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

2701 lines
74 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<199>;
.reg .b16 %rs<177>;
.reg .f32 %f<1577>;
.reg .b32 %r<497>;
.reg .f64 %fd<5>;
.reg .b64 %rd<127>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r92), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r93), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u32 %r3, [params+616];
abs.s32 %r4, %r3;
ld.const.u32 %r5, [params+540];
shl.b32 %r6, %r5, 1;
ld.const.u64 %rd34, [params+400];
cvta.to.global.u64 %rd35, %rd34;
ld.const.u32 %r98, [params+392];
mad.lo.s32 %r99, %r98, %r93, %r92;
mul.wide.u32 %rd36, %r99, 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 %f1429, 0f00000000;
mov.u16 %rs176, 0;
mov.f32 %f1430, %f1429;
mov.f32 %f1431, %f1429;
@%p11 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f375, %rs7;
div.rn.f32 %f376, %f375, 0f437F0000;
fma.rn.f32 %f377, %f376, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs176, 255;
cvt.rn.f32.u16 %f378, %rs13;
div.rn.f32 %f379, %f378, 0f437F0000;
fma.rn.f32 %f380, %f379, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f381, %rs175;
div.rn.f32 %f382, %f381, 0f437F0000;
fma.rn.f32 %f383, %f382, 0f40000000, 0fBF800000;
mul.f32 %f384, %f380, %f380;
fma.rn.f32 %f385, %f377, %f377, %f384;
fma.rn.f32 %f386, %f383, %f383, %f385;
sqrt.rn.f32 %f387, %f386;
rcp.rn.f32 %f388, %f387;
mul.f32 %f1431, %f388, %f383;
mul.f32 %f1430, %f388, %f380;
mul.f32 %f1429, %f377, %f388;
$L__BB0_4:
ld.const.v2.u32 {%r100, %r101}, [params];
add.s32 %r7, %r100, %r92;
add.s32 %r8, %r101, %r93;
setp.eq.f32 %p12, %f1429, 0f00000000;
setp.eq.f32 %p13, %f1430, 0f00000000;
and.pred %p14, %p12, %p13;
setp.eq.f32 %p15, %f1431, 0f00000000;
and.pred %p16, %p15, %p14;
@%p16 bra $L__BB0_160;
bra.uni $L__BB0_5;
$L__BB0_160:
ld.const.u32 %r89, [params+104];
and.b32 %r461, %r89, 1;
setp.eq.b32 %p187, %r461, 1;
mov.pred %p188, 0;
xor.pred %p189, %p187, %p188;
not.pred %p190, %p189;
@%p190 bra $L__BB0_162;
ld.const.u64 %rd97, [params+144];
cvta.to.global.u64 %rd98, %rd97;
ld.const.u32 %r462, [params+136];
mad.lo.s32 %r463, %r462, %r8, %r7;
mul.wide.u32 %rd99, %r463, 4;
add.s64 %rd100, %rd98, %rd99;
mov.u16 %rs97, 0;
st.global.v4.u8 [%rd100], {%rs97, %rs97, %rs97, %rs97};
$L__BB0_162:
and.b32 %r464, %r89, 8;
setp.eq.s32 %p191, %r464, 0;
@%p191 bra $L__BB0_164;
ld.const.u64 %rd101, [params+192];
cvta.to.global.u64 %rd102, %rd101;
ld.const.u32 %r465, [params+184];
mad.lo.s32 %r466, %r465, %r8, %r7;
mov.f32 %f1377, 0f00000000;
cvt.rzi.u32.f32 %r467, %f1377;
mul.wide.u32 %rd103, %r466, 2;
add.s64 %rd104, %rd102, %rd103;
mov.u16 %rs98, 0;
cvt.u16.u32 %rs99, %r467;
st.global.v2.u8 [%rd104], {%rs99, %rs98};
$L__BB0_164:
and.b32 %r468, %r89, 4;
setp.eq.s32 %p192, %r468, 0;
ld.const.u32 %r496, [params+108];
@%p192 bra $L__BB0_168;
setp.eq.s32 %p193, %r496, 0;
ld.const.u64 %rd105, [params+224];
cvta.to.global.u64 %rd106, %rd105;
ld.const.u32 %r469, [params+216];
mad.lo.s32 %r470, %r469, %r8, %r7;
mul.wide.u32 %rd107, %r470, 8;
add.s64 %rd28, %rd106, %rd107;
@%p193 bra $L__BB0_167;
ld.global.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd28];
// begin inline asm
{ cvt.f32.f16 %f1378, %rs106;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1379, %rs107;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1380, %rs108;}
// end inline asm
add.f32 %f1381, %f1378, 0f00000000;
add.f32 %f1382, %f1379, 0f00000000;
add.f32 %f1383, %f1380, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs105, %f1383;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs104, %f1382;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs103, %f1381;}
// end inline asm
mov.u16 %rs110, 0;
st.global.v4.u16 [%rd28], {%rs103, %rs104, %rs105, %rs110};
bra.uni $L__BB0_168;
$L__BB0_5:
ld.const.u64 %rd37, [params+432];
cvta.to.global.u64 %rd38, %rd37;
ld.const.u32 %r105, [params+424];
mad.lo.s32 %r106, %r105, %r93, %r92;
mul.wide.u32 %rd39, %r106, 12;
add.s64 %rd40, %rd38, %rd39;
ld.global.f32 %f10, [%rd40];
mul.f32 %f405, %f10, 0f3456BF95;
ld.global.f32 %f11, [%rd40+4];
mul.f32 %f406, %f11, 0f3456BF95;
ld.global.f32 %f12, [%rd40+8];
mul.f32 %f407, %f12, 0f3456BF95;
abs.f32 %f408, %f1429;
div.rn.f32 %f409, %f405, %f408;
abs.f32 %f410, %f1430;
div.rn.f32 %f411, %f406, %f410;
abs.f32 %f412, %f1431;
div.rn.f32 %f413, %f407, %f412;
abs.f32 %f414, %f409;
abs.f32 %f415, %f411;
abs.f32 %f416, %f413;
mov.f32 %f417, 0f38D1B717;
max.f32 %f418, %f414, %f417;
max.f32 %f419, %f415, %f417;
max.f32 %f420, %f416, %f417;
fma.rn.f32 %f13, %f1429, %f418, %f10;
fma.rn.f32 %f14, %f1430, %f419, %f11;
fma.rn.f32 %f15, %f1431, %f420, %f12;
ld.const.u64 %rd41, [params+128];
cvta.to.global.u64 %rd42, %rd41;
ld.const.u32 %r107, [params+120];
mad.lo.s32 %r108, %r107, %r93, %r92;
mul.wide.u32 %rd43, %r108, 4;
add.s64 %rd44, %rd42, %rd43;
ld.global.u32 %r487, [%rd44];
setp.lt.s32 %p18, %r4, 1;
mov.pred %p17, 0;
mov.f32 %f1451, 0f00000000;
mov.u32 %r483, 0;
mov.f32 %f1452, %f1451;
mov.f32 %f1453, %f1451;
mov.f32 %f1454, %f1451;
mov.f32 %f1455, %f1451;
mov.f32 %f1456, %f1451;
mov.f32 %f1457, %f1451;
mov.f32 %f1458, %f1451;
mov.f32 %f1459, %f1451;
mov.f32 %f1460, %f1451;
mov.f32 %f1461, %f1451;
mov.f32 %f1462, %f1451;
mov.f32 %f1463, %f1451;
mov.f32 %f1464, %f1451;
mov.f32 %f1465, %f1451;
mov.f32 %f1466, %f1451;
mov.pred %p198, %p17;
@%p18 bra $L__BB0_23;
ld.const.f32 %f16, [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 %r110, [params+632];
and.b32 %r111, %r93, 255;
and.b32 %r112, %r92, 255;
mad.lo.s32 %r113, %r110, %r111, %r112;
mul.wide.u32 %rd48, %r113, 3;
add.s64 %rd4, %rd47, %rd48;
ld.const.f32 %f17, [params+660];
mul.f32 %f18, %f13, 0f3456BF95;
mul.f32 %f19, %f14, 0f3456BF95;
mul.f32 %f20, %f15, 0f3456BF95;
ld.const.u64 %rd5, [params+96];
abs.f32 %f564, %f19;
abs.f32 %f565, %f18;
max.f32 %f566, %f565, %f564;
abs.f32 %f567, %f20;
max.f32 %f568, %f566, %f567;
$L__BB0_7:
shl.b32 %r114, %r483, 1;
mul.wide.s32 %rd49, %r114, 12;
add.s64 %rd6, %rd3, %rd49;
ld.global.f32 %f437, [%rd6];
sub.f32 %f438, %f437, %f10;
ld.global.f32 %f439, [%rd6+4];
sub.f32 %f440, %f439, %f11;
ld.global.f32 %f441, [%rd6+8];
sub.f32 %f442, %f441, %f12;
mul.f32 %f443, %f440, %f440;
fma.rn.f32 %f444, %f438, %f438, %f443;
fma.rn.f32 %f445, %f442, %f442, %f444;
sqrt.rn.f32 %f37, %f445;
rcp.rn.f32 %f446, %f37;
mul.f32 %f38, %f438, %f446;
mul.f32 %f39, %f440, %f446;
mul.f32 %f40, %f442, %f446;
mul.f32 %f447, %f37, %f37;
div.rn.f32 %f41, %f16, %f447;
ld.global.u8 %rs15, [%rd4];
cvt.rn.f32.u16 %f448, %rs15;
div.rn.f32 %f449, %f448, 0fC37F0000;
fma.rn.f32 %f450, %f449, 0f3F333333, 0f3F800000;
setp.gt.f32 %p20, %f41, %f450;
setp.ne.s32 %p21, %r5, 0;
and.pred %p22, %p21, %p20;
mov.pred %p198, -1;
@%p22 bra $L__BB0_23;
mul.f32 %f42, %f37, %f17;
mov.f32 %f455, 0f40800000;
abs.f32 %f44, %f42;
setp.lt.f32 %p23, %f44, 0f00800000;
mul.f32 %f457, %f44, 0f4B800000;
selp.f32 %f458, %f457, %f44, %p23;
selp.f32 %f459, 0fC3170000, 0fC2FE0000, %p23;
mov.b32 %r115, %f458;
and.b32 %r116, %r115, 8388607;
or.b32 %r117, %r116, 1065353216;
mov.b32 %f460, %r117;
shr.u32 %r118, %r115, 23;
cvt.rn.f32.u32 %f461, %r118;
add.f32 %f462, %f459, %f461;
setp.gt.f32 %p24, %f460, 0f3FB504F3;
mul.f32 %f463, %f460, 0f3F000000;
add.f32 %f464, %f462, 0f3F800000;
selp.f32 %f465, %f464, %f462, %p24;
selp.f32 %f466, %f463, %f460, %p24;
add.f32 %f467, %f466, 0fBF800000;
add.f32 %f468, %f466, 0f3F800000;
rcp.approx.ftz.f32 %f469, %f468;
add.f32 %f470, %f467, %f467;
mul.f32 %f471, %f470, %f469;
mul.f32 %f472, %f471, %f471;
mov.f32 %f473, 0f3C4CAF63;
mov.f32 %f474, 0f3B18F0FE;
fma.rn.f32 %f475, %f474, %f472, %f473;
mov.f32 %f476, 0f3DAAAABD;
fma.rn.f32 %f477, %f475, %f472, %f476;
mul.rn.f32 %f478, %f477, %f472;
mul.rn.f32 %f479, %f478, %f471;
sub.f32 %f480, %f467, %f471;
add.f32 %f481, %f480, %f480;
neg.f32 %f482, %f471;
fma.rn.f32 %f483, %f482, %f467, %f481;
mul.rn.f32 %f484, %f469, %f483;
add.f32 %f485, %f479, %f471;
sub.f32 %f486, %f471, %f485;
add.f32 %f487, %f479, %f486;
add.f32 %f488, %f484, %f487;
add.f32 %f489, %f485, %f488;
sub.f32 %f490, %f485, %f489;
add.f32 %f491, %f488, %f490;
mov.f32 %f492, 0f3F317200;
mul.rn.f32 %f493, %f465, %f492;
mov.f32 %f494, 0f35BFBE8E;
mul.rn.f32 %f495, %f465, %f494;
add.f32 %f496, %f493, %f489;
sub.f32 %f497, %f493, %f496;
add.f32 %f498, %f489, %f497;
add.f32 %f499, %f491, %f498;
add.f32 %f500, %f495, %f499;
add.f32 %f501, %f496, %f500;
sub.f32 %f502, %f496, %f501;
add.f32 %f503, %f500, %f502;
mul.rn.f32 %f504, %f455, %f501;
neg.f32 %f505, %f504;
fma.rn.f32 %f506, %f455, %f501, %f505;
fma.rn.f32 %f507, %f455, %f503, %f506;
mov.f32 %f508, 0f00000000;
fma.rn.f32 %f509, %f508, %f501, %f507;
add.rn.f32 %f510, %f504, %f509;
neg.f32 %f511, %f510;
add.rn.f32 %f512, %f504, %f511;
add.rn.f32 %f513, %f512, %f509;
mov.b32 %r119, %f510;
setp.eq.s32 %p25, %r119, 1118925336;
add.s32 %r120, %r119, -1;
mov.b32 %f514, %r120;
add.f32 %f515, %f513, 0f37000000;
selp.f32 %f45, %f515, %f513, %p25;
selp.f32 %f516, %f514, %f510, %p25;
mov.f32 %f517, 0f3FB8AA3B;
mul.rn.f32 %f518, %f516, %f517;
cvt.rzi.f32.f32 %f519, %f518;
abs.f32 %f520, %f519;
setp.gt.f32 %p26, %f520, 0f42FC0000;
mov.b32 %r121, %f519;
and.b32 %r122, %r121, -2147483648;
or.b32 %r123, %r122, 1123811328;
mov.b32 %f521, %r123;
selp.f32 %f522, %f521, %f519, %p26;
mov.f32 %f523, 0fBF317218;
fma.rn.f32 %f524, %f522, %f523, %f516;
mov.f32 %f525, 0f3102E308;
fma.rn.f32 %f526, %f522, %f525, %f524;
mul.f32 %f527, %f526, 0f3FB8AA3B;
add.f32 %f528, %f522, 0f4B40007F;
mov.b32 %r124, %f528;
shl.b32 %r125, %r124, 23;
mov.b32 %f529, %r125;
ex2.approx.ftz.f32 %f530, %f527;
mul.f32 %f46, %f530, %f529;
setp.eq.f32 %p27, %f46, 0f7F800000;
mov.f32 %f1448, 0f7F800000;
@%p27 bra $L__BB0_10;
fma.rn.f32 %f1448, %f46, %f45, %f46;
$L__BB0_10:
mov.f32 %f1428, 0f40000000;
cvt.rzi.f32.f32 %f1427, %f1428;
add.f32 %f1426, %f1427, %f1427;
mov.f32 %f1425, 0f40800000;
sub.f32 %f1424, %f1425, %f1426;
abs.f32 %f1423, %f1424;
setp.lt.f32 %p28, %f42, 0f00000000;
setp.eq.f32 %p29, %f1423, 0f3F800000;
and.pred %p1, %p28, %p29;
setp.eq.f32 %p30, %f42, 0f00000000;
@%p30 bra $L__BB0_14;
bra.uni $L__BB0_11;
$L__BB0_14:
add.f32 %f535, %f42, %f42;
selp.f32 %f1450, %f535, 0f00000000, %p29;
bra.uni $L__BB0_15;
$L__BB0_11:
mov.b32 %r126, %f1448;
xor.b32 %r127, %r126, -2147483648;
mov.b32 %f531, %r127;
selp.f32 %f1450, %f531, %f1448, %p1;
setp.geu.f32 %p31, %f42, 0f00000000;
@%p31 bra $L__BB0_15;
mov.f32 %f532, 0f40800000;
cvt.rzi.f32.f32 %f533, %f532;
setp.eq.f32 %p32, %f533, 0f40800000;
@%p32 bra $L__BB0_15;
mov.f32 %f1450, 0f7FFFFFFF;
$L__BB0_15:
add.f32 %f536, %f44, 0f40800000;
mov.b32 %r128, %f536;
setp.lt.s32 %p34, %r128, 2139095040;
@%p34 bra $L__BB0_20;
setp.gtu.f32 %p35, %f44, 0f7F800000;
@%p35 bra $L__BB0_19;
bra.uni $L__BB0_17;
$L__BB0_19:
add.f32 %f1450, %f42, 0f40800000;
bra.uni $L__BB0_20;
$L__BB0_17:
setp.neu.f32 %p36, %f44, 0f7F800000;
@%p36 bra $L__BB0_20;
selp.f32 %f1450, 0fFF800000, 0f7F800000, %p1;
$L__BB0_20:
shl.b32 %r480, %r483, 1;
mul.wide.s32 %rd121, %r480, 12;
add.s64 %rd120, %rd3, %rd121;
mov.f32 %f537, 0f3F800000;
sub.f32 %f538, %f537, %f1450;
setp.eq.f32 %p37, %f42, 0f3F800000;
selp.f32 %f539, 0f00000000, %f538, %p37;
cvt.sat.f32.f32 %f540, %f539;
mul.f32 %f541, %f41, %f540;
ld.global.f32 %f542, [%rd120+12];
mul.f32 %f543, %f38, %f542;
ld.global.f32 %f544, [%rd120+16];
mul.f32 %f545, %f39, %f544;
neg.f32 %f546, %f545;
sub.f32 %f547, %f546, %f543;
ld.global.f32 %f548, [%rd120+20];
mul.f32 %f549, %f40, %f548;
sub.f32 %f550, %f547, %f549;
cvt.sat.f32.f32 %f551, %f550;
mul.f32 %f55, %f541, %f551;
mul.f32 %f552, %f1430, %f39;
fma.rn.f32 %f553, %f1429, %f38, %f552;
fma.rn.f32 %f56, %f1431, %f40, %f553;
setp.leu.f32 %p38, %f55, 0f3727C5AC;
@%p38 bra $L__BB0_22;
cvt.sat.f32.f32 %f563, %f56;
mov.f32 %f569, 0f38D1B717;
max.f32 %f560, %f568, %f569;
sub.f32 %f561, %f37, %f560;
mov.f32 %f562, 0f00000000;
mov.u32 %r165, 2;
mov.u32 %r166, 1;
mov.u32 %r167, 3;
mov.u32 %r170, 1065353216;
mov.u32 %r199, 0;
// begin inline asm
call(%r129,%r130,%r131,%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),_optix_trace_typed_32,(%r199,%rd5,%f13,%f14,%f15,%f38,%f39,%f40,%f560,%f561,%f562,%r166,%r199,%r166,%r165,%r166,%r167,%r170,%r170,%r170,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199,%r199);
// end inline asm
mov.b32 %f570, %r129;
mov.b32 %f571, %r130;
mov.b32 %f572, %r131;
mul.f32 %f573, %f55, 0f3EA2F983;
mul.f32 %f574, %f573, %f570;
mul.f32 %f575, %f573, %f571;
mul.f32 %f576, %f573, %f572;
fma.rn.f32 %f1463, %f563, %f574, %f1463;
fma.rn.f32 %f1464, %f563, %f575, %f1464;
fma.rn.f32 %f1465, %f563, %f576, %f1465;
fma.rn.f32 %f1457, %f38, %f574, %f1457;
fma.rn.f32 %f1458, %f38, %f575, %f1458;
fma.rn.f32 %f1459, %f38, %f576, %f1459;
fma.rn.f32 %f1454, %f39, %f574, %f1454;
fma.rn.f32 %f1455, %f39, %f575, %f1455;
fma.rn.f32 %f1456, %f39, %f576, %f1456;
fma.rn.f32 %f1451, %f40, %f574, %f1451;
fma.rn.f32 %f1452, %f40, %f575, %f1452;
fma.rn.f32 %f1453, %f40, %f576, %f1453;
add.f32 %f1460, %f1460, %f574;
add.f32 %f1461, %f1461, %f575;
add.f32 %f1462, %f1462, %f576;
add.f32 %f1466, %f1466, %f570;
$L__BB0_22:
add.s32 %r483, %r483, 1;
setp.lt.s32 %p40, %r483, %r4;
mov.pred %p198, %p17;
@%p40 bra $L__BB0_7;
$L__BB0_23:
cvt.rn.f32.s32 %f577, %r483;
mov.f32 %f578, 0f3F800000;
max.f32 %f579, %f577, %f578;
rcp.rn.f32 %f580, %f579;
mul.f32 %f1552, %f1463, %f580;
mul.f32 %f1553, %f1464, %f580;
mul.f32 %f1554, %f1465, %f580;
div.rn.f32 %f1555, %f1466, %f579;
mul.f32 %f1549, %f1460, %f580;
mul.f32 %f1550, %f1461, %f580;
mul.f32 %f1551, %f1462, %f580;
mul.f32 %f1546, %f1457, %f580;
mul.f32 %f1547, %f1458, %f580;
mul.f32 %f1548, %f1459, %f580;
mul.f32 %f1543, %f1454, %f580;
mul.f32 %f1544, %f1455, %f580;
mul.f32 %f1545, %f1456, %f580;
mul.f32 %f1540, %f1451, %f580;
mul.f32 %f1541, %f1452, %f580;
mul.f32 %f1542, %f1453, %f580;
not.pred %p41, %p198;
@%p41 bra $L__BB0_56;
setp.lt.s32 %p42, %r5, 1;
mov.f32 %f1514, 0f00000000;
mov.f32 %f1513, %f1514;
mov.f32 %f1512, %f1514;
mov.f32 %f1511, %f1514;
mov.f32 %f1510, %f1514;
mov.f32 %f1509, %f1514;
mov.f32 %f1508, %f1514;
mov.f32 %f1507, %f1514;
mov.f32 %f1506, %f1514;
mov.f32 %f1505, %f1514;
mov.f32 %f1504, %f1514;
mov.f32 %f1503, %f1514;
mov.f32 %f1502, %f1514;
mov.f32 %f1501, %f1514;
mov.f32 %f1500, %f1514;
mov.f32 %f1499, %f1514;
@%p42 bra $L__BB0_55;
cvt.rn.f32.s32 %f613, %r6;
rcp.rn.f32 %f121, %f613;
mul.f32 %f122, %f13, 0f3456BF95;
mul.f32 %f123, %f14, 0f3456BF95;
mul.f32 %f124, %f15, 0f3456BF95;
mov.u32 %r200, 0;
ld.const.u64 %rd7, [params+96];
add.s64 %rd8, %rd1, 24;
mov.u64 %rd51, __cudart_i2opi_f;
abs.f32 %f674, %f123;
abs.f32 %f675, %f122;
max.f32 %f676, %f675, %f674;
abs.f32 %f677, %f124;
max.f32 %f678, %f676, %f677;
mov.u32 %r484, %r200;
$L__BB0_26:
cvt.rn.f32.s32 %f141, %r484;
mov.u32 %r486, %r200;
$L__BB0_27:
mad.lo.s32 %r202, %r487, 1664525, 1013904223;
and.b32 %r203, %r202, 16777215;
cvt.rn.f32.u32 %f614, %r203;
fma.rn.f32 %f615, %f614, 0f33800000, %f141;
mul.f32 %f616, %f121, %f615;
mad.lo.s32 %r487, %r202, 1664525, 1013904223;
and.b32 %r204, %r487, 16777215;
cvt.rn.f32.u32 %f617, %r204;
cvt.rn.f32.s32 %f618, %r486;
fma.rn.f32 %f619, %f617, 0f33800000, %f618;
mul.f32 %f620, %f121, %f619;
fma.rn.f32 %f621, %f121, %f615, %f616;
mov.f32 %f622, 0f3F800000;
sub.f32 %f158, %f622, %f621;
mul.f32 %f623, %f158, %f158;
sub.f32 %f624, %f622, %f623;
mov.f32 %f625, 0f00000000;
max.f32 %f626, %f625, %f624;
sqrt.rn.f32 %f159, %f626;
mul.f32 %f160, %f620, 0f40C90FDB;
mul.f32 %f627, %f160, 0f3F22F983;
cvt.rni.s32.f32 %r494, %f627;
cvt.rn.f32.s32 %f628, %r494;
mov.f32 %f629, 0fBFC90FDA;
fma.rn.f32 %f630, %f628, %f629, %f160;
mov.f32 %f631, 0fB3A22168;
fma.rn.f32 %f632, %f628, %f631, %f630;
mov.f32 %f633, 0fA7C234C5;
fma.rn.f32 %f1518, %f628, %f633, %f632;
abs.f32 %f162, %f160;
setp.ltu.f32 %p43, %f162, 0f47CE4780;
mov.u32 %r491, %r494;
mov.f32 %f1515, %f1518;
@%p43 bra $L__BB0_35;
setp.eq.f32 %p44, %f162, 0f7F800000;
@%p44 bra $L__BB0_34;
bra.uni $L__BB0_29;
$L__BB0_34:
mov.f32 %f636, 0f00000000;
mul.rn.f32 %f1515, %f160, %f636;
mov.u32 %r491, 0;
bra.uni $L__BB0_35;
$L__BB0_29:
mov.b32 %r19, %f160;
bfe.u32 %r206, %r19, 23, 8;
add.s32 %r20, %r206, -128;
shl.b32 %r207, %r19, 8;
or.b32 %r21, %r207, -2147483648;
shr.u32 %r22, %r20, 5;
mov.u64 %rd124, 0;
mov.u32 %r488, 0;
mov.u64 %rd122, %rd1;
mov.u64 %rd123, %rd51;
$L__BB0_30:
.pragma "nounroll";
ld.global.nc.u32 %r208, [%rd123];
mad.wide.u32 %rd53, %r208, %r21, %rd124;
shr.u64 %rd124, %rd53, 32;
st.local.u32 [%rd122], %rd53;
add.s64 %rd123, %rd123, 4;
add.s64 %rd122, %rd122, 4;
add.s32 %r488, %r488, 1;
setp.ne.s32 %p45, %r488, 6;
@%p45 bra $L__BB0_30;
st.local.u32 [%rd8], %rd124;
mov.u32 %r209, 4;
sub.s32 %r25, %r209, %r22;
mov.u32 %r210, 6;
sub.s32 %r211, %r210, %r22;
mul.wide.s32 %rd54, %r211, 4;
add.s64 %rd55, %rd1, %rd54;
ld.local.u32 %r489, [%rd55];
ld.local.u32 %r490, [%rd55+-4];
and.b32 %r28, %r20, 31;
setp.eq.s32 %p46, %r28, 0;
@%p46 bra $L__BB0_33;
mov.u32 %r212, 32;
sub.s32 %r213, %r212, %r28;
shr.u32 %r214, %r490, %r213;
shl.b32 %r215, %r489, %r28;
add.s32 %r489, %r214, %r215;
mul.wide.s32 %rd56, %r25, 4;
add.s64 %rd57, %rd1, %rd56;
ld.local.u32 %r216, [%rd57];
shr.u32 %r217, %r216, %r213;
shl.b32 %r218, %r490, %r28;
add.s32 %r490, %r217, %r218;
$L__BB0_33:
and.b32 %r219, %r19, -2147483648;
shr.u32 %r220, %r490, 30;
shl.b32 %r221, %r489, 2;
or.b32 %r222, %r220, %r221;
shr.u32 %r223, %r222, 31;
shr.u32 %r224, %r489, 30;
add.s32 %r225, %r223, %r224;
neg.s32 %r226, %r225;
setp.eq.s32 %p47, %r219, 0;
selp.b32 %r491, %r225, %r226, %p47;
setp.ne.s32 %p48, %r223, 0;
xor.b32 %r227, %r219, -2147483648;
selp.b32 %r228, %r227, %r219, %p48;
selp.b32 %r229, -1, 0, %p48;
xor.b32 %r230, %r222, %r229;
shl.b32 %r231, %r490, 2;
xor.b32 %r232, %r231, %r229;
cvt.u64.u32 %rd58, %r230;
cvt.u64.u32 %rd59, %r232;
bfi.b64 %rd60, %rd58, %rd59, 32, 32;
cvt.rn.f64.s64 %fd1, %rd60;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f634, %fd2;
setp.eq.s32 %p49, %r228, 0;
neg.f32 %f635, %f634;
selp.f32 %f1515, %f634, %f635, %p49;
$L__BB0_35:
add.s32 %r35, %r491, 1;
and.b32 %r36, %r35, 1;
setp.eq.s32 %p50, %r36, 0;
selp.f32 %f166, %f1515, 0f3F800000, %p50;
mul.rn.f32 %f167, %f1515, %f1515;
mov.f32 %f1516, 0fB94D4153;
@%p50 bra $L__BB0_37;
mov.f32 %f638, 0fBAB607ED;
mov.f32 %f639, 0f37CBAC00;
fma.rn.f32 %f1516, %f639, %f167, %f638;
$L__BB0_37:
selp.f32 %f640, 0f3C0885E4, 0f3D2AAABB, %p50;
fma.rn.f32 %f641, %f1516, %f167, %f640;
selp.f32 %f642, 0fBE2AAAA8, 0fBEFFFFFF, %p50;
fma.rn.f32 %f643, %f641, %f167, %f642;
mov.f32 %f644, 0f00000000;
fma.rn.f32 %f645, %f167, %f166, %f644;
fma.rn.f32 %f1517, %f643, %f645, %f166;
and.b32 %r234, %r35, 2;
setp.eq.s32 %p52, %r234, 0;
@%p52 bra $L__BB0_39;
mov.f32 %f647, 0fBF800000;
fma.rn.f32 %f1517, %f1517, %f647, %f644;
$L__BB0_39:
mul.f32 %f173, %f159, %f1517;
@%p43 bra $L__BB0_47;
setp.eq.f32 %p54, %f162, 0f7F800000;
@%p54 bra $L__BB0_46;
bra.uni $L__BB0_41;
$L__BB0_46:
mov.f32 %f650, 0f00000000;
mul.rn.f32 %f1518, %f160, %f650;
mov.u32 %r494, 0;
bra.uni $L__BB0_47;
$L__BB0_41:
mov.b32 %r37, %f160;
bfe.u32 %r235, %r37, 23, 8;
add.s32 %r38, %r235, -128;
shl.b32 %r236, %r37, 8;
or.b32 %r39, %r236, -2147483648;
shr.u32 %r40, %r38, 5;
mov.u64 %rd125, 0;
mov.u64 %rd126, %rd125;
$L__BB0_42:
.pragma "nounroll";
shl.b64 %rd63, %rd125, 2;
mov.u64 %rd64, __cudart_i2opi_f;
add.s64 %rd65, %rd64, %rd63;
ld.global.nc.u32 %r237, [%rd65];
mad.wide.u32 %rd66, %r237, %r39, %rd126;
shr.u64 %rd126, %rd66, 32;
add.s64 %rd67, %rd1, %rd63;
st.local.u32 [%rd67], %rd66;
cvt.u32.u64 %r238, %rd125;
add.s32 %r239, %r238, 1;
cvt.s64.s32 %rd125, %r239;
setp.ne.s32 %p55, %r239, 6;
@%p55 bra $L__BB0_42;
st.local.u32 [%rd8], %rd126;
mov.u32 %r240, 4;
sub.s32 %r41, %r240, %r40;
mov.u32 %r241, 6;
sub.s32 %r242, %r241, %r40;
mul.wide.s32 %rd68, %r242, 4;
add.s64 %rd69, %rd1, %rd68;
ld.local.u32 %r492, [%rd69];
ld.local.u32 %r493, [%rd69+-4];
and.b32 %r44, %r38, 31;
setp.eq.s32 %p56, %r44, 0;
@%p56 bra $L__BB0_45;
mov.u32 %r243, 32;
sub.s32 %r244, %r243, %r44;
shr.u32 %r245, %r493, %r244;
shl.b32 %r246, %r492, %r44;
add.s32 %r492, %r245, %r246;
mul.wide.s32 %rd70, %r41, 4;
add.s64 %rd71, %rd1, %rd70;
ld.local.u32 %r247, [%rd71];
shr.u32 %r248, %r247, %r244;
shl.b32 %r249, %r493, %r44;
add.s32 %r493, %r248, %r249;
$L__BB0_45:
and.b32 %r250, %r37, -2147483648;
shr.u32 %r251, %r493, 30;
shl.b32 %r252, %r492, 2;
or.b32 %r253, %r251, %r252;
shr.u32 %r254, %r253, 31;
shr.u32 %r255, %r492, 30;
add.s32 %r256, %r254, %r255;
neg.s32 %r257, %r256;
setp.eq.s32 %p57, %r250, 0;
selp.b32 %r494, %r256, %r257, %p57;
setp.ne.s32 %p58, %r254, 0;
xor.b32 %r258, %r250, -2147483648;
selp.b32 %r259, %r258, %r250, %p58;
selp.b32 %r260, -1, 0, %p58;
xor.b32 %r261, %r253, %r260;
shl.b32 %r262, %r493, 2;
xor.b32 %r263, %r262, %r260;
cvt.u64.u32 %rd72, %r261;
cvt.u64.u32 %rd73, %r263;
bfi.b64 %rd74, %rd72, %rd73, 32, 32;
cvt.rn.f64.s64 %fd3, %rd74;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f648, %fd4;
setp.eq.s32 %p59, %r259, 0;
neg.f32 %f649, %f648;
selp.f32 %f1518, %f648, %f649, %p59;
$L__BB0_47:
and.b32 %r51, %r494, 1;
setp.eq.s32 %p60, %r51, 0;
selp.f32 %f177, %f1518, 0f3F800000, %p60;
mul.rn.f32 %f178, %f1518, %f1518;
mov.f32 %f1519, 0fB94D4153;
@%p60 bra $L__BB0_49;
mov.f32 %f652, 0fBAB607ED;
mov.f32 %f653, 0f37CBAC00;
fma.rn.f32 %f1519, %f653, %f178, %f652;
$L__BB0_49:
selp.f32 %f654, 0f3C0885E4, 0f3D2AAABB, %p60;
fma.rn.f32 %f655, %f1519, %f178, %f654;
selp.f32 %f656, 0fBE2AAAA8, 0fBEFFFFFF, %p60;
fma.rn.f32 %f657, %f655, %f178, %f656;
mov.f32 %f1521, 0f00000000;
fma.rn.f32 %f659, %f178, %f177, %f1521;
fma.rn.f32 %f1520, %f657, %f659, %f177;
and.b32 %r265, %r494, 2;
setp.eq.s32 %p62, %r265, 0;
@%p62 bra $L__BB0_51;
mov.f32 %f660, 0f00000000;
mov.f32 %f661, 0fBF800000;
fma.rn.f32 %f1520, %f1520, %f661, %f660;
$L__BB0_51:
mov.f32 %f679, 0f38D1B717;
max.f32 %f668, %f678, %f679;
mul.f32 %f666, %f159, %f1520;
mov.f32 %f669, 0f6C4ECB8F;
mov.u32 %r299, 1;
mov.u32 %r302, 2;
mov.u32 %r304, 4;
mov.u32 %r307, 1065353216;
mov.u32 %r308, 2139095039;
mov.u32 %r336, 0;
// begin inline asm
call(%r266,%r267,%r268,%r269,%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),_optix_trace_typed_32,(%r336,%rd7,%f13,%f14,%f15,%f173,%f666,%f158,%f668,%f669,%f1521,%r299,%r336,%r336,%r302,%r336,%r304,%r307,%r307,%r307,%r308,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336,%r336);
// end inline asm
mov.b32 %f185, %r269;
setp.ge.f32 %p63, %f185, 0f00000000;
mov.f32 %f1522, %f1521;
mov.f32 %f1523, %f1521;
@%p63 bra $L__BB0_53;
mov.b32 %f1523, %r266;
mov.b32 %f1522, %r267;
mov.b32 %f1521, %r268;
$L__BB0_53:
setp.ltu.f32 %p64, %f185, 0f00000000;
selp.f32 %f680, 0f3F800000, 0f00000000, %p64;
add.f32 %f1499, %f1499, %f680;
fma.rn.f32 %f1508, %f173, %f1523, %f1508;
fma.rn.f32 %f1507, %f173, %f1522, %f1507;
fma.rn.f32 %f1506, %f173, %f1521, %f1506;
fma.rn.f32 %f1511, %f666, %f1523, %f1511;
fma.rn.f32 %f1510, %f666, %f1522, %f1510;
fma.rn.f32 %f1509, %f666, %f1521, %f1509;
fma.rn.f32 %f1514, %f158, %f1523, %f1514;
fma.rn.f32 %f1513, %f158, %f1522, %f1513;
fma.rn.f32 %f1512, %f158, %f1521, %f1512;
add.f32 %f1505, %f1505, %f1523;
add.f32 %f1504, %f1504, %f1522;
add.f32 %f1503, %f1503, %f1521;
mul.f32 %f681, %f1430, %f666;
fma.rn.f32 %f682, %f1429, %f173, %f681;
fma.rn.f32 %f683, %f1431, %f158, %f682;
cvt.sat.f32.f32 %f684, %f683;
fma.rn.f32 %f1502, %f1523, %f684, %f1502;
fma.rn.f32 %f1501, %f1522, %f684, %f1501;
fma.rn.f32 %f1500, %f1521, %f684, %f1500;
add.s32 %r486, %r486, 1;
setp.lt.s32 %p65, %r486, %r6;
@%p65 bra $L__BB0_27;
add.s32 %r484, %r484, 1;
setp.lt.s32 %p66, %r484, %r6;
@%p66 bra $L__BB0_26;
$L__BB0_55:
mul.lo.s32 %r337, %r6, %r6;
cvt.rn.f32.s32 %f685, %r337;
rcp.rn.f32 %f686, %f685;
mul.f32 %f687, %f686, %f1502;
mul.f32 %f688, %f686, %f1501;
mul.f32 %f689, %f686, %f1500;
div.rn.f32 %f1555, %f1499, %f685;
mul.f32 %f1549, %f686, %f1505;
mul.f32 %f1550, %f686, %f1504;
mul.f32 %f1551, %f686, %f1503;
mul.f32 %f1546, %f686, %f1508;
mul.f32 %f1547, %f686, %f1507;
mul.f32 %f1548, %f686, %f1506;
mul.f32 %f1543, %f686, %f1511;
mul.f32 %f1544, %f686, %f1510;
mul.f32 %f1545, %f686, %f1509;
mul.f32 %f1540, %f686, %f1514;
mul.f32 %f1541, %f686, %f1513;
mul.f32 %f1542, %f686, %f1512;
fma.rn.f32 %f1552, %f686, %f1502, %f687;
fma.rn.f32 %f1553, %f686, %f1501, %f688;
fma.rn.f32 %f1554, %f686, %f1500, %f689;
$L__BB0_56:
ld.const.u32 %r481, [params+616];
setp.lt.s32 %p67, %r481, 0;
selp.f32 %f256, %f1552, %f1555, %p67;
ld.const.u32 %r86, [params+104];
and.b32 %r338, %r86, 8;
setp.eq.s32 %p68, %r338, 0;
@%p68 bra $L__BB0_70;
ld.const.u64 %rd76, [params+192];
cvta.to.global.u64 %rd19, %rd76;
ld.const.u32 %r339, [params+184];
mad.lo.s32 %r340, %r339, %r8, %r7;
cvt.u64.u32 %rd20, %r340;
mov.f32 %f691, 0f3E68BA2E;
cvt.rzi.f32.f32 %f692, %f691;
add.f32 %f693, %f692, %f692;
mov.f32 %f694, 0f3EE8BA2E;
sub.f32 %f695, %f694, %f693;
abs.f32 %f257, %f695;
abs.f32 %f258, %f256;
setp.lt.f32 %p69, %f258, 0f00800000;
mul.f32 %f696, %f258, 0f4B800000;
selp.f32 %f697, %f696, %f258, %p69;
selp.f32 %f698, 0fC3170000, 0fC2FE0000, %p69;
mov.b32 %r341, %f697;
and.b32 %r342, %r341, 8388607;
or.b32 %r343, %r342, 1065353216;
mov.b32 %f699, %r343;
shr.u32 %r344, %r341, 23;
cvt.rn.f32.u32 %f700, %r344;
add.f32 %f701, %f698, %f700;
setp.gt.f32 %p70, %f699, 0f3FB504F3;
mul.f32 %f702, %f699, 0f3F000000;
add.f32 %f703, %f701, 0f3F800000;
selp.f32 %f704, %f703, %f701, %p70;
selp.f32 %f705, %f702, %f699, %p70;
add.f32 %f706, %f705, 0fBF800000;
add.f32 %f707, %f705, 0f3F800000;
rcp.approx.ftz.f32 %f708, %f707;
add.f32 %f709, %f706, %f706;
mul.f32 %f710, %f709, %f708;
mul.f32 %f711, %f710, %f710;
mov.f32 %f712, 0f3C4CAF63;
mov.f32 %f713, 0f3B18F0FE;
fma.rn.f32 %f714, %f713, %f711, %f712;
mov.f32 %f715, 0f3DAAAABD;
fma.rn.f32 %f716, %f714, %f711, %f715;
mul.rn.f32 %f717, %f716, %f711;
mul.rn.f32 %f718, %f717, %f710;
sub.f32 %f719, %f706, %f710;
add.f32 %f720, %f719, %f719;
neg.f32 %f721, %f710;
fma.rn.f32 %f722, %f721, %f706, %f720;
mul.rn.f32 %f723, %f708, %f722;
add.f32 %f724, %f718, %f710;
sub.f32 %f725, %f710, %f724;
add.f32 %f726, %f718, %f725;
add.f32 %f727, %f723, %f726;
add.f32 %f728, %f724, %f727;
sub.f32 %f729, %f724, %f728;
add.f32 %f730, %f727, %f729;
mov.f32 %f731, 0f3F317200;
mul.rn.f32 %f732, %f704, %f731;
mov.f32 %f733, 0f35BFBE8E;
mul.rn.f32 %f734, %f704, %f733;
add.f32 %f735, %f732, %f728;
sub.f32 %f736, %f732, %f735;
add.f32 %f737, %f728, %f736;
add.f32 %f738, %f730, %f737;
add.f32 %f739, %f734, %f738;
add.f32 %f740, %f735, %f739;
sub.f32 %f741, %f735, %f740;
add.f32 %f742, %f739, %f741;
mul.rn.f32 %f743, %f694, %f740;
neg.f32 %f744, %f743;
fma.rn.f32 %f745, %f694, %f740, %f744;
fma.rn.f32 %f746, %f694, %f742, %f745;
mov.f32 %f747, 0f00000000;
fma.rn.f32 %f748, %f747, %f740, %f746;
add.rn.f32 %f749, %f743, %f748;
neg.f32 %f750, %f749;
add.rn.f32 %f751, %f743, %f750;
add.rn.f32 %f752, %f751, %f748;
mov.b32 %r345, %f749;
setp.eq.s32 %p71, %r345, 1118925336;
add.s32 %r346, %r345, -1;
mov.b32 %f753, %r346;
add.f32 %f754, %f752, 0f37000000;
selp.f32 %f259, %f754, %f752, %p71;
selp.f32 %f755, %f753, %f749, %p71;
mov.f32 %f756, 0f3FB8AA3B;
mul.rn.f32 %f757, %f755, %f756;
cvt.rzi.f32.f32 %f758, %f757;
abs.f32 %f759, %f758;
setp.gt.f32 %p72, %f759, 0f42FC0000;
mov.b32 %r347, %f758;
and.b32 %r348, %r347, -2147483648;
or.b32 %r349, %r348, 1123811328;
mov.b32 %f760, %r349;
selp.f32 %f761, %f760, %f758, %p72;
mov.f32 %f762, 0fBF317218;
fma.rn.f32 %f763, %f761, %f762, %f755;
mov.f32 %f764, 0f3102E308;
fma.rn.f32 %f765, %f761, %f764, %f763;
mul.f32 %f766, %f765, 0f3FB8AA3B;
add.f32 %f767, %f761, 0f4B40007F;
mov.b32 %r350, %f767;
shl.b32 %r351, %r350, 23;
mov.b32 %f768, %r351;
ex2.approx.ftz.f32 %f769, %f766;
mul.f32 %f260, %f769, %f768;
setp.eq.f32 %p73, %f260, 0f7F800000;
mov.f32 %f1556, 0f7F800000;
@%p73 bra $L__BB0_59;
fma.rn.f32 %f1556, %f260, %f259, %f260;
$L__BB0_59:
setp.lt.f32 %p74, %f256, 0f00000000;
setp.eq.f32 %p75, %f257, 0f3F800000;
and.pred %p3, %p74, %p75;
setp.eq.f32 %p76, %f256, 0f00000000;
@%p76 bra $L__BB0_63;
bra.uni $L__BB0_60;
$L__BB0_63:
add.f32 %f774, %f256, %f256;
selp.f32 %f1558, %f774, 0f00000000, %p75;
bra.uni $L__BB0_64;
$L__BB0_167:
mov.f32 %f1386, 0f00000000;
mov.u32 %r496, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f1386;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs112, %f1386;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs111, %f1386;}
// end inline asm
mov.u16 %rs114, 0;
st.global.v4.u16 [%rd28], {%rs111, %rs112, %rs113, %rs114};
$L__BB0_168:
ld.const.u64 %rd108, [params+256];
cvta.to.global.u64 %rd109, %rd108;
ld.const.u32 %r472, [params+248];
mad.lo.s32 %r473, %r472, %r8, %r7;
mul.wide.u32 %rd110, %r473, 8;
add.s64 %rd29, %rd109, %rd110;
setp.eq.s32 %p194, %r496, 0;
@%p194 bra $L__BB0_170;
ld.global.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd29];
// begin inline asm
{ cvt.f32.f16 %f1387, %rs121;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1388, %rs122;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1389, %rs123;}
// end inline asm
add.f32 %f1390, %f1387, 0f00000000;
add.f32 %f1391, %f1388, 0f00000000;
add.f32 %f1392, %f1389, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs120, %f1392;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs119, %f1391;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs118, %f1390;}
// end inline asm
mov.u16 %rs125, 0;
st.global.v4.u16 [%rd29], {%rs118, %rs119, %rs120, %rs125};
bra.uni $L__BB0_171;
$L__BB0_170:
mov.f32 %f1395, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1395;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs127, %f1395;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs126, %f1395;}
// end inline asm
mov.u16 %rs129, 0;
st.global.v4.u16 [%rd29], {%rs126, %rs127, %rs128, %rs129};
$L__BB0_171:
ld.const.u64 %rd111, [params+272];
cvta.to.global.u64 %rd112, %rd111;
ld.const.u32 %r474, [params+264];
mad.lo.s32 %r475, %r474, %r8, %r7;
mul.wide.u32 %rd113, %r475, 8;
add.s64 %rd30, %rd112, %rd113;
@%p194 bra $L__BB0_173;
ld.global.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd30];
// begin inline asm
{ cvt.f32.f16 %f1396, %rs136;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1397, %rs137;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1398, %rs138;}
// end inline asm
add.f32 %f1399, %f1396, 0f00000000;
add.f32 %f1400, %f1397, 0f00000000;
add.f32 %f1401, %f1398, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs135, %f1401;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs134, %f1400;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs133, %f1399;}
// end inline asm
mov.u16 %rs140, 0;
st.global.v4.u16 [%rd30], {%rs133, %rs134, %rs135, %rs140};
bra.uni $L__BB0_174;
$L__BB0_173:
mov.f32 %f1404, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1404;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs142, %f1404;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs141, %f1404;}
// end inline asm
mov.u16 %rs144, 0;
st.global.v4.u16 [%rd30], {%rs141, %rs142, %rs143, %rs144};
$L__BB0_174:
ld.const.u64 %rd114, [params+288];
cvta.to.global.u64 %rd115, %rd114;
ld.const.u32 %r476, [params+280];
mad.lo.s32 %r477, %r476, %r8, %r7;
mul.wide.u32 %rd116, %r477, 8;
add.s64 %rd31, %rd115, %rd116;
@%p194 bra $L__BB0_176;
ld.global.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd31];
// begin inline asm
{ cvt.f32.f16 %f1405, %rs151;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1406, %rs152;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1407, %rs153;}
// end inline asm
add.f32 %f1408, %f1405, 0f00000000;
add.f32 %f1409, %f1406, 0f00000000;
add.f32 %f1410, %f1407, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs150, %f1410;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs149, %f1409;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs148, %f1408;}
// end inline asm
mov.u16 %rs155, 0;
st.global.v4.u16 [%rd31], {%rs148, %rs149, %rs150, %rs155};
bra.uni $L__BB0_177;
$L__BB0_176:
mov.f32 %f1413, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1413;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs157, %f1413;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs156, %f1413;}
// end inline asm
mov.u16 %rs159, 0;
st.global.v4.u16 [%rd31], {%rs156, %rs157, %rs158, %rs159};
$L__BB0_177:
ld.const.u64 %rd117, [params+304];
cvta.to.global.u64 %rd118, %rd117;
ld.const.u32 %r478, [params+296];
mad.lo.s32 %r479, %r478, %r8, %r7;
mul.wide.u32 %rd119, %r479, 8;
add.s64 %rd32, %rd118, %rd119;
@%p194 bra $L__BB0_179;
ld.global.v4.u16 {%rs166, %rs167, %rs168, %rs169}, [%rd32];
// begin inline asm
{ cvt.f32.f16 %f1414, %rs166;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1415, %rs167;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1416, %rs168;}
// end inline asm
add.f32 %f1417, %f1414, 0f00000000;
add.f32 %f1418, %f1415, 0f00000000;
add.f32 %f1419, %f1416, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs165, %f1419;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs164, %f1418;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs163, %f1417;}
// end inline asm
mov.u16 %rs170, 0;
st.global.v4.u16 [%rd32], {%rs163, %rs164, %rs165, %rs170};
bra.uni $L__BB0_180;
$L__BB0_179:
mov.f32 %f1422, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs173, %f1422;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs172, %f1422;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs171, %f1422;}
// end inline asm
mov.u16 %rs174, 0;
st.global.v4.u16 [%rd32], {%rs171, %rs172, %rs173, %rs174};
bra.uni $L__BB0_180;
$L__BB0_60:
mov.b32 %r352, %f1556;
xor.b32 %r353, %r352, -2147483648;
mov.b32 %f770, %r353;
selp.f32 %f1558, %f770, %f1556, %p3;
setp.geu.f32 %p77, %f256, 0f00000000;
@%p77 bra $L__BB0_64;
mov.f32 %f771, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f772, %f771;
setp.eq.f32 %p78, %f772, 0f3EE8BA2E;
@%p78 bra $L__BB0_64;
mov.f32 %f1558, 0f7FFFFFFF;
$L__BB0_64:
add.f32 %f775, %f258, 0f3EE8BA2E;
mov.b32 %r354, %f775;
setp.lt.s32 %p80, %r354, 2139095040;
@%p80 bra $L__BB0_69;
setp.gtu.f32 %p81, %f258, 0f7F800000;
@%p81 bra $L__BB0_68;
bra.uni $L__BB0_66;
$L__BB0_68:
add.f32 %f1558, %f256, 0f3EE8BA2E;
bra.uni $L__BB0_69;
$L__BB0_66:
setp.neu.f32 %p82, %f258, 0f7F800000;
@%p82 bra $L__BB0_69;
selp.f32 %f1558, 0fFF800000, 0f7F800000, %p3;
$L__BB0_69:
mul.f32 %f776, %f1558, 0f437F0000;
setp.eq.f32 %p83, %f256, 0f3F800000;
selp.f32 %f777, 0f437F0000, %f776, %p83;
cvt.rzi.u32.f32 %r355, %f777;
shl.b64 %rd77, %rd20, 1;
add.s64 %rd78, %rd19, %rd77;
cvt.u16.u32 %rs16, %r355;
mov.u16 %rs17, 255;
st.global.v2.u8 [%rd78], {%rs16, %rs17};
$L__BB0_70:
ld.const.v2.f32 {%f778, %f779}, [params+648];
mul.f32 %f271, %f1552, %f778;
mul.f32 %f272, %f1553, %f779;
ld.const.f32 %f273, [params+656];
mul.f32 %f274, %f1554, %f273;
and.b32 %r356, %r86, 1;
setp.eq.b32 %p84, %r356, 1;
mov.pred %p85, 0;
xor.pred %p86, %p84, %p85;
not.pred %p87, %p86;
@%p87 bra $L__BB0_144;
mov.f32 %f781, 0f3E666666;
cvt.rzi.f32.f32 %f782, %f781;
add.f32 %f783, %f782, %f782;
mov.f32 %f784, 0f3EE66666;
sub.f32 %f785, %f784, %f783;
abs.f32 %f275, %f785;
abs.f32 %f276, %f271;
setp.lt.f32 %p88, %f276, 0f00800000;
mul.f32 %f786, %f276, 0f4B800000;
selp.f32 %f787, %f786, %f276, %p88;
selp.f32 %f788, 0fC3170000, 0fC2FE0000, %p88;
mov.b32 %r357, %f787;
and.b32 %r358, %r357, 8388607;
or.b32 %r359, %r358, 1065353216;
mov.b32 %f789, %r359;
shr.u32 %r360, %r357, 23;
cvt.rn.f32.u32 %f790, %r360;
add.f32 %f791, %f788, %f790;
setp.gt.f32 %p89, %f789, 0f3FB504F3;
mul.f32 %f792, %f789, 0f3F000000;
add.f32 %f793, %f791, 0f3F800000;
selp.f32 %f794, %f793, %f791, %p89;
selp.f32 %f795, %f792, %f789, %p89;
add.f32 %f796, %f795, 0fBF800000;
add.f32 %f797, %f795, 0f3F800000;
rcp.approx.ftz.f32 %f798, %f797;
add.f32 %f799, %f796, %f796;
mul.f32 %f800, %f799, %f798;
mul.f32 %f801, %f800, %f800;
mov.f32 %f802, 0f3C4CAF63;
mov.f32 %f803, 0f3B18F0FE;
fma.rn.f32 %f804, %f803, %f801, %f802;
mov.f32 %f805, 0f3DAAAABD;
fma.rn.f32 %f806, %f804, %f801, %f805;
mul.rn.f32 %f807, %f806, %f801;
mul.rn.f32 %f808, %f807, %f800;
sub.f32 %f809, %f796, %f800;
add.f32 %f810, %f809, %f809;
neg.f32 %f811, %f800;
fma.rn.f32 %f812, %f811, %f796, %f810;
mul.rn.f32 %f813, %f798, %f812;
add.f32 %f814, %f808, %f800;
sub.f32 %f815, %f800, %f814;
add.f32 %f816, %f808, %f815;
add.f32 %f817, %f813, %f816;
add.f32 %f818, %f814, %f817;
sub.f32 %f819, %f814, %f818;
add.f32 %f820, %f817, %f819;
mov.f32 %f821, 0f3F317200;
mul.rn.f32 %f822, %f794, %f821;
mov.f32 %f823, 0f35BFBE8E;
mul.rn.f32 %f824, %f794, %f823;
add.f32 %f825, %f822, %f818;
sub.f32 %f826, %f822, %f825;
add.f32 %f827, %f818, %f826;
add.f32 %f828, %f820, %f827;
add.f32 %f829, %f824, %f828;
add.f32 %f830, %f825, %f829;
sub.f32 %f831, %f825, %f830;
add.f32 %f832, %f829, %f831;
mul.rn.f32 %f833, %f784, %f830;
neg.f32 %f834, %f833;
fma.rn.f32 %f835, %f784, %f830, %f834;
fma.rn.f32 %f836, %f784, %f832, %f835;
mov.f32 %f837, 0f00000000;
fma.rn.f32 %f838, %f837, %f830, %f836;
add.rn.f32 %f839, %f833, %f838;
neg.f32 %f840, %f839;
add.rn.f32 %f841, %f833, %f840;
add.rn.f32 %f842, %f841, %f838;
mov.b32 %r361, %f839;
setp.eq.s32 %p90, %r361, 1118925336;
add.s32 %r362, %r361, -1;
mov.b32 %f843, %r362;
add.f32 %f844, %f842, 0f37000000;
selp.f32 %f277, %f844, %f842, %p90;
selp.f32 %f845, %f843, %f839, %p90;
mov.f32 %f846, 0f3FB8AA3B;
mul.rn.f32 %f847, %f845, %f846;
cvt.rzi.f32.f32 %f848, %f847;
abs.f32 %f849, %f848;
setp.gt.f32 %p91, %f849, 0f42FC0000;
mov.b32 %r363, %f848;
and.b32 %r364, %r363, -2147483648;
or.b32 %r365, %r364, 1123811328;
mov.b32 %f850, %r365;
selp.f32 %f851, %f850, %f848, %p91;
mov.f32 %f852, 0fBF317218;
fma.rn.f32 %f853, %f851, %f852, %f845;
mov.f32 %f854, 0f3102E308;
fma.rn.f32 %f855, %f851, %f854, %f853;
mul.f32 %f856, %f855, 0f3FB8AA3B;
add.f32 %f857, %f851, 0f4B40007F;
mov.b32 %r366, %f857;
shl.b32 %r367, %r366, 23;
mov.b32 %f858, %r367;
ex2.approx.ftz.f32 %f859, %f856;
mul.f32 %f278, %f859, %f858;
setp.eq.f32 %p92, %f278, 0f7F800000;
mov.f32 %f1559, 0f7F800000;
@%p92 bra $L__BB0_73;
fma.rn.f32 %f1559, %f278, %f277, %f278;
$L__BB0_73:
setp.lt.f32 %p93, %f271, 0f00000000;
setp.eq.f32 %p94, %f275, 0f3F800000;
and.pred %p4, %p93, %p94;
setp.eq.f32 %p95, %f271, 0f00000000;
@%p95 bra $L__BB0_77;
bra.uni $L__BB0_74;
$L__BB0_77:
add.f32 %f864, %f271, %f271;
selp.f32 %f1561, %f864, 0f00000000, %p94;
bra.uni $L__BB0_78;
$L__BB0_74:
mov.b32 %r368, %f1559;
xor.b32 %r369, %r368, -2147483648;
mov.b32 %f860, %r369;
selp.f32 %f1561, %f860, %f1559, %p4;
setp.geu.f32 %p96, %f271, 0f00000000;
@%p96 bra $L__BB0_78;
mov.f32 %f861, 0f3EE66666;
cvt.rzi.f32.f32 %f862, %f861;
setp.eq.f32 %p97, %f862, 0f3EE66666;
@%p97 bra $L__BB0_78;
mov.f32 %f1561, 0f7FFFFFFF;
$L__BB0_78:
add.f32 %f865, %f276, 0f3EE66666;
mov.b32 %r370, %f865;
setp.lt.s32 %p99, %r370, 2139095040;
@%p99 bra $L__BB0_83;
setp.gtu.f32 %p100, %f276, 0f7F800000;
@%p100 bra $L__BB0_82;
bra.uni $L__BB0_80;
$L__BB0_82:
add.f32 %f1561, %f271, 0f3EE66666;
bra.uni $L__BB0_83;
$L__BB0_80:
setp.neu.f32 %p101, %f276, 0f7F800000;
@%p101 bra $L__BB0_83;
selp.f32 %f1561, 0fFF800000, 0f7F800000, %p4;
$L__BB0_83:
setp.eq.f32 %p102, %f271, 0f3F800000;
selp.f32 %f287, 0f3F800000, %f1561, %p102;
abs.f32 %f288, %f272;
setp.lt.f32 %p103, %f288, 0f00800000;
mul.f32 %f867, %f288, 0f4B800000;
selp.f32 %f868, %f867, %f288, %p103;
selp.f32 %f869, 0fC3170000, 0fC2FE0000, %p103;
mov.b32 %r371, %f868;
and.b32 %r372, %r371, 8388607;
or.b32 %r373, %r372, 1065353216;
mov.b32 %f870, %r373;
shr.u32 %r374, %r371, 23;
cvt.rn.f32.u32 %f871, %r374;
add.f32 %f872, %f869, %f871;
setp.gt.f32 %p104, %f870, 0f3FB504F3;
mul.f32 %f873, %f870, 0f3F000000;
add.f32 %f874, %f872, 0f3F800000;
selp.f32 %f875, %f874, %f872, %p104;
selp.f32 %f876, %f873, %f870, %p104;
add.f32 %f877, %f876, 0fBF800000;
add.f32 %f878, %f876, 0f3F800000;
rcp.approx.ftz.f32 %f879, %f878;
add.f32 %f880, %f877, %f877;
mul.f32 %f881, %f880, %f879;
mul.f32 %f882, %f881, %f881;
mov.f32 %f883, 0f3C4CAF63;
mov.f32 %f884, 0f3B18F0FE;
fma.rn.f32 %f885, %f884, %f882, %f883;
mov.f32 %f886, 0f3DAAAABD;
fma.rn.f32 %f887, %f885, %f882, %f886;
mul.rn.f32 %f888, %f887, %f882;
mul.rn.f32 %f889, %f888, %f881;
sub.f32 %f890, %f877, %f881;
add.f32 %f891, %f890, %f890;
neg.f32 %f892, %f881;
fma.rn.f32 %f893, %f892, %f877, %f891;
mul.rn.f32 %f894, %f879, %f893;
add.f32 %f895, %f889, %f881;
sub.f32 %f896, %f881, %f895;
add.f32 %f897, %f889, %f896;
add.f32 %f898, %f894, %f897;
add.f32 %f899, %f895, %f898;
sub.f32 %f900, %f895, %f899;
add.f32 %f901, %f898, %f900;
mov.f32 %f902, 0f3F317200;
mul.rn.f32 %f903, %f875, %f902;
mov.f32 %f904, 0f35BFBE8E;
mul.rn.f32 %f905, %f875, %f904;
add.f32 %f906, %f903, %f899;
sub.f32 %f907, %f903, %f906;
add.f32 %f908, %f899, %f907;
add.f32 %f909, %f901, %f908;
add.f32 %f910, %f905, %f909;
add.f32 %f911, %f906, %f910;
sub.f32 %f912, %f906, %f911;
add.f32 %f913, %f910, %f912;
mov.f32 %f914, 0f3EE66666;
mul.rn.f32 %f915, %f914, %f911;
neg.f32 %f916, %f915;
fma.rn.f32 %f917, %f914, %f911, %f916;
fma.rn.f32 %f918, %f914, %f913, %f917;
mov.f32 %f919, 0f00000000;
fma.rn.f32 %f920, %f919, %f911, %f918;
add.rn.f32 %f921, %f915, %f920;
neg.f32 %f922, %f921;
add.rn.f32 %f923, %f915, %f922;
add.rn.f32 %f924, %f923, %f920;
mov.b32 %r375, %f921;
setp.eq.s32 %p105, %r375, 1118925336;
add.s32 %r376, %r375, -1;
mov.b32 %f925, %r376;
add.f32 %f926, %f924, 0f37000000;
selp.f32 %f289, %f926, %f924, %p105;
selp.f32 %f927, %f925, %f921, %p105;
mov.f32 %f928, 0f3FB8AA3B;
mul.rn.f32 %f929, %f927, %f928;
cvt.rzi.f32.f32 %f930, %f929;
abs.f32 %f931, %f930;
setp.gt.f32 %p106, %f931, 0f42FC0000;
mov.b32 %r377, %f930;
and.b32 %r378, %r377, -2147483648;
or.b32 %r379, %r378, 1123811328;
mov.b32 %f932, %r379;
selp.f32 %f933, %f932, %f930, %p106;
mov.f32 %f934, 0fBF317218;
fma.rn.f32 %f935, %f933, %f934, %f927;
mov.f32 %f936, 0f3102E308;
fma.rn.f32 %f937, %f933, %f936, %f935;
mul.f32 %f938, %f937, 0f3FB8AA3B;
add.f32 %f939, %f933, 0f4B40007F;
mov.b32 %r380, %f939;
shl.b32 %r381, %r380, 23;
mov.b32 %f940, %r381;
ex2.approx.ftz.f32 %f941, %f938;
mul.f32 %f290, %f941, %f940;
setp.eq.f32 %p107, %f290, 0f7F800000;
mov.f32 %f1562, 0f7F800000;
@%p107 bra $L__BB0_85;
fma.rn.f32 %f1562, %f290, %f289, %f290;
$L__BB0_85:
setp.lt.f32 %p108, %f272, 0f00000000;
and.pred %p5, %p108, %p94;
setp.eq.f32 %p110, %f272, 0f00000000;
@%p110 bra $L__BB0_89;
bra.uni $L__BB0_86;
$L__BB0_89:
add.f32 %f946, %f272, %f272;
selp.f32 %f1564, %f946, 0f00000000, %p94;
bra.uni $L__BB0_90;
$L__BB0_86:
mov.b32 %r382, %f1562;
xor.b32 %r383, %r382, -2147483648;
mov.b32 %f942, %r383;
selp.f32 %f1564, %f942, %f1562, %p5;
setp.geu.f32 %p111, %f272, 0f00000000;
@%p111 bra $L__BB0_90;
mov.f32 %f943, 0f3EE66666;
cvt.rzi.f32.f32 %f944, %f943;
setp.eq.f32 %p112, %f944, 0f3EE66666;
@%p112 bra $L__BB0_90;
mov.f32 %f1564, 0f7FFFFFFF;
$L__BB0_90:
add.f32 %f947, %f288, 0f3EE66666;
mov.b32 %r384, %f947;
setp.lt.s32 %p114, %r384, 2139095040;
@%p114 bra $L__BB0_95;
setp.gtu.f32 %p115, %f288, 0f7F800000;
@%p115 bra $L__BB0_94;
bra.uni $L__BB0_92;
$L__BB0_94:
add.f32 %f1564, %f272, 0f3EE66666;
bra.uni $L__BB0_95;
$L__BB0_92:
setp.neu.f32 %p116, %f288, 0f7F800000;
@%p116 bra $L__BB0_95;
selp.f32 %f1564, 0fFF800000, 0f7F800000, %p5;
$L__BB0_95:
setp.eq.f32 %p117, %f272, 0f3F800000;
selp.f32 %f299, 0f3F800000, %f1564, %p117;
abs.f32 %f300, %f274;
setp.lt.f32 %p118, %f300, 0f00800000;
mul.f32 %f949, %f300, 0f4B800000;
selp.f32 %f950, %f949, %f300, %p118;
selp.f32 %f951, 0fC3170000, 0fC2FE0000, %p118;
mov.b32 %r385, %f950;
and.b32 %r386, %r385, 8388607;
or.b32 %r387, %r386, 1065353216;
mov.b32 %f952, %r387;
shr.u32 %r388, %r385, 23;
cvt.rn.f32.u32 %f953, %r388;
add.f32 %f954, %f951, %f953;
setp.gt.f32 %p119, %f952, 0f3FB504F3;
mul.f32 %f955, %f952, 0f3F000000;
add.f32 %f956, %f954, 0f3F800000;
selp.f32 %f957, %f956, %f954, %p119;
selp.f32 %f958, %f955, %f952, %p119;
add.f32 %f959, %f958, 0fBF800000;
add.f32 %f960, %f958, 0f3F800000;
rcp.approx.ftz.f32 %f961, %f960;
add.f32 %f962, %f959, %f959;
mul.f32 %f963, %f962, %f961;
mul.f32 %f964, %f963, %f963;
mov.f32 %f965, 0f3C4CAF63;
mov.f32 %f966, 0f3B18F0FE;
fma.rn.f32 %f967, %f966, %f964, %f965;
mov.f32 %f968, 0f3DAAAABD;
fma.rn.f32 %f969, %f967, %f964, %f968;
mul.rn.f32 %f970, %f969, %f964;
mul.rn.f32 %f971, %f970, %f963;
sub.f32 %f972, %f959, %f963;
add.f32 %f973, %f972, %f972;
neg.f32 %f974, %f963;
fma.rn.f32 %f975, %f974, %f959, %f973;
mul.rn.f32 %f976, %f961, %f975;
add.f32 %f977, %f971, %f963;
sub.f32 %f978, %f963, %f977;
add.f32 %f979, %f971, %f978;
add.f32 %f980, %f976, %f979;
add.f32 %f981, %f977, %f980;
sub.f32 %f982, %f977, %f981;
add.f32 %f983, %f980, %f982;
mov.f32 %f984, 0f3F317200;
mul.rn.f32 %f985, %f957, %f984;
mov.f32 %f986, 0f35BFBE8E;
mul.rn.f32 %f987, %f957, %f986;
add.f32 %f988, %f985, %f981;
sub.f32 %f989, %f985, %f988;
add.f32 %f990, %f981, %f989;
add.f32 %f991, %f983, %f990;
add.f32 %f992, %f987, %f991;
add.f32 %f993, %f988, %f992;
sub.f32 %f994, %f988, %f993;
add.f32 %f995, %f992, %f994;
mov.f32 %f996, 0f3EE66666;
mul.rn.f32 %f997, %f996, %f993;
neg.f32 %f998, %f997;
fma.rn.f32 %f999, %f996, %f993, %f998;
fma.rn.f32 %f1000, %f996, %f995, %f999;
mov.f32 %f1001, 0f00000000;
fma.rn.f32 %f1002, %f1001, %f993, %f1000;
add.rn.f32 %f1003, %f997, %f1002;
neg.f32 %f1004, %f1003;
add.rn.f32 %f1005, %f997, %f1004;
add.rn.f32 %f1006, %f1005, %f1002;
mov.b32 %r389, %f1003;
setp.eq.s32 %p120, %r389, 1118925336;
add.s32 %r390, %r389, -1;
mov.b32 %f1007, %r390;
add.f32 %f1008, %f1006, 0f37000000;
selp.f32 %f301, %f1008, %f1006, %p120;
selp.f32 %f1009, %f1007, %f1003, %p120;
mov.f32 %f1010, 0f3FB8AA3B;
mul.rn.f32 %f1011, %f1009, %f1010;
cvt.rzi.f32.f32 %f1012, %f1011;
abs.f32 %f1013, %f1012;
setp.gt.f32 %p121, %f1013, 0f42FC0000;
mov.b32 %r391, %f1012;
and.b32 %r392, %r391, -2147483648;
or.b32 %r393, %r392, 1123811328;
mov.b32 %f1014, %r393;
selp.f32 %f1015, %f1014, %f1012, %p121;
mov.f32 %f1016, 0fBF317218;
fma.rn.f32 %f1017, %f1015, %f1016, %f1009;
mov.f32 %f1018, 0f3102E308;
fma.rn.f32 %f1019, %f1015, %f1018, %f1017;
mul.f32 %f1020, %f1019, 0f3FB8AA3B;
add.f32 %f1021, %f1015, 0f4B40007F;
mov.b32 %r394, %f1021;
shl.b32 %r395, %r394, 23;
mov.b32 %f1022, %r395;
ex2.approx.ftz.f32 %f1023, %f1020;
mul.f32 %f302, %f1023, %f1022;
setp.eq.f32 %p122, %f302, 0f7F800000;
mov.f32 %f1565, 0f7F800000;
@%p122 bra $L__BB0_97;
fma.rn.f32 %f1565, %f302, %f301, %f302;
$L__BB0_97:
setp.lt.f32 %p123, %f274, 0f00000000;
and.pred %p6, %p123, %p94;
setp.eq.f32 %p125, %f274, 0f00000000;
@%p125 bra $L__BB0_101;
bra.uni $L__BB0_98;
$L__BB0_101:
add.f32 %f1028, %f274, %f274;
selp.f32 %f1567, %f1028, 0f00000000, %p94;
bra.uni $L__BB0_102;
$L__BB0_98:
mov.b32 %r396, %f1565;
xor.b32 %r397, %r396, -2147483648;
mov.b32 %f1024, %r397;
selp.f32 %f1567, %f1024, %f1565, %p6;
setp.geu.f32 %p126, %f274, 0f00000000;
@%p126 bra $L__BB0_102;
mov.f32 %f1025, 0f3EE66666;
cvt.rzi.f32.f32 %f1026, %f1025;
setp.eq.f32 %p127, %f1026, 0f3EE66666;
@%p127 bra $L__BB0_102;
mov.f32 %f1567, 0f7FFFFFFF;
$L__BB0_102:
add.f32 %f1029, %f300, 0f3EE66666;
mov.b32 %r398, %f1029;
setp.lt.s32 %p129, %r398, 2139095040;
@%p129 bra $L__BB0_107;
setp.gtu.f32 %p130, %f300, 0f7F800000;
@%p130 bra $L__BB0_106;
bra.uni $L__BB0_104;
$L__BB0_106:
add.f32 %f1567, %f274, 0f3EE66666;
bra.uni $L__BB0_107;
$L__BB0_104:
setp.neu.f32 %p131, %f300, 0f7F800000;
@%p131 bra $L__BB0_107;
selp.f32 %f1567, 0fFF800000, 0f7F800000, %p6;
$L__BB0_107:
setp.eq.f32 %p132, %f274, 0f3F800000;
mov.f32 %f1031, 0f3F800000;
selp.f32 %f1032, 0f3F800000, %f1567, %p132;
ld.const.u64 %rd79, [params+144];
cvta.to.global.u64 %rd21, %rd79;
ld.const.u32 %r399, [params+136];
mad.lo.s32 %r400, %r399, %r8, %r7;
cvt.u64.u32 %rd22, %r400;
min.f32 %f1033, %f287, %f1031;
mov.f32 %f1034, 0f00000000;
max.f32 %f311, %f1034, %f1033;
min.f32 %f1035, %f299, %f1031;
max.f32 %f312, %f1034, %f1035;
min.f32 %f1036, %f1032, %f1031;
max.f32 %f313, %f1034, %f1036;
mov.f32 %f1037, 0f3E555555;
cvt.rzi.f32.f32 %f1038, %f1037;
add.f32 %f1039, %f1038, %f1038;
mov.f32 %f1040, 0f3ED55555;
sub.f32 %f1041, %f1040, %f1039;
abs.f32 %f314, %f1041;
abs.f32 %f315, %f311;
setp.lt.f32 %p133, %f315, 0f00800000;
mul.f32 %f1042, %f315, 0f4B800000;
selp.f32 %f1043, %f1042, %f315, %p133;
selp.f32 %f1044, 0fC3170000, 0fC2FE0000, %p133;
mov.b32 %r401, %f1043;
and.b32 %r402, %r401, 8388607;
or.b32 %r403, %r402, 1065353216;
mov.b32 %f1045, %r403;
shr.u32 %r404, %r401, 23;
cvt.rn.f32.u32 %f1046, %r404;
add.f32 %f1047, %f1044, %f1046;
setp.gt.f32 %p134, %f1045, 0f3FB504F3;
mul.f32 %f1048, %f1045, 0f3F000000;
add.f32 %f1049, %f1047, 0f3F800000;
selp.f32 %f1050, %f1049, %f1047, %p134;
selp.f32 %f1051, %f1048, %f1045, %p134;
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;
mul.rn.f32 %f1089, %f1040, %f1086;
neg.f32 %f1090, %f1089;
fma.rn.f32 %f1091, %f1040, %f1086, %f1090;
fma.rn.f32 %f1092, %f1040, %f1088, %f1091;
fma.rn.f32 %f1093, %f1034, %f1086, %f1092;
add.rn.f32 %f1094, %f1089, %f1093;
neg.f32 %f1095, %f1094;
add.rn.f32 %f1096, %f1089, %f1095;
add.rn.f32 %f1097, %f1096, %f1093;
mov.b32 %r405, %f1094;
setp.eq.s32 %p135, %r405, 1118925336;
add.s32 %r406, %r405, -1;
mov.b32 %f1098, %r406;
add.f32 %f1099, %f1097, 0f37000000;
selp.f32 %f316, %f1099, %f1097, %p135;
selp.f32 %f1100, %f1098, %f1094, %p135;
mov.f32 %f1101, 0f3FB8AA3B;
mul.rn.f32 %f1102, %f1100, %f1101;
cvt.rzi.f32.f32 %f1103, %f1102;
abs.f32 %f1104, %f1103;
setp.gt.f32 %p136, %f1104, 0f42FC0000;
mov.b32 %r407, %f1103;
and.b32 %r408, %r407, -2147483648;
or.b32 %r409, %r408, 1123811328;
mov.b32 %f1105, %r409;
selp.f32 %f1106, %f1105, %f1103, %p136;
mov.f32 %f1107, 0fBF317218;
fma.rn.f32 %f1108, %f1106, %f1107, %f1100;
mov.f32 %f1109, 0f3102E308;
fma.rn.f32 %f1110, %f1106, %f1109, %f1108;
mul.f32 %f1111, %f1110, 0f3FB8AA3B;
add.f32 %f1112, %f1106, 0f4B40007F;
mov.b32 %r410, %f1112;
shl.b32 %r411, %r410, 23;
mov.b32 %f1113, %r411;
ex2.approx.ftz.f32 %f1114, %f1111;
mul.f32 %f317, %f1114, %f1113;
setp.eq.f32 %p137, %f317, 0f7F800000;
mov.f32 %f1568, 0f7F800000;
@%p137 bra $L__BB0_109;
fma.rn.f32 %f1568, %f317, %f316, %f317;
$L__BB0_109:
setp.lt.f32 %p138, %f311, 0f00000000;
setp.eq.f32 %p139, %f314, 0f3F800000;
and.pred %p7, %p138, %p139;
setp.eq.f32 %p140, %f311, 0f00000000;
@%p140 bra $L__BB0_113;
bra.uni $L__BB0_110;
$L__BB0_113:
add.f32 %f1119, %f311, %f311;
selp.f32 %f1570, %f1119, 0f00000000, %p139;
bra.uni $L__BB0_114;
$L__BB0_110:
mov.b32 %r412, %f1568;
xor.b32 %r413, %r412, -2147483648;
mov.b32 %f1115, %r413;
selp.f32 %f1570, %f1115, %f1568, %p7;
setp.geu.f32 %p141, %f311, 0f00000000;
@%p141 bra $L__BB0_114;
mov.f32 %f1116, 0f3ED55555;
cvt.rzi.f32.f32 %f1117, %f1116;
setp.eq.f32 %p142, %f1117, 0f3ED55555;
@%p142 bra $L__BB0_114;
mov.f32 %f1570, 0f7FFFFFFF;
$L__BB0_114:
add.f32 %f1120, %f315, 0f3ED55555;
mov.b32 %r414, %f1120;
setp.lt.s32 %p144, %r414, 2139095040;
@%p144 bra $L__BB0_119;
setp.gtu.f32 %p145, %f315, 0f7F800000;
@%p145 bra $L__BB0_118;
bra.uni $L__BB0_116;
$L__BB0_118:
add.f32 %f1570, %f311, 0f3ED55555;
bra.uni $L__BB0_119;
$L__BB0_116:
setp.neu.f32 %p146, %f315, 0f7F800000;
@%p146 bra $L__BB0_119;
selp.f32 %f1570, 0fFF800000, 0f7F800000, %p7;
$L__BB0_119:
abs.f32 %f326, %f312;
setp.lt.f32 %p147, %f326, 0f00800000;
mul.f32 %f1122, %f326, 0f4B800000;
selp.f32 %f1123, %f1122, %f326, %p147;
selp.f32 %f1124, 0fC3170000, 0fC2FE0000, %p147;
mov.b32 %r415, %f1123;
and.b32 %r416, %r415, 8388607;
or.b32 %r417, %r416, 1065353216;
mov.b32 %f1125, %r417;
shr.u32 %r418, %r415, 23;
cvt.rn.f32.u32 %f1126, %r418;
add.f32 %f1127, %f1124, %f1126;
setp.gt.f32 %p148, %f1125, 0f3FB504F3;
mul.f32 %f1128, %f1125, 0f3F000000;
add.f32 %f1129, %f1127, 0f3F800000;
selp.f32 %f1130, %f1129, %f1127, %p148;
selp.f32 %f1131, %f1128, %f1125, %p148;
add.f32 %f1132, %f1131, 0fBF800000;
add.f32 %f1133, %f1131, 0f3F800000;
rcp.approx.ftz.f32 %f1134, %f1133;
add.f32 %f1135, %f1132, %f1132;
mul.f32 %f1136, %f1135, %f1134;
mul.f32 %f1137, %f1136, %f1136;
mov.f32 %f1138, 0f3C4CAF63;
mov.f32 %f1139, 0f3B18F0FE;
fma.rn.f32 %f1140, %f1139, %f1137, %f1138;
mov.f32 %f1141, 0f3DAAAABD;
fma.rn.f32 %f1142, %f1140, %f1137, %f1141;
mul.rn.f32 %f1143, %f1142, %f1137;
mul.rn.f32 %f1144, %f1143, %f1136;
sub.f32 %f1145, %f1132, %f1136;
add.f32 %f1146, %f1145, %f1145;
neg.f32 %f1147, %f1136;
fma.rn.f32 %f1148, %f1147, %f1132, %f1146;
mul.rn.f32 %f1149, %f1134, %f1148;
add.f32 %f1150, %f1144, %f1136;
sub.f32 %f1151, %f1136, %f1150;
add.f32 %f1152, %f1144, %f1151;
add.f32 %f1153, %f1149, %f1152;
add.f32 %f1154, %f1150, %f1153;
sub.f32 %f1155, %f1150, %f1154;
add.f32 %f1156, %f1153, %f1155;
mov.f32 %f1157, 0f3F317200;
mul.rn.f32 %f1158, %f1130, %f1157;
mov.f32 %f1159, 0f35BFBE8E;
mul.rn.f32 %f1160, %f1130, %f1159;
add.f32 %f1161, %f1158, %f1154;
sub.f32 %f1162, %f1158, %f1161;
add.f32 %f1163, %f1154, %f1162;
add.f32 %f1164, %f1156, %f1163;
add.f32 %f1165, %f1160, %f1164;
add.f32 %f1166, %f1161, %f1165;
sub.f32 %f1167, %f1161, %f1166;
add.f32 %f1168, %f1165, %f1167;
mov.f32 %f1169, 0f3ED55555;
mul.rn.f32 %f1170, %f1169, %f1166;
neg.f32 %f1171, %f1170;
fma.rn.f32 %f1172, %f1169, %f1166, %f1171;
fma.rn.f32 %f1173, %f1169, %f1168, %f1172;
mov.f32 %f1174, 0f00000000;
fma.rn.f32 %f1175, %f1174, %f1166, %f1173;
add.rn.f32 %f1176, %f1170, %f1175;
neg.f32 %f1177, %f1176;
add.rn.f32 %f1178, %f1170, %f1177;
add.rn.f32 %f1179, %f1178, %f1175;
mov.b32 %r419, %f1176;
setp.eq.s32 %p149, %r419, 1118925336;
add.s32 %r420, %r419, -1;
mov.b32 %f1180, %r420;
add.f32 %f1181, %f1179, 0f37000000;
selp.f32 %f327, %f1181, %f1179, %p149;
selp.f32 %f1182, %f1180, %f1176, %p149;
mov.f32 %f1183, 0f3FB8AA3B;
mul.rn.f32 %f1184, %f1182, %f1183;
cvt.rzi.f32.f32 %f1185, %f1184;
abs.f32 %f1186, %f1185;
setp.gt.f32 %p150, %f1186, 0f42FC0000;
mov.b32 %r421, %f1185;
and.b32 %r422, %r421, -2147483648;
or.b32 %r423, %r422, 1123811328;
mov.b32 %f1187, %r423;
selp.f32 %f1188, %f1187, %f1185, %p150;
mov.f32 %f1189, 0fBF317218;
fma.rn.f32 %f1190, %f1188, %f1189, %f1182;
mov.f32 %f1191, 0f3102E308;
fma.rn.f32 %f1192, %f1188, %f1191, %f1190;
mul.f32 %f1193, %f1192, 0f3FB8AA3B;
add.f32 %f1194, %f1188, 0f4B40007F;
mov.b32 %r424, %f1194;
shl.b32 %r425, %r424, 23;
mov.b32 %f1195, %r425;
ex2.approx.ftz.f32 %f1196, %f1193;
mul.f32 %f328, %f1196, %f1195;
setp.eq.f32 %p151, %f328, 0f7F800000;
mov.f32 %f1571, 0f7F800000;
@%p151 bra $L__BB0_121;
fma.rn.f32 %f1571, %f328, %f327, %f328;
$L__BB0_121:
setp.lt.f32 %p152, %f312, 0f00000000;
and.pred %p8, %p152, %p139;
setp.eq.f32 %p154, %f312, 0f00000000;
@%p154 bra $L__BB0_125;
bra.uni $L__BB0_122;
$L__BB0_125:
add.f32 %f1201, %f312, %f312;
selp.f32 %f1573, %f1201, 0f00000000, %p139;
bra.uni $L__BB0_126;
$L__BB0_122:
mov.b32 %r426, %f1571;
xor.b32 %r427, %r426, -2147483648;
mov.b32 %f1197, %r427;
selp.f32 %f1573, %f1197, %f1571, %p8;
setp.geu.f32 %p155, %f312, 0f00000000;
@%p155 bra $L__BB0_126;
mov.f32 %f1198, 0f3ED55555;
cvt.rzi.f32.f32 %f1199, %f1198;
setp.eq.f32 %p156, %f1199, 0f3ED55555;
@%p156 bra $L__BB0_126;
mov.f32 %f1573, 0f7FFFFFFF;
$L__BB0_126:
add.f32 %f1202, %f326, 0f3ED55555;
mov.b32 %r428, %f1202;
setp.lt.s32 %p158, %r428, 2139095040;
@%p158 bra $L__BB0_131;
setp.gtu.f32 %p159, %f326, 0f7F800000;
@%p159 bra $L__BB0_130;
bra.uni $L__BB0_128;
$L__BB0_130:
add.f32 %f1573, %f312, 0f3ED55555;
bra.uni $L__BB0_131;
$L__BB0_128:
setp.neu.f32 %p160, %f326, 0f7F800000;
@%p160 bra $L__BB0_131;
selp.f32 %f1573, 0fFF800000, 0f7F800000, %p8;
$L__BB0_131:
abs.f32 %f337, %f313;
setp.lt.f32 %p161, %f337, 0f00800000;
mul.f32 %f1204, %f337, 0f4B800000;
selp.f32 %f1205, %f1204, %f337, %p161;
selp.f32 %f1206, 0fC3170000, 0fC2FE0000, %p161;
mov.b32 %r429, %f1205;
and.b32 %r430, %r429, 8388607;
or.b32 %r431, %r430, 1065353216;
mov.b32 %f1207, %r431;
shr.u32 %r432, %r429, 23;
cvt.rn.f32.u32 %f1208, %r432;
add.f32 %f1209, %f1206, %f1208;
setp.gt.f32 %p162, %f1207, 0f3FB504F3;
mul.f32 %f1210, %f1207, 0f3F000000;
add.f32 %f1211, %f1209, 0f3F800000;
selp.f32 %f1212, %f1211, %f1209, %p162;
selp.f32 %f1213, %f1210, %f1207, %p162;
add.f32 %f1214, %f1213, 0fBF800000;
add.f32 %f1215, %f1213, 0f3F800000;
rcp.approx.ftz.f32 %f1216, %f1215;
add.f32 %f1217, %f1214, %f1214;
mul.f32 %f1218, %f1217, %f1216;
mul.f32 %f1219, %f1218, %f1218;
mov.f32 %f1220, 0f3C4CAF63;
mov.f32 %f1221, 0f3B18F0FE;
fma.rn.f32 %f1222, %f1221, %f1219, %f1220;
mov.f32 %f1223, 0f3DAAAABD;
fma.rn.f32 %f1224, %f1222, %f1219, %f1223;
mul.rn.f32 %f1225, %f1224, %f1219;
mul.rn.f32 %f1226, %f1225, %f1218;
sub.f32 %f1227, %f1214, %f1218;
add.f32 %f1228, %f1227, %f1227;
neg.f32 %f1229, %f1218;
fma.rn.f32 %f1230, %f1229, %f1214, %f1228;
mul.rn.f32 %f1231, %f1216, %f1230;
add.f32 %f1232, %f1226, %f1218;
sub.f32 %f1233, %f1218, %f1232;
add.f32 %f1234, %f1226, %f1233;
add.f32 %f1235, %f1231, %f1234;
add.f32 %f1236, %f1232, %f1235;
sub.f32 %f1237, %f1232, %f1236;
add.f32 %f1238, %f1235, %f1237;
mov.f32 %f1239, 0f3F317200;
mul.rn.f32 %f1240, %f1212, %f1239;
mov.f32 %f1241, 0f35BFBE8E;
mul.rn.f32 %f1242, %f1212, %f1241;
add.f32 %f1243, %f1240, %f1236;
sub.f32 %f1244, %f1240, %f1243;
add.f32 %f1245, %f1236, %f1244;
add.f32 %f1246, %f1238, %f1245;
add.f32 %f1247, %f1242, %f1246;
add.f32 %f1248, %f1243, %f1247;
sub.f32 %f1249, %f1243, %f1248;
add.f32 %f1250, %f1247, %f1249;
mov.f32 %f1251, 0f3ED55555;
mul.rn.f32 %f1252, %f1251, %f1248;
neg.f32 %f1253, %f1252;
fma.rn.f32 %f1254, %f1251, %f1248, %f1253;
fma.rn.f32 %f1255, %f1251, %f1250, %f1254;
mov.f32 %f1256, 0f00000000;
fma.rn.f32 %f1257, %f1256, %f1248, %f1255;
add.rn.f32 %f1258, %f1252, %f1257;
neg.f32 %f1259, %f1258;
add.rn.f32 %f1260, %f1252, %f1259;
add.rn.f32 %f1261, %f1260, %f1257;
mov.b32 %r433, %f1258;
setp.eq.s32 %p163, %r433, 1118925336;
add.s32 %r434, %r433, -1;
mov.b32 %f1262, %r434;
add.f32 %f1263, %f1261, 0f37000000;
selp.f32 %f338, %f1263, %f1261, %p163;
selp.f32 %f1264, %f1262, %f1258, %p163;
mov.f32 %f1265, 0f3FB8AA3B;
mul.rn.f32 %f1266, %f1264, %f1265;
cvt.rzi.f32.f32 %f1267, %f1266;
abs.f32 %f1268, %f1267;
setp.gt.f32 %p164, %f1268, 0f42FC0000;
mov.b32 %r435, %f1267;
and.b32 %r436, %r435, -2147483648;
or.b32 %r437, %r436, 1123811328;
mov.b32 %f1269, %r437;
selp.f32 %f1270, %f1269, %f1267, %p164;
mov.f32 %f1271, 0fBF317218;
fma.rn.f32 %f1272, %f1270, %f1271, %f1264;
mov.f32 %f1273, 0f3102E308;
fma.rn.f32 %f1274, %f1270, %f1273, %f1272;
mul.f32 %f1275, %f1274, 0f3FB8AA3B;
add.f32 %f1276, %f1270, 0f4B40007F;
mov.b32 %r438, %f1276;
shl.b32 %r439, %r438, 23;
mov.b32 %f1277, %r439;
ex2.approx.ftz.f32 %f1278, %f1275;
mul.f32 %f339, %f1278, %f1277;
setp.eq.f32 %p165, %f339, 0f7F800000;
mov.f32 %f1574, 0f7F800000;
@%p165 bra $L__BB0_133;
fma.rn.f32 %f1574, %f339, %f338, %f339;
$L__BB0_133:
setp.lt.f32 %p166, %f313, 0f00000000;
and.pred %p9, %p166, %p139;
setp.eq.f32 %p168, %f313, 0f00000000;
@%p168 bra $L__BB0_137;
bra.uni $L__BB0_134;
$L__BB0_137:
add.f32 %f1283, %f313, %f313;
selp.f32 %f1576, %f1283, 0f00000000, %p139;
bra.uni $L__BB0_138;
$L__BB0_134:
mov.b32 %r440, %f1574;
xor.b32 %r441, %r440, -2147483648;
mov.b32 %f1279, %r441;
selp.f32 %f1576, %f1279, %f1574, %p9;
setp.geu.f32 %p169, %f313, 0f00000000;
@%p169 bra $L__BB0_138;
mov.f32 %f1280, 0f3ED55555;
cvt.rzi.f32.f32 %f1281, %f1280;
setp.eq.f32 %p170, %f1281, 0f3ED55555;
@%p170 bra $L__BB0_138;
mov.f32 %f1576, 0f7FFFFFFF;
$L__BB0_138:
add.f32 %f1284, %f337, 0f3ED55555;
mov.b32 %r442, %f1284;
setp.lt.s32 %p172, %r442, 2139095040;
@%p172 bra $L__BB0_143;
setp.gtu.f32 %p173, %f337, 0f7F800000;
@%p173 bra $L__BB0_142;
bra.uni $L__BB0_140;
$L__BB0_142:
add.f32 %f1576, %f313, 0f3ED55555;
bra.uni $L__BB0_143;
$L__BB0_140:
setp.neu.f32 %p174, %f337, 0f7F800000;
@%p174 bra $L__BB0_143;
selp.f32 %f1576, 0fFF800000, 0f7F800000, %p9;
$L__BB0_143:
fma.rn.f32 %f1285, %f1570, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p175, %f311, 0f3F800000;
mov.f32 %f1286, 0f3F800000;
selp.f32 %f1287, 0f3F7FFFFF, %f1285, %p175;
mul.f32 %f1288, %f311, 0f414EB852;
setp.lt.f32 %p176, %f311, 0f3B4D2E1C;
selp.f32 %f1289, %f1288, %f1287, %p176;
fma.rn.f32 %f1290, %f1573, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p177, %f312, 0f3F800000;
selp.f32 %f1291, 0f3F7FFFFF, %f1290, %p177;
mul.f32 %f1292, %f312, 0f414EB852;
setp.lt.f32 %p178, %f312, 0f3B4D2E1C;
selp.f32 %f1293, %f1292, %f1291, %p178;
fma.rn.f32 %f1294, %f1576, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p179, %f313, 0f3F800000;
selp.f32 %f1295, 0f3F7FFFFF, %f1294, %p179;
mul.f32 %f1296, %f313, 0f414EB852;
setp.lt.f32 %p180, %f313, 0f3B4D2E1C;
selp.f32 %f1297, %f1296, %f1295, %p180;
min.f32 %f1298, %f1289, %f1286;
mov.f32 %f1299, 0f00000000;
max.f32 %f1300, %f1299, %f1298;
mul.f32 %f1301, %f1300, 0f43800000;
cvt.rzi.u32.f32 %r443, %f1301;
min.u32 %r444, %r443, 255;
min.f32 %f1302, %f1293, %f1286;
max.f32 %f1303, %f1299, %f1302;
mul.f32 %f1304, %f1303, 0f43800000;
cvt.rzi.u32.f32 %r445, %f1304;
min.u32 %r446, %r445, 255;
min.f32 %f1305, %f1297, %f1286;
max.f32 %f1306, %f1299, %f1305;
mul.f32 %f1307, %f1306, 0f43800000;
cvt.rzi.u32.f32 %r447, %f1307;
min.u32 %r448, %r447, 255;
shl.b64 %rd80, %rd22, 2;
add.s64 %rd81, %rd21, %rd80;
cvt.u16.u32 %rs18, %r448;
cvt.u16.u32 %rs19, %r446;
cvt.u16.u32 %rs20, %r444;
mov.u16 %rs21, 255;
st.global.v4.u8 [%rd81], {%rs20, %rs19, %rs18, %rs21};
$L__BB0_144:
and.b32 %r449, %r86, 4;
setp.eq.s32 %p181, %r449, 0;
ld.const.u32 %r495, [params+108];
@%p181 bra $L__BB0_148;
setp.eq.s32 %p182, %r495, 0;
ld.const.u64 %rd82, [params+224];
cvta.to.global.u64 %rd83, %rd82;
ld.const.u32 %r450, [params+216];
mad.lo.s32 %r451, %r450, %r8, %r7;
mul.wide.u32 %rd84, %r451, 8;
add.s64 %rd23, %rd83, %rd84;
@%p182 bra $L__BB0_147;
ld.global.v4.u16 {%rs29, %rs30, %rs31, %rs32}, [%rd23];
// begin inline asm
{ cvt.f32.f16 %f1308, %rs29;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1309, %rs30;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1310, %rs31;}
// end inline asm
add.f32 %f1311, %f271, %f1308;
add.f32 %f1312, %f272, %f1309;
add.f32 %f1313, %f274, %f1310;
mov.f32 %f1314, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1313;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs26, %f1312;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs25, %f1311;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1314;}
// end inline asm
st.global.v4.u16 [%rd23], {%rs25, %rs26, %rs27, %rs28};
bra.uni $L__BB0_148;
$L__BB0_147:
mov.f32 %f1318, 0f3F800000;
mov.u32 %r495, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f1318;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f274;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs34, %f272;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs33, %f271;}
// end inline asm
st.global.v4.u16 [%rd23], {%rs33, %rs34, %rs35, %rs36};
$L__BB0_148:
selp.f32 %f1319, 0f3F800000, 0f3E800000, %p198;
mul.f32 %f348, %f1319, %f1549;
mul.f32 %f349, %f1319, %f1550;
mul.f32 %f350, %f1319, %f1551;
mul.f32 %f351, %f1319, %f1546;
mul.f32 %f352, %f1319, %f1547;
mul.f32 %f353, %f1319, %f1548;
mul.f32 %f354, %f1319, %f1543;
mul.f32 %f355, %f1319, %f1544;
mul.f32 %f356, %f1319, %f1545;
mul.f32 %f357, %f1319, %f1540;
mul.f32 %f358, %f1319, %f1541;
mul.f32 %f359, %f1319, %f1542;
mul.f32 %f360, %f348, %f778;
mul.f32 %f361, %f349, %f779;
mul.f32 %f362, %f350, %f273;
ld.const.u64 %rd85, [params+256];
cvta.to.global.u64 %rd86, %rd85;
ld.const.u32 %r453, [params+248];
mad.lo.s32 %r454, %r453, %r8, %r7;
mul.wide.u32 %rd87, %r454, 8;
add.s64 %rd24, %rd86, %rd87;
setp.eq.s32 %p183, %r495, 0;
@%p183 bra $L__BB0_150;
ld.global.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd24];
// begin inline asm
{ cvt.f32.f16 %f1320, %rs44;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1321, %rs45;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1322, %rs46;}
// end inline asm
add.f32 %f1323, %f360, %f1320;
add.f32 %f1324, %f361, %f1321;
add.f32 %f1325, %f362, %f1322;
mov.f32 %f1326, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs42, %f1325;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs41, %f1324;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs40, %f1323;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f1326;}
// end inline asm
st.global.v4.u16 [%rd24], {%rs40, %rs41, %rs42, %rs43};
bra.uni $L__BB0_151;
$L__BB0_150:
mov.f32 %f1330, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f1330;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f362;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs49, %f361;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs48, %f360;}
// end inline asm
st.global.v4.u16 [%rd24], {%rs48, %rs49, %rs50, %rs51};
$L__BB0_151:
mov.f32 %f1331, 0f34000000;
max.f32 %f1332, %f348, %f1331;
div.rn.f32 %f1333, %f351, %f1332;
max.f32 %f1334, %f349, %f1331;
div.rn.f32 %f1335, %f352, %f1334;
max.f32 %f1336, %f350, %f1331;
div.rn.f32 %f1337, %f353, %f1336;
fma.rn.f32 %f363, %f1333, 0f3F000000, 0f3F000000;
fma.rn.f32 %f364, %f1335, 0f3F000000, 0f3F000000;
fma.rn.f32 %f365, %f1337, 0f3F000000, 0f3F000000;
div.rn.f32 %f1338, %f354, %f1332;
div.rn.f32 %f1339, %f355, %f1334;
div.rn.f32 %f1340, %f356, %f1336;
fma.rn.f32 %f366, %f1338, 0f3F000000, 0f3F000000;
fma.rn.f32 %f367, %f1339, 0f3F000000, 0f3F000000;
fma.rn.f32 %f368, %f1340, 0f3F000000, 0f3F000000;
div.rn.f32 %f1341, %f357, %f1332;
div.rn.f32 %f1342, %f358, %f1334;
div.rn.f32 %f1343, %f359, %f1336;
fma.rn.f32 %f369, %f1341, 0f3F000000, 0f3F000000;
fma.rn.f32 %f370, %f1342, 0f3F000000, 0f3F000000;
fma.rn.f32 %f371, %f1343, 0f3F000000, 0f3F000000;
ld.const.u64 %rd88, [params+272];
cvta.to.global.u64 %rd89, %rd88;
ld.const.u32 %r455, [params+264];
mad.lo.s32 %r456, %r455, %r8, %r7;
mul.wide.u32 %rd90, %r456, 8;
add.s64 %rd25, %rd89, %rd90;
@%p183 bra $L__BB0_153;
ld.global.v4.u16 {%rs59, %rs60, %rs61, %rs62}, [%rd25];
// begin inline asm
{ cvt.f32.f16 %f1344, %rs59;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1345, %rs60;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1346, %rs61;}
// end inline asm
add.f32 %f1347, %f363, %f1344;
add.f32 %f1348, %f364, %f1345;
add.f32 %f1349, %f365, %f1346;
mov.f32 %f1350, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs57, %f1349;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs56, %f1348;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs55, %f1347;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs58, %f1350;}
// end inline asm
st.global.v4.u16 [%rd25], {%rs55, %rs56, %rs57, %rs58};
bra.uni $L__BB0_154;
$L__BB0_153:
mov.f32 %f1354, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs66, %f1354;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs65, %f365;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs64, %f364;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs63, %f363;}
// end inline asm
st.global.v4.u16 [%rd25], {%rs63, %rs64, %rs65, %rs66};
$L__BB0_154:
ld.const.u64 %rd91, [params+288];
cvta.to.global.u64 %rd92, %rd91;
ld.const.u32 %r457, [params+280];
mad.lo.s32 %r458, %r457, %r8, %r7;
mul.wide.u32 %rd93, %r458, 8;
add.s64 %rd26, %rd92, %rd93;
@%p183 bra $L__BB0_156;
ld.global.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd26];
// begin inline asm
{ cvt.f32.f16 %f1355, %rs74;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1356, %rs75;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1357, %rs76;}
// end inline asm
add.f32 %f1358, %f366, %f1355;
add.f32 %f1359, %f367, %f1356;
add.f32 %f1360, %f368, %f1357;
mov.f32 %f1361, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs72, %f1360;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs71, %f1359;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs70, %f1358;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs73, %f1361;}
// end inline asm
st.global.v4.u16 [%rd26], {%rs70, %rs71, %rs72, %rs73};
bra.uni $L__BB0_157;
$L__BB0_156:
mov.f32 %f1365, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs81, %f1365;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs80, %f368;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs79, %f367;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs78, %f366;}
// end inline asm
st.global.v4.u16 [%rd26], {%rs78, %rs79, %rs80, %rs81};
$L__BB0_157:
ld.const.u64 %rd94, [params+304];
cvta.to.global.u64 %rd95, %rd94;
ld.const.u32 %r459, [params+296];
mad.lo.s32 %r460, %r459, %r8, %r7;
mul.wide.u32 %rd96, %r460, 8;
add.s64 %rd27, %rd95, %rd96;
@%p183 bra $L__BB0_159;
ld.global.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd27];
// begin inline asm
{ cvt.f32.f16 %f1366, %rs89;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1367, %rs90;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1368, %rs91;}
// end inline asm
add.f32 %f1369, %f369, %f1366;
add.f32 %f1370, %f370, %f1367;
add.f32 %f1371, %f371, %f1368;
mov.f32 %f1372, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs87, %f1371;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs86, %f1370;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs85, %f1369;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs88, %f1372;}
// end inline asm
st.global.v4.u16 [%rd27], {%rs85, %rs86, %rs87, %rs88};
bra.uni $L__BB0_180;
$L__BB0_159:
mov.f32 %f1376, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs96, %f1376;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs95, %f371;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs94, %f370;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs93, %f369;}
// end inline asm
st.global.v4.u16 [%rd27], {%rs93, %rs94, %rs95, %rs96};
$L__BB0_180:
ret;
}