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

2626 lines
73 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<1632>;
.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 %f1518, 0f00000000;
mov.u16 %rs178, 0;
mov.f32 %f1519, %f1518;
mov.f32 %f1520, %f1518;
@%p11 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f312, %rs9;
div.rn.f32 %f313, %f312, 0f437F0000;
fma.rn.f32 %f314, %f313, 0f40000000, 0fBF800000;
and.b16 %rs15, %rs178, 255;
cvt.rn.f32.u16 %f315, %rs15;
div.rn.f32 %f316, %f315, 0f437F0000;
fma.rn.f32 %f317, %f316, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f318, %rs177;
div.rn.f32 %f319, %f318, 0f437F0000;
fma.rn.f32 %f320, %f319, 0f40000000, 0fBF800000;
mul.f32 %f321, %f317, %f317;
fma.rn.f32 %f322, %f314, %f314, %f321;
fma.rn.f32 %f323, %f320, %f320, %f322;
sqrt.rn.f32 %f324, %f323;
rcp.rn.f32 %f325, %f324;
mul.f32 %f1520, %f325, %f320;
mul.f32 %f1519, %f325, %f317;
mul.f32 %f1518, %f314, %f325;
$L__BB0_4:
ld.const.v2.u32 {%r34, %r35}, [params];
add.s32 %r3, %r34, %r26;
add.s32 %r4, %r35, %r27;
setp.eq.f32 %p12, %f1518, 0f00000000;
setp.eq.f32 %p13, %f1519, 0f00000000;
and.pred %p14, %p12, %p13;
setp.eq.f32 %p15, %f1520, 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 %f1469, 0f00000000;
cvt.rzi.u32.f32 %r410, %f1469;
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 %f1470, %rs108;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1471, %rs109;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1472, %rs110;}
// end inline asm
add.f32 %f1473, %f1470, 0f00000000;
add.f32 %f1474, %f1471, 0f00000000;
add.f32 %f1475, %f1472, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs107, %f1475;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs106, %f1474;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs105, %f1473;}
// 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 %f326, %f10, 0f3456BF95;
ld.global.f32 %f11, [%rd27+4];
mul.f32 %f327, %f11, 0f3456BF95;
ld.global.f32 %f12, [%rd27+8];
mul.f32 %f328, %f12, 0f3456BF95;
abs.f32 %f329, %f1518;
div.rn.f32 %f330, %f326, %f329;
abs.f32 %f331, %f1519;
div.rn.f32 %f332, %f327, %f331;
abs.f32 %f333, %f1520;
div.rn.f32 %f334, %f328, %f333;
abs.f32 %f335, %f330;
abs.f32 %f336, %f332;
abs.f32 %f337, %f334;
mov.f32 %f338, 0f38D1B717;
max.f32 %f339, %f335, %f338;
max.f32 %f340, %f336, %f338;
max.f32 %f341, %f337, %f338;
fma.rn.f32 %f13, %f1518, %f339, %f10;
fma.rn.f32 %f14, %f1519, %f340, %f11;
fma.rn.f32 %f15, %f1520, %f341, %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 %rd28, [params+608];
cvta.to.global.u64 %rd2, %rd28;
mov.f32 %f358, 0f40000000;
cvt.rzi.f32.f32 %f359, %f358;
add.f32 %f360, %f359, %f359;
mov.f32 %f361, 0f40800000;
sub.f32 %f362, %f361, %f360;
abs.f32 %f18, %f362;
ld.const.u32 %r6, [params+544];
ld.const.u32 %r427, [params+104];
mul.f32 %f19, %f13, 0f3456BF95;
mul.f32 %f20, %f14, 0f3456BF95;
mul.f32 %f21, %f15, 0f3456BF95;
mov.f32 %f357, 0f00000000;
mov.u32 %r423, 0;
ld.const.u64 %rd3, [params+96];
abs.f32 %f567, %f19;
abs.f32 %f568, %f20;
max.f32 %f569, %f567, %f568;
abs.f32 %f570, %f21;
max.f32 %f571, %f569, %f570;
mov.f32 %f1594, %f357;
mov.f32 %f1593, %f357;
mov.f32 %f1592, %f357;
mov.f32 %f1591, %f357;
mov.f32 %f1590, %f357;
mov.f32 %f1589, %f357;
mov.f32 %f1588, %f357;
mov.f32 %f1587, %f357;
mov.f32 %f1586, %f357;
mov.f32 %f1585, %f357;
mov.f32 %f1584, %f357;
mov.f32 %f1583, %f357;
mov.f32 %f1582, %f357;
mov.f32 %f1581, %f357;
mov.f32 %f1580, %f357;
mov.f32 %f1579, %f357;
$L__BB0_7:
mul.wide.u32 %rd29, %r423, 112;
add.s64 %rd5, %rd2, %rd29;
ld.global.v4.f32 {%f364, %f365, %f366, %f367}, [%rd5];
ld.global.v4.f32 {%f368, %f369, %f370, %f371}, [%rd5+16];
ld.global.v4.f32 {%f374, %f1545, %f1546, %f377}, [%rd5+32];
ld.global.v4.f32 {%f378, %f379, %f380, %f1543}, [%rd5+48];
ld.global.v4.f32 {%f382, %f383, %f384, %f385}, [%rd5+64];
ld.global.v4.u32 {%r41, %r42, %r43, %r44}, [%rd5+80];
mov.b32 %f57, %r43;
mov.b32 %f56, %r42;
mov.b32 %f55, %r41;
ld.global.u64 %rd6, [%rd5+96];
sub.f32 %f386, %f365, %f10;
sub.f32 %f387, %f366, %f11;
sub.f32 %f388, %f367, %f12;
mul.f32 %f389, %f387, %f387;
fma.rn.f32 %f390, %f386, %f386, %f389;
fma.rn.f32 %f391, %f388, %f388, %f390;
sqrt.rn.f32 %f392, %f391;
rcp.rn.f32 %f393, %f392;
mul.f32 %f59, %f386, %f393;
mul.f32 %f60, %f387, %f393;
mul.f32 %f61, %f388, %f393;
mul.f32 %f62, %f392, %f369;
mul.f32 %f63, %f392, %f371;
abs.f32 %f64, %f63;
setp.lt.f32 %p18, %f64, 0f00800000;
mul.f32 %f394, %f64, 0f4B800000;
selp.f32 %f395, %f394, %f64, %p18;
selp.f32 %f396, 0fC3170000, 0fC2FE0000, %p18;
mov.b32 %r48, %f395;
and.b32 %r49, %r48, 8388607;
or.b32 %r50, %r49, 1065353216;
mov.b32 %f397, %r50;
shr.u32 %r51, %r48, 23;
cvt.rn.f32.u32 %f398, %r51;
add.f32 %f399, %f396, %f398;
setp.gt.f32 %p19, %f397, 0f3FB504F3;
mul.f32 %f400, %f397, 0f3F000000;
add.f32 %f401, %f399, 0f3F800000;
selp.f32 %f402, %f401, %f399, %p19;
selp.f32 %f403, %f400, %f397, %p19;
add.f32 %f404, %f403, 0fBF800000;
add.f32 %f405, %f403, 0f3F800000;
rcp.approx.ftz.f32 %f406, %f405;
add.f32 %f407, %f404, %f404;
mul.f32 %f408, %f407, %f406;
mul.f32 %f409, %f408, %f408;
mov.f32 %f410, 0f3C4CAF63;
mov.f32 %f411, 0f3B18F0FE;
fma.rn.f32 %f412, %f411, %f409, %f410;
mov.f32 %f413, 0f3DAAAABD;
fma.rn.f32 %f414, %f412, %f409, %f413;
mul.rn.f32 %f415, %f414, %f409;
mul.rn.f32 %f416, %f415, %f408;
sub.f32 %f417, %f404, %f408;
add.f32 %f418, %f417, %f417;
neg.f32 %f419, %f408;
fma.rn.f32 %f420, %f419, %f404, %f418;
mul.rn.f32 %f421, %f406, %f420;
add.f32 %f422, %f416, %f408;
sub.f32 %f423, %f408, %f422;
add.f32 %f424, %f416, %f423;
add.f32 %f425, %f421, %f424;
add.f32 %f426, %f422, %f425;
sub.f32 %f427, %f422, %f426;
add.f32 %f428, %f425, %f427;
mov.f32 %f429, 0f3F317200;
mul.rn.f32 %f430, %f402, %f429;
mov.f32 %f431, 0f35BFBE8E;
mul.rn.f32 %f432, %f402, %f431;
add.f32 %f433, %f430, %f426;
sub.f32 %f434, %f430, %f433;
add.f32 %f435, %f426, %f434;
add.f32 %f436, %f428, %f435;
add.f32 %f437, %f432, %f436;
add.f32 %f438, %f433, %f437;
sub.f32 %f439, %f433, %f438;
add.f32 %f440, %f437, %f439;
mul.rn.f32 %f442, %f361, %f438;
neg.f32 %f443, %f442;
fma.rn.f32 %f444, %f361, %f438, %f443;
fma.rn.f32 %f445, %f361, %f440, %f444;
fma.rn.f32 %f447, %f357, %f438, %f445;
add.rn.f32 %f448, %f442, %f447;
neg.f32 %f449, %f448;
add.rn.f32 %f450, %f442, %f449;
add.rn.f32 %f451, %f450, %f447;
mov.b32 %r52, %f448;
setp.eq.s32 %p20, %r52, 1118925336;
add.s32 %r53, %r52, -1;
mov.b32 %f452, %r53;
add.f32 %f453, %f451, 0f37000000;
selp.f32 %f65, %f453, %f451, %p20;
selp.f32 %f454, %f452, %f448, %p20;
mov.f32 %f455, 0f3FB8AA3B;
mul.rn.f32 %f456, %f454, %f455;
cvt.rzi.f32.f32 %f457, %f456;
abs.f32 %f458, %f457;
setp.gt.f32 %p21, %f458, 0f42FC0000;
mov.b32 %r54, %f457;
and.b32 %r55, %r54, -2147483648;
or.b32 %r56, %r55, 1123811328;
mov.b32 %f459, %r56;
selp.f32 %f460, %f459, %f457, %p21;
mov.f32 %f461, 0fBF317218;
fma.rn.f32 %f462, %f460, %f461, %f454;
mov.f32 %f463, 0f3102E308;
fma.rn.f32 %f464, %f460, %f463, %f462;
mul.f32 %f465, %f464, 0f3FB8AA3B;
add.f32 %f466, %f460, 0f4B40007F;
mov.b32 %r57, %f466;
shl.b32 %r58, %r57, 23;
mov.b32 %f467, %r58;
ex2.approx.ftz.f32 %f468, %f465;
mul.f32 %f66, %f468, %f467;
setp.eq.f32 %p22, %f66, 0f7F800000;
mov.f32 %f1537, 0f7F800000;
@%p22 bra $L__BB0_9;
fma.rn.f32 %f1537, %f66, %f65, %f66;
$L__BB0_9:
setp.lt.f32 %p23, %f63, 0f00000000;
setp.eq.f32 %p24, %f18, 0f3F800000;
and.pred %p1, %p23, %p24;
setp.eq.f32 %p25, %f63, 0f00000000;
@%p25 bra $L__BB0_13;
bra.uni $L__BB0_10;
$L__BB0_13:
add.f32 %f473, %f63, %f63;
selp.f32 %f1539, %f473, 0f00000000, %p24;
bra.uni $L__BB0_14;
$L__BB0_10:
mov.b32 %r59, %f1537;
xor.b32 %r60, %r59, -2147483648;
mov.b32 %f469, %r60;
selp.f32 %f1539, %f469, %f1537, %p1;
setp.geu.f32 %p26, %f63, 0f00000000;
@%p26 bra $L__BB0_14;
mov.f32 %f470, 0f40800000;
cvt.rzi.f32.f32 %f471, %f470;
setp.eq.f32 %p27, %f471, 0f40800000;
@%p27 bra $L__BB0_14;
mov.f32 %f1539, 0f7FFFFFFF;
$L__BB0_14:
abs.f32 %f1515, %f63;
add.f32 %f474, %f1515, 0f40800000;
mov.b32 %r61, %f474;
setp.lt.s32 %p29, %r61, 2139095040;
@%p29 bra $L__BB0_19;
abs.f32 %f1516, %f63;
setp.gtu.f32 %p30, %f1516, 0f7F800000;
@%p30 bra $L__BB0_18;
bra.uni $L__BB0_16;
$L__BB0_18:
add.f32 %f1539, %f63, 0f40800000;
bra.uni $L__BB0_19;
$L__BB0_16:
abs.f32 %f1517, %f63;
setp.neu.f32 %p31, %f1517, 0f7F800000;
@%p31 bra $L__BB0_19;
selp.f32 %f1539, 0fFF800000, 0f7F800000, %p1;
$L__BB0_19:
mov.f32 %f475, 0f3F800000;
sub.f32 %f476, %f475, %f1539;
setp.eq.f32 %p32, %f63, 0f3F800000;
selp.f32 %f477, 0f00000000, %f476, %p32;
cvt.sat.f32.f32 %f478, %f477;
fma.rn.f32 %f479, %f62, %f62, %f370;
div.rn.f32 %f75, %f478, %f479;
mul.f32 %f480, %f1519, %f60;
fma.rn.f32 %f481, %f1518, %f59, %f480;
fma.rn.f32 %f76, %f1520, %f61, %f481;
setp.eq.f32 %p33, %f364, 0f3F800000;
@%p33 bra $L__BB0_25;
bra.uni $L__BB0_20;
$L__BB0_25:
setp.eq.s64 %p37, %rd6, 0;
mov.f32 %f1544, %f374;
mov.f32 %f1547, %f75;
@%p37 bra $L__BB0_29;
setp.geu.f32 %p38, %f1543, 0f00000000;
mov.f32 %f1540, %f59;
mov.f32 %f1541, %f60;
mov.f32 %f1542, %f61;
@%p38 bra $L__BB0_28;
mul.f32 %f506, %f56, %f387;
fma.rn.f32 %f507, %f55, %f386, %f506;
fma.rn.f32 %f509, %f57, %f388, %f507;
rcp.rn.f32 %f510, %f509;
mul.f32 %f1540, %f386, %f510;
mul.f32 %f1541, %f387, %f510;
mul.f32 %f1542, %f388, %f510;
neg.f32 %f1543, %f1543;
$L__BB0_28:
mul.f32 %f511, %f57, %f61;
mul.f32 %f512, %f56, %f60;
neg.f32 %f513, %f512;
mul.f32 %f514, %f55, %f59;
sub.f32 %f515, %f513, %f514;
sub.f32 %f516, %f515, %f511;
setp.gt.f32 %p39, %f516, 0f00000000;
selp.f32 %f517, 0f3F800000, 0f00000000, %p39;
mov.f32 %f518, 0f3F800000;
mul.f32 %f519, %f379, %f1541;
mul.f32 %f520, %f383, %f1541;
fma.rn.f32 %f521, %f378, %f1540, %f519;
fma.rn.f32 %f522, %f382, %f1540, %f520;
fma.rn.f32 %f523, %f380, %f1542, %f521;
fma.rn.f32 %f524, %f384, %f1542, %f522;
fma.rn.f32 %f525, %f1543, %f523, 0f3F000000;
sub.f32 %f526, %f518, %f525;
fma.rn.f32 %f527, %f1543, %f524, 0f3F000000;
tex.2d.v4.f32.f32 {%f528, %f529, %f530, %f531}, [%rd6, {%f526, %f527}];
mul.f32 %f532, %f517, %f528;
mul.f32 %f533, %f517, %f529;
mul.f32 %f534, %f517, %f530;
mul.f32 %f1544, %f374, %f532;
mul.f32 %f1545, %f1545, %f533;
mul.f32 %f1546, %f1546, %f534;
mov.f32 %f1547, %f75;
bra.uni $L__BB0_29;
$L__BB0_20:
setp.eq.f32 %p34, %f364, 0f40000000;
@%p34 bra $L__BB0_23;
bra.uni $L__BB0_21;
$L__BB0_23:
setp.eq.s64 %p36, %rd6, 0;
mov.f32 %f1544, %f374;
mov.f32 %f1547, %f75;
@%p36 bra $L__BB0_29;
mul.f32 %f491, %f379, %f60;
mul.f32 %f492, %f383, %f60;
mul.f32 %f493, %f56, %f60;
fma.rn.f32 %f494, %f378, %f59, %f491;
fma.rn.f32 %f495, %f382, %f59, %f492;
fma.rn.f32 %f496, %f55, %f59, %f493;
fma.rn.f32 %f497, %f380, %f61, %f494;
fma.rn.f32 %f498, %f384, %f61, %f495;
fma.rn.f32 %f499, %f57, %f61, %f496;
tex.cube.v4.f32.f32 {%f500, %f501, %f502, %f503}, [%rd6, {%f497, %f498, %f499, %f499}];
mul.f32 %f1544, %f374, %f500;
mul.f32 %f1545, %f1545, %f501;
mul.f32 %f1546, %f1546, %f502;
mov.f32 %f1547, %f75;
bra.uni $L__BB0_29;
$L__BB0_21:
setp.neu.f32 %p35, %f364, 0f40800000;
mov.f32 %f1544, %f374;
mov.f32 %f1547, %f75;
@%p35 bra $L__BB0_29;
mul.f32 %f482, %f55, %f59;
mul.f32 %f483, %f56, %f60;
neg.f32 %f484, %f483;
sub.f32 %f485, %f484, %f482;
mul.f32 %f486, %f57, %f61;
sub.f32 %f487, %f485, %f486;
fma.rn.f32 %f488, %f377, %f487, %f1543;
cvt.sat.f32.f32 %f489, %f488;
mul.f32 %f490, %f489, %f489;
mul.f32 %f1547, %f75, %f490;
mov.f32 %f1544, %f374;
$L__BB0_29:
max.f32 %f550, %f1544, %f1545;
max.f32 %f551, %f550, %f1546;
mul.f32 %f552, %f1547, %f551;
setp.lt.f32 %p41, %f552, 0f3727C5AC;
mov.pred %p182, -1;
mov.f32 %f1548, 0f00000000;
mov.f32 %f1549, %f1548;
mov.f32 %f1550, %f1548;
mov.f32 %f1551, %f1548;
mov.f32 %f1552, %f1548;
mov.f32 %f1553, %f1548;
mov.f32 %f1554, %f1548;
mov.f32 %f1555, %f1548;
mov.f32 %f1556, %f1548;
mov.f32 %f1557, %f1548;
mov.f32 %f1558, %f1548;
mov.f32 %f1559, %f1548;
mov.f32 %f1560, %f1548;
mov.f32 %f1561, %f1548;
mov.f32 %f1562, %f1548;
@%p41 bra $L__BB0_31;
setp.eq.s32 %p43, %r6, 0;
selp.f32 %f553, %f76, 0f3F800000, %p43;
cvt.sat.f32.f32 %f554, %f553;
mul.f32 %f555, %f1547, %f554;
mul.f32 %f1548, %f1544, %f555;
mul.f32 %f1549, %f1545, %f555;
mul.f32 %f1550, %f1546, %f555;
mul.f32 %f556, %f1547, 0f3E800000;
mul.f32 %f1551, %f1544, %f556;
mul.f32 %f1552, %f1545, %f556;
mul.f32 %f1553, %f1546, %f556;
mul.f32 %f1554, %f59, %f1551;
mul.f32 %f1555, %f59, %f1552;
mul.f32 %f1556, %f59, %f1553;
mul.f32 %f1557, %f60, %f1551;
mul.f32 %f1558, %f60, %f1552;
mul.f32 %f1559, %f60, %f1553;
mul.f32 %f1560, %f61, %f1551;
mul.f32 %f1561, %f61, %f1552;
mul.f32 %f1562, %f61, %f1553;
mov.pred %p182, 0;
$L__BB0_31:
@%p182 bra $L__BB0_43;
setp.eq.s32 %p44, %r44, 0;
mov.u16 %rs179, 0;
mov.f32 %f1578, 0f3F800000;
mov.f32 %f1576, %f1578;
mov.f32 %f1577, %f1578;
@%p44 bra $L__BB0_40;
abs.s32 %r10, %r44;
setp.lt.s32 %p45, %r10, 1;
mov.f32 %f1566, 0f00000000;
mov.f32 %f1567, %f1566;
mov.f32 %f1568, %f1566;
@%p45 bra $L__BB0_39;
mov.f32 %f572, 0f38D1B717;
max.f32 %f573, %f571, %f572;
and.b32 %r63, %r427, 32;
setp.eq.s32 %p46, %r63, 0;
mov.u32 %r426, 0;
selp.f32 %f574, 0f3F800000, 0f41200000, %p46;
mul.f32 %f139, %f574, %f573;
and.b32 %r11, %r10, 1;
setp.eq.s32 %p47, %r10, 1;
mov.f32 %f1568, 0f00000000;
mov.f32 %f1567, %f1568;
mov.f32 %f1566, %f1568;
@%p47 bra $L__BB0_37;
sub.s32 %r425, %r10, %r11;
$L__BB0_36:
cvt.rn.f32.s32 %f596, %r426;
add.f32 %f597, %f16, %f596;
sub.f32 %f598, %f17, %f596;
mul.f32 %f599, %f597, 0f3DD32618;
cvt.rmi.f32.f32 %f600, %f599;
sub.f32 %f601, %f599, %f600;
mul.f32 %f602, %f598, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f603, %f602;
sub.f32 %f604, %f602, %f603;
mul.f32 %f605, %f596, 0f3DC74539;
cvt.rmi.f32.f32 %f606, %f605;
sub.f32 %f607, %f605, %f606;
add.f32 %f608, %f604, 0f420551EC;
add.f32 %f609, %f601, 0f420551EC;
add.f32 %f610, %f607, 0f420551EC;
mul.f32 %f611, %f601, %f608;
fma.rn.f32 %f612, %f604, %f609, %f611;
fma.rn.f32 %f613, %f607, %f610, %f612;
add.f32 %f614, %f601, %f613;
add.f32 %f615, %f604, %f613;
add.f32 %f616, %f607, %f613;
add.f32 %f617, %f614, %f615;
mul.f32 %f618, %f616, %f617;
cvt.rmi.f32.f32 %f619, %f618;
sub.f32 %f620, %f618, %f619;
add.f32 %f621, %f614, %f614;
mul.f32 %f622, %f615, %f621;
cvt.rmi.f32.f32 %f623, %f622;
sub.f32 %f624, %f622, %f623;
mul.f32 %f625, %f614, %f617;
cvt.rmi.f32.f32 %f626, %f625;
sub.f32 %f627, %f625, %f626;
fma.rn.f32 %f628, %f620, 0f40000000, 0fBF800000;
fma.rn.f32 %f629, %f624, 0f40000000, 0fBF800000;
fma.rn.f32 %f630, %f627, 0f40000000, 0fBF800000;
ld.global.f32 %f631, [%rd5+16];
ld.global.f32 %f632, [%rd5+4];
fma.rn.f32 %f633, %f631, %f628, %f632;
ld.global.f32 %f634, [%rd5+8];
fma.rn.f32 %f635, %f631, %f629, %f634;
ld.global.f32 %f636, [%rd5+12];
fma.rn.f32 %f637, %f631, %f630, %f636;
sub.f32 %f638, %f633, %f10;
sub.f32 %f639, %f635, %f11;
sub.f32 %f640, %f637, %f12;
mul.f32 %f641, %f639, %f639;
fma.rn.f32 %f642, %f638, %f638, %f641;
fma.rn.f32 %f643, %f640, %f640, %f642;
sqrt.rn.f32 %f585, %f643;
rcp.rn.f32 %f644, %f585;
mul.f32 %f581, %f644, %f638;
mul.f32 %f582, %f644, %f639;
mul.f32 %f583, %f644, %f640;
mov.f32 %f595, 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,%f581,%f582,%f583,%f139,%f585,%f595,%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 %f645, %r65;
mov.b32 %f646, %r66;
mov.b32 %f647, %r67;
add.f32 %f648, %f1566, %f645;
add.f32 %f649, %f1567, %f646;
add.f32 %f650, %f1568, %f647;
add.s32 %r207, %r426, 1;
cvt.rn.f32.s32 %f651, %r207;
add.f32 %f652, %f16, %f651;
sub.f32 %f653, %f17, %f651;
mul.f32 %f654, %f652, 0f3DD32618;
cvt.rmi.f32.f32 %f655, %f654;
sub.f32 %f656, %f654, %f655;
mul.f32 %f657, %f653, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f658, %f657;
sub.f32 %f659, %f657, %f658;
mul.f32 %f660, %f651, 0f3DC74539;
cvt.rmi.f32.f32 %f661, %f660;
sub.f32 %f662, %f660, %f661;
add.f32 %f663, %f659, 0f420551EC;
add.f32 %f664, %f656, 0f420551EC;
add.f32 %f665, %f662, 0f420551EC;
mul.f32 %f666, %f656, %f663;
fma.rn.f32 %f667, %f659, %f664, %f666;
fma.rn.f32 %f668, %f662, %f665, %f667;
add.f32 %f669, %f656, %f668;
add.f32 %f670, %f659, %f668;
add.f32 %f671, %f662, %f668;
add.f32 %f672, %f669, %f670;
mul.f32 %f673, %f671, %f672;
cvt.rmi.f32.f32 %f674, %f673;
sub.f32 %f675, %f673, %f674;
add.f32 %f676, %f669, %f669;
mul.f32 %f677, %f670, %f676;
cvt.rmi.f32.f32 %f678, %f677;
sub.f32 %f679, %f677, %f678;
mul.f32 %f680, %f669, %f672;
cvt.rmi.f32.f32 %f681, %f680;
sub.f32 %f682, %f680, %f681;
fma.rn.f32 %f683, %f675, 0f40000000, 0fBF800000;
fma.rn.f32 %f684, %f679, 0f40000000, 0fBF800000;
fma.rn.f32 %f685, %f682, 0f40000000, 0fBF800000;
ld.global.f32 %f686, [%rd5+16];
ld.global.f32 %f687, [%rd5+4];
fma.rn.f32 %f688, %f686, %f683, %f687;
ld.global.f32 %f689, [%rd5+8];
fma.rn.f32 %f690, %f686, %f684, %f689;
ld.global.f32 %f691, [%rd5+12];
fma.rn.f32 %f692, %f686, %f685, %f691;
sub.f32 %f693, %f688, %f10;
sub.f32 %f694, %f690, %f11;
sub.f32 %f695, %f692, %f12;
mul.f32 %f696, %f694, %f694;
fma.rn.f32 %f697, %f693, %f693, %f696;
fma.rn.f32 %f698, %f695, %f695, %f697;
sqrt.rn.f32 %f594, %f698;
rcp.rn.f32 %f699, %f594;
mul.f32 %f590, %f699, %f693;
mul.f32 %f591, %f699, %f694;
mul.f32 %f592, %f699, %f695;
// 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,%f590,%f591,%f592,%f139,%f594,%f595,%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 %f700, %r136;
mov.b32 %f701, %r137;
mov.b32 %f702, %r138;
add.f32 %f1566, %f648, %f700;
add.f32 %f1567, %f649, %f701;
add.f32 %f1568, %f650, %f702;
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 %f712, %r426;
add.f32 %f713, %f16, %f712;
sub.f32 %f714, %f17, %f712;
mul.f32 %f715, %f713, 0f3DD32618;
cvt.rmi.f32.f32 %f716, %f715;
sub.f32 %f717, %f715, %f716;
mul.f32 %f718, %f714, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f719, %f718;
sub.f32 %f720, %f718, %f719;
mul.f32 %f721, %f712, 0f3DC74539;
cvt.rmi.f32.f32 %f722, %f721;
sub.f32 %f723, %f721, %f722;
add.f32 %f724, %f720, 0f420551EC;
add.f32 %f725, %f717, 0f420551EC;
add.f32 %f726, %f723, 0f420551EC;
mul.f32 %f727, %f717, %f724;
fma.rn.f32 %f728, %f720, %f725, %f727;
fma.rn.f32 %f729, %f723, %f726, %f728;
add.f32 %f730, %f717, %f729;
add.f32 %f731, %f720, %f729;
add.f32 %f732, %f723, %f729;
add.f32 %f733, %f730, %f731;
mul.f32 %f734, %f732, %f733;
cvt.rmi.f32.f32 %f735, %f734;
sub.f32 %f736, %f734, %f735;
add.f32 %f737, %f730, %f730;
mul.f32 %f738, %f731, %f737;
cvt.rmi.f32.f32 %f739, %f738;
sub.f32 %f740, %f738, %f739;
mul.f32 %f741, %f730, %f733;
cvt.rmi.f32.f32 %f742, %f741;
sub.f32 %f743, %f741, %f742;
fma.rn.f32 %f744, %f736, 0f40000000, 0fBF800000;
fma.rn.f32 %f745, %f740, 0f40000000, 0fBF800000;
fma.rn.f32 %f746, %f743, 0f40000000, 0fBF800000;
ld.global.f32 %f747, [%rd5+16];
ld.global.f32 %f748, [%rd5+4];
fma.rn.f32 %f749, %f747, %f744, %f748;
ld.global.f32 %f750, [%rd5+8];
fma.rn.f32 %f751, %f747, %f745, %f750;
ld.global.f32 %f752, [%rd5+12];
fma.rn.f32 %f753, %f747, %f746, %f752;
sub.f32 %f754, %f749, %f10;
sub.f32 %f755, %f751, %f11;
sub.f32 %f756, %f753, %f12;
mul.f32 %f757, %f755, %f755;
fma.rn.f32 %f758, %f754, %f754, %f757;
fma.rn.f32 %f759, %f756, %f756, %f758;
sqrt.rn.f32 %f710, %f759;
rcp.rn.f32 %f760, %f710;
mul.f32 %f706, %f760, %f754;
mul.f32 %f707, %f760, %f755;
mul.f32 %f708, %f760, %f756;
mov.f32 %f711, 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,%f706,%f707,%f708,%f139,%f710,%f711,%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 %f761, %r208;
mov.b32 %f762, %r209;
mov.b32 %f763, %r210;
add.f32 %f1566, %f1566, %f761;
add.f32 %f1567, %f1567, %f762;
add.f32 %f1568, %f1568, %f763;
$L__BB0_39:
cvt.rn.f32.s32 %f764, %r10;
rcp.rn.f32 %f765, %f764;
mul.f32 %f1578, %f765, %f1566;
mul.f32 %f1576, %f765, %f1567;
mul.f32 %f1577, %f765, %f1568;
shr.u32 %r279, %r44, 31;
cvt.u16.u32 %rs179, %r279;
$L__BB0_40:
fma.rn.f32 %f1579, %f1548, %f1578, %f1579;
fma.rn.f32 %f1580, %f1549, %f1576, %f1580;
fma.rn.f32 %f1581, %f1550, %f1577, %f1581;
fma.rn.f32 %f1582, %f1551, %f1578, %f1582;
fma.rn.f32 %f1583, %f1552, %f1576, %f1583;
fma.rn.f32 %f1584, %f1553, %f1577, %f1584;
fma.rn.f32 %f1585, %f1554, %f1578, %f1585;
fma.rn.f32 %f1586, %f1555, %f1576, %f1586;
fma.rn.f32 %f1587, %f1556, %f1577, %f1587;
fma.rn.f32 %f1588, %f1557, %f1578, %f1588;
fma.rn.f32 %f1589, %f1558, %f1576, %f1589;
fma.rn.f32 %f1590, %f1559, %f1577, %f1590;
fma.rn.f32 %f1591, %f1560, %f1578, %f1591;
fma.rn.f32 %f1592, %f1561, %f1576, %f1592;
fma.rn.f32 %f1593, %f1562, %f1577, %f1593;
setp.eq.s16 %p50, %rs179, 0;
@%p50 bra $L__BB0_42;
div.rn.f32 %f766, %f1548, %f374;
div.rn.f32 %f767, %f766, %f75;
cvt.sat.f32.f32 %f768, %f767;
mul.f32 %f1578, %f1578, %f768;
$L__BB0_42:
add.f32 %f1594, %f1594, %f1578;
$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 %f1579, 0f00000000;
ld.const.u32 %r427, [params+104];
mov.f32 %f1580, %f1579;
mov.f32 %f1581, %f1579;
mov.f32 %f1582, %f1579;
mov.f32 %f1583, %f1579;
mov.f32 %f1584, %f1579;
mov.f32 %f1585, %f1579;
mov.f32 %f1586, %f1579;
mov.f32 %f1587, %f1579;
mov.f32 %f1588, %f1579;
mov.f32 %f1589, %f1579;
mov.f32 %f1590, %f1579;
mov.f32 %f1591, %f1579;
mov.f32 %f1592, %f1579;
mov.f32 %f1593, %f1579;
mov.f32 %f1594, %f1579;
$L__BB0_45:
and.b32 %r281, %r427, 8;
setp.eq.s32 %p52, %r281, 0;
@%p52 bra $L__BB0_59;
cvt.sat.f32.f32 %f214, %f1594;
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 %f786, 0f3E68BA2E;
cvt.rzi.f32.f32 %f787, %f786;
add.f32 %f788, %f787, %f787;
mov.f32 %f789, 0f3EE8BA2E;
sub.f32 %f790, %f789, %f788;
abs.f32 %f215, %f790;
abs.f32 %f216, %f214;
setp.lt.f32 %p53, %f216, 0f00800000;
mul.f32 %f791, %f216, 0f4B800000;
selp.f32 %f792, %f791, %f216, %p53;
selp.f32 %f793, 0fC3170000, 0fC2FE0000, %p53;
mov.b32 %r284, %f792;
and.b32 %r285, %r284, 8388607;
or.b32 %r286, %r285, 1065353216;
mov.b32 %f794, %r286;
shr.u32 %r287, %r284, 23;
cvt.rn.f32.u32 %f795, %r287;
add.f32 %f796, %f793, %f795;
setp.gt.f32 %p54, %f794, 0f3FB504F3;
mul.f32 %f797, %f794, 0f3F000000;
add.f32 %f798, %f796, 0f3F800000;
selp.f32 %f799, %f798, %f796, %p54;
selp.f32 %f800, %f797, %f794, %p54;
add.f32 %f801, %f800, 0fBF800000;
add.f32 %f802, %f800, 0f3F800000;
rcp.approx.ftz.f32 %f803, %f802;
add.f32 %f804, %f801, %f801;
mul.f32 %f805, %f804, %f803;
mul.f32 %f806, %f805, %f805;
mov.f32 %f807, 0f3C4CAF63;
mov.f32 %f808, 0f3B18F0FE;
fma.rn.f32 %f809, %f808, %f806, %f807;
mov.f32 %f810, 0f3DAAAABD;
fma.rn.f32 %f811, %f809, %f806, %f810;
mul.rn.f32 %f812, %f811, %f806;
mul.rn.f32 %f813, %f812, %f805;
sub.f32 %f814, %f801, %f805;
add.f32 %f815, %f814, %f814;
neg.f32 %f816, %f805;
fma.rn.f32 %f817, %f816, %f801, %f815;
mul.rn.f32 %f818, %f803, %f817;
add.f32 %f819, %f813, %f805;
sub.f32 %f820, %f805, %f819;
add.f32 %f821, %f813, %f820;
add.f32 %f822, %f818, %f821;
add.f32 %f823, %f819, %f822;
sub.f32 %f824, %f819, %f823;
add.f32 %f825, %f822, %f824;
mov.f32 %f826, 0f3F317200;
mul.rn.f32 %f827, %f799, %f826;
mov.f32 %f828, 0f35BFBE8E;
mul.rn.f32 %f829, %f799, %f828;
add.f32 %f830, %f827, %f823;
sub.f32 %f831, %f827, %f830;
add.f32 %f832, %f823, %f831;
add.f32 %f833, %f825, %f832;
add.f32 %f834, %f829, %f833;
add.f32 %f835, %f830, %f834;
sub.f32 %f836, %f830, %f835;
add.f32 %f837, %f834, %f836;
mul.rn.f32 %f838, %f789, %f835;
neg.f32 %f839, %f838;
fma.rn.f32 %f840, %f789, %f835, %f839;
fma.rn.f32 %f841, %f789, %f837, %f840;
mov.f32 %f842, 0f00000000;
fma.rn.f32 %f843, %f842, %f835, %f841;
add.rn.f32 %f844, %f838, %f843;
neg.f32 %f845, %f844;
add.rn.f32 %f846, %f838, %f845;
add.rn.f32 %f847, %f846, %f843;
mov.b32 %r288, %f844;
setp.eq.s32 %p55, %r288, 1118925336;
add.s32 %r289, %r288, -1;
mov.b32 %f848, %r289;
add.f32 %f849, %f847, 0f37000000;
selp.f32 %f217, %f849, %f847, %p55;
selp.f32 %f850, %f848, %f844, %p55;
mov.f32 %f851, 0f3FB8AA3B;
mul.rn.f32 %f852, %f850, %f851;
cvt.rzi.f32.f32 %f853, %f852;
abs.f32 %f854, %f853;
setp.gt.f32 %p56, %f854, 0f42FC0000;
mov.b32 %r290, %f853;
and.b32 %r291, %r290, -2147483648;
or.b32 %r292, %r291, 1123811328;
mov.b32 %f855, %r292;
selp.f32 %f856, %f855, %f853, %p56;
mov.f32 %f857, 0fBF317218;
fma.rn.f32 %f858, %f856, %f857, %f850;
mov.f32 %f859, 0f3102E308;
fma.rn.f32 %f860, %f856, %f859, %f858;
mul.f32 %f861, %f860, 0f3FB8AA3B;
add.f32 %f862, %f856, 0f4B40007F;
mov.b32 %r293, %f862;
shl.b32 %r294, %r293, 23;
mov.b32 %f863, %r294;
ex2.approx.ftz.f32 %f864, %f861;
mul.f32 %f218, %f864, %f863;
setp.eq.f32 %p57, %f218, 0f7F800000;
mov.f32 %f1611, 0f7F800000;
@%p57 bra $L__BB0_48;
fma.rn.f32 %f1611, %f218, %f217, %f218;
$L__BB0_48:
setp.lt.f32 %p58, %f214, 0f00000000;
setp.eq.f32 %p59, %f215, 0f3F800000;
and.pred %p3, %p58, %p59;
setp.eq.f32 %p60, %f214, 0f00000000;
@%p60 bra $L__BB0_52;
bra.uni $L__BB0_49;
$L__BB0_52:
add.f32 %f869, %f214, %f214;
selp.f32 %f1613, %f869, 0f00000000, %p59;
bra.uni $L__BB0_53;
$L__BB0_156:
mov.f32 %f1478, 0f00000000;
mov.u32 %r429, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs115, %f1478;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs114, %f1478;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f1478;}
// 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 %f1479, %rs123;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1480, %rs124;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1481, %rs125;}
// end inline asm
add.f32 %f1482, %f1479, 0f00000000;
add.f32 %f1483, %f1480, 0f00000000;
add.f32 %f1484, %f1481, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs122, %f1484;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs121, %f1483;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs120, %f1482;}
// 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 %f1487, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs130, %f1487;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs129, %f1487;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1487;}
// 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 %f1488, %rs138;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1489, %rs139;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1490, %rs140;}
// end inline asm
add.f32 %f1491, %f1488, 0f00000000;
add.f32 %f1492, %f1489, 0f00000000;
add.f32 %f1493, %f1490, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs137, %f1493;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs136, %f1492;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs135, %f1491;}
// 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 %f1496, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs145, %f1496;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs144, %f1496;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1496;}
// 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 %f1497, %rs153;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1498, %rs154;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1499, %rs155;}
// end inline asm
add.f32 %f1500, %f1497, 0f00000000;
add.f32 %f1501, %f1498, 0f00000000;
add.f32 %f1502, %f1499, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs152, %f1502;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs151, %f1501;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs150, %f1500;}
// 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 %f1505, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs160, %f1505;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs159, %f1505;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1505;}
// 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 %f1506, %rs168;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1507, %rs169;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1508, %rs170;}
// end inline asm
add.f32 %f1509, %f1506, 0f00000000;
add.f32 %f1510, %f1507, 0f00000000;
add.f32 %f1511, %f1508, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs167, %f1511;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs166, %f1510;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs165, %f1509;}
// 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 %f1514, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs175, %f1514;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs174, %f1514;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs173, %f1514;}
// 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, %f1611;
xor.b32 %r296, %r295, -2147483648;
mov.b32 %f865, %r296;
selp.f32 %f1613, %f865, %f1611, %p3;
setp.geu.f32 %p61, %f214, 0f00000000;
@%p61 bra $L__BB0_53;
mov.f32 %f866, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f867, %f866;
setp.eq.f32 %p62, %f867, 0f3EE8BA2E;
@%p62 bra $L__BB0_53;
mov.f32 %f1613, 0f7FFFFFFF;
$L__BB0_53:
add.f32 %f870, %f216, 0f3EE8BA2E;
mov.b32 %r297, %f870;
setp.lt.s32 %p64, %r297, 2139095040;
@%p64 bra $L__BB0_58;
setp.gtu.f32 %p65, %f216, 0f7F800000;
@%p65 bra $L__BB0_57;
bra.uni $L__BB0_55;
$L__BB0_57:
add.f32 %f1613, %f214, 0f3EE8BA2E;
bra.uni $L__BB0_58;
$L__BB0_55:
setp.neu.f32 %p66, %f216, 0f7F800000;
@%p66 bra $L__BB0_58;
selp.f32 %f1613, 0fFF800000, 0f7F800000, %p3;
$L__BB0_58:
mul.f32 %f871, %f1613, 0f437F0000;
setp.eq.f32 %p67, %f214, 0f3F800000;
selp.f32 %f872, 0f437F0000, %f871, %p67;
cvt.rzi.u32.f32 %r298, %f872;
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 %f874, 0f3E666666;
cvt.rzi.f32.f32 %f875, %f874;
add.f32 %f876, %f875, %f875;
mov.f32 %f877, 0f3EE66666;
sub.f32 %f878, %f877, %f876;
abs.f32 %f227, %f878;
abs.f32 %f228, %f1579;
setp.lt.f32 %p72, %f228, 0f00800000;
mul.f32 %f879, %f228, 0f4B800000;
selp.f32 %f880, %f879, %f228, %p72;
selp.f32 %f881, 0fC3170000, 0fC2FE0000, %p72;
mov.b32 %r300, %f880;
and.b32 %r301, %r300, 8388607;
or.b32 %r302, %r301, 1065353216;
mov.b32 %f882, %r302;
shr.u32 %r303, %r300, 23;
cvt.rn.f32.u32 %f883, %r303;
add.f32 %f884, %f881, %f883;
setp.gt.f32 %p73, %f882, 0f3FB504F3;
mul.f32 %f885, %f882, 0f3F000000;
add.f32 %f886, %f884, 0f3F800000;
selp.f32 %f887, %f886, %f884, %p73;
selp.f32 %f888, %f885, %f882, %p73;
add.f32 %f889, %f888, 0fBF800000;
add.f32 %f890, %f888, 0f3F800000;
rcp.approx.ftz.f32 %f891, %f890;
add.f32 %f892, %f889, %f889;
mul.f32 %f893, %f892, %f891;
mul.f32 %f894, %f893, %f893;
mov.f32 %f895, 0f3C4CAF63;
mov.f32 %f896, 0f3B18F0FE;
fma.rn.f32 %f897, %f896, %f894, %f895;
mov.f32 %f898, 0f3DAAAABD;
fma.rn.f32 %f899, %f897, %f894, %f898;
mul.rn.f32 %f900, %f899, %f894;
mul.rn.f32 %f901, %f900, %f893;
sub.f32 %f902, %f889, %f893;
add.f32 %f903, %f902, %f902;
neg.f32 %f904, %f893;
fma.rn.f32 %f905, %f904, %f889, %f903;
mul.rn.f32 %f906, %f891, %f905;
add.f32 %f907, %f901, %f893;
sub.f32 %f908, %f893, %f907;
add.f32 %f909, %f901, %f908;
add.f32 %f910, %f906, %f909;
add.f32 %f911, %f907, %f910;
sub.f32 %f912, %f907, %f911;
add.f32 %f913, %f910, %f912;
mov.f32 %f914, 0f3F317200;
mul.rn.f32 %f915, %f887, %f914;
mov.f32 %f916, 0f35BFBE8E;
mul.rn.f32 %f917, %f887, %f916;
add.f32 %f918, %f915, %f911;
sub.f32 %f919, %f915, %f918;
add.f32 %f920, %f911, %f919;
add.f32 %f921, %f913, %f920;
add.f32 %f922, %f917, %f921;
add.f32 %f923, %f918, %f922;
sub.f32 %f924, %f918, %f923;
add.f32 %f925, %f922, %f924;
mul.rn.f32 %f926, %f877, %f923;
neg.f32 %f927, %f926;
fma.rn.f32 %f928, %f877, %f923, %f927;
fma.rn.f32 %f929, %f877, %f925, %f928;
mov.f32 %f930, 0f00000000;
fma.rn.f32 %f931, %f930, %f923, %f929;
add.rn.f32 %f932, %f926, %f931;
neg.f32 %f933, %f932;
add.rn.f32 %f934, %f926, %f933;
add.rn.f32 %f935, %f934, %f931;
mov.b32 %r304, %f932;
setp.eq.s32 %p74, %r304, 1118925336;
add.s32 %r305, %r304, -1;
mov.b32 %f936, %r305;
add.f32 %f937, %f935, 0f37000000;
selp.f32 %f229, %f937, %f935, %p74;
selp.f32 %f938, %f936, %f932, %p74;
mov.f32 %f939, 0f3FB8AA3B;
mul.rn.f32 %f940, %f938, %f939;
cvt.rzi.f32.f32 %f941, %f940;
abs.f32 %f942, %f941;
setp.gt.f32 %p75, %f942, 0f42FC0000;
mov.b32 %r306, %f941;
and.b32 %r307, %r306, -2147483648;
or.b32 %r308, %r307, 1123811328;
mov.b32 %f943, %r308;
selp.f32 %f944, %f943, %f941, %p75;
mov.f32 %f945, 0fBF317218;
fma.rn.f32 %f946, %f944, %f945, %f938;
mov.f32 %f947, 0f3102E308;
fma.rn.f32 %f948, %f944, %f947, %f946;
mul.f32 %f949, %f948, 0f3FB8AA3B;
add.f32 %f950, %f944, 0f4B40007F;
mov.b32 %r309, %f950;
shl.b32 %r310, %r309, 23;
mov.b32 %f951, %r310;
ex2.approx.ftz.f32 %f952, %f949;
mul.f32 %f230, %f952, %f951;
setp.eq.f32 %p76, %f230, 0f7F800000;
mov.f32 %f1614, 0f7F800000;
@%p76 bra $L__BB0_62;
fma.rn.f32 %f1614, %f230, %f229, %f230;
$L__BB0_62:
setp.lt.f32 %p77, %f1579, 0f00000000;
setp.eq.f32 %p78, %f227, 0f3F800000;
and.pred %p4, %p77, %p78;
setp.eq.f32 %p79, %f1579, 0f00000000;
@%p79 bra $L__BB0_66;
bra.uni $L__BB0_63;
$L__BB0_66:
add.f32 %f957, %f1579, %f1579;
selp.f32 %f1616, %f957, 0f00000000, %p78;
bra.uni $L__BB0_67;
$L__BB0_63:
mov.b32 %r311, %f1614;
xor.b32 %r312, %r311, -2147483648;
mov.b32 %f953, %r312;
selp.f32 %f1616, %f953, %f1614, %p4;
setp.geu.f32 %p80, %f1579, 0f00000000;
@%p80 bra $L__BB0_67;
mov.f32 %f954, 0f3EE66666;
cvt.rzi.f32.f32 %f955, %f954;
setp.eq.f32 %p81, %f955, 0f3EE66666;
@%p81 bra $L__BB0_67;
mov.f32 %f1616, 0f7FFFFFFF;
$L__BB0_67:
add.f32 %f958, %f228, 0f3EE66666;
mov.b32 %r313, %f958;
setp.lt.s32 %p83, %r313, 2139095040;
@%p83 bra $L__BB0_72;
setp.gtu.f32 %p84, %f228, 0f7F800000;
@%p84 bra $L__BB0_71;
bra.uni $L__BB0_69;
$L__BB0_71:
add.f32 %f1616, %f1579, 0f3EE66666;
bra.uni $L__BB0_72;
$L__BB0_69:
setp.neu.f32 %p85, %f228, 0f7F800000;
@%p85 bra $L__BB0_72;
selp.f32 %f1616, 0fFF800000, 0f7F800000, %p4;
$L__BB0_72:
setp.eq.f32 %p86, %f1579, 0f3F800000;
selp.f32 %f239, 0f3F800000, %f1616, %p86;
abs.f32 %f240, %f1580;
setp.lt.f32 %p87, %f240, 0f00800000;
mul.f32 %f960, %f240, 0f4B800000;
selp.f32 %f961, %f960, %f240, %p87;
selp.f32 %f962, 0fC3170000, 0fC2FE0000, %p87;
mov.b32 %r314, %f961;
and.b32 %r315, %r314, 8388607;
or.b32 %r316, %r315, 1065353216;
mov.b32 %f963, %r316;
shr.u32 %r317, %r314, 23;
cvt.rn.f32.u32 %f964, %r317;
add.f32 %f965, %f962, %f964;
setp.gt.f32 %p88, %f963, 0f3FB504F3;
mul.f32 %f966, %f963, 0f3F000000;
add.f32 %f967, %f965, 0f3F800000;
selp.f32 %f968, %f967, %f965, %p88;
selp.f32 %f969, %f966, %f963, %p88;
add.f32 %f970, %f969, 0fBF800000;
add.f32 %f971, %f969, 0f3F800000;
rcp.approx.ftz.f32 %f972, %f971;
add.f32 %f973, %f970, %f970;
mul.f32 %f974, %f973, %f972;
mul.f32 %f975, %f974, %f974;
mov.f32 %f976, 0f3C4CAF63;
mov.f32 %f977, 0f3B18F0FE;
fma.rn.f32 %f978, %f977, %f975, %f976;
mov.f32 %f979, 0f3DAAAABD;
fma.rn.f32 %f980, %f978, %f975, %f979;
mul.rn.f32 %f981, %f980, %f975;
mul.rn.f32 %f982, %f981, %f974;
sub.f32 %f983, %f970, %f974;
add.f32 %f984, %f983, %f983;
neg.f32 %f985, %f974;
fma.rn.f32 %f986, %f985, %f970, %f984;
mul.rn.f32 %f987, %f972, %f986;
add.f32 %f988, %f982, %f974;
sub.f32 %f989, %f974, %f988;
add.f32 %f990, %f982, %f989;
add.f32 %f991, %f987, %f990;
add.f32 %f992, %f988, %f991;
sub.f32 %f993, %f988, %f992;
add.f32 %f994, %f991, %f993;
mov.f32 %f995, 0f3F317200;
mul.rn.f32 %f996, %f968, %f995;
mov.f32 %f997, 0f35BFBE8E;
mul.rn.f32 %f998, %f968, %f997;
add.f32 %f999, %f996, %f992;
sub.f32 %f1000, %f996, %f999;
add.f32 %f1001, %f992, %f1000;
add.f32 %f1002, %f994, %f1001;
add.f32 %f1003, %f998, %f1002;
add.f32 %f1004, %f999, %f1003;
sub.f32 %f1005, %f999, %f1004;
add.f32 %f1006, %f1003, %f1005;
mov.f32 %f1007, 0f3EE66666;
mul.rn.f32 %f1008, %f1007, %f1004;
neg.f32 %f1009, %f1008;
fma.rn.f32 %f1010, %f1007, %f1004, %f1009;
fma.rn.f32 %f1011, %f1007, %f1006, %f1010;
mov.f32 %f1012, 0f00000000;
fma.rn.f32 %f1013, %f1012, %f1004, %f1011;
add.rn.f32 %f1014, %f1008, %f1013;
neg.f32 %f1015, %f1014;
add.rn.f32 %f1016, %f1008, %f1015;
add.rn.f32 %f1017, %f1016, %f1013;
mov.b32 %r318, %f1014;
setp.eq.s32 %p89, %r318, 1118925336;
add.s32 %r319, %r318, -1;
mov.b32 %f1018, %r319;
add.f32 %f1019, %f1017, 0f37000000;
selp.f32 %f241, %f1019, %f1017, %p89;
selp.f32 %f1020, %f1018, %f1014, %p89;
mov.f32 %f1021, 0f3FB8AA3B;
mul.rn.f32 %f1022, %f1020, %f1021;
cvt.rzi.f32.f32 %f1023, %f1022;
abs.f32 %f1024, %f1023;
setp.gt.f32 %p90, %f1024, 0f42FC0000;
mov.b32 %r320, %f1023;
and.b32 %r321, %r320, -2147483648;
or.b32 %r322, %r321, 1123811328;
mov.b32 %f1025, %r322;
selp.f32 %f1026, %f1025, %f1023, %p90;
mov.f32 %f1027, 0fBF317218;
fma.rn.f32 %f1028, %f1026, %f1027, %f1020;
mov.f32 %f1029, 0f3102E308;
fma.rn.f32 %f1030, %f1026, %f1029, %f1028;
mul.f32 %f1031, %f1030, 0f3FB8AA3B;
add.f32 %f1032, %f1026, 0f4B40007F;
mov.b32 %r323, %f1032;
shl.b32 %r324, %r323, 23;
mov.b32 %f1033, %r324;
ex2.approx.ftz.f32 %f1034, %f1031;
mul.f32 %f242, %f1034, %f1033;
setp.eq.f32 %p91, %f242, 0f7F800000;
mov.f32 %f1617, 0f7F800000;
@%p91 bra $L__BB0_74;
fma.rn.f32 %f1617, %f242, %f241, %f242;
$L__BB0_74:
setp.lt.f32 %p92, %f1580, 0f00000000;
and.pred %p5, %p92, %p78;
setp.eq.f32 %p94, %f1580, 0f00000000;
@%p94 bra $L__BB0_78;
bra.uni $L__BB0_75;
$L__BB0_78:
add.f32 %f1039, %f1580, %f1580;
selp.f32 %f1619, %f1039, 0f00000000, %p78;
bra.uni $L__BB0_79;
$L__BB0_75:
mov.b32 %r325, %f1617;
xor.b32 %r326, %r325, -2147483648;
mov.b32 %f1035, %r326;
selp.f32 %f1619, %f1035, %f1617, %p5;
setp.geu.f32 %p95, %f1580, 0f00000000;
@%p95 bra $L__BB0_79;
mov.f32 %f1036, 0f3EE66666;
cvt.rzi.f32.f32 %f1037, %f1036;
setp.eq.f32 %p96, %f1037, 0f3EE66666;
@%p96 bra $L__BB0_79;
mov.f32 %f1619, 0f7FFFFFFF;
$L__BB0_79:
add.f32 %f1040, %f240, 0f3EE66666;
mov.b32 %r327, %f1040;
setp.lt.s32 %p98, %r327, 2139095040;
@%p98 bra $L__BB0_84;
setp.gtu.f32 %p99, %f240, 0f7F800000;
@%p99 bra $L__BB0_83;
bra.uni $L__BB0_81;
$L__BB0_83:
add.f32 %f1619, %f1580, 0f3EE66666;
bra.uni $L__BB0_84;
$L__BB0_81:
setp.neu.f32 %p100, %f240, 0f7F800000;
@%p100 bra $L__BB0_84;
selp.f32 %f1619, 0fFF800000, 0f7F800000, %p5;
$L__BB0_84:
setp.eq.f32 %p101, %f1580, 0f3F800000;
selp.f32 %f251, 0f3F800000, %f1619, %p101;
abs.f32 %f252, %f1581;
setp.lt.f32 %p102, %f252, 0f00800000;
mul.f32 %f1042, %f252, 0f4B800000;
selp.f32 %f1043, %f1042, %f252, %p102;
selp.f32 %f1044, 0fC3170000, 0fC2FE0000, %p102;
mov.b32 %r328, %f1043;
and.b32 %r329, %r328, 8388607;
or.b32 %r330, %r329, 1065353216;
mov.b32 %f1045, %r330;
shr.u32 %r331, %r328, 23;
cvt.rn.f32.u32 %f1046, %r331;
add.f32 %f1047, %f1044, %f1046;
setp.gt.f32 %p103, %f1045, 0f3FB504F3;
mul.f32 %f1048, %f1045, 0f3F000000;
add.f32 %f1049, %f1047, 0f3F800000;
selp.f32 %f1050, %f1049, %f1047, %p103;
selp.f32 %f1051, %f1048, %f1045, %p103;
add.f32 %f1052, %f1051, 0fBF800000;
add.f32 %f1053, %f1051, 0f3F800000;
rcp.approx.ftz.f32 %f1054, %f1053;
add.f32 %f1055, %f1052, %f1052;
mul.f32 %f1056, %f1055, %f1054;
mul.f32 %f1057, %f1056, %f1056;
mov.f32 %f1058, 0f3C4CAF63;
mov.f32 %f1059, 0f3B18F0FE;
fma.rn.f32 %f1060, %f1059, %f1057, %f1058;
mov.f32 %f1061, 0f3DAAAABD;
fma.rn.f32 %f1062, %f1060, %f1057, %f1061;
mul.rn.f32 %f1063, %f1062, %f1057;
mul.rn.f32 %f1064, %f1063, %f1056;
sub.f32 %f1065, %f1052, %f1056;
add.f32 %f1066, %f1065, %f1065;
neg.f32 %f1067, %f1056;
fma.rn.f32 %f1068, %f1067, %f1052, %f1066;
mul.rn.f32 %f1069, %f1054, %f1068;
add.f32 %f1070, %f1064, %f1056;
sub.f32 %f1071, %f1056, %f1070;
add.f32 %f1072, %f1064, %f1071;
add.f32 %f1073, %f1069, %f1072;
add.f32 %f1074, %f1070, %f1073;
sub.f32 %f1075, %f1070, %f1074;
add.f32 %f1076, %f1073, %f1075;
mov.f32 %f1077, 0f3F317200;
mul.rn.f32 %f1078, %f1050, %f1077;
mov.f32 %f1079, 0f35BFBE8E;
mul.rn.f32 %f1080, %f1050, %f1079;
add.f32 %f1081, %f1078, %f1074;
sub.f32 %f1082, %f1078, %f1081;
add.f32 %f1083, %f1074, %f1082;
add.f32 %f1084, %f1076, %f1083;
add.f32 %f1085, %f1080, %f1084;
add.f32 %f1086, %f1081, %f1085;
sub.f32 %f1087, %f1081, %f1086;
add.f32 %f1088, %f1085, %f1087;
mov.f32 %f1089, 0f3EE66666;
mul.rn.f32 %f1090, %f1089, %f1086;
neg.f32 %f1091, %f1090;
fma.rn.f32 %f1092, %f1089, %f1086, %f1091;
fma.rn.f32 %f1093, %f1089, %f1088, %f1092;
mov.f32 %f1094, 0f00000000;
fma.rn.f32 %f1095, %f1094, %f1086, %f1093;
add.rn.f32 %f1096, %f1090, %f1095;
neg.f32 %f1097, %f1096;
add.rn.f32 %f1098, %f1090, %f1097;
add.rn.f32 %f1099, %f1098, %f1095;
mov.b32 %r332, %f1096;
setp.eq.s32 %p104, %r332, 1118925336;
add.s32 %r333, %r332, -1;
mov.b32 %f1100, %r333;
add.f32 %f1101, %f1099, 0f37000000;
selp.f32 %f253, %f1101, %f1099, %p104;
selp.f32 %f1102, %f1100, %f1096, %p104;
mov.f32 %f1103, 0f3FB8AA3B;
mul.rn.f32 %f1104, %f1102, %f1103;
cvt.rzi.f32.f32 %f1105, %f1104;
abs.f32 %f1106, %f1105;
setp.gt.f32 %p105, %f1106, 0f42FC0000;
mov.b32 %r334, %f1105;
and.b32 %r335, %r334, -2147483648;
or.b32 %r336, %r335, 1123811328;
mov.b32 %f1107, %r336;
selp.f32 %f1108, %f1107, %f1105, %p105;
mov.f32 %f1109, 0fBF317218;
fma.rn.f32 %f1110, %f1108, %f1109, %f1102;
mov.f32 %f1111, 0f3102E308;
fma.rn.f32 %f1112, %f1108, %f1111, %f1110;
mul.f32 %f1113, %f1112, 0f3FB8AA3B;
add.f32 %f1114, %f1108, 0f4B40007F;
mov.b32 %r337, %f1114;
shl.b32 %r338, %r337, 23;
mov.b32 %f1115, %r338;
ex2.approx.ftz.f32 %f1116, %f1113;
mul.f32 %f254, %f1116, %f1115;
setp.eq.f32 %p106, %f254, 0f7F800000;
mov.f32 %f1620, 0f7F800000;
@%p106 bra $L__BB0_86;
fma.rn.f32 %f1620, %f254, %f253, %f254;
$L__BB0_86:
setp.lt.f32 %p107, %f1581, 0f00000000;
and.pred %p6, %p107, %p78;
setp.eq.f32 %p109, %f1581, 0f00000000;
@%p109 bra $L__BB0_90;
bra.uni $L__BB0_87;
$L__BB0_90:
add.f32 %f1121, %f1581, %f1581;
selp.f32 %f1622, %f1121, 0f00000000, %p78;
bra.uni $L__BB0_91;
$L__BB0_87:
mov.b32 %r339, %f1620;
xor.b32 %r340, %r339, -2147483648;
mov.b32 %f1117, %r340;
selp.f32 %f1622, %f1117, %f1620, %p6;
setp.geu.f32 %p110, %f1581, 0f00000000;
@%p110 bra $L__BB0_91;
mov.f32 %f1118, 0f3EE66666;
cvt.rzi.f32.f32 %f1119, %f1118;
setp.eq.f32 %p111, %f1119, 0f3EE66666;
@%p111 bra $L__BB0_91;
mov.f32 %f1622, 0f7FFFFFFF;
$L__BB0_91:
add.f32 %f1122, %f252, 0f3EE66666;
mov.b32 %r341, %f1122;
setp.lt.s32 %p113, %r341, 2139095040;
@%p113 bra $L__BB0_96;
setp.gtu.f32 %p114, %f252, 0f7F800000;
@%p114 bra $L__BB0_95;
bra.uni $L__BB0_93;
$L__BB0_95:
add.f32 %f1622, %f1581, 0f3EE66666;
bra.uni $L__BB0_96;
$L__BB0_93:
setp.neu.f32 %p115, %f252, 0f7F800000;
@%p115 bra $L__BB0_96;
selp.f32 %f1622, 0fFF800000, 0f7F800000, %p6;
$L__BB0_96:
setp.eq.f32 %p116, %f1581, 0f3F800000;
mov.f32 %f1124, 0f3F800000;
selp.f32 %f1125, 0f3F800000, %f1622, %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 %f1126, %f239, %f1124;
mov.f32 %f1127, 0f00000000;
max.f32 %f263, %f1127, %f1126;
min.f32 %f1128, %f251, %f1124;
max.f32 %f264, %f1127, %f1128;
min.f32 %f1129, %f1125, %f1124;
max.f32 %f265, %f1127, %f1129;
mov.f32 %f1130, 0f3E555555;
cvt.rzi.f32.f32 %f1131, %f1130;
add.f32 %f1132, %f1131, %f1131;
mov.f32 %f1133, 0f3ED55555;
sub.f32 %f1134, %f1133, %f1132;
abs.f32 %f266, %f1134;
abs.f32 %f267, %f263;
setp.lt.f32 %p117, %f267, 0f00800000;
mul.f32 %f1135, %f267, 0f4B800000;
selp.f32 %f1136, %f1135, %f267, %p117;
selp.f32 %f1137, 0fC3170000, 0fC2FE0000, %p117;
mov.b32 %r344, %f1136;
and.b32 %r345, %r344, 8388607;
or.b32 %r346, %r345, 1065353216;
mov.b32 %f1138, %r346;
shr.u32 %r347, %r344, 23;
cvt.rn.f32.u32 %f1139, %r347;
add.f32 %f1140, %f1137, %f1139;
setp.gt.f32 %p118, %f1138, 0f3FB504F3;
mul.f32 %f1141, %f1138, 0f3F000000;
add.f32 %f1142, %f1140, 0f3F800000;
selp.f32 %f1143, %f1142, %f1140, %p118;
selp.f32 %f1144, %f1141, %f1138, %p118;
add.f32 %f1145, %f1144, 0fBF800000;
add.f32 %f1146, %f1144, 0f3F800000;
rcp.approx.ftz.f32 %f1147, %f1146;
add.f32 %f1148, %f1145, %f1145;
mul.f32 %f1149, %f1148, %f1147;
mul.f32 %f1150, %f1149, %f1149;
mov.f32 %f1151, 0f3C4CAF63;
mov.f32 %f1152, 0f3B18F0FE;
fma.rn.f32 %f1153, %f1152, %f1150, %f1151;
mov.f32 %f1154, 0f3DAAAABD;
fma.rn.f32 %f1155, %f1153, %f1150, %f1154;
mul.rn.f32 %f1156, %f1155, %f1150;
mul.rn.f32 %f1157, %f1156, %f1149;
sub.f32 %f1158, %f1145, %f1149;
add.f32 %f1159, %f1158, %f1158;
neg.f32 %f1160, %f1149;
fma.rn.f32 %f1161, %f1160, %f1145, %f1159;
mul.rn.f32 %f1162, %f1147, %f1161;
add.f32 %f1163, %f1157, %f1149;
sub.f32 %f1164, %f1149, %f1163;
add.f32 %f1165, %f1157, %f1164;
add.f32 %f1166, %f1162, %f1165;
add.f32 %f1167, %f1163, %f1166;
sub.f32 %f1168, %f1163, %f1167;
add.f32 %f1169, %f1166, %f1168;
mov.f32 %f1170, 0f3F317200;
mul.rn.f32 %f1171, %f1143, %f1170;
mov.f32 %f1172, 0f35BFBE8E;
mul.rn.f32 %f1173, %f1143, %f1172;
add.f32 %f1174, %f1171, %f1167;
sub.f32 %f1175, %f1171, %f1174;
add.f32 %f1176, %f1167, %f1175;
add.f32 %f1177, %f1169, %f1176;
add.f32 %f1178, %f1173, %f1177;
add.f32 %f1179, %f1174, %f1178;
sub.f32 %f1180, %f1174, %f1179;
add.f32 %f1181, %f1178, %f1180;
mul.rn.f32 %f1182, %f1133, %f1179;
neg.f32 %f1183, %f1182;
fma.rn.f32 %f1184, %f1133, %f1179, %f1183;
fma.rn.f32 %f1185, %f1133, %f1181, %f1184;
fma.rn.f32 %f1186, %f1127, %f1179, %f1185;
add.rn.f32 %f1187, %f1182, %f1186;
neg.f32 %f1188, %f1187;
add.rn.f32 %f1189, %f1182, %f1188;
add.rn.f32 %f1190, %f1189, %f1186;
mov.b32 %r348, %f1187;
setp.eq.s32 %p119, %r348, 1118925336;
add.s32 %r349, %r348, -1;
mov.b32 %f1191, %r349;
add.f32 %f1192, %f1190, 0f37000000;
selp.f32 %f268, %f1192, %f1190, %p119;
selp.f32 %f1193, %f1191, %f1187, %p119;
mov.f32 %f1194, 0f3FB8AA3B;
mul.rn.f32 %f1195, %f1193, %f1194;
cvt.rzi.f32.f32 %f1196, %f1195;
abs.f32 %f1197, %f1196;
setp.gt.f32 %p120, %f1197, 0f42FC0000;
mov.b32 %r350, %f1196;
and.b32 %r351, %r350, -2147483648;
or.b32 %r352, %r351, 1123811328;
mov.b32 %f1198, %r352;
selp.f32 %f1199, %f1198, %f1196, %p120;
mov.f32 %f1200, 0fBF317218;
fma.rn.f32 %f1201, %f1199, %f1200, %f1193;
mov.f32 %f1202, 0f3102E308;
fma.rn.f32 %f1203, %f1199, %f1202, %f1201;
mul.f32 %f1204, %f1203, 0f3FB8AA3B;
add.f32 %f1205, %f1199, 0f4B40007F;
mov.b32 %r353, %f1205;
shl.b32 %r354, %r353, 23;
mov.b32 %f1206, %r354;
ex2.approx.ftz.f32 %f1207, %f1204;
mul.f32 %f269, %f1207, %f1206;
setp.eq.f32 %p121, %f269, 0f7F800000;
mov.f32 %f1623, 0f7F800000;
@%p121 bra $L__BB0_98;
fma.rn.f32 %f1623, %f269, %f268, %f269;
$L__BB0_98:
setp.lt.f32 %p122, %f263, 0f00000000;
setp.eq.f32 %p123, %f266, 0f3F800000;
and.pred %p7, %p122, %p123;
setp.eq.f32 %p124, %f263, 0f00000000;
@%p124 bra $L__BB0_102;
bra.uni $L__BB0_99;
$L__BB0_102:
add.f32 %f1212, %f263, %f263;
selp.f32 %f1625, %f1212, 0f00000000, %p123;
bra.uni $L__BB0_103;
$L__BB0_99:
mov.b32 %r355, %f1623;
xor.b32 %r356, %r355, -2147483648;
mov.b32 %f1208, %r356;
selp.f32 %f1625, %f1208, %f1623, %p7;
setp.geu.f32 %p125, %f263, 0f00000000;
@%p125 bra $L__BB0_103;
mov.f32 %f1209, 0f3ED55555;
cvt.rzi.f32.f32 %f1210, %f1209;
setp.eq.f32 %p126, %f1210, 0f3ED55555;
@%p126 bra $L__BB0_103;
mov.f32 %f1625, 0f7FFFFFFF;
$L__BB0_103:
add.f32 %f1213, %f267, 0f3ED55555;
mov.b32 %r357, %f1213;
setp.lt.s32 %p128, %r357, 2139095040;
@%p128 bra $L__BB0_108;
setp.gtu.f32 %p129, %f267, 0f7F800000;
@%p129 bra $L__BB0_107;
bra.uni $L__BB0_105;
$L__BB0_107:
add.f32 %f1625, %f263, 0f3ED55555;
bra.uni $L__BB0_108;
$L__BB0_105:
setp.neu.f32 %p130, %f267, 0f7F800000;
@%p130 bra $L__BB0_108;
selp.f32 %f1625, 0fFF800000, 0f7F800000, %p7;
$L__BB0_108:
abs.f32 %f278, %f264;
setp.lt.f32 %p131, %f278, 0f00800000;
mul.f32 %f1215, %f278, 0f4B800000;
selp.f32 %f1216, %f1215, %f278, %p131;
selp.f32 %f1217, 0fC3170000, 0fC2FE0000, %p131;
mov.b32 %r358, %f1216;
and.b32 %r359, %r358, 8388607;
or.b32 %r360, %r359, 1065353216;
mov.b32 %f1218, %r360;
shr.u32 %r361, %r358, 23;
cvt.rn.f32.u32 %f1219, %r361;
add.f32 %f1220, %f1217, %f1219;
setp.gt.f32 %p132, %f1218, 0f3FB504F3;
mul.f32 %f1221, %f1218, 0f3F000000;
add.f32 %f1222, %f1220, 0f3F800000;
selp.f32 %f1223, %f1222, %f1220, %p132;
selp.f32 %f1224, %f1221, %f1218, %p132;
add.f32 %f1225, %f1224, 0fBF800000;
add.f32 %f1226, %f1224, 0f3F800000;
rcp.approx.ftz.f32 %f1227, %f1226;
add.f32 %f1228, %f1225, %f1225;
mul.f32 %f1229, %f1228, %f1227;
mul.f32 %f1230, %f1229, %f1229;
mov.f32 %f1231, 0f3C4CAF63;
mov.f32 %f1232, 0f3B18F0FE;
fma.rn.f32 %f1233, %f1232, %f1230, %f1231;
mov.f32 %f1234, 0f3DAAAABD;
fma.rn.f32 %f1235, %f1233, %f1230, %f1234;
mul.rn.f32 %f1236, %f1235, %f1230;
mul.rn.f32 %f1237, %f1236, %f1229;
sub.f32 %f1238, %f1225, %f1229;
add.f32 %f1239, %f1238, %f1238;
neg.f32 %f1240, %f1229;
fma.rn.f32 %f1241, %f1240, %f1225, %f1239;
mul.rn.f32 %f1242, %f1227, %f1241;
add.f32 %f1243, %f1237, %f1229;
sub.f32 %f1244, %f1229, %f1243;
add.f32 %f1245, %f1237, %f1244;
add.f32 %f1246, %f1242, %f1245;
add.f32 %f1247, %f1243, %f1246;
sub.f32 %f1248, %f1243, %f1247;
add.f32 %f1249, %f1246, %f1248;
mov.f32 %f1250, 0f3F317200;
mul.rn.f32 %f1251, %f1223, %f1250;
mov.f32 %f1252, 0f35BFBE8E;
mul.rn.f32 %f1253, %f1223, %f1252;
add.f32 %f1254, %f1251, %f1247;
sub.f32 %f1255, %f1251, %f1254;
add.f32 %f1256, %f1247, %f1255;
add.f32 %f1257, %f1249, %f1256;
add.f32 %f1258, %f1253, %f1257;
add.f32 %f1259, %f1254, %f1258;
sub.f32 %f1260, %f1254, %f1259;
add.f32 %f1261, %f1258, %f1260;
mov.f32 %f1262, 0f3ED55555;
mul.rn.f32 %f1263, %f1262, %f1259;
neg.f32 %f1264, %f1263;
fma.rn.f32 %f1265, %f1262, %f1259, %f1264;
fma.rn.f32 %f1266, %f1262, %f1261, %f1265;
mov.f32 %f1267, 0f00000000;
fma.rn.f32 %f1268, %f1267, %f1259, %f1266;
add.rn.f32 %f1269, %f1263, %f1268;
neg.f32 %f1270, %f1269;
add.rn.f32 %f1271, %f1263, %f1270;
add.rn.f32 %f1272, %f1271, %f1268;
mov.b32 %r362, %f1269;
setp.eq.s32 %p133, %r362, 1118925336;
add.s32 %r363, %r362, -1;
mov.b32 %f1273, %r363;
add.f32 %f1274, %f1272, 0f37000000;
selp.f32 %f279, %f1274, %f1272, %p133;
selp.f32 %f1275, %f1273, %f1269, %p133;
mov.f32 %f1276, 0f3FB8AA3B;
mul.rn.f32 %f1277, %f1275, %f1276;
cvt.rzi.f32.f32 %f1278, %f1277;
abs.f32 %f1279, %f1278;
setp.gt.f32 %p134, %f1279, 0f42FC0000;
mov.b32 %r364, %f1278;
and.b32 %r365, %r364, -2147483648;
or.b32 %r366, %r365, 1123811328;
mov.b32 %f1280, %r366;
selp.f32 %f1281, %f1280, %f1278, %p134;
mov.f32 %f1282, 0fBF317218;
fma.rn.f32 %f1283, %f1281, %f1282, %f1275;
mov.f32 %f1284, 0f3102E308;
fma.rn.f32 %f1285, %f1281, %f1284, %f1283;
mul.f32 %f1286, %f1285, 0f3FB8AA3B;
add.f32 %f1287, %f1281, 0f4B40007F;
mov.b32 %r367, %f1287;
shl.b32 %r368, %r367, 23;
mov.b32 %f1288, %r368;
ex2.approx.ftz.f32 %f1289, %f1286;
mul.f32 %f280, %f1289, %f1288;
setp.eq.f32 %p135, %f280, 0f7F800000;
mov.f32 %f1626, 0f7F800000;
@%p135 bra $L__BB0_110;
fma.rn.f32 %f1626, %f280, %f279, %f280;
$L__BB0_110:
setp.lt.f32 %p136, %f264, 0f00000000;
and.pred %p8, %p136, %p123;
setp.eq.f32 %p138, %f264, 0f00000000;
@%p138 bra $L__BB0_114;
bra.uni $L__BB0_111;
$L__BB0_114:
add.f32 %f1294, %f264, %f264;
selp.f32 %f1628, %f1294, 0f00000000, %p123;
bra.uni $L__BB0_115;
$L__BB0_111:
mov.b32 %r369, %f1626;
xor.b32 %r370, %r369, -2147483648;
mov.b32 %f1290, %r370;
selp.f32 %f1628, %f1290, %f1626, %p8;
setp.geu.f32 %p139, %f264, 0f00000000;
@%p139 bra $L__BB0_115;
mov.f32 %f1291, 0f3ED55555;
cvt.rzi.f32.f32 %f1292, %f1291;
setp.eq.f32 %p140, %f1292, 0f3ED55555;
@%p140 bra $L__BB0_115;
mov.f32 %f1628, 0f7FFFFFFF;
$L__BB0_115:
add.f32 %f1295, %f278, 0f3ED55555;
mov.b32 %r371, %f1295;
setp.lt.s32 %p142, %r371, 2139095040;
@%p142 bra $L__BB0_120;
setp.gtu.f32 %p143, %f278, 0f7F800000;
@%p143 bra $L__BB0_119;
bra.uni $L__BB0_117;
$L__BB0_119:
add.f32 %f1628, %f264, 0f3ED55555;
bra.uni $L__BB0_120;
$L__BB0_117:
setp.neu.f32 %p144, %f278, 0f7F800000;
@%p144 bra $L__BB0_120;
selp.f32 %f1628, 0fFF800000, 0f7F800000, %p8;
$L__BB0_120:
abs.f32 %f289, %f265;
setp.lt.f32 %p145, %f289, 0f00800000;
mul.f32 %f1297, %f289, 0f4B800000;
selp.f32 %f1298, %f1297, %f289, %p145;
selp.f32 %f1299, 0fC3170000, 0fC2FE0000, %p145;
mov.b32 %r372, %f1298;
and.b32 %r373, %r372, 8388607;
or.b32 %r374, %r373, 1065353216;
mov.b32 %f1300, %r374;
shr.u32 %r375, %r372, 23;
cvt.rn.f32.u32 %f1301, %r375;
add.f32 %f1302, %f1299, %f1301;
setp.gt.f32 %p146, %f1300, 0f3FB504F3;
mul.f32 %f1303, %f1300, 0f3F000000;
add.f32 %f1304, %f1302, 0f3F800000;
selp.f32 %f1305, %f1304, %f1302, %p146;
selp.f32 %f1306, %f1303, %f1300, %p146;
add.f32 %f1307, %f1306, 0fBF800000;
add.f32 %f1308, %f1306, 0f3F800000;
rcp.approx.ftz.f32 %f1309, %f1308;
add.f32 %f1310, %f1307, %f1307;
mul.f32 %f1311, %f1310, %f1309;
mul.f32 %f1312, %f1311, %f1311;
mov.f32 %f1313, 0f3C4CAF63;
mov.f32 %f1314, 0f3B18F0FE;
fma.rn.f32 %f1315, %f1314, %f1312, %f1313;
mov.f32 %f1316, 0f3DAAAABD;
fma.rn.f32 %f1317, %f1315, %f1312, %f1316;
mul.rn.f32 %f1318, %f1317, %f1312;
mul.rn.f32 %f1319, %f1318, %f1311;
sub.f32 %f1320, %f1307, %f1311;
add.f32 %f1321, %f1320, %f1320;
neg.f32 %f1322, %f1311;
fma.rn.f32 %f1323, %f1322, %f1307, %f1321;
mul.rn.f32 %f1324, %f1309, %f1323;
add.f32 %f1325, %f1319, %f1311;
sub.f32 %f1326, %f1311, %f1325;
add.f32 %f1327, %f1319, %f1326;
add.f32 %f1328, %f1324, %f1327;
add.f32 %f1329, %f1325, %f1328;
sub.f32 %f1330, %f1325, %f1329;
add.f32 %f1331, %f1328, %f1330;
mov.f32 %f1332, 0f3F317200;
mul.rn.f32 %f1333, %f1305, %f1332;
mov.f32 %f1334, 0f35BFBE8E;
mul.rn.f32 %f1335, %f1305, %f1334;
add.f32 %f1336, %f1333, %f1329;
sub.f32 %f1337, %f1333, %f1336;
add.f32 %f1338, %f1329, %f1337;
add.f32 %f1339, %f1331, %f1338;
add.f32 %f1340, %f1335, %f1339;
add.f32 %f1341, %f1336, %f1340;
sub.f32 %f1342, %f1336, %f1341;
add.f32 %f1343, %f1340, %f1342;
mov.f32 %f1344, 0f3ED55555;
mul.rn.f32 %f1345, %f1344, %f1341;
neg.f32 %f1346, %f1345;
fma.rn.f32 %f1347, %f1344, %f1341, %f1346;
fma.rn.f32 %f1348, %f1344, %f1343, %f1347;
mov.f32 %f1349, 0f00000000;
fma.rn.f32 %f1350, %f1349, %f1341, %f1348;
add.rn.f32 %f1351, %f1345, %f1350;
neg.f32 %f1352, %f1351;
add.rn.f32 %f1353, %f1345, %f1352;
add.rn.f32 %f1354, %f1353, %f1350;
mov.b32 %r376, %f1351;
setp.eq.s32 %p147, %r376, 1118925336;
add.s32 %r377, %r376, -1;
mov.b32 %f1355, %r377;
add.f32 %f1356, %f1354, 0f37000000;
selp.f32 %f290, %f1356, %f1354, %p147;
selp.f32 %f1357, %f1355, %f1351, %p147;
mov.f32 %f1358, 0f3FB8AA3B;
mul.rn.f32 %f1359, %f1357, %f1358;
cvt.rzi.f32.f32 %f1360, %f1359;
abs.f32 %f1361, %f1360;
setp.gt.f32 %p148, %f1361, 0f42FC0000;
mov.b32 %r378, %f1360;
and.b32 %r379, %r378, -2147483648;
or.b32 %r380, %r379, 1123811328;
mov.b32 %f1362, %r380;
selp.f32 %f1363, %f1362, %f1360, %p148;
mov.f32 %f1364, 0fBF317218;
fma.rn.f32 %f1365, %f1363, %f1364, %f1357;
mov.f32 %f1366, 0f3102E308;
fma.rn.f32 %f1367, %f1363, %f1366, %f1365;
mul.f32 %f1368, %f1367, 0f3FB8AA3B;
add.f32 %f1369, %f1363, 0f4B40007F;
mov.b32 %r381, %f1369;
shl.b32 %r382, %r381, 23;
mov.b32 %f1370, %r382;
ex2.approx.ftz.f32 %f1371, %f1368;
mul.f32 %f291, %f1371, %f1370;
setp.eq.f32 %p149, %f291, 0f7F800000;
mov.f32 %f1629, 0f7F800000;
@%p149 bra $L__BB0_122;
fma.rn.f32 %f1629, %f291, %f290, %f291;
$L__BB0_122:
setp.lt.f32 %p150, %f265, 0f00000000;
and.pred %p9, %p150, %p123;
setp.eq.f32 %p152, %f265, 0f00000000;
@%p152 bra $L__BB0_126;
bra.uni $L__BB0_123;
$L__BB0_126:
add.f32 %f1376, %f265, %f265;
selp.f32 %f1631, %f1376, 0f00000000, %p123;
bra.uni $L__BB0_127;
$L__BB0_123:
mov.b32 %r383, %f1629;
xor.b32 %r384, %r383, -2147483648;
mov.b32 %f1372, %r384;
selp.f32 %f1631, %f1372, %f1629, %p9;
setp.geu.f32 %p153, %f265, 0f00000000;
@%p153 bra $L__BB0_127;
mov.f32 %f1373, 0f3ED55555;
cvt.rzi.f32.f32 %f1374, %f1373;
setp.eq.f32 %p154, %f1374, 0f3ED55555;
@%p154 bra $L__BB0_127;
mov.f32 %f1631, 0f7FFFFFFF;
$L__BB0_127:
add.f32 %f1377, %f289, 0f3ED55555;
mov.b32 %r385, %f1377;
setp.lt.s32 %p156, %r385, 2139095040;
@%p156 bra $L__BB0_132;
setp.gtu.f32 %p157, %f289, 0f7F800000;
@%p157 bra $L__BB0_131;
bra.uni $L__BB0_129;
$L__BB0_131:
add.f32 %f1631, %f265, 0f3ED55555;
bra.uni $L__BB0_132;
$L__BB0_129:
setp.neu.f32 %p158, %f289, 0f7F800000;
@%p158 bra $L__BB0_132;
selp.f32 %f1631, 0fFF800000, 0f7F800000, %p9;
$L__BB0_132:
fma.rn.f32 %f1378, %f1625, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p159, %f263, 0f3F800000;
mov.f32 %f1379, 0f3F800000;
selp.f32 %f1380, 0f3F7FFFFF, %f1378, %p159;
mul.f32 %f1381, %f263, 0f414EB852;
setp.lt.f32 %p160, %f263, 0f3B4D2E1C;
selp.f32 %f1382, %f1381, %f1380, %p160;
fma.rn.f32 %f1383, %f1628, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p161, %f264, 0f3F800000;
selp.f32 %f1384, 0f3F7FFFFF, %f1383, %p161;
mul.f32 %f1385, %f264, 0f414EB852;
setp.lt.f32 %p162, %f264, 0f3B4D2E1C;
selp.f32 %f1386, %f1385, %f1384, %p162;
fma.rn.f32 %f1387, %f1631, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p163, %f265, 0f3F800000;
selp.f32 %f1388, 0f3F7FFFFF, %f1387, %p163;
mul.f32 %f1389, %f265, 0f414EB852;
setp.lt.f32 %p164, %f265, 0f3B4D2E1C;
selp.f32 %f1390, %f1389, %f1388, %p164;
min.f32 %f1391, %f1382, %f1379;
mov.f32 %f1392, 0f00000000;
max.f32 %f1393, %f1392, %f1391;
mul.f32 %f1394, %f1393, 0f43800000;
cvt.rzi.u32.f32 %r386, %f1394;
min.u32 %r387, %r386, 255;
min.f32 %f1395, %f1386, %f1379;
max.f32 %f1396, %f1392, %f1395;
mul.f32 %f1397, %f1396, 0f43800000;
cvt.rzi.u32.f32 %r388, %f1397;
min.u32 %r389, %r388, 255;
min.f32 %f1398, %f1390, %f1379;
max.f32 %f1399, %f1392, %f1398;
mul.f32 %f1400, %f1399, 0f43800000;
cvt.rzi.u32.f32 %r390, %f1400;
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 %f1401, %rs31;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1402, %rs32;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1403, %rs33;}
// end inline asm
add.f32 %f1404, %f1579, %f1401;
add.f32 %f1405, %f1580, %f1402;
add.f32 %f1406, %f1581, %f1403;
mov.f32 %f1407, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f1406;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1405;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1404;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f1407;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs27, %rs28, %rs29, %rs30};
bra.uni $L__BB0_137;
$L__BB0_136:
mov.f32 %f1411, 0f3F800000;
mov.u32 %r428, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f1411;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f1581;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f1580;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f1579;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs35, %rs36, %rs37, %rs38};
$L__BB0_137:
mov.f32 %f1412, 0f34000000;
max.f32 %f1413, %f1582, %f1412;
div.rn.f32 %f1414, %f1585, %f1413;
max.f32 %f1415, %f1583, %f1412;
div.rn.f32 %f1416, %f1586, %f1415;
max.f32 %f1417, %f1584, %f1412;
div.rn.f32 %f1418, %f1587, %f1417;
fma.rn.f32 %f300, %f1414, 0f3F000000, 0f3F000000;
fma.rn.f32 %f301, %f1416, 0f3F000000, 0f3F000000;
fma.rn.f32 %f302, %f1418, 0f3F000000, 0f3F000000;
div.rn.f32 %f1419, %f1588, %f1413;
div.rn.f32 %f1420, %f1589, %f1415;
div.rn.f32 %f1421, %f1590, %f1417;
fma.rn.f32 %f303, %f1419, 0f3F000000, 0f3F000000;
fma.rn.f32 %f304, %f1420, 0f3F000000, 0f3F000000;
fma.rn.f32 %f305, %f1421, 0f3F000000, 0f3F000000;
div.rn.f32 %f1422, %f1591, %f1413;
div.rn.f32 %f1423, %f1592, %f1415;
div.rn.f32 %f1424, %f1593, %f1417;
fma.rn.f32 %f306, %f1422, 0f3F000000, 0f3F000000;
fma.rn.f32 %f307, %f1423, 0f3F000000, 0f3F000000;
fma.rn.f32 %f308, %f1424, 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 %f1425, %rs46;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1426, %rs47;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1427, %rs48;}
// end inline asm
add.f32 %f1428, %f1582, %f1425;
add.f32 %f1429, %f1583, %f1426;
add.f32 %f1430, %f1584, %f1427;
mov.f32 %f1431, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f1430;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f1429;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs42, %f1428;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f1431;}
// end inline asm
st.global.v4.u16 [%rd12], {%rs42, %rs43, %rs44, %rs45};
bra.uni $L__BB0_140;
$L__BB0_139:
mov.f32 %f1435, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs53, %f1435;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs52, %f1584;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f1583;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f1582;}
// 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 %f1436, %rs61;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1437, %rs62;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1438, %rs63;}
// end inline asm
add.f32 %f1439, %f300, %f1436;
add.f32 %f1440, %f301, %f1437;
add.f32 %f1441, %f302, %f1438;
mov.f32 %f1442, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs59, %f1441;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs58, %f1440;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs57, %f1439;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs60, %f1442;}
// end inline asm
st.global.v4.u16 [%rd13], {%rs57, %rs58, %rs59, %rs60};
bra.uni $L__BB0_143;
$L__BB0_142:
mov.f32 %f1446, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs68, %f1446;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs67, %f302;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs66, %f301;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs65, %f300;}
// 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 %f1447, %rs76;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1448, %rs77;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1449, %rs78;}
// end inline asm
add.f32 %f1450, %f303, %f1447;
add.f32 %f1451, %f304, %f1448;
add.f32 %f1452, %f305, %f1449;
mov.f32 %f1453, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs74, %f1452;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs73, %f1451;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs72, %f1450;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs75, %f1453;}
// end inline asm
st.global.v4.u16 [%rd14], {%rs72, %rs73, %rs74, %rs75};
bra.uni $L__BB0_146;
$L__BB0_145:
mov.f32 %f1457, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs83, %f1457;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs82, %f305;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs81, %f304;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs80, %f303;}
// 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 %f1458, %rs91;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1459, %rs92;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1460, %rs93;}
// end inline asm
add.f32 %f1461, %f306, %f1458;
add.f32 %f1462, %f307, %f1459;
add.f32 %f1463, %f308, %f1460;
mov.f32 %f1464, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs89, %f1463;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs88, %f1462;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs87, %f1461;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs90, %f1464;}
// end inline asm
st.global.v4.u16 [%rd15], {%rs87, %rs88, %rs89, %rs90};
bra.uni $L__BB0_169;
$L__BB0_148:
mov.f32 %f1468, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs98, %f1468;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs97, %f308;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs96, %f307;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs95, %f306;}
// end inline asm
st.global.v4.u16 [%rd15], {%rs95, %rs96, %rs97, %rs98};
$L__BB0_169:
ret;
}