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

2115 lines
61 KiB
Plaintext

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-31968024
// Cuda compilation tools, release 12.0, V12.0.76
// Based on NVVM 7.0.1
//
.version 8.0
.target sm_52
.address_size 64
// .globl __raygen__oxMain
.const .align 16 .b8 params[1184];
.visible .entry __raygen__oxMain()
{
.reg .pred %p<179>;
.reg .b16 %rs<65>;
.reg .f32 %f<1412>;
.reg .b32 %r<418>;
.reg .b64 %rd<54>;
// begin inline asm
call (%r23), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r24), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u64 %rd13, [params+400];
cvta.to.global.u64 %rd14, %rd13;
ld.const.u32 %r29, [params+392];
mad.lo.s32 %r30, %r29, %r24, %r23;
mul.wide.u32 %rd15, %r30, 4;
add.s64 %rd1, %rd14, %rd15;
ld.global.v2.u8 {%rs9, %rs63}, [%rd1];
or.b16 %rs11, %rs9, %rs63;
and.b16 %rs12, %rs11, 255;
setp.eq.s16 %p10, %rs12, 0;
@%p10 bra $L__BB0_2;
ld.global.u8 %rs62, [%rd1+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs62, [%rd1+2];
setp.eq.s16 %p11, %rs62, 0;
mov.f32 %f1331, 0f00000000;
mov.u16 %rs63, 0;
mov.f32 %f1332, %f1331;
mov.f32 %f1333, %f1331;
@%p11 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f237, %rs9;
div.rn.f32 %f238, %f237, 0f437F0000;
fma.rn.f32 %f239, %f238, 0f40000000, 0fBF800000;
and.b16 %rs15, %rs63, 255;
cvt.rn.f32.u16 %f240, %rs15;
div.rn.f32 %f241, %f240, 0f437F0000;
fma.rn.f32 %f242, %f241, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f243, %rs62;
div.rn.f32 %f244, %f243, 0f437F0000;
fma.rn.f32 %f245, %f244, 0f40000000, 0fBF800000;
mul.f32 %f246, %f242, %f242;
fma.rn.f32 %f247, %f239, %f239, %f246;
fma.rn.f32 %f248, %f245, %f245, %f247;
sqrt.rn.f32 %f249, %f248;
rcp.rn.f32 %f250, %f249;
mul.f32 %f1333, %f250, %f245;
mul.f32 %f1332, %f250, %f242;
mul.f32 %f1331, %f239, %f250;
$L__BB0_4:
ld.const.v2.u32 {%r31, %r32}, [params];
add.s32 %r3, %r31, %r23;
add.s32 %r4, %r32, %r24;
setp.eq.f32 %p12, %f1331, 0f00000000;
setp.eq.f32 %p13, %f1332, 0f00000000;
and.pred %p14, %p12, %p13;
setp.eq.f32 %p15, %f1333, 0f00000000;
and.pred %p16, %p15, %p14;
@%p16 bra $L__BB0_141;
bra.uni $L__BB0_5;
$L__BB0_141:
ld.const.u32 %r22, [params+104];
and.b32 %r399, %r22, 1;
setp.eq.b32 %p170, %r399, 1;
mov.pred %p171, 0;
xor.pred %p172, %p170, %p171;
not.pred %p173, %p172;
@%p173 bra $L__BB0_143;
ld.const.u64 %rd38, [params+144];
cvta.to.global.u64 %rd39, %rd38;
ld.const.u32 %r400, [params+136];
mad.lo.s32 %r401, %r400, %r4, %r3;
mul.wide.u32 %rd40, %r401, 4;
add.s64 %rd41, %rd39, %rd40;
mov.u16 %rs43, 0;
st.global.v4.u8 [%rd41], {%rs43, %rs43, %rs43, %rs43};
$L__BB0_143:
and.b32 %r402, %r22, 8;
setp.eq.s32 %p174, %r402, 0;
@%p174 bra $L__BB0_145;
ld.const.u64 %rd42, [params+192];
cvta.to.global.u64 %rd43, %rd42;
ld.const.u32 %r403, [params+184];
mad.lo.s32 %r404, %r403, %r4, %r3;
mov.f32 %f1318, 0f00000000;
cvt.rzi.u32.f32 %r405, %f1318;
mul.wide.u32 %rd44, %r404, 2;
add.s64 %rd45, %rd43, %rd44;
mov.u16 %rs44, 0;
cvt.u16.u32 %rs45, %r405;
st.global.v2.u8 [%rd45], {%rs45, %rs44};
$L__BB0_145:
and.b32 %r406, %r22, 4;
setp.eq.s32 %p175, %r406, 0;
@%p175 bra $L__BB0_149;
ld.const.u32 %r407, [params+108];
setp.eq.s32 %p176, %r407, 0;
ld.const.u64 %rd46, [params+224];
cvta.to.global.u64 %rd47, %rd46;
ld.const.u32 %r408, [params+216];
mad.lo.s32 %r409, %r408, %r4, %r3;
mul.wide.u32 %rd48, %r409, 8;
add.s64 %rd12, %rd47, %rd48;
@%p176 bra $L__BB0_148;
ld.global.v4.u16 {%rs52, %rs53, %rs54, %rs55}, [%rd12];
// begin inline asm
{ cvt.f32.f16 %f1319, %rs52;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1320, %rs53;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1321, %rs54;}
// end inline asm
add.f32 %f1322, %f1319, 0f00000000;
add.f32 %f1323, %f1320, 0f00000000;
add.f32 %f1324, %f1321, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f1324;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f1323;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs49, %f1322;}
// end inline asm
mov.u16 %rs56, 0;
st.global.v4.u16 [%rd12], {%rs49, %rs50, %rs51, %rs56};
bra.uni $L__BB0_149;
$L__BB0_5:
ld.const.u64 %rd16, [params+432];
cvta.to.global.u64 %rd17, %rd16;
ld.const.u32 %r35, [params+424];
mad.lo.s32 %r36, %r35, %r24, %r23;
mul.wide.u32 %rd18, %r36, 12;
add.s64 %rd19, %rd17, %rd18;
ld.global.f32 %f10, [%rd19];
mul.f32 %f251, %f10, 0f3456BF95;
ld.global.f32 %f11, [%rd19+4];
mul.f32 %f252, %f11, 0f3456BF95;
ld.global.f32 %f12, [%rd19+8];
mul.f32 %f253, %f12, 0f3456BF95;
abs.f32 %f254, %f1331;
div.rn.f32 %f255, %f251, %f254;
abs.f32 %f256, %f1332;
div.rn.f32 %f257, %f252, %f256;
abs.f32 %f258, %f1333;
div.rn.f32 %f259, %f253, %f258;
abs.f32 %f260, %f255;
abs.f32 %f261, %f257;
abs.f32 %f262, %f259;
mov.f32 %f263, 0f38D1B717;
max.f32 %f264, %f260, %f263;
max.f32 %f265, %f261, %f263;
max.f32 %f266, %f262, %f263;
fma.rn.f32 %f13, %f1331, %f264, %f10;
fma.rn.f32 %f14, %f1332, %f265, %f11;
fma.rn.f32 %f15, %f1333, %f266, %f12;
ld.const.u32 %r5, [params+600];
setp.eq.s32 %p17, %r5, 0;
@%p17 bra $L__BB0_46;
ld.const.u64 %rd20, [params+608];
cvta.to.global.u64 %rd2, %rd20;
mov.f32 %f274, 0f40000000;
cvt.rzi.f32.f32 %f275, %f274;
add.f32 %f276, %f275, %f275;
mov.f32 %f277, 0f40800000;
sub.f32 %f278, %f277, %f276;
abs.f32 %f16, %f278;
ld.const.u32 %r6, [params+544];
ld.const.u32 %r417, [params+104];
and.b32 %r8, %r417, 64;
and.b32 %r38, %r417, 32;
setp.eq.s32 %p18, %r38, 0;
mov.u32 %r413, 0;
selp.f32 %f17, 0f3F800000, 0f41200000, %p18;
mul.f32 %f18, %f13, 0f3456BF95;
mul.f32 %f19, %f14, 0f3456BF95;
mul.f32 %f20, %f15, 0f3456BF95;
mov.f32 %f273, 0f00000000;
ld.const.u64 %rd3, [params+96];
abs.f32 %f469, %f18;
abs.f32 %f470, %f19;
max.f32 %f471, %f469, %f470;
abs.f32 %f472, %f20;
max.f32 %f473, %f471, %f472;
mov.f32 %f1383, %f273;
mov.f32 %f1382, %f273;
mov.f32 %f1381, %f273;
mov.f32 %f1380, %f273;
mov.f32 %f1379, %f273;
mov.f32 %f1378, %f273;
mov.f32 %f1377, %f273;
$L__BB0_7:
mul.wide.u32 %rd21, %r413, 112;
add.s64 %rd5, %rd2, %rd21;
ld.global.v4.f32 {%f280, %f281, %f282, %f283}, [%rd5];
ld.global.v4.f32 {%f287, %f288, %f289, %f290}, [%rd5+16];
ld.global.v4.f32 {%f293, %f1349, %f1350, %f296}, [%rd5+32];
ld.global.v4.f32 {%f297, %f298, %f299, %f1347}, [%rd5+48];
ld.global.v4.f32 {%f301, %f302, %f303, %f304}, [%rd5+64];
ld.global.v4.u32 {%r39, %r40, %r41, %r42}, [%rd5+80];
mov.b32 %f41, %r39;
mov.b32 %f42, %r40;
mov.b32 %f43, %r41;
ld.global.u64 %rd6, [%rd5+96];
sub.f32 %f44, %f281, %f10;
sub.f32 %f45, %f282, %f11;
sub.f32 %f46, %f283, %f12;
mul.f32 %f305, %f45, %f45;
fma.rn.f32 %f306, %f44, %f44, %f305;
fma.rn.f32 %f307, %f46, %f46, %f306;
sqrt.rn.f32 %f308, %f307;
rcp.rn.f32 %f309, %f308;
mul.f32 %f47, %f44, %f309;
mul.f32 %f48, %f45, %f309;
mul.f32 %f49, %f46, %f309;
mul.f32 %f50, %f308, %f288;
mul.f32 %f51, %f308, %f290;
abs.f32 %f52, %f51;
setp.lt.f32 %p19, %f52, 0f00800000;
mul.f32 %f310, %f52, 0f4B800000;
selp.f32 %f311, %f310, %f52, %p19;
selp.f32 %f312, 0fC3170000, 0fC2FE0000, %p19;
mov.b32 %r46, %f311;
and.b32 %r47, %r46, 8388607;
or.b32 %r48, %r47, 1065353216;
mov.b32 %f313, %r48;
shr.u32 %r49, %r46, 23;
cvt.rn.f32.u32 %f314, %r49;
add.f32 %f315, %f312, %f314;
setp.gt.f32 %p20, %f313, 0f3FB504F3;
mul.f32 %f316, %f313, 0f3F000000;
add.f32 %f317, %f315, 0f3F800000;
selp.f32 %f318, %f317, %f315, %p20;
selp.f32 %f319, %f316, %f313, %p20;
add.f32 %f320, %f319, 0fBF800000;
add.f32 %f321, %f319, 0f3F800000;
rcp.approx.ftz.f32 %f322, %f321;
add.f32 %f323, %f320, %f320;
mul.f32 %f324, %f323, %f322;
mul.f32 %f325, %f324, %f324;
mov.f32 %f326, 0f3C4CAF63;
mov.f32 %f327, 0f3B18F0FE;
fma.rn.f32 %f328, %f327, %f325, %f326;
mov.f32 %f329, 0f3DAAAABD;
fma.rn.f32 %f330, %f328, %f325, %f329;
mul.rn.f32 %f331, %f330, %f325;
mul.rn.f32 %f332, %f331, %f324;
sub.f32 %f333, %f320, %f324;
add.f32 %f334, %f333, %f333;
neg.f32 %f335, %f324;
fma.rn.f32 %f336, %f335, %f320, %f334;
mul.rn.f32 %f337, %f322, %f336;
add.f32 %f338, %f332, %f324;
sub.f32 %f339, %f324, %f338;
add.f32 %f340, %f332, %f339;
add.f32 %f341, %f337, %f340;
add.f32 %f342, %f338, %f341;
sub.f32 %f343, %f338, %f342;
add.f32 %f344, %f341, %f343;
mov.f32 %f345, 0f3F317200;
mul.rn.f32 %f346, %f318, %f345;
mov.f32 %f347, 0f35BFBE8E;
mul.rn.f32 %f348, %f318, %f347;
add.f32 %f349, %f346, %f342;
sub.f32 %f350, %f346, %f349;
add.f32 %f351, %f342, %f350;
add.f32 %f352, %f344, %f351;
add.f32 %f353, %f348, %f352;
add.f32 %f354, %f349, %f353;
sub.f32 %f355, %f349, %f354;
add.f32 %f356, %f353, %f355;
mul.rn.f32 %f358, %f277, %f354;
neg.f32 %f359, %f358;
fma.rn.f32 %f360, %f277, %f354, %f359;
fma.rn.f32 %f361, %f277, %f356, %f360;
fma.rn.f32 %f363, %f273, %f354, %f361;
add.rn.f32 %f364, %f358, %f363;
neg.f32 %f365, %f364;
add.rn.f32 %f366, %f358, %f365;
add.rn.f32 %f367, %f366, %f363;
mov.b32 %r50, %f364;
setp.eq.s32 %p21, %r50, 1118925336;
add.s32 %r51, %r50, -1;
mov.b32 %f368, %r51;
add.f32 %f369, %f367, 0f37000000;
selp.f32 %f53, %f369, %f367, %p21;
selp.f32 %f370, %f368, %f364, %p21;
mov.f32 %f371, 0f3FB8AA3B;
mul.rn.f32 %f372, %f370, %f371;
cvt.rzi.f32.f32 %f373, %f372;
abs.f32 %f374, %f373;
setp.gt.f32 %p22, %f374, 0f42FC0000;
mov.b32 %r52, %f373;
and.b32 %r53, %r52, -2147483648;
or.b32 %r54, %r53, 1123811328;
mov.b32 %f375, %r54;
selp.f32 %f376, %f375, %f373, %p22;
mov.f32 %f377, 0fBF317218;
fma.rn.f32 %f378, %f376, %f377, %f370;
mov.f32 %f379, 0f3102E308;
fma.rn.f32 %f380, %f376, %f379, %f378;
mul.f32 %f381, %f380, 0f3FB8AA3B;
add.f32 %f382, %f376, 0f4B40007F;
mov.b32 %r55, %f382;
shl.b32 %r56, %r55, 23;
mov.b32 %f383, %r56;
ex2.approx.ftz.f32 %f384, %f381;
mul.f32 %f54, %f384, %f383;
setp.eq.f32 %p23, %f54, 0f7F800000;
mov.f32 %f1341, 0f7F800000;
@%p23 bra $L__BB0_9;
fma.rn.f32 %f1341, %f54, %f53, %f54;
$L__BB0_9:
setp.lt.f32 %p24, %f51, 0f00000000;
setp.eq.f32 %p25, %f16, 0f3F800000;
and.pred %p1, %p24, %p25;
setp.eq.f32 %p26, %f51, 0f00000000;
@%p26 bra $L__BB0_13;
bra.uni $L__BB0_10;
$L__BB0_13:
add.f32 %f389, %f51, %f51;
selp.f32 %f1343, %f389, 0f00000000, %p25;
bra.uni $L__BB0_14;
$L__BB0_10:
mov.b32 %r57, %f1341;
xor.b32 %r58, %r57, -2147483648;
mov.b32 %f385, %r58;
selp.f32 %f1343, %f385, %f1341, %p1;
setp.geu.f32 %p27, %f51, 0f00000000;
@%p27 bra $L__BB0_14;
mov.f32 %f386, 0f40800000;
cvt.rzi.f32.f32 %f387, %f386;
setp.eq.f32 %p28, %f387, 0f40800000;
@%p28 bra $L__BB0_14;
mov.f32 %f1343, 0f7FFFFFFF;
$L__BB0_14:
abs.f32 %f1328, %f51;
add.f32 %f390, %f1328, 0f40800000;
mov.b32 %r59, %f390;
setp.lt.s32 %p30, %r59, 2139095040;
@%p30 bra $L__BB0_19;
abs.f32 %f1329, %f51;
setp.gtu.f32 %p31, %f1329, 0f7F800000;
@%p31 bra $L__BB0_18;
bra.uni $L__BB0_16;
$L__BB0_18:
add.f32 %f1343, %f51, 0f40800000;
bra.uni $L__BB0_19;
$L__BB0_16:
abs.f32 %f1330, %f51;
setp.neu.f32 %p32, %f1330, 0f7F800000;
@%p32 bra $L__BB0_19;
selp.f32 %f1343, 0fFF800000, 0f7F800000, %p1;
$L__BB0_19:
mov.f32 %f391, 0f3F800000;
sub.f32 %f392, %f391, %f1343;
setp.eq.f32 %p33, %f51, 0f3F800000;
selp.f32 %f393, 0f00000000, %f392, %p33;
cvt.sat.f32.f32 %f394, %f393;
fma.rn.f32 %f395, %f50, %f50, %f289;
div.rn.f32 %f63, %f394, %f395;
mul.f32 %f396, %f1332, %f48;
fma.rn.f32 %f397, %f1331, %f47, %f396;
fma.rn.f32 %f64, %f1333, %f49, %f397;
setp.eq.f32 %p34, %f280, 0f3F800000;
@%p34 bra $L__BB0_25;
bra.uni $L__BB0_20;
$L__BB0_25:
setp.eq.s64 %p38, %rd6, 0;
mov.f32 %f1348, %f293;
mov.f32 %f1351, %f63;
@%p38 bra $L__BB0_29;
setp.geu.f32 %p39, %f1347, 0f00000000;
mov.f32 %f1344, %f47;
mov.f32 %f1345, %f48;
mov.f32 %f1346, %f49;
@%p39 bra $L__BB0_28;
mul.f32 %f420, %f42, %f45;
fma.rn.f32 %f421, %f41, %f44, %f420;
fma.rn.f32 %f422, %f43, %f46, %f421;
rcp.rn.f32 %f423, %f422;
mul.f32 %f1344, %f44, %f423;
mul.f32 %f1345, %f45, %f423;
mul.f32 %f1346, %f46, %f423;
neg.f32 %f1347, %f1347;
$L__BB0_28:
mul.f32 %f424, %f43, %f49;
mul.f32 %f425, %f42, %f48;
neg.f32 %f426, %f425;
mul.f32 %f427, %f41, %f47;
sub.f32 %f428, %f426, %f427;
sub.f32 %f429, %f428, %f424;
setp.gt.f32 %p40, %f429, 0f00000000;
selp.f32 %f430, 0f3F800000, 0f00000000, %p40;
mov.f32 %f431, 0f3F800000;
mul.f32 %f432, %f298, %f1345;
fma.rn.f32 %f433, %f297, %f1344, %f432;
mul.f32 %f434, %f302, %f1345;
fma.rn.f32 %f435, %f301, %f1344, %f434;
fma.rn.f32 %f436, %f299, %f1346, %f433;
fma.rn.f32 %f437, %f303, %f1346, %f435;
fma.rn.f32 %f438, %f1347, %f436, 0f3F000000;
sub.f32 %f439, %f431, %f438;
fma.rn.f32 %f440, %f1347, %f437, 0f3F000000;
tex.2d.v4.f32.f32 {%f441, %f442, %f443, %f444}, [%rd6, {%f439, %f440}];
mul.f32 %f445, %f430, %f441;
mul.f32 %f446, %f430, %f442;
mul.f32 %f447, %f430, %f443;
mul.f32 %f1348, %f293, %f445;
mul.f32 %f1349, %f1349, %f446;
mul.f32 %f1350, %f1350, %f447;
mov.f32 %f1351, %f63;
bra.uni $L__BB0_29;
$L__BB0_20:
setp.eq.f32 %p35, %f280, 0f40000000;
@%p35 bra $L__BB0_23;
bra.uni $L__BB0_21;
$L__BB0_23:
setp.eq.s64 %p37, %rd6, 0;
mov.f32 %f1348, %f293;
mov.f32 %f1351, %f63;
@%p37 bra $L__BB0_29;
mul.f32 %f407, %f298, %f48;
fma.rn.f32 %f408, %f297, %f47, %f407;
mul.f32 %f409, %f302, %f48;
fma.rn.f32 %f410, %f301, %f47, %f409;
mul.f32 %f411, %f42, %f48;
fma.rn.f32 %f412, %f41, %f47, %f411;
fma.rn.f32 %f413, %f299, %f49, %f408;
fma.rn.f32 %f414, %f303, %f49, %f410;
fma.rn.f32 %f415, %f43, %f49, %f412;
tex.cube.v4.f32.f32 {%f416, %f417, %f418, %f419}, [%rd6, {%f413, %f414, %f415, %f415}];
mul.f32 %f1348, %f293, %f416;
mul.f32 %f1349, %f1349, %f417;
mul.f32 %f1350, %f1350, %f418;
mov.f32 %f1351, %f63;
bra.uni $L__BB0_29;
$L__BB0_21:
setp.neu.f32 %p36, %f280, 0f40800000;
mov.f32 %f1348, %f293;
mov.f32 %f1351, %f63;
@%p36 bra $L__BB0_29;
mul.f32 %f398, %f41, %f47;
mul.f32 %f399, %f42, %f48;
neg.f32 %f400, %f399;
sub.f32 %f401, %f400, %f398;
mul.f32 %f402, %f43, %f49;
sub.f32 %f403, %f401, %f402;
fma.rn.f32 %f404, %f296, %f403, %f1347;
cvt.sat.f32.f32 %f405, %f404;
mul.f32 %f406, %f405, %f405;
mul.f32 %f1351, %f63, %f406;
mov.f32 %f1348, %f293;
$L__BB0_29:
max.f32 %f454, %f1348, %f1349;
max.f32 %f455, %f454, %f1350;
setp.eq.s32 %p42, %r6, 0;
selp.f32 %f456, %f64, 0f3F800000, %p42;
cvt.sat.f32.f32 %f457, %f456;
mul.f32 %f84, %f457, %f1351;
mul.f32 %f458, %f84, %f455;
setp.lt.f32 %p43, %f458, 0f3727C5AC;
mov.pred %p178, -1;
mov.f32 %f1352, 0f00000000;
mov.f32 %f1353, %f1352;
mov.f32 %f1354, %f1352;
mov.f32 %f1355, %f1352;
mov.f32 %f1356, %f1352;
mov.f32 %f1357, %f1352;
@%p43 bra $L__BB0_31;
setp.eq.s32 %p45, %r8, 0;
mul.f32 %f1352, %f1348, %f84;
mul.f32 %f1353, %f1349, %f84;
mul.f32 %f1354, %f1350, %f84;
selp.f32 %f1355, 0f00000000, %f47, %p45;
selp.f32 %f1356, 0f00000000, %f48, %p45;
selp.f32 %f1357, 0f00000000, %f49, %p45;
mov.pred %p178, 0;
$L__BB0_31:
@%p178 bra $L__BB0_45;
setp.eq.s32 %p46, %r42, 0;
mov.u16 %rs64, 0;
mov.f32 %f1376, 0f3F800000;
mov.f32 %f1371, %f1376;
mov.f32 %f1372, %f1376;
@%p46 bra $L__BB0_40;
abs.s32 %r11, %r42;
setp.lt.s32 %p47, %r11, 1;
mov.f32 %f1361, 0f00000000;
mov.f32 %f1362, %f1361;
mov.f32 %f1363, %f1361;
@%p47 bra $L__BB0_39;
mov.f32 %f474, 0f38D1B717;
max.f32 %f475, %f473, %f474;
mul.f32 %f97, %f17, %f475;
and.b32 %r12, %r11, 1;
setp.eq.s32 %p48, %r11, 1;
mov.f32 %f1363, 0f00000000;
mov.u32 %r416, 0;
mov.f32 %f1362, %f1363;
mov.f32 %f1361, %f1363;
@%p48 bra $L__BB0_37;
sub.s32 %r415, %r11, %r12;
$L__BB0_36:
cvt.rn.f32.s32 %f497, %r416;
mul.f32 %f498, %f497, 0f3DD32618;
cvt.rmi.f32.f32 %f499, %f498;
sub.f32 %f500, %f498, %f499;
mul.f32 %f501, %f497, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f502, %f501;
sub.f32 %f503, %f501, %f502;
mul.f32 %f504, %f497, 0f3DC74539;
cvt.rmi.f32.f32 %f505, %f504;
sub.f32 %f506, %f504, %f505;
add.f32 %f507, %f503, 0f4199851F;
add.f32 %f508, %f506, 0f4199851F;
add.f32 %f509, %f500, 0f4199851F;
mul.f32 %f510, %f503, %f508;
fma.rn.f32 %f511, %f500, %f507, %f510;
fma.rn.f32 %f512, %f509, %f506, %f511;
add.f32 %f513, %f500, %f512;
add.f32 %f514, %f503, %f512;
add.f32 %f515, %f506, %f512;
add.f32 %f516, %f513, %f514;
mul.f32 %f517, %f515, %f516;
cvt.rmi.f32.f32 %f518, %f517;
sub.f32 %f519, %f517, %f518;
add.f32 %f520, %f513, %f515;
mul.f32 %f521, %f514, %f520;
cvt.rmi.f32.f32 %f522, %f521;
sub.f32 %f523, %f521, %f522;
add.f32 %f524, %f514, %f515;
mul.f32 %f525, %f513, %f524;
cvt.rmi.f32.f32 %f526, %f525;
sub.f32 %f527, %f525, %f526;
fma.rn.f32 %f528, %f519, 0f40000000, 0fBF800000;
fma.rn.f32 %f529, %f523, 0f40000000, 0fBF800000;
fma.rn.f32 %f530, %f527, 0f40000000, 0fBF800000;
ld.global.f32 %f531, [%rd5+16];
ld.global.f32 %f532, [%rd5+4];
fma.rn.f32 %f533, %f531, %f528, %f532;
ld.global.f32 %f534, [%rd5+8];
fma.rn.f32 %f535, %f531, %f529, %f534;
ld.global.f32 %f536, [%rd5+12];
fma.rn.f32 %f537, %f531, %f530, %f536;
sub.f32 %f538, %f533, %f10;
sub.f32 %f539, %f535, %f11;
sub.f32 %f540, %f537, %f12;
mul.f32 %f541, %f539, %f539;
fma.rn.f32 %f542, %f538, %f538, %f541;
fma.rn.f32 %f543, %f540, %f540, %f542;
sqrt.rn.f32 %f486, %f543;
rcp.rn.f32 %f544, %f486;
mul.f32 %f482, %f544, %f538;
mul.f32 %f483, %f544, %f539;
mul.f32 %f484, %f544, %f540;
mov.f32 %f496, 0f00000000;
mov.u32 %r169, 2;
mov.u32 %r170, 1;
mov.u32 %r171, 3;
mov.u32 %r174, 1065353216;
mov.u32 %r203, 0;
// begin inline asm
call(%r62,%r63,%r64,%r65,%r66,%r67,%r68,%r69,%r70,%r71,%r72,%r73,%r74,%r75,%r76,%r77,%r78,%r79,%r80,%r81,%r82,%r83,%r84,%r85,%r86,%r87,%r88,%r89,%r90,%r91,%r92,%r93),_optix_trace_typed_32,(%r203,%rd3,%f13,%f14,%f15,%f482,%f483,%f484,%f97,%f486,%f496,%r170,%r203,%r170,%r169,%r170,%r171,%r174,%r174,%r174,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203);
// end inline asm
mov.b32 %f545, %r62;
mov.b32 %f546, %r63;
mov.b32 %f547, %r64;
add.f32 %f548, %f1361, %f545;
add.f32 %f549, %f1362, %f546;
add.f32 %f550, %f1363, %f547;
add.s32 %r204, %r416, 1;
cvt.rn.f32.s32 %f551, %r204;
mul.f32 %f552, %f551, 0f3DD32618;
cvt.rmi.f32.f32 %f553, %f552;
sub.f32 %f554, %f552, %f553;
mul.f32 %f555, %f551, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f556, %f555;
sub.f32 %f557, %f555, %f556;
mul.f32 %f558, %f551, 0f3DC74539;
cvt.rmi.f32.f32 %f559, %f558;
sub.f32 %f560, %f558, %f559;
add.f32 %f561, %f557, 0f4199851F;
add.f32 %f562, %f560, 0f4199851F;
add.f32 %f563, %f554, 0f4199851F;
mul.f32 %f564, %f557, %f562;
fma.rn.f32 %f565, %f554, %f561, %f564;
fma.rn.f32 %f566, %f563, %f560, %f565;
add.f32 %f567, %f554, %f566;
add.f32 %f568, %f557, %f566;
add.f32 %f569, %f560, %f566;
add.f32 %f570, %f567, %f568;
mul.f32 %f571, %f569, %f570;
cvt.rmi.f32.f32 %f572, %f571;
sub.f32 %f573, %f571, %f572;
add.f32 %f574, %f567, %f569;
mul.f32 %f575, %f568, %f574;
cvt.rmi.f32.f32 %f576, %f575;
sub.f32 %f577, %f575, %f576;
add.f32 %f578, %f568, %f569;
mul.f32 %f579, %f567, %f578;
cvt.rmi.f32.f32 %f580, %f579;
sub.f32 %f581, %f579, %f580;
fma.rn.f32 %f582, %f573, 0f40000000, 0fBF800000;
fma.rn.f32 %f583, %f577, 0f40000000, 0fBF800000;
fma.rn.f32 %f584, %f581, 0f40000000, 0fBF800000;
ld.global.f32 %f585, [%rd5+16];
ld.global.f32 %f586, [%rd5+4];
fma.rn.f32 %f587, %f585, %f582, %f586;
ld.global.f32 %f588, [%rd5+8];
fma.rn.f32 %f589, %f585, %f583, %f588;
ld.global.f32 %f590, [%rd5+12];
fma.rn.f32 %f591, %f585, %f584, %f590;
sub.f32 %f592, %f587, %f10;
sub.f32 %f593, %f589, %f11;
sub.f32 %f594, %f591, %f12;
mul.f32 %f595, %f593, %f593;
fma.rn.f32 %f596, %f592, %f592, %f595;
fma.rn.f32 %f597, %f594, %f594, %f596;
sqrt.rn.f32 %f495, %f597;
rcp.rn.f32 %f598, %f495;
mul.f32 %f491, %f598, %f592;
mul.f32 %f492, %f598, %f593;
mul.f32 %f493, %f598, %f594;
// begin inline asm
call(%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,%r164),_optix_trace_typed_32,(%r203,%rd3,%f13,%f14,%f15,%f491,%f492,%f493,%f97,%f495,%f496,%r170,%r203,%r170,%r169,%r170,%r171,%r174,%r174,%r174,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203);
// end inline asm
mov.b32 %f599, %r133;
mov.b32 %f600, %r134;
mov.b32 %f601, %r135;
add.f32 %f1361, %f548, %f599;
add.f32 %f1362, %f549, %f600;
add.f32 %f1363, %f550, %f601;
add.s32 %r416, %r416, 2;
add.s32 %r415, %r415, -2;
setp.ne.s32 %p49, %r415, 0;
@%p49 bra $L__BB0_36;
$L__BB0_37:
setp.eq.s32 %p50, %r12, 0;
@%p50 bra $L__BB0_39;
cvt.rn.f32.s32 %f611, %r416;
mul.f32 %f612, %f611, 0f3DD32618;
cvt.rmi.f32.f32 %f613, %f612;
sub.f32 %f614, %f612, %f613;
mul.f32 %f615, %f611, 0f3DD2F1AA;
cvt.rmi.f32.f32 %f616, %f615;
sub.f32 %f617, %f615, %f616;
mul.f32 %f618, %f611, 0f3DC74539;
cvt.rmi.f32.f32 %f619, %f618;
sub.f32 %f620, %f618, %f619;
add.f32 %f621, %f617, 0f4199851F;
add.f32 %f622, %f620, 0f4199851F;
add.f32 %f623, %f614, 0f4199851F;
mul.f32 %f624, %f617, %f622;
fma.rn.f32 %f625, %f614, %f621, %f624;
fma.rn.f32 %f626, %f623, %f620, %f625;
add.f32 %f627, %f614, %f626;
add.f32 %f628, %f617, %f626;
add.f32 %f629, %f620, %f626;
add.f32 %f630, %f627, %f628;
mul.f32 %f631, %f629, %f630;
cvt.rmi.f32.f32 %f632, %f631;
sub.f32 %f633, %f631, %f632;
add.f32 %f634, %f627, %f629;
mul.f32 %f635, %f628, %f634;
cvt.rmi.f32.f32 %f636, %f635;
sub.f32 %f637, %f635, %f636;
add.f32 %f638, %f628, %f629;
mul.f32 %f639, %f627, %f638;
cvt.rmi.f32.f32 %f640, %f639;
sub.f32 %f641, %f639, %f640;
fma.rn.f32 %f642, %f633, 0f40000000, 0fBF800000;
fma.rn.f32 %f643, %f637, 0f40000000, 0fBF800000;
fma.rn.f32 %f644, %f641, 0f40000000, 0fBF800000;
ld.global.f32 %f645, [%rd5+16];
ld.global.f32 %f646, [%rd5+4];
fma.rn.f32 %f647, %f645, %f642, %f646;
ld.global.f32 %f648, [%rd5+8];
fma.rn.f32 %f649, %f645, %f643, %f648;
ld.global.f32 %f650, [%rd5+12];
fma.rn.f32 %f651, %f645, %f644, %f650;
sub.f32 %f652, %f647, %f10;
sub.f32 %f653, %f649, %f11;
sub.f32 %f654, %f651, %f12;
mul.f32 %f655, %f653, %f653;
fma.rn.f32 %f656, %f652, %f652, %f655;
fma.rn.f32 %f657, %f654, %f654, %f656;
sqrt.rn.f32 %f609, %f657;
rcp.rn.f32 %f658, %f609;
mul.f32 %f605, %f658, %f652;
mul.f32 %f606, %f658, %f653;
mul.f32 %f607, %f658, %f654;
mov.f32 %f610, 0f00000000;
mov.u32 %r241, 2;
mov.u32 %r242, 1;
mov.u32 %r243, 3;
mov.u32 %r246, 1065353216;
mov.u32 %r275, 0;
// begin inline asm
call(%r205,%r206,%r207,%r208,%r209,%r210,%r211,%r212,%r213,%r214,%r215,%r216,%r217,%r218,%r219,%r220,%r221,%r222,%r223,%r224,%r225,%r226,%r227,%r228,%r229,%r230,%r231,%r232,%r233,%r234,%r235,%r236),_optix_trace_typed_32,(%r275,%rd3,%f13,%f14,%f15,%f605,%f606,%f607,%f97,%f609,%f610,%r242,%r275,%r242,%r241,%r242,%r243,%r246,%r246,%r246,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275);
// end inline asm
mov.b32 %f659, %r205;
mov.b32 %f660, %r206;
mov.b32 %f661, %r207;
add.f32 %f1361, %f1361, %f659;
add.f32 %f1362, %f1362, %f660;
add.f32 %f1363, %f1363, %f661;
$L__BB0_39:
cvt.rn.f32.s32 %f662, %r11;
rcp.rn.f32 %f663, %f662;
mul.f32 %f1376, %f663, %f1361;
mul.f32 %f1371, %f663, %f1362;
mul.f32 %f1372, %f663, %f1363;
shr.u32 %r276, %r42, 31;
cvt.u16.u32 %rs64, %r276;
$L__BB0_40:
fma.rn.f32 %f1377, %f1352, %f1376, %f1377;
fma.rn.f32 %f1378, %f1353, %f1371, %f1378;
fma.rn.f32 %f1379, %f1354, %f1372, %f1379;
setp.eq.s32 %p51, %r8, 0;
@%p51 bra $L__BB0_42;
mul.f32 %f664, %f1353, 0f3F372474;
fma.rn.f32 %f665, %f1352, 0f3E59999A, %f664;
fma.rn.f32 %f666, %f1354, 0f3D93A92A, %f665;
fma.rn.f32 %f1380, %f1355, %f666, %f1380;
fma.rn.f32 %f1381, %f1356, %f666, %f1381;
fma.rn.f32 %f1382, %f666, %f1357, %f1382;
$L__BB0_42:
setp.eq.s16 %p52, %rs64, 0;
@%p52 bra $L__BB0_44;
div.rn.f32 %f667, %f1352, %f293;
div.rn.f32 %f668, %f667, %f63;
cvt.sat.f32.f32 %f669, %f668;
mul.f32 %f1376, %f1376, %f669;
$L__BB0_44:
add.f32 %f1383, %f1383, %f1376;
$L__BB0_45:
cvt.u64.u32 %rd53, %r413;
cvt.u32.u64 %r277, %rd53;
add.s32 %r413, %r277, 1;
setp.lt.u32 %p53, %r413, %r5;
@%p53 bra $L__BB0_7;
bra.uni $L__BB0_47;
$L__BB0_46:
mov.f32 %f1377, 0f00000000;
ld.const.u32 %r417, [params+104];
mov.f32 %f1378, %f1377;
mov.f32 %f1379, %f1377;
mov.f32 %f1380, %f1377;
mov.f32 %f1381, %f1377;
mov.f32 %f1382, %f1377;
mov.f32 %f1383, %f1377;
$L__BB0_47:
and.b32 %r278, %r417, 8;
setp.eq.s32 %p54, %r278, 0;
@%p54 bra $L__BB0_61;
cvt.sat.f32.f32 %f148, %f1383;
ld.const.u64 %rd25, [params+192];
cvta.to.global.u64 %rd7, %rd25;
ld.const.u32 %r279, [params+184];
mad.lo.s32 %r280, %r279, %r4, %r3;
cvt.u64.u32 %rd8, %r280;
mov.f32 %f678, 0f3E68BA2E;
cvt.rzi.f32.f32 %f679, %f678;
add.f32 %f680, %f679, %f679;
mov.f32 %f681, 0f3EE8BA2E;
sub.f32 %f682, %f681, %f680;
abs.f32 %f149, %f682;
abs.f32 %f150, %f148;
setp.lt.f32 %p55, %f150, 0f00800000;
mul.f32 %f683, %f150, 0f4B800000;
selp.f32 %f684, %f683, %f150, %p55;
selp.f32 %f685, 0fC3170000, 0fC2FE0000, %p55;
mov.b32 %r281, %f684;
and.b32 %r282, %r281, 8388607;
or.b32 %r283, %r282, 1065353216;
mov.b32 %f686, %r283;
shr.u32 %r284, %r281, 23;
cvt.rn.f32.u32 %f687, %r284;
add.f32 %f688, %f685, %f687;
setp.gt.f32 %p56, %f686, 0f3FB504F3;
mul.f32 %f689, %f686, 0f3F000000;
add.f32 %f690, %f688, 0f3F800000;
selp.f32 %f691, %f690, %f688, %p56;
selp.f32 %f692, %f689, %f686, %p56;
add.f32 %f693, %f692, 0fBF800000;
add.f32 %f694, %f692, 0f3F800000;
rcp.approx.ftz.f32 %f695, %f694;
add.f32 %f696, %f693, %f693;
mul.f32 %f697, %f696, %f695;
mul.f32 %f698, %f697, %f697;
mov.f32 %f699, 0f3C4CAF63;
mov.f32 %f700, 0f3B18F0FE;
fma.rn.f32 %f701, %f700, %f698, %f699;
mov.f32 %f702, 0f3DAAAABD;
fma.rn.f32 %f703, %f701, %f698, %f702;
mul.rn.f32 %f704, %f703, %f698;
mul.rn.f32 %f705, %f704, %f697;
sub.f32 %f706, %f693, %f697;
add.f32 %f707, %f706, %f706;
neg.f32 %f708, %f697;
fma.rn.f32 %f709, %f708, %f693, %f707;
mul.rn.f32 %f710, %f695, %f709;
add.f32 %f711, %f705, %f697;
sub.f32 %f712, %f697, %f711;
add.f32 %f713, %f705, %f712;
add.f32 %f714, %f710, %f713;
add.f32 %f715, %f711, %f714;
sub.f32 %f716, %f711, %f715;
add.f32 %f717, %f714, %f716;
mov.f32 %f718, 0f3F317200;
mul.rn.f32 %f719, %f691, %f718;
mov.f32 %f720, 0f35BFBE8E;
mul.rn.f32 %f721, %f691, %f720;
add.f32 %f722, %f719, %f715;
sub.f32 %f723, %f719, %f722;
add.f32 %f724, %f715, %f723;
add.f32 %f725, %f717, %f724;
add.f32 %f726, %f721, %f725;
add.f32 %f727, %f722, %f726;
sub.f32 %f728, %f722, %f727;
add.f32 %f729, %f726, %f728;
mul.rn.f32 %f730, %f681, %f727;
neg.f32 %f731, %f730;
fma.rn.f32 %f732, %f681, %f727, %f731;
fma.rn.f32 %f733, %f681, %f729, %f732;
mov.f32 %f734, 0f00000000;
fma.rn.f32 %f735, %f734, %f727, %f733;
add.rn.f32 %f736, %f730, %f735;
neg.f32 %f737, %f736;
add.rn.f32 %f738, %f730, %f737;
add.rn.f32 %f739, %f738, %f735;
mov.b32 %r285, %f736;
setp.eq.s32 %p57, %r285, 1118925336;
add.s32 %r286, %r285, -1;
mov.b32 %f740, %r286;
add.f32 %f741, %f739, 0f37000000;
selp.f32 %f151, %f741, %f739, %p57;
selp.f32 %f742, %f740, %f736, %p57;
mov.f32 %f743, 0f3FB8AA3B;
mul.rn.f32 %f744, %f742, %f743;
cvt.rzi.f32.f32 %f745, %f744;
abs.f32 %f746, %f745;
setp.gt.f32 %p58, %f746, 0f42FC0000;
mov.b32 %r287, %f745;
and.b32 %r288, %r287, -2147483648;
or.b32 %r289, %r288, 1123811328;
mov.b32 %f747, %r289;
selp.f32 %f748, %f747, %f745, %p58;
mov.f32 %f749, 0fBF317218;
fma.rn.f32 %f750, %f748, %f749, %f742;
mov.f32 %f751, 0f3102E308;
fma.rn.f32 %f752, %f748, %f751, %f750;
mul.f32 %f753, %f752, 0f3FB8AA3B;
add.f32 %f754, %f748, 0f4B40007F;
mov.b32 %r290, %f754;
shl.b32 %r291, %r290, 23;
mov.b32 %f755, %r291;
ex2.approx.ftz.f32 %f756, %f753;
mul.f32 %f152, %f756, %f755;
setp.eq.f32 %p59, %f152, 0f7F800000;
mov.f32 %f1391, 0f7F800000;
@%p59 bra $L__BB0_50;
fma.rn.f32 %f1391, %f152, %f151, %f152;
$L__BB0_50:
setp.lt.f32 %p60, %f148, 0f00000000;
setp.eq.f32 %p61, %f149, 0f3F800000;
and.pred %p3, %p60, %p61;
setp.eq.f32 %p62, %f148, 0f00000000;
@%p62 bra $L__BB0_54;
bra.uni $L__BB0_51;
$L__BB0_54:
add.f32 %f761, %f148, %f148;
selp.f32 %f1393, %f761, 0f00000000, %p61;
bra.uni $L__BB0_55;
$L__BB0_148:
mov.f32 %f1327, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs59, %f1327;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs58, %f1327;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs57, %f1327;}
// end inline asm
mov.u16 %rs60, 0;
st.global.v4.u16 [%rd12], {%rs57, %rs58, %rs59, %rs60};
$L__BB0_149:
and.b32 %r410, %r22, 64;
setp.eq.s32 %p177, %r410, 0;
@%p177 bra $L__BB0_151;
ld.const.u64 %rd49, [params+208];
cvta.to.global.u64 %rd50, %rd49;
ld.const.u32 %r411, [params+200];
mad.lo.s32 %r412, %r411, %r4, %r3;
mul.wide.u32 %rd51, %r412, 4;
add.s64 %rd52, %rd50, %rd51;
mov.u16 %rs61, 0;
st.global.v4.u8 [%rd52], {%rs61, %rs61, %rs61, %rs61};
bra.uni $L__BB0_151;
$L__BB0_51:
mov.b32 %r292, %f1391;
xor.b32 %r293, %r292, -2147483648;
mov.b32 %f757, %r293;
selp.f32 %f1393, %f757, %f1391, %p3;
setp.geu.f32 %p63, %f148, 0f00000000;
@%p63 bra $L__BB0_55;
mov.f32 %f758, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f759, %f758;
setp.eq.f32 %p64, %f759, 0f3EE8BA2E;
@%p64 bra $L__BB0_55;
mov.f32 %f1393, 0f7FFFFFFF;
$L__BB0_55:
add.f32 %f762, %f150, 0f3EE8BA2E;
mov.b32 %r294, %f762;
setp.lt.s32 %p66, %r294, 2139095040;
@%p66 bra $L__BB0_60;
setp.gtu.f32 %p67, %f150, 0f7F800000;
@%p67 bra $L__BB0_59;
bra.uni $L__BB0_57;
$L__BB0_59:
add.f32 %f1393, %f148, 0f3EE8BA2E;
bra.uni $L__BB0_60;
$L__BB0_57:
setp.neu.f32 %p68, %f150, 0f7F800000;
@%p68 bra $L__BB0_60;
selp.f32 %f1393, 0fFF800000, 0f7F800000, %p3;
$L__BB0_60:
mul.f32 %f763, %f1393, 0f437F0000;
setp.eq.f32 %p69, %f148, 0f3F800000;
selp.f32 %f764, 0f437F0000, %f763, %p69;
cvt.rzi.u32.f32 %r295, %f764;
shl.b64 %rd26, %rd8, 1;
add.s64 %rd27, %rd7, %rd26;
cvt.u16.u32 %rs18, %r295;
mov.u16 %rs19, 255;
st.global.v2.u8 [%rd27], {%rs18, %rs19};
$L__BB0_61:
and.b32 %r296, %r417, 1;
setp.eq.b32 %p70, %r296, 1;
mov.pred %p71, 0;
xor.pred %p72, %p70, %p71;
not.pred %p73, %p72;
@%p73 bra $L__BB0_135;
mov.f32 %f766, 0f3E666666;
cvt.rzi.f32.f32 %f767, %f766;
add.f32 %f768, %f767, %f767;
mov.f32 %f769, 0f3EE66666;
sub.f32 %f770, %f769, %f768;
abs.f32 %f161, %f770;
abs.f32 %f162, %f1377;
setp.lt.f32 %p74, %f162, 0f00800000;
mul.f32 %f771, %f162, 0f4B800000;
selp.f32 %f772, %f771, %f162, %p74;
selp.f32 %f773, 0fC3170000, 0fC2FE0000, %p74;
mov.b32 %r297, %f772;
and.b32 %r298, %r297, 8388607;
or.b32 %r299, %r298, 1065353216;
mov.b32 %f774, %r299;
shr.u32 %r300, %r297, 23;
cvt.rn.f32.u32 %f775, %r300;
add.f32 %f776, %f773, %f775;
setp.gt.f32 %p75, %f774, 0f3FB504F3;
mul.f32 %f777, %f774, 0f3F000000;
add.f32 %f778, %f776, 0f3F800000;
selp.f32 %f779, %f778, %f776, %p75;
selp.f32 %f780, %f777, %f774, %p75;
add.f32 %f781, %f780, 0fBF800000;
add.f32 %f782, %f780, 0f3F800000;
rcp.approx.ftz.f32 %f783, %f782;
add.f32 %f784, %f781, %f781;
mul.f32 %f785, %f784, %f783;
mul.f32 %f786, %f785, %f785;
mov.f32 %f787, 0f3C4CAF63;
mov.f32 %f788, 0f3B18F0FE;
fma.rn.f32 %f789, %f788, %f786, %f787;
mov.f32 %f790, 0f3DAAAABD;
fma.rn.f32 %f791, %f789, %f786, %f790;
mul.rn.f32 %f792, %f791, %f786;
mul.rn.f32 %f793, %f792, %f785;
sub.f32 %f794, %f781, %f785;
add.f32 %f795, %f794, %f794;
neg.f32 %f796, %f785;
fma.rn.f32 %f797, %f796, %f781, %f795;
mul.rn.f32 %f798, %f783, %f797;
add.f32 %f799, %f793, %f785;
sub.f32 %f800, %f785, %f799;
add.f32 %f801, %f793, %f800;
add.f32 %f802, %f798, %f801;
add.f32 %f803, %f799, %f802;
sub.f32 %f804, %f799, %f803;
add.f32 %f805, %f802, %f804;
mov.f32 %f806, 0f3F317200;
mul.rn.f32 %f807, %f779, %f806;
mov.f32 %f808, 0f35BFBE8E;
mul.rn.f32 %f809, %f779, %f808;
add.f32 %f810, %f807, %f803;
sub.f32 %f811, %f807, %f810;
add.f32 %f812, %f803, %f811;
add.f32 %f813, %f805, %f812;
add.f32 %f814, %f809, %f813;
add.f32 %f815, %f810, %f814;
sub.f32 %f816, %f810, %f815;
add.f32 %f817, %f814, %f816;
mul.rn.f32 %f818, %f769, %f815;
neg.f32 %f819, %f818;
fma.rn.f32 %f820, %f769, %f815, %f819;
fma.rn.f32 %f821, %f769, %f817, %f820;
mov.f32 %f822, 0f00000000;
fma.rn.f32 %f823, %f822, %f815, %f821;
add.rn.f32 %f824, %f818, %f823;
neg.f32 %f825, %f824;
add.rn.f32 %f826, %f818, %f825;
add.rn.f32 %f827, %f826, %f823;
mov.b32 %r301, %f824;
setp.eq.s32 %p76, %r301, 1118925336;
add.s32 %r302, %r301, -1;
mov.b32 %f828, %r302;
add.f32 %f829, %f827, 0f37000000;
selp.f32 %f163, %f829, %f827, %p76;
selp.f32 %f830, %f828, %f824, %p76;
mov.f32 %f831, 0f3FB8AA3B;
mul.rn.f32 %f832, %f830, %f831;
cvt.rzi.f32.f32 %f833, %f832;
abs.f32 %f834, %f833;
setp.gt.f32 %p77, %f834, 0f42FC0000;
mov.b32 %r303, %f833;
and.b32 %r304, %r303, -2147483648;
or.b32 %r305, %r304, 1123811328;
mov.b32 %f835, %r305;
selp.f32 %f836, %f835, %f833, %p77;
mov.f32 %f837, 0fBF317218;
fma.rn.f32 %f838, %f836, %f837, %f830;
mov.f32 %f839, 0f3102E308;
fma.rn.f32 %f840, %f836, %f839, %f838;
mul.f32 %f841, %f840, 0f3FB8AA3B;
add.f32 %f842, %f836, 0f4B40007F;
mov.b32 %r306, %f842;
shl.b32 %r307, %r306, 23;
mov.b32 %f843, %r307;
ex2.approx.ftz.f32 %f844, %f841;
mul.f32 %f164, %f844, %f843;
setp.eq.f32 %p78, %f164, 0f7F800000;
mov.f32 %f1394, 0f7F800000;
@%p78 bra $L__BB0_64;
fma.rn.f32 %f1394, %f164, %f163, %f164;
$L__BB0_64:
setp.lt.f32 %p79, %f1377, 0f00000000;
setp.eq.f32 %p80, %f161, 0f3F800000;
and.pred %p4, %p79, %p80;
setp.eq.f32 %p81, %f1377, 0f00000000;
@%p81 bra $L__BB0_68;
bra.uni $L__BB0_65;
$L__BB0_68:
add.f32 %f849, %f1377, %f1377;
selp.f32 %f1396, %f849, 0f00000000, %p80;
bra.uni $L__BB0_69;
$L__BB0_65:
mov.b32 %r308, %f1394;
xor.b32 %r309, %r308, -2147483648;
mov.b32 %f845, %r309;
selp.f32 %f1396, %f845, %f1394, %p4;
setp.geu.f32 %p82, %f1377, 0f00000000;
@%p82 bra $L__BB0_69;
mov.f32 %f846, 0f3EE66666;
cvt.rzi.f32.f32 %f847, %f846;
setp.eq.f32 %p83, %f847, 0f3EE66666;
@%p83 bra $L__BB0_69;
mov.f32 %f1396, 0f7FFFFFFF;
$L__BB0_69:
add.f32 %f850, %f162, 0f3EE66666;
mov.b32 %r310, %f850;
setp.lt.s32 %p85, %r310, 2139095040;
@%p85 bra $L__BB0_74;
setp.gtu.f32 %p86, %f162, 0f7F800000;
@%p86 bra $L__BB0_73;
bra.uni $L__BB0_71;
$L__BB0_73:
add.f32 %f1396, %f1377, 0f3EE66666;
bra.uni $L__BB0_74;
$L__BB0_71:
setp.neu.f32 %p87, %f162, 0f7F800000;
@%p87 bra $L__BB0_74;
selp.f32 %f1396, 0fFF800000, 0f7F800000, %p4;
$L__BB0_74:
setp.eq.f32 %p88, %f1377, 0f3F800000;
selp.f32 %f173, 0f3F800000, %f1396, %p88;
abs.f32 %f174, %f1378;
setp.lt.f32 %p89, %f174, 0f00800000;
mul.f32 %f852, %f174, 0f4B800000;
selp.f32 %f853, %f852, %f174, %p89;
selp.f32 %f854, 0fC3170000, 0fC2FE0000, %p89;
mov.b32 %r311, %f853;
and.b32 %r312, %r311, 8388607;
or.b32 %r313, %r312, 1065353216;
mov.b32 %f855, %r313;
shr.u32 %r314, %r311, 23;
cvt.rn.f32.u32 %f856, %r314;
add.f32 %f857, %f854, %f856;
setp.gt.f32 %p90, %f855, 0f3FB504F3;
mul.f32 %f858, %f855, 0f3F000000;
add.f32 %f859, %f857, 0f3F800000;
selp.f32 %f860, %f859, %f857, %p90;
selp.f32 %f861, %f858, %f855, %p90;
add.f32 %f862, %f861, 0fBF800000;
add.f32 %f863, %f861, 0f3F800000;
rcp.approx.ftz.f32 %f864, %f863;
add.f32 %f865, %f862, %f862;
mul.f32 %f866, %f865, %f864;
mul.f32 %f867, %f866, %f866;
mov.f32 %f868, 0f3C4CAF63;
mov.f32 %f869, 0f3B18F0FE;
fma.rn.f32 %f870, %f869, %f867, %f868;
mov.f32 %f871, 0f3DAAAABD;
fma.rn.f32 %f872, %f870, %f867, %f871;
mul.rn.f32 %f873, %f872, %f867;
mul.rn.f32 %f874, %f873, %f866;
sub.f32 %f875, %f862, %f866;
add.f32 %f876, %f875, %f875;
neg.f32 %f877, %f866;
fma.rn.f32 %f878, %f877, %f862, %f876;
mul.rn.f32 %f879, %f864, %f878;
add.f32 %f880, %f874, %f866;
sub.f32 %f881, %f866, %f880;
add.f32 %f882, %f874, %f881;
add.f32 %f883, %f879, %f882;
add.f32 %f884, %f880, %f883;
sub.f32 %f885, %f880, %f884;
add.f32 %f886, %f883, %f885;
mov.f32 %f887, 0f3F317200;
mul.rn.f32 %f888, %f860, %f887;
mov.f32 %f889, 0f35BFBE8E;
mul.rn.f32 %f890, %f860, %f889;
add.f32 %f891, %f888, %f884;
sub.f32 %f892, %f888, %f891;
add.f32 %f893, %f884, %f892;
add.f32 %f894, %f886, %f893;
add.f32 %f895, %f890, %f894;
add.f32 %f896, %f891, %f895;
sub.f32 %f897, %f891, %f896;
add.f32 %f898, %f895, %f897;
mov.f32 %f899, 0f3EE66666;
mul.rn.f32 %f900, %f899, %f896;
neg.f32 %f901, %f900;
fma.rn.f32 %f902, %f899, %f896, %f901;
fma.rn.f32 %f903, %f899, %f898, %f902;
mov.f32 %f904, 0f00000000;
fma.rn.f32 %f905, %f904, %f896, %f903;
add.rn.f32 %f906, %f900, %f905;
neg.f32 %f907, %f906;
add.rn.f32 %f908, %f900, %f907;
add.rn.f32 %f909, %f908, %f905;
mov.b32 %r315, %f906;
setp.eq.s32 %p91, %r315, 1118925336;
add.s32 %r316, %r315, -1;
mov.b32 %f910, %r316;
add.f32 %f911, %f909, 0f37000000;
selp.f32 %f175, %f911, %f909, %p91;
selp.f32 %f912, %f910, %f906, %p91;
mov.f32 %f913, 0f3FB8AA3B;
mul.rn.f32 %f914, %f912, %f913;
cvt.rzi.f32.f32 %f915, %f914;
abs.f32 %f916, %f915;
setp.gt.f32 %p92, %f916, 0f42FC0000;
mov.b32 %r317, %f915;
and.b32 %r318, %r317, -2147483648;
or.b32 %r319, %r318, 1123811328;
mov.b32 %f917, %r319;
selp.f32 %f918, %f917, %f915, %p92;
mov.f32 %f919, 0fBF317218;
fma.rn.f32 %f920, %f918, %f919, %f912;
mov.f32 %f921, 0f3102E308;
fma.rn.f32 %f922, %f918, %f921, %f920;
mul.f32 %f923, %f922, 0f3FB8AA3B;
add.f32 %f924, %f918, 0f4B40007F;
mov.b32 %r320, %f924;
shl.b32 %r321, %r320, 23;
mov.b32 %f925, %r321;
ex2.approx.ftz.f32 %f926, %f923;
mul.f32 %f176, %f926, %f925;
setp.eq.f32 %p93, %f176, 0f7F800000;
mov.f32 %f1397, 0f7F800000;
@%p93 bra $L__BB0_76;
fma.rn.f32 %f1397, %f176, %f175, %f176;
$L__BB0_76:
setp.lt.f32 %p94, %f1378, 0f00000000;
and.pred %p5, %p94, %p80;
setp.eq.f32 %p96, %f1378, 0f00000000;
@%p96 bra $L__BB0_80;
bra.uni $L__BB0_77;
$L__BB0_80:
add.f32 %f931, %f1378, %f1378;
selp.f32 %f1399, %f931, 0f00000000, %p80;
bra.uni $L__BB0_81;
$L__BB0_77:
mov.b32 %r322, %f1397;
xor.b32 %r323, %r322, -2147483648;
mov.b32 %f927, %r323;
selp.f32 %f1399, %f927, %f1397, %p5;
setp.geu.f32 %p97, %f1378, 0f00000000;
@%p97 bra $L__BB0_81;
mov.f32 %f928, 0f3EE66666;
cvt.rzi.f32.f32 %f929, %f928;
setp.eq.f32 %p98, %f929, 0f3EE66666;
@%p98 bra $L__BB0_81;
mov.f32 %f1399, 0f7FFFFFFF;
$L__BB0_81:
add.f32 %f932, %f174, 0f3EE66666;
mov.b32 %r324, %f932;
setp.lt.s32 %p100, %r324, 2139095040;
@%p100 bra $L__BB0_86;
setp.gtu.f32 %p101, %f174, 0f7F800000;
@%p101 bra $L__BB0_85;
bra.uni $L__BB0_83;
$L__BB0_85:
add.f32 %f1399, %f1378, 0f3EE66666;
bra.uni $L__BB0_86;
$L__BB0_83:
setp.neu.f32 %p102, %f174, 0f7F800000;
@%p102 bra $L__BB0_86;
selp.f32 %f1399, 0fFF800000, 0f7F800000, %p5;
$L__BB0_86:
setp.eq.f32 %p103, %f1378, 0f3F800000;
selp.f32 %f185, 0f3F800000, %f1399, %p103;
abs.f32 %f186, %f1379;
setp.lt.f32 %p104, %f186, 0f00800000;
mul.f32 %f934, %f186, 0f4B800000;
selp.f32 %f935, %f934, %f186, %p104;
selp.f32 %f936, 0fC3170000, 0fC2FE0000, %p104;
mov.b32 %r325, %f935;
and.b32 %r326, %r325, 8388607;
or.b32 %r327, %r326, 1065353216;
mov.b32 %f937, %r327;
shr.u32 %r328, %r325, 23;
cvt.rn.f32.u32 %f938, %r328;
add.f32 %f939, %f936, %f938;
setp.gt.f32 %p105, %f937, 0f3FB504F3;
mul.f32 %f940, %f937, 0f3F000000;
add.f32 %f941, %f939, 0f3F800000;
selp.f32 %f942, %f941, %f939, %p105;
selp.f32 %f943, %f940, %f937, %p105;
add.f32 %f944, %f943, 0fBF800000;
add.f32 %f945, %f943, 0f3F800000;
rcp.approx.ftz.f32 %f946, %f945;
add.f32 %f947, %f944, %f944;
mul.f32 %f948, %f947, %f946;
mul.f32 %f949, %f948, %f948;
mov.f32 %f950, 0f3C4CAF63;
mov.f32 %f951, 0f3B18F0FE;
fma.rn.f32 %f952, %f951, %f949, %f950;
mov.f32 %f953, 0f3DAAAABD;
fma.rn.f32 %f954, %f952, %f949, %f953;
mul.rn.f32 %f955, %f954, %f949;
mul.rn.f32 %f956, %f955, %f948;
sub.f32 %f957, %f944, %f948;
add.f32 %f958, %f957, %f957;
neg.f32 %f959, %f948;
fma.rn.f32 %f960, %f959, %f944, %f958;
mul.rn.f32 %f961, %f946, %f960;
add.f32 %f962, %f956, %f948;
sub.f32 %f963, %f948, %f962;
add.f32 %f964, %f956, %f963;
add.f32 %f965, %f961, %f964;
add.f32 %f966, %f962, %f965;
sub.f32 %f967, %f962, %f966;
add.f32 %f968, %f965, %f967;
mov.f32 %f969, 0f3F317200;
mul.rn.f32 %f970, %f942, %f969;
mov.f32 %f971, 0f35BFBE8E;
mul.rn.f32 %f972, %f942, %f971;
add.f32 %f973, %f970, %f966;
sub.f32 %f974, %f970, %f973;
add.f32 %f975, %f966, %f974;
add.f32 %f976, %f968, %f975;
add.f32 %f977, %f972, %f976;
add.f32 %f978, %f973, %f977;
sub.f32 %f979, %f973, %f978;
add.f32 %f980, %f977, %f979;
mov.f32 %f981, 0f3EE66666;
mul.rn.f32 %f982, %f981, %f978;
neg.f32 %f983, %f982;
fma.rn.f32 %f984, %f981, %f978, %f983;
fma.rn.f32 %f985, %f981, %f980, %f984;
mov.f32 %f986, 0f00000000;
fma.rn.f32 %f987, %f986, %f978, %f985;
add.rn.f32 %f988, %f982, %f987;
neg.f32 %f989, %f988;
add.rn.f32 %f990, %f982, %f989;
add.rn.f32 %f991, %f990, %f987;
mov.b32 %r329, %f988;
setp.eq.s32 %p106, %r329, 1118925336;
add.s32 %r330, %r329, -1;
mov.b32 %f992, %r330;
add.f32 %f993, %f991, 0f37000000;
selp.f32 %f187, %f993, %f991, %p106;
selp.f32 %f994, %f992, %f988, %p106;
mov.f32 %f995, 0f3FB8AA3B;
mul.rn.f32 %f996, %f994, %f995;
cvt.rzi.f32.f32 %f997, %f996;
abs.f32 %f998, %f997;
setp.gt.f32 %p107, %f998, 0f42FC0000;
mov.b32 %r331, %f997;
and.b32 %r332, %r331, -2147483648;
or.b32 %r333, %r332, 1123811328;
mov.b32 %f999, %r333;
selp.f32 %f1000, %f999, %f997, %p107;
mov.f32 %f1001, 0fBF317218;
fma.rn.f32 %f1002, %f1000, %f1001, %f994;
mov.f32 %f1003, 0f3102E308;
fma.rn.f32 %f1004, %f1000, %f1003, %f1002;
mul.f32 %f1005, %f1004, 0f3FB8AA3B;
add.f32 %f1006, %f1000, 0f4B40007F;
mov.b32 %r334, %f1006;
shl.b32 %r335, %r334, 23;
mov.b32 %f1007, %r335;
ex2.approx.ftz.f32 %f1008, %f1005;
mul.f32 %f188, %f1008, %f1007;
setp.eq.f32 %p108, %f188, 0f7F800000;
mov.f32 %f1400, 0f7F800000;
@%p108 bra $L__BB0_88;
fma.rn.f32 %f1400, %f188, %f187, %f188;
$L__BB0_88:
setp.lt.f32 %p109, %f1379, 0f00000000;
and.pred %p6, %p109, %p80;
setp.eq.f32 %p111, %f1379, 0f00000000;
@%p111 bra $L__BB0_92;
bra.uni $L__BB0_89;
$L__BB0_92:
add.f32 %f1013, %f1379, %f1379;
selp.f32 %f1402, %f1013, 0f00000000, %p80;
bra.uni $L__BB0_93;
$L__BB0_89:
mov.b32 %r336, %f1400;
xor.b32 %r337, %r336, -2147483648;
mov.b32 %f1009, %r337;
selp.f32 %f1402, %f1009, %f1400, %p6;
setp.geu.f32 %p112, %f1379, 0f00000000;
@%p112 bra $L__BB0_93;
mov.f32 %f1010, 0f3EE66666;
cvt.rzi.f32.f32 %f1011, %f1010;
setp.eq.f32 %p113, %f1011, 0f3EE66666;
@%p113 bra $L__BB0_93;
mov.f32 %f1402, 0f7FFFFFFF;
$L__BB0_93:
add.f32 %f1014, %f186, 0f3EE66666;
mov.b32 %r338, %f1014;
setp.lt.s32 %p115, %r338, 2139095040;
@%p115 bra $L__BB0_98;
setp.gtu.f32 %p116, %f186, 0f7F800000;
@%p116 bra $L__BB0_97;
bra.uni $L__BB0_95;
$L__BB0_97:
add.f32 %f1402, %f1379, 0f3EE66666;
bra.uni $L__BB0_98;
$L__BB0_95:
setp.neu.f32 %p117, %f186, 0f7F800000;
@%p117 bra $L__BB0_98;
selp.f32 %f1402, 0fFF800000, 0f7F800000, %p6;
$L__BB0_98:
setp.eq.f32 %p118, %f1379, 0f3F800000;
mov.f32 %f1016, 0f3F800000;
selp.f32 %f1017, 0f3F800000, %f1402, %p118;
ld.const.u64 %rd28, [params+144];
cvta.to.global.u64 %rd9, %rd28;
ld.const.u32 %r339, [params+136];
mad.lo.s32 %r340, %r339, %r4, %r3;
cvt.u64.u32 %rd10, %r340;
min.f32 %f1018, %f173, %f1016;
mov.f32 %f1019, 0f00000000;
max.f32 %f197, %f1019, %f1018;
min.f32 %f1020, %f185, %f1016;
max.f32 %f198, %f1019, %f1020;
min.f32 %f1021, %f1017, %f1016;
max.f32 %f199, %f1019, %f1021;
mov.f32 %f1022, 0f3E555555;
cvt.rzi.f32.f32 %f1023, %f1022;
add.f32 %f1024, %f1023, %f1023;
mov.f32 %f1025, 0f3ED55555;
sub.f32 %f1026, %f1025, %f1024;
abs.f32 %f200, %f1026;
abs.f32 %f201, %f197;
setp.lt.f32 %p119, %f201, 0f00800000;
mul.f32 %f1027, %f201, 0f4B800000;
selp.f32 %f1028, %f1027, %f201, %p119;
selp.f32 %f1029, 0fC3170000, 0fC2FE0000, %p119;
mov.b32 %r341, %f1028;
and.b32 %r342, %r341, 8388607;
or.b32 %r343, %r342, 1065353216;
mov.b32 %f1030, %r343;
shr.u32 %r344, %r341, 23;
cvt.rn.f32.u32 %f1031, %r344;
add.f32 %f1032, %f1029, %f1031;
setp.gt.f32 %p120, %f1030, 0f3FB504F3;
mul.f32 %f1033, %f1030, 0f3F000000;
add.f32 %f1034, %f1032, 0f3F800000;
selp.f32 %f1035, %f1034, %f1032, %p120;
selp.f32 %f1036, %f1033, %f1030, %p120;
add.f32 %f1037, %f1036, 0fBF800000;
add.f32 %f1038, %f1036, 0f3F800000;
rcp.approx.ftz.f32 %f1039, %f1038;
add.f32 %f1040, %f1037, %f1037;
mul.f32 %f1041, %f1040, %f1039;
mul.f32 %f1042, %f1041, %f1041;
mov.f32 %f1043, 0f3C4CAF63;
mov.f32 %f1044, 0f3B18F0FE;
fma.rn.f32 %f1045, %f1044, %f1042, %f1043;
mov.f32 %f1046, 0f3DAAAABD;
fma.rn.f32 %f1047, %f1045, %f1042, %f1046;
mul.rn.f32 %f1048, %f1047, %f1042;
mul.rn.f32 %f1049, %f1048, %f1041;
sub.f32 %f1050, %f1037, %f1041;
add.f32 %f1051, %f1050, %f1050;
neg.f32 %f1052, %f1041;
fma.rn.f32 %f1053, %f1052, %f1037, %f1051;
mul.rn.f32 %f1054, %f1039, %f1053;
add.f32 %f1055, %f1049, %f1041;
sub.f32 %f1056, %f1041, %f1055;
add.f32 %f1057, %f1049, %f1056;
add.f32 %f1058, %f1054, %f1057;
add.f32 %f1059, %f1055, %f1058;
sub.f32 %f1060, %f1055, %f1059;
add.f32 %f1061, %f1058, %f1060;
mov.f32 %f1062, 0f3F317200;
mul.rn.f32 %f1063, %f1035, %f1062;
mov.f32 %f1064, 0f35BFBE8E;
mul.rn.f32 %f1065, %f1035, %f1064;
add.f32 %f1066, %f1063, %f1059;
sub.f32 %f1067, %f1063, %f1066;
add.f32 %f1068, %f1059, %f1067;
add.f32 %f1069, %f1061, %f1068;
add.f32 %f1070, %f1065, %f1069;
add.f32 %f1071, %f1066, %f1070;
sub.f32 %f1072, %f1066, %f1071;
add.f32 %f1073, %f1070, %f1072;
mul.rn.f32 %f1074, %f1025, %f1071;
neg.f32 %f1075, %f1074;
fma.rn.f32 %f1076, %f1025, %f1071, %f1075;
fma.rn.f32 %f1077, %f1025, %f1073, %f1076;
fma.rn.f32 %f1078, %f1019, %f1071, %f1077;
add.rn.f32 %f1079, %f1074, %f1078;
neg.f32 %f1080, %f1079;
add.rn.f32 %f1081, %f1074, %f1080;
add.rn.f32 %f1082, %f1081, %f1078;
mov.b32 %r345, %f1079;
setp.eq.s32 %p121, %r345, 1118925336;
add.s32 %r346, %r345, -1;
mov.b32 %f1083, %r346;
add.f32 %f1084, %f1082, 0f37000000;
selp.f32 %f202, %f1084, %f1082, %p121;
selp.f32 %f1085, %f1083, %f1079, %p121;
mov.f32 %f1086, 0f3FB8AA3B;
mul.rn.f32 %f1087, %f1085, %f1086;
cvt.rzi.f32.f32 %f1088, %f1087;
abs.f32 %f1089, %f1088;
setp.gt.f32 %p122, %f1089, 0f42FC0000;
mov.b32 %r347, %f1088;
and.b32 %r348, %r347, -2147483648;
or.b32 %r349, %r348, 1123811328;
mov.b32 %f1090, %r349;
selp.f32 %f1091, %f1090, %f1088, %p122;
mov.f32 %f1092, 0fBF317218;
fma.rn.f32 %f1093, %f1091, %f1092, %f1085;
mov.f32 %f1094, 0f3102E308;
fma.rn.f32 %f1095, %f1091, %f1094, %f1093;
mul.f32 %f1096, %f1095, 0f3FB8AA3B;
add.f32 %f1097, %f1091, 0f4B40007F;
mov.b32 %r350, %f1097;
shl.b32 %r351, %r350, 23;
mov.b32 %f1098, %r351;
ex2.approx.ftz.f32 %f1099, %f1096;
mul.f32 %f203, %f1099, %f1098;
setp.eq.f32 %p123, %f203, 0f7F800000;
mov.f32 %f1403, 0f7F800000;
@%p123 bra $L__BB0_100;
fma.rn.f32 %f1403, %f203, %f202, %f203;
$L__BB0_100:
setp.lt.f32 %p124, %f197, 0f00000000;
setp.eq.f32 %p125, %f200, 0f3F800000;
and.pred %p7, %p124, %p125;
setp.eq.f32 %p126, %f197, 0f00000000;
@%p126 bra $L__BB0_104;
bra.uni $L__BB0_101;
$L__BB0_104:
add.f32 %f1104, %f197, %f197;
selp.f32 %f1405, %f1104, 0f00000000, %p125;
bra.uni $L__BB0_105;
$L__BB0_101:
mov.b32 %r352, %f1403;
xor.b32 %r353, %r352, -2147483648;
mov.b32 %f1100, %r353;
selp.f32 %f1405, %f1100, %f1403, %p7;
setp.geu.f32 %p127, %f197, 0f00000000;
@%p127 bra $L__BB0_105;
mov.f32 %f1101, 0f3ED55555;
cvt.rzi.f32.f32 %f1102, %f1101;
setp.eq.f32 %p128, %f1102, 0f3ED55555;
@%p128 bra $L__BB0_105;
mov.f32 %f1405, 0f7FFFFFFF;
$L__BB0_105:
add.f32 %f1105, %f201, 0f3ED55555;
mov.b32 %r354, %f1105;
setp.lt.s32 %p130, %r354, 2139095040;
@%p130 bra $L__BB0_110;
setp.gtu.f32 %p131, %f201, 0f7F800000;
@%p131 bra $L__BB0_109;
bra.uni $L__BB0_107;
$L__BB0_109:
add.f32 %f1405, %f197, 0f3ED55555;
bra.uni $L__BB0_110;
$L__BB0_107:
setp.neu.f32 %p132, %f201, 0f7F800000;
@%p132 bra $L__BB0_110;
selp.f32 %f1405, 0fFF800000, 0f7F800000, %p7;
$L__BB0_110:
abs.f32 %f212, %f198;
setp.lt.f32 %p133, %f212, 0f00800000;
mul.f32 %f1107, %f212, 0f4B800000;
selp.f32 %f1108, %f1107, %f212, %p133;
selp.f32 %f1109, 0fC3170000, 0fC2FE0000, %p133;
mov.b32 %r355, %f1108;
and.b32 %r356, %r355, 8388607;
or.b32 %r357, %r356, 1065353216;
mov.b32 %f1110, %r357;
shr.u32 %r358, %r355, 23;
cvt.rn.f32.u32 %f1111, %r358;
add.f32 %f1112, %f1109, %f1111;
setp.gt.f32 %p134, %f1110, 0f3FB504F3;
mul.f32 %f1113, %f1110, 0f3F000000;
add.f32 %f1114, %f1112, 0f3F800000;
selp.f32 %f1115, %f1114, %f1112, %p134;
selp.f32 %f1116, %f1113, %f1110, %p134;
add.f32 %f1117, %f1116, 0fBF800000;
add.f32 %f1118, %f1116, 0f3F800000;
rcp.approx.ftz.f32 %f1119, %f1118;
add.f32 %f1120, %f1117, %f1117;
mul.f32 %f1121, %f1120, %f1119;
mul.f32 %f1122, %f1121, %f1121;
mov.f32 %f1123, 0f3C4CAF63;
mov.f32 %f1124, 0f3B18F0FE;
fma.rn.f32 %f1125, %f1124, %f1122, %f1123;
mov.f32 %f1126, 0f3DAAAABD;
fma.rn.f32 %f1127, %f1125, %f1122, %f1126;
mul.rn.f32 %f1128, %f1127, %f1122;
mul.rn.f32 %f1129, %f1128, %f1121;
sub.f32 %f1130, %f1117, %f1121;
add.f32 %f1131, %f1130, %f1130;
neg.f32 %f1132, %f1121;
fma.rn.f32 %f1133, %f1132, %f1117, %f1131;
mul.rn.f32 %f1134, %f1119, %f1133;
add.f32 %f1135, %f1129, %f1121;
sub.f32 %f1136, %f1121, %f1135;
add.f32 %f1137, %f1129, %f1136;
add.f32 %f1138, %f1134, %f1137;
add.f32 %f1139, %f1135, %f1138;
sub.f32 %f1140, %f1135, %f1139;
add.f32 %f1141, %f1138, %f1140;
mov.f32 %f1142, 0f3F317200;
mul.rn.f32 %f1143, %f1115, %f1142;
mov.f32 %f1144, 0f35BFBE8E;
mul.rn.f32 %f1145, %f1115, %f1144;
add.f32 %f1146, %f1143, %f1139;
sub.f32 %f1147, %f1143, %f1146;
add.f32 %f1148, %f1139, %f1147;
add.f32 %f1149, %f1141, %f1148;
add.f32 %f1150, %f1145, %f1149;
add.f32 %f1151, %f1146, %f1150;
sub.f32 %f1152, %f1146, %f1151;
add.f32 %f1153, %f1150, %f1152;
mov.f32 %f1154, 0f3ED55555;
mul.rn.f32 %f1155, %f1154, %f1151;
neg.f32 %f1156, %f1155;
fma.rn.f32 %f1157, %f1154, %f1151, %f1156;
fma.rn.f32 %f1158, %f1154, %f1153, %f1157;
mov.f32 %f1159, 0f00000000;
fma.rn.f32 %f1160, %f1159, %f1151, %f1158;
add.rn.f32 %f1161, %f1155, %f1160;
neg.f32 %f1162, %f1161;
add.rn.f32 %f1163, %f1155, %f1162;
add.rn.f32 %f1164, %f1163, %f1160;
mov.b32 %r359, %f1161;
setp.eq.s32 %p135, %r359, 1118925336;
add.s32 %r360, %r359, -1;
mov.b32 %f1165, %r360;
add.f32 %f1166, %f1164, 0f37000000;
selp.f32 %f213, %f1166, %f1164, %p135;
selp.f32 %f1167, %f1165, %f1161, %p135;
mov.f32 %f1168, 0f3FB8AA3B;
mul.rn.f32 %f1169, %f1167, %f1168;
cvt.rzi.f32.f32 %f1170, %f1169;
abs.f32 %f1171, %f1170;
setp.gt.f32 %p136, %f1171, 0f42FC0000;
mov.b32 %r361, %f1170;
and.b32 %r362, %r361, -2147483648;
or.b32 %r363, %r362, 1123811328;
mov.b32 %f1172, %r363;
selp.f32 %f1173, %f1172, %f1170, %p136;
mov.f32 %f1174, 0fBF317218;
fma.rn.f32 %f1175, %f1173, %f1174, %f1167;
mov.f32 %f1176, 0f3102E308;
fma.rn.f32 %f1177, %f1173, %f1176, %f1175;
mul.f32 %f1178, %f1177, 0f3FB8AA3B;
add.f32 %f1179, %f1173, 0f4B40007F;
mov.b32 %r364, %f1179;
shl.b32 %r365, %r364, 23;
mov.b32 %f1180, %r365;
ex2.approx.ftz.f32 %f1181, %f1178;
mul.f32 %f214, %f1181, %f1180;
setp.eq.f32 %p137, %f214, 0f7F800000;
mov.f32 %f1406, 0f7F800000;
@%p137 bra $L__BB0_112;
fma.rn.f32 %f1406, %f214, %f213, %f214;
$L__BB0_112:
setp.lt.f32 %p138, %f198, 0f00000000;
and.pred %p8, %p138, %p125;
setp.eq.f32 %p140, %f198, 0f00000000;
@%p140 bra $L__BB0_116;
bra.uni $L__BB0_113;
$L__BB0_116:
add.f32 %f1186, %f198, %f198;
selp.f32 %f1408, %f1186, 0f00000000, %p125;
bra.uni $L__BB0_117;
$L__BB0_113:
mov.b32 %r366, %f1406;
xor.b32 %r367, %r366, -2147483648;
mov.b32 %f1182, %r367;
selp.f32 %f1408, %f1182, %f1406, %p8;
setp.geu.f32 %p141, %f198, 0f00000000;
@%p141 bra $L__BB0_117;
mov.f32 %f1183, 0f3ED55555;
cvt.rzi.f32.f32 %f1184, %f1183;
setp.eq.f32 %p142, %f1184, 0f3ED55555;
@%p142 bra $L__BB0_117;
mov.f32 %f1408, 0f7FFFFFFF;
$L__BB0_117:
add.f32 %f1187, %f212, 0f3ED55555;
mov.b32 %r368, %f1187;
setp.lt.s32 %p144, %r368, 2139095040;
@%p144 bra $L__BB0_122;
setp.gtu.f32 %p145, %f212, 0f7F800000;
@%p145 bra $L__BB0_121;
bra.uni $L__BB0_119;
$L__BB0_121:
add.f32 %f1408, %f198, 0f3ED55555;
bra.uni $L__BB0_122;
$L__BB0_119:
setp.neu.f32 %p146, %f212, 0f7F800000;
@%p146 bra $L__BB0_122;
selp.f32 %f1408, 0fFF800000, 0f7F800000, %p8;
$L__BB0_122:
abs.f32 %f223, %f199;
setp.lt.f32 %p147, %f223, 0f00800000;
mul.f32 %f1189, %f223, 0f4B800000;
selp.f32 %f1190, %f1189, %f223, %p147;
selp.f32 %f1191, 0fC3170000, 0fC2FE0000, %p147;
mov.b32 %r369, %f1190;
and.b32 %r370, %r369, 8388607;
or.b32 %r371, %r370, 1065353216;
mov.b32 %f1192, %r371;
shr.u32 %r372, %r369, 23;
cvt.rn.f32.u32 %f1193, %r372;
add.f32 %f1194, %f1191, %f1193;
setp.gt.f32 %p148, %f1192, 0f3FB504F3;
mul.f32 %f1195, %f1192, 0f3F000000;
add.f32 %f1196, %f1194, 0f3F800000;
selp.f32 %f1197, %f1196, %f1194, %p148;
selp.f32 %f1198, %f1195, %f1192, %p148;
add.f32 %f1199, %f1198, 0fBF800000;
add.f32 %f1200, %f1198, 0f3F800000;
rcp.approx.ftz.f32 %f1201, %f1200;
add.f32 %f1202, %f1199, %f1199;
mul.f32 %f1203, %f1202, %f1201;
mul.f32 %f1204, %f1203, %f1203;
mov.f32 %f1205, 0f3C4CAF63;
mov.f32 %f1206, 0f3B18F0FE;
fma.rn.f32 %f1207, %f1206, %f1204, %f1205;
mov.f32 %f1208, 0f3DAAAABD;
fma.rn.f32 %f1209, %f1207, %f1204, %f1208;
mul.rn.f32 %f1210, %f1209, %f1204;
mul.rn.f32 %f1211, %f1210, %f1203;
sub.f32 %f1212, %f1199, %f1203;
add.f32 %f1213, %f1212, %f1212;
neg.f32 %f1214, %f1203;
fma.rn.f32 %f1215, %f1214, %f1199, %f1213;
mul.rn.f32 %f1216, %f1201, %f1215;
add.f32 %f1217, %f1211, %f1203;
sub.f32 %f1218, %f1203, %f1217;
add.f32 %f1219, %f1211, %f1218;
add.f32 %f1220, %f1216, %f1219;
add.f32 %f1221, %f1217, %f1220;
sub.f32 %f1222, %f1217, %f1221;
add.f32 %f1223, %f1220, %f1222;
mov.f32 %f1224, 0f3F317200;
mul.rn.f32 %f1225, %f1197, %f1224;
mov.f32 %f1226, 0f35BFBE8E;
mul.rn.f32 %f1227, %f1197, %f1226;
add.f32 %f1228, %f1225, %f1221;
sub.f32 %f1229, %f1225, %f1228;
add.f32 %f1230, %f1221, %f1229;
add.f32 %f1231, %f1223, %f1230;
add.f32 %f1232, %f1227, %f1231;
add.f32 %f1233, %f1228, %f1232;
sub.f32 %f1234, %f1228, %f1233;
add.f32 %f1235, %f1232, %f1234;
mov.f32 %f1236, 0f3ED55555;
mul.rn.f32 %f1237, %f1236, %f1233;
neg.f32 %f1238, %f1237;
fma.rn.f32 %f1239, %f1236, %f1233, %f1238;
fma.rn.f32 %f1240, %f1236, %f1235, %f1239;
mov.f32 %f1241, 0f00000000;
fma.rn.f32 %f1242, %f1241, %f1233, %f1240;
add.rn.f32 %f1243, %f1237, %f1242;
neg.f32 %f1244, %f1243;
add.rn.f32 %f1245, %f1237, %f1244;
add.rn.f32 %f1246, %f1245, %f1242;
mov.b32 %r373, %f1243;
setp.eq.s32 %p149, %r373, 1118925336;
add.s32 %r374, %r373, -1;
mov.b32 %f1247, %r374;
add.f32 %f1248, %f1246, 0f37000000;
selp.f32 %f224, %f1248, %f1246, %p149;
selp.f32 %f1249, %f1247, %f1243, %p149;
mov.f32 %f1250, 0f3FB8AA3B;
mul.rn.f32 %f1251, %f1249, %f1250;
cvt.rzi.f32.f32 %f1252, %f1251;
abs.f32 %f1253, %f1252;
setp.gt.f32 %p150, %f1253, 0f42FC0000;
mov.b32 %r375, %f1252;
and.b32 %r376, %r375, -2147483648;
or.b32 %r377, %r376, 1123811328;
mov.b32 %f1254, %r377;
selp.f32 %f1255, %f1254, %f1252, %p150;
mov.f32 %f1256, 0fBF317218;
fma.rn.f32 %f1257, %f1255, %f1256, %f1249;
mov.f32 %f1258, 0f3102E308;
fma.rn.f32 %f1259, %f1255, %f1258, %f1257;
mul.f32 %f1260, %f1259, 0f3FB8AA3B;
add.f32 %f1261, %f1255, 0f4B40007F;
mov.b32 %r378, %f1261;
shl.b32 %r379, %r378, 23;
mov.b32 %f1262, %r379;
ex2.approx.ftz.f32 %f1263, %f1260;
mul.f32 %f225, %f1263, %f1262;
setp.eq.f32 %p151, %f225, 0f7F800000;
mov.f32 %f1409, 0f7F800000;
@%p151 bra $L__BB0_124;
fma.rn.f32 %f1409, %f225, %f224, %f225;
$L__BB0_124:
setp.lt.f32 %p152, %f199, 0f00000000;
and.pred %p9, %p152, %p125;
setp.eq.f32 %p154, %f199, 0f00000000;
@%p154 bra $L__BB0_128;
bra.uni $L__BB0_125;
$L__BB0_128:
add.f32 %f1268, %f199, %f199;
selp.f32 %f1411, %f1268, 0f00000000, %p125;
bra.uni $L__BB0_129;
$L__BB0_125:
mov.b32 %r380, %f1409;
xor.b32 %r381, %r380, -2147483648;
mov.b32 %f1264, %r381;
selp.f32 %f1411, %f1264, %f1409, %p9;
setp.geu.f32 %p155, %f199, 0f00000000;
@%p155 bra $L__BB0_129;
mov.f32 %f1265, 0f3ED55555;
cvt.rzi.f32.f32 %f1266, %f1265;
setp.eq.f32 %p156, %f1266, 0f3ED55555;
@%p156 bra $L__BB0_129;
mov.f32 %f1411, 0f7FFFFFFF;
$L__BB0_129:
add.f32 %f1269, %f223, 0f3ED55555;
mov.b32 %r382, %f1269;
setp.lt.s32 %p158, %r382, 2139095040;
@%p158 bra $L__BB0_134;
setp.gtu.f32 %p159, %f223, 0f7F800000;
@%p159 bra $L__BB0_133;
bra.uni $L__BB0_131;
$L__BB0_133:
add.f32 %f1411, %f199, 0f3ED55555;
bra.uni $L__BB0_134;
$L__BB0_131:
setp.neu.f32 %p160, %f223, 0f7F800000;
@%p160 bra $L__BB0_134;
selp.f32 %f1411, 0fFF800000, 0f7F800000, %p9;
$L__BB0_134:
fma.rn.f32 %f1270, %f1405, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p161, %f197, 0f3F800000;
mov.f32 %f1271, 0f3F800000;
selp.f32 %f1272, 0f3F7FFFFF, %f1270, %p161;
mul.f32 %f1273, %f197, 0f414EB852;
setp.lt.f32 %p162, %f197, 0f3B4D2E1C;
selp.f32 %f1274, %f1273, %f1272, %p162;
fma.rn.f32 %f1275, %f1408, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p163, %f198, 0f3F800000;
selp.f32 %f1276, 0f3F7FFFFF, %f1275, %p163;
mul.f32 %f1277, %f198, 0f414EB852;
setp.lt.f32 %p164, %f198, 0f3B4D2E1C;
selp.f32 %f1278, %f1277, %f1276, %p164;
fma.rn.f32 %f1279, %f1411, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p165, %f199, 0f3F800000;
selp.f32 %f1280, 0f3F7FFFFF, %f1279, %p165;
mul.f32 %f1281, %f199, 0f414EB852;
setp.lt.f32 %p166, %f199, 0f3B4D2E1C;
selp.f32 %f1282, %f1281, %f1280, %p166;
min.f32 %f1283, %f1274, %f1271;
mov.f32 %f1284, 0f00000000;
max.f32 %f1285, %f1284, %f1283;
mul.f32 %f1286, %f1285, 0f43800000;
cvt.rzi.u32.f32 %r383, %f1286;
min.u32 %r384, %r383, 255;
min.f32 %f1287, %f1278, %f1271;
max.f32 %f1288, %f1284, %f1287;
mul.f32 %f1289, %f1288, 0f43800000;
cvt.rzi.u32.f32 %r385, %f1289;
min.u32 %r386, %r385, 255;
min.f32 %f1290, %f1282, %f1271;
max.f32 %f1291, %f1284, %f1290;
mul.f32 %f1292, %f1291, 0f43800000;
cvt.rzi.u32.f32 %r387, %f1292;
min.u32 %r388, %r387, 255;
shl.b64 %rd29, %rd10, 2;
add.s64 %rd30, %rd9, %rd29;
cvt.u16.u32 %rs20, %r388;
cvt.u16.u32 %rs21, %r386;
cvt.u16.u32 %rs22, %r384;
mov.u16 %rs23, 255;
st.global.v4.u8 [%rd30], {%rs22, %rs21, %rs20, %rs23};
$L__BB0_135:
and.b32 %r389, %r417, 4;
setp.eq.s32 %p167, %r389, 0;
@%p167 bra $L__BB0_139;
ld.const.u32 %r390, [params+108];
setp.eq.s32 %p168, %r390, 0;
ld.const.u64 %rd31, [params+224];
cvta.to.global.u64 %rd32, %rd31;
ld.const.u32 %r391, [params+216];
mad.lo.s32 %r392, %r391, %r4, %r3;
mul.wide.u32 %rd33, %r392, 8;
add.s64 %rd11, %rd32, %rd33;
@%p168 bra $L__BB0_138;
ld.global.v4.u16 {%rs31, %rs32, %rs33, %rs34}, [%rd11];
// begin inline asm
{ cvt.f32.f16 %f1293, %rs31;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1294, %rs32;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1295, %rs33;}
// end inline asm
add.f32 %f1296, %f1377, %f1293;
add.f32 %f1297, %f1378, %f1294;
add.f32 %f1298, %f1379, %f1295;
mov.f32 %f1299, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f1298;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1297;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1296;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f1299;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs27, %rs28, %rs29, %rs30};
bra.uni $L__BB0_139;
$L__BB0_138:
mov.f32 %f1303, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f1303;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f1379;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f1378;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f1377;}
// end inline asm
st.global.v4.u16 [%rd11], {%rs35, %rs36, %rs37, %rs38};
$L__BB0_139:
and.b32 %r393, %r417, 64;
setp.eq.s32 %p169, %r393, 0;
@%p169 bra $L__BB0_151;
mul.f32 %f1304, %f1381, %f1381;
fma.rn.f32 %f1305, %f1380, %f1380, %f1304;
fma.rn.f32 %f1306, %f1382, %f1382, %f1305;
sqrt.rn.f32 %f1307, %f1306;
rcp.rn.f32 %f1308, %f1307;
mul.f32 %f1309, %f1380, %f1308;
mul.f32 %f1310, %f1381, %f1308;
mul.f32 %f1311, %f1382, %f1308;
ld.const.u64 %rd34, [params+208];
cvta.to.global.u64 %rd35, %rd34;
ld.const.u32 %r394, [params+200];
mad.lo.s32 %r395, %r394, %r4, %r3;
fma.rn.f32 %f1312, %f1309, 0f3F000000, 0f3F000000;
mul.f32 %f1313, %f1312, 0f437F0000;
cvt.rzi.u32.f32 %r396, %f1313;
fma.rn.f32 %f1314, %f1310, 0f3F000000, 0f3F000000;
mul.f32 %f1315, %f1314, 0f437F0000;
cvt.rzi.u32.f32 %r397, %f1315;
fma.rn.f32 %f1316, %f1311, 0f3F000000, 0f3F000000;
mul.f32 %f1317, %f1316, 0f437F0000;
cvt.rzi.u32.f32 %r398, %f1317;
mul.wide.u32 %rd36, %r395, 4;
add.s64 %rd37, %rd35, %rd36;
cvt.u16.u32 %rs39, %r398;
cvt.u16.u32 %rs40, %r397;
cvt.u16.u32 %rs41, %r396;
mov.u16 %rs42, 255;
st.global.v4.u8 [%rd37], {%rs41, %rs40, %rs39, %rs42};
$L__BB0_151:
ret;
}