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

2744 lines
76 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];
.global .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
.visible .entry __raygen__oxMain()
{
.local .align 4 .b8 __local_depot0[28];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<201>;
.reg .b16 %rs<177>;
.reg .f32 %f<1645>;
.reg .b32 %r<503>;
.reg .f64 %fd<5>;
.reg .b64 %rd<127>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r93), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r94), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u32 %r3, [params+616];
ld.const.u32 %r4, [params+540];
ld.const.u64 %rd34, [params+400];
cvta.to.global.u64 %rd35, %rd34;
ld.const.u32 %r99, [params+392];
mad.lo.s32 %r100, %r99, %r94, %r93;
mul.wide.u32 %rd36, %r100, 4;
add.s64 %rd2, %rd35, %rd36;
ld.global.v2.u8 {%rs7, %rs176}, [%rd2];
or.b16 %rs9, %rs7, %rs176;
and.b16 %rs10, %rs9, 255;
setp.eq.s16 %p10, %rs10, 0;
@%p10 bra $L__BB0_2;
ld.global.u8 %rs175, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs175, [%rd2+2];
setp.eq.s16 %p11, %rs175, 0;
mov.f32 %f1481, 0f00000000;
mov.u16 %rs176, 0;
mov.f32 %f1482, %f1481;
mov.f32 %f1483, %f1481;
@%p11 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f401, %rs7;
div.rn.f32 %f402, %f401, 0f437F0000;
fma.rn.f32 %f403, %f402, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs176, 255;
cvt.rn.f32.u16 %f404, %rs13;
div.rn.f32 %f405, %f404, 0f437F0000;
fma.rn.f32 %f406, %f405, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f407, %rs175;
div.rn.f32 %f408, %f407, 0f437F0000;
fma.rn.f32 %f409, %f408, 0f40000000, 0fBF800000;
mul.f32 %f410, %f406, %f406;
fma.rn.f32 %f411, %f403, %f403, %f410;
fma.rn.f32 %f412, %f409, %f409, %f411;
sqrt.rn.f32 %f413, %f412;
rcp.rn.f32 %f414, %f413;
mul.f32 %f1483, %f414, %f409;
mul.f32 %f1482, %f414, %f406;
mul.f32 %f1481, %f403, %f414;
$L__BB0_4:
ld.const.v2.u32 {%r101, %r102}, [params];
add.s32 %r5, %r101, %r93;
add.s32 %r6, %r102, %r94;
setp.eq.f32 %p12, %f1481, 0f00000000;
setp.eq.f32 %p13, %f1482, 0f00000000;
and.pred %p14, %p12, %p13;
setp.eq.f32 %p15, %f1483, 0f00000000;
and.pred %p16, %p15, %p14;
@%p16 bra $L__BB0_162;
bra.uni $L__BB0_5;
$L__BB0_162:
ld.const.u32 %r90, [params+104];
and.b32 %r465, %r90, 1;
setp.eq.b32 %p189, %r465, 1;
mov.pred %p190, 0;
xor.pred %p191, %p189, %p190;
not.pred %p192, %p191;
@%p192 bra $L__BB0_164;
ld.const.u64 %rd97, [params+144];
cvta.to.global.u64 %rd98, %rd97;
ld.const.u32 %r466, [params+136];
mad.lo.s32 %r467, %r466, %r6, %r5;
mul.wide.u32 %rd99, %r467, 4;
add.s64 %rd100, %rd98, %rd99;
mov.u16 %rs97, 0;
st.global.v4.u8 [%rd100], {%rs97, %rs97, %rs97, %rs97};
$L__BB0_164:
and.b32 %r468, %r90, 8;
setp.eq.s32 %p193, %r468, 0;
@%p193 bra $L__BB0_166;
ld.const.u64 %rd101, [params+192];
cvta.to.global.u64 %rd102, %rd101;
ld.const.u32 %r469, [params+184];
mad.lo.s32 %r470, %r469, %r6, %r5;
mov.f32 %f1427, 0f00000000;
cvt.rzi.u32.f32 %r471, %f1427;
mul.wide.u32 %rd103, %r470, 2;
add.s64 %rd104, %rd102, %rd103;
mov.u16 %rs98, 0;
cvt.u16.u32 %rs99, %r471;
st.global.v2.u8 [%rd104], {%rs99, %rs98};
$L__BB0_166:
and.b32 %r472, %r90, 4;
setp.eq.s32 %p194, %r472, 0;
ld.const.u32 %r502, [params+108];
@%p194 bra $L__BB0_170;
setp.eq.s32 %p195, %r502, 0;
ld.const.u64 %rd105, [params+224];
cvta.to.global.u64 %rd106, %rd105;
ld.const.u32 %r473, [params+216];
mad.lo.s32 %r474, %r473, %r6, %r5;
mul.wide.u32 %rd107, %r474, 8;
add.s64 %rd28, %rd106, %rd107;
@%p195 bra $L__BB0_169;
ld.global.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd28];
// begin inline asm
{ cvt.f32.f16 %f1428, %rs106;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1429, %rs107;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1430, %rs108;}
// end inline asm
add.f32 %f1431, %f1428, 0f00000000;
add.f32 %f1432, %f1429, 0f00000000;
add.f32 %f1433, %f1430, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs105, %f1433;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs104, %f1432;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs103, %f1431;}
// end inline asm
mov.u16 %rs110, 0;
st.global.v4.u16 [%rd28], {%rs103, %rs104, %rs105, %rs110};
bra.uni $L__BB0_170;
$L__BB0_5:
ld.const.u64 %rd37, [params+432];
cvta.to.global.u64 %rd38, %rd37;
ld.const.u32 %r106, [params+424];
mad.lo.s32 %r107, %r106, %r94, %r93;
mul.wide.u32 %rd39, %r107, 12;
add.s64 %rd40, %rd38, %rd39;
ld.global.f32 %f10, [%rd40];
mul.f32 %f431, %f10, 0f3456BF95;
ld.global.f32 %f11, [%rd40+4];
mul.f32 %f432, %f11, 0f3456BF95;
ld.global.f32 %f12, [%rd40+8];
mul.f32 %f433, %f12, 0f3456BF95;
abs.f32 %f13, %f1481;
div.rn.f32 %f434, %f431, %f13;
abs.f32 %f435, %f1482;
div.rn.f32 %f436, %f432, %f435;
abs.f32 %f14, %f1483;
div.rn.f32 %f437, %f433, %f14;
abs.f32 %f438, %f434;
abs.f32 %f439, %f436;
abs.f32 %f440, %f437;
mov.f32 %f441, 0f38D1B717;
max.f32 %f442, %f438, %f441;
max.f32 %f443, %f439, %f441;
max.f32 %f444, %f440, %f441;
fma.rn.f32 %f15, %f1481, %f442, %f10;
fma.rn.f32 %f16, %f1482, %f443, %f11;
fma.rn.f32 %f17, %f1483, %f444, %f12;
ld.const.u64 %rd41, [params+128];
cvta.to.global.u64 %rd42, %rd41;
ld.const.u32 %r108, [params+120];
mad.lo.s32 %r109, %r108, %r94, %r93;
mul.wide.u32 %rd43, %r109, 4;
add.s64 %rd44, %rd42, %rd43;
ld.global.u32 %r493, [%rd44];
mov.u32 %r489, 0;
abs.s32 %r110, %r3;
setp.lt.s32 %p18, %r110, 1;
mov.pred %p17, 0;
mov.f32 %f1503, 0f00000000;
mov.f32 %f1504, %f1503;
mov.f32 %f1505, %f1503;
mov.f32 %f1506, %f1503;
mov.f32 %f1507, %f1503;
mov.f32 %f1508, %f1503;
mov.f32 %f1509, %f1503;
mov.f32 %f1510, %f1503;
mov.f32 %f1511, %f1503;
mov.f32 %f1512, %f1503;
mov.f32 %f1513, %f1503;
mov.f32 %f1514, %f1503;
mov.f32 %f1515, %f1503;
mov.f32 %f1516, %f1503;
mov.f32 %f1517, %f1503;
mov.f32 %f1518, %f1503;
mov.pred %p200, %p17;
@%p18 bra $L__BB0_25;
ld.const.f32 %f18, [params+620];
ld.const.u64 %rd45, [params+624];
cvta.to.global.u64 %rd3, %rd45;
ld.const.u64 %rd46, [params+640];
cvta.to.global.u64 %rd47, %rd46;
ld.const.u32 %r113, [params+632];
and.b32 %r114, %r94, 255;
and.b32 %r115, %r93, 255;
mad.lo.s32 %r116, %r113, %r114, %r115;
mul.wide.u32 %rd48, %r116, 3;
add.s64 %rd4, %rd47, %rd48;
ld.const.f32 %f19, [params+660];
mul.f32 %f20, %f15, 0f3456BF95;
mul.f32 %f21, %f16, 0f3456BF95;
mul.f32 %f22, %f17, 0f3456BF95;
ld.const.u64 %rd5, [params+96];
abs.f32 %f588, %f21;
abs.f32 %f589, %f20;
max.f32 %f590, %f589, %f588;
abs.f32 %f591, %f22;
max.f32 %f592, %f590, %f591;
mov.u32 %r486, %r489;
$L__BB0_7:
shl.b32 %r117, %r486, 1;
mul.wide.s32 %rd49, %r117, 12;
add.s64 %rd6, %rd3, %rd49;
ld.global.f32 %f461, [%rd6];
sub.f32 %f462, %f461, %f10;
ld.global.f32 %f463, [%rd6+4];
sub.f32 %f464, %f463, %f11;
ld.global.f32 %f465, [%rd6+8];
sub.f32 %f466, %f465, %f12;
mul.f32 %f467, %f464, %f464;
fma.rn.f32 %f468, %f462, %f462, %f467;
fma.rn.f32 %f469, %f466, %f466, %f468;
sqrt.rn.f32 %f39, %f469;
rcp.rn.f32 %f470, %f39;
mul.f32 %f40, %f462, %f470;
mul.f32 %f41, %f464, %f470;
mul.f32 %f42, %f466, %f470;
mul.f32 %f471, %f1482, %f41;
fma.rn.f32 %f472, %f1481, %f40, %f471;
fma.rn.f32 %f43, %f1483, %f42, %f472;
setp.leu.f32 %p19, %f43, 0f00000000;
@%p19 bra $L__BB0_24;
setp.ne.s32 %p21, %r4, 0;
mul.f32 %f473, %f39, %f39;
div.rn.f32 %f44, %f18, %f473;
ld.global.u8 %rs15, [%rd4];
cvt.rn.f32.u16 %f474, %rs15;
div.rn.f32 %f475, %f474, 0fC37F0000;
fma.rn.f32 %f476, %f475, 0f3F333333, 0f3F800000;
setp.gt.f32 %p22, %f44, %f476;
and.pred %p23, %p21, %p22;
mov.pred %p200, -1;
@%p23 bra $L__BB0_25;
mul.f32 %f45, %f39, %f19;
mov.f32 %f481, 0f40800000;
abs.f32 %f47, %f45;
setp.lt.f32 %p24, %f47, 0f00800000;
mul.f32 %f483, %f47, 0f4B800000;
selp.f32 %f484, %f483, %f47, %p24;
selp.f32 %f485, 0fC3170000, 0fC2FE0000, %p24;
mov.b32 %r118, %f484;
and.b32 %r119, %r118, 8388607;
or.b32 %r120, %r119, 1065353216;
mov.b32 %f486, %r120;
shr.u32 %r121, %r118, 23;
cvt.rn.f32.u32 %f487, %r121;
add.f32 %f488, %f485, %f487;
setp.gt.f32 %p25, %f486, 0f3FB504F3;
mul.f32 %f489, %f486, 0f3F000000;
add.f32 %f490, %f488, 0f3F800000;
selp.f32 %f491, %f490, %f488, %p25;
selp.f32 %f492, %f489, %f486, %p25;
add.f32 %f493, %f492, 0fBF800000;
add.f32 %f494, %f492, 0f3F800000;
rcp.approx.ftz.f32 %f495, %f494;
add.f32 %f496, %f493, %f493;
mul.f32 %f497, %f496, %f495;
mul.f32 %f498, %f497, %f497;
mov.f32 %f499, 0f3C4CAF63;
mov.f32 %f500, 0f3B18F0FE;
fma.rn.f32 %f501, %f500, %f498, %f499;
mov.f32 %f502, 0f3DAAAABD;
fma.rn.f32 %f503, %f501, %f498, %f502;
mul.rn.f32 %f504, %f503, %f498;
mul.rn.f32 %f505, %f504, %f497;
sub.f32 %f506, %f493, %f497;
add.f32 %f507, %f506, %f506;
neg.f32 %f508, %f497;
fma.rn.f32 %f509, %f508, %f493, %f507;
mul.rn.f32 %f510, %f495, %f509;
add.f32 %f511, %f505, %f497;
sub.f32 %f512, %f497, %f511;
add.f32 %f513, %f505, %f512;
add.f32 %f514, %f510, %f513;
add.f32 %f515, %f511, %f514;
sub.f32 %f516, %f511, %f515;
add.f32 %f517, %f514, %f516;
mov.f32 %f518, 0f3F317200;
mul.rn.f32 %f519, %f491, %f518;
mov.f32 %f520, 0f35BFBE8E;
mul.rn.f32 %f521, %f491, %f520;
add.f32 %f522, %f519, %f515;
sub.f32 %f523, %f519, %f522;
add.f32 %f524, %f515, %f523;
add.f32 %f525, %f517, %f524;
add.f32 %f526, %f521, %f525;
add.f32 %f527, %f522, %f526;
sub.f32 %f528, %f522, %f527;
add.f32 %f529, %f526, %f528;
mul.rn.f32 %f530, %f481, %f527;
neg.f32 %f531, %f530;
fma.rn.f32 %f532, %f481, %f527, %f531;
fma.rn.f32 %f533, %f481, %f529, %f532;
mov.f32 %f534, 0f00000000;
fma.rn.f32 %f535, %f534, %f527, %f533;
add.rn.f32 %f536, %f530, %f535;
neg.f32 %f537, %f536;
add.rn.f32 %f538, %f530, %f537;
add.rn.f32 %f539, %f538, %f535;
mov.b32 %r122, %f536;
setp.eq.s32 %p26, %r122, 1118925336;
add.s32 %r123, %r122, -1;
mov.b32 %f540, %r123;
add.f32 %f541, %f539, 0f37000000;
selp.f32 %f48, %f541, %f539, %p26;
selp.f32 %f542, %f540, %f536, %p26;
mov.f32 %f543, 0f3FB8AA3B;
mul.rn.f32 %f544, %f542, %f543;
cvt.rzi.f32.f32 %f545, %f544;
abs.f32 %f546, %f545;
setp.gt.f32 %p27, %f546, 0f42FC0000;
mov.b32 %r124, %f545;
and.b32 %r125, %r124, -2147483648;
or.b32 %r126, %r125, 1123811328;
mov.b32 %f547, %r126;
selp.f32 %f548, %f547, %f545, %p27;
mov.f32 %f549, 0fBF317218;
fma.rn.f32 %f550, %f548, %f549, %f542;
mov.f32 %f551, 0f3102E308;
fma.rn.f32 %f552, %f548, %f551, %f550;
mul.f32 %f553, %f552, 0f3FB8AA3B;
add.f32 %f554, %f548, 0f4B40007F;
mov.b32 %r127, %f554;
shl.b32 %r128, %r127, 23;
mov.b32 %f555, %r128;
ex2.approx.ftz.f32 %f556, %f553;
mul.f32 %f49, %f556, %f555;
setp.eq.f32 %p28, %f49, 0f7F800000;
mov.f32 %f1500, 0f7F800000;
@%p28 bra $L__BB0_11;
fma.rn.f32 %f1500, %f49, %f48, %f49;
$L__BB0_11:
mov.f32 %f1478, 0f40000000;
cvt.rzi.f32.f32 %f1477, %f1478;
add.f32 %f1476, %f1477, %f1477;
mov.f32 %f1475, 0f40800000;
sub.f32 %f1474, %f1475, %f1476;
abs.f32 %f1473, %f1474;
setp.lt.f32 %p29, %f45, 0f00000000;
setp.eq.f32 %p30, %f1473, 0f3F800000;
and.pred %p1, %p29, %p30;
setp.eq.f32 %p31, %f45, 0f00000000;
@%p31 bra $L__BB0_15;
bra.uni $L__BB0_12;
$L__BB0_15:
add.f32 %f561, %f45, %f45;
selp.f32 %f1502, %f561, 0f00000000, %p30;
bra.uni $L__BB0_16;
$L__BB0_12:
mov.b32 %r129, %f1500;
xor.b32 %r130, %r129, -2147483648;
mov.b32 %f557, %r130;
selp.f32 %f1502, %f557, %f1500, %p1;
setp.geu.f32 %p32, %f45, 0f00000000;
@%p32 bra $L__BB0_16;
mov.f32 %f558, 0f40800000;
cvt.rzi.f32.f32 %f559, %f558;
setp.eq.f32 %p33, %f559, 0f40800000;
@%p33 bra $L__BB0_16;
mov.f32 %f1502, 0f7FFFFFFF;
$L__BB0_16:
add.f32 %f562, %f47, 0f40800000;
mov.b32 %r131, %f562;
setp.lt.s32 %p35, %r131, 2139095040;
@%p35 bra $L__BB0_21;
setp.gtu.f32 %p36, %f47, 0f7F800000;
@%p36 bra $L__BB0_20;
bra.uni $L__BB0_18;
$L__BB0_20:
add.f32 %f1502, %f45, 0f40800000;
bra.uni $L__BB0_21;
$L__BB0_18:
setp.neu.f32 %p37, %f47, 0f7F800000;
@%p37 bra $L__BB0_21;
selp.f32 %f1502, 0fFF800000, 0f7F800000, %p1;
$L__BB0_21:
shl.b32 %r484, %r486, 1;
mul.wide.s32 %rd121, %r484, 12;
add.s64 %rd120, %rd3, %rd121;
mov.f32 %f563, 0f3F800000;
sub.f32 %f564, %f563, %f1502;
setp.eq.f32 %p38, %f45, 0f3F800000;
selp.f32 %f565, 0f00000000, %f564, %p38;
cvt.sat.f32.f32 %f566, %f565;
mul.f32 %f567, %f44, %f566;
ld.global.f32 %f568, [%rd120+12];
mul.f32 %f569, %f40, %f568;
ld.global.f32 %f570, [%rd120+16];
mul.f32 %f571, %f41, %f570;
neg.f32 %f572, %f571;
sub.f32 %f573, %f572, %f569;
ld.global.f32 %f574, [%rd120+20];
mul.f32 %f575, %f42, %f574;
sub.f32 %f576, %f573, %f575;
cvt.sat.f32.f32 %f577, %f576;
mul.f32 %f58, %f567, %f577;
setp.leu.f32 %p39, %f58, 0f3727C5AC;
@%p39 bra $L__BB0_23;
cvt.sat.f32.f32 %f587, %f43;
mov.f32 %f593, 0f38D1B717;
max.f32 %f584, %f592, %f593;
sub.f32 %f585, %f39, %f584;
mov.f32 %f586, 0f00000000;
mov.u32 %r168, 2;
mov.u32 %r169, 1;
mov.u32 %r170, 3;
mov.u32 %r173, 1065353216;
mov.u32 %r202, 0;
// begin inline asm
call(%r132,%r133,%r134,%r135,%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163),_optix_trace_typed_32,(%r202,%rd5,%f15,%f16,%f17,%f40,%f41,%f42,%f584,%f585,%f586,%r169,%r202,%r169,%r168,%r169,%r170,%r173,%r173,%r173,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202);
// end inline asm
mov.b32 %f594, %r132;
mov.b32 %f595, %r133;
mov.b32 %f596, %r134;
mul.f32 %f597, %f58, 0f3EA2F983;
mul.f32 %f598, %f597, %f594;
mul.f32 %f599, %f597, %f595;
mul.f32 %f600, %f597, %f596;
fma.rn.f32 %f1515, %f587, %f598, %f1515;
fma.rn.f32 %f1516, %f587, %f599, %f1516;
fma.rn.f32 %f1517, %f587, %f600, %f1517;
mul.f32 %f601, %f43, 0f40800000;
cvt.sat.f32.f32 %f602, %f601;
mul.f32 %f603, %f602, %f598;
mul.f32 %f604, %f602, %f599;
mul.f32 %f605, %f602, %f600;
fma.rn.f32 %f1509, %f40, %f603, %f1509;
fma.rn.f32 %f1510, %f40, %f604, %f1510;
fma.rn.f32 %f1511, %f40, %f605, %f1511;
fma.rn.f32 %f1506, %f41, %f603, %f1506;
fma.rn.f32 %f1507, %f41, %f604, %f1507;
fma.rn.f32 %f1508, %f41, %f605, %f1508;
fma.rn.f32 %f1503, %f42, %f603, %f1503;
fma.rn.f32 %f1504, %f42, %f604, %f1504;
fma.rn.f32 %f1505, %f42, %f605, %f1505;
add.f32 %f1512, %f1512, %f603;
add.f32 %f1513, %f1513, %f604;
add.f32 %f1514, %f1514, %f605;
add.f32 %f1518, %f1518, %f594;
$L__BB0_23:
add.s32 %r489, %r489, 1;
$L__BB0_24:
add.s32 %r486, %r486, 1;
setp.lt.s32 %p41, %r486, %r110;
mov.pred %p200, %p17;
@%p41 bra $L__BB0_7;
$L__BB0_25:
cvt.rn.f32.s32 %f606, %r489;
mov.f32 %f607, 0f3F800000;
max.f32 %f608, %f606, %f607;
rcp.rn.f32 %f609, %f608;
mul.f32 %f1620, %f1515, %f609;
mul.f32 %f1621, %f1516, %f609;
mul.f32 %f1622, %f1517, %f609;
div.rn.f32 %f1623, %f1518, %f608;
mul.f32 %f1617, %f1512, %f609;
mul.f32 %f1618, %f1513, %f609;
mul.f32 %f1619, %f1514, %f609;
mul.f32 %f1614, %f1509, %f609;
mul.f32 %f1615, %f1510, %f609;
mul.f32 %f1616, %f1511, %f609;
mul.f32 %f1611, %f1506, %f609;
mul.f32 %f1612, %f1507, %f609;
mul.f32 %f1613, %f1508, %f609;
mul.f32 %f1608, %f1503, %f609;
mul.f32 %f1609, %f1504, %f609;
mul.f32 %f1610, %f1505, %f609;
not.pred %p42, %p200;
@%p42 bra $L__BB0_58;
abs.f32 %f1480, %f1483;
abs.f32 %f1479, %f1481;
setp.gt.f32 %p43, %f1479, %f1480;
selp.f32 %f626, 0f00000000, %f1482, %p43;
mov.f32 %f1582, 0f00000000;
neg.f32 %f627, %f1483;
selp.f32 %f628, %f1481, %f627, %p43;
neg.f32 %f629, %f1482;
selp.f32 %f630, %f629, 0f00000000, %p43;
mul.f32 %f631, %f628, %f628;
fma.rn.f32 %f632, %f630, %f630, %f631;
fma.rn.f32 %f633, %f626, %f626, %f632;
sqrt.rn.f32 %f634, %f633;
rcp.rn.f32 %f635, %f634;
mul.f32 %f139, %f630, %f635;
mul.f32 %f140, %f628, %f635;
mul.f32 %f141, %f626, %f635;
setp.lt.s32 %p44, %r4, 1;
mov.f32 %f1581, %f1582;
mov.f32 %f1580, %f1582;
mov.f32 %f1579, %f1582;
mov.f32 %f1578, %f1582;
mov.f32 %f1577, %f1582;
mov.f32 %f1576, %f1582;
mov.f32 %f1575, %f1582;
mov.f32 %f1574, %f1582;
mov.f32 %f1573, %f1582;
mov.f32 %f1572, %f1582;
mov.f32 %f1571, %f1582;
mov.f32 %f1570, %f1582;
mov.f32 %f1569, %f1582;
mov.f32 %f1568, %f1582;
mov.f32 %f1567, %f1582;
@%p44 bra $L__BB0_57;
cvt.rn.f32.s32 %f652, %r4;
rcp.rn.f32 %f142, %f652;
mul.f32 %f143, %f15, 0f3456BF95;
mul.f32 %f144, %f16, 0f3456BF95;
mul.f32 %f145, %f17, 0f3456BF95;
ld.const.u64 %rd7, [params+96];
mul.f32 %f653, %f1481, %f140;
mul.f32 %f654, %f1482, %f139;
sub.f32 %f146, %f654, %f653;
mul.f32 %f655, %f1483, %f139;
mul.f32 %f656, %f1481, %f141;
sub.f32 %f147, %f656, %f655;
mul.f32 %f657, %f1482, %f141;
mul.f32 %f658, %f1483, %f140;
sub.f32 %f148, %f658, %f657;
mov.u32 %r204, 0;
add.s64 %rd8, %rd1, 24;
mov.u64 %rd51, __cudart_i2opi_f;
abs.f32 %f724, %f144;
abs.f32 %f725, %f143;
max.f32 %f726, %f725, %f724;
abs.f32 %f727, %f145;
max.f32 %f728, %f726, %f727;
mov.u32 %r490, %r204;
$L__BB0_28:
cvt.rn.f32.s32 %f165, %r490;
mov.u32 %r492, %r204;
$L__BB0_29:
mad.lo.s32 %r206, %r493, 1664525, 1013904223;
and.b32 %r207, %r206, 16777215;
cvt.rn.f32.u32 %f659, %r207;
fma.rn.f32 %f660, %f659, 0f33800000, %f165;
mul.f32 %f182, %f142, %f660;
mad.lo.s32 %r493, %r206, 1664525, 1013904223;
and.b32 %r208, %r493, 16777215;
cvt.rn.f32.u32 %f661, %r208;
cvt.rn.f32.s32 %f662, %r492;
fma.rn.f32 %f663, %f661, 0f33800000, %f662;
mul.f32 %f664, %f142, %f663;
mul.f32 %f665, %f182, %f182;
mov.f32 %f666, 0f3F800000;
sub.f32 %f667, %f666, %f665;
mov.f32 %f668, 0f00000000;
max.f32 %f669, %f668, %f667;
sqrt.rn.f32 %f183, %f669;
mul.f32 %f184, %f664, 0f40C90FDB;
mul.f32 %f670, %f184, 0f3F22F983;
cvt.rni.s32.f32 %r500, %f670;
cvt.rn.f32.s32 %f671, %r500;
mov.f32 %f672, 0fBFC90FDA;
fma.rn.f32 %f673, %f671, %f672, %f184;
mov.f32 %f674, 0fB3A22168;
fma.rn.f32 %f675, %f671, %f674, %f673;
mov.f32 %f676, 0fA7C234C5;
fma.rn.f32 %f1586, %f671, %f676, %f675;
abs.f32 %f186, %f184;
setp.ltu.f32 %p45, %f186, 0f47CE4780;
mov.u32 %r497, %r500;
mov.f32 %f1583, %f1586;
@%p45 bra $L__BB0_37;
setp.eq.f32 %p46, %f186, 0f7F800000;
@%p46 bra $L__BB0_36;
bra.uni $L__BB0_31;
$L__BB0_36:
mov.f32 %f679, 0f00000000;
mul.rn.f32 %f1583, %f184, %f679;
mov.u32 %r497, 0;
bra.uni $L__BB0_37;
$L__BB0_31:
mov.b32 %r20, %f184;
bfe.u32 %r210, %r20, 23, 8;
add.s32 %r21, %r210, -128;
shl.b32 %r211, %r20, 8;
or.b32 %r22, %r211, -2147483648;
shr.u32 %r23, %r21, 5;
mov.u64 %rd124, 0;
mov.u32 %r494, 0;
mov.u64 %rd122, %rd1;
mov.u64 %rd123, %rd51;
$L__BB0_32:
.pragma "nounroll";
ld.global.nc.u32 %r212, [%rd123];
mad.wide.u32 %rd53, %r212, %r22, %rd124;
shr.u64 %rd124, %rd53, 32;
st.local.u32 [%rd122], %rd53;
add.s64 %rd123, %rd123, 4;
add.s64 %rd122, %rd122, 4;
add.s32 %r494, %r494, 1;
setp.ne.s32 %p47, %r494, 6;
@%p47 bra $L__BB0_32;
st.local.u32 [%rd8], %rd124;
mov.u32 %r213, 4;
sub.s32 %r26, %r213, %r23;
mov.u32 %r214, 6;
sub.s32 %r215, %r214, %r23;
mul.wide.s32 %rd54, %r215, 4;
add.s64 %rd55, %rd1, %rd54;
ld.local.u32 %r495, [%rd55];
ld.local.u32 %r496, [%rd55+-4];
and.b32 %r29, %r21, 31;
setp.eq.s32 %p48, %r29, 0;
@%p48 bra $L__BB0_35;
mov.u32 %r216, 32;
sub.s32 %r217, %r216, %r29;
shr.u32 %r218, %r496, %r217;
shl.b32 %r219, %r495, %r29;
add.s32 %r495, %r218, %r219;
mul.wide.s32 %rd56, %r26, 4;
add.s64 %rd57, %rd1, %rd56;
ld.local.u32 %r220, [%rd57];
shr.u32 %r221, %r220, %r217;
shl.b32 %r222, %r496, %r29;
add.s32 %r496, %r221, %r222;
$L__BB0_35:
and.b32 %r223, %r20, -2147483648;
shr.u32 %r224, %r496, 30;
shl.b32 %r225, %r495, 2;
or.b32 %r226, %r224, %r225;
shr.u32 %r227, %r226, 31;
shr.u32 %r228, %r495, 30;
add.s32 %r229, %r227, %r228;
neg.s32 %r230, %r229;
setp.eq.s32 %p49, %r223, 0;
selp.b32 %r497, %r229, %r230, %p49;
setp.ne.s32 %p50, %r227, 0;
xor.b32 %r231, %r223, -2147483648;
selp.b32 %r232, %r231, %r223, %p50;
selp.b32 %r233, -1, 0, %p50;
xor.b32 %r234, %r226, %r233;
shl.b32 %r235, %r496, 2;
xor.b32 %r236, %r235, %r233;
cvt.u64.u32 %rd58, %r234;
cvt.u64.u32 %rd59, %r236;
bfi.b64 %rd60, %rd58, %rd59, 32, 32;
cvt.rn.f64.s64 %fd1, %rd60;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f677, %fd2;
setp.eq.s32 %p51, %r232, 0;
neg.f32 %f678, %f677;
selp.f32 %f1583, %f677, %f678, %p51;
$L__BB0_37:
add.s32 %r36, %r497, 1;
and.b32 %r37, %r36, 1;
setp.eq.s32 %p52, %r37, 0;
selp.f32 %f190, %f1583, 0f3F800000, %p52;
mul.rn.f32 %f191, %f1583, %f1583;
mov.f32 %f1584, 0fB94D4153;
@%p52 bra $L__BB0_39;
mov.f32 %f681, 0fBAB607ED;
mov.f32 %f682, 0f37CBAC00;
fma.rn.f32 %f1584, %f682, %f191, %f681;
$L__BB0_39:
selp.f32 %f683, 0f3C0885E4, 0f3D2AAABB, %p52;
fma.rn.f32 %f684, %f1584, %f191, %f683;
selp.f32 %f685, 0fBE2AAAA8, 0fBEFFFFFF, %p52;
fma.rn.f32 %f686, %f684, %f191, %f685;
mov.f32 %f687, 0f00000000;
fma.rn.f32 %f688, %f191, %f190, %f687;
fma.rn.f32 %f1585, %f686, %f688, %f190;
and.b32 %r238, %r36, 2;
setp.eq.s32 %p54, %r238, 0;
@%p54 bra $L__BB0_41;
mov.f32 %f690, 0fBF800000;
fma.rn.f32 %f1585, %f1585, %f690, %f687;
$L__BB0_41:
@%p45 bra $L__BB0_49;
setp.eq.f32 %p56, %f186, 0f7F800000;
@%p56 bra $L__BB0_48;
bra.uni $L__BB0_43;
$L__BB0_48:
mov.f32 %f693, 0f00000000;
mul.rn.f32 %f1586, %f184, %f693;
mov.u32 %r500, 0;
bra.uni $L__BB0_49;
$L__BB0_43:
mov.b32 %r38, %f184;
bfe.u32 %r239, %r38, 23, 8;
add.s32 %r39, %r239, -128;
shl.b32 %r240, %r38, 8;
or.b32 %r40, %r240, -2147483648;
shr.u32 %r41, %r39, 5;
mov.u64 %rd125, 0;
mov.u64 %rd126, %rd125;
$L__BB0_44:
.pragma "nounroll";
shl.b64 %rd63, %rd125, 2;
mov.u64 %rd64, __cudart_i2opi_f;
add.s64 %rd65, %rd64, %rd63;
ld.global.nc.u32 %r241, [%rd65];
mad.wide.u32 %rd66, %r241, %r40, %rd126;
shr.u64 %rd126, %rd66, 32;
add.s64 %rd67, %rd1, %rd63;
st.local.u32 [%rd67], %rd66;
cvt.u32.u64 %r242, %rd125;
add.s32 %r243, %r242, 1;
cvt.s64.s32 %rd125, %r243;
setp.ne.s32 %p57, %r243, 6;
@%p57 bra $L__BB0_44;
st.local.u32 [%rd8], %rd126;
mov.u32 %r244, 4;
sub.s32 %r42, %r244, %r41;
mov.u32 %r245, 6;
sub.s32 %r246, %r245, %r41;
mul.wide.s32 %rd68, %r246, 4;
add.s64 %rd69, %rd1, %rd68;
ld.local.u32 %r498, [%rd69];
ld.local.u32 %r499, [%rd69+-4];
and.b32 %r45, %r39, 31;
setp.eq.s32 %p58, %r45, 0;
@%p58 bra $L__BB0_47;
mov.u32 %r247, 32;
sub.s32 %r248, %r247, %r45;
shr.u32 %r249, %r499, %r248;
shl.b32 %r250, %r498, %r45;
add.s32 %r498, %r249, %r250;
mul.wide.s32 %rd70, %r42, 4;
add.s64 %rd71, %rd1, %rd70;
ld.local.u32 %r251, [%rd71];
shr.u32 %r252, %r251, %r248;
shl.b32 %r253, %r499, %r45;
add.s32 %r499, %r252, %r253;
$L__BB0_47:
and.b32 %r254, %r38, -2147483648;
shr.u32 %r255, %r499, 30;
shl.b32 %r256, %r498, 2;
or.b32 %r257, %r255, %r256;
shr.u32 %r258, %r257, 31;
shr.u32 %r259, %r498, 30;
add.s32 %r260, %r258, %r259;
neg.s32 %r261, %r260;
setp.eq.s32 %p59, %r254, 0;
selp.b32 %r500, %r260, %r261, %p59;
setp.ne.s32 %p60, %r258, 0;
xor.b32 %r262, %r254, -2147483648;
selp.b32 %r263, %r262, %r254, %p60;
selp.b32 %r264, -1, 0, %p60;
xor.b32 %r265, %r257, %r264;
shl.b32 %r266, %r499, 2;
xor.b32 %r267, %r266, %r264;
cvt.u64.u32 %rd72, %r265;
cvt.u64.u32 %rd73, %r267;
bfi.b64 %rd74, %rd72, %rd73, 32, 32;
cvt.rn.f64.s64 %fd3, %rd74;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f691, %fd4;
setp.eq.s32 %p61, %r263, 0;
neg.f32 %f692, %f691;
selp.f32 %f1586, %f691, %f692, %p61;
$L__BB0_49:
mul.f32 %f200, %f183, %f1585;
and.b32 %r52, %r500, 1;
setp.eq.s32 %p62, %r52, 0;
selp.f32 %f201, %f1586, 0f3F800000, %p62;
mul.rn.f32 %f202, %f1586, %f1586;
mov.f32 %f1587, 0fB94D4153;
@%p62 bra $L__BB0_51;
mov.f32 %f695, 0fBAB607ED;
mov.f32 %f696, 0f37CBAC00;
fma.rn.f32 %f1587, %f696, %f202, %f695;
$L__BB0_51:
selp.f32 %f697, 0f3C0885E4, 0f3D2AAABB, %p62;
fma.rn.f32 %f698, %f1587, %f202, %f697;
selp.f32 %f699, 0fBE2AAAA8, 0fBEFFFFFF, %p62;
fma.rn.f32 %f700, %f698, %f202, %f699;
mov.f32 %f1589, 0f00000000;
fma.rn.f32 %f702, %f202, %f201, %f1589;
fma.rn.f32 %f1588, %f700, %f702, %f201;
and.b32 %r269, %r500, 2;
setp.eq.s32 %p64, %r269, 0;
@%p64 bra $L__BB0_53;
mov.f32 %f703, 0f00000000;
mov.f32 %f704, 0fBF800000;
fma.rn.f32 %f1588, %f1588, %f704, %f703;
$L__BB0_53:
mul.f32 %f717, %f183, %f1588;
mul.f32 %f718, %f139, %f717;
mul.f32 %f719, %f140, %f717;
mul.f32 %f720, %f141, %f717;
fma.rn.f32 %f721, %f148, %f200, %f718;
fma.rn.f32 %f722, %f147, %f200, %f719;
fma.rn.f32 %f723, %f146, %f200, %f720;
fma.rn.f32 %f208, %f1481, %f182, %f721;
fma.rn.f32 %f209, %f1482, %f182, %f722;
fma.rn.f32 %f210, %f1483, %f182, %f723;
mov.f32 %f729, 0f38D1B717;
max.f32 %f711, %f728, %f729;
mov.f32 %f712, 0f6C4ECB8F;
mov.u32 %r303, 1;
mov.u32 %r306, 2;
mov.u32 %r308, 4;
mov.u32 %r311, 1065353216;
mov.u32 %r312, 2139095039;
mov.u32 %r340, 0;
// begin inline asm
call(%r270,%r271,%r272,%r273,%r274,%r275,%r276,%r277,%r278,%r279,%r280,%r281,%r282,%r283,%r284,%r285,%r286,%r287,%r288,%r289,%r290,%r291,%r292,%r293,%r294,%r295,%r296,%r297,%r298,%r299,%r300,%r301),_optix_trace_typed_32,(%r340,%rd7,%f15,%f16,%f17,%f208,%f209,%f210,%f711,%f712,%f1589,%r303,%r340,%r340,%r306,%r340,%r308,%r311,%r311,%r311,%r312,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340,%r340);
// end inline asm
mov.b32 %f211, %r273;
setp.ge.f32 %p65, %f211, 0f00000000;
mov.f32 %f1590, %f1589;
mov.f32 %f1591, %f1589;
@%p65 bra $L__BB0_55;
mov.b32 %f1591, %r270;
mov.b32 %f1590, %r271;
mov.b32 %f1589, %r272;
$L__BB0_55:
setp.ltu.f32 %p66, %f211, 0f00000000;
selp.f32 %f730, 0f3F800000, 0f00000000, %p66;
add.f32 %f1567, %f1567, %f730;
fma.rn.f32 %f1576, %f208, %f1591, %f1576;
fma.rn.f32 %f1575, %f208, %f1590, %f1575;
fma.rn.f32 %f1574, %f208, %f1589, %f1574;
fma.rn.f32 %f1579, %f209, %f1591, %f1579;
fma.rn.f32 %f1578, %f209, %f1590, %f1578;
fma.rn.f32 %f1577, %f209, %f1589, %f1577;
fma.rn.f32 %f1582, %f210, %f1591, %f1582;
fma.rn.f32 %f1581, %f210, %f1590, %f1581;
fma.rn.f32 %f1580, %f210, %f1589, %f1580;
add.f32 %f1573, %f1573, %f1591;
add.f32 %f1572, %f1572, %f1590;
add.f32 %f1571, %f1571, %f1589;
mul.f32 %f731, %f1482, %f209;
fma.rn.f32 %f732, %f1481, %f208, %f731;
fma.rn.f32 %f733, %f1483, %f210, %f732;
cvt.sat.f32.f32 %f734, %f733;
fma.rn.f32 %f1570, %f1591, %f734, %f1570;
fma.rn.f32 %f1569, %f1590, %f734, %f1569;
fma.rn.f32 %f1568, %f1589, %f734, %f1568;
add.s32 %r492, %r492, 1;
setp.lt.s32 %p67, %r492, %r4;
@%p67 bra $L__BB0_29;
add.s32 %r490, %r490, 1;
setp.lt.s32 %p68, %r490, %r4;
@%p68 bra $L__BB0_28;
$L__BB0_57:
mul.lo.s32 %r341, %r4, %r4;
cvt.rn.f32.s32 %f735, %r341;
rcp.rn.f32 %f736, %f735;
mul.f32 %f737, %f736, %f1570;
mul.f32 %f738, %f736, %f1569;
mul.f32 %f739, %f736, %f1568;
div.rn.f32 %f1623, %f1567, %f735;
mul.f32 %f1617, %f736, %f1573;
mul.f32 %f1618, %f736, %f1572;
mul.f32 %f1619, %f736, %f1571;
mul.f32 %f1614, %f736, %f1576;
mul.f32 %f1615, %f736, %f1575;
mul.f32 %f1616, %f736, %f1574;
mul.f32 %f1611, %f736, %f1579;
mul.f32 %f1612, %f736, %f1578;
mul.f32 %f1613, %f736, %f1577;
mul.f32 %f1608, %f736, %f1582;
mul.f32 %f1609, %f736, %f1581;
mul.f32 %f1610, %f736, %f1580;
fma.rn.f32 %f1620, %f736, %f1570, %f737;
fma.rn.f32 %f1621, %f736, %f1569, %f738;
fma.rn.f32 %f1622, %f736, %f1568, %f739;
$L__BB0_58:
ld.const.u32 %r485, [params+616];
setp.lt.s32 %p69, %r485, 0;
selp.f32 %f282, %f1620, %f1623, %p69;
ld.const.u32 %r87, [params+104];
and.b32 %r342, %r87, 8;
setp.eq.s32 %p70, %r342, 0;
@%p70 bra $L__BB0_72;
ld.const.u64 %rd76, [params+192];
cvta.to.global.u64 %rd19, %rd76;
ld.const.u32 %r343, [params+184];
mad.lo.s32 %r344, %r343, %r6, %r5;
cvt.u64.u32 %rd20, %r344;
mov.f32 %f741, 0f3E68BA2E;
cvt.rzi.f32.f32 %f742, %f741;
add.f32 %f743, %f742, %f742;
mov.f32 %f744, 0f3EE8BA2E;
sub.f32 %f745, %f744, %f743;
abs.f32 %f283, %f745;
abs.f32 %f284, %f282;
setp.lt.f32 %p71, %f284, 0f00800000;
mul.f32 %f746, %f284, 0f4B800000;
selp.f32 %f747, %f746, %f284, %p71;
selp.f32 %f748, 0fC3170000, 0fC2FE0000, %p71;
mov.b32 %r345, %f747;
and.b32 %r346, %r345, 8388607;
or.b32 %r347, %r346, 1065353216;
mov.b32 %f749, %r347;
shr.u32 %r348, %r345, 23;
cvt.rn.f32.u32 %f750, %r348;
add.f32 %f751, %f748, %f750;
setp.gt.f32 %p72, %f749, 0f3FB504F3;
mul.f32 %f752, %f749, 0f3F000000;
add.f32 %f753, %f751, 0f3F800000;
selp.f32 %f754, %f753, %f751, %p72;
selp.f32 %f755, %f752, %f749, %p72;
add.f32 %f756, %f755, 0fBF800000;
add.f32 %f757, %f755, 0f3F800000;
rcp.approx.ftz.f32 %f758, %f757;
add.f32 %f759, %f756, %f756;
mul.f32 %f760, %f759, %f758;
mul.f32 %f761, %f760, %f760;
mov.f32 %f762, 0f3C4CAF63;
mov.f32 %f763, 0f3B18F0FE;
fma.rn.f32 %f764, %f763, %f761, %f762;
mov.f32 %f765, 0f3DAAAABD;
fma.rn.f32 %f766, %f764, %f761, %f765;
mul.rn.f32 %f767, %f766, %f761;
mul.rn.f32 %f768, %f767, %f760;
sub.f32 %f769, %f756, %f760;
add.f32 %f770, %f769, %f769;
neg.f32 %f771, %f760;
fma.rn.f32 %f772, %f771, %f756, %f770;
mul.rn.f32 %f773, %f758, %f772;
add.f32 %f774, %f768, %f760;
sub.f32 %f775, %f760, %f774;
add.f32 %f776, %f768, %f775;
add.f32 %f777, %f773, %f776;
add.f32 %f778, %f774, %f777;
sub.f32 %f779, %f774, %f778;
add.f32 %f780, %f777, %f779;
mov.f32 %f781, 0f3F317200;
mul.rn.f32 %f782, %f754, %f781;
mov.f32 %f783, 0f35BFBE8E;
mul.rn.f32 %f784, %f754, %f783;
add.f32 %f785, %f782, %f778;
sub.f32 %f786, %f782, %f785;
add.f32 %f787, %f778, %f786;
add.f32 %f788, %f780, %f787;
add.f32 %f789, %f784, %f788;
add.f32 %f790, %f785, %f789;
sub.f32 %f791, %f785, %f790;
add.f32 %f792, %f789, %f791;
mul.rn.f32 %f793, %f744, %f790;
neg.f32 %f794, %f793;
fma.rn.f32 %f795, %f744, %f790, %f794;
fma.rn.f32 %f796, %f744, %f792, %f795;
mov.f32 %f797, 0f00000000;
fma.rn.f32 %f798, %f797, %f790, %f796;
add.rn.f32 %f799, %f793, %f798;
neg.f32 %f800, %f799;
add.rn.f32 %f801, %f793, %f800;
add.rn.f32 %f802, %f801, %f798;
mov.b32 %r349, %f799;
setp.eq.s32 %p73, %r349, 1118925336;
add.s32 %r350, %r349, -1;
mov.b32 %f803, %r350;
add.f32 %f804, %f802, 0f37000000;
selp.f32 %f285, %f804, %f802, %p73;
selp.f32 %f805, %f803, %f799, %p73;
mov.f32 %f806, 0f3FB8AA3B;
mul.rn.f32 %f807, %f805, %f806;
cvt.rzi.f32.f32 %f808, %f807;
abs.f32 %f809, %f808;
setp.gt.f32 %p74, %f809, 0f42FC0000;
mov.b32 %r351, %f808;
and.b32 %r352, %r351, -2147483648;
or.b32 %r353, %r352, 1123811328;
mov.b32 %f810, %r353;
selp.f32 %f811, %f810, %f808, %p74;
mov.f32 %f812, 0fBF317218;
fma.rn.f32 %f813, %f811, %f812, %f805;
mov.f32 %f814, 0f3102E308;
fma.rn.f32 %f815, %f811, %f814, %f813;
mul.f32 %f816, %f815, 0f3FB8AA3B;
add.f32 %f817, %f811, 0f4B40007F;
mov.b32 %r354, %f817;
shl.b32 %r355, %r354, 23;
mov.b32 %f818, %r355;
ex2.approx.ftz.f32 %f819, %f816;
mul.f32 %f286, %f819, %f818;
setp.eq.f32 %p75, %f286, 0f7F800000;
mov.f32 %f1624, 0f7F800000;
@%p75 bra $L__BB0_61;
fma.rn.f32 %f1624, %f286, %f285, %f286;
$L__BB0_61:
setp.lt.f32 %p76, %f282, 0f00000000;
setp.eq.f32 %p77, %f283, 0f3F800000;
and.pred %p3, %p76, %p77;
setp.eq.f32 %p78, %f282, 0f00000000;
@%p78 bra $L__BB0_65;
bra.uni $L__BB0_62;
$L__BB0_65:
add.f32 %f824, %f282, %f282;
selp.f32 %f1626, %f824, 0f00000000, %p77;
bra.uni $L__BB0_66;
$L__BB0_169:
mov.f32 %f1436, 0f00000000;
mov.u32 %r502, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f1436;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs112, %f1436;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs111, %f1436;}
// end inline asm
mov.u16 %rs114, 0;
st.global.v4.u16 [%rd28], {%rs111, %rs112, %rs113, %rs114};
$L__BB0_170:
ld.const.u64 %rd108, [params+256];
cvta.to.global.u64 %rd109, %rd108;
ld.const.u32 %r476, [params+248];
mad.lo.s32 %r477, %r476, %r6, %r5;
mul.wide.u32 %rd110, %r477, 8;
add.s64 %rd29, %rd109, %rd110;
setp.eq.s32 %p196, %r502, 0;
@%p196 bra $L__BB0_172;
ld.global.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd29];
// begin inline asm
{ cvt.f32.f16 %f1437, %rs121;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1438, %rs122;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1439, %rs123;}
// end inline asm
add.f32 %f1440, %f1437, 0f00000000;
add.f32 %f1441, %f1438, 0f00000000;
add.f32 %f1442, %f1439, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs120, %f1442;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs119, %f1441;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs118, %f1440;}
// end inline asm
mov.u16 %rs125, 0;
st.global.v4.u16 [%rd29], {%rs118, %rs119, %rs120, %rs125};
bra.uni $L__BB0_173;
$L__BB0_172:
mov.f32 %f1445, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1445;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs127, %f1445;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs126, %f1445;}
// end inline asm
mov.u16 %rs129, 0;
st.global.v4.u16 [%rd29], {%rs126, %rs127, %rs128, %rs129};
$L__BB0_173:
ld.const.u64 %rd111, [params+272];
cvta.to.global.u64 %rd112, %rd111;
ld.const.u32 %r478, [params+264];
mad.lo.s32 %r479, %r478, %r6, %r5;
mul.wide.u32 %rd113, %r479, 8;
add.s64 %rd30, %rd112, %rd113;
@%p196 bra $L__BB0_175;
ld.global.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd30];
// begin inline asm
{ cvt.f32.f16 %f1446, %rs136;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1447, %rs137;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1448, %rs138;}
// end inline asm
add.f32 %f1449, %f1446, 0f00000000;
add.f32 %f1450, %f1447, 0f00000000;
add.f32 %f1451, %f1448, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs135, %f1451;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs134, %f1450;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs133, %f1449;}
// end inline asm
mov.u16 %rs140, 0;
st.global.v4.u16 [%rd30], {%rs133, %rs134, %rs135, %rs140};
bra.uni $L__BB0_176;
$L__BB0_175:
mov.f32 %f1454, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1454;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs142, %f1454;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs141, %f1454;}
// end inline asm
mov.u16 %rs144, 0;
st.global.v4.u16 [%rd30], {%rs141, %rs142, %rs143, %rs144};
$L__BB0_176:
ld.const.u64 %rd114, [params+288];
cvta.to.global.u64 %rd115, %rd114;
ld.const.u32 %r480, [params+280];
mad.lo.s32 %r481, %r480, %r6, %r5;
mul.wide.u32 %rd116, %r481, 8;
add.s64 %rd31, %rd115, %rd116;
@%p196 bra $L__BB0_178;
ld.global.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd31];
// begin inline asm
{ cvt.f32.f16 %f1455, %rs151;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1456, %rs152;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1457, %rs153;}
// end inline asm
add.f32 %f1458, %f1455, 0f00000000;
add.f32 %f1459, %f1456, 0f00000000;
add.f32 %f1460, %f1457, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs150, %f1460;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs149, %f1459;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs148, %f1458;}
// end inline asm
mov.u16 %rs155, 0;
st.global.v4.u16 [%rd31], {%rs148, %rs149, %rs150, %rs155};
bra.uni $L__BB0_179;
$L__BB0_178:
mov.f32 %f1463, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1463;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs157, %f1463;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs156, %f1463;}
// end inline asm
mov.u16 %rs159, 0;
st.global.v4.u16 [%rd31], {%rs156, %rs157, %rs158, %rs159};
$L__BB0_179:
ld.const.u64 %rd117, [params+304];
cvta.to.global.u64 %rd118, %rd117;
ld.const.u32 %r482, [params+296];
mad.lo.s32 %r483, %r482, %r6, %r5;
mul.wide.u32 %rd119, %r483, 8;
add.s64 %rd32, %rd118, %rd119;
@%p196 bra $L__BB0_181;
ld.global.v4.u16 {%rs166, %rs167, %rs168, %rs169}, [%rd32];
// begin inline asm
{ cvt.f32.f16 %f1464, %rs166;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1465, %rs167;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1466, %rs168;}
// end inline asm
add.f32 %f1467, %f1464, 0f00000000;
add.f32 %f1468, %f1465, 0f00000000;
add.f32 %f1469, %f1466, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs165, %f1469;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs164, %f1468;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs163, %f1467;}
// end inline asm
mov.u16 %rs170, 0;
st.global.v4.u16 [%rd32], {%rs163, %rs164, %rs165, %rs170};
bra.uni $L__BB0_182;
$L__BB0_181:
mov.f32 %f1472, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs173, %f1472;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs172, %f1472;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs171, %f1472;}
// end inline asm
mov.u16 %rs174, 0;
st.global.v4.u16 [%rd32], {%rs171, %rs172, %rs173, %rs174};
bra.uni $L__BB0_182;
$L__BB0_62:
mov.b32 %r356, %f1624;
xor.b32 %r357, %r356, -2147483648;
mov.b32 %f820, %r357;
selp.f32 %f1626, %f820, %f1624, %p3;
setp.geu.f32 %p79, %f282, 0f00000000;
@%p79 bra $L__BB0_66;
mov.f32 %f821, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f822, %f821;
setp.eq.f32 %p80, %f822, 0f3EE8BA2E;
@%p80 bra $L__BB0_66;
mov.f32 %f1626, 0f7FFFFFFF;
$L__BB0_66:
add.f32 %f825, %f284, 0f3EE8BA2E;
mov.b32 %r358, %f825;
setp.lt.s32 %p82, %r358, 2139095040;
@%p82 bra $L__BB0_71;
setp.gtu.f32 %p83, %f284, 0f7F800000;
@%p83 bra $L__BB0_70;
bra.uni $L__BB0_68;
$L__BB0_70:
add.f32 %f1626, %f282, 0f3EE8BA2E;
bra.uni $L__BB0_71;
$L__BB0_68:
setp.neu.f32 %p84, %f284, 0f7F800000;
@%p84 bra $L__BB0_71;
selp.f32 %f1626, 0fFF800000, 0f7F800000, %p3;
$L__BB0_71:
mul.f32 %f826, %f1626, 0f437F0000;
setp.eq.f32 %p85, %f282, 0f3F800000;
selp.f32 %f827, 0f437F0000, %f826, %p85;
cvt.rzi.u32.f32 %r359, %f827;
shl.b64 %rd77, %rd20, 1;
add.s64 %rd78, %rd19, %rd77;
cvt.u16.u32 %rs16, %r359;
mov.u16 %rs17, 255;
st.global.v2.u8 [%rd78], {%rs16, %rs17};
$L__BB0_72:
ld.const.v2.f32 {%f828, %f829}, [params+648];
mul.f32 %f297, %f1620, %f828;
mul.f32 %f298, %f1621, %f829;
ld.const.f32 %f299, [params+656];
mul.f32 %f300, %f1622, %f299;
and.b32 %r360, %r87, 1;
setp.eq.b32 %p86, %r360, 1;
mov.pred %p87, 0;
xor.pred %p88, %p86, %p87;
not.pred %p89, %p88;
@%p89 bra $L__BB0_146;
mov.f32 %f831, 0f3E666666;
cvt.rzi.f32.f32 %f832, %f831;
add.f32 %f833, %f832, %f832;
mov.f32 %f834, 0f3EE66666;
sub.f32 %f835, %f834, %f833;
abs.f32 %f301, %f835;
abs.f32 %f302, %f297;
setp.lt.f32 %p90, %f302, 0f00800000;
mul.f32 %f836, %f302, 0f4B800000;
selp.f32 %f837, %f836, %f302, %p90;
selp.f32 %f838, 0fC3170000, 0fC2FE0000, %p90;
mov.b32 %r361, %f837;
and.b32 %r362, %r361, 8388607;
or.b32 %r363, %r362, 1065353216;
mov.b32 %f839, %r363;
shr.u32 %r364, %r361, 23;
cvt.rn.f32.u32 %f840, %r364;
add.f32 %f841, %f838, %f840;
setp.gt.f32 %p91, %f839, 0f3FB504F3;
mul.f32 %f842, %f839, 0f3F000000;
add.f32 %f843, %f841, 0f3F800000;
selp.f32 %f844, %f843, %f841, %p91;
selp.f32 %f845, %f842, %f839, %p91;
add.f32 %f846, %f845, 0fBF800000;
add.f32 %f847, %f845, 0f3F800000;
rcp.approx.ftz.f32 %f848, %f847;
add.f32 %f849, %f846, %f846;
mul.f32 %f850, %f849, %f848;
mul.f32 %f851, %f850, %f850;
mov.f32 %f852, 0f3C4CAF63;
mov.f32 %f853, 0f3B18F0FE;
fma.rn.f32 %f854, %f853, %f851, %f852;
mov.f32 %f855, 0f3DAAAABD;
fma.rn.f32 %f856, %f854, %f851, %f855;
mul.rn.f32 %f857, %f856, %f851;
mul.rn.f32 %f858, %f857, %f850;
sub.f32 %f859, %f846, %f850;
add.f32 %f860, %f859, %f859;
neg.f32 %f861, %f850;
fma.rn.f32 %f862, %f861, %f846, %f860;
mul.rn.f32 %f863, %f848, %f862;
add.f32 %f864, %f858, %f850;
sub.f32 %f865, %f850, %f864;
add.f32 %f866, %f858, %f865;
add.f32 %f867, %f863, %f866;
add.f32 %f868, %f864, %f867;
sub.f32 %f869, %f864, %f868;
add.f32 %f870, %f867, %f869;
mov.f32 %f871, 0f3F317200;
mul.rn.f32 %f872, %f844, %f871;
mov.f32 %f873, 0f35BFBE8E;
mul.rn.f32 %f874, %f844, %f873;
add.f32 %f875, %f872, %f868;
sub.f32 %f876, %f872, %f875;
add.f32 %f877, %f868, %f876;
add.f32 %f878, %f870, %f877;
add.f32 %f879, %f874, %f878;
add.f32 %f880, %f875, %f879;
sub.f32 %f881, %f875, %f880;
add.f32 %f882, %f879, %f881;
mul.rn.f32 %f883, %f834, %f880;
neg.f32 %f884, %f883;
fma.rn.f32 %f885, %f834, %f880, %f884;
fma.rn.f32 %f886, %f834, %f882, %f885;
mov.f32 %f887, 0f00000000;
fma.rn.f32 %f888, %f887, %f880, %f886;
add.rn.f32 %f889, %f883, %f888;
neg.f32 %f890, %f889;
add.rn.f32 %f891, %f883, %f890;
add.rn.f32 %f892, %f891, %f888;
mov.b32 %r365, %f889;
setp.eq.s32 %p92, %r365, 1118925336;
add.s32 %r366, %r365, -1;
mov.b32 %f893, %r366;
add.f32 %f894, %f892, 0f37000000;
selp.f32 %f303, %f894, %f892, %p92;
selp.f32 %f895, %f893, %f889, %p92;
mov.f32 %f896, 0f3FB8AA3B;
mul.rn.f32 %f897, %f895, %f896;
cvt.rzi.f32.f32 %f898, %f897;
abs.f32 %f899, %f898;
setp.gt.f32 %p93, %f899, 0f42FC0000;
mov.b32 %r367, %f898;
and.b32 %r368, %r367, -2147483648;
or.b32 %r369, %r368, 1123811328;
mov.b32 %f900, %r369;
selp.f32 %f901, %f900, %f898, %p93;
mov.f32 %f902, 0fBF317218;
fma.rn.f32 %f903, %f901, %f902, %f895;
mov.f32 %f904, 0f3102E308;
fma.rn.f32 %f905, %f901, %f904, %f903;
mul.f32 %f906, %f905, 0f3FB8AA3B;
add.f32 %f907, %f901, 0f4B40007F;
mov.b32 %r370, %f907;
shl.b32 %r371, %r370, 23;
mov.b32 %f908, %r371;
ex2.approx.ftz.f32 %f909, %f906;
mul.f32 %f304, %f909, %f908;
setp.eq.f32 %p94, %f304, 0f7F800000;
mov.f32 %f1627, 0f7F800000;
@%p94 bra $L__BB0_75;
fma.rn.f32 %f1627, %f304, %f303, %f304;
$L__BB0_75:
setp.lt.f32 %p95, %f297, 0f00000000;
setp.eq.f32 %p96, %f301, 0f3F800000;
and.pred %p4, %p95, %p96;
setp.eq.f32 %p97, %f297, 0f00000000;
@%p97 bra $L__BB0_79;
bra.uni $L__BB0_76;
$L__BB0_79:
add.f32 %f914, %f297, %f297;
selp.f32 %f1629, %f914, 0f00000000, %p96;
bra.uni $L__BB0_80;
$L__BB0_76:
mov.b32 %r372, %f1627;
xor.b32 %r373, %r372, -2147483648;
mov.b32 %f910, %r373;
selp.f32 %f1629, %f910, %f1627, %p4;
setp.geu.f32 %p98, %f297, 0f00000000;
@%p98 bra $L__BB0_80;
mov.f32 %f911, 0f3EE66666;
cvt.rzi.f32.f32 %f912, %f911;
setp.eq.f32 %p99, %f912, 0f3EE66666;
@%p99 bra $L__BB0_80;
mov.f32 %f1629, 0f7FFFFFFF;
$L__BB0_80:
add.f32 %f915, %f302, 0f3EE66666;
mov.b32 %r374, %f915;
setp.lt.s32 %p101, %r374, 2139095040;
@%p101 bra $L__BB0_85;
setp.gtu.f32 %p102, %f302, 0f7F800000;
@%p102 bra $L__BB0_84;
bra.uni $L__BB0_82;
$L__BB0_84:
add.f32 %f1629, %f297, 0f3EE66666;
bra.uni $L__BB0_85;
$L__BB0_82:
setp.neu.f32 %p103, %f302, 0f7F800000;
@%p103 bra $L__BB0_85;
selp.f32 %f1629, 0fFF800000, 0f7F800000, %p4;
$L__BB0_85:
setp.eq.f32 %p104, %f297, 0f3F800000;
selp.f32 %f313, 0f3F800000, %f1629, %p104;
abs.f32 %f314, %f298;
setp.lt.f32 %p105, %f314, 0f00800000;
mul.f32 %f917, %f314, 0f4B800000;
selp.f32 %f918, %f917, %f314, %p105;
selp.f32 %f919, 0fC3170000, 0fC2FE0000, %p105;
mov.b32 %r375, %f918;
and.b32 %r376, %r375, 8388607;
or.b32 %r377, %r376, 1065353216;
mov.b32 %f920, %r377;
shr.u32 %r378, %r375, 23;
cvt.rn.f32.u32 %f921, %r378;
add.f32 %f922, %f919, %f921;
setp.gt.f32 %p106, %f920, 0f3FB504F3;
mul.f32 %f923, %f920, 0f3F000000;
add.f32 %f924, %f922, 0f3F800000;
selp.f32 %f925, %f924, %f922, %p106;
selp.f32 %f926, %f923, %f920, %p106;
add.f32 %f927, %f926, 0fBF800000;
add.f32 %f928, %f926, 0f3F800000;
rcp.approx.ftz.f32 %f929, %f928;
add.f32 %f930, %f927, %f927;
mul.f32 %f931, %f930, %f929;
mul.f32 %f932, %f931, %f931;
mov.f32 %f933, 0f3C4CAF63;
mov.f32 %f934, 0f3B18F0FE;
fma.rn.f32 %f935, %f934, %f932, %f933;
mov.f32 %f936, 0f3DAAAABD;
fma.rn.f32 %f937, %f935, %f932, %f936;
mul.rn.f32 %f938, %f937, %f932;
mul.rn.f32 %f939, %f938, %f931;
sub.f32 %f940, %f927, %f931;
add.f32 %f941, %f940, %f940;
neg.f32 %f942, %f931;
fma.rn.f32 %f943, %f942, %f927, %f941;
mul.rn.f32 %f944, %f929, %f943;
add.f32 %f945, %f939, %f931;
sub.f32 %f946, %f931, %f945;
add.f32 %f947, %f939, %f946;
add.f32 %f948, %f944, %f947;
add.f32 %f949, %f945, %f948;
sub.f32 %f950, %f945, %f949;
add.f32 %f951, %f948, %f950;
mov.f32 %f952, 0f3F317200;
mul.rn.f32 %f953, %f925, %f952;
mov.f32 %f954, 0f35BFBE8E;
mul.rn.f32 %f955, %f925, %f954;
add.f32 %f956, %f953, %f949;
sub.f32 %f957, %f953, %f956;
add.f32 %f958, %f949, %f957;
add.f32 %f959, %f951, %f958;
add.f32 %f960, %f955, %f959;
add.f32 %f961, %f956, %f960;
sub.f32 %f962, %f956, %f961;
add.f32 %f963, %f960, %f962;
mov.f32 %f964, 0f3EE66666;
mul.rn.f32 %f965, %f964, %f961;
neg.f32 %f966, %f965;
fma.rn.f32 %f967, %f964, %f961, %f966;
fma.rn.f32 %f968, %f964, %f963, %f967;
mov.f32 %f969, 0f00000000;
fma.rn.f32 %f970, %f969, %f961, %f968;
add.rn.f32 %f971, %f965, %f970;
neg.f32 %f972, %f971;
add.rn.f32 %f973, %f965, %f972;
add.rn.f32 %f974, %f973, %f970;
mov.b32 %r379, %f971;
setp.eq.s32 %p107, %r379, 1118925336;
add.s32 %r380, %r379, -1;
mov.b32 %f975, %r380;
add.f32 %f976, %f974, 0f37000000;
selp.f32 %f315, %f976, %f974, %p107;
selp.f32 %f977, %f975, %f971, %p107;
mov.f32 %f978, 0f3FB8AA3B;
mul.rn.f32 %f979, %f977, %f978;
cvt.rzi.f32.f32 %f980, %f979;
abs.f32 %f981, %f980;
setp.gt.f32 %p108, %f981, 0f42FC0000;
mov.b32 %r381, %f980;
and.b32 %r382, %r381, -2147483648;
or.b32 %r383, %r382, 1123811328;
mov.b32 %f982, %r383;
selp.f32 %f983, %f982, %f980, %p108;
mov.f32 %f984, 0fBF317218;
fma.rn.f32 %f985, %f983, %f984, %f977;
mov.f32 %f986, 0f3102E308;
fma.rn.f32 %f987, %f983, %f986, %f985;
mul.f32 %f988, %f987, 0f3FB8AA3B;
add.f32 %f989, %f983, 0f4B40007F;
mov.b32 %r384, %f989;
shl.b32 %r385, %r384, 23;
mov.b32 %f990, %r385;
ex2.approx.ftz.f32 %f991, %f988;
mul.f32 %f316, %f991, %f990;
setp.eq.f32 %p109, %f316, 0f7F800000;
mov.f32 %f1630, 0f7F800000;
@%p109 bra $L__BB0_87;
fma.rn.f32 %f1630, %f316, %f315, %f316;
$L__BB0_87:
setp.lt.f32 %p110, %f298, 0f00000000;
and.pred %p5, %p110, %p96;
setp.eq.f32 %p112, %f298, 0f00000000;
@%p112 bra $L__BB0_91;
bra.uni $L__BB0_88;
$L__BB0_91:
add.f32 %f996, %f298, %f298;
selp.f32 %f1632, %f996, 0f00000000, %p96;
bra.uni $L__BB0_92;
$L__BB0_88:
mov.b32 %r386, %f1630;
xor.b32 %r387, %r386, -2147483648;
mov.b32 %f992, %r387;
selp.f32 %f1632, %f992, %f1630, %p5;
setp.geu.f32 %p113, %f298, 0f00000000;
@%p113 bra $L__BB0_92;
mov.f32 %f993, 0f3EE66666;
cvt.rzi.f32.f32 %f994, %f993;
setp.eq.f32 %p114, %f994, 0f3EE66666;
@%p114 bra $L__BB0_92;
mov.f32 %f1632, 0f7FFFFFFF;
$L__BB0_92:
add.f32 %f997, %f314, 0f3EE66666;
mov.b32 %r388, %f997;
setp.lt.s32 %p116, %r388, 2139095040;
@%p116 bra $L__BB0_97;
setp.gtu.f32 %p117, %f314, 0f7F800000;
@%p117 bra $L__BB0_96;
bra.uni $L__BB0_94;
$L__BB0_96:
add.f32 %f1632, %f298, 0f3EE66666;
bra.uni $L__BB0_97;
$L__BB0_94:
setp.neu.f32 %p118, %f314, 0f7F800000;
@%p118 bra $L__BB0_97;
selp.f32 %f1632, 0fFF800000, 0f7F800000, %p5;
$L__BB0_97:
setp.eq.f32 %p119, %f298, 0f3F800000;
selp.f32 %f325, 0f3F800000, %f1632, %p119;
abs.f32 %f326, %f300;
setp.lt.f32 %p120, %f326, 0f00800000;
mul.f32 %f999, %f326, 0f4B800000;
selp.f32 %f1000, %f999, %f326, %p120;
selp.f32 %f1001, 0fC3170000, 0fC2FE0000, %p120;
mov.b32 %r389, %f1000;
and.b32 %r390, %r389, 8388607;
or.b32 %r391, %r390, 1065353216;
mov.b32 %f1002, %r391;
shr.u32 %r392, %r389, 23;
cvt.rn.f32.u32 %f1003, %r392;
add.f32 %f1004, %f1001, %f1003;
setp.gt.f32 %p121, %f1002, 0f3FB504F3;
mul.f32 %f1005, %f1002, 0f3F000000;
add.f32 %f1006, %f1004, 0f3F800000;
selp.f32 %f1007, %f1006, %f1004, %p121;
selp.f32 %f1008, %f1005, %f1002, %p121;
add.f32 %f1009, %f1008, 0fBF800000;
add.f32 %f1010, %f1008, 0f3F800000;
rcp.approx.ftz.f32 %f1011, %f1010;
add.f32 %f1012, %f1009, %f1009;
mul.f32 %f1013, %f1012, %f1011;
mul.f32 %f1014, %f1013, %f1013;
mov.f32 %f1015, 0f3C4CAF63;
mov.f32 %f1016, 0f3B18F0FE;
fma.rn.f32 %f1017, %f1016, %f1014, %f1015;
mov.f32 %f1018, 0f3DAAAABD;
fma.rn.f32 %f1019, %f1017, %f1014, %f1018;
mul.rn.f32 %f1020, %f1019, %f1014;
mul.rn.f32 %f1021, %f1020, %f1013;
sub.f32 %f1022, %f1009, %f1013;
add.f32 %f1023, %f1022, %f1022;
neg.f32 %f1024, %f1013;
fma.rn.f32 %f1025, %f1024, %f1009, %f1023;
mul.rn.f32 %f1026, %f1011, %f1025;
add.f32 %f1027, %f1021, %f1013;
sub.f32 %f1028, %f1013, %f1027;
add.f32 %f1029, %f1021, %f1028;
add.f32 %f1030, %f1026, %f1029;
add.f32 %f1031, %f1027, %f1030;
sub.f32 %f1032, %f1027, %f1031;
add.f32 %f1033, %f1030, %f1032;
mov.f32 %f1034, 0f3F317200;
mul.rn.f32 %f1035, %f1007, %f1034;
mov.f32 %f1036, 0f35BFBE8E;
mul.rn.f32 %f1037, %f1007, %f1036;
add.f32 %f1038, %f1035, %f1031;
sub.f32 %f1039, %f1035, %f1038;
add.f32 %f1040, %f1031, %f1039;
add.f32 %f1041, %f1033, %f1040;
add.f32 %f1042, %f1037, %f1041;
add.f32 %f1043, %f1038, %f1042;
sub.f32 %f1044, %f1038, %f1043;
add.f32 %f1045, %f1042, %f1044;
mov.f32 %f1046, 0f3EE66666;
mul.rn.f32 %f1047, %f1046, %f1043;
neg.f32 %f1048, %f1047;
fma.rn.f32 %f1049, %f1046, %f1043, %f1048;
fma.rn.f32 %f1050, %f1046, %f1045, %f1049;
mov.f32 %f1051, 0f00000000;
fma.rn.f32 %f1052, %f1051, %f1043, %f1050;
add.rn.f32 %f1053, %f1047, %f1052;
neg.f32 %f1054, %f1053;
add.rn.f32 %f1055, %f1047, %f1054;
add.rn.f32 %f1056, %f1055, %f1052;
mov.b32 %r393, %f1053;
setp.eq.s32 %p122, %r393, 1118925336;
add.s32 %r394, %r393, -1;
mov.b32 %f1057, %r394;
add.f32 %f1058, %f1056, 0f37000000;
selp.f32 %f327, %f1058, %f1056, %p122;
selp.f32 %f1059, %f1057, %f1053, %p122;
mov.f32 %f1060, 0f3FB8AA3B;
mul.rn.f32 %f1061, %f1059, %f1060;
cvt.rzi.f32.f32 %f1062, %f1061;
abs.f32 %f1063, %f1062;
setp.gt.f32 %p123, %f1063, 0f42FC0000;
mov.b32 %r395, %f1062;
and.b32 %r396, %r395, -2147483648;
or.b32 %r397, %r396, 1123811328;
mov.b32 %f1064, %r397;
selp.f32 %f1065, %f1064, %f1062, %p123;
mov.f32 %f1066, 0fBF317218;
fma.rn.f32 %f1067, %f1065, %f1066, %f1059;
mov.f32 %f1068, 0f3102E308;
fma.rn.f32 %f1069, %f1065, %f1068, %f1067;
mul.f32 %f1070, %f1069, 0f3FB8AA3B;
add.f32 %f1071, %f1065, 0f4B40007F;
mov.b32 %r398, %f1071;
shl.b32 %r399, %r398, 23;
mov.b32 %f1072, %r399;
ex2.approx.ftz.f32 %f1073, %f1070;
mul.f32 %f328, %f1073, %f1072;
setp.eq.f32 %p124, %f328, 0f7F800000;
mov.f32 %f1633, 0f7F800000;
@%p124 bra $L__BB0_99;
fma.rn.f32 %f1633, %f328, %f327, %f328;
$L__BB0_99:
setp.lt.f32 %p125, %f300, 0f00000000;
and.pred %p6, %p125, %p96;
setp.eq.f32 %p127, %f300, 0f00000000;
@%p127 bra $L__BB0_103;
bra.uni $L__BB0_100;
$L__BB0_103:
add.f32 %f1078, %f300, %f300;
selp.f32 %f1635, %f1078, 0f00000000, %p96;
bra.uni $L__BB0_104;
$L__BB0_100:
mov.b32 %r400, %f1633;
xor.b32 %r401, %r400, -2147483648;
mov.b32 %f1074, %r401;
selp.f32 %f1635, %f1074, %f1633, %p6;
setp.geu.f32 %p128, %f300, 0f00000000;
@%p128 bra $L__BB0_104;
mov.f32 %f1075, 0f3EE66666;
cvt.rzi.f32.f32 %f1076, %f1075;
setp.eq.f32 %p129, %f1076, 0f3EE66666;
@%p129 bra $L__BB0_104;
mov.f32 %f1635, 0f7FFFFFFF;
$L__BB0_104:
add.f32 %f1079, %f326, 0f3EE66666;
mov.b32 %r402, %f1079;
setp.lt.s32 %p131, %r402, 2139095040;
@%p131 bra $L__BB0_109;
setp.gtu.f32 %p132, %f326, 0f7F800000;
@%p132 bra $L__BB0_108;
bra.uni $L__BB0_106;
$L__BB0_108:
add.f32 %f1635, %f300, 0f3EE66666;
bra.uni $L__BB0_109;
$L__BB0_106:
setp.neu.f32 %p133, %f326, 0f7F800000;
@%p133 bra $L__BB0_109;
selp.f32 %f1635, 0fFF800000, 0f7F800000, %p6;
$L__BB0_109:
setp.eq.f32 %p134, %f300, 0f3F800000;
mov.f32 %f1081, 0f3F800000;
selp.f32 %f1082, 0f3F800000, %f1635, %p134;
ld.const.u64 %rd79, [params+144];
cvta.to.global.u64 %rd21, %rd79;
ld.const.u32 %r403, [params+136];
mad.lo.s32 %r404, %r403, %r6, %r5;
cvt.u64.u32 %rd22, %r404;
min.f32 %f1083, %f313, %f1081;
mov.f32 %f1084, 0f00000000;
max.f32 %f337, %f1084, %f1083;
min.f32 %f1085, %f325, %f1081;
max.f32 %f338, %f1084, %f1085;
min.f32 %f1086, %f1082, %f1081;
max.f32 %f339, %f1084, %f1086;
mov.f32 %f1087, 0f3E555555;
cvt.rzi.f32.f32 %f1088, %f1087;
add.f32 %f1089, %f1088, %f1088;
mov.f32 %f1090, 0f3ED55555;
sub.f32 %f1091, %f1090, %f1089;
abs.f32 %f340, %f1091;
abs.f32 %f341, %f337;
setp.lt.f32 %p135, %f341, 0f00800000;
mul.f32 %f1092, %f341, 0f4B800000;
selp.f32 %f1093, %f1092, %f341, %p135;
selp.f32 %f1094, 0fC3170000, 0fC2FE0000, %p135;
mov.b32 %r405, %f1093;
and.b32 %r406, %r405, 8388607;
or.b32 %r407, %r406, 1065353216;
mov.b32 %f1095, %r407;
shr.u32 %r408, %r405, 23;
cvt.rn.f32.u32 %f1096, %r408;
add.f32 %f1097, %f1094, %f1096;
setp.gt.f32 %p136, %f1095, 0f3FB504F3;
mul.f32 %f1098, %f1095, 0f3F000000;
add.f32 %f1099, %f1097, 0f3F800000;
selp.f32 %f1100, %f1099, %f1097, %p136;
selp.f32 %f1101, %f1098, %f1095, %p136;
add.f32 %f1102, %f1101, 0fBF800000;
add.f32 %f1103, %f1101, 0f3F800000;
rcp.approx.ftz.f32 %f1104, %f1103;
add.f32 %f1105, %f1102, %f1102;
mul.f32 %f1106, %f1105, %f1104;
mul.f32 %f1107, %f1106, %f1106;
mov.f32 %f1108, 0f3C4CAF63;
mov.f32 %f1109, 0f3B18F0FE;
fma.rn.f32 %f1110, %f1109, %f1107, %f1108;
mov.f32 %f1111, 0f3DAAAABD;
fma.rn.f32 %f1112, %f1110, %f1107, %f1111;
mul.rn.f32 %f1113, %f1112, %f1107;
mul.rn.f32 %f1114, %f1113, %f1106;
sub.f32 %f1115, %f1102, %f1106;
add.f32 %f1116, %f1115, %f1115;
neg.f32 %f1117, %f1106;
fma.rn.f32 %f1118, %f1117, %f1102, %f1116;
mul.rn.f32 %f1119, %f1104, %f1118;
add.f32 %f1120, %f1114, %f1106;
sub.f32 %f1121, %f1106, %f1120;
add.f32 %f1122, %f1114, %f1121;
add.f32 %f1123, %f1119, %f1122;
add.f32 %f1124, %f1120, %f1123;
sub.f32 %f1125, %f1120, %f1124;
add.f32 %f1126, %f1123, %f1125;
mov.f32 %f1127, 0f3F317200;
mul.rn.f32 %f1128, %f1100, %f1127;
mov.f32 %f1129, 0f35BFBE8E;
mul.rn.f32 %f1130, %f1100, %f1129;
add.f32 %f1131, %f1128, %f1124;
sub.f32 %f1132, %f1128, %f1131;
add.f32 %f1133, %f1124, %f1132;
add.f32 %f1134, %f1126, %f1133;
add.f32 %f1135, %f1130, %f1134;
add.f32 %f1136, %f1131, %f1135;
sub.f32 %f1137, %f1131, %f1136;
add.f32 %f1138, %f1135, %f1137;
mul.rn.f32 %f1139, %f1090, %f1136;
neg.f32 %f1140, %f1139;
fma.rn.f32 %f1141, %f1090, %f1136, %f1140;
fma.rn.f32 %f1142, %f1090, %f1138, %f1141;
fma.rn.f32 %f1143, %f1084, %f1136, %f1142;
add.rn.f32 %f1144, %f1139, %f1143;
neg.f32 %f1145, %f1144;
add.rn.f32 %f1146, %f1139, %f1145;
add.rn.f32 %f1147, %f1146, %f1143;
mov.b32 %r409, %f1144;
setp.eq.s32 %p137, %r409, 1118925336;
add.s32 %r410, %r409, -1;
mov.b32 %f1148, %r410;
add.f32 %f1149, %f1147, 0f37000000;
selp.f32 %f342, %f1149, %f1147, %p137;
selp.f32 %f1150, %f1148, %f1144, %p137;
mov.f32 %f1151, 0f3FB8AA3B;
mul.rn.f32 %f1152, %f1150, %f1151;
cvt.rzi.f32.f32 %f1153, %f1152;
abs.f32 %f1154, %f1153;
setp.gt.f32 %p138, %f1154, 0f42FC0000;
mov.b32 %r411, %f1153;
and.b32 %r412, %r411, -2147483648;
or.b32 %r413, %r412, 1123811328;
mov.b32 %f1155, %r413;
selp.f32 %f1156, %f1155, %f1153, %p138;
mov.f32 %f1157, 0fBF317218;
fma.rn.f32 %f1158, %f1156, %f1157, %f1150;
mov.f32 %f1159, 0f3102E308;
fma.rn.f32 %f1160, %f1156, %f1159, %f1158;
mul.f32 %f1161, %f1160, 0f3FB8AA3B;
add.f32 %f1162, %f1156, 0f4B40007F;
mov.b32 %r414, %f1162;
shl.b32 %r415, %r414, 23;
mov.b32 %f1163, %r415;
ex2.approx.ftz.f32 %f1164, %f1161;
mul.f32 %f343, %f1164, %f1163;
setp.eq.f32 %p139, %f343, 0f7F800000;
mov.f32 %f1636, 0f7F800000;
@%p139 bra $L__BB0_111;
fma.rn.f32 %f1636, %f343, %f342, %f343;
$L__BB0_111:
setp.lt.f32 %p140, %f337, 0f00000000;
setp.eq.f32 %p141, %f340, 0f3F800000;
and.pred %p7, %p140, %p141;
setp.eq.f32 %p142, %f337, 0f00000000;
@%p142 bra $L__BB0_115;
bra.uni $L__BB0_112;
$L__BB0_115:
add.f32 %f1169, %f337, %f337;
selp.f32 %f1638, %f1169, 0f00000000, %p141;
bra.uni $L__BB0_116;
$L__BB0_112:
mov.b32 %r416, %f1636;
xor.b32 %r417, %r416, -2147483648;
mov.b32 %f1165, %r417;
selp.f32 %f1638, %f1165, %f1636, %p7;
setp.geu.f32 %p143, %f337, 0f00000000;
@%p143 bra $L__BB0_116;
mov.f32 %f1166, 0f3ED55555;
cvt.rzi.f32.f32 %f1167, %f1166;
setp.eq.f32 %p144, %f1167, 0f3ED55555;
@%p144 bra $L__BB0_116;
mov.f32 %f1638, 0f7FFFFFFF;
$L__BB0_116:
add.f32 %f1170, %f341, 0f3ED55555;
mov.b32 %r418, %f1170;
setp.lt.s32 %p146, %r418, 2139095040;
@%p146 bra $L__BB0_121;
setp.gtu.f32 %p147, %f341, 0f7F800000;
@%p147 bra $L__BB0_120;
bra.uni $L__BB0_118;
$L__BB0_120:
add.f32 %f1638, %f337, 0f3ED55555;
bra.uni $L__BB0_121;
$L__BB0_118:
setp.neu.f32 %p148, %f341, 0f7F800000;
@%p148 bra $L__BB0_121;
selp.f32 %f1638, 0fFF800000, 0f7F800000, %p7;
$L__BB0_121:
abs.f32 %f352, %f338;
setp.lt.f32 %p149, %f352, 0f00800000;
mul.f32 %f1172, %f352, 0f4B800000;
selp.f32 %f1173, %f1172, %f352, %p149;
selp.f32 %f1174, 0fC3170000, 0fC2FE0000, %p149;
mov.b32 %r419, %f1173;
and.b32 %r420, %r419, 8388607;
or.b32 %r421, %r420, 1065353216;
mov.b32 %f1175, %r421;
shr.u32 %r422, %r419, 23;
cvt.rn.f32.u32 %f1176, %r422;
add.f32 %f1177, %f1174, %f1176;
setp.gt.f32 %p150, %f1175, 0f3FB504F3;
mul.f32 %f1178, %f1175, 0f3F000000;
add.f32 %f1179, %f1177, 0f3F800000;
selp.f32 %f1180, %f1179, %f1177, %p150;
selp.f32 %f1181, %f1178, %f1175, %p150;
add.f32 %f1182, %f1181, 0fBF800000;
add.f32 %f1183, %f1181, 0f3F800000;
rcp.approx.ftz.f32 %f1184, %f1183;
add.f32 %f1185, %f1182, %f1182;
mul.f32 %f1186, %f1185, %f1184;
mul.f32 %f1187, %f1186, %f1186;
mov.f32 %f1188, 0f3C4CAF63;
mov.f32 %f1189, 0f3B18F0FE;
fma.rn.f32 %f1190, %f1189, %f1187, %f1188;
mov.f32 %f1191, 0f3DAAAABD;
fma.rn.f32 %f1192, %f1190, %f1187, %f1191;
mul.rn.f32 %f1193, %f1192, %f1187;
mul.rn.f32 %f1194, %f1193, %f1186;
sub.f32 %f1195, %f1182, %f1186;
add.f32 %f1196, %f1195, %f1195;
neg.f32 %f1197, %f1186;
fma.rn.f32 %f1198, %f1197, %f1182, %f1196;
mul.rn.f32 %f1199, %f1184, %f1198;
add.f32 %f1200, %f1194, %f1186;
sub.f32 %f1201, %f1186, %f1200;
add.f32 %f1202, %f1194, %f1201;
add.f32 %f1203, %f1199, %f1202;
add.f32 %f1204, %f1200, %f1203;
sub.f32 %f1205, %f1200, %f1204;
add.f32 %f1206, %f1203, %f1205;
mov.f32 %f1207, 0f3F317200;
mul.rn.f32 %f1208, %f1180, %f1207;
mov.f32 %f1209, 0f35BFBE8E;
mul.rn.f32 %f1210, %f1180, %f1209;
add.f32 %f1211, %f1208, %f1204;
sub.f32 %f1212, %f1208, %f1211;
add.f32 %f1213, %f1204, %f1212;
add.f32 %f1214, %f1206, %f1213;
add.f32 %f1215, %f1210, %f1214;
add.f32 %f1216, %f1211, %f1215;
sub.f32 %f1217, %f1211, %f1216;
add.f32 %f1218, %f1215, %f1217;
mov.f32 %f1219, 0f3ED55555;
mul.rn.f32 %f1220, %f1219, %f1216;
neg.f32 %f1221, %f1220;
fma.rn.f32 %f1222, %f1219, %f1216, %f1221;
fma.rn.f32 %f1223, %f1219, %f1218, %f1222;
mov.f32 %f1224, 0f00000000;
fma.rn.f32 %f1225, %f1224, %f1216, %f1223;
add.rn.f32 %f1226, %f1220, %f1225;
neg.f32 %f1227, %f1226;
add.rn.f32 %f1228, %f1220, %f1227;
add.rn.f32 %f1229, %f1228, %f1225;
mov.b32 %r423, %f1226;
setp.eq.s32 %p151, %r423, 1118925336;
add.s32 %r424, %r423, -1;
mov.b32 %f1230, %r424;
add.f32 %f1231, %f1229, 0f37000000;
selp.f32 %f353, %f1231, %f1229, %p151;
selp.f32 %f1232, %f1230, %f1226, %p151;
mov.f32 %f1233, 0f3FB8AA3B;
mul.rn.f32 %f1234, %f1232, %f1233;
cvt.rzi.f32.f32 %f1235, %f1234;
abs.f32 %f1236, %f1235;
setp.gt.f32 %p152, %f1236, 0f42FC0000;
mov.b32 %r425, %f1235;
and.b32 %r426, %r425, -2147483648;
or.b32 %r427, %r426, 1123811328;
mov.b32 %f1237, %r427;
selp.f32 %f1238, %f1237, %f1235, %p152;
mov.f32 %f1239, 0fBF317218;
fma.rn.f32 %f1240, %f1238, %f1239, %f1232;
mov.f32 %f1241, 0f3102E308;
fma.rn.f32 %f1242, %f1238, %f1241, %f1240;
mul.f32 %f1243, %f1242, 0f3FB8AA3B;
add.f32 %f1244, %f1238, 0f4B40007F;
mov.b32 %r428, %f1244;
shl.b32 %r429, %r428, 23;
mov.b32 %f1245, %r429;
ex2.approx.ftz.f32 %f1246, %f1243;
mul.f32 %f354, %f1246, %f1245;
setp.eq.f32 %p153, %f354, 0f7F800000;
mov.f32 %f1639, 0f7F800000;
@%p153 bra $L__BB0_123;
fma.rn.f32 %f1639, %f354, %f353, %f354;
$L__BB0_123:
setp.lt.f32 %p154, %f338, 0f00000000;
and.pred %p8, %p154, %p141;
setp.eq.f32 %p156, %f338, 0f00000000;
@%p156 bra $L__BB0_127;
bra.uni $L__BB0_124;
$L__BB0_127:
add.f32 %f1251, %f338, %f338;
selp.f32 %f1641, %f1251, 0f00000000, %p141;
bra.uni $L__BB0_128;
$L__BB0_124:
mov.b32 %r430, %f1639;
xor.b32 %r431, %r430, -2147483648;
mov.b32 %f1247, %r431;
selp.f32 %f1641, %f1247, %f1639, %p8;
setp.geu.f32 %p157, %f338, 0f00000000;
@%p157 bra $L__BB0_128;
mov.f32 %f1248, 0f3ED55555;
cvt.rzi.f32.f32 %f1249, %f1248;
setp.eq.f32 %p158, %f1249, 0f3ED55555;
@%p158 bra $L__BB0_128;
mov.f32 %f1641, 0f7FFFFFFF;
$L__BB0_128:
add.f32 %f1252, %f352, 0f3ED55555;
mov.b32 %r432, %f1252;
setp.lt.s32 %p160, %r432, 2139095040;
@%p160 bra $L__BB0_133;
setp.gtu.f32 %p161, %f352, 0f7F800000;
@%p161 bra $L__BB0_132;
bra.uni $L__BB0_130;
$L__BB0_132:
add.f32 %f1641, %f338, 0f3ED55555;
bra.uni $L__BB0_133;
$L__BB0_130:
setp.neu.f32 %p162, %f352, 0f7F800000;
@%p162 bra $L__BB0_133;
selp.f32 %f1641, 0fFF800000, 0f7F800000, %p8;
$L__BB0_133:
abs.f32 %f363, %f339;
setp.lt.f32 %p163, %f363, 0f00800000;
mul.f32 %f1254, %f363, 0f4B800000;
selp.f32 %f1255, %f1254, %f363, %p163;
selp.f32 %f1256, 0fC3170000, 0fC2FE0000, %p163;
mov.b32 %r433, %f1255;
and.b32 %r434, %r433, 8388607;
or.b32 %r435, %r434, 1065353216;
mov.b32 %f1257, %r435;
shr.u32 %r436, %r433, 23;
cvt.rn.f32.u32 %f1258, %r436;
add.f32 %f1259, %f1256, %f1258;
setp.gt.f32 %p164, %f1257, 0f3FB504F3;
mul.f32 %f1260, %f1257, 0f3F000000;
add.f32 %f1261, %f1259, 0f3F800000;
selp.f32 %f1262, %f1261, %f1259, %p164;
selp.f32 %f1263, %f1260, %f1257, %p164;
add.f32 %f1264, %f1263, 0fBF800000;
add.f32 %f1265, %f1263, 0f3F800000;
rcp.approx.ftz.f32 %f1266, %f1265;
add.f32 %f1267, %f1264, %f1264;
mul.f32 %f1268, %f1267, %f1266;
mul.f32 %f1269, %f1268, %f1268;
mov.f32 %f1270, 0f3C4CAF63;
mov.f32 %f1271, 0f3B18F0FE;
fma.rn.f32 %f1272, %f1271, %f1269, %f1270;
mov.f32 %f1273, 0f3DAAAABD;
fma.rn.f32 %f1274, %f1272, %f1269, %f1273;
mul.rn.f32 %f1275, %f1274, %f1269;
mul.rn.f32 %f1276, %f1275, %f1268;
sub.f32 %f1277, %f1264, %f1268;
add.f32 %f1278, %f1277, %f1277;
neg.f32 %f1279, %f1268;
fma.rn.f32 %f1280, %f1279, %f1264, %f1278;
mul.rn.f32 %f1281, %f1266, %f1280;
add.f32 %f1282, %f1276, %f1268;
sub.f32 %f1283, %f1268, %f1282;
add.f32 %f1284, %f1276, %f1283;
add.f32 %f1285, %f1281, %f1284;
add.f32 %f1286, %f1282, %f1285;
sub.f32 %f1287, %f1282, %f1286;
add.f32 %f1288, %f1285, %f1287;
mov.f32 %f1289, 0f3F317200;
mul.rn.f32 %f1290, %f1262, %f1289;
mov.f32 %f1291, 0f35BFBE8E;
mul.rn.f32 %f1292, %f1262, %f1291;
add.f32 %f1293, %f1290, %f1286;
sub.f32 %f1294, %f1290, %f1293;
add.f32 %f1295, %f1286, %f1294;
add.f32 %f1296, %f1288, %f1295;
add.f32 %f1297, %f1292, %f1296;
add.f32 %f1298, %f1293, %f1297;
sub.f32 %f1299, %f1293, %f1298;
add.f32 %f1300, %f1297, %f1299;
mov.f32 %f1301, 0f3ED55555;
mul.rn.f32 %f1302, %f1301, %f1298;
neg.f32 %f1303, %f1302;
fma.rn.f32 %f1304, %f1301, %f1298, %f1303;
fma.rn.f32 %f1305, %f1301, %f1300, %f1304;
mov.f32 %f1306, 0f00000000;
fma.rn.f32 %f1307, %f1306, %f1298, %f1305;
add.rn.f32 %f1308, %f1302, %f1307;
neg.f32 %f1309, %f1308;
add.rn.f32 %f1310, %f1302, %f1309;
add.rn.f32 %f1311, %f1310, %f1307;
mov.b32 %r437, %f1308;
setp.eq.s32 %p165, %r437, 1118925336;
add.s32 %r438, %r437, -1;
mov.b32 %f1312, %r438;
add.f32 %f1313, %f1311, 0f37000000;
selp.f32 %f364, %f1313, %f1311, %p165;
selp.f32 %f1314, %f1312, %f1308, %p165;
mov.f32 %f1315, 0f3FB8AA3B;
mul.rn.f32 %f1316, %f1314, %f1315;
cvt.rzi.f32.f32 %f1317, %f1316;
abs.f32 %f1318, %f1317;
setp.gt.f32 %p166, %f1318, 0f42FC0000;
mov.b32 %r439, %f1317;
and.b32 %r440, %r439, -2147483648;
or.b32 %r441, %r440, 1123811328;
mov.b32 %f1319, %r441;
selp.f32 %f1320, %f1319, %f1317, %p166;
mov.f32 %f1321, 0fBF317218;
fma.rn.f32 %f1322, %f1320, %f1321, %f1314;
mov.f32 %f1323, 0f3102E308;
fma.rn.f32 %f1324, %f1320, %f1323, %f1322;
mul.f32 %f1325, %f1324, 0f3FB8AA3B;
add.f32 %f1326, %f1320, 0f4B40007F;
mov.b32 %r442, %f1326;
shl.b32 %r443, %r442, 23;
mov.b32 %f1327, %r443;
ex2.approx.ftz.f32 %f1328, %f1325;
mul.f32 %f365, %f1328, %f1327;
setp.eq.f32 %p167, %f365, 0f7F800000;
mov.f32 %f1642, 0f7F800000;
@%p167 bra $L__BB0_135;
fma.rn.f32 %f1642, %f365, %f364, %f365;
$L__BB0_135:
setp.lt.f32 %p168, %f339, 0f00000000;
and.pred %p9, %p168, %p141;
setp.eq.f32 %p170, %f339, 0f00000000;
@%p170 bra $L__BB0_139;
bra.uni $L__BB0_136;
$L__BB0_139:
add.f32 %f1333, %f339, %f339;
selp.f32 %f1644, %f1333, 0f00000000, %p141;
bra.uni $L__BB0_140;
$L__BB0_136:
mov.b32 %r444, %f1642;
xor.b32 %r445, %r444, -2147483648;
mov.b32 %f1329, %r445;
selp.f32 %f1644, %f1329, %f1642, %p9;
setp.geu.f32 %p171, %f339, 0f00000000;
@%p171 bra $L__BB0_140;
mov.f32 %f1330, 0f3ED55555;
cvt.rzi.f32.f32 %f1331, %f1330;
setp.eq.f32 %p172, %f1331, 0f3ED55555;
@%p172 bra $L__BB0_140;
mov.f32 %f1644, 0f7FFFFFFF;
$L__BB0_140:
add.f32 %f1334, %f363, 0f3ED55555;
mov.b32 %r446, %f1334;
setp.lt.s32 %p174, %r446, 2139095040;
@%p174 bra $L__BB0_145;
setp.gtu.f32 %p175, %f363, 0f7F800000;
@%p175 bra $L__BB0_144;
bra.uni $L__BB0_142;
$L__BB0_144:
add.f32 %f1644, %f339, 0f3ED55555;
bra.uni $L__BB0_145;
$L__BB0_142:
setp.neu.f32 %p176, %f363, 0f7F800000;
@%p176 bra $L__BB0_145;
selp.f32 %f1644, 0fFF800000, 0f7F800000, %p9;
$L__BB0_145:
fma.rn.f32 %f1335, %f1638, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p177, %f337, 0f3F800000;
mov.f32 %f1336, 0f3F800000;
selp.f32 %f1337, 0f3F7FFFFF, %f1335, %p177;
mul.f32 %f1338, %f337, 0f414EB852;
setp.lt.f32 %p178, %f337, 0f3B4D2E1C;
selp.f32 %f1339, %f1338, %f1337, %p178;
fma.rn.f32 %f1340, %f1641, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p179, %f338, 0f3F800000;
selp.f32 %f1341, 0f3F7FFFFF, %f1340, %p179;
mul.f32 %f1342, %f338, 0f414EB852;
setp.lt.f32 %p180, %f338, 0f3B4D2E1C;
selp.f32 %f1343, %f1342, %f1341, %p180;
fma.rn.f32 %f1344, %f1644, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p181, %f339, 0f3F800000;
selp.f32 %f1345, 0f3F7FFFFF, %f1344, %p181;
mul.f32 %f1346, %f339, 0f414EB852;
setp.lt.f32 %p182, %f339, 0f3B4D2E1C;
selp.f32 %f1347, %f1346, %f1345, %p182;
min.f32 %f1348, %f1339, %f1336;
mov.f32 %f1349, 0f00000000;
max.f32 %f1350, %f1349, %f1348;
mul.f32 %f1351, %f1350, 0f43800000;
cvt.rzi.u32.f32 %r447, %f1351;
min.u32 %r448, %r447, 255;
min.f32 %f1352, %f1343, %f1336;
max.f32 %f1353, %f1349, %f1352;
mul.f32 %f1354, %f1353, 0f43800000;
cvt.rzi.u32.f32 %r449, %f1354;
min.u32 %r450, %r449, 255;
min.f32 %f1355, %f1347, %f1336;
max.f32 %f1356, %f1349, %f1355;
mul.f32 %f1357, %f1356, 0f43800000;
cvt.rzi.u32.f32 %r451, %f1357;
min.u32 %r452, %r451, 255;
shl.b64 %rd80, %rd22, 2;
add.s64 %rd81, %rd21, %rd80;
cvt.u16.u32 %rs18, %r452;
cvt.u16.u32 %rs19, %r450;
cvt.u16.u32 %rs20, %r448;
mov.u16 %rs21, 255;
st.global.v4.u8 [%rd81], {%rs20, %rs19, %rs18, %rs21};
$L__BB0_146:
and.b32 %r453, %r87, 4;
setp.eq.s32 %p183, %r453, 0;
ld.const.u32 %r501, [params+108];
@%p183 bra $L__BB0_150;
setp.eq.s32 %p184, %r501, 0;
ld.const.u64 %rd82, [params+224];
cvta.to.global.u64 %rd83, %rd82;
ld.const.u32 %r454, [params+216];
mad.lo.s32 %r455, %r454, %r6, %r5;
mul.wide.u32 %rd84, %r455, 8;
add.s64 %rd23, %rd83, %rd84;
@%p184 bra $L__BB0_149;
ld.global.v4.u16 {%rs29, %rs30, %rs31, %rs32}, [%rd23];
// begin inline asm
{ cvt.f32.f16 %f1358, %rs29;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1359, %rs30;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1360, %rs31;}
// end inline asm
add.f32 %f1361, %f297, %f1358;
add.f32 %f1362, %f298, %f1359;
add.f32 %f1363, %f300, %f1360;
mov.f32 %f1364, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1363;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs26, %f1362;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs25, %f1361;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1364;}
// end inline asm
st.global.v4.u16 [%rd23], {%rs25, %rs26, %rs27, %rs28};
bra.uni $L__BB0_150;
$L__BB0_149:
mov.f32 %f1368, 0f3F800000;
mov.u32 %r501, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f1368;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f300;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs34, %f298;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs33, %f297;}
// end inline asm
st.global.v4.u16 [%rd23], {%rs33, %rs34, %rs35, %rs36};
$L__BB0_150:
selp.f32 %f1369, 0f3F000000, 0f3E800000, %p200;
mul.f32 %f374, %f1369, %f1617;
mul.f32 %f375, %f1369, %f1618;
mul.f32 %f376, %f1369, %f1619;
mul.f32 %f377, %f1369, %f1614;
mul.f32 %f378, %f1369, %f1615;
mul.f32 %f379, %f1369, %f1616;
mul.f32 %f380, %f1369, %f1611;
mul.f32 %f381, %f1369, %f1612;
mul.f32 %f382, %f1369, %f1613;
mul.f32 %f383, %f1369, %f1608;
mul.f32 %f384, %f1369, %f1609;
mul.f32 %f385, %f1369, %f1610;
mul.f32 %f386, %f374, %f828;
mul.f32 %f387, %f375, %f829;
mul.f32 %f388, %f376, %f299;
ld.const.u64 %rd85, [params+256];
cvta.to.global.u64 %rd86, %rd85;
ld.const.u32 %r457, [params+248];
mad.lo.s32 %r458, %r457, %r6, %r5;
mul.wide.u32 %rd87, %r458, 8;
add.s64 %rd24, %rd86, %rd87;
setp.eq.s32 %p185, %r501, 0;
@%p185 bra $L__BB0_152;
ld.global.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd24];
// begin inline asm
{ cvt.f32.f16 %f1370, %rs44;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1371, %rs45;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1372, %rs46;}
// end inline asm
add.f32 %f1373, %f386, %f1370;
add.f32 %f1374, %f387, %f1371;
add.f32 %f1375, %f388, %f1372;
mov.f32 %f1376, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs42, %f1375;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs41, %f1374;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs40, %f1373;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f1376;}
// end inline asm
st.global.v4.u16 [%rd24], {%rs40, %rs41, %rs42, %rs43};
bra.uni $L__BB0_153;
$L__BB0_152:
mov.f32 %f1380, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f1380;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f388;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs49, %f387;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs48, %f386;}
// end inline asm
st.global.v4.u16 [%rd24], {%rs48, %rs49, %rs50, %rs51};
$L__BB0_153:
mov.f32 %f1381, 0f34000000;
max.f32 %f1382, %f374, %f1381;
div.rn.f32 %f1383, %f377, %f1382;
max.f32 %f1384, %f375, %f1381;
div.rn.f32 %f1385, %f378, %f1384;
max.f32 %f1386, %f376, %f1381;
div.rn.f32 %f1387, %f379, %f1386;
fma.rn.f32 %f389, %f1383, 0f3F000000, 0f3F000000;
fma.rn.f32 %f390, %f1385, 0f3F000000, 0f3F000000;
fma.rn.f32 %f391, %f1387, 0f3F000000, 0f3F000000;
div.rn.f32 %f1388, %f380, %f1382;
div.rn.f32 %f1389, %f381, %f1384;
div.rn.f32 %f1390, %f382, %f1386;
fma.rn.f32 %f392, %f1388, 0f3F000000, 0f3F000000;
fma.rn.f32 %f393, %f1389, 0f3F000000, 0f3F000000;
fma.rn.f32 %f394, %f1390, 0f3F000000, 0f3F000000;
div.rn.f32 %f1391, %f383, %f1382;
div.rn.f32 %f1392, %f384, %f1384;
div.rn.f32 %f1393, %f385, %f1386;
fma.rn.f32 %f395, %f1391, 0f3F000000, 0f3F000000;
fma.rn.f32 %f396, %f1392, 0f3F000000, 0f3F000000;
fma.rn.f32 %f397, %f1393, 0f3F000000, 0f3F000000;
ld.const.u64 %rd88, [params+272];
cvta.to.global.u64 %rd89, %rd88;
ld.const.u32 %r459, [params+264];
mad.lo.s32 %r460, %r459, %r6, %r5;
mul.wide.u32 %rd90, %r460, 8;
add.s64 %rd25, %rd89, %rd90;
@%p185 bra $L__BB0_155;
ld.global.v4.u16 {%rs59, %rs60, %rs61, %rs62}, [%rd25];
// begin inline asm
{ cvt.f32.f16 %f1394, %rs59;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1395, %rs60;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1396, %rs61;}
// end inline asm
add.f32 %f1397, %f389, %f1394;
add.f32 %f1398, %f390, %f1395;
add.f32 %f1399, %f391, %f1396;
mov.f32 %f1400, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs57, %f1399;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs56, %f1398;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs55, %f1397;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs58, %f1400;}
// end inline asm
st.global.v4.u16 [%rd25], {%rs55, %rs56, %rs57, %rs58};
bra.uni $L__BB0_156;
$L__BB0_155:
mov.f32 %f1404, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs66, %f1404;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs65, %f391;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs64, %f390;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs63, %f389;}
// end inline asm
st.global.v4.u16 [%rd25], {%rs63, %rs64, %rs65, %rs66};
$L__BB0_156:
ld.const.u64 %rd91, [params+288];
cvta.to.global.u64 %rd92, %rd91;
ld.const.u32 %r461, [params+280];
mad.lo.s32 %r462, %r461, %r6, %r5;
mul.wide.u32 %rd93, %r462, 8;
add.s64 %rd26, %rd92, %rd93;
@%p185 bra $L__BB0_158;
ld.global.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd26];
// begin inline asm
{ cvt.f32.f16 %f1405, %rs74;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1406, %rs75;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1407, %rs76;}
// end inline asm
add.f32 %f1408, %f392, %f1405;
add.f32 %f1409, %f393, %f1406;
add.f32 %f1410, %f394, %f1407;
mov.f32 %f1411, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs72, %f1410;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs71, %f1409;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs70, %f1408;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs73, %f1411;}
// end inline asm
st.global.v4.u16 [%rd26], {%rs70, %rs71, %rs72, %rs73};
bra.uni $L__BB0_159;
$L__BB0_158:
mov.f32 %f1415, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs81, %f1415;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs80, %f394;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs79, %f393;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs78, %f392;}
// end inline asm
st.global.v4.u16 [%rd26], {%rs78, %rs79, %rs80, %rs81};
$L__BB0_159:
ld.const.u64 %rd94, [params+304];
cvta.to.global.u64 %rd95, %rd94;
ld.const.u32 %r463, [params+296];
mad.lo.s32 %r464, %r463, %r6, %r5;
mul.wide.u32 %rd96, %r464, 8;
add.s64 %rd27, %rd95, %rd96;
@%p185 bra $L__BB0_161;
ld.global.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd27];
// begin inline asm
{ cvt.f32.f16 %f1416, %rs89;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1417, %rs90;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1418, %rs91;}
// end inline asm
add.f32 %f1419, %f395, %f1416;
add.f32 %f1420, %f396, %f1417;
add.f32 %f1421, %f397, %f1418;
mov.f32 %f1422, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs87, %f1421;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs86, %f1420;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs85, %f1419;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs88, %f1422;}
// end inline asm
st.global.v4.u16 [%rd27], {%rs85, %rs86, %rs87, %rs88};
bra.uni $L__BB0_182;
$L__BB0_161:
mov.f32 %f1426, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs96, %f1426;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs95, %f397;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs94, %f396;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs93, %f395;}
// end inline asm
st.global.v4.u16 [%rd27], {%rs93, %rs94, %rs95, %rs96};
$L__BB0_182:
ret;
}