Files
2026-08-06 18:56:32 +03:00

2281 lines
62 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<161>;
.reg .b16 %rs<175>;
.reg .f32 %f<1260>;
.reg .b32 %r<359>;
.reg .f64 %fd<5>;
.reg .b64 %rd<106>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r55), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r56), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u32 %r3, [params+540];
ld.const.u64 %rd28, [params+400];
cvta.to.global.u64 %rd29, %rd28;
ld.const.u32 %r61, [params+392];
mad.lo.s32 %r62, %r61, %r56, %r55;
mul.wide.u32 %rd30, %r62, 4;
add.s64 %rd2, %rd29, %rd30;
ld.global.v2.u8 {%rs16, %rs170}, [%rd2];
or.b16 %rs18, %rs16, %rs170;
and.b16 %rs19, %rs18, 255;
setp.eq.s16 %p7, %rs19, 0;
@%p7 bra $L__BB0_2;
ld.global.u8 %rs169, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs169, [%rd2+2];
setp.eq.s16 %p8, %rs169, 0;
mov.f32 %f1176, 0f00000000;
mov.u16 %rs170, 0;
mov.f32 %f1177, %f1176;
mov.f32 %f1178, %f1176;
@%p8 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f244, %rs16;
div.rn.f32 %f245, %f244, 0f437F0000;
fma.rn.f32 %f246, %f245, 0f40000000, 0fBF800000;
and.b16 %rs22, %rs170, 255;
cvt.rn.f32.u16 %f247, %rs22;
div.rn.f32 %f248, %f247, 0f437F0000;
fma.rn.f32 %f249, %f248, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f250, %rs169;
div.rn.f32 %f251, %f250, 0f437F0000;
fma.rn.f32 %f252, %f251, 0f40000000, 0fBF800000;
mul.f32 %f253, %f249, %f249;
fma.rn.f32 %f254, %f246, %f246, %f253;
fma.rn.f32 %f255, %f252, %f252, %f254;
sqrt.rn.f32 %f256, %f255;
rcp.rn.f32 %f257, %f256;
mul.f32 %f1178, %f257, %f252;
mul.f32 %f1177, %f257, %f249;
mul.f32 %f1176, %f246, %f257;
$L__BB0_4:
ld.const.v2.u32 {%r63, %r64}, [params];
add.s32 %r4, %r63, %r55;
add.s32 %r5, %r64, %r56;
setp.eq.f32 %p9, %f1176, 0f00000000;
setp.eq.f32 %p10, %f1177, 0f00000000;
and.pred %p11, %p9, %p10;
setp.eq.f32 %p12, %f1178, 0f00000000;
and.pred %p13, %p12, %p11;
@%p13 bra $L__BB0_131;
bra.uni $L__BB0_5;
$L__BB0_131:
ld.const.u32 %r52, [params+104];
and.b32 %r314, %r52, 1;
setp.eq.b32 %p151, %r314, 1;
mov.pred %p152, 0;
xor.pred %p153, %p151, %p152;
not.pred %p154, %p153;
@%p154 bra $L__BB0_133;
ld.const.u64 %rd82, [params+144];
cvta.to.global.u64 %rd83, %rd82;
ld.const.u32 %r315, [params+136];
mad.lo.s32 %r316, %r315, %r5, %r4;
mul.wide.u32 %rd84, %r316, 4;
add.s64 %rd85, %rd83, %rd84;
mov.u16 %rs108, 0;
st.global.v4.u8 [%rd85], {%rs108, %rs108, %rs108, %rs108};
$L__BB0_133:
and.b32 %r317, %r52, 4;
setp.eq.s32 %p155, %r317, 0;
ld.const.u32 %r358, [params+108];
@%p155 bra $L__BB0_137;
setp.eq.s32 %p156, %r358, 0;
ld.const.u64 %rd86, [params+224];
cvta.to.global.u64 %rd87, %rd86;
ld.const.u32 %r318, [params+216];
mad.lo.s32 %r319, %r318, %r5, %r4;
mul.wide.u32 %rd88, %r319, 8;
add.s64 %rd23, %rd87, %rd88;
@%p156 bra $L__BB0_136;
ld.global.v4.u16 {%rs115, %rs116, %rs117, %rs118}, [%rd23];
// begin inline asm
{ cvt.f32.f16 %f1131, %rs115;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1132, %rs116;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1133, %rs117;}
// end inline asm
add.f32 %f1134, %f1131, 0f00000000;
add.f32 %f1135, %f1132, 0f00000000;
add.f32 %f1136, %f1133, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs114, %f1136;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f1135;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs112, %f1134;}
// end inline asm
mov.u16 %rs119, 0;
st.global.v4.u16 [%rd23], {%rs112, %rs113, %rs114, %rs119};
bra.uni $L__BB0_137;
$L__BB0_5:
ld.const.u64 %rd31, [params+432];
cvta.to.global.u64 %rd32, %rd31;
ld.const.u32 %r67, [params+424];
mad.lo.s32 %r68, %r67, %r56, %r55;
mul.wide.u32 %rd33, %r68, 12;
add.s64 %rd34, %rd32, %rd33;
ld.global.f32 %f258, [%rd34];
mul.f32 %f259, %f258, 0f3456BF95;
ld.global.f32 %f260, [%rd34+4];
mul.f32 %f261, %f260, 0f3456BF95;
ld.global.f32 %f262, [%rd34+8];
mul.f32 %f263, %f262, 0f3456BF95;
abs.f32 %f264, %f1176;
div.rn.f32 %f265, %f259, %f264;
abs.f32 %f266, %f1177;
div.rn.f32 %f267, %f261, %f266;
abs.f32 %f268, %f1178;
div.rn.f32 %f269, %f263, %f268;
abs.f32 %f270, %f265;
abs.f32 %f271, %f267;
abs.f32 %f272, %f269;
mov.f32 %f273, 0f38D1B717;
max.f32 %f274, %f270, %f273;
max.f32 %f275, %f271, %f273;
max.f32 %f276, %f272, %f273;
fma.rn.f32 %f10, %f1176, %f274, %f258;
fma.rn.f32 %f11, %f1177, %f275, %f260;
fma.rn.f32 %f12, %f1178, %f276, %f262;
ld.const.u32 %r6, [params+588];
setp.gt.f32 %p14, %f264, %f268;
neg.f32 %f277, %f1177;
selp.f32 %f278, %f277, 0f00000000, %p14;
neg.f32 %f279, %f1178;
selp.f32 %f280, %f1176, %f279, %p14;
selp.f32 %f281, 0f00000000, %f1177, %p14;
mul.f32 %f282, %f280, %f280;
fma.rn.f32 %f283, %f278, %f278, %f282;
fma.rn.f32 %f284, %f281, %f281, %f283;
sqrt.rn.f32 %f285, %f284;
rcp.rn.f32 %f286, %f285;
mul.f32 %f13, %f278, %f286;
mul.f32 %f14, %f280, %f286;
mul.f32 %f15, %f281, %f286;
ld.const.u64 %rd35, [params+128];
cvta.to.global.u64 %rd36, %rd35;
ld.const.u32 %r69, [params+120];
mad.lo.s32 %r70, %r69, %r56, %r55;
mul.wide.u32 %rd37, %r70, 4;
add.s64 %rd38, %rd36, %rd37;
ld.global.u32 %r349, [%rd38];
ld.const.u64 %rd39, [params+464];
cvta.to.global.u64 %rd40, %rd39;
ld.const.u32 %r71, [params+456];
mad.lo.s32 %r72, %r71, %r56, %r55;
mul.wide.u32 %rd41, %r72, 4;
add.s64 %rd3, %rd40, %rd41;
ld.global.v2.u8 {%rs24, %rs173}, [%rd3];
or.b16 %rs26, %rs24, %rs173;
and.b16 %rs27, %rs26, 255;
setp.eq.s16 %p15, %rs27, 0;
@%p15 bra $L__BB0_7;
ld.global.u8 %rs171, [%rd3+2];
bra.uni $L__BB0_8;
$L__BB0_7:
ld.global.u8 %rs171, [%rd3+2];
setp.eq.s16 %p16, %rs171, 0;
mov.f32 %f1179, 0f00000000;
mov.u16 %rs173, 0;
mov.u16 %rs174, %rs173;
mov.f32 %f1180, %f1179;
mov.f32 %f1181, %f1179;
@%p16 bra $L__BB0_9;
$L__BB0_8:
mov.u16 %rs174, %rs173;
cvt.rn.f32.u16 %f290, %rs24;
div.rn.f32 %f291, %f290, 0f437F0000;
fma.rn.f32 %f292, %f291, 0f40000000, 0fBF800000;
and.b16 %rs32, %rs174, 255;
cvt.rn.f32.u16 %f293, %rs32;
div.rn.f32 %f294, %f293, 0f437F0000;
fma.rn.f32 %f295, %f294, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f296, %rs171;
div.rn.f32 %f297, %f296, 0f437F0000;
fma.rn.f32 %f298, %f297, 0f40000000, 0fBF800000;
mul.f32 %f299, %f295, %f295;
fma.rn.f32 %f300, %f292, %f292, %f299;
fma.rn.f32 %f301, %f298, %f298, %f300;
sqrt.rn.f32 %f302, %f301;
rcp.rn.f32 %f303, %f302;
mul.f32 %f1181, %f303, %f298;
mul.f32 %f1180, %f303, %f295;
mul.f32 %f1179, %f292, %f303;
mov.u16 %rs173, %rs171;
$L__BB0_9:
mul.f32 %f307, %f1177, %f1181;
mul.f32 %f308, %f1178, %f1180;
sub.f32 %f309, %f307, %f308;
mul.f32 %f310, %f1178, %f1179;
mul.f32 %f311, %f1176, %f1181;
sub.f32 %f312, %f310, %f311;
mul.f32 %f313, %f1176, %f1180;
mul.f32 %f314, %f1177, %f1179;
sub.f32 %f315, %f313, %f314;
ld.global.u8 %rs34, [%rd3+3];
setp.eq.s16 %p17, %rs34, 0;
selp.f32 %f22, 0fBF800000, 0f3F800000, %p17;
mul.f32 %f316, %f309, %f22;
mul.f32 %f317, %f312, %f22;
mul.f32 %f318, %f315, %f22;
mul.f32 %f319, %f316, 0f00000000;
mul.f32 %f320, %f317, 0f00000000;
mul.f32 %f321, %f318, 0f00000000;
fma.rn.f32 %f322, %f1179, 0f3F5105EC, %f319;
fma.rn.f32 %f323, %f1180, 0f3F5105EC, %f320;
fma.rn.f32 %f324, %f1181, 0f3F5105EC, %f321;
mul.f32 %f23, %f1176, 0f3F13CD3A;
add.f32 %f24, %f23, %f322;
mul.f32 %f25, %f1177, 0f3F13CD3A;
add.f32 %f26, %f25, %f323;
mul.f32 %f27, %f1178, 0f3F13CD3A;
add.f32 %f28, %f27, %f324;
or.b16 %rs35, %rs24, %rs174;
or.b16 %rs15, %rs35, %rs173;
and.b16 %rs36, %rs15, 255;
setp.eq.s16 %p18, %rs36, 0;
mov.f32 %f1185, 0f00000000;
mov.f32 %f1182, %f1185;
mov.f32 %f1183, %f1185;
mov.f32 %f1184, %f1185;
@%p18 bra $L__BB0_11;
cvt.rn.f32.u16 %f325, %rs24;
div.rn.f32 %f326, %f325, 0f437F0000;
fma.rn.f32 %f327, %f326, 0f40000000, 0fBF800000;
and.b16 %rs38, %rs174, 255;
cvt.rn.f32.u16 %f328, %rs38;
div.rn.f32 %f329, %f328, 0f437F0000;
fma.rn.f32 %f330, %f329, 0f40000000, 0fBF800000;
and.b16 %rs39, %rs173, 255;
cvt.rn.f32.u16 %f331, %rs39;
div.rn.f32 %f332, %f331, 0f437F0000;
fma.rn.f32 %f333, %f332, 0f40000000, 0fBF800000;
mul.f32 %f334, %f330, %f330;
fma.rn.f32 %f335, %f327, %f327, %f334;
fma.rn.f32 %f336, %f333, %f333, %f335;
sqrt.rn.f32 %f337, %f336;
rcp.rn.f32 %f338, %f337;
mul.f32 %f1184, %f338, %f333;
mul.f32 %f1183, %f338, %f330;
mul.f32 %f1182, %f327, %f338;
$L__BB0_11:
mul.f32 %f342, %f1177, %f1184;
mul.f32 %f343, %f1178, %f1183;
sub.f32 %f344, %f342, %f343;
mul.f32 %f345, %f1178, %f1182;
mul.f32 %f346, %f1176, %f1184;
sub.f32 %f347, %f345, %f346;
mul.f32 %f348, %f1176, %f1183;
mul.f32 %f349, %f1177, %f1182;
sub.f32 %f350, %f348, %f349;
mul.f32 %f351, %f344, %f22;
mul.f32 %f352, %f347, %f22;
mul.f32 %f353, %f350, %f22;
mul.f32 %f354, %f1182, 0f3ED105EC;
mul.f32 %f355, %f1183, 0f3ED105EC;
mul.f32 %f356, %f1184, 0f3ED105EC;
mul.f32 %f357, %f351, 0f3F3504F3;
mul.f32 %f358, %f352, 0f3F3504F3;
mul.f32 %f359, %f353, 0f3F3504F3;
sub.f32 %f360, %f357, %f354;
sub.f32 %f361, %f358, %f355;
sub.f32 %f362, %f359, %f356;
add.f32 %f35, %f23, %f360;
add.f32 %f36, %f25, %f361;
add.f32 %f37, %f27, %f362;
mov.f32 %f1186, %f1185;
mov.f32 %f1187, %f1185;
@%p18 bra $L__BB0_13;
cvt.rn.f32.u16 %f363, %rs24;
div.rn.f32 %f364, %f363, 0f437F0000;
fma.rn.f32 %f365, %f364, 0f40000000, 0fBF800000;
and.b16 %rs42, %rs174, 255;
cvt.rn.f32.u16 %f366, %rs42;
div.rn.f32 %f367, %f366, 0f437F0000;
fma.rn.f32 %f368, %f367, 0f40000000, 0fBF800000;
and.b16 %rs43, %rs173, 255;
cvt.rn.f32.u16 %f369, %rs43;
div.rn.f32 %f370, %f369, 0f437F0000;
fma.rn.f32 %f371, %f370, 0f40000000, 0fBF800000;
mul.f32 %f372, %f368, %f368;
fma.rn.f32 %f373, %f365, %f365, %f372;
fma.rn.f32 %f374, %f371, %f371, %f373;
sqrt.rn.f32 %f375, %f374;
rcp.rn.f32 %f376, %f375;
mul.f32 %f1187, %f376, %f371;
mul.f32 %f1186, %f376, %f368;
mul.f32 %f1185, %f365, %f376;
$L__BB0_13:
mul.f32 %f389, %f1177, %f1187;
mul.f32 %f390, %f1178, %f1186;
sub.f32 %f391, %f389, %f390;
mul.f32 %f392, %f1178, %f1185;
mul.f32 %f393, %f1176, %f1187;
sub.f32 %f394, %f392, %f393;
mul.f32 %f395, %f1176, %f1186;
mul.f32 %f396, %f1177, %f1185;
sub.f32 %f397, %f395, %f396;
mul.f32 %f398, %f391, %f22;
mul.f32 %f399, %f394, %f22;
mul.f32 %f400, %f397, %f22;
mul.f32 %f401, %f1185, 0f3ED105EC;
mul.f32 %f402, %f1186, 0f3ED105EC;
mul.f32 %f403, %f1187, 0f3ED105EC;
mul.f32 %f404, %f398, 0fBF3504F3;
mul.f32 %f405, %f399, 0fBF3504F3;
mul.f32 %f406, %f400, 0fBF3504F3;
sub.f32 %f407, %f404, %f401;
sub.f32 %f408, %f405, %f402;
sub.f32 %f409, %f406, %f403;
add.f32 %f44, %f23, %f407;
add.f32 %f45, %f25, %f408;
add.f32 %f46, %f27, %f409;
setp.lt.s32 %p20, %r3, 1;
mov.f32 %f1218, 0f00000000;
mov.f32 %f1219, %f1218;
mov.f32 %f1220, %f1218;
mov.f32 %f1221, %f1218;
mov.f32 %f1222, %f1218;
mov.f32 %f1223, %f1218;
mov.f32 %f1224, %f1218;
mov.f32 %f1225, %f1218;
mov.f32 %f1226, %f1218;
mov.f32 %f1227, %f1218;
mov.f32 %f1228, %f1218;
mov.f32 %f1229, %f1218;
@%p20 bra $L__BB0_44;
cvt.rn.f32.s32 %f422, %r3;
rcp.rn.f32 %f47, %f422;
mul.f32 %f48, %f10, 0f3456BF95;
mul.f32 %f49, %f11, 0f3456BF95;
mul.f32 %f50, %f12, 0f3456BF95;
ld.const.u64 %rd4, [params+592];
ld.const.u64 %rd5, [params+96];
mul.f32 %f423, %f1178, %f14;
mul.f32 %f424, %f1177, %f15;
sub.f32 %f51, %f423, %f424;
mul.f32 %f425, %f1176, %f15;
mul.f32 %f426, %f1178, %f13;
sub.f32 %f52, %f425, %f426;
mul.f32 %f427, %f1177, %f13;
mul.f32 %f428, %f1176, %f14;
sub.f32 %f53, %f427, %f428;
mov.u32 %r73, 0;
add.s64 %rd6, %rd1, 24;
mov.u64 %rd42, __cudart_i2opi_f;
abs.f32 %f491, %f48;
abs.f32 %f492, %f49;
max.f32 %f493, %f491, %f492;
abs.f32 %f494, %f50;
max.f32 %f495, %f493, %f494;
mov.u32 %r346, %r73;
$L__BB0_15:
mov.u32 %r348, %r73;
$L__BB0_16:
cvt.rn.f32.s32 %f1168, %r346;
mad.lo.s32 %r75, %r349, 1664525, 1013904223;
and.b32 %r76, %r75, 16777215;
cvt.rn.f32.u32 %f429, %r76;
fma.rn.f32 %f430, %f429, 0f33800000, %f1168;
mul.f32 %f79, %f47, %f430;
mad.lo.s32 %r349, %r75, 1664525, 1013904223;
and.b32 %r77, %r349, 16777215;
cvt.rn.f32.u32 %f431, %r77;
cvt.rn.f32.s32 %f432, %r348;
fma.rn.f32 %f433, %f431, 0f33800000, %f432;
mul.f32 %f434, %f47, %f433;
mul.f32 %f435, %f79, %f79;
mov.f32 %f436, 0f3F800000;
sub.f32 %f437, %f436, %f435;
mov.f32 %f438, 0f00000000;
max.f32 %f439, %f438, %f437;
sqrt.rn.f32 %f80, %f439;
mul.f32 %f81, %f434, 0f40C90FDB;
mul.f32 %f440, %f81, 0f3F22F983;
cvt.rni.s32.f32 %r353, %f440;
cvt.rn.f32.s32 %f441, %r353;
mov.f32 %f442, 0fBFC90FDA;
fma.rn.f32 %f443, %f441, %f442, %f81;
mov.f32 %f444, 0fB3A22168;
fma.rn.f32 %f445, %f441, %f444, %f443;
mov.f32 %f446, 0fA7C234C5;
fma.rn.f32 %f1215, %f441, %f446, %f445;
abs.f32 %f83, %f81;
setp.ltu.f32 %p21, %f83, 0f47CE4780;
mov.f32 %f1212, %f1215;
@%p21 bra $L__BB0_24;
setp.eq.f32 %p22, %f83, 0f7F800000;
@%p22 bra $L__BB0_23;
bra.uni $L__BB0_18;
$L__BB0_23:
mov.f32 %f449, 0f00000000;
mul.rn.f32 %f1212, %f81, %f449;
mov.u32 %r353, 0;
bra.uni $L__BB0_24;
$L__BB0_18:
mov.b32 %r14, %f81;
bfe.u32 %r79, %r14, 23, 8;
mov.u64 %rd103, 0;
mov.u32 %r350, 0;
mov.u64 %rd101, %rd1;
mov.u64 %rd102, %rd42;
$L__BB0_19:
.pragma "nounroll";
mov.b32 %r329, %f81;
shl.b32 %r328, %r329, 8;
or.b32 %r327, %r328, -2147483648;
ld.global.nc.u32 %r81, [%rd102];
mad.wide.u32 %rd44, %r81, %r327, %rd103;
shr.u64 %rd103, %rd44, 32;
st.local.u32 [%rd101], %rd44;
add.s64 %rd102, %rd102, 4;
add.s64 %rd101, %rd101, 4;
add.s32 %r350, %r350, 1;
setp.ne.s32 %p23, %r350, 6;
@%p23 bra $L__BB0_19;
add.s32 %r334, %r79, -128;
mov.b32 %r333, %f81;
bfe.u32 %r332, %r333, 23, 8;
add.s32 %r331, %r332, -128;
shr.u32 %r330, %r331, 5;
st.local.u32 [%rd6], %rd103;
mov.u32 %r83, 6;
sub.s32 %r84, %r83, %r330;
mul.wide.s32 %rd45, %r84, 4;
add.s64 %rd46, %rd1, %rd45;
ld.local.u32 %r351, [%rd46];
ld.local.u32 %r352, [%rd46+-4];
and.b32 %r23, %r331, 31;
setp.eq.s32 %p24, %r23, 0;
@%p24 bra $L__BB0_22;
mov.b32 %r342, %f81;
bfe.u32 %r341, %r342, 23, 8;
add.s32 %r340, %r341, -128;
shr.u32 %r339, %r340, 5;
mov.u32 %r338, 4;
sub.s32 %r337, %r338, %r339;
mov.u32 %r85, 32;
sub.s32 %r86, %r85, %r23;
shr.u32 %r87, %r352, %r86;
shl.b32 %r88, %r351, %r23;
add.s32 %r351, %r87, %r88;
mul.wide.s32 %rd47, %r337, 4;
add.s64 %rd48, %rd1, %rd47;
ld.local.u32 %r89, [%rd48];
shr.u32 %r90, %r89, %r86;
shl.b32 %r91, %r352, %r23;
add.s32 %r352, %r90, %r91;
$L__BB0_22:
mov.b32 %r335, %f81;
and.b32 %r92, %r335, -2147483648;
shr.u32 %r93, %r352, 30;
shl.b32 %r94, %r351, 2;
or.b32 %r95, %r93, %r94;
shr.u32 %r96, %r95, 31;
shr.u32 %r97, %r351, 30;
add.s32 %r98, %r96, %r97;
neg.s32 %r99, %r98;
setp.eq.s32 %p25, %r92, 0;
selp.b32 %r353, %r98, %r99, %p25;
setp.ne.s32 %p26, %r96, 0;
xor.b32 %r100, %r92, -2147483648;
selp.b32 %r101, %r100, %r92, %p26;
selp.b32 %r102, -1, 0, %p26;
xor.b32 %r103, %r95, %r102;
shl.b32 %r104, %r352, 2;
xor.b32 %r105, %r104, %r102;
cvt.u64.u32 %rd49, %r103;
cvt.u64.u32 %rd50, %r105;
bfi.b64 %rd51, %rd49, %rd50, 32, 32;
cvt.rn.f64.s64 %fd1, %rd51;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f447, %fd2;
setp.eq.s32 %p27, %r101, 0;
neg.f32 %f448, %f447;
selp.f32 %f1212, %f447, %f448, %p27;
$L__BB0_24:
add.s32 %r30, %r353, 1;
and.b32 %r31, %r30, 1;
setp.eq.s32 %p28, %r31, 0;
selp.f32 %f87, %f1212, 0f3F800000, %p28;
mul.rn.f32 %f88, %f1212, %f1212;
mov.f32 %f1213, 0fB94D4153;
@%p28 bra $L__BB0_26;
mov.f32 %f451, 0fBAB607ED;
mov.f32 %f452, 0f37CBAC00;
fma.rn.f32 %f1213, %f452, %f88, %f451;
$L__BB0_26:
selp.f32 %f453, 0f3C0885E4, 0f3D2AAABB, %p28;
fma.rn.f32 %f454, %f1213, %f88, %f453;
selp.f32 %f455, 0fBE2AAAA8, 0fBEFFFFFF, %p28;
fma.rn.f32 %f456, %f454, %f88, %f455;
mov.f32 %f457, 0f00000000;
fma.rn.f32 %f458, %f88, %f87, %f457;
fma.rn.f32 %f1214, %f456, %f458, %f87;
and.b32 %r107, %r30, 2;
setp.eq.s32 %p30, %r107, 0;
@%p30 bra $L__BB0_28;
mov.f32 %f460, 0fBF800000;
fma.rn.f32 %f1214, %f1214, %f460, %f457;
$L__BB0_28:
mul.f32 %f1167, %f81, 0f3F22F983;
cvt.rni.s32.f32 %r356, %f1167;
setp.ltu.f32 %p160, %f83, 0f47CE4780;
@%p160 bra $L__BB0_36;
setp.eq.f32 %p32, %f83, 0f7F800000;
@%p32 bra $L__BB0_35;
bra.uni $L__BB0_30;
$L__BB0_35:
mov.f32 %f463, 0f00000000;
mul.rn.f32 %f1215, %f81, %f463;
mov.u32 %r356, 0;
bra.uni $L__BB0_36;
$L__BB0_30:
mov.b32 %r32, %f81;
bfe.u32 %r108, %r32, 23, 8;
add.s32 %r33, %r108, -128;
shl.b32 %r109, %r32, 8;
or.b32 %r34, %r109, -2147483648;
shr.u32 %r35, %r33, 5;
mov.u64 %rd104, 0;
mov.u64 %rd105, %rd104;
$L__BB0_31:
.pragma "nounroll";
shl.b64 %rd54, %rd104, 2;
mov.u64 %rd55, __cudart_i2opi_f;
add.s64 %rd56, %rd55, %rd54;
ld.global.nc.u32 %r110, [%rd56];
mad.wide.u32 %rd57, %r110, %r34, %rd105;
shr.u64 %rd105, %rd57, 32;
add.s64 %rd58, %rd1, %rd54;
st.local.u32 [%rd58], %rd57;
cvt.u32.u64 %r111, %rd104;
add.s32 %r112, %r111, 1;
cvt.s64.s32 %rd104, %r112;
setp.ne.s32 %p33, %r112, 6;
@%p33 bra $L__BB0_31;
st.local.u32 [%rd6], %rd105;
mov.u32 %r113, 4;
sub.s32 %r36, %r113, %r35;
mov.u32 %r114, 6;
sub.s32 %r115, %r114, %r35;
mul.wide.s32 %rd59, %r115, 4;
add.s64 %rd60, %rd1, %rd59;
ld.local.u32 %r354, [%rd60];
ld.local.u32 %r355, [%rd60+-4];
and.b32 %r39, %r33, 31;
setp.eq.s32 %p34, %r39, 0;
@%p34 bra $L__BB0_34;
mov.u32 %r116, 32;
sub.s32 %r117, %r116, %r39;
shr.u32 %r118, %r355, %r117;
shl.b32 %r119, %r354, %r39;
add.s32 %r354, %r118, %r119;
mul.wide.s32 %rd61, %r36, 4;
add.s64 %rd62, %rd1, %rd61;
ld.local.u32 %r120, [%rd62];
shr.u32 %r121, %r120, %r117;
shl.b32 %r122, %r355, %r39;
add.s32 %r355, %r121, %r122;
$L__BB0_34:
and.b32 %r123, %r32, -2147483648;
shr.u32 %r124, %r355, 30;
shl.b32 %r125, %r354, 2;
or.b32 %r126, %r124, %r125;
shr.u32 %r127, %r126, 31;
shr.u32 %r128, %r354, 30;
add.s32 %r129, %r127, %r128;
neg.s32 %r130, %r129;
setp.eq.s32 %p35, %r123, 0;
selp.b32 %r356, %r129, %r130, %p35;
setp.ne.s32 %p36, %r127, 0;
xor.b32 %r131, %r123, -2147483648;
selp.b32 %r132, %r131, %r123, %p36;
selp.b32 %r133, -1, 0, %p36;
xor.b32 %r134, %r126, %r133;
shl.b32 %r135, %r355, 2;
xor.b32 %r136, %r135, %r133;
cvt.u64.u32 %rd63, %r134;
cvt.u64.u32 %rd64, %r136;
bfi.b64 %rd65, %rd63, %rd64, 32, 32;
cvt.rn.f64.s64 %fd3, %rd65;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f461, %fd4;
setp.eq.s32 %p37, %r132, 0;
neg.f32 %f462, %f461;
selp.f32 %f1215, %f461, %f462, %p37;
$L__BB0_36:
mul.f32 %f97, %f80, %f1214;
and.b32 %r46, %r356, 1;
setp.eq.s32 %p38, %r46, 0;
mul.rn.f32 %f99, %f1215, %f1215;
mov.f32 %f1216, 0fB94D4153;
@%p38 bra $L__BB0_38;
mov.f32 %f465, 0fBAB607ED;
mov.f32 %f466, 0f37CBAC00;
fma.rn.f32 %f1216, %f466, %f99, %f465;
$L__BB0_38:
selp.f32 %f1169, %f1215, 0f3F800000, %p38;
selp.f32 %f467, 0f3C0885E4, 0f3D2AAABB, %p38;
fma.rn.f32 %f468, %f1216, %f99, %f467;
selp.f32 %f469, 0fBE2AAAA8, 0fBEFFFFFF, %p38;
fma.rn.f32 %f470, %f468, %f99, %f469;
mov.f32 %f471, 0f00000000;
fma.rn.f32 %f472, %f99, %f1169, %f471;
fma.rn.f32 %f1217, %f470, %f472, %f1169;
and.b32 %r138, %r356, 2;
setp.eq.s32 %p40, %r138, 0;
@%p40 bra $L__BB0_40;
mov.f32 %f474, 0fBF800000;
fma.rn.f32 %f1217, %f1217, %f474, %f471;
$L__BB0_40:
mul.f32 %f475, %f80, %f1217;
mul.f32 %f476, %f13, %f475;
mul.f32 %f477, %f14, %f475;
mul.f32 %f478, %f15, %f475;
fma.rn.f32 %f479, %f51, %f97, %f476;
fma.rn.f32 %f480, %f52, %f97, %f477;
fma.rn.f32 %f481, %f53, %f97, %f478;
fma.rn.f32 %f105, %f1176, %f79, %f479;
fma.rn.f32 %f106, %f1177, %f79, %f480;
fma.rn.f32 %f107, %f1178, %f79, %f481;
setp.leu.f32 %p41, %f106, 0f00000000;
setp.ne.s32 %p42, %r6, 0;
and.pred %p43, %p42, %p41;
@%p43 bra $L__BB0_42;
mov.f32 %f496, 0f38D1B717;
max.f32 %f488, %f495, %f496;
neg.f32 %f497, %f105;
neg.f32 %f498, %f106;
neg.f32 %f499, %f107;
tex.cube.v4.f32.f32 {%f500, %f501, %f502, %f503}, [%rd4, {%f499, %f498, %f497, %f497}];
mov.b32 %r178, %f500;
mov.b32 %r179, %f501;
mov.b32 %r180, %f502;
mov.f32 %f489, 0f6C4ECB8F;
mov.f32 %f490, 0f00000000;
mov.u32 %r175, 2;
mov.u32 %r176, 1;
mov.u32 %r177, 3;
mov.u32 %r209, 0;
// begin inline asm
call(%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167,%r168,%r169,%r170),_optix_trace_typed_32,(%r209,%rd5,%f10,%f11,%f12,%f105,%f106,%f107,%f488,%f489,%f490,%r176,%r209,%r176,%r175,%r176,%r177,%r178,%r179,%r180,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209,%r209);
// end inline asm
mov.b32 %f504, %r139;
mov.b32 %f505, %r140;
mov.b32 %f506, %r141;
mul.f32 %f507, %f26, %f106;
fma.rn.f32 %f508, %f24, %f105, %f507;
fma.rn.f32 %f509, %f28, %f107, %f508;
cvt.sat.f32.f32 %f510, %f509;
fma.rn.f32 %f1224, %f510, %f504, %f1224;
fma.rn.f32 %f1225, %f510, %f505, %f1225;
fma.rn.f32 %f1226, %f510, %f506, %f1226;
mul.f32 %f511, %f36, %f106;
fma.rn.f32 %f512, %f35, %f105, %f511;
fma.rn.f32 %f513, %f37, %f107, %f512;
cvt.sat.f32.f32 %f514, %f513;
fma.rn.f32 %f1221, %f514, %f504, %f1221;
fma.rn.f32 %f1222, %f514, %f505, %f1222;
fma.rn.f32 %f1223, %f514, %f506, %f1223;
mul.f32 %f515, %f45, %f106;
fma.rn.f32 %f516, %f44, %f105, %f515;
fma.rn.f32 %f517, %f46, %f107, %f516;
cvt.sat.f32.f32 %f518, %f517;
fma.rn.f32 %f1218, %f518, %f504, %f1218;
fma.rn.f32 %f1219, %f518, %f505, %f1219;
fma.rn.f32 %f1220, %f518, %f506, %f1220;
mul.f32 %f519, %f1177, %f106;
fma.rn.f32 %f520, %f1176, %f105, %f519;
fma.rn.f32 %f521, %f1178, %f107, %f520;
cvt.sat.f32.f32 %f522, %f521;
fma.rn.f32 %f1227, %f522, %f504, %f1227;
fma.rn.f32 %f1228, %f522, %f505, %f1228;
fma.rn.f32 %f1229, %f522, %f506, %f1229;
$L__BB0_42:
add.s32 %r348, %r348, 1;
setp.lt.s32 %p44, %r348, %r3;
@%p44 bra $L__BB0_16;
add.s32 %r346, %r346, 1;
setp.lt.s32 %p45, %r346, %r3;
@%p45 bra $L__BB0_15;
$L__BB0_44:
mul.lo.s32 %r210, %r3, %r3;
cvt.rn.f32.s32 %f523, %r210;
rcp.rn.f32 %f524, %f523;
mul.f32 %f525, %f524, %f1227;
mul.f32 %f526, %f524, %f1228;
mul.f32 %f527, %f524, %f1229;
mul.f32 %f144, %f524, %f1224;
mul.f32 %f145, %f524, %f1225;
mul.f32 %f146, %f524, %f1226;
mul.f32 %f147, %f524, %f1221;
mul.f32 %f148, %f524, %f1222;
mul.f32 %f149, %f524, %f1223;
mul.f32 %f150, %f524, %f1218;
mul.f32 %f151, %f524, %f1219;
mul.f32 %f152, %f524, %f1220;
fma.rn.f32 %f528, %f524, %f1227, %f525;
fma.rn.f32 %f529, %f524, %f1228, %f526;
fma.rn.f32 %f530, %f524, %f1229, %f527;
ld.const.v4.f32 {%f531, %f532, %f533, %f534}, [params+576];
mul.f32 %f156, %f528, %f531;
mul.f32 %f157, %f529, %f532;
mul.f32 %f158, %f530, %f533;
ld.const.u32 %r49, [params+104];
and.b32 %r211, %r49, 1;
setp.eq.b32 %p46, %r211, 1;
mov.pred %p47, 0;
xor.pred %p48, %p46, %p47;
not.pred %p49, %p48;
@%p49 bra $L__BB0_118;
mov.f32 %f536, 0f3E666666;
cvt.rzi.f32.f32 %f537, %f536;
add.f32 %f538, %f537, %f537;
mov.f32 %f539, 0f3EE66666;
sub.f32 %f540, %f539, %f538;
abs.f32 %f159, %f540;
abs.f32 %f160, %f156;
setp.lt.f32 %p50, %f160, 0f00800000;
mul.f32 %f541, %f160, 0f4B800000;
selp.f32 %f542, %f541, %f160, %p50;
selp.f32 %f543, 0fC3170000, 0fC2FE0000, %p50;
mov.b32 %r212, %f542;
and.b32 %r213, %r212, 8388607;
or.b32 %r214, %r213, 1065353216;
mov.b32 %f544, %r214;
shr.u32 %r215, %r212, 23;
cvt.rn.f32.u32 %f545, %r215;
add.f32 %f546, %f543, %f545;
setp.gt.f32 %p51, %f544, 0f3FB504F3;
mul.f32 %f547, %f544, 0f3F000000;
add.f32 %f548, %f546, 0f3F800000;
selp.f32 %f549, %f548, %f546, %p51;
selp.f32 %f550, %f547, %f544, %p51;
add.f32 %f551, %f550, 0fBF800000;
add.f32 %f552, %f550, 0f3F800000;
rcp.approx.ftz.f32 %f553, %f552;
add.f32 %f554, %f551, %f551;
mul.f32 %f555, %f554, %f553;
mul.f32 %f556, %f555, %f555;
mov.f32 %f557, 0f3C4CAF63;
mov.f32 %f558, 0f3B18F0FE;
fma.rn.f32 %f559, %f558, %f556, %f557;
mov.f32 %f560, 0f3DAAAABD;
fma.rn.f32 %f561, %f559, %f556, %f560;
mul.rn.f32 %f562, %f561, %f556;
mul.rn.f32 %f563, %f562, %f555;
sub.f32 %f564, %f551, %f555;
add.f32 %f565, %f564, %f564;
neg.f32 %f566, %f555;
fma.rn.f32 %f567, %f566, %f551, %f565;
mul.rn.f32 %f568, %f553, %f567;
add.f32 %f569, %f563, %f555;
sub.f32 %f570, %f555, %f569;
add.f32 %f571, %f563, %f570;
add.f32 %f572, %f568, %f571;
add.f32 %f573, %f569, %f572;
sub.f32 %f574, %f569, %f573;
add.f32 %f575, %f572, %f574;
mov.f32 %f576, 0f3F317200;
mul.rn.f32 %f577, %f549, %f576;
mov.f32 %f578, 0f35BFBE8E;
mul.rn.f32 %f579, %f549, %f578;
add.f32 %f580, %f577, %f573;
sub.f32 %f581, %f577, %f580;
add.f32 %f582, %f573, %f581;
add.f32 %f583, %f575, %f582;
add.f32 %f584, %f579, %f583;
add.f32 %f585, %f580, %f584;
sub.f32 %f586, %f580, %f585;
add.f32 %f587, %f584, %f586;
mul.rn.f32 %f588, %f539, %f585;
neg.f32 %f589, %f588;
fma.rn.f32 %f590, %f539, %f585, %f589;
fma.rn.f32 %f591, %f539, %f587, %f590;
mov.f32 %f592, 0f00000000;
fma.rn.f32 %f593, %f592, %f585, %f591;
add.rn.f32 %f594, %f588, %f593;
neg.f32 %f595, %f594;
add.rn.f32 %f596, %f588, %f595;
add.rn.f32 %f597, %f596, %f593;
mov.b32 %r216, %f594;
setp.eq.s32 %p52, %r216, 1118925336;
add.s32 %r217, %r216, -1;
mov.b32 %f598, %r217;
add.f32 %f599, %f597, 0f37000000;
selp.f32 %f161, %f599, %f597, %p52;
selp.f32 %f600, %f598, %f594, %p52;
mov.f32 %f601, 0f3FB8AA3B;
mul.rn.f32 %f602, %f600, %f601;
cvt.rzi.f32.f32 %f603, %f602;
abs.f32 %f604, %f603;
setp.gt.f32 %p53, %f604, 0f42FC0000;
mov.b32 %r218, %f603;
and.b32 %r219, %r218, -2147483648;
or.b32 %r220, %r219, 1123811328;
mov.b32 %f605, %r220;
selp.f32 %f606, %f605, %f603, %p53;
mov.f32 %f607, 0fBF317218;
fma.rn.f32 %f608, %f606, %f607, %f600;
mov.f32 %f609, 0f3102E308;
fma.rn.f32 %f610, %f606, %f609, %f608;
mul.f32 %f611, %f610, 0f3FB8AA3B;
add.f32 %f612, %f606, 0f4B40007F;
mov.b32 %r221, %f612;
shl.b32 %r222, %r221, 23;
mov.b32 %f613, %r222;
ex2.approx.ftz.f32 %f614, %f611;
mul.f32 %f162, %f614, %f613;
setp.eq.f32 %p54, %f162, 0f7F800000;
mov.f32 %f1242, 0f7F800000;
@%p54 bra $L__BB0_47;
fma.rn.f32 %f1242, %f162, %f161, %f162;
$L__BB0_47:
setp.lt.f32 %p55, %f156, 0f00000000;
setp.eq.f32 %p56, %f159, 0f3F800000;
and.pred %p1, %p55, %p56;
setp.eq.f32 %p57, %f156, 0f00000000;
@%p57 bra $L__BB0_51;
bra.uni $L__BB0_48;
$L__BB0_51:
add.f32 %f619, %f156, %f156;
selp.f32 %f1244, %f619, 0f00000000, %p56;
bra.uni $L__BB0_52;
$L__BB0_136:
mov.f32 %f1139, 0f00000000;
mov.u32 %r358, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs122, %f1139;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs121, %f1139;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs120, %f1139;}
// end inline asm
mov.u16 %rs123, 0;
st.global.v4.u16 [%rd23], {%rs120, %rs121, %rs122, %rs123};
$L__BB0_137:
ld.const.u64 %rd89, [params+256];
cvta.to.global.u64 %rd90, %rd89;
ld.const.u32 %r321, [params+248];
mad.lo.s32 %r322, %r321, %r5, %r4;
mul.wide.u32 %rd91, %r322, 8;
add.s64 %rd24, %rd90, %rd91;
setp.eq.s32 %p157, %r358, 0;
@%p157 bra $L__BB0_139;
ld.global.v4.u16 {%rs130, %rs131, %rs132, %rs133}, [%rd24];
// begin inline asm
{ cvt.f32.f16 %f1140, %rs130;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1141, %rs131;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1142, %rs132;}
// end inline asm
add.f32 %f1143, %f1140, 0f00000000;
add.f32 %f1144, %f1141, 0f00000000;
add.f32 %f1145, %f1142, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs129, %f1145;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1144;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs127, %f1143;}
// end inline asm
mov.u16 %rs134, 0;
st.global.v4.u16 [%rd24], {%rs127, %rs128, %rs129, %rs134};
bra.uni $L__BB0_140;
$L__BB0_139:
mov.f32 %f1148, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs137, %f1148;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs136, %f1148;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs135, %f1148;}
// end inline asm
mov.u16 %rs138, 0;
st.global.v4.u16 [%rd24], {%rs135, %rs136, %rs137, %rs138};
$L__BB0_140:
ld.const.u64 %rd92, [params+272];
cvta.to.global.u64 %rd93, %rd92;
ld.const.u32 %r323, [params+264];
mad.lo.s32 %r324, %r323, %r5, %r4;
mul.wide.u32 %rd94, %r324, 8;
add.s64 %rd25, %rd93, %rd94;
@%p157 bra $L__BB0_142;
ld.global.v4.u16 {%rs145, %rs146, %rs147, %rs148}, [%rd25];
// begin inline asm
{ cvt.f32.f16 %f1149, %rs145;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1150, %rs146;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1151, %rs147;}
// end inline asm
add.f32 %f1152, %f1149, 0f00000000;
add.f32 %f1153, %f1150, 0f00000000;
add.f32 %f1154, %f1151, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs144, %f1154;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1153;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs142, %f1152;}
// end inline asm
mov.u16 %rs149, 0;
st.global.v4.u16 [%rd25], {%rs142, %rs143, %rs144, %rs149};
bra.uni $L__BB0_143;
$L__BB0_142:
mov.f32 %f1157, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs152, %f1157;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs151, %f1157;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs150, %f1157;}
// end inline asm
mov.u16 %rs153, 0;
st.global.v4.u16 [%rd25], {%rs150, %rs151, %rs152, %rs153};
$L__BB0_143:
ld.const.u64 %rd95, [params+288];
cvta.to.global.u64 %rd96, %rd95;
ld.const.u32 %r325, [params+280];
mad.lo.s32 %r326, %r325, %r5, %r4;
mul.wide.u32 %rd97, %r326, 8;
add.s64 %rd26, %rd96, %rd97;
@%p157 bra $L__BB0_145;
ld.global.v4.u16 {%rs160, %rs161, %rs162, %rs163}, [%rd26];
// begin inline asm
{ cvt.f32.f16 %f1158, %rs160;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1159, %rs161;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1160, %rs162;}
// end inline asm
add.f32 %f1161, %f1158, 0f00000000;
add.f32 %f1162, %f1159, 0f00000000;
add.f32 %f1163, %f1160, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs159, %f1163;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1162;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs157, %f1161;}
// end inline asm
mov.u16 %rs164, 0;
st.global.v4.u16 [%rd26], {%rs157, %rs158, %rs159, %rs164};
bra.uni $L__BB0_146;
$L__BB0_145:
mov.f32 %f1166, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs167, %f1166;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs166, %f1166;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs165, %f1166;}
// end inline asm
mov.u16 %rs168, 0;
st.global.v4.u16 [%rd26], {%rs165, %rs166, %rs167, %rs168};
bra.uni $L__BB0_146;
$L__BB0_48:
mov.b32 %r223, %f1242;
xor.b32 %r224, %r223, -2147483648;
mov.b32 %f615, %r224;
selp.f32 %f1244, %f615, %f1242, %p1;
setp.geu.f32 %p58, %f156, 0f00000000;
@%p58 bra $L__BB0_52;
mov.f32 %f616, 0f3EE66666;
cvt.rzi.f32.f32 %f617, %f616;
setp.eq.f32 %p59, %f617, 0f3EE66666;
@%p59 bra $L__BB0_52;
mov.f32 %f1244, 0f7FFFFFFF;
$L__BB0_52:
add.f32 %f620, %f160, 0f3EE66666;
mov.b32 %r225, %f620;
setp.lt.s32 %p61, %r225, 2139095040;
@%p61 bra $L__BB0_57;
setp.gtu.f32 %p62, %f160, 0f7F800000;
@%p62 bra $L__BB0_56;
bra.uni $L__BB0_54;
$L__BB0_56:
add.f32 %f1244, %f156, 0f3EE66666;
bra.uni $L__BB0_57;
$L__BB0_54:
setp.neu.f32 %p63, %f160, 0f7F800000;
@%p63 bra $L__BB0_57;
selp.f32 %f1244, 0fFF800000, 0f7F800000, %p1;
$L__BB0_57:
setp.eq.f32 %p64, %f156, 0f3F800000;
selp.f32 %f171, 0f3F800000, %f1244, %p64;
abs.f32 %f172, %f157;
setp.lt.f32 %p65, %f172, 0f00800000;
mul.f32 %f622, %f172, 0f4B800000;
selp.f32 %f623, %f622, %f172, %p65;
selp.f32 %f624, 0fC3170000, 0fC2FE0000, %p65;
mov.b32 %r226, %f623;
and.b32 %r227, %r226, 8388607;
or.b32 %r228, %r227, 1065353216;
mov.b32 %f625, %r228;
shr.u32 %r229, %r226, 23;
cvt.rn.f32.u32 %f626, %r229;
add.f32 %f627, %f624, %f626;
setp.gt.f32 %p66, %f625, 0f3FB504F3;
mul.f32 %f628, %f625, 0f3F000000;
add.f32 %f629, %f627, 0f3F800000;
selp.f32 %f630, %f629, %f627, %p66;
selp.f32 %f631, %f628, %f625, %p66;
add.f32 %f632, %f631, 0fBF800000;
add.f32 %f633, %f631, 0f3F800000;
rcp.approx.ftz.f32 %f634, %f633;
add.f32 %f635, %f632, %f632;
mul.f32 %f636, %f635, %f634;
mul.f32 %f637, %f636, %f636;
mov.f32 %f638, 0f3C4CAF63;
mov.f32 %f639, 0f3B18F0FE;
fma.rn.f32 %f640, %f639, %f637, %f638;
mov.f32 %f641, 0f3DAAAABD;
fma.rn.f32 %f642, %f640, %f637, %f641;
mul.rn.f32 %f643, %f642, %f637;
mul.rn.f32 %f644, %f643, %f636;
sub.f32 %f645, %f632, %f636;
add.f32 %f646, %f645, %f645;
neg.f32 %f647, %f636;
fma.rn.f32 %f648, %f647, %f632, %f646;
mul.rn.f32 %f649, %f634, %f648;
add.f32 %f650, %f644, %f636;
sub.f32 %f651, %f636, %f650;
add.f32 %f652, %f644, %f651;
add.f32 %f653, %f649, %f652;
add.f32 %f654, %f650, %f653;
sub.f32 %f655, %f650, %f654;
add.f32 %f656, %f653, %f655;
mov.f32 %f657, 0f3F317200;
mul.rn.f32 %f658, %f630, %f657;
mov.f32 %f659, 0f35BFBE8E;
mul.rn.f32 %f660, %f630, %f659;
add.f32 %f661, %f658, %f654;
sub.f32 %f662, %f658, %f661;
add.f32 %f663, %f654, %f662;
add.f32 %f664, %f656, %f663;
add.f32 %f665, %f660, %f664;
add.f32 %f666, %f661, %f665;
sub.f32 %f667, %f661, %f666;
add.f32 %f668, %f665, %f667;
mov.f32 %f669, 0f3EE66666;
mul.rn.f32 %f670, %f669, %f666;
neg.f32 %f671, %f670;
fma.rn.f32 %f672, %f669, %f666, %f671;
fma.rn.f32 %f673, %f669, %f668, %f672;
mov.f32 %f674, 0f00000000;
fma.rn.f32 %f675, %f674, %f666, %f673;
add.rn.f32 %f676, %f670, %f675;
neg.f32 %f677, %f676;
add.rn.f32 %f678, %f670, %f677;
add.rn.f32 %f679, %f678, %f675;
mov.b32 %r230, %f676;
setp.eq.s32 %p67, %r230, 1118925336;
add.s32 %r231, %r230, -1;
mov.b32 %f680, %r231;
add.f32 %f681, %f679, 0f37000000;
selp.f32 %f173, %f681, %f679, %p67;
selp.f32 %f682, %f680, %f676, %p67;
mov.f32 %f683, 0f3FB8AA3B;
mul.rn.f32 %f684, %f682, %f683;
cvt.rzi.f32.f32 %f685, %f684;
abs.f32 %f686, %f685;
setp.gt.f32 %p68, %f686, 0f42FC0000;
mov.b32 %r232, %f685;
and.b32 %r233, %r232, -2147483648;
or.b32 %r234, %r233, 1123811328;
mov.b32 %f687, %r234;
selp.f32 %f688, %f687, %f685, %p68;
mov.f32 %f689, 0fBF317218;
fma.rn.f32 %f690, %f688, %f689, %f682;
mov.f32 %f691, 0f3102E308;
fma.rn.f32 %f692, %f688, %f691, %f690;
mul.f32 %f693, %f692, 0f3FB8AA3B;
add.f32 %f694, %f688, 0f4B40007F;
mov.b32 %r235, %f694;
shl.b32 %r236, %r235, 23;
mov.b32 %f695, %r236;
ex2.approx.ftz.f32 %f696, %f693;
mul.f32 %f174, %f696, %f695;
setp.eq.f32 %p69, %f174, 0f7F800000;
mov.f32 %f1245, 0f7F800000;
@%p69 bra $L__BB0_59;
fma.rn.f32 %f1245, %f174, %f173, %f174;
$L__BB0_59:
setp.lt.f32 %p70, %f157, 0f00000000;
and.pred %p2, %p70, %p56;
setp.eq.f32 %p72, %f157, 0f00000000;
@%p72 bra $L__BB0_63;
bra.uni $L__BB0_60;
$L__BB0_63:
add.f32 %f701, %f157, %f157;
selp.f32 %f1247, %f701, 0f00000000, %p56;
bra.uni $L__BB0_64;
$L__BB0_60:
mov.b32 %r237, %f1245;
xor.b32 %r238, %r237, -2147483648;
mov.b32 %f697, %r238;
selp.f32 %f1247, %f697, %f1245, %p2;
setp.geu.f32 %p73, %f157, 0f00000000;
@%p73 bra $L__BB0_64;
mov.f32 %f698, 0f3EE66666;
cvt.rzi.f32.f32 %f699, %f698;
setp.eq.f32 %p74, %f699, 0f3EE66666;
@%p74 bra $L__BB0_64;
mov.f32 %f1247, 0f7FFFFFFF;
$L__BB0_64:
add.f32 %f702, %f172, 0f3EE66666;
mov.b32 %r239, %f702;
setp.lt.s32 %p76, %r239, 2139095040;
@%p76 bra $L__BB0_69;
setp.gtu.f32 %p77, %f172, 0f7F800000;
@%p77 bra $L__BB0_68;
bra.uni $L__BB0_66;
$L__BB0_68:
add.f32 %f1247, %f157, 0f3EE66666;
bra.uni $L__BB0_69;
$L__BB0_66:
setp.neu.f32 %p78, %f172, 0f7F800000;
@%p78 bra $L__BB0_69;
selp.f32 %f1247, 0fFF800000, 0f7F800000, %p2;
$L__BB0_69:
setp.eq.f32 %p79, %f157, 0f3F800000;
selp.f32 %f183, 0f3F800000, %f1247, %p79;
abs.f32 %f184, %f158;
setp.lt.f32 %p80, %f184, 0f00800000;
mul.f32 %f704, %f184, 0f4B800000;
selp.f32 %f705, %f704, %f184, %p80;
selp.f32 %f706, 0fC3170000, 0fC2FE0000, %p80;
mov.b32 %r240, %f705;
and.b32 %r241, %r240, 8388607;
or.b32 %r242, %r241, 1065353216;
mov.b32 %f707, %r242;
shr.u32 %r243, %r240, 23;
cvt.rn.f32.u32 %f708, %r243;
add.f32 %f709, %f706, %f708;
setp.gt.f32 %p81, %f707, 0f3FB504F3;
mul.f32 %f710, %f707, 0f3F000000;
add.f32 %f711, %f709, 0f3F800000;
selp.f32 %f712, %f711, %f709, %p81;
selp.f32 %f713, %f710, %f707, %p81;
add.f32 %f714, %f713, 0fBF800000;
add.f32 %f715, %f713, 0f3F800000;
rcp.approx.ftz.f32 %f716, %f715;
add.f32 %f717, %f714, %f714;
mul.f32 %f718, %f717, %f716;
mul.f32 %f719, %f718, %f718;
mov.f32 %f720, 0f3C4CAF63;
mov.f32 %f721, 0f3B18F0FE;
fma.rn.f32 %f722, %f721, %f719, %f720;
mov.f32 %f723, 0f3DAAAABD;
fma.rn.f32 %f724, %f722, %f719, %f723;
mul.rn.f32 %f725, %f724, %f719;
mul.rn.f32 %f726, %f725, %f718;
sub.f32 %f727, %f714, %f718;
add.f32 %f728, %f727, %f727;
neg.f32 %f729, %f718;
fma.rn.f32 %f730, %f729, %f714, %f728;
mul.rn.f32 %f731, %f716, %f730;
add.f32 %f732, %f726, %f718;
sub.f32 %f733, %f718, %f732;
add.f32 %f734, %f726, %f733;
add.f32 %f735, %f731, %f734;
add.f32 %f736, %f732, %f735;
sub.f32 %f737, %f732, %f736;
add.f32 %f738, %f735, %f737;
mov.f32 %f739, 0f3F317200;
mul.rn.f32 %f740, %f712, %f739;
mov.f32 %f741, 0f35BFBE8E;
mul.rn.f32 %f742, %f712, %f741;
add.f32 %f743, %f740, %f736;
sub.f32 %f744, %f740, %f743;
add.f32 %f745, %f736, %f744;
add.f32 %f746, %f738, %f745;
add.f32 %f747, %f742, %f746;
add.f32 %f748, %f743, %f747;
sub.f32 %f749, %f743, %f748;
add.f32 %f750, %f747, %f749;
mov.f32 %f751, 0f3EE66666;
mul.rn.f32 %f752, %f751, %f748;
neg.f32 %f753, %f752;
fma.rn.f32 %f754, %f751, %f748, %f753;
fma.rn.f32 %f755, %f751, %f750, %f754;
mov.f32 %f756, 0f00000000;
fma.rn.f32 %f757, %f756, %f748, %f755;
add.rn.f32 %f758, %f752, %f757;
neg.f32 %f759, %f758;
add.rn.f32 %f760, %f752, %f759;
add.rn.f32 %f761, %f760, %f757;
mov.b32 %r244, %f758;
setp.eq.s32 %p82, %r244, 1118925336;
add.s32 %r245, %r244, -1;
mov.b32 %f762, %r245;
add.f32 %f763, %f761, 0f37000000;
selp.f32 %f185, %f763, %f761, %p82;
selp.f32 %f764, %f762, %f758, %p82;
mov.f32 %f765, 0f3FB8AA3B;
mul.rn.f32 %f766, %f764, %f765;
cvt.rzi.f32.f32 %f767, %f766;
abs.f32 %f768, %f767;
setp.gt.f32 %p83, %f768, 0f42FC0000;
mov.b32 %r246, %f767;
and.b32 %r247, %r246, -2147483648;
or.b32 %r248, %r247, 1123811328;
mov.b32 %f769, %r248;
selp.f32 %f770, %f769, %f767, %p83;
mov.f32 %f771, 0fBF317218;
fma.rn.f32 %f772, %f770, %f771, %f764;
mov.f32 %f773, 0f3102E308;
fma.rn.f32 %f774, %f770, %f773, %f772;
mul.f32 %f775, %f774, 0f3FB8AA3B;
add.f32 %f776, %f770, 0f4B40007F;
mov.b32 %r249, %f776;
shl.b32 %r250, %r249, 23;
mov.b32 %f777, %r250;
ex2.approx.ftz.f32 %f778, %f775;
mul.f32 %f186, %f778, %f777;
setp.eq.f32 %p84, %f186, 0f7F800000;
mov.f32 %f1248, 0f7F800000;
@%p84 bra $L__BB0_71;
fma.rn.f32 %f1248, %f186, %f185, %f186;
$L__BB0_71:
setp.lt.f32 %p85, %f158, 0f00000000;
and.pred %p3, %p85, %p56;
setp.eq.f32 %p87, %f158, 0f00000000;
@%p87 bra $L__BB0_75;
bra.uni $L__BB0_72;
$L__BB0_75:
add.f32 %f783, %f158, %f158;
selp.f32 %f1250, %f783, 0f00000000, %p56;
bra.uni $L__BB0_76;
$L__BB0_72:
mov.b32 %r251, %f1248;
xor.b32 %r252, %r251, -2147483648;
mov.b32 %f779, %r252;
selp.f32 %f1250, %f779, %f1248, %p3;
setp.geu.f32 %p88, %f158, 0f00000000;
@%p88 bra $L__BB0_76;
mov.f32 %f780, 0f3EE66666;
cvt.rzi.f32.f32 %f781, %f780;
setp.eq.f32 %p89, %f781, 0f3EE66666;
@%p89 bra $L__BB0_76;
mov.f32 %f1250, 0f7FFFFFFF;
$L__BB0_76:
add.f32 %f784, %f184, 0f3EE66666;
mov.b32 %r253, %f784;
setp.lt.s32 %p91, %r253, 2139095040;
@%p91 bra $L__BB0_81;
setp.gtu.f32 %p92, %f184, 0f7F800000;
@%p92 bra $L__BB0_80;
bra.uni $L__BB0_78;
$L__BB0_80:
add.f32 %f1250, %f158, 0f3EE66666;
bra.uni $L__BB0_81;
$L__BB0_78:
setp.neu.f32 %p93, %f184, 0f7F800000;
@%p93 bra $L__BB0_81;
selp.f32 %f1250, 0fFF800000, 0f7F800000, %p3;
$L__BB0_81:
setp.eq.f32 %p94, %f158, 0f3F800000;
mov.f32 %f786, 0f3F800000;
selp.f32 %f787, 0f3F800000, %f1250, %p94;
min.f32 %f788, %f171, %f786;
mov.f32 %f789, 0f00000000;
max.f32 %f195, %f789, %f788;
min.f32 %f790, %f183, %f786;
max.f32 %f196, %f789, %f790;
min.f32 %f791, %f787, %f786;
max.f32 %f197, %f789, %f791;
mov.f32 %f795, 0f3ED55555;
abs.f32 %f199, %f195;
setp.lt.f32 %p95, %f199, 0f00800000;
mul.f32 %f797, %f199, 0f4B800000;
selp.f32 %f798, %f797, %f199, %p95;
selp.f32 %f799, 0fC3170000, 0fC2FE0000, %p95;
mov.b32 %r256, %f798;
and.b32 %r257, %r256, 8388607;
or.b32 %r258, %r257, 1065353216;
mov.b32 %f800, %r258;
shr.u32 %r259, %r256, 23;
cvt.rn.f32.u32 %f801, %r259;
add.f32 %f802, %f799, %f801;
setp.gt.f32 %p96, %f800, 0f3FB504F3;
mul.f32 %f803, %f800, 0f3F000000;
add.f32 %f804, %f802, 0f3F800000;
selp.f32 %f805, %f804, %f802, %p96;
selp.f32 %f806, %f803, %f800, %p96;
add.f32 %f807, %f806, 0fBF800000;
add.f32 %f808, %f806, 0f3F800000;
rcp.approx.ftz.f32 %f809, %f808;
add.f32 %f810, %f807, %f807;
mul.f32 %f811, %f810, %f809;
mul.f32 %f812, %f811, %f811;
mov.f32 %f813, 0f3C4CAF63;
mov.f32 %f814, 0f3B18F0FE;
fma.rn.f32 %f815, %f814, %f812, %f813;
mov.f32 %f816, 0f3DAAAABD;
fma.rn.f32 %f817, %f815, %f812, %f816;
mul.rn.f32 %f818, %f817, %f812;
mul.rn.f32 %f819, %f818, %f811;
sub.f32 %f820, %f807, %f811;
add.f32 %f821, %f820, %f820;
neg.f32 %f822, %f811;
fma.rn.f32 %f823, %f822, %f807, %f821;
mul.rn.f32 %f824, %f809, %f823;
add.f32 %f825, %f819, %f811;
sub.f32 %f826, %f811, %f825;
add.f32 %f827, %f819, %f826;
add.f32 %f828, %f824, %f827;
add.f32 %f829, %f825, %f828;
sub.f32 %f830, %f825, %f829;
add.f32 %f831, %f828, %f830;
mov.f32 %f832, 0f3F317200;
mul.rn.f32 %f833, %f805, %f832;
mov.f32 %f834, 0f35BFBE8E;
mul.rn.f32 %f835, %f805, %f834;
add.f32 %f836, %f833, %f829;
sub.f32 %f837, %f833, %f836;
add.f32 %f838, %f829, %f837;
add.f32 %f839, %f831, %f838;
add.f32 %f840, %f835, %f839;
add.f32 %f841, %f836, %f840;
sub.f32 %f842, %f836, %f841;
add.f32 %f843, %f840, %f842;
mul.rn.f32 %f844, %f795, %f841;
neg.f32 %f845, %f844;
fma.rn.f32 %f846, %f795, %f841, %f845;
fma.rn.f32 %f847, %f795, %f843, %f846;
fma.rn.f32 %f848, %f789, %f841, %f847;
add.rn.f32 %f849, %f844, %f848;
neg.f32 %f850, %f849;
add.rn.f32 %f851, %f844, %f850;
add.rn.f32 %f852, %f851, %f848;
mov.b32 %r260, %f849;
setp.eq.s32 %p97, %r260, 1118925336;
add.s32 %r261, %r260, -1;
mov.b32 %f853, %r261;
add.f32 %f854, %f852, 0f37000000;
selp.f32 %f200, %f854, %f852, %p97;
selp.f32 %f855, %f853, %f849, %p97;
mov.f32 %f856, 0f3FB8AA3B;
mul.rn.f32 %f857, %f855, %f856;
cvt.rzi.f32.f32 %f858, %f857;
abs.f32 %f859, %f858;
setp.gt.f32 %p98, %f859, 0f42FC0000;
mov.b32 %r262, %f858;
and.b32 %r263, %r262, -2147483648;
or.b32 %r264, %r263, 1123811328;
mov.b32 %f860, %r264;
selp.f32 %f861, %f860, %f858, %p98;
mov.f32 %f862, 0fBF317218;
fma.rn.f32 %f863, %f861, %f862, %f855;
mov.f32 %f864, 0f3102E308;
fma.rn.f32 %f865, %f861, %f864, %f863;
mul.f32 %f866, %f865, 0f3FB8AA3B;
add.f32 %f867, %f861, 0f4B40007F;
mov.b32 %r265, %f867;
shl.b32 %r266, %r265, 23;
mov.b32 %f868, %r266;
ex2.approx.ftz.f32 %f869, %f866;
mul.f32 %f201, %f869, %f868;
setp.eq.f32 %p99, %f201, 0f7F800000;
mov.f32 %f1251, 0f7F800000;
@%p99 bra $L__BB0_83;
fma.rn.f32 %f1251, %f201, %f200, %f201;
$L__BB0_83:
mov.f32 %f1175, 0f3E555555;
cvt.rzi.f32.f32 %f1174, %f1175;
add.f32 %f1173, %f1174, %f1174;
mov.f32 %f1172, 0f3ED55555;
sub.f32 %f1171, %f1172, %f1173;
abs.f32 %f1170, %f1171;
setp.lt.f32 %p100, %f195, 0f00000000;
setp.eq.f32 %p101, %f1170, 0f3F800000;
and.pred %p4, %p100, %p101;
setp.eq.f32 %p102, %f195, 0f00000000;
@%p102 bra $L__BB0_87;
bra.uni $L__BB0_84;
$L__BB0_87:
add.f32 %f874, %f195, %f195;
selp.f32 %f1253, %f874, 0f00000000, %p101;
bra.uni $L__BB0_88;
$L__BB0_84:
mov.b32 %r267, %f1251;
xor.b32 %r268, %r267, -2147483648;
mov.b32 %f870, %r268;
selp.f32 %f1253, %f870, %f1251, %p4;
setp.geu.f32 %p103, %f195, 0f00000000;
@%p103 bra $L__BB0_88;
mov.f32 %f871, 0f3ED55555;
cvt.rzi.f32.f32 %f872, %f871;
setp.eq.f32 %p104, %f872, 0f3ED55555;
@%p104 bra $L__BB0_88;
mov.f32 %f1253, 0f7FFFFFFF;
$L__BB0_88:
add.f32 %f875, %f199, 0f3ED55555;
mov.b32 %r269, %f875;
setp.lt.s32 %p106, %r269, 2139095040;
@%p106 bra $L__BB0_93;
setp.gtu.f32 %p107, %f199, 0f7F800000;
@%p107 bra $L__BB0_92;
bra.uni $L__BB0_90;
$L__BB0_92:
add.f32 %f1253, %f195, 0f3ED55555;
bra.uni $L__BB0_93;
$L__BB0_90:
setp.neu.f32 %p108, %f199, 0f7F800000;
@%p108 bra $L__BB0_93;
selp.f32 %f1253, 0fFF800000, 0f7F800000, %p4;
$L__BB0_93:
abs.f32 %f210, %f196;
setp.lt.f32 %p109, %f210, 0f00800000;
mul.f32 %f877, %f210, 0f4B800000;
selp.f32 %f878, %f877, %f210, %p109;
selp.f32 %f879, 0fC3170000, 0fC2FE0000, %p109;
mov.b32 %r270, %f878;
and.b32 %r271, %r270, 8388607;
or.b32 %r272, %r271, 1065353216;
mov.b32 %f880, %r272;
shr.u32 %r273, %r270, 23;
cvt.rn.f32.u32 %f881, %r273;
add.f32 %f882, %f879, %f881;
setp.gt.f32 %p110, %f880, 0f3FB504F3;
mul.f32 %f883, %f880, 0f3F000000;
add.f32 %f884, %f882, 0f3F800000;
selp.f32 %f885, %f884, %f882, %p110;
selp.f32 %f886, %f883, %f880, %p110;
add.f32 %f887, %f886, 0fBF800000;
add.f32 %f888, %f886, 0f3F800000;
rcp.approx.ftz.f32 %f889, %f888;
add.f32 %f890, %f887, %f887;
mul.f32 %f891, %f890, %f889;
mul.f32 %f892, %f891, %f891;
mov.f32 %f893, 0f3C4CAF63;
mov.f32 %f894, 0f3B18F0FE;
fma.rn.f32 %f895, %f894, %f892, %f893;
mov.f32 %f896, 0f3DAAAABD;
fma.rn.f32 %f897, %f895, %f892, %f896;
mul.rn.f32 %f898, %f897, %f892;
mul.rn.f32 %f899, %f898, %f891;
sub.f32 %f900, %f887, %f891;
add.f32 %f901, %f900, %f900;
neg.f32 %f902, %f891;
fma.rn.f32 %f903, %f902, %f887, %f901;
mul.rn.f32 %f904, %f889, %f903;
add.f32 %f905, %f899, %f891;
sub.f32 %f906, %f891, %f905;
add.f32 %f907, %f899, %f906;
add.f32 %f908, %f904, %f907;
add.f32 %f909, %f905, %f908;
sub.f32 %f910, %f905, %f909;
add.f32 %f911, %f908, %f910;
mov.f32 %f912, 0f3F317200;
mul.rn.f32 %f913, %f885, %f912;
mov.f32 %f914, 0f35BFBE8E;
mul.rn.f32 %f915, %f885, %f914;
add.f32 %f916, %f913, %f909;
sub.f32 %f917, %f913, %f916;
add.f32 %f918, %f909, %f917;
add.f32 %f919, %f911, %f918;
add.f32 %f920, %f915, %f919;
add.f32 %f921, %f916, %f920;
sub.f32 %f922, %f916, %f921;
add.f32 %f923, %f920, %f922;
mov.f32 %f924, 0f3ED55555;
mul.rn.f32 %f925, %f924, %f921;
neg.f32 %f926, %f925;
fma.rn.f32 %f927, %f924, %f921, %f926;
fma.rn.f32 %f928, %f924, %f923, %f927;
mov.f32 %f929, 0f00000000;
fma.rn.f32 %f930, %f929, %f921, %f928;
add.rn.f32 %f931, %f925, %f930;
neg.f32 %f932, %f931;
add.rn.f32 %f933, %f925, %f932;
add.rn.f32 %f934, %f933, %f930;
mov.b32 %r274, %f931;
setp.eq.s32 %p111, %r274, 1118925336;
add.s32 %r275, %r274, -1;
mov.b32 %f935, %r275;
add.f32 %f936, %f934, 0f37000000;
selp.f32 %f211, %f936, %f934, %p111;
selp.f32 %f937, %f935, %f931, %p111;
mov.f32 %f938, 0f3FB8AA3B;
mul.rn.f32 %f939, %f937, %f938;
cvt.rzi.f32.f32 %f940, %f939;
abs.f32 %f941, %f940;
setp.gt.f32 %p112, %f941, 0f42FC0000;
mov.b32 %r276, %f940;
and.b32 %r277, %r276, -2147483648;
or.b32 %r278, %r277, 1123811328;
mov.b32 %f942, %r278;
selp.f32 %f943, %f942, %f940, %p112;
mov.f32 %f944, 0fBF317218;
fma.rn.f32 %f945, %f943, %f944, %f937;
mov.f32 %f946, 0f3102E308;
fma.rn.f32 %f947, %f943, %f946, %f945;
mul.f32 %f948, %f947, 0f3FB8AA3B;
add.f32 %f949, %f943, 0f4B40007F;
mov.b32 %r279, %f949;
shl.b32 %r280, %r279, 23;
mov.b32 %f950, %r280;
ex2.approx.ftz.f32 %f951, %f948;
mul.f32 %f212, %f951, %f950;
setp.eq.f32 %p113, %f212, 0f7F800000;
mov.f32 %f1254, 0f7F800000;
@%p113 bra $L__BB0_95;
fma.rn.f32 %f1254, %f212, %f211, %f212;
$L__BB0_95:
setp.lt.f32 %p114, %f196, 0f00000000;
and.pred %p5, %p114, %p101;
setp.eq.f32 %p116, %f196, 0f00000000;
@%p116 bra $L__BB0_99;
bra.uni $L__BB0_96;
$L__BB0_99:
add.f32 %f956, %f196, %f196;
selp.f32 %f1256, %f956, 0f00000000, %p101;
bra.uni $L__BB0_100;
$L__BB0_96:
mov.b32 %r281, %f1254;
xor.b32 %r282, %r281, -2147483648;
mov.b32 %f952, %r282;
selp.f32 %f1256, %f952, %f1254, %p5;
setp.geu.f32 %p117, %f196, 0f00000000;
@%p117 bra $L__BB0_100;
mov.f32 %f953, 0f3ED55555;
cvt.rzi.f32.f32 %f954, %f953;
setp.eq.f32 %p118, %f954, 0f3ED55555;
@%p118 bra $L__BB0_100;
mov.f32 %f1256, 0f7FFFFFFF;
$L__BB0_100:
add.f32 %f957, %f210, 0f3ED55555;
mov.b32 %r283, %f957;
setp.lt.s32 %p120, %r283, 2139095040;
@%p120 bra $L__BB0_105;
setp.gtu.f32 %p121, %f210, 0f7F800000;
@%p121 bra $L__BB0_104;
bra.uni $L__BB0_102;
$L__BB0_104:
add.f32 %f1256, %f196, 0f3ED55555;
bra.uni $L__BB0_105;
$L__BB0_102:
setp.neu.f32 %p122, %f210, 0f7F800000;
@%p122 bra $L__BB0_105;
selp.f32 %f1256, 0fFF800000, 0f7F800000, %p5;
$L__BB0_105:
abs.f32 %f221, %f197;
setp.lt.f32 %p123, %f221, 0f00800000;
mul.f32 %f959, %f221, 0f4B800000;
selp.f32 %f960, %f959, %f221, %p123;
selp.f32 %f961, 0fC3170000, 0fC2FE0000, %p123;
mov.b32 %r284, %f960;
and.b32 %r285, %r284, 8388607;
or.b32 %r286, %r285, 1065353216;
mov.b32 %f962, %r286;
shr.u32 %r287, %r284, 23;
cvt.rn.f32.u32 %f963, %r287;
add.f32 %f964, %f961, %f963;
setp.gt.f32 %p124, %f962, 0f3FB504F3;
mul.f32 %f965, %f962, 0f3F000000;
add.f32 %f966, %f964, 0f3F800000;
selp.f32 %f967, %f966, %f964, %p124;
selp.f32 %f968, %f965, %f962, %p124;
add.f32 %f969, %f968, 0fBF800000;
add.f32 %f970, %f968, 0f3F800000;
rcp.approx.ftz.f32 %f971, %f970;
add.f32 %f972, %f969, %f969;
mul.f32 %f973, %f972, %f971;
mul.f32 %f974, %f973, %f973;
mov.f32 %f975, 0f3C4CAF63;
mov.f32 %f976, 0f3B18F0FE;
fma.rn.f32 %f977, %f976, %f974, %f975;
mov.f32 %f978, 0f3DAAAABD;
fma.rn.f32 %f979, %f977, %f974, %f978;
mul.rn.f32 %f980, %f979, %f974;
mul.rn.f32 %f981, %f980, %f973;
sub.f32 %f982, %f969, %f973;
add.f32 %f983, %f982, %f982;
neg.f32 %f984, %f973;
fma.rn.f32 %f985, %f984, %f969, %f983;
mul.rn.f32 %f986, %f971, %f985;
add.f32 %f987, %f981, %f973;
sub.f32 %f988, %f973, %f987;
add.f32 %f989, %f981, %f988;
add.f32 %f990, %f986, %f989;
add.f32 %f991, %f987, %f990;
sub.f32 %f992, %f987, %f991;
add.f32 %f993, %f990, %f992;
mov.f32 %f994, 0f3F317200;
mul.rn.f32 %f995, %f967, %f994;
mov.f32 %f996, 0f35BFBE8E;
mul.rn.f32 %f997, %f967, %f996;
add.f32 %f998, %f995, %f991;
sub.f32 %f999, %f995, %f998;
add.f32 %f1000, %f991, %f999;
add.f32 %f1001, %f993, %f1000;
add.f32 %f1002, %f997, %f1001;
add.f32 %f1003, %f998, %f1002;
sub.f32 %f1004, %f998, %f1003;
add.f32 %f1005, %f1002, %f1004;
mov.f32 %f1006, 0f3ED55555;
mul.rn.f32 %f1007, %f1006, %f1003;
neg.f32 %f1008, %f1007;
fma.rn.f32 %f1009, %f1006, %f1003, %f1008;
fma.rn.f32 %f1010, %f1006, %f1005, %f1009;
mov.f32 %f1011, 0f00000000;
fma.rn.f32 %f1012, %f1011, %f1003, %f1010;
add.rn.f32 %f1013, %f1007, %f1012;
neg.f32 %f1014, %f1013;
add.rn.f32 %f1015, %f1007, %f1014;
add.rn.f32 %f1016, %f1015, %f1012;
mov.b32 %r288, %f1013;
setp.eq.s32 %p125, %r288, 1118925336;
add.s32 %r289, %r288, -1;
mov.b32 %f1017, %r289;
add.f32 %f1018, %f1016, 0f37000000;
selp.f32 %f222, %f1018, %f1016, %p125;
selp.f32 %f1019, %f1017, %f1013, %p125;
mov.f32 %f1020, 0f3FB8AA3B;
mul.rn.f32 %f1021, %f1019, %f1020;
cvt.rzi.f32.f32 %f1022, %f1021;
abs.f32 %f1023, %f1022;
setp.gt.f32 %p126, %f1023, 0f42FC0000;
mov.b32 %r290, %f1022;
and.b32 %r291, %r290, -2147483648;
or.b32 %r292, %r291, 1123811328;
mov.b32 %f1024, %r292;
selp.f32 %f1025, %f1024, %f1022, %p126;
mov.f32 %f1026, 0fBF317218;
fma.rn.f32 %f1027, %f1025, %f1026, %f1019;
mov.f32 %f1028, 0f3102E308;
fma.rn.f32 %f1029, %f1025, %f1028, %f1027;
mul.f32 %f1030, %f1029, 0f3FB8AA3B;
add.f32 %f1031, %f1025, 0f4B40007F;
mov.b32 %r293, %f1031;
shl.b32 %r294, %r293, 23;
mov.b32 %f1032, %r294;
ex2.approx.ftz.f32 %f1033, %f1030;
mul.f32 %f223, %f1033, %f1032;
setp.eq.f32 %p127, %f223, 0f7F800000;
mov.f32 %f1257, 0f7F800000;
@%p127 bra $L__BB0_107;
fma.rn.f32 %f1257, %f223, %f222, %f223;
$L__BB0_107:
setp.lt.f32 %p128, %f197, 0f00000000;
and.pred %p6, %p128, %p101;
setp.eq.f32 %p130, %f197, 0f00000000;
@%p130 bra $L__BB0_111;
bra.uni $L__BB0_108;
$L__BB0_111:
add.f32 %f1038, %f197, %f197;
selp.f32 %f1259, %f1038, 0f00000000, %p101;
bra.uni $L__BB0_112;
$L__BB0_108:
mov.b32 %r295, %f1257;
xor.b32 %r296, %r295, -2147483648;
mov.b32 %f1034, %r296;
selp.f32 %f1259, %f1034, %f1257, %p6;
setp.geu.f32 %p131, %f197, 0f00000000;
@%p131 bra $L__BB0_112;
mov.f32 %f1035, 0f3ED55555;
cvt.rzi.f32.f32 %f1036, %f1035;
setp.eq.f32 %p132, %f1036, 0f3ED55555;
@%p132 bra $L__BB0_112;
mov.f32 %f1259, 0f7FFFFFFF;
$L__BB0_112:
add.f32 %f1039, %f221, 0f3ED55555;
mov.b32 %r297, %f1039;
setp.lt.s32 %p134, %r297, 2139095040;
@%p134 bra $L__BB0_117;
setp.gtu.f32 %p135, %f221, 0f7F800000;
@%p135 bra $L__BB0_116;
bra.uni $L__BB0_114;
$L__BB0_116:
add.f32 %f1259, %f197, 0f3ED55555;
bra.uni $L__BB0_117;
$L__BB0_114:
setp.neu.f32 %p136, %f221, 0f7F800000;
@%p136 bra $L__BB0_117;
selp.f32 %f1259, 0fFF800000, 0f7F800000, %p6;
$L__BB0_117:
ld.const.u64 %rd100, [params+144];
cvta.to.global.u64 %rd99, %rd100;
ld.const.u32 %r344, [params+136];
mad.lo.s32 %r343, %r344, %r5, %r4;
cvt.u64.u32 %rd98, %r343;
fma.rn.f32 %f1040, %f1253, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p137, %f195, 0f3F800000;
mov.f32 %f1041, 0f3F800000;
selp.f32 %f1042, 0f3F7FFFFF, %f1040, %p137;
mul.f32 %f1043, %f195, 0f414EB852;
setp.lt.f32 %p138, %f195, 0f3B4D2E1C;
selp.f32 %f1044, %f1043, %f1042, %p138;
fma.rn.f32 %f1045, %f1256, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p139, %f196, 0f3F800000;
selp.f32 %f1046, 0f3F7FFFFF, %f1045, %p139;
mul.f32 %f1047, %f196, 0f414EB852;
setp.lt.f32 %p140, %f196, 0f3B4D2E1C;
selp.f32 %f1048, %f1047, %f1046, %p140;
fma.rn.f32 %f1049, %f1259, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p141, %f197, 0f3F800000;
selp.f32 %f1050, 0f3F7FFFFF, %f1049, %p141;
mul.f32 %f1051, %f197, 0f414EB852;
setp.lt.f32 %p142, %f197, 0f3B4D2E1C;
selp.f32 %f1052, %f1051, %f1050, %p142;
min.f32 %f1053, %f1044, %f1041;
mov.f32 %f1054, 0f00000000;
max.f32 %f1055, %f1054, %f1053;
mul.f32 %f1056, %f1055, 0f43800000;
cvt.rzi.u32.f32 %r298, %f1056;
min.u32 %r299, %r298, 255;
min.f32 %f1057, %f1048, %f1041;
max.f32 %f1058, %f1054, %f1057;
mul.f32 %f1059, %f1058, 0f43800000;
cvt.rzi.u32.f32 %r300, %f1059;
min.u32 %r301, %r300, 255;
min.f32 %f1060, %f1052, %f1041;
max.f32 %f1061, %f1054, %f1060;
mul.f32 %f1062, %f1061, 0f43800000;
cvt.rzi.u32.f32 %r302, %f1062;
min.u32 %r303, %r302, 255;
shl.b64 %rd68, %rd98, 2;
add.s64 %rd69, %rd99, %rd68;
cvt.u16.u32 %rs44, %r303;
cvt.u16.u32 %rs45, %r301;
cvt.u16.u32 %rs46, %r299;
mov.u16 %rs47, 255;
st.global.v4.u8 [%rd69], {%rs46, %rs45, %rs44, %rs47};
$L__BB0_118:
ld.const.u32 %r345, [params+104];
and.b32 %r304, %r345, 4;
setp.eq.s32 %p143, %r304, 0;
ld.const.u32 %r357, [params+108];
@%p143 bra $L__BB0_122;
setp.eq.s32 %p144, %r357, 0;
ld.const.u64 %rd70, [params+224];
cvta.to.global.u64 %rd71, %rd70;
ld.const.u32 %r305, [params+216];
mad.lo.s32 %r306, %r305, %r5, %r4;
mul.wide.u32 %rd72, %r306, 8;
add.s64 %rd19, %rd71, %rd72;
@%p144 bra $L__BB0_121;
ld.global.v4.u16 {%rs55, %rs56, %rs57, %rs58}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f1063, %rs55;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1064, %rs56;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1065, %rs57;}
// end inline asm
add.f32 %f1066, %f156, %f1063;
add.f32 %f1067, %f157, %f1064;
add.f32 %f1068, %f158, %f1065;
mov.f32 %f1069, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs53, %f1068;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs52, %f1067;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f1066;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs54, %f1069;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs51, %rs52, %rs53, %rs54};
bra.uni $L__BB0_122;
$L__BB0_121:
mov.f32 %f1073, 0f3F800000;
mov.u32 %r357, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs62, %f1073;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs61, %f158;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs60, %f157;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs59, %f156;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs59, %rs60, %rs61, %rs62};
$L__BB0_122:
mul.f32 %f1074, %f147, %f531;
mul.f32 %f1075, %f144, %f531;
add.f32 %f1076, %f1075, %f1074;
mul.f32 %f1077, %f148, %f532;
mul.f32 %f1078, %f145, %f532;
add.f32 %f1079, %f1078, %f1077;
mul.f32 %f1080, %f149, %f533;
mul.f32 %f1081, %f146, %f533;
add.f32 %f1082, %f1081, %f1080;
mul.f32 %f1083, %f150, %f531;
add.f32 %f1084, %f1076, %f1083;
mul.f32 %f1085, %f151, %f532;
add.f32 %f1086, %f1079, %f1085;
mul.f32 %f1087, %f152, %f533;
add.f32 %f1088, %f1082, %f1087;
mul.f32 %f1089, %f1084, 0f3F13CD3A;
mul.f32 %f1090, %f1086, 0f3F13CD3A;
mul.f32 %f1091, %f1088, 0f3F13CD3A;
div.rn.f32 %f1092, %f156, %f1089;
div.rn.f32 %f1093, %f157, %f1090;
div.rn.f32 %f1094, %f158, %f1091;
setp.eq.f32 %p145, %f156, 0f00000000;
selp.f32 %f1095, 0f00000000, %f1092, %p145;
setp.eq.f32 %p146, %f157, 0f00000000;
selp.f32 %f1096, 0f00000000, %f1093, %p146;
setp.eq.f32 %p147, %f158, 0f00000000;
selp.f32 %f1097, 0f00000000, %f1094, %p147;
mul.f32 %f232, %f1075, %f1095;
mul.f32 %f233, %f1078, %f1096;
mul.f32 %f234, %f1081, %f1097;
mul.f32 %f235, %f1074, %f1095;
mul.f32 %f236, %f1077, %f1096;
mul.f32 %f237, %f1080, %f1097;
mul.f32 %f238, %f1083, %f1095;
mul.f32 %f239, %f1085, %f1096;
mul.f32 %f240, %f1087, %f1097;
ld.const.u64 %rd73, [params+256];
cvta.to.global.u64 %rd74, %rd73;
ld.const.u32 %r308, [params+248];
mad.lo.s32 %r309, %r308, %r5, %r4;
mul.wide.u32 %rd75, %r309, 8;
add.s64 %rd20, %rd74, %rd75;
setp.eq.s32 %p148, %r357, 0;
@%p148 bra $L__BB0_124;
ld.global.v4.u16 {%rs70, %rs71, %rs72, %rs73}, [%rd20];
// begin inline asm
{ cvt.f32.f16 %f1098, %rs70;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1099, %rs71;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1100, %rs72;}
// end inline asm
add.f32 %f1101, %f232, %f1098;
add.f32 %f1102, %f233, %f1099;
add.f32 %f1103, %f234, %f1100;
mov.f32 %f1104, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs68, %f1103;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs67, %f1102;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs66, %f1101;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs69, %f1104;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs66, %rs67, %rs68, %rs69};
bra.uni $L__BB0_125;
$L__BB0_124:
mov.f32 %f1108, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs77, %f1108;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs76, %f234;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs75, %f233;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs74, %f232;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs74, %rs75, %rs76, %rs77};
$L__BB0_125:
ld.const.u64 %rd76, [params+272];
cvta.to.global.u64 %rd77, %rd76;
ld.const.u32 %r310, [params+264];
mad.lo.s32 %r311, %r310, %r5, %r4;
mul.wide.u32 %rd78, %r311, 8;
add.s64 %rd21, %rd77, %rd78;
@%p148 bra $L__BB0_127;
ld.global.v4.u16 {%rs85, %rs86, %rs87, %rs88}, [%rd21];
// begin inline asm
{ cvt.f32.f16 %f1109, %rs85;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1110, %rs86;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1111, %rs87;}
// end inline asm
add.f32 %f1112, %f235, %f1109;
add.f32 %f1113, %f236, %f1110;
add.f32 %f1114, %f237, %f1111;
mov.f32 %f1115, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs83, %f1114;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs82, %f1113;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs81, %f1112;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs84, %f1115;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs81, %rs82, %rs83, %rs84};
bra.uni $L__BB0_128;
$L__BB0_127:
mov.f32 %f1119, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs92, %f1119;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs91, %f237;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs90, %f236;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs89, %f235;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs89, %rs90, %rs91, %rs92};
$L__BB0_128:
ld.const.u64 %rd79, [params+288];
cvta.to.global.u64 %rd80, %rd79;
ld.const.u32 %r312, [params+280];
mad.lo.s32 %r313, %r312, %r5, %r4;
mul.wide.u32 %rd81, %r313, 8;
add.s64 %rd22, %rd80, %rd81;
@%p148 bra $L__BB0_130;
ld.global.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd22];
// begin inline asm
{ cvt.f32.f16 %f1120, %rs100;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1121, %rs101;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1122, %rs102;}
// end inline asm
add.f32 %f1123, %f238, %f1120;
add.f32 %f1124, %f239, %f1121;
add.f32 %f1125, %f240, %f1122;
mov.f32 %f1126, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs98, %f1125;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs97, %f1124;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs96, %f1123;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs99, %f1126;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs96, %rs97, %rs98, %rs99};
bra.uni $L__BB0_146;
$L__BB0_130:
mov.f32 %f1130, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs107, %f1130;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs106, %f240;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs105, %f239;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs104, %f238;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs104, %rs105, %rs106, %rs107};
$L__BB0_146:
ret;
}