn.f32 %f288, %f1384, 0f3F000000, 0f3F000000;
fma.rn.f32 %f289, %f1385, 0f3F000000, 0f3F000000;
fma.rn.f32 %f290, %f1386, 0f3F000000, 0f3F000000;
ld.const.u64 %rd42, [params+256];
cvta.to.global.u64 %rd43, %rd42;
ld.const.u32 %r396, [params+248];
mad.lo.s32 %r397, %r396, %r4, %r3;
mul.wide.u32 %rd44, %r397, 8;
add.s64 %rd12, %rd43, %rd44;
setp.eq.s32 %p167, %r428, 0;
@%p167 bra $L__BB0_139;
ld.global.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd12];
// begin inline asm
{ cvt.f32.f16 %f1387, %rs46;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1388, %rs47;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1389, %rs48;}
// end inline asm
add.f32 %f1390, %f1534, %f1387;
add.f32 %f1391, %f1535, %f1388;
add.f32 %f1392, %f1536, %f1389;
mov.f32 %f1393, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f1392;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f1391;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs42, %f1390;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f1393;}
// end inline asm
st.global.v4.u16 [%rd12], {%rs42, %rs43, %rs44, %rs45};
bra.uni $L__BB0_140;
$L__BB0_139:
mov.f32 %f1397, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs53, %f1397;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs52, %f1536;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f1535;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f1534;}
// end inline asm
st.global.v4.u16 [%rd12], {%rs50, %rs51, %rs52, %rs53};
$L__BB0_140:
ld.const.u64 %rd45, [params+272];
cvta.to.global.u64 %rd46, %rd45;
ld.const.u32 %r398, [params+264];
mad.lo.s32 %r399, %r398, %r4, %r3;
mul.wide.u32 %rd47, %r399, 8;
add.s64 %rd13, %rd46, %rd47;
@%p167 bra $L__BB0_142;
ld.global.v4.u16 {%rs61, %rs62, %rs63, %rs64}, [%rd13];
// begin inline asm
{ cvt.f32.f16 %f1398, %rs61;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1399, %rs62;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1400, %rs63;}
// end inline asm
add.f32 %f1401, %f282, %f1398;
add.f32 %f1402, %f283, %f1399;
add.f32 %f1403, %f284, %f1400;
mov.f32 %f1404, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs59, %f1403;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs58, %f1402;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs57, %f1401;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs60, %f1404;}
// end inline asm
st.global.v4.u16 [%rd13], {%rs57, %rs58, %rs59, %rs60};
bra.uni $L__BB0_143;
$L__BB0_142:
mov.f32 %f1408, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs68, %f1408;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs67, %f284;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs66, %f283;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs65, %f282;}
// end inline asm
st.global.v4.u16 [%rd13], {%rs65, %rs66, %rs67, %rs68};
$L__BB0_143:
ld.const.u64 %rd48, [params+288];
cvta.to.global.u64 %rd49, %rd48;
ld.const.u32 %r400, [params+280];
mad.lo.s32 %r401, %r400, %r4, %r3;
mul.wide.u32 %rd50, %r401, 8;
add.s64 %rd14, %rd49, %rd50;
@%p167 bra $L__BB0_145;
ld.global.v4.u16 {%rs76, %rs77, %rs78, %rs79}, [%rd14];
// begin inline asm
{ cvt.f32.f16 %f1409, %rs76;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1410, %rs77;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1411, %rs78;}
// end inline asm
add.f32 %f1412, %f285, %f1409;
add.f32 %f1413, %f286, %f1410;
add.f32 %f1414, %f287, %f1411;
mov.f32 %f1415, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs74, %f1414;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs73, %f1413;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs72, %f1412;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs75, %f1415;}
// end inline asm
st.global.v4.u16 [%rd14], {%rs72, %rs73, %rs74, %rs75};
bra.uni $L__BB0_146;
$L__BB0_145:
mov.f32 %f1419, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs83, %f1419;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs82, %f287;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs81, %f286;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs80, %f285;}
// end inline asm
st.global.v4.u16 [%rd14], {%rs80, %rs81, %rs82, %rs83};
$L__BB0_146:
ld.const.u64 %rd51, [params+304];
cvta.to.global.u64 %rd52, %rd51;
ld.const.u32 %r402, [params+296];
mad.lo.s32 %r403, %r402, %r4, %r3;
mul.wide.u32 %rd53, %r403, 8;
add.s64 %rd15, %rd52, %rd53;
@%p167 bra $L__BB0_148;
ld.global.v4.u16 {%rs91, %rs92, %rs93, %rs94}, [%rd15];
// begin inline asm
{ cvt.f32.f16 %f1420, %rs91;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1421, %rs92;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1422, %rs93;}
// end inline asm
add.f32 %f1423, %f288, %f1420;
add.f32 %f1424, %f289, %f1421;
add.f32 %f1425, %f290, %f1422;
mov.f32 %f1426, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs89, %f1425;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs88, %f1424;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs87, %f1423;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs90, %f1426;}
// end inline asm
st.global.v4.u16 [%rd15], {%rs87, %rs88, %rs89, %rs90};
bra.uni $L__BB0_169;
$L__BB0_148:
mov.f32 %f1430, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs98, %f1430;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs97, %f290;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs96, %f289;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs95, %f288;}
// end inline asm
st.global.v4.u16 [%rd15], {%rs95, %rs96, %rs97, %rs98};
$L__BB0_169:
ret;
}