Files
FORT-BO/Assets/Editor/x64/Bakery/optix9/lmBatchPointLightProbeSHL2Translucent.ptx
2026-08-06 18:56:32 +03:00

3330 lines
90 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<193>;
.reg .b16 %rs<330>;
.reg .f32 %f<1923>;
.reg .b32 %r<450>;
.reg .b64 %rd<118>;
// 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 %rd31, [params+400];
cvta.to.global.u64 %rd32, %rd31;
ld.const.u32 %r32, [params+392];
mad.lo.s32 %r33, %r32, %r27, %r26;
mul.wide.u32 %rd33, %r33, 4;
add.s64 %rd1, %rd32, %rd33;
ld.global.v2.u8 {%rs9, %rs328}, [%rd1];
or.b16 %rs11, %rs9, %rs328;
and.b16 %rs12, %rs11, 255;
setp.eq.s16 %p10, %rs12, 0;
@%p10 bra $L__BB0_2;
ld.global.u8 %rs327, [%rd1+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs327, [%rd1+2];
setp.eq.s16 %p11, %rs327, 0;
mov.f32 %f1749, 0f00000000;
mov.u16 %rs328, 0;
mov.f32 %f1750, %f1749;
mov.f32 %f1751, %f1749;
@%p11 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f405, %rs9;
div.rn.f32 %f406, %f405, 0f437F0000;
fma.rn.f32 %f407, %f406, 0f40000000, 0fBF800000;
and.b16 %rs15, %rs328, 255;
cvt.rn.f32.u16 %f408, %rs15;
div.rn.f32 %f409, %f408, 0f437F0000;
fma.rn.f32 %f410, %f409, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f411, %rs327;
div.rn.f32 %f412, %f411, 0f437F0000;
fma.rn.f32 %f413, %f412, 0f40000000, 0fBF800000;
mul.f32 %f414, %f410, %f410;
fma.rn.f32 %f415, %f407, %f407, %f414;
fma.rn.f32 %f416, %f413, %f413, %f415;
sqrt.rn.f32 %f417, %f416;
rcp.rn.f32 %f418, %f417;
mul.f32 %f1751, %f418, %f413;
mul.f32 %f1750, %f418, %f410;
mul.f32 %f1749, %f407, %f418;
$L__BB0_4:
ld.const.v2.u32 {%r34, %r35}, [params];
add.s32 %r3, %r34, %r26;
add.s32 %r4, %r35, %r27;
setp.eq.f32 %p12, %f1749, 0f00000000;
setp.eq.f32 %p13, %f1750, 0f00000000;
and.pred %p14, %p12, %p13;
setp.eq.f32 %p15, %f1751, 0f00000000;
and.pred %p16, %p15, %p14;
@%p16 bra $L__BB0_164;
bra.uni $L__BB0_5;
$L__BB0_164:
ld.const.u32 %r23, [params+104];
and.b32 %r414, %r23, 1;
setp.eq.b32 %p176, %r414, 1;
mov.pred %p177, 0;
xor.pred %p178, %p176, %p177;
not.pred %p179, %p178;
@%p179 bra $L__BB0_166;
ld.const.u64 %rd79, [params+144];
cvta.to.global.u64 %rd80, %rd79;
ld.const.u32 %r415, [params+136];
mad.lo.s32 %r416, %r415, %r4, %r3;
mul.wide.u32 %rd81, %r416, 4;
add.s64 %rd82, %rd80, %rd81;
mov.u16 %rs174, 0;
st.global.v4.u8 [%rd82], {%rs174, %rs174, %rs174, %rs174};
$L__BB0_166:
and.b32 %r417, %r23, 8;
setp.eq.s32 %p180, %r417, 0;
@%p180 bra $L__BB0_168;
ld.const.u64 %rd83, [params+192];
cvta.to.global.u64 %rd84, %rd83;
ld.const.u32 %r418, [params+184];
mad.lo.s32 %r419, %r418, %r4, %r3;
mov.f32 %f1655, 0f00000000;
cvt.rzi.u32.f32 %r420, %f1655;
mul.wide.u32 %rd85, %r419, 2;
add.s64 %rd86, %rd84, %rd85;
mov.u16 %rs175, 0;
cvt.u16.u32 %rs176, %r420;
st.global.v2.u8 [%rd86], {%rs176, %rs175};
$L__BB0_168:
and.b32 %r421, %r23, 4;
setp.eq.s32 %p181, %r421, 0;
ld.const.u32 %r449, [params+108];
@%p181 bra $L__BB0_172;
setp.eq.s32 %p182, %r449, 0;
ld.const.u64 %rd87, [params+224];
cvta.to.global.u64 %rd88, %rd87;
ld.const.u32 %r422, [params+216];
mad.lo.s32 %r423, %r422, %r4, %r3;
mul.wide.u32 %rd89, %r423, 8;
add.s64 %rd21, %rd88, %rd89;
@%p182 bra $L__BB0_171;
ld.global.v4.u16 {%rs183, %rs184, %rs185, %rs186}, [%rd21];
// begin inline asm
{ cvt.f32.f16 %f1656, %rs183;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1657, %rs184;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1658, %rs185;}
// end inline asm
add.f32 %f1659, %f1656, 0f00000000;
add.f32 %f1660, %f1657, 0f00000000;
add.f32 %f1661, %f1658, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs182, %f1661;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs181, %f1660;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs180, %f1659;}
// end inline asm
mov.u16 %rs187, 0;
st.global.v4.u16 [%rd21], {%rs180, %rs181, %rs182, %rs187};
bra.uni $L__BB0_172;
$L__BB0_5:
ld.const.u64 %rd34, [params+432];
cvta.to.global.u64 %rd35, %rd34;
ld.const.u32 %r38, [params+424];
mad.lo.s32 %r39, %r38, %r27, %r26;
mul.wide.u32 %rd36, %r39, 12;
add.s64 %rd37, %rd35, %rd36;
ld.global.f32 %f10, [%rd37];
mul.f32 %f419, %f10, 0f3456BF95;
ld.global.f32 %f11, [%rd37+4];
mul.f32 %f420, %f11, 0f3456BF95;
ld.global.f32 %f12, [%rd37+8];
mul.f32 %f421, %f12, 0f3456BF95;
abs.f32 %f422, %f1749;
div.rn.f32 %f423, %f419, %f422;
abs.f32 %f424, %f1750;
div.rn.f32 %f425, %f420, %f424;
abs.f32 %f426, %f1751;
div.rn.f32 %f427, %f421, %f426;
abs.f32 %f428, %f423;
abs.f32 %f429, %f425;
abs.f32 %f430, %f427;
mov.f32 %f431, 0f38D1B717;
max.f32 %f432, %f428, %f431;
max.f32 %f433, %f429, %f431;
max.f32 %f434, %f430, %f431;
fma.rn.f32 %f13, %f1749, %f432, %f10;
fma.rn.f32 %f14, %f1750, %f433, %f11;
fma.rn.f32 %f15, %f1751, %f434, %f12;
cvt.rn.f32.u32 %f16, %r26;
cvt.rn.f32.u32 %f17, %r27;
ld.const.u32 %r5, [params+600];
setp.eq.s32 %p17, %r5, 0;
@%p17 bra $L__BB0_44;
ld.const.u64 %rd38, [params+608];
cvta.to.global.u64 %rd2, %rd38;
mov.f32 %f466, 0f40000000;
cvt.rzi.f32.f32 %f467, %f466;
add.f32 %f468, %f467, %f467;
mov.f32 %f469, 0f40800000;
sub.f32 %f470, %f469, %f468;
abs.f32 %f18, %f470;
ld.const.u32 %r6, [params+544];
ld.const.u32 %r447, [params+104];
mul.f32 %f19, %f13, 0f3456BF95;
mul.f32 %f20, %f14, 0f3456BF95;
mul.f32 %f21, %f15, 0f3456BF95;
mov.f32 %f465, 0f00000000;
mov.u32 %r443, 0;
ld.const.u64 %rd3, [params+96];
abs.f32 %f696, %f19;
abs.f32 %f697, %f20;
max.f32 %f698, %f696, %f697;
abs.f32 %f699, %f21;
max.f32 %f700, %f698, %f699;
mov.f32 %f1870, %f465;
mov.f32 %f1869, %f465;
mov.f32 %f1868, %f465;
mov.f32 %f1867, %f465;
mov.f32 %f1866, %f465;
mov.f32 %f1865, %f465;
mov.f32 %f1864, %f465;
mov.f32 %f1863, %f465;
mov.f32 %f1862, %f465;
mov.f32 %f1861, %f465;
mov.f32 %f1860, %f465;
mov.f32 %f1859, %f465;
mov.f32 %f1858, %f465;
mov.f32 %f1857, %f465;
mov.f32 %f1856, %f465;
mov.f32 %f1855, %f465;
mov.f32 %f1854, %f465;
mov.f32 %f1853, %f465;
mov.f32 %f1852, %f465;
mov.f32 %f1851, %f465;
mov.f32 %f1850, %f465;
mov.f32 %f1849, %f465;
mov.f32 %f1848, %f465;
mov.f32 %f1847, %f465;
mov.f32 %f1846, %f465;
mov.f32 %f1845, %f465;
mov.f32 %f1844, %f465;
mov.f32 %f1843, %f465;
mov.f32 %f1842, %f465;
mov.f32 %f1841, %f465;
mov.f32 %f1840, %f465;
$L__BB0_7:
mul.wide.u32 %rd39, %r443, 112;
add.s64 %rd5, %rd2, %rd39;
ld.global.v4.f32 {%f472, %f473, %f474, %f475}, [%rd5];
ld.global.v4.f32 {%f479, %f480, %f481, %f482}, [%rd5+16];
ld.global.v4.f32 {%f485, %f1791, %f1792, %f488}, [%rd5+32];
ld.global.v4.f32 {%f489, %f490, %f491, %f1789}, [%rd5+48];
ld.global.v4.f32 {%f493, %f494, %f495, %f496}, [%rd5+64];
ld.global.v4.u32 {%r41, %r42, %r43, %r44}, [%rd5+80];
mov.b32 %f66, %r41;
mov.b32 %f67, %r42;
mov.b32 %f68, %r43;
ld.global.u64 %rd6, [%rd5+96];
sub.f32 %f69, %f473, %f10;
sub.f32 %f70, %f474, %f11;
sub.f32 %f71, %f475, %f12;
mul.f32 %f497, %f70, %f70;
fma.rn.f32 %f498, %f69, %f69, %f497;
fma.rn.f32 %f499, %f71, %f71, %f498;
sqrt.rn.f32 %f500, %f499;
rcp.rn.f32 %f501, %f500;
mul.f32 %f72, %f69, %f501;
mul.f32 %f73, %f70, %f501;
mul.f32 %f74, %f71, %f501;
mul.f32 %f75, %f500, %f480;
mul.f32 %f76, %f500, %f482;
abs.f32 %f77, %f76;
setp.lt.f32 %p18, %f77, 0f00800000;
mul.f32 %f502, %f77, 0f4B800000;
selp.f32 %f503, %f502, %f77, %p18;
selp.f32 %f504, 0fC3170000, 0fC2FE0000, %p18;
mov.b32 %r48, %f503;
and.b32 %r49, %r48, 8388607;
or.b32 %r50, %r49, 1065353216;
mov.b32 %f505, %r50;
shr.u32 %r51, %r48, 23;
cvt.rn.f32.u32 %f506, %r51;
add.f32 %f507, %f504, %f506;
setp.gt.f32 %p19, %f505, 0f3FB504F3;
mul.f32 %f508, %f505, 0f3F000000;
add.f32 %f509, %f507, 0f3F800000;
selp.f32 %f510, %f509, %f507, %p19;
selp.f32 %f511, %f508, %f505, %p19;
add.f32 %f512, %f511, 0fBF800000;
add.f32 %f513, %f511, 0f3F800000;
rcp.approx.ftz.f32 %f514, %f513;
add.f32 %f515, %f512, %f512;
mul.f32 %f516, %f515, %f514;
mul.f32 %f517, %f516, %f516;
mov.f32 %f518, 0f3C4CAF63;
mov.f32 %f519, 0f3B18F0FE;
fma.rn.f32 %f520, %f519, %f517, %f518;
mov.f32 %f521, 0f3DAAAABD;
fma.rn.f32 %f522, %f520, %f517, %f521;
mul.rn.f32 %f523, %f522, %f517;
mul.rn.f32 %f524, %f523, %f516;
sub.f32 %f525, %f512, %f516;
add.f32 %f526, %f525, %f525;
neg.f32 %f527, %f516;
fma.rn.f32 %f528, %f527, %f512, %f526;
mul.rn.f32 %f529, %f514, %f528;
add.f32 %f530, %f524, %f516;
sub.f32 %f531, %f516, %f530;
add.f32 %f532, %f524, %f531;
add.f32 %f533, %f529, %f532;
add.f32 %f534, %f530, %f533;
sub.f32 %f535, %f530, %f534;
add.f32 %f536, %f533, %f535;
mov.f32 %f537, 0f3F317200;
mul.rn.f32 %f538, %f510, %f537;
mov.f32 %f539, 0f35BFBE8E;
mul.rn.f32 %f540, %f510, %f539;
add.f32 %f541, %f538, %f534;
sub.f32 %f542, %f538, %f541;
add.f32 %f543, %f534, %f542;
add.f32 %f544, %f536, %f543;
add.f32 %f545, %f540, %f544;
add.f32 %f546, %f541, %f545;
sub.f32 %f547, %f541, %f546;
add.f32 %f548, %f545, %f547;
mul.rn.f32 %f550, %f469, %f546;
neg.f32 %f551, %f550;
fma.rn.f32 %f552, %f469, %f546, %f551;
fma.rn.f32 %f553, %f469, %f548, %f552;
fma.rn.f32 %f555, %f465, %f546, %f553;
add.rn.f32 %f556, %f550, %f555;
neg.f32 %f557, %f556;
add.rn.f32 %f558, %f550, %f557;
add.rn.f32 %f559, %f558, %f555;
mov.b32 %r52, %f556;
setp.eq.s32 %p20, %r52, 1118925336;
add.s32 %r53, %r52, -1;
mov.b32 %f560, %r53;
add.f32 %f561, %f559, 0f37000000;
selp.f32 %f78, %f561, %f559, %p20;
selp.f32 %f562, %f560, %f556, %p20;
mov.f32 %f563, 0f3FB8AA3B;
mul.rn.f32 %f564, %f562, %f563;
cvt.rzi.f32.f32 %f565, %f564;
abs.f32 %f566, %f565;
setp.gt.f32 %p21, %f566, 0f42FC0000;
mov.b32 %r54, %f565;
and.b32 %r55, %r54, -2147483648;
or.b32 %r56, %r55, 1123811328;
mov.b32 %f567, %r56;
selp.f32 %f568, %f567, %f565, %p21;
mov.f32 %f569, 0fBF317218;
fma.rn.f32 %f570, %f568, %f569, %f562;
mov.f32 %f571, 0f3102E308;
fma.rn.f32 %f572, %f568, %f571, %f570;
mul.f32 %f573, %f572, 0f3FB8AA3B;
add.f32 %f574, %f568, 0f4B40007F;
mov.b32 %r57, %f574;
shl.b32 %r58, %r57, 23;
mov.b32 %f575, %r58;
ex2.approx.ftz.f32 %f576, %f573;
mul.f32 %f79, %f576, %f575;
setp.eq.f32 %p22, %f79, 0f7F800000;
mov.f32 %f1783, 0f7F800000;
@%p22 bra $L__BB0_9;
fma.rn.f32 %f1783, %f79, %f78, %f79;
$L__BB0_9:
setp.lt.f32 %p23, %f76, 0f00000000;
setp.eq.f32 %p24, %f18, 0f3F800000;
and.pred %p1, %p23, %p24;
setp.eq.f32 %p25, %f76, 0f00000000;
@%p25 bra $L__BB0_13;
bra.uni $L__BB0_10;
$L__BB0_13:
add.f32 %f581, %f76, %f76;
selp.f32 %f1785, %f581, 0f00000000, %p24;
bra.uni $L__BB0_14;
$L__BB0_10:
mov.b32 %r59, %f1783;
xor.b32 %r60, %r59, -2147483648;
mov.b32 %f577, %r60;
selp.f32 %f1785, %f577, %f1783, %p1;
setp.geu.f32 %p26, %f76, 0f00000000;
@%p26 bra $L__BB0_14;
mov.f32 %f578, 0f40800000;
cvt.rzi.f32.f32 %f579, %f578;
setp.eq.f32 %p27, %f579, 0f40800000;
@%p27 bra $L__BB0_14;
mov.f32 %f1785, 0f7FFFFFFF;
$L__BB0_14:
abs.f32 %f1746, %f76;
add.f32 %f582, %f1746, 0f40800000;
mov.b32 %r61, %f582;
setp.lt.s32 %p29, %r61, 2139095040;
@%p29 bra $L__BB0_19;
abs.f32 %f1747, %f76;
setp.gtu.f32 %p30, %f1747, 0f7F800000;
@%p30 bra $L__BB0_18;
bra.uni $L__BB0_16;
$L__BB0_18:
add.f32 %f1785, %f76, 0f40800000;
bra.uni $L__BB0_19;
$L__BB0_16:
abs.f32 %f1748, %f76;
setp.neu.f32 %p31, %f1748, 0f7F800000;
@%p31 bra $L__BB0_19;
selp.f32 %f1785, 0fFF800000, 0f7F800000, %p1;
$L__BB0_19:
mov.f32 %f583, 0f3F800000;
sub.f32 %f584, %f583, %f1785;
setp.eq.f32 %p32, %f76, 0f3F800000;
selp.f32 %f585, 0f00000000, %f584, %p32;
cvt.sat.f32.f32 %f586, %f585;
fma.rn.f32 %f587, %f75, %f75, %f481;
div.rn.f32 %f88, %f586, %f587;
mul.f32 %f588, %f1750, %f73;
fma.rn.f32 %f589, %f1749, %f72, %f588;
fma.rn.f32 %f89, %f1751, %f74, %f589;
setp.eq.f32 %p33, %f472, 0f3F800000;
@%p33 bra $L__BB0_25;
bra.uni $L__BB0_20;
$L__BB0_25:
setp.eq.s64 %p37, %rd6, 0;
mov.f32 %f1790, %f485;
mov.f32 %f1793, %f88;
@%p37 bra $L__BB0_29;
setp.geu.f32 %p38, %f1789, 0f00000000;
mov.f32 %f1786, %f72;
mov.f32 %f1787, %f73;
mov.f32 %f1788, %f74;
@%p38 bra $L__BB0_28;
mul.f32 %f612, %f67, %f70;
fma.rn.f32 %f613, %f66, %f69, %f612;
fma.rn.f32 %f614, %f68, %f71, %f613;
rcp.rn.f32 %f615, %f614;
mul.f32 %f1786, %f69, %f615;
mul.f32 %f1787, %f70, %f615;
mul.f32 %f1788, %f71, %f615;
neg.f32 %f1789, %f1789;
$L__BB0_28:
mul.f32 %f616, %f68, %f74;
mul.f32 %f617, %f67, %f73;
neg.f32 %f618, %f617;
mul.f32 %f619, %f66, %f72;
sub.f32 %f620, %f618, %f619;
sub.f32 %f621, %f620, %f616;
setp.gt.f32 %p39, %f621, 0f00000000;
selp.f32 %f622, 0f3F800000, 0f00000000, %p39;
mov.f32 %f623, 0f3F800000;
mul.f32 %f624, %f490, %f1787;
fma.rn.f32 %f625, %f489, %f1786, %f624;
mul.f32 %f626, %f494, %f1787;
fma.rn.f32 %f627, %f493, %f1786, %f626;
fma.rn.f32 %f628, %f491, %f1788, %f625;
fma.rn.f32 %f629, %f495, %f1788, %f627;
fma.rn.f32 %f630, %f1789, %f628, 0f3F000000;
sub.f32 %f631, %f623, %f630;
fma.rn.f32 %f632, %f1789, %f629, 0f3F000000;
tex.2d.v4.f32.f32 {%f633, %f634, %f635, %f636}, [%rd6, {%f631, %f632}];
mul.f32 %f637, %f622, %f633;
mul.f32 %f638, %f622, %f634;
mul.f32 %f639, %f622, %f635;
mul.f32 %f1790, %f485, %f637;
mul.f32 %f1791, %f1791, %f638;
mul.f32 %f1792, %f1792, %f639;
mov.f32 %f1793, %f88;
bra.uni $L__BB0_29;
$L__BB0_20:
setp.eq.f32 %p34, %f472, 0f40000000;
@%p34 bra $L__BB0_23;
bra.uni $L__BB0_21;
$L__BB0_23:
setp.eq.s64 %p36, %rd6, 0;
mov.f32 %f1790, %f485;
mov.f32 %f1793, %f88;
@%p36 bra $L__BB0_29;
mul.f32 %f599, %f490, %f73;
fma.rn.f32 %f600, %f489, %f72, %f599;
mul.f32 %f601, %f494, %f73;
fma.rn.f32 %f602, %f493, %f72, %f601;
mul.f32 %f603, %f67, %f73;
fma.rn.f32 %f604, %f66, %f72, %f603;
fma.rn.f32 %f605, %f491, %f74, %f600;
fma.rn.f32 %f606, %f495, %f74, %f602;
fma.rn.f32 %f607, %f68, %f74, %f604;
tex.cube.v4.f32.f32 {%f608, %f609, %f610, %f611}, [%rd6, {%f605, %f606, %f607, %f607}];
mul.f32 %f1790, %f485, %f608;
mul.f32 %f1791, %f1791, %f609;
mul.f32 %f1792, %f1792, %f610;
mov.f32 %f1793, %f88;
bra.uni $L__BB0_29;
$L__BB0_21:
setp.neu.f32 %p35, %f472, 0f40800000;
mov.f32 %f1790, %f485;
mov.f32 %f1793, %f88;
@%p35 bra $L__BB0_29;
mul.f32 %f590, %f66, %f72;
mul.f32 %f591, %f67, %f73;
neg.f32 %f592, %f591;
sub.f32 %f593, %f592, %f590;
mul.f32 %f594, %f68, %f74;
sub.f32 %f595, %f593, %f594;
fma.rn.f32 %f596, %f488, %f595, %f1789;
cvt.sat.f32.f32 %f597, %f596;
mul.f32 %f598, %f597, %f597;
mul.f32 %f1793, %f88, %f598;
mov.f32 %f1790, %f485;
$L__BB0_29:
max.f32 %f670, %f1790, %f1791;
max.f32 %f671, %f670, %f1792;
mul.f32 %f672, %f1793, %f671;
setp.lt.f32 %p41, %f672, 0f3727C5AC;
mov.pred %p192, -1;
mov.f32 %f1794, 0f00000000;
mov.f32 %f1795, %f1794;
mov.f32 %f1796, %f1794;
mov.f32 %f1797, %f1794;
mov.f32 %f1798, %f1794;
mov.f32 %f1799, %f1794;
mov.f32 %f1800, %f1794;
mov.f32 %f1801, %f1794;
mov.f32 %f1802, %f1794;
mov.f32 %f1803, %f1794;
mov.f32 %f1804, %f1794;
mov.f32 %f1805, %f1794;
mov.f32 %f1806, %f1794;
mov.f32 %f1807, %f1794;
mov.f32 %f1808, %f1794;
mov.f32 %f1809, %f1794;
mov.f32 %f1810, %f1794;
mov.f32 %f1811, %f1794;
mov.f32 %f1812, %f1794;
mov.f32 %f1813, %f1794;
mov.f32 %f1814, %f1794;
mov.f32 %f1815, %f1794;
mov.f32 %f1816, %f1794;
mov.f32 %f1817, %f1794;
mov.f32 %f1818, %f1794;
mov.f32 %f1819, %f1794;
mov.f32 %f1820, %f1794;
mov.f32 %f1821, %f1794;
mov.f32 %f1822, %f1794;
mov.f32 %f1823, %f1794;
@%p41 bra $L__BB0_31;
setp.eq.s32 %p43, %r6, 0;
selp.f32 %f673, %f89, 0f3F800000, %p43;
cvt.sat.f32.f32 %f674, %f673;
mul.f32 %f675, %f1793, %f674;
mul.f32 %f1820, %f1790, %f675;
mul.f32 %f1819, %f1791, %f675;
mul.f32 %f1818, %f1792, %f675;
mul.f32 %f676, %f1790, %f1793;
add.f32 %f1817, %f676, 0f00000000;
mul.f32 %f677, %f1791, %f1793;
add.f32 %f1816, %f677, 0f00000000;
mul.f32 %f678, %f1792, %f1793;
add.f32 %f1815, %f678, 0f00000000;
fma.rn.f32 %f1814, %f73, %f676, 0f00000000;
fma.rn.f32 %f1813, %f73, %f677, 0f00000000;
fma.rn.f32 %f1812, %f73, %f678, 0f00000000;
fma.rn.f32 %f1811, %f74, %f676, 0f00000000;
fma.rn.f32 %f1810, %f74, %f677, 0f00000000;
fma.rn.f32 %f1809, %f74, %f678, 0f00000000;
fma.rn.f32 %f1808, %f72, %f676, 0f00000000;
fma.rn.f32 %f1807, %f72, %f677, 0f00000000;
fma.rn.f32 %f1806, %f72, %f678, 0f00000000;
mul.f32 %f679, %f72, %f73;
fma.rn.f32 %f1805, %f679, %f676, 0f00000000;
fma.rn.f32 %f1804, %f679, %f677, 0f00000000;
fma.rn.f32 %f1803, %f679, %f678, 0f00000000;
mul.f32 %f680, %f73, %f74;
fma.rn.f32 %f1802, %f680, %f676, 0f00000000;
fma.rn.f32 %f1801, %f680, %f677, 0f00000000;
fma.rn.f32 %f1800, %f680, %f678, 0f00000000;
fma.rn.f32 %f681, %f74, %f74, 0fBEAAAAAB;
fma.rn.f32 %f1799, %f681, %f676, 0f00000000;
fma.rn.f32 %f1798, %f681, %f677, 0f00000000;
fma.rn.f32 %f1797, %f681, %f678, 0f00000000;
mul.f32 %f682, %f72, %f74;
fma.rn.f32 %f1796, %f682, %f676, 0f00000000;
fma.rn.f32 %f1795, %f682, %f677, 0f00000000;
fma.rn.f32 %f1794, %f682, %f678, 0f00000000;
mul.f32 %f683, %f73, %f73;
mul.f32 %f684, %f72, %f72;
sub.f32 %f685, %f684, %f683;
fma.rn.f32 %f1821, %f685, %f676, 0f00000000;
fma.rn.f32 %f1822, %f685, %f677, 0f00000000;
fma.rn.f32 %f1823, %f685, %f678, 0f00000000;
mov.pred %p192, 0;
$L__BB0_31:
@%p192 bra $L__BB0_43;
setp.eq.s32 %p44, %r44, 0;
mov.u16 %rs329, 0;
mov.f32 %f1836, 0f3F800000;
mov.f32 %f1837, %f1836;
mov.f32 %f1839, %f1836;
@%p44 bra $L__BB0_40;
abs.s32 %r10, %r44;
setp.lt.s32 %p45, %r10, 1;
mov.f32 %f1829, 0f00000000;
mov.f32 %f1828, %f1829;
mov.f32 %f1827, %f1829;
@%p45 bra $L__BB0_39;
mov.f32 %f701, 0f38D1B717;
max.f32 %f702, %f700, %f701;
and.b32 %r63, %r447, 32;
setp.eq.s32 %p46, %r63, 0;
mov.u32 %r446, 0;
selp.f32 %f703, 0f3F800000, 0f41200000, %p46;
mul.f32 %f169, %f703, %f702;
and.b32 %r11, %r10, 1;
setp.eq.s32 %p47, %r10, 1;
mov.f32 %f1827, 0f00000000;
mov.f32 %f1828, %f1827;
mov.f32 %f1829, %f1827;
@%p47 bra $L__BB0_37;
sub.s32 %r445, %r10, %r11;
$L__BB0_36:
cvt.rn.f32.s32 %f725, %r446;
add.f32 %f726, %f16, %f725;
sub.f32 %f727, %f17, %f725;
mul.f32 %f728, %f726, 0f3DD32618;
cvt.rmi.f32.f32 %f729, %f728;
sub.f32 %f730, %f728, %f729;
mul.f32 %f731, %f727, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f732, %f731;
sub.f32 %f733, %f731, %f732;
mul.f32 %f734, %f725, 0f3DC74539;
cvt.rmi.f32.f32 %f735, %f734;
sub.f32 %f736, %f734, %f735;
add.f32 %f737, %f733, 0f420551EC;
add.f32 %f738, %f730, 0f420551EC;
add.f32 %f739, %f736, 0f420551EC;
mul.f32 %f740, %f730, %f737;
fma.rn.f32 %f741, %f733, %f738, %f740;
fma.rn.f32 %f742, %f736, %f739, %f741;
add.f32 %f743, %f730, %f742;
add.f32 %f744, %f733, %f742;
add.f32 %f745, %f736, %f742;
add.f32 %f746, %f743, %f744;
mul.f32 %f747, %f745, %f746;
cvt.rmi.f32.f32 %f748, %f747;
sub.f32 %f749, %f747, %f748;
add.f32 %f750, %f743, %f743;
mul.f32 %f751, %f744, %f750;
cvt.rmi.f32.f32 %f752, %f751;
sub.f32 %f753, %f751, %f752;
mul.f32 %f754, %f743, %f746;
cvt.rmi.f32.f32 %f755, %f754;
sub.f32 %f756, %f754, %f755;
fma.rn.f32 %f757, %f749, 0f40000000, 0fBF800000;
fma.rn.f32 %f758, %f753, 0f40000000, 0fBF800000;
fma.rn.f32 %f759, %f756, 0f40000000, 0fBF800000;
ld.global.f32 %f760, [%rd5+16];
ld.global.f32 %f761, [%rd5+4];
fma.rn.f32 %f762, %f760, %f757, %f761;
ld.global.f32 %f763, [%rd5+8];
fma.rn.f32 %f764, %f760, %f758, %f763;
ld.global.f32 %f765, [%rd5+12];
fma.rn.f32 %f766, %f760, %f759, %f765;
sub.f32 %f767, %f762, %f10;
sub.f32 %f768, %f764, %f11;
sub.f32 %f769, %f766, %f12;
mul.f32 %f770, %f768, %f768;
fma.rn.f32 %f771, %f767, %f767, %f770;
fma.rn.f32 %f772, %f769, %f769, %f771;
sqrt.rn.f32 %f714, %f772;
rcp.rn.f32 %f773, %f714;
mul.f32 %f710, %f773, %f767;
mul.f32 %f711, %f773, %f768;
mul.f32 %f712, %f773, %f769;
mov.f32 %f724, 0f00000000;
mov.u32 %r172, 2;
mov.u32 %r173, 1;
mov.u32 %r174, 3;
mov.u32 %r177, 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,%f710,%f711,%f712,%f169,%f714,%f724,%r173,%r206,%r173,%r172,%r173,%r174,%r177,%r177,%r177,%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 %f774, %r65;
mov.b32 %f775, %r66;
mov.b32 %f776, %r67;
add.f32 %f777, %f1827, %f774;
add.f32 %f778, %f1828, %f775;
add.f32 %f779, %f1829, %f776;
add.s32 %r207, %r446, 1;
cvt.rn.f32.s32 %f780, %r207;
add.f32 %f781, %f16, %f780;
sub.f32 %f782, %f17, %f780;
mul.f32 %f783, %f781, 0f3DD32618;
cvt.rmi.f32.f32 %f784, %f783;
sub.f32 %f785, %f783, %f784;
mul.f32 %f786, %f782, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f787, %f786;
sub.f32 %f788, %f786, %f787;
mul.f32 %f789, %f780, 0f3DC74539;
cvt.rmi.f32.f32 %f790, %f789;
sub.f32 %f791, %f789, %f790;
add.f32 %f792, %f788, 0f420551EC;
add.f32 %f793, %f785, 0f420551EC;
add.f32 %f794, %f791, 0f420551EC;
mul.f32 %f795, %f785, %f792;
fma.rn.f32 %f796, %f788, %f793, %f795;
fma.rn.f32 %f797, %f791, %f794, %f796;
add.f32 %f798, %f785, %f797;
add.f32 %f799, %f788, %f797;
add.f32 %f800, %f791, %f797;
add.f32 %f801, %f798, %f799;
mul.f32 %f802, %f800, %f801;
cvt.rmi.f32.f32 %f803, %f802;
sub.f32 %f804, %f802, %f803;
add.f32 %f805, %f798, %f798;
mul.f32 %f806, %f799, %f805;
cvt.rmi.f32.f32 %f807, %f806;
sub.f32 %f808, %f806, %f807;
mul.f32 %f809, %f798, %f801;
cvt.rmi.f32.f32 %f810, %f809;
sub.f32 %f811, %f809, %f810;
fma.rn.f32 %f812, %f804, 0f40000000, 0fBF800000;
fma.rn.f32 %f813, %f808, 0f40000000, 0fBF800000;
fma.rn.f32 %f814, %f811, 0f40000000, 0fBF800000;
ld.global.f32 %f815, [%rd5+16];
ld.global.f32 %f816, [%rd5+4];
fma.rn.f32 %f817, %f815, %f812, %f816;
ld.global.f32 %f818, [%rd5+8];
fma.rn.f32 %f819, %f815, %f813, %f818;
ld.global.f32 %f820, [%rd5+12];
fma.rn.f32 %f821, %f815, %f814, %f820;
sub.f32 %f822, %f817, %f10;
sub.f32 %f823, %f819, %f11;
sub.f32 %f824, %f821, %f12;
mul.f32 %f825, %f823, %f823;
fma.rn.f32 %f826, %f822, %f822, %f825;
fma.rn.f32 %f827, %f824, %f824, %f826;
sqrt.rn.f32 %f723, %f827;
rcp.rn.f32 %f828, %f723;
mul.f32 %f719, %f828, %f822;
mul.f32 %f720, %f828, %f823;
mul.f32 %f721, %f828, %f824;
// 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,%f719,%f720,%f721,%f169,%f723,%f724,%r173,%r206,%r173,%r172,%r173,%r174,%r177,%r177,%r177,%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 %f829, %r136;
mov.b32 %f830, %r137;
mov.b32 %f831, %r138;
add.f32 %f1827, %f777, %f829;
add.f32 %f1828, %f778, %f830;
add.f32 %f1829, %f779, %f831;
add.s32 %r446, %r446, 2;
add.s32 %r445, %r445, -2;
setp.ne.s32 %p48, %r445, 0;
@%p48 bra $L__BB0_36;
$L__BB0_37:
setp.eq.s32 %p49, %r11, 0;
@%p49 bra $L__BB0_39;
cvt.rn.f32.s32 %f841, %r446;
add.f32 %f842, %f16, %f841;
sub.f32 %f843, %f17, %f841;
mul.f32 %f844, %f842, 0f3DD32618;
cvt.rmi.f32.f32 %f845, %f844;
sub.f32 %f846, %f844, %f845;
mul.f32 %f847, %f843, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f848, %f847;
sub.f32 %f849, %f847, %f848;
mul.f32 %f850, %f841, 0f3DC74539;
cvt.rmi.f32.f32 %f851, %f850;
sub.f32 %f852, %f850, %f851;
add.f32 %f853, %f849, 0f420551EC;
add.f32 %f854, %f846, 0f420551EC;
add.f32 %f855, %f852, 0f420551EC;
mul.f32 %f856, %f846, %f853;
fma.rn.f32 %f857, %f849, %f854, %f856;
fma.rn.f32 %f858, %f852, %f855, %f857;
add.f32 %f859, %f846, %f858;
add.f32 %f860, %f849, %f858;
add.f32 %f861, %f852, %f858;
add.f32 %f862, %f859, %f860;
mul.f32 %f863, %f861, %f862;
cvt.rmi.f32.f32 %f864, %f863;
sub.f32 %f865, %f863, %f864;
add.f32 %f866, %f859, %f859;
mul.f32 %f867, %f860, %f866;
cvt.rmi.f32.f32 %f868, %f867;
sub.f32 %f869, %f867, %f868;
mul.f32 %f870, %f859, %f862;
cvt.rmi.f32.f32 %f871, %f870;
sub.f32 %f872, %f870, %f871;
fma.rn.f32 %f873, %f865, 0f40000000, 0fBF800000;
fma.rn.f32 %f874, %f869, 0f40000000, 0fBF800000;
fma.rn.f32 %f875, %f872, 0f40000000, 0fBF800000;
ld.global.f32 %f876, [%rd5+16];
ld.global.f32 %f877, [%rd5+4];
fma.rn.f32 %f878, %f876, %f873, %f877;
ld.global.f32 %f879, [%rd5+8];
fma.rn.f32 %f880, %f876, %f874, %f879;
ld.global.f32 %f881, [%rd5+12];
fma.rn.f32 %f882, %f876, %f875, %f881;
sub.f32 %f883, %f878, %f10;
sub.f32 %f884, %f880, %f11;
sub.f32 %f885, %f882, %f12;
mul.f32 %f886, %f884, %f884;
fma.rn.f32 %f887, %f883, %f883, %f886;
fma.rn.f32 %f888, %f885, %f885, %f887;
sqrt.rn.f32 %f839, %f888;
rcp.rn.f32 %f889, %f839;
mul.f32 %f835, %f889, %f883;
mul.f32 %f836, %f889, %f884;
mul.f32 %f837, %f889, %f885;
mov.f32 %f840, 0f00000000;
mov.u32 %r244, 2;
mov.u32 %r245, 1;
mov.u32 %r246, 3;
mov.u32 %r249, 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,%f835,%f836,%f837,%f169,%f839,%f840,%r245,%r278,%r245,%r244,%r245,%r246,%r249,%r249,%r249,%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 %f890, %r208;
mov.b32 %f891, %r209;
mov.b32 %f892, %r210;
add.f32 %f1827, %f1827, %f890;
add.f32 %f1828, %f1828, %f891;
add.f32 %f1829, %f1829, %f892;
$L__BB0_39:
cvt.rn.f32.s32 %f893, %r10;
rcp.rn.f32 %f894, %f893;
mul.f32 %f1839, %f894, %f1827;
mul.f32 %f1837, %f894, %f1828;
mul.f32 %f1836, %f894, %f1829;
shr.u32 %r279, %r44, 31;
cvt.u16.u32 %rs329, %r279;
$L__BB0_40:
fma.rn.f32 %f1840, %f1820, %f1839, %f1840;
fma.rn.f32 %f1841, %f1819, %f1837, %f1841;
fma.rn.f32 %f1842, %f1818, %f1836, %f1842;
fma.rn.f32 %f1843, %f1817, %f1839, %f1843;
fma.rn.f32 %f1844, %f1816, %f1837, %f1844;
fma.rn.f32 %f1845, %f1815, %f1836, %f1845;
fma.rn.f32 %f1846, %f1814, %f1839, %f1846;
fma.rn.f32 %f1847, %f1813, %f1837, %f1847;
fma.rn.f32 %f1848, %f1812, %f1836, %f1848;
fma.rn.f32 %f1849, %f1811, %f1839, %f1849;
fma.rn.f32 %f1850, %f1810, %f1837, %f1850;
fma.rn.f32 %f1851, %f1809, %f1836, %f1851;
fma.rn.f32 %f1852, %f1808, %f1839, %f1852;
fma.rn.f32 %f1853, %f1807, %f1837, %f1853;
fma.rn.f32 %f1854, %f1806, %f1836, %f1854;
fma.rn.f32 %f1855, %f1805, %f1839, %f1855;
fma.rn.f32 %f1856, %f1804, %f1837, %f1856;
fma.rn.f32 %f1857, %f1803, %f1836, %f1857;
fma.rn.f32 %f1858, %f1802, %f1839, %f1858;
fma.rn.f32 %f1859, %f1801, %f1837, %f1859;
fma.rn.f32 %f1860, %f1800, %f1836, %f1860;
fma.rn.f32 %f1861, %f1799, %f1839, %f1861;
fma.rn.f32 %f1862, %f1798, %f1837, %f1862;
fma.rn.f32 %f1863, %f1797, %f1836, %f1863;
fma.rn.f32 %f1864, %f1796, %f1839, %f1864;
fma.rn.f32 %f1865, %f1795, %f1837, %f1865;
fma.rn.f32 %f1866, %f1794, %f1836, %f1866;
fma.rn.f32 %f1867, %f1821, %f1839, %f1867;
fma.rn.f32 %f1868, %f1822, %f1837, %f1868;
fma.rn.f32 %f1869, %f1823, %f1836, %f1869;
setp.eq.s16 %p50, %rs329, 0;
@%p50 bra $L__BB0_42;
div.rn.f32 %f895, %f1820, %f485;
div.rn.f32 %f896, %f895, %f88;
cvt.sat.f32.f32 %f897, %f896;
mul.f32 %f1839, %f1839, %f897;
$L__BB0_42:
add.f32 %f1870, %f1870, %f1839;
$L__BB0_43:
cvt.u64.u32 %rd117, %r443;
cvt.u32.u64 %r280, %rd117;
add.s32 %r443, %r280, 1;
setp.lt.u32 %p51, %r443, %r5;
@%p51 bra $L__BB0_7;
bra.uni $L__BB0_45;
$L__BB0_44:
mov.f32 %f1840, 0f00000000;
ld.const.u32 %r447, [params+104];
mov.f32 %f1841, %f1840;
mov.f32 %f1842, %f1840;
mov.f32 %f1843, %f1840;
mov.f32 %f1844, %f1840;
mov.f32 %f1845, %f1840;
mov.f32 %f1846, %f1840;
mov.f32 %f1847, %f1840;
mov.f32 %f1848, %f1840;
mov.f32 %f1849, %f1840;
mov.f32 %f1850, %f1840;
mov.f32 %f1851, %f1840;
mov.f32 %f1852, %f1840;
mov.f32 %f1853, %f1840;
mov.f32 %f1854, %f1840;
mov.f32 %f1855, %f1840;
mov.f32 %f1856, %f1840;
mov.f32 %f1857, %f1840;
mov.f32 %f1858, %f1840;
mov.f32 %f1859, %f1840;
mov.f32 %f1860, %f1840;
mov.f32 %f1861, %f1840;
mov.f32 %f1862, %f1840;
mov.f32 %f1863, %f1840;
mov.f32 %f1864, %f1840;
mov.f32 %f1865, %f1840;
mov.f32 %f1866, %f1840;
mov.f32 %f1867, %f1840;
mov.f32 %f1868, %f1840;
mov.f32 %f1869, %f1840;
mov.f32 %f1870, %f1840;
$L__BB0_45:
and.b32 %r281, %r447, 8;
setp.eq.s32 %p52, %r281, 0;
@%p52 bra $L__BB0_59;
cvt.sat.f32.f32 %f289, %f1870;
ld.const.u64 %rd43, [params+192];
cvta.to.global.u64 %rd7, %rd43;
ld.const.u32 %r282, [params+184];
mad.lo.s32 %r283, %r282, %r4, %r3;
cvt.u64.u32 %rd8, %r283;
mov.f32 %f930, 0f3E68BA2E;
cvt.rzi.f32.f32 %f931, %f930;
add.f32 %f932, %f931, %f931;
mov.f32 %f933, 0f3EE8BA2E;
sub.f32 %f934, %f933, %f932;
abs.f32 %f290, %f934;
abs.f32 %f291, %f289;
setp.lt.f32 %p53, %f291, 0f00800000;
mul.f32 %f935, %f291, 0f4B800000;
selp.f32 %f936, %f935, %f291, %p53;
selp.f32 %f937, 0fC3170000, 0fC2FE0000, %p53;
mov.b32 %r284, %f936;
and.b32 %r285, %r284, 8388607;
or.b32 %r286, %r285, 1065353216;
mov.b32 %f938, %r286;
shr.u32 %r287, %r284, 23;
cvt.rn.f32.u32 %f939, %r287;
add.f32 %f940, %f937, %f939;
setp.gt.f32 %p54, %f938, 0f3FB504F3;
mul.f32 %f941, %f938, 0f3F000000;
add.f32 %f942, %f940, 0f3F800000;
selp.f32 %f943, %f942, %f940, %p54;
selp.f32 %f944, %f941, %f938, %p54;
add.f32 %f945, %f944, 0fBF800000;
add.f32 %f946, %f944, 0f3F800000;
rcp.approx.ftz.f32 %f947, %f946;
add.f32 %f948, %f945, %f945;
mul.f32 %f949, %f948, %f947;
mul.f32 %f950, %f949, %f949;
mov.f32 %f951, 0f3C4CAF63;
mov.f32 %f952, 0f3B18F0FE;
fma.rn.f32 %f953, %f952, %f950, %f951;
mov.f32 %f954, 0f3DAAAABD;
fma.rn.f32 %f955, %f953, %f950, %f954;
mul.rn.f32 %f956, %f955, %f950;
mul.rn.f32 %f957, %f956, %f949;
sub.f32 %f958, %f945, %f949;
add.f32 %f959, %f958, %f958;
neg.f32 %f960, %f949;
fma.rn.f32 %f961, %f960, %f945, %f959;
mul.rn.f32 %f962, %f947, %f961;
add.f32 %f963, %f957, %f949;
sub.f32 %f964, %f949, %f963;
add.f32 %f965, %f957, %f964;
add.f32 %f966, %f962, %f965;
add.f32 %f967, %f963, %f966;
sub.f32 %f968, %f963, %f967;
add.f32 %f969, %f966, %f968;
mov.f32 %f970, 0f3F317200;
mul.rn.f32 %f971, %f943, %f970;
mov.f32 %f972, 0f35BFBE8E;
mul.rn.f32 %f973, %f943, %f972;
add.f32 %f974, %f971, %f967;
sub.f32 %f975, %f971, %f974;
add.f32 %f976, %f967, %f975;
add.f32 %f977, %f969, %f976;
add.f32 %f978, %f973, %f977;
add.f32 %f979, %f974, %f978;
sub.f32 %f980, %f974, %f979;
add.f32 %f981, %f978, %f980;
mul.rn.f32 %f982, %f933, %f979;
neg.f32 %f983, %f982;
fma.rn.f32 %f984, %f933, %f979, %f983;
fma.rn.f32 %f985, %f933, %f981, %f984;
mov.f32 %f986, 0f00000000;
fma.rn.f32 %f987, %f986, %f979, %f985;
add.rn.f32 %f988, %f982, %f987;
neg.f32 %f989, %f988;
add.rn.f32 %f990, %f982, %f989;
add.rn.f32 %f991, %f990, %f987;
mov.b32 %r288, %f988;
setp.eq.s32 %p55, %r288, 1118925336;
add.s32 %r289, %r288, -1;
mov.b32 %f992, %r289;
add.f32 %f993, %f991, 0f37000000;
selp.f32 %f292, %f993, %f991, %p55;
selp.f32 %f994, %f992, %f988, %p55;
mov.f32 %f995, 0f3FB8AA3B;
mul.rn.f32 %f996, %f994, %f995;
cvt.rzi.f32.f32 %f997, %f996;
abs.f32 %f998, %f997;
setp.gt.f32 %p56, %f998, 0f42FC0000;
mov.b32 %r290, %f997;
and.b32 %r291, %r290, -2147483648;
or.b32 %r292, %r291, 1123811328;
mov.b32 %f999, %r292;
selp.f32 %f1000, %f999, %f997, %p56;
mov.f32 %f1001, 0fBF317218;
fma.rn.f32 %f1002, %f1000, %f1001, %f994;
mov.f32 %f1003, 0f3102E308;
fma.rn.f32 %f1004, %f1000, %f1003, %f1002;
mul.f32 %f1005, %f1004, 0f3FB8AA3B;
add.f32 %f1006, %f1000, 0f4B40007F;
mov.b32 %r293, %f1006;
shl.b32 %r294, %r293, 23;
mov.b32 %f1007, %r294;
ex2.approx.ftz.f32 %f1008, %f1005;
mul.f32 %f293, %f1008, %f1007;
setp.eq.f32 %p57, %f293, 0f7F800000;
mov.f32 %f1902, 0f7F800000;
@%p57 bra $L__BB0_48;
fma.rn.f32 %f1902, %f293, %f292, %f293;
$L__BB0_48:
setp.lt.f32 %p58, %f289, 0f00000000;
setp.eq.f32 %p59, %f290, 0f3F800000;
and.pred %p3, %p58, %p59;
setp.eq.f32 %p60, %f289, 0f00000000;
@%p60 bra $L__BB0_52;
bra.uni $L__BB0_49;
$L__BB0_52:
add.f32 %f1013, %f289, %f289;
selp.f32 %f1904, %f1013, 0f00000000, %p59;
bra.uni $L__BB0_53;
$L__BB0_171:
mov.f32 %f1664, 0f00000000;
mov.u32 %r449, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs190, %f1664;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs189, %f1664;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs188, %f1664;}
// end inline asm
mov.u16 %rs191, 0;
st.global.v4.u16 [%rd21], {%rs188, %rs189, %rs190, %rs191};
$L__BB0_172:
ld.const.u64 %rd90, [params+256];
cvta.to.global.u64 %rd91, %rd90;
ld.const.u32 %r425, [params+248];
mad.lo.s32 %r426, %r425, %r4, %r3;
mul.wide.u32 %rd92, %r426, 8;
add.s64 %rd22, %rd91, %rd92;
setp.eq.s32 %p183, %r449, 0;
@%p183 bra $L__BB0_174;
ld.global.v4.u16 {%rs198, %rs199, %rs200, %rs201}, [%rd22];
// begin inline asm
{ cvt.f32.f16 %f1665, %rs198;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1666, %rs199;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1667, %rs200;}
// end inline asm
add.f32 %f1668, %f1665, 0f00000000;
add.f32 %f1669, %f1666, 0f00000000;
add.f32 %f1670, %f1667, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs197, %f1670;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs196, %f1669;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs195, %f1668;}
// end inline asm
mov.u16 %rs202, 0;
st.global.v4.u16 [%rd22], {%rs195, %rs196, %rs197, %rs202};
bra.uni $L__BB0_175;
$L__BB0_174:
mov.f32 %f1673, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs205, %f1673;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs204, %f1673;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs203, %f1673;}
// end inline asm
mov.u16 %rs206, 0;
st.global.v4.u16 [%rd22], {%rs203, %rs204, %rs205, %rs206};
$L__BB0_175:
ld.const.u64 %rd93, [params+272];
cvta.to.global.u64 %rd94, %rd93;
ld.const.u32 %r427, [params+264];
mad.lo.s32 %r428, %r427, %r4, %r3;
mul.wide.u32 %rd95, %r428, 8;
add.s64 %rd23, %rd94, %rd95;
@%p183 bra $L__BB0_177;
ld.global.v4.u16 {%rs213, %rs214, %rs215, %rs216}, [%rd23];
// begin inline asm
{ cvt.f32.f16 %f1674, %rs213;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1675, %rs214;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1676, %rs215;}
// end inline asm
add.f32 %f1677, %f1674, 0f00000000;
add.f32 %f1678, %f1675, 0f00000000;
add.f32 %f1679, %f1676, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs212, %f1679;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs211, %f1678;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs210, %f1677;}
// end inline asm
mov.u16 %rs217, 0;
st.global.v4.u16 [%rd23], {%rs210, %rs211, %rs212, %rs217};
bra.uni $L__BB0_178;
$L__BB0_177:
mov.f32 %f1682, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs220, %f1682;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs219, %f1682;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs218, %f1682;}
// end inline asm
mov.u16 %rs221, 0;
st.global.v4.u16 [%rd23], {%rs218, %rs219, %rs220, %rs221};
$L__BB0_178:
ld.const.u64 %rd96, [params+288];
cvta.to.global.u64 %rd97, %rd96;
ld.const.u32 %r429, [params+280];
mad.lo.s32 %r430, %r429, %r4, %r3;
mul.wide.u32 %rd98, %r430, 8;
add.s64 %rd24, %rd97, %rd98;
@%p183 bra $L__BB0_180;
ld.global.v4.u16 {%rs228, %rs229, %rs230, %rs231}, [%rd24];
// begin inline asm
{ cvt.f32.f16 %f1683, %rs228;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1684, %rs229;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1685, %rs230;}
// end inline asm
add.f32 %f1686, %f1683, 0f00000000;
add.f32 %f1687, %f1684, 0f00000000;
add.f32 %f1688, %f1685, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs227, %f1688;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs226, %f1687;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs225, %f1686;}
// end inline asm
mov.u16 %rs232, 0;
st.global.v4.u16 [%rd24], {%rs225, %rs226, %rs227, %rs232};
bra.uni $L__BB0_181;
$L__BB0_180:
mov.f32 %f1691, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs235, %f1691;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs234, %f1691;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs233, %f1691;}
// end inline asm
mov.u16 %rs236, 0;
st.global.v4.u16 [%rd24], {%rs233, %rs234, %rs235, %rs236};
$L__BB0_181:
ld.const.u64 %rd99, [params+304];
cvta.to.global.u64 %rd100, %rd99;
ld.const.u32 %r431, [params+296];
mad.lo.s32 %r432, %r431, %r4, %r3;
mul.wide.u32 %rd101, %r432, 8;
add.s64 %rd25, %rd100, %rd101;
@%p183 bra $L__BB0_183;
ld.global.v4.u16 {%rs243, %rs244, %rs245, %rs246}, [%rd25];
// begin inline asm
{ cvt.f32.f16 %f1692, %rs243;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1693, %rs244;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1694, %rs245;}
// end inline asm
add.f32 %f1695, %f1692, 0f00000000;
add.f32 %f1696, %f1693, 0f00000000;
add.f32 %f1697, %f1694, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs242, %f1697;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs241, %f1696;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs240, %f1695;}
// end inline asm
mov.u16 %rs247, 0;
st.global.v4.u16 [%rd25], {%rs240, %rs241, %rs242, %rs247};
bra.uni $L__BB0_184;
$L__BB0_183:
mov.f32 %f1700, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs250, %f1700;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs249, %f1700;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs248, %f1700;}
// end inline asm
mov.u16 %rs251, 0;
st.global.v4.u16 [%rd25], {%rs248, %rs249, %rs250, %rs251};
$L__BB0_184:
ld.const.u64 %rd102, [params+320];
cvta.to.global.u64 %rd103, %rd102;
ld.const.u32 %r433, [params+312];
mad.lo.s32 %r434, %r433, %r4, %r3;
mul.wide.u32 %rd104, %r434, 8;
add.s64 %rd26, %rd103, %rd104;
@%p183 bra $L__BB0_186;
ld.global.v4.u16 {%rs258, %rs259, %rs260, %rs261}, [%rd26];
// begin inline asm
{ cvt.f32.f16 %f1701, %rs258;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1702, %rs259;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1703, %rs260;}
// end inline asm
add.f32 %f1704, %f1701, 0f00000000;
add.f32 %f1705, %f1702, 0f00000000;
add.f32 %f1706, %f1703, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs257, %f1706;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs256, %f1705;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs255, %f1704;}
// end inline asm
mov.u16 %rs262, 0;
st.global.v4.u16 [%rd26], {%rs255, %rs256, %rs257, %rs262};
bra.uni $L__BB0_187;
$L__BB0_186:
mov.f32 %f1709, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs265, %f1709;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs264, %f1709;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs263, %f1709;}
// end inline asm
mov.u16 %rs266, 0;
st.global.v4.u16 [%rd26], {%rs263, %rs264, %rs265, %rs266};
$L__BB0_187:
ld.const.u64 %rd105, [params+336];
cvta.to.global.u64 %rd106, %rd105;
ld.const.u32 %r435, [params+328];
mad.lo.s32 %r436, %r435, %r4, %r3;
mul.wide.u32 %rd107, %r436, 8;
add.s64 %rd27, %rd106, %rd107;
@%p183 bra $L__BB0_189;
ld.global.v4.u16 {%rs273, %rs274, %rs275, %rs276}, [%rd27];
// begin inline asm
{ cvt.f32.f16 %f1710, %rs273;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1711, %rs274;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1712, %rs275;}
// end inline asm
add.f32 %f1713, %f1710, 0f00000000;
add.f32 %f1714, %f1711, 0f00000000;
add.f32 %f1715, %f1712, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs272, %f1715;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs271, %f1714;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs270, %f1713;}
// end inline asm
mov.u16 %rs277, 0;
st.global.v4.u16 [%rd27], {%rs270, %rs271, %rs272, %rs277};
bra.uni $L__BB0_190;
$L__BB0_189:
mov.f32 %f1718, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs280, %f1718;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs279, %f1718;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs278, %f1718;}
// end inline asm
mov.u16 %rs281, 0;
st.global.v4.u16 [%rd27], {%rs278, %rs279, %rs280, %rs281};
$L__BB0_190:
ld.const.u64 %rd108, [params+352];
cvta.to.global.u64 %rd109, %rd108;
ld.const.u32 %r437, [params+344];
mad.lo.s32 %r438, %r437, %r4, %r3;
mul.wide.u32 %rd110, %r438, 8;
add.s64 %rd28, %rd109, %rd110;
@%p183 bra $L__BB0_192;
ld.global.v4.u16 {%rs288, %rs289, %rs290, %rs291}, [%rd28];
// begin inline asm
{ cvt.f32.f16 %f1719, %rs288;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1720, %rs289;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1721, %rs290;}
// end inline asm
add.f32 %f1722, %f1719, 0f00000000;
add.f32 %f1723, %f1720, 0f00000000;
add.f32 %f1724, %f1721, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs287, %f1724;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs286, %f1723;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs285, %f1722;}
// end inline asm
mov.u16 %rs292, 0;
st.global.v4.u16 [%rd28], {%rs285, %rs286, %rs287, %rs292};
bra.uni $L__BB0_193;
$L__BB0_192:
mov.f32 %f1727, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs295, %f1727;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs294, %f1727;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs293, %f1727;}
// end inline asm
mov.u16 %rs296, 0;
st.global.v4.u16 [%rd28], {%rs293, %rs294, %rs295, %rs296};
$L__BB0_193:
ld.const.u64 %rd111, [params+368];
cvta.to.global.u64 %rd112, %rd111;
ld.const.u32 %r439, [params+360];
mad.lo.s32 %r440, %r439, %r4, %r3;
mul.wide.u32 %rd113, %r440, 8;
add.s64 %rd29, %rd112, %rd113;
@%p183 bra $L__BB0_195;
ld.global.v4.u16 {%rs303, %rs304, %rs305, %rs306}, [%rd29];
// begin inline asm
{ cvt.f32.f16 %f1728, %rs303;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1729, %rs304;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1730, %rs305;}
// end inline asm
add.f32 %f1731, %f1728, 0f00000000;
add.f32 %f1732, %f1729, 0f00000000;
add.f32 %f1733, %f1730, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs302, %f1733;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs301, %f1732;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs300, %f1731;}
// end inline asm
mov.u16 %rs307, 0;
st.global.v4.u16 [%rd29], {%rs300, %rs301, %rs302, %rs307};
bra.uni $L__BB0_196;
$L__BB0_195:
mov.f32 %f1736, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs310, %f1736;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs309, %f1736;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs308, %f1736;}
// end inline asm
mov.u16 %rs311, 0;
st.global.v4.u16 [%rd29], {%rs308, %rs309, %rs310, %rs311};
$L__BB0_196:
ld.const.u64 %rd114, [params+384];
cvta.to.global.u64 %rd115, %rd114;
ld.const.u32 %r441, [params+376];
mad.lo.s32 %r442, %r441, %r4, %r3;
mul.wide.u32 %rd116, %r442, 8;
add.s64 %rd30, %rd115, %rd116;
@%p183 bra $L__BB0_198;
ld.global.v4.u16 {%rs318, %rs319, %rs320, %rs321}, [%rd30];
// begin inline asm
{ cvt.f32.f16 %f1737, %rs318;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1738, %rs319;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1739, %rs320;}
// end inline asm
add.f32 %f1740, %f1737, 0f00000000;
add.f32 %f1741, %f1738, 0f00000000;
add.f32 %f1742, %f1739, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs317, %f1742;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs316, %f1741;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs315, %f1740;}
// end inline asm
mov.u16 %rs322, 0;
st.global.v4.u16 [%rd30], {%rs315, %rs316, %rs317, %rs322};
bra.uni $L__BB0_199;
$L__BB0_198:
mov.f32 %f1745, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs325, %f1745;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs324, %f1745;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs323, %f1745;}
// end inline asm
mov.u16 %rs326, 0;
st.global.v4.u16 [%rd30], {%rs323, %rs324, %rs325, %rs326};
bra.uni $L__BB0_199;
$L__BB0_49:
mov.b32 %r295, %f1902;
xor.b32 %r296, %r295, -2147483648;
mov.b32 %f1009, %r296;
selp.f32 %f1904, %f1009, %f1902, %p3;
setp.geu.f32 %p61, %f289, 0f00000000;
@%p61 bra $L__BB0_53;
mov.f32 %f1010, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f1011, %f1010;
setp.eq.f32 %p62, %f1011, 0f3EE8BA2E;
@%p62 bra $L__BB0_53;
mov.f32 %f1904, 0f7FFFFFFF;
$L__BB0_53:
add.f32 %f1014, %f291, 0f3EE8BA2E;
mov.b32 %r297, %f1014;
setp.lt.s32 %p64, %r297, 2139095040;
@%p64 bra $L__BB0_58;
setp.gtu.f32 %p65, %f291, 0f7F800000;
@%p65 bra $L__BB0_57;
bra.uni $L__BB0_55;
$L__BB0_57:
add.f32 %f1904, %f289, 0f3EE8BA2E;
bra.uni $L__BB0_58;
$L__BB0_55:
setp.neu.f32 %p66, %f291, 0f7F800000;
@%p66 bra $L__BB0_58;
selp.f32 %f1904, 0fFF800000, 0f7F800000, %p3;
$L__BB0_58:
mul.f32 %f1015, %f1904, 0f437F0000;
setp.eq.f32 %p67, %f289, 0f3F800000;
selp.f32 %f1016, 0f437F0000, %f1015, %p67;
cvt.rzi.u32.f32 %r298, %f1016;
shl.b64 %rd44, %rd8, 1;
add.s64 %rd45, %rd7, %rd44;
cvt.u16.u32 %rs18, %r298;
mov.u16 %rs19, 255;
st.global.v2.u8 [%rd45], {%rs18, %rs19};
$L__BB0_59:
and.b32 %r299, %r447, 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 %f1018, 0f3E666666;
cvt.rzi.f32.f32 %f1019, %f1018;
add.f32 %f1020, %f1019, %f1019;
mov.f32 %f1021, 0f3EE66666;
sub.f32 %f1022, %f1021, %f1020;
abs.f32 %f302, %f1022;
abs.f32 %f303, %f1840;
setp.lt.f32 %p72, %f303, 0f00800000;
mul.f32 %f1023, %f303, 0f4B800000;
selp.f32 %f1024, %f1023, %f303, %p72;
selp.f32 %f1025, 0fC3170000, 0fC2FE0000, %p72;
mov.b32 %r300, %f1024;
and.b32 %r301, %r300, 8388607;
or.b32 %r302, %r301, 1065353216;
mov.b32 %f1026, %r302;
shr.u32 %r303, %r300, 23;
cvt.rn.f32.u32 %f1027, %r303;
add.f32 %f1028, %f1025, %f1027;
setp.gt.f32 %p73, %f1026, 0f3FB504F3;
mul.f32 %f1029, %f1026, 0f3F000000;
add.f32 %f1030, %f1028, 0f3F800000;
selp.f32 %f1031, %f1030, %f1028, %p73;
selp.f32 %f1032, %f1029, %f1026, %p73;
add.f32 %f1033, %f1032, 0fBF800000;
add.f32 %f1034, %f1032, 0f3F800000;
rcp.approx.ftz.f32 %f1035, %f1034;
add.f32 %f1036, %f1033, %f1033;
mul.f32 %f1037, %f1036, %f1035;
mul.f32 %f1038, %f1037, %f1037;
mov.f32 %f1039, 0f3C4CAF63;
mov.f32 %f1040, 0f3B18F0FE;
fma.rn.f32 %f1041, %f1040, %f1038, %f1039;
mov.f32 %f1042, 0f3DAAAABD;
fma.rn.f32 %f1043, %f1041, %f1038, %f1042;
mul.rn.f32 %f1044, %f1043, %f1038;
mul.rn.f32 %f1045, %f1044, %f1037;
sub.f32 %f1046, %f1033, %f1037;
add.f32 %f1047, %f1046, %f1046;
neg.f32 %f1048, %f1037;
fma.rn.f32 %f1049, %f1048, %f1033, %f1047;
mul.rn.f32 %f1050, %f1035, %f1049;
add.f32 %f1051, %f1045, %f1037;
sub.f32 %f1052, %f1037, %f1051;
add.f32 %f1053, %f1045, %f1052;
add.f32 %f1054, %f1050, %f1053;
add.f32 %f1055, %f1051, %f1054;
sub.f32 %f1056, %f1051, %f1055;
add.f32 %f1057, %f1054, %f1056;
mov.f32 %f1058, 0f3F317200;
mul.rn.f32 %f1059, %f1031, %f1058;
mov.f32 %f1060, 0f35BFBE8E;
mul.rn.f32 %f1061, %f1031, %f1060;
add.f32 %f1062, %f1059, %f1055;
sub.f32 %f1063, %f1059, %f1062;
add.f32 %f1064, %f1055, %f1063;
add.f32 %f1065, %f1057, %f1064;
add.f32 %f1066, %f1061, %f1065;
add.f32 %f1067, %f1062, %f1066;
sub.f32 %f1068, %f1062, %f1067;
add.f32 %f1069, %f1066, %f1068;
mul.rn.f32 %f1070, %f1021, %f1067;
neg.f32 %f1071, %f1070;
fma.rn.f32 %f1072, %f1021, %f1067, %f1071;
fma.rn.f32 %f1073, %f1021, %f1069, %f1072;
mov.f32 %f1074, 0f00000000;
fma.rn.f32 %f1075, %f1074, %f1067, %f1073;
add.rn.f32 %f1076, %f1070, %f1075;
neg.f32 %f1077, %f1076;
add.rn.f32 %f1078, %f1070, %f1077;
add.rn.f32 %f1079, %f1078, %f1075;
mov.b32 %r304, %f1076;
setp.eq.s32 %p74, %r304, 1118925336;
add.s32 %r305, %r304, -1;
mov.b32 %f1080, %r305;
add.f32 %f1081, %f1079, 0f37000000;
selp.f32 %f304, %f1081, %f1079, %p74;
selp.f32 %f1082, %f1080, %f1076, %p74;
mov.f32 %f1083, 0f3FB8AA3B;
mul.rn.f32 %f1084, %f1082, %f1083;
cvt.rzi.f32.f32 %f1085, %f1084;
abs.f32 %f1086, %f1085;
setp.gt.f32 %p75, %f1086, 0f42FC0000;
mov.b32 %r306, %f1085;
and.b32 %r307, %r306, -2147483648;
or.b32 %r308, %r307, 1123811328;
mov.b32 %f1087, %r308;
selp.f32 %f1088, %f1087, %f1085, %p75;
mov.f32 %f1089, 0fBF317218;
fma.rn.f32 %f1090, %f1088, %f1089, %f1082;
mov.f32 %f1091, 0f3102E308;
fma.rn.f32 %f1092, %f1088, %f1091, %f1090;
mul.f32 %f1093, %f1092, 0f3FB8AA3B;
add.f32 %f1094, %f1088, 0f4B40007F;
mov.b32 %r309, %f1094;
shl.b32 %r310, %r309, 23;
mov.b32 %f1095, %r310;
ex2.approx.ftz.f32 %f1096, %f1093;
mul.f32 %f305, %f1096, %f1095;
setp.eq.f32 %p76, %f305, 0f7F800000;
mov.f32 %f1905, 0f7F800000;
@%p76 bra $L__BB0_62;
fma.rn.f32 %f1905, %f305, %f304, %f305;
$L__BB0_62:
setp.lt.f32 %p77, %f1840, 0f00000000;
setp.eq.f32 %p78, %f302, 0f3F800000;
and.pred %p4, %p77, %p78;
setp.eq.f32 %p79, %f1840, 0f00000000;
@%p79 bra $L__BB0_66;
bra.uni $L__BB0_63;
$L__BB0_66:
add.f32 %f1101, %f1840, %f1840;
selp.f32 %f1907, %f1101, 0f00000000, %p78;
bra.uni $L__BB0_67;
$L__BB0_63:
mov.b32 %r311, %f1905;
xor.b32 %r312, %r311, -2147483648;
mov.b32 %f1097, %r312;
selp.f32 %f1907, %f1097, %f1905, %p4;
setp.geu.f32 %p80, %f1840, 0f00000000;
@%p80 bra $L__BB0_67;
mov.f32 %f1098, 0f3EE66666;
cvt.rzi.f32.f32 %f1099, %f1098;
setp.eq.f32 %p81, %f1099, 0f3EE66666;
@%p81 bra $L__BB0_67;
mov.f32 %f1907, 0f7FFFFFFF;
$L__BB0_67:
add.f32 %f1102, %f303, 0f3EE66666;
mov.b32 %r313, %f1102;
setp.lt.s32 %p83, %r313, 2139095040;
@%p83 bra $L__BB0_72;
setp.gtu.f32 %p84, %f303, 0f7F800000;
@%p84 bra $L__BB0_71;
bra.uni $L__BB0_69;
$L__BB0_71:
add.f32 %f1907, %f1840, 0f3EE66666;
bra.uni $L__BB0_72;
$L__BB0_69:
setp.neu.f32 %p85, %f303, 0f7F800000;
@%p85 bra $L__BB0_72;
selp.f32 %f1907, 0fFF800000, 0f7F800000, %p4;
$L__BB0_72:
setp.eq.f32 %p86, %f1840, 0f3F800000;
selp.f32 %f314, 0f3F800000, %f1907, %p86;
abs.f32 %f315, %f1841;
setp.lt.f32 %p87, %f315, 0f00800000;
mul.f32 %f1104, %f315, 0f4B800000;
selp.f32 %f1105, %f1104, %f315, %p87;
selp.f32 %f1106, 0fC3170000, 0fC2FE0000, %p87;
mov.b32 %r314, %f1105;
and.b32 %r315, %r314, 8388607;
or.b32 %r316, %r315, 1065353216;
mov.b32 %f1107, %r316;
shr.u32 %r317, %r314, 23;
cvt.rn.f32.u32 %f1108, %r317;
add.f32 %f1109, %f1106, %f1108;
setp.gt.f32 %p88, %f1107, 0f3FB504F3;
mul.f32 %f1110, %f1107, 0f3F000000;
add.f32 %f1111, %f1109, 0f3F800000;
selp.f32 %f1112, %f1111, %f1109, %p88;
selp.f32 %f1113, %f1110, %f1107, %p88;
add.f32 %f1114, %f1113, 0fBF800000;
add.f32 %f1115, %f1113, 0f3F800000;
rcp.approx.ftz.f32 %f1116, %f1115;
add.f32 %f1117, %f1114, %f1114;
mul.f32 %f1118, %f1117, %f1116;
mul.f32 %f1119, %f1118, %f1118;
mov.f32 %f1120, 0f3C4CAF63;
mov.f32 %f1121, 0f3B18F0FE;
fma.rn.f32 %f1122, %f1121, %f1119, %f1120;
mov.f32 %f1123, 0f3DAAAABD;
fma.rn.f32 %f1124, %f1122, %f1119, %f1123;
mul.rn.f32 %f1125, %f1124, %f1119;
mul.rn.f32 %f1126, %f1125, %f1118;
sub.f32 %f1127, %f1114, %f1118;
add.f32 %f1128, %f1127, %f1127;
neg.f32 %f1129, %f1118;
fma.rn.f32 %f1130, %f1129, %f1114, %f1128;
mul.rn.f32 %f1131, %f1116, %f1130;
add.f32 %f1132, %f1126, %f1118;
sub.f32 %f1133, %f1118, %f1132;
add.f32 %f1134, %f1126, %f1133;
add.f32 %f1135, %f1131, %f1134;
add.f32 %f1136, %f1132, %f1135;
sub.f32 %f1137, %f1132, %f1136;
add.f32 %f1138, %f1135, %f1137;
mov.f32 %f1139, 0f3F317200;
mul.rn.f32 %f1140, %f1112, %f1139;
mov.f32 %f1141, 0f35BFBE8E;
mul.rn.f32 %f1142, %f1112, %f1141;
add.f32 %f1143, %f1140, %f1136;
sub.f32 %f1144, %f1140, %f1143;
add.f32 %f1145, %f1136, %f1144;
add.f32 %f1146, %f1138, %f1145;
add.f32 %f1147, %f1142, %f1146;
add.f32 %f1148, %f1143, %f1147;
sub.f32 %f1149, %f1143, %f1148;
add.f32 %f1150, %f1147, %f1149;
mov.f32 %f1151, 0f3EE66666;
mul.rn.f32 %f1152, %f1151, %f1148;
neg.f32 %f1153, %f1152;
fma.rn.f32 %f1154, %f1151, %f1148, %f1153;
fma.rn.f32 %f1155, %f1151, %f1150, %f1154;
mov.f32 %f1156, 0f00000000;
fma.rn.f32 %f1157, %f1156, %f1148, %f1155;
add.rn.f32 %f1158, %f1152, %f1157;
neg.f32 %f1159, %f1158;
add.rn.f32 %f1160, %f1152, %f1159;
add.rn.f32 %f1161, %f1160, %f1157;
mov.b32 %r318, %f1158;
setp.eq.s32 %p89, %r318, 1118925336;
add.s32 %r319, %r318, -1;
mov.b32 %f1162, %r319;
add.f32 %f1163, %f1161, 0f37000000;
selp.f32 %f316, %f1163, %f1161, %p89;
selp.f32 %f1164, %f1162, %f1158, %p89;
mov.f32 %f1165, 0f3FB8AA3B;
mul.rn.f32 %f1166, %f1164, %f1165;
cvt.rzi.f32.f32 %f1167, %f1166;
abs.f32 %f1168, %f1167;
setp.gt.f32 %p90, %f1168, 0f42FC0000;
mov.b32 %r320, %f1167;
and.b32 %r321, %r320, -2147483648;
or.b32 %r322, %r321, 1123811328;
mov.b32 %f1169, %r322;
selp.f32 %f1170, %f1169, %f1167, %p90;
mov.f32 %f1171, 0fBF317218;
fma.rn.f32 %f1172, %f1170, %f1171, %f1164;
mov.f32 %f1173, 0f3102E308;
fma.rn.f32 %f1174, %f1170, %f1173, %f1172;
mul.f32 %f1175, %f1174, 0f3FB8AA3B;
add.f32 %f1176, %f1170, 0f4B40007F;
mov.b32 %r323, %f1176;
shl.b32 %r324, %r323, 23;
mov.b32 %f1177, %r324;
ex2.approx.ftz.f32 %f1178, %f1175;
mul.f32 %f317, %f1178, %f1177;
setp.eq.f32 %p91, %f317, 0f7F800000;
mov.f32 %f1908, 0f7F800000;
@%p91 bra $L__BB0_74;
fma.rn.f32 %f1908, %f317, %f316, %f317;
$L__BB0_74:
setp.lt.f32 %p92, %f1841, 0f00000000;
and.pred %p5, %p92, %p78;
setp.eq.f32 %p94, %f1841, 0f00000000;
@%p94 bra $L__BB0_78;
bra.uni $L__BB0_75;
$L__BB0_78:
add.f32 %f1183, %f1841, %f1841;
selp.f32 %f1910, %f1183, 0f00000000, %p78;
bra.uni $L__BB0_79;
$L__BB0_75:
mov.b32 %r325, %f1908;
xor.b32 %r326, %r325, -2147483648;
mov.b32 %f1179, %r326;
selp.f32 %f1910, %f1179, %f1908, %p5;
setp.geu.f32 %p95, %f1841, 0f00000000;
@%p95 bra $L__BB0_79;
mov.f32 %f1180, 0f3EE66666;
cvt.rzi.f32.f32 %f1181, %f1180;
setp.eq.f32 %p96, %f1181, 0f3EE66666;
@%p96 bra $L__BB0_79;
mov.f32 %f1910, 0f7FFFFFFF;
$L__BB0_79:
add.f32 %f1184, %f315, 0f3EE66666;
mov.b32 %r327, %f1184;
setp.lt.s32 %p98, %r327, 2139095040;
@%p98 bra $L__BB0_84;
setp.gtu.f32 %p99, %f315, 0f7F800000;
@%p99 bra $L__BB0_83;
bra.uni $L__BB0_81;
$L__BB0_83:
add.f32 %f1910, %f1841, 0f3EE66666;
bra.uni $L__BB0_84;
$L__BB0_81:
setp.neu.f32 %p100, %f315, 0f7F800000;
@%p100 bra $L__BB0_84;
selp.f32 %f1910, 0fFF800000, 0f7F800000, %p5;
$L__BB0_84:
setp.eq.f32 %p101, %f1841, 0f3F800000;
selp.f32 %f326, 0f3F800000, %f1910, %p101;
abs.f32 %f327, %f1842;
setp.lt.f32 %p102, %f327, 0f00800000;
mul.f32 %f1186, %f327, 0f4B800000;
selp.f32 %f1187, %f1186, %f327, %p102;
selp.f32 %f1188, 0fC3170000, 0fC2FE0000, %p102;
mov.b32 %r328, %f1187;
and.b32 %r329, %r328, 8388607;
or.b32 %r330, %r329, 1065353216;
mov.b32 %f1189, %r330;
shr.u32 %r331, %r328, 23;
cvt.rn.f32.u32 %f1190, %r331;
add.f32 %f1191, %f1188, %f1190;
setp.gt.f32 %p103, %f1189, 0f3FB504F3;
mul.f32 %f1192, %f1189, 0f3F000000;
add.f32 %f1193, %f1191, 0f3F800000;
selp.f32 %f1194, %f1193, %f1191, %p103;
selp.f32 %f1195, %f1192, %f1189, %p103;
add.f32 %f1196, %f1195, 0fBF800000;
add.f32 %f1197, %f1195, 0f3F800000;
rcp.approx.ftz.f32 %f1198, %f1197;
add.f32 %f1199, %f1196, %f1196;
mul.f32 %f1200, %f1199, %f1198;
mul.f32 %f1201, %f1200, %f1200;
mov.f32 %f1202, 0f3C4CAF63;
mov.f32 %f1203, 0f3B18F0FE;
fma.rn.f32 %f1204, %f1203, %f1201, %f1202;
mov.f32 %f1205, 0f3DAAAABD;
fma.rn.f32 %f1206, %f1204, %f1201, %f1205;
mul.rn.f32 %f1207, %f1206, %f1201;
mul.rn.f32 %f1208, %f1207, %f1200;
sub.f32 %f1209, %f1196, %f1200;
add.f32 %f1210, %f1209, %f1209;
neg.f32 %f1211, %f1200;
fma.rn.f32 %f1212, %f1211, %f1196, %f1210;
mul.rn.f32 %f1213, %f1198, %f1212;
add.f32 %f1214, %f1208, %f1200;
sub.f32 %f1215, %f1200, %f1214;
add.f32 %f1216, %f1208, %f1215;
add.f32 %f1217, %f1213, %f1216;
add.f32 %f1218, %f1214, %f1217;
sub.f32 %f1219, %f1214, %f1218;
add.f32 %f1220, %f1217, %f1219;
mov.f32 %f1221, 0f3F317200;
mul.rn.f32 %f1222, %f1194, %f1221;
mov.f32 %f1223, 0f35BFBE8E;
mul.rn.f32 %f1224, %f1194, %f1223;
add.f32 %f1225, %f1222, %f1218;
sub.f32 %f1226, %f1222, %f1225;
add.f32 %f1227, %f1218, %f1226;
add.f32 %f1228, %f1220, %f1227;
add.f32 %f1229, %f1224, %f1228;
add.f32 %f1230, %f1225, %f1229;
sub.f32 %f1231, %f1225, %f1230;
add.f32 %f1232, %f1229, %f1231;
mov.f32 %f1233, 0f3EE66666;
mul.rn.f32 %f1234, %f1233, %f1230;
neg.f32 %f1235, %f1234;
fma.rn.f32 %f1236, %f1233, %f1230, %f1235;
fma.rn.f32 %f1237, %f1233, %f1232, %f1236;
mov.f32 %f1238, 0f00000000;
fma.rn.f32 %f1239, %f1238, %f1230, %f1237;
add.rn.f32 %f1240, %f1234, %f1239;
neg.f32 %f1241, %f1240;
add.rn.f32 %f1242, %f1234, %f1241;
add.rn.f32 %f1243, %f1242, %f1239;
mov.b32 %r332, %f1240;
setp.eq.s32 %p104, %r332, 1118925336;
add.s32 %r333, %r332, -1;
mov.b32 %f1244, %r333;
add.f32 %f1245, %f1243, 0f37000000;
selp.f32 %f328, %f1245, %f1243, %p104;
selp.f32 %f1246, %f1244, %f1240, %p104;
mov.f32 %f1247, 0f3FB8AA3B;
mul.rn.f32 %f1248, %f1246, %f1247;
cvt.rzi.f32.f32 %f1249, %f1248;
abs.f32 %f1250, %f1249;
setp.gt.f32 %p105, %f1250, 0f42FC0000;
mov.b32 %r334, %f1249;
and.b32 %r335, %r334, -2147483648;
or.b32 %r336, %r335, 1123811328;
mov.b32 %f1251, %r336;
selp.f32 %f1252, %f1251, %f1249, %p105;
mov.f32 %f1253, 0fBF317218;
fma.rn.f32 %f1254, %f1252, %f1253, %f1246;
mov.f32 %f1255, 0f3102E308;
fma.rn.f32 %f1256, %f1252, %f1255, %f1254;
mul.f32 %f1257, %f1256, 0f3FB8AA3B;
add.f32 %f1258, %f1252, 0f4B40007F;
mov.b32 %r337, %f1258;
shl.b32 %r338, %r337, 23;
mov.b32 %f1259, %r338;
ex2.approx.ftz.f32 %f1260, %f1257;
mul.f32 %f329, %f1260, %f1259;
setp.eq.f32 %p106, %f329, 0f7F800000;
mov.f32 %f1911, 0f7F800000;
@%p106 bra $L__BB0_86;
fma.rn.f32 %f1911, %f329, %f328, %f329;
$L__BB0_86:
setp.lt.f32 %p107, %f1842, 0f00000000;
and.pred %p6, %p107, %p78;
setp.eq.f32 %p109, %f1842, 0f00000000;
@%p109 bra $L__BB0_90;
bra.uni $L__BB0_87;
$L__BB0_90:
add.f32 %f1265, %f1842, %f1842;
selp.f32 %f1913, %f1265, 0f00000000, %p78;
bra.uni $L__BB0_91;
$L__BB0_87:
mov.b32 %r339, %f1911;
xor.b32 %r340, %r339, -2147483648;
mov.b32 %f1261, %r340;
selp.f32 %f1913, %f1261, %f1911, %p6;
setp.geu.f32 %p110, %f1842, 0f00000000;
@%p110 bra $L__BB0_91;
mov.f32 %f1262, 0f3EE66666;
cvt.rzi.f32.f32 %f1263, %f1262;
setp.eq.f32 %p111, %f1263, 0f3EE66666;
@%p111 bra $L__BB0_91;
mov.f32 %f1913, 0f7FFFFFFF;
$L__BB0_91:
add.f32 %f1266, %f327, 0f3EE66666;
mov.b32 %r341, %f1266;
setp.lt.s32 %p113, %r341, 2139095040;
@%p113 bra $L__BB0_96;
setp.gtu.f32 %p114, %f327, 0f7F800000;
@%p114 bra $L__BB0_95;
bra.uni $L__BB0_93;
$L__BB0_95:
add.f32 %f1913, %f1842, 0f3EE66666;
bra.uni $L__BB0_96;
$L__BB0_93:
setp.neu.f32 %p115, %f327, 0f7F800000;
@%p115 bra $L__BB0_96;
selp.f32 %f1913, 0fFF800000, 0f7F800000, %p6;
$L__BB0_96:
setp.eq.f32 %p116, %f1842, 0f3F800000;
mov.f32 %f1268, 0f3F800000;
selp.f32 %f1269, 0f3F800000, %f1913, %p116;
ld.const.u64 %rd46, [params+144];
cvta.to.global.u64 %rd9, %rd46;
ld.const.u32 %r342, [params+136];
mad.lo.s32 %r343, %r342, %r4, %r3;
cvt.u64.u32 %rd10, %r343;
min.f32 %f1270, %f314, %f1268;
mov.f32 %f1271, 0f00000000;
max.f32 %f338, %f1271, %f1270;
min.f32 %f1272, %f326, %f1268;
max.f32 %f339, %f1271, %f1272;
min.f32 %f1273, %f1269, %f1268;
max.f32 %f340, %f1271, %f1273;
mov.f32 %f1274, 0f3E555555;
cvt.rzi.f32.f32 %f1275, %f1274;
add.f32 %f1276, %f1275, %f1275;
mov.f32 %f1277, 0f3ED55555;
sub.f32 %f1278, %f1277, %f1276;
abs.f32 %f341, %f1278;
abs.f32 %f342, %f338;
setp.lt.f32 %p117, %f342, 0f00800000;
mul.f32 %f1279, %f342, 0f4B800000;
selp.f32 %f1280, %f1279, %f342, %p117;
selp.f32 %f1281, 0fC3170000, 0fC2FE0000, %p117;
mov.b32 %r344, %f1280;
and.b32 %r345, %r344, 8388607;
or.b32 %r346, %r345, 1065353216;
mov.b32 %f1282, %r346;
shr.u32 %r347, %r344, 23;
cvt.rn.f32.u32 %f1283, %r347;
add.f32 %f1284, %f1281, %f1283;
setp.gt.f32 %p118, %f1282, 0f3FB504F3;
mul.f32 %f1285, %f1282, 0f3F000000;
add.f32 %f1286, %f1284, 0f3F800000;
selp.f32 %f1287, %f1286, %f1284, %p118;
selp.f32 %f1288, %f1285, %f1282, %p118;
add.f32 %f1289, %f1288, 0fBF800000;
add.f32 %f1290, %f1288, 0f3F800000;
rcp.approx.ftz.f32 %f1291, %f1290;
add.f32 %f1292, %f1289, %f1289;
mul.f32 %f1293, %f1292, %f1291;
mul.f32 %f1294, %f1293, %f1293;
mov.f32 %f1295, 0f3C4CAF63;
mov.f32 %f1296, 0f3B18F0FE;
fma.rn.f32 %f1297, %f1296, %f1294, %f1295;
mov.f32 %f1298, 0f3DAAAABD;
fma.rn.f32 %f1299, %f1297, %f1294, %f1298;
mul.rn.f32 %f1300, %f1299, %f1294;
mul.rn.f32 %f1301, %f1300, %f1293;
sub.f32 %f1302, %f1289, %f1293;
add.f32 %f1303, %f1302, %f1302;
neg.f32 %f1304, %f1293;
fma.rn.f32 %f1305, %f1304, %f1289, %f1303;
mul.rn.f32 %f1306, %f1291, %f1305;
add.f32 %f1307, %f1301, %f1293;
sub.f32 %f1308, %f1293, %f1307;
add.f32 %f1309, %f1301, %f1308;
add.f32 %f1310, %f1306, %f1309;
add.f32 %f1311, %f1307, %f1310;
sub.f32 %f1312, %f1307, %f1311;
add.f32 %f1313, %f1310, %f1312;
mov.f32 %f1314, 0f3F317200;
mul.rn.f32 %f1315, %f1287, %f1314;
mov.f32 %f1316, 0f35BFBE8E;
mul.rn.f32 %f1317, %f1287, %f1316;
add.f32 %f1318, %f1315, %f1311;
sub.f32 %f1319, %f1315, %f1318;
add.f32 %f1320, %f1311, %f1319;
add.f32 %f1321, %f1313, %f1320;
add.f32 %f1322, %f1317, %f1321;
add.f32 %f1323, %f1318, %f1322;
sub.f32 %f1324, %f1318, %f1323;
add.f32 %f1325, %f1322, %f1324;
mul.rn.f32 %f1326, %f1277, %f1323;
neg.f32 %f1327, %f1326;
fma.rn.f32 %f1328, %f1277, %f1323, %f1327;
fma.rn.f32 %f1329, %f1277, %f1325, %f1328;
fma.rn.f32 %f1330, %f1271, %f1323, %f1329;
add.rn.f32 %f1331, %f1326, %f1330;
neg.f32 %f1332, %f1331;
add.rn.f32 %f1333, %f1326, %f1332;
add.rn.f32 %f1334, %f1333, %f1330;
mov.b32 %r348, %f1331;
setp.eq.s32 %p119, %r348, 1118925336;
add.s32 %r349, %r348, -1;
mov.b32 %f1335, %r349;
add.f32 %f1336, %f1334, 0f37000000;
selp.f32 %f343, %f1336, %f1334, %p119;
selp.f32 %f1337, %f1335, %f1331, %p119;
mov.f32 %f1338, 0f3FB8AA3B;
mul.rn.f32 %f1339, %f1337, %f1338;
cvt.rzi.f32.f32 %f1340, %f1339;
abs.f32 %f1341, %f1340;
setp.gt.f32 %p120, %f1341, 0f42FC0000;
mov.b32 %r350, %f1340;
and.b32 %r351, %r350, -2147483648;
or.b32 %r352, %r351, 1123811328;
mov.b32 %f1342, %r352;
selp.f32 %f1343, %f1342, %f1340, %p120;
mov.f32 %f1344, 0fBF317218;
fma.rn.f32 %f1345, %f1343, %f1344, %f1337;
mov.f32 %f1346, 0f3102E308;
fma.rn.f32 %f1347, %f1343, %f1346, %f1345;
mul.f32 %f1348, %f1347, 0f3FB8AA3B;
add.f32 %f1349, %f1343, 0f4B40007F;
mov.b32 %r353, %f1349;
shl.b32 %r354, %r353, 23;
mov.b32 %f1350, %r354;
ex2.approx.ftz.f32 %f1351, %f1348;
mul.f32 %f344, %f1351, %f1350;
setp.eq.f32 %p121, %f344, 0f7F800000;
mov.f32 %f1914, 0f7F800000;
@%p121 bra $L__BB0_98;
fma.rn.f32 %f1914, %f344, %f343, %f344;
$L__BB0_98:
setp.lt.f32 %p122, %f338, 0f00000000;
setp.eq.f32 %p123, %f341, 0f3F800000;
and.pred %p7, %p122, %p123;
setp.eq.f32 %p124, %f338, 0f00000000;
@%p124 bra $L__BB0_102;
bra.uni $L__BB0_99;
$L__BB0_102:
add.f32 %f1356, %f338, %f338;
selp.f32 %f1916, %f1356, 0f00000000, %p123;
bra.uni $L__BB0_103;
$L__BB0_99:
mov.b32 %r355, %f1914;
xor.b32 %r356, %r355, -2147483648;
mov.b32 %f1352, %r356;
selp.f32 %f1916, %f1352, %f1914, %p7;
setp.geu.f32 %p125, %f338, 0f00000000;
@%p125 bra $L__BB0_103;
mov.f32 %f1353, 0f3ED55555;
cvt.rzi.f32.f32 %f1354, %f1353;
setp.eq.f32 %p126, %f1354, 0f3ED55555;
@%p126 bra $L__BB0_103;
mov.f32 %f1916, 0f7FFFFFFF;
$L__BB0_103:
add.f32 %f1357, %f342, 0f3ED55555;
mov.b32 %r357, %f1357;
setp.lt.s32 %p128, %r357, 2139095040;
@%p128 bra $L__BB0_108;
setp.gtu.f32 %p129, %f342, 0f7F800000;
@%p129 bra $L__BB0_107;
bra.uni $L__BB0_105;
$L__BB0_107:
add.f32 %f1916, %f338, 0f3ED55555;
bra.uni $L__BB0_108;
$L__BB0_105:
setp.neu.f32 %p130, %f342, 0f7F800000;
@%p130 bra $L__BB0_108;
selp.f32 %f1916, 0fFF800000, 0f7F800000, %p7;
$L__BB0_108:
abs.f32 %f353, %f339;
setp.lt.f32 %p131, %f353, 0f00800000;
mul.f32 %f1359, %f353, 0f4B800000;
selp.f32 %f1360, %f1359, %f353, %p131;
selp.f32 %f1361, 0fC3170000, 0fC2FE0000, %p131;
mov.b32 %r358, %f1360;
and.b32 %r359, %r358, 8388607;
or.b32 %r360, %r359, 1065353216;
mov.b32 %f1362, %r360;
shr.u32 %r361, %r358, 23;
cvt.rn.f32.u32 %f1363, %r361;
add.f32 %f1364, %f1361, %f1363;
setp.gt.f32 %p132, %f1362, 0f3FB504F3;
mul.f32 %f1365, %f1362, 0f3F000000;
add.f32 %f1366, %f1364, 0f3F800000;
selp.f32 %f1367, %f1366, %f1364, %p132;
selp.f32 %f1368, %f1365, %f1362, %p132;
add.f32 %f1369, %f1368, 0fBF800000;
add.f32 %f1370, %f1368, 0f3F800000;
rcp.approx.ftz.f32 %f1371, %f1370;
add.f32 %f1372, %f1369, %f1369;
mul.f32 %f1373, %f1372, %f1371;
mul.f32 %f1374, %f1373, %f1373;
mov.f32 %f1375, 0f3C4CAF63;
mov.f32 %f1376, 0f3B18F0FE;
fma.rn.f32 %f1377, %f1376, %f1374, %f1375;
mov.f32 %f1378, 0f3DAAAABD;
fma.rn.f32 %f1379, %f1377, %f1374, %f1378;
mul.rn.f32 %f1380, %f1379, %f1374;
mul.rn.f32 %f1381, %f1380, %f1373;
sub.f32 %f1382, %f1369, %f1373;
add.f32 %f1383, %f1382, %f1382;
neg.f32 %f1384, %f1373;
fma.rn.f32 %f1385, %f1384, %f1369, %f1383;
mul.rn.f32 %f1386, %f1371, %f1385;
add.f32 %f1387, %f1381, %f1373;
sub.f32 %f1388, %f1373, %f1387;
add.f32 %f1389, %f1381, %f1388;
add.f32 %f1390, %f1386, %f1389;
add.f32 %f1391, %f1387, %f1390;
sub.f32 %f1392, %f1387, %f1391;
add.f32 %f1393, %f1390, %f1392;
mov.f32 %f1394, 0f3F317200;
mul.rn.f32 %f1395, %f1367, %f1394;
mov.f32 %f1396, 0f35BFBE8E;
mul.rn.f32 %f1397, %f1367, %f1396;
add.f32 %f1398, %f1395, %f1391;
sub.f32 %f1399, %f1395, %f1398;
add.f32 %f1400, %f1391, %f1399;
add.f32 %f1401, %f1393, %f1400;
add.f32 %f1402, %f1397, %f1401;
add.f32 %f1403, %f1398, %f1402;
sub.f32 %f1404, %f1398, %f1403;
add.f32 %f1405, %f1402, %f1404;
mov.f32 %f1406, 0f3ED55555;
mul.rn.f32 %f1407, %f1406, %f1403;
neg.f32 %f1408, %f1407;
fma.rn.f32 %f1409, %f1406, %f1403, %f1408;
fma.rn.f32 %f1410, %f1406, %f1405, %f1409;
mov.f32 %f1411, 0f00000000;
fma.rn.f32 %f1412, %f1411, %f1403, %f1410;
add.rn.f32 %f1413, %f1407, %f1412;
neg.f32 %f1414, %f1413;
add.rn.f32 %f1415, %f1407, %f1414;
add.rn.f32 %f1416, %f1415, %f1412;
mov.b32 %r362, %f1413;
setp.eq.s32 %p133, %r362, 1118925336;
add.s32 %r363, %r362, -1;
mov.b32 %f1417, %r363;
add.f32 %f1418, %f1416, 0f37000000;
selp.f32 %f354, %f1418, %f1416, %p133;
selp.f32 %f1419, %f1417, %f1413, %p133;
mov.f32 %f1420, 0f3FB8AA3B;
mul.rn.f32 %f1421, %f1419, %f1420;
cvt.rzi.f32.f32 %f1422, %f1421;
abs.f32 %f1423, %f1422;
setp.gt.f32 %p134, %f1423, 0f42FC0000;
mov.b32 %r364, %f1422;
and.b32 %r365, %r364, -2147483648;
or.b32 %r366, %r365, 1123811328;
mov.b32 %f1424, %r366;
selp.f32 %f1425, %f1424, %f1422, %p134;
mov.f32 %f1426, 0fBF317218;
fma.rn.f32 %f1427, %f1425, %f1426, %f1419;
mov.f32 %f1428, 0f3102E308;
fma.rn.f32 %f1429, %f1425, %f1428, %f1427;
mul.f32 %f1430, %f1429, 0f3FB8AA3B;
add.f32 %f1431, %f1425, 0f4B40007F;
mov.b32 %r367, %f1431;
shl.b32 %r368, %r367, 23;
mov.b32 %f1432, %r368;
ex2.approx.ftz.f32 %f1433, %f1430;
mul.f32 %f355, %f1433, %f1432;
setp.eq.f32 %p135, %f355, 0f7F800000;
mov.f32 %f1917, 0f7F800000;
@%p135 bra $L__BB0_110;
fma.rn.f32 %f1917, %f355, %f354, %f355;
$L__BB0_110:
setp.lt.f32 %p136, %f339, 0f00000000;
and.pred %p8, %p136, %p123;
setp.eq.f32 %p138, %f339, 0f00000000;
@%p138 bra $L__BB0_114;
bra.uni $L__BB0_111;
$L__BB0_114:
add.f32 %f1438, %f339, %f339;
selp.f32 %f1919, %f1438, 0f00000000, %p123;
bra.uni $L__BB0_115;
$L__BB0_111:
mov.b32 %r369, %f1917;
xor.b32 %r370, %r369, -2147483648;
mov.b32 %f1434, %r370;
selp.f32 %f1919, %f1434, %f1917, %p8;
setp.geu.f32 %p139, %f339, 0f00000000;
@%p139 bra $L__BB0_115;
mov.f32 %f1435, 0f3ED55555;
cvt.rzi.f32.f32 %f1436, %f1435;
setp.eq.f32 %p140, %f1436, 0f3ED55555;
@%p140 bra $L__BB0_115;
mov.f32 %f1919, 0f7FFFFFFF;
$L__BB0_115:
add.f32 %f1439, %f353, 0f3ED55555;
mov.b32 %r371, %f1439;
setp.lt.s32 %p142, %r371, 2139095040;
@%p142 bra $L__BB0_120;
setp.gtu.f32 %p143, %f353, 0f7F800000;
@%p143 bra $L__BB0_119;
bra.uni $L__BB0_117;
$L__BB0_119:
add.f32 %f1919, %f339, 0f3ED55555;
bra.uni $L__BB0_120;
$L__BB0_117:
setp.neu.f32 %p144, %f353, 0f7F800000;
@%p144 bra $L__BB0_120;
selp.f32 %f1919, 0fFF800000, 0f7F800000, %p8;
$L__BB0_120:
abs.f32 %f364, %f340;
setp.lt.f32 %p145, %f364, 0f00800000;
mul.f32 %f1441, %f364, 0f4B800000;
selp.f32 %f1442, %f1441, %f364, %p145;
selp.f32 %f1443, 0fC3170000, 0fC2FE0000, %p145;
mov.b32 %r372, %f1442;
and.b32 %r373, %r372, 8388607;
or.b32 %r374, %r373, 1065353216;
mov.b32 %f1444, %r374;
shr.u32 %r375, %r372, 23;
cvt.rn.f32.u32 %f1445, %r375;
add.f32 %f1446, %f1443, %f1445;
setp.gt.f32 %p146, %f1444, 0f3FB504F3;
mul.f32 %f1447, %f1444, 0f3F000000;
add.f32 %f1448, %f1446, 0f3F800000;
selp.f32 %f1449, %f1448, %f1446, %p146;
selp.f32 %f1450, %f1447, %f1444, %p146;
add.f32 %f1451, %f1450, 0fBF800000;
add.f32 %f1452, %f1450, 0f3F800000;
rcp.approx.ftz.f32 %f1453, %f1452;
add.f32 %f1454, %f1451, %f1451;
mul.f32 %f1455, %f1454, %f1453;
mul.f32 %f1456, %f1455, %f1455;
mov.f32 %f1457, 0f3C4CAF63;
mov.f32 %f1458, 0f3B18F0FE;
fma.rn.f32 %f1459, %f1458, %f1456, %f1457;
mov.f32 %f1460, 0f3DAAAABD;
fma.rn.f32 %f1461, %f1459, %f1456, %f1460;
mul.rn.f32 %f1462, %f1461, %f1456;
mul.rn.f32 %f1463, %f1462, %f1455;
sub.f32 %f1464, %f1451, %f1455;
add.f32 %f1465, %f1464, %f1464;
neg.f32 %f1466, %f1455;
fma.rn.f32 %f1467, %f1466, %f1451, %f1465;
mul.rn.f32 %f1468, %f1453, %f1467;
add.f32 %f1469, %f1463, %f1455;
sub.f32 %f1470, %f1455, %f1469;
add.f32 %f1471, %f1463, %f1470;
add.f32 %f1472, %f1468, %f1471;
add.f32 %f1473, %f1469, %f1472;
sub.f32 %f1474, %f1469, %f1473;
add.f32 %f1475, %f1472, %f1474;
mov.f32 %f1476, 0f3F317200;
mul.rn.f32 %f1477, %f1449, %f1476;
mov.f32 %f1478, 0f35BFBE8E;
mul.rn.f32 %f1479, %f1449, %f1478;
add.f32 %f1480, %f1477, %f1473;
sub.f32 %f1481, %f1477, %f1480;
add.f32 %f1482, %f1473, %f1481;
add.f32 %f1483, %f1475, %f1482;
add.f32 %f1484, %f1479, %f1483;
add.f32 %f1485, %f1480, %f1484;
sub.f32 %f1486, %f1480, %f1485;
add.f32 %f1487, %f1484, %f1486;
mov.f32 %f1488, 0f3ED55555;
mul.rn.f32 %f1489, %f1488, %f1485;
neg.f32 %f1490, %f1489;
fma.rn.f32 %f1491, %f1488, %f1485, %f1490;
fma.rn.f32 %f1492, %f1488, %f1487, %f1491;
mov.f32 %f1493, 0f00000000;
fma.rn.f32 %f1494, %f1493, %f1485, %f1492;
add.rn.f32 %f1495, %f1489, %f1494;
neg.f32 %f1496, %f1495;
add.rn.f32 %f1497, %f1489, %f1496;
add.rn.f32 %f1498, %f1497, %f1494;
mov.b32 %r376, %f1495;
setp.eq.s32 %p147, %r376, 1118925336;
add.s32 %r377, %r376, -1;
mov.b32 %f1499, %r377;
add.f32 %f1500, %f1498, 0f37000000;
selp.f32 %f365, %f1500, %f1498, %p147;
selp.f32 %f1501, %f1499, %f1495, %p147;
mov.f32 %f1502, 0f3FB8AA3B;
mul.rn.f32 %f1503, %f1501, %f1502;
cvt.rzi.f32.f32 %f1504, %f1503;
abs.f32 %f1505, %f1504;
setp.gt.f32 %p148, %f1505, 0f42FC0000;
mov.b32 %r378, %f1504;
and.b32 %r379, %r378, -2147483648;
or.b32 %r380, %r379, 1123811328;
mov.b32 %f1506, %r380;
selp.f32 %f1507, %f1506, %f1504, %p148;
mov.f32 %f1508, 0fBF317218;
fma.rn.f32 %f1509, %f1507, %f1508, %f1501;
mov.f32 %f1510, 0f3102E308;
fma.rn.f32 %f1511, %f1507, %f1510, %f1509;
mul.f32 %f1512, %f1511, 0f3FB8AA3B;
add.f32 %f1513, %f1507, 0f4B40007F;
mov.b32 %r381, %f1513;
shl.b32 %r382, %r381, 23;
mov.b32 %f1514, %r382;
ex2.approx.ftz.f32 %f1515, %f1512;
mul.f32 %f366, %f1515, %f1514;
setp.eq.f32 %p149, %f366, 0f7F800000;
mov.f32 %f1920, 0f7F800000;
@%p149 bra $L__BB0_122;
fma.rn.f32 %f1920, %f366, %f365, %f366;
$L__BB0_122:
setp.lt.f32 %p150, %f340, 0f00000000;
and.pred %p9, %p150, %p123;
setp.eq.f32 %p152, %f340, 0f00000000;
@%p152 bra $L__BB0_126;
bra.uni $L__BB0_123;
$L__BB0_126:
add.f32 %f1520, %f340, %f340;
selp.f32 %f1922, %f1520, 0f00000000, %p123;
bra.uni $L__BB0_127;
$L__BB0_123:
mov.b32 %r383, %f1920;
xor.b32 %r384, %r383, -2147483648;
mov.b32 %f1516, %r384;
selp.f32 %f1922, %f1516, %f1920, %p9;
setp.geu.f32 %p153, %f340, 0f00000000;
@%p153 bra $L__BB0_127;
mov.f32 %f1517, 0f3ED55555;
cvt.rzi.f32.f32 %f1518, %f1517;
setp.eq.f32 %p154, %f1518, 0f3ED55555;
@%p154 bra $L__BB0_127;
mov.f32 %f1922, 0f7FFFFFFF;
$L__BB0_127:
add.f32 %f1521, %f364, 0f3ED55555;
mov.b32 %r385, %f1521;
setp.lt.s32 %p156, %r385, 2139095040;
@%p156 bra $L__BB0_132;
setp.gtu.f32 %p157, %f364, 0f7F800000;
@%p157 bra $L__BB0_131;
bra.uni $L__BB0_129;
$L__BB0_131:
add.f32 %f1922, %f340, 0f3ED55555;
bra.uni $L__BB0_132;
$L__BB0_129:
setp.neu.f32 %p158, %f364, 0f7F800000;
@%p158 bra $L__BB0_132;
selp.f32 %f1922, 0fFF800000, 0f7F800000, %p9;
$L__BB0_132:
fma.rn.f32 %f1522, %f1916, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p159, %f338, 0f3F800000;
mov.f32 %f1523, 0f3F800000;
selp.f32 %f1524, 0f3F7FFFFF, %f1522, %p159;
mul.f32 %f1525, %f338, 0f414EB852;
setp.lt.f32 %p160, %f338, 0f3B4D2E1C;
selp.f32 %f1526, %f1525, %f1524, %p160;
fma.rn.f32 %f1527, %f1919, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p161, %f339, 0f3F800000;
selp.f32 %f1528, 0f3F7FFFFF, %f1527, %p161;
mul.f32 %f1529, %f339, 0f414EB852;
setp.lt.f32 %p162, %f339, 0f3B4D2E1C;
selp.f32 %f1530, %f1529, %f1528, %p162;
fma.rn.f32 %f1531, %f1922, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p163, %f340, 0f3F800000;
selp.f32 %f1532, 0f3F7FFFFF, %f1531, %p163;
mul.f32 %f1533, %f340, 0f414EB852;
setp.lt.f32 %p164, %f340, 0f3B4D2E1C;
selp.f32 %f1534, %f1533, %f1532, %p164;
min.f32 %f1535, %f1526, %f1523;
mov.f32 %f1536, 0f00000000;
max.f32 %f1537, %f1536, %f1535;
mul.f32 %f1538, %f1537, 0f43800000;
cvt.rzi.u32.f32 %r386, %f1538;
min.u32 %r387, %r386, 255;
min.f32 %f1539, %f1530, %f1523;
max.f32 %f1540, %f1536, %f1539;
mul.f32 %f1541, %f1540, 0f43800000;
cvt.rzi.u32.f32 %r388, %f1541;
min.u32 %r389, %r388, 255;
min.f32 %f1542, %f1534, %f1523;
max.f32 %f1543, %f1536, %f1542;
mul.f32 %f1544, %f1543, 0f43800000;
cvt.rzi.u32.f32 %r390, %f1544;
min.u32 %r391, %r390, 255;
shl.b64 %rd47, %rd10, 2;
add.s64 %rd48, %rd9, %rd47;
cvt.u16.u32 %rs20, %r391;
cvt.u16.u32 %rs21, %r389;
cvt.u16.u32 %rs22, %r387;
mov.u16 %rs23, 255;
st.global.v4.u8 [%rd48], {%rs22, %rs21, %rs20, %rs23};
$L__BB0_133:
and.b32 %r392, %r447, 4;
setp.eq.s32 %p165, %r392, 0;
ld.const.u32 %r448, [params+108];
@%p165 bra $L__BB0_137;
setp.eq.s32 %p166, %r448, 0;
ld.const.u64 %rd49, [params+224];
cvta.to.global.u64 %rd50, %rd49;
ld.const.u32 %r393, [params+216];
mad.lo.s32 %r394, %r393, %r4, %r3;
mul.wide.u32 %rd51, %r394, 8;
add.s64 %rd11, %rd50, %rd51;
@%p166 bra $L__BB0_136;
ld.global.v4.u16 {%rs31, %rs32, %rs33, %rs34}, [%rd11];
// begin inline asm
{ cvt.f32.f16 %f1545, %rs31;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1546, %rs32;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1547, %rs33;}
// end inline asm
add.f32 %f1548, %f1840, %f1545;
add.f32 %f1549, %f1841, %f1546;
add.f32 %f1550, %f1842, %f1547;
mov.f32 %f1551, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f1550;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1549;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1548;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f1551;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs27, %rs28, %rs29, %rs30};
bra.uni $L__BB0_137;
$L__BB0_136:
mov.f32 %f1555, 0f3F800000;
mov.u32 %r448, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f1555;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f1842;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f1841;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f1840;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs35, %rs36, %rs37, %rs38};
$L__BB0_137:
mul.f32 %f375, %f1843, 0f3F558750;
mul.f32 %f376, %f1844, 0f3F558750;
mul.f32 %f377, %f1845, 0f3F558750;
ld.const.u64 %rd52, [params+256];
cvta.to.global.u64 %rd53, %rd52;
ld.const.u32 %r396, [params+248];
mad.lo.s32 %r397, %r396, %r4, %r3;
mul.wide.u32 %rd54, %r397, 8;
add.s64 %rd12, %rd53, %rd54;
setp.eq.s32 %p167, %r448, 0;
@%p167 bra $L__BB0_139;
ld.global.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd12];
// begin inline asm
{ cvt.f32.f16 %f1556, %rs46;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1557, %rs47;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1558, %rs48;}
// end inline asm
add.f32 %f1559, %f375, %f1556;
add.f32 %f1560, %f376, %f1557;
add.f32 %f1561, %f377, %f1558;
mov.f32 %f1562, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f1561;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f1560;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs42, %f1559;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f1562;}
// end inline asm
st.global.v4.u16 [%rd12], {%rs42, %rs43, %rs44, %rs45};
bra.uni $L__BB0_140;
$L__BB0_139:
mov.f32 %f1566, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs53, %f1566;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs52, %f377;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f376;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f375;}
// end inline asm
st.global.v4.u16 [%rd12], {%rs50, %rs51, %rs52, %rs53};
$L__BB0_140:
mul.f32 %f378, %f1846, 0f3FB8EBD1;
mul.f32 %f379, %f1847, 0f3FB8EBD1;
mul.f32 %f380, %f1848, 0f3FB8EBD1;
ld.const.u64 %rd55, [params+272];
cvta.to.global.u64 %rd56, %rd55;
ld.const.u32 %r398, [params+264];
mad.lo.s32 %r399, %r398, %r4, %r3;
mul.wide.u32 %rd57, %r399, 8;
add.s64 %rd13, %rd56, %rd57;
@%p167 bra $L__BB0_142;
ld.global.v4.u16 {%rs61, %rs62, %rs63, %rs64}, [%rd13];
// begin inline asm
{ cvt.f32.f16 %f1567, %rs61;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1568, %rs62;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1569, %rs63;}
// end inline asm
add.f32 %f1570, %f378, %f1567;
add.f32 %f1571, %f379, %f1568;
add.f32 %f1572, %f380, %f1569;
mov.f32 %f1573, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs59, %f1572;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs58, %f1571;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs57, %f1570;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs60, %f1573;}
// end inline asm
st.global.v4.u16 [%rd13], {%rs57, %rs58, %rs59, %rs60};
bra.uni $L__BB0_143;
$L__BB0_142:
mov.f32 %f1577, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs68, %f1577;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs67, %f380;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs66, %f379;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs65, %f378;}
// end inline asm
st.global.v4.u16 [%rd13], {%rs65, %rs66, %rs67, %rs68};
$L__BB0_143:
mul.f32 %f381, %f1849, 0f3FB8EBD1;
mul.f32 %f382, %f1850, 0f3FB8EBD1;
mul.f32 %f383, %f1851, 0f3FB8EBD1;
ld.const.u64 %rd58, [params+288];
cvta.to.global.u64 %rd59, %rd58;
ld.const.u32 %r400, [params+280];
mad.lo.s32 %r401, %r400, %r4, %r3;
mul.wide.u32 %rd60, %r401, 8;
add.s64 %rd14, %rd59, %rd60;
@%p167 bra $L__BB0_145;
ld.global.v4.u16 {%rs76, %rs77, %rs78, %rs79}, [%rd14];
// begin inline asm
{ cvt.f32.f16 %f1578, %rs76;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1579, %rs77;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1580, %rs78;}
// end inline asm
add.f32 %f1581, %f381, %f1578;
add.f32 %f1582, %f382, %f1579;
add.f32 %f1583, %f383, %f1580;
mov.f32 %f1584, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs74, %f1583;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs73, %f1582;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs72, %f1581;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs75, %f1584;}
// end inline asm
st.global.v4.u16 [%rd14], {%rs72, %rs73, %rs74, %rs75};
bra.uni $L__BB0_146;
$L__BB0_145:
mov.f32 %f1588, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs83, %f1588;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs82, %f383;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs81, %f382;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs80, %f381;}
// end inline asm
st.global.v4.u16 [%rd14], {%rs80, %rs81, %rs82, %rs83};
$L__BB0_146:
mul.f32 %f384, %f1852, 0f3FB8EBD1;
mul.f32 %f385, %f1853, 0f3FB8EBD1;
mul.f32 %f386, %f1854, 0f3FB8EBD1;
ld.const.u64 %rd61, [params+304];
cvta.to.global.u64 %rd62, %rd61;
ld.const.u32 %r402, [params+296];
mad.lo.s32 %r403, %r402, %r4, %r3;
mul.wide.u32 %rd63, %r403, 8;
add.s64 %rd15, %rd62, %rd63;
@%p167 bra $L__BB0_148;
ld.global.v4.u16 {%rs91, %rs92, %rs93, %rs94}, [%rd15];
// begin inline asm
{ cvt.f32.f16 %f1589, %rs91;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1590, %rs92;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1591, %rs93;}
// end inline asm
add.f32 %f1592, %f384, %f1589;
add.f32 %f1593, %f385, %f1590;
add.f32 %f1594, %f386, %f1591;
mov.f32 %f1595, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs89, %f1594;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs88, %f1593;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs87, %f1592;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs90, %f1595;}
// end inline asm
st.global.v4.u16 [%rd15], {%rs87, %rs88, %rs89, %rs90};
bra.uni $L__BB0_149;
$L__BB0_148:
mov.f32 %f1599, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs98, %f1599;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs97, %f386;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs96, %f385;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs95, %f384;}
// end inline asm
st.global.v4.u16 [%rd15], {%rs95, %rs96, %rs97, %rs98};
$L__BB0_149:
mul.f32 %f387, %f1855, 0f404EBF87;
mul.f32 %f388, %f1856, 0f404EBF87;
mul.f32 %f389, %f1857, 0f404EBF87;
ld.const.u64 %rd64, [params+320];
cvta.to.global.u64 %rd65, %rd64;
ld.const.u32 %r404, [params+312];
mad.lo.s32 %r405, %r404, %r4, %r3;
mul.wide.u32 %rd66, %r405, 8;
add.s64 %rd16, %rd65, %rd66;
@%p167 bra $L__BB0_151;
ld.global.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd16];
// begin inline asm
{ cvt.f32.f16 %f1600, %rs106;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1601, %rs107;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1602, %rs108;}
// end inline asm
add.f32 %f1603, %f387, %f1600;
add.f32 %f1604, %f388, %f1601;
add.f32 %f1605, %f389, %f1602;
mov.f32 %f1606, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs104, %f1605;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs103, %f1604;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs102, %f1603;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs105, %f1606;}
// end inline asm
st.global.v4.u16 [%rd16], {%rs102, %rs103, %rs104, %rs105};
bra.uni $L__BB0_152;
$L__BB0_151:
mov.f32 %f1610, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f1610;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs112, %f389;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs111, %f388;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs110, %f387;}
// end inline asm
st.global.v4.u16 [%rd16], {%rs110, %rs111, %rs112, %rs113};
$L__BB0_152:
mul.f32 %f390, %f1858, 0f404EBF87;
mul.f32 %f391, %f1859, 0f404EBF87;
mul.f32 %f392, %f1860, 0f404EBF87;
ld.const.u64 %rd67, [params+336];
cvta.to.global.u64 %rd68, %rd67;
ld.const.u32 %r406, [params+328];
mad.lo.s32 %r407, %r406, %r4, %r3;
mul.wide.u32 %rd69, %r407, 8;
add.s64 %rd17, %rd68, %rd69;
@%p167 bra $L__BB0_154;
ld.global.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd17];
// begin inline asm
{ cvt.f32.f16 %f1611, %rs121;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1612, %rs122;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1613, %rs123;}
// end inline asm
add.f32 %f1614, %f390, %f1611;
add.f32 %f1615, %f391, %f1612;
add.f32 %f1616, %f392, %f1613;
mov.f32 %f1617, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs119, %f1616;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs118, %f1615;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs117, %f1614;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs120, %f1617;}
// end inline asm
st.global.v4.u16 [%rd17], {%rs117, %rs118, %rs119, %rs120};
bra.uni $L__BB0_155;
$L__BB0_154:
mov.f32 %f1621, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1621;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs127, %f392;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs126, %f391;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs125, %f390;}
// end inline asm
st.global.v4.u16 [%rd17], {%rs125, %rs126, %rs127, %rs128};
$L__BB0_155:
mul.f32 %f393, %f1861, 0f40330C94;
mul.f32 %f394, %f1862, 0f40330C94;
mul.f32 %f395, %f1863, 0f40330C94;
ld.const.u64 %rd70, [params+352];
cvta.to.global.u64 %rd71, %rd70;
ld.const.u32 %r408, [params+344];
mad.lo.s32 %r409, %r408, %r4, %r3;
mul.wide.u32 %rd72, %r409, 8;
add.s64 %rd18, %rd71, %rd72;
@%p167 bra $L__BB0_157;
ld.global.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd18];
// begin inline asm
{ cvt.f32.f16 %f1622, %rs136;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1623, %rs137;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1624, %rs138;}
// end inline asm
add.f32 %f1625, %f393, %f1622;
add.f32 %f1626, %f394, %f1623;
add.f32 %f1627, %f395, %f1624;
mov.f32 %f1628, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs134, %f1627;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs133, %f1626;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs132, %f1625;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs135, %f1628;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs132, %rs133, %rs134, %rs135};
bra.uni $L__BB0_158;
$L__BB0_157:
mov.f32 %f1632, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1632;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs142, %f395;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs141, %f394;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs140, %f393;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs140, %rs141, %rs142, %rs143};
$L__BB0_158:
mul.f32 %f396, %f1864, 0f404EBF87;
mul.f32 %f397, %f1865, 0f404EBF87;
mul.f32 %f398, %f1866, 0f404EBF87;
ld.const.u64 %rd73, [params+368];
cvta.to.global.u64 %rd74, %rd73;
ld.const.u32 %r410, [params+360];
mad.lo.s32 %r411, %r410, %r4, %r3;
mul.wide.u32 %rd75, %r411, 8;
add.s64 %rd19, %rd74, %rd75;
@%p167 bra $L__BB0_160;
ld.global.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f1633, %rs151;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1634, %rs152;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1635, %rs153;}
// end inline asm
add.f32 %f1636, %f396, %f1633;
add.f32 %f1637, %f397, %f1634;
add.f32 %f1638, %f398, %f1635;
mov.f32 %f1639, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs149, %f1638;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs148, %f1637;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs147, %f1636;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs150, %f1639;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs147, %rs148, %rs149, %rs150};
bra.uni $L__BB0_161;
$L__BB0_160:
mov.f32 %f1643, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1643;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs157, %f398;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs156, %f397;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs155, %f396;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs155, %rs156, %rs157, %rs158};
$L__BB0_161:
mul.f32 %f399, %f1867, 0f3FCEBF87;
mul.f32 %f400, %f1868, 0f3FCEBF87;
mul.f32 %f401, %f1869, 0f3FCEBF87;
ld.const.u64 %rd76, [params+384];
cvta.to.global.u64 %rd77, %rd76;
ld.const.u32 %r412, [params+376];
mad.lo.s32 %r413, %r412, %r4, %r3;
mul.wide.u32 %rd78, %r413, 8;
add.s64 %rd20, %rd77, %rd78;
@%p167 bra $L__BB0_163;
ld.global.v4.u16 {%rs166, %rs167, %rs168, %rs169}, [%rd20];
// begin inline asm
{ cvt.f32.f16 %f1644, %rs166;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1645, %rs167;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1646, %rs168;}
// end inline asm
add.f32 %f1647, %f399, %f1644;
add.f32 %f1648, %f400, %f1645;
add.f32 %f1649, %f401, %f1646;
mov.f32 %f1650, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs164, %f1649;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs163, %f1648;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs162, %f1647;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs165, %f1650;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs162, %rs163, %rs164, %rs165};
bra.uni $L__BB0_199;
$L__BB0_163:
mov.f32 %f1654, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs173, %f1654;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs172, %f401;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs171, %f400;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs170, %f399;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs170, %rs171, %rs172, %rs173};
$L__BB0_199:
ret;
}