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

2400 lines
64 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<171>;
.reg .b16 %rs<178>;
.reg .f32 %f<1159>;
.reg .b32 %r<369>;
.reg .f64 %fd<5>;
.reg .b64 %rd<121>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r54), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r55), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u64 %rd31, [params+400];
cvta.to.global.u64 %rd32, %rd31;
ld.const.u32 %r60, [params+392];
mad.lo.s32 %r61, %r60, %r55, %r54;
mul.wide.u32 %rd33, %r61, 4;
add.s64 %rd2, %rd32, %rd33;
ld.global.v2.u8 {%rs7, %rs177}, [%rd2];
or.b16 %rs9, %rs7, %rs177;
and.b16 %rs10, %rs9, 255;
setp.eq.s16 %p8, %rs10, 0;
@%p8 bra $L__BB0_2;
ld.global.u8 %rs176, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs176, [%rd2+2];
setp.eq.s16 %p9, %rs176, 0;
mov.f32 %f1117, 0f00000000;
mov.u16 %rs177, 0;
mov.f32 %f1118, %f1117;
mov.f32 %f1119, %f1117;
@%p9 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f170, %rs7;
div.rn.f32 %f171, %f170, 0f437F0000;
fma.rn.f32 %f172, %f171, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs177, 255;
cvt.rn.f32.u16 %f173, %rs13;
div.rn.f32 %f174, %f173, 0f437F0000;
fma.rn.f32 %f175, %f174, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f176, %rs176;
div.rn.f32 %f177, %f176, 0f437F0000;
fma.rn.f32 %f178, %f177, 0f40000000, 0fBF800000;
mul.f32 %f179, %f175, %f175;
fma.rn.f32 %f180, %f172, %f172, %f179;
fma.rn.f32 %f181, %f178, %f178, %f180;
sqrt.rn.f32 %f182, %f181;
rcp.rn.f32 %f183, %f182;
mul.f32 %f1119, %f183, %f178;
mul.f32 %f1118, %f183, %f175;
mul.f32 %f1117, %f172, %f183;
$L__BB0_4:
ld.const.v2.u32 {%r62, %r63}, [params];
add.s32 %r3, %r62, %r54;
add.s32 %r4, %r63, %r55;
setp.eq.f32 %p10, %f1117, 0f00000000;
setp.eq.f32 %p11, %f1118, 0f00000000;
and.pred %p12, %p10, %p11;
setp.eq.f32 %p13, %f1119, 0f00000000;
and.pred %p14, %p13, %p12;
@%p14 bra $L__BB0_140;
bra.uni $L__BB0_5;
$L__BB0_140:
ld.const.u32 %r51, [params+104];
and.b32 %r331, %r51, 1;
setp.eq.b32 %p160, %r331, 1;
mov.pred %p161, 0;
xor.pred %p162, %p160, %p161;
not.pred %p163, %p162;
@%p163 bra $L__BB0_142;
ld.const.u64 %rd87, [params+144];
cvta.to.global.u64 %rd88, %rd87;
ld.const.u32 %r332, [params+136];
mad.lo.s32 %r333, %r332, %r4, %r3;
mul.wide.u32 %rd89, %r333, 4;
add.s64 %rd90, %rd88, %rd89;
mov.u16 %rs98, 0;
st.global.v4.u8 [%rd90], {%rs98, %rs98, %rs98, %rs98};
$L__BB0_142:
and.b32 %r334, %r51, 8;
setp.eq.s32 %p164, %r334, 0;
@%p164 bra $L__BB0_144;
ld.const.u64 %rd91, [params+192];
cvta.to.global.u64 %rd92, %rd91;
ld.const.u32 %r335, [params+184];
mad.lo.s32 %r336, %r335, %r4, %r3;
mov.f32 %f1022, 0f00000000;
cvt.rzi.u32.f32 %r337, %f1022;
mul.wide.u32 %rd93, %r336, 2;
add.s64 %rd94, %rd92, %rd93;
mov.u16 %rs99, 0;
cvt.u16.u32 %rs100, %r337;
st.global.v2.u8 [%rd94], {%rs100, %rs99};
$L__BB0_144:
and.b32 %r338, %r51, 4;
setp.eq.s32 %p165, %r338, 0;
ld.const.u32 %r368, [params+108];
@%p165 bra $L__BB0_148;
setp.eq.s32 %p166, %r368, 0;
ld.const.u64 %rd95, [params+224];
cvta.to.global.u64 %rd96, %rd95;
ld.const.u32 %r339, [params+216];
mad.lo.s32 %r340, %r339, %r4, %r3;
mul.wide.u32 %rd97, %r340, 8;
add.s64 %rd25, %rd96, %rd97;
@%p166 bra $L__BB0_147;
ld.global.v4.u16 {%rs107, %rs108, %rs109, %rs110}, [%rd25];
// begin inline asm
{ cvt.f32.f16 %f1023, %rs107;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1024, %rs108;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1025, %rs109;}
// end inline asm
add.f32 %f1026, %f1023, 0f00000000;
add.f32 %f1027, %f1024, 0f00000000;
add.f32 %f1028, %f1025, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs106, %f1028;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs105, %f1027;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs104, %f1026;}
// end inline asm
mov.u16 %rs111, 0;
st.global.v4.u16 [%rd25], {%rs104, %rs105, %rs106, %rs111};
bra.uni $L__BB0_148;
$L__BB0_5:
ld.const.v4.f32 {%f187, %f188, %f189, %f190}, [params+512];
neg.f32 %f191, %f188;
neg.f32 %f192, %f189;
mul.f32 %f193, %f1117, %f187;
mul.f32 %f194, %f1118, %f191;
sub.f32 %f195, %f194, %f193;
mul.f32 %f196, %f1119, %f189;
sub.f32 %f13, %f195, %f196;
ld.const.u64 %rd34, [params+432];
cvta.to.global.u64 %rd35, %rd34;
ld.const.u32 %r66, [params+424];
mad.lo.s32 %r67, %r66, %r55, %r54;
mul.wide.u32 %rd36, %r67, 12;
add.s64 %rd37, %rd35, %rd36;
ld.global.f32 %f197, [%rd37];
mul.f32 %f198, %f197, 0f3456BF95;
ld.global.f32 %f199, [%rd37+4];
mul.f32 %f200, %f199, 0f3456BF95;
ld.global.f32 %f201, [%rd37+8];
mul.f32 %f202, %f201, 0f3456BF95;
abs.f32 %f203, %f1117;
div.rn.f32 %f204, %f198, %f203;
abs.f32 %f205, %f1118;
div.rn.f32 %f206, %f200, %f205;
abs.f32 %f207, %f1119;
div.rn.f32 %f208, %f202, %f207;
abs.f32 %f209, %f204;
abs.f32 %f210, %f206;
abs.f32 %f211, %f208;
mov.f32 %f212, 0f38D1B717;
max.f32 %f213, %f209, %f212;
max.f32 %f214, %f210, %f212;
max.f32 %f215, %f211, %f212;
fma.rn.f32 %f14, %f1117, %f213, %f197;
fma.rn.f32 %f15, %f1118, %f214, %f199;
fma.rn.f32 %f16, %f1119, %f215, %f201;
abs.f32 %f216, %f187;
abs.f32 %f217, %f189;
setp.gt.f32 %p15, %f216, %f217;
selp.f32 %f218, %f191, 0f00000000, %p15;
mov.f32 %f1125, 0f00000000;
selp.f32 %f219, %f187, %f192, %p15;
selp.f32 %f220, 0f00000000, %f188, %p15;
mul.f32 %f221, %f219, %f219;
fma.rn.f32 %f222, %f218, %f218, %f221;
fma.rn.f32 %f223, %f220, %f220, %f222;
sqrt.rn.f32 %f224, %f223;
rcp.rn.f32 %f225, %f224;
mul.f32 %f17, %f218, %f225;
mul.f32 %f18, %f219, %f225;
mul.f32 %f19, %f220, %f225;
ld.const.u64 %rd38, [params+128];
cvta.to.global.u64 %rd39, %rd38;
ld.const.u32 %r68, [params+120];
mad.lo.s32 %r69, %r68, %r55, %r54;
mul.wide.u32 %rd40, %r69, 4;
add.s64 %rd3, %rd39, %rd40;
ld.const.u32 %r5, [params+540];
setp.lt.s32 %p16, %r5, 1;
mov.f32 %f1124, %f1125;
mov.f32 %f1123, %f1125;
@%p16 bra $L__BB0_34;
cvt.rn.f32.s32 %f229, %r5;
rcp.rn.f32 %f20, %f229;
ld.global.u32 %r359, [%rd3];
ld.const.f32 %f21, [params+536];
ld.const.u8 %rs15, [params+104];
and.b16 %rs16, %rs15, 32;
setp.eq.s16 %p17, %rs16, 0;
mov.u32 %r70, 0;
selp.f32 %f22, 0f3F800000, 0f41200000, %p17;
mul.f32 %f23, %f14, 0f3456BF95;
mul.f32 %f24, %f15, 0f3456BF95;
mul.f32 %f25, %f16, 0f3456BF95;
ld.const.u64 %rd4, [params+96];
mul.f32 %f230, %f187, %f18;
mul.f32 %f231, %f188, %f17;
sub.f32 %f26, %f231, %f230;
mul.f32 %f232, %f189, %f17;
mul.f32 %f233, %f187, %f19;
sub.f32 %f27, %f233, %f232;
mul.f32 %f234, %f188, %f19;
mul.f32 %f235, %f189, %f18;
sub.f32 %f28, %f235, %f234;
add.s64 %rd5, %rd1, 24;
mov.u64 %rd41, __cudart_i2opi_f;
abs.f32 %f318, %f24;
abs.f32 %f319, %f23;
max.f32 %f320, %f319, %f318;
abs.f32 %f321, %f25;
max.f32 %f322, %f320, %f321;
mov.u32 %r356, %r70;
$L__BB0_7:
mov.u32 %r358, %r70;
$L__BB0_8:
cvt.rn.f32.s32 %f1074, %r356;
mad.lo.s32 %r72, %r359, 1664525, 1013904223;
and.b32 %r73, %r72, 16777215;
cvt.rn.f32.u32 %f236, %r73;
fma.rn.f32 %f237, %f236, 0f33800000, %f1074;
mul.f32 %f238, %f20, %f237;
mad.lo.s32 %r359, %r72, 1664525, 1013904223;
and.b32 %r74, %r359, 16777215;
cvt.rn.f32.u32 %f239, %r74;
cvt.rn.f32.s32 %f240, %r358;
fma.rn.f32 %f241, %f239, 0f33800000, %f240;
mul.f32 %f242, %f20, %f241;
sqrt.rn.f32 %f36, %f238;
mul.f32 %f37, %f242, 0f40C90FDB;
mul.f32 %f243, %f37, 0f3F22F983;
cvt.rni.s32.f32 %r366, %f243;
cvt.rn.f32.s32 %f244, %r366;
mov.f32 %f245, 0fBFC90FDA;
fma.rn.f32 %f246, %f244, %f245, %f37;
mov.f32 %f247, 0fB3A22168;
fma.rn.f32 %f248, %f244, %f247, %f246;
mov.f32 %f249, 0fA7C234C5;
fma.rn.f32 %f1129, %f244, %f249, %f248;
abs.f32 %f39, %f37;
setp.ltu.f32 %p18, %f39, 0f47CE4780;
mov.u32 %r363, %r366;
mov.f32 %f1126, %f1129;
@%p18 bra $L__BB0_16;
setp.eq.f32 %p19, %f39, 0f7F800000;
@%p19 bra $L__BB0_15;
bra.uni $L__BB0_10;
$L__BB0_15:
mov.f32 %f252, 0f00000000;
mul.rn.f32 %f1126, %f37, %f252;
mov.u32 %r363, 0;
bra.uni $L__BB0_16;
$L__BB0_10:
mov.b32 %r13, %f37;
bfe.u32 %r76, %r13, 23, 8;
add.s32 %r14, %r76, -128;
shl.b32 %r77, %r13, 8;
or.b32 %r15, %r77, -2147483648;
shr.u32 %r16, %r14, 5;
mov.u64 %rd118, 0;
mov.u32 %r360, 0;
mov.u64 %rd116, %rd1;
mov.u64 %rd117, %rd41;
$L__BB0_11:
.pragma "nounroll";
ld.global.nc.u32 %r78, [%rd117];
mad.wide.u32 %rd43, %r78, %r15, %rd118;
shr.u64 %rd118, %rd43, 32;
st.local.u32 [%rd116], %rd43;
add.s64 %rd117, %rd117, 4;
add.s64 %rd116, %rd116, 4;
add.s32 %r360, %r360, 1;
setp.ne.s32 %p20, %r360, 6;
@%p20 bra $L__BB0_11;
st.local.u32 [%rd5], %rd118;
mov.u32 %r79, 4;
sub.s32 %r19, %r79, %r16;
mov.u32 %r80, 6;
sub.s32 %r81, %r80, %r16;
mul.wide.s32 %rd44, %r81, 4;
add.s64 %rd45, %rd1, %rd44;
ld.local.u32 %r361, [%rd45];
ld.local.u32 %r362, [%rd45+-4];
and.b32 %r22, %r14, 31;
setp.eq.s32 %p21, %r22, 0;
@%p21 bra $L__BB0_14;
mov.u32 %r82, 32;
sub.s32 %r83, %r82, %r22;
shr.u32 %r84, %r362, %r83;
shl.b32 %r85, %r361, %r22;
add.s32 %r361, %r84, %r85;
mul.wide.s32 %rd46, %r19, 4;
add.s64 %rd47, %rd1, %rd46;
ld.local.u32 %r86, [%rd47];
shr.u32 %r87, %r86, %r83;
shl.b32 %r88, %r362, %r22;
add.s32 %r362, %r87, %r88;
$L__BB0_14:
and.b32 %r89, %r13, -2147483648;
shr.u32 %r90, %r362, 30;
shl.b32 %r91, %r361, 2;
or.b32 %r92, %r90, %r91;
shr.u32 %r93, %r92, 31;
shr.u32 %r94, %r361, 30;
add.s32 %r95, %r93, %r94;
neg.s32 %r96, %r95;
setp.eq.s32 %p22, %r89, 0;
selp.b32 %r363, %r95, %r96, %p22;
setp.ne.s32 %p23, %r93, 0;
xor.b32 %r97, %r89, -2147483648;
selp.b32 %r98, %r97, %r89, %p23;
selp.b32 %r99, -1, 0, %p23;
xor.b32 %r100, %r92, %r99;
shl.b32 %r101, %r362, 2;
xor.b32 %r102, %r101, %r99;
cvt.u64.u32 %rd48, %r100;
cvt.u64.u32 %rd49, %r102;
bfi.b64 %rd50, %rd48, %rd49, 32, 32;
cvt.rn.f64.s64 %fd1, %rd50;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f250, %fd2;
setp.eq.s32 %p24, %r98, 0;
neg.f32 %f251, %f250;
selp.f32 %f1126, %f250, %f251, %p24;
$L__BB0_16:
add.s32 %r29, %r363, 1;
and.b32 %r30, %r29, 1;
setp.eq.s32 %p25, %r30, 0;
selp.f32 %f43, %f1126, 0f3F800000, %p25;
mul.rn.f32 %f44, %f1126, %f1126;
mov.f32 %f1127, 0fB94D4153;
@%p25 bra $L__BB0_18;
mov.f32 %f254, 0fBAB607ED;
mov.f32 %f255, 0f37CBAC00;
fma.rn.f32 %f1127, %f255, %f44, %f254;
$L__BB0_18:
selp.f32 %f256, 0f3C0885E4, 0f3D2AAABB, %p25;
fma.rn.f32 %f257, %f1127, %f44, %f256;
selp.f32 %f258, 0fBE2AAAA8, 0fBEFFFFFF, %p25;
fma.rn.f32 %f259, %f257, %f44, %f258;
mov.f32 %f260, 0f00000000;
fma.rn.f32 %f261, %f44, %f43, %f260;
fma.rn.f32 %f1128, %f259, %f261, %f43;
and.b32 %r104, %r29, 2;
setp.eq.s32 %p27, %r104, 0;
@%p27 bra $L__BB0_20;
mov.f32 %f263, 0fBF800000;
fma.rn.f32 %f1128, %f1128, %f263, %f260;
$L__BB0_20:
@%p18 bra $L__BB0_28;
setp.eq.f32 %p29, %f39, 0f7F800000;
@%p29 bra $L__BB0_27;
bra.uni $L__BB0_22;
$L__BB0_27:
mov.f32 %f266, 0f00000000;
mul.rn.f32 %f1129, %f37, %f266;
mov.u32 %r366, 0;
bra.uni $L__BB0_28;
$L__BB0_22:
mov.b32 %r31, %f37;
bfe.u32 %r105, %r31, 23, 8;
add.s32 %r32, %r105, -128;
shl.b32 %r106, %r31, 8;
or.b32 %r33, %r106, -2147483648;
shr.u32 %r34, %r32, 5;
mov.u64 %rd119, 0;
mov.u64 %rd120, %rd119;
$L__BB0_23:
.pragma "nounroll";
shl.b64 %rd53, %rd119, 2;
mov.u64 %rd54, __cudart_i2opi_f;
add.s64 %rd55, %rd54, %rd53;
ld.global.nc.u32 %r107, [%rd55];
mad.wide.u32 %rd56, %r107, %r33, %rd120;
shr.u64 %rd120, %rd56, 32;
add.s64 %rd57, %rd1, %rd53;
st.local.u32 [%rd57], %rd56;
cvt.u32.u64 %r108, %rd119;
add.s32 %r109, %r108, 1;
cvt.s64.s32 %rd119, %r109;
setp.ne.s32 %p30, %r109, 6;
@%p30 bra $L__BB0_23;
st.local.u32 [%rd5], %rd120;
mov.u32 %r110, 4;
sub.s32 %r35, %r110, %r34;
mov.u32 %r111, 6;
sub.s32 %r112, %r111, %r34;
mul.wide.s32 %rd58, %r112, 4;
add.s64 %rd59, %rd1, %rd58;
ld.local.u32 %r364, [%rd59];
ld.local.u32 %r365, [%rd59+-4];
and.b32 %r38, %r32, 31;
setp.eq.s32 %p31, %r38, 0;
@%p31 bra $L__BB0_26;
mov.u32 %r113, 32;
sub.s32 %r114, %r113, %r38;
shr.u32 %r115, %r365, %r114;
shl.b32 %r116, %r364, %r38;
add.s32 %r364, %r115, %r116;
mul.wide.s32 %rd60, %r35, 4;
add.s64 %rd61, %rd1, %rd60;
ld.local.u32 %r117, [%rd61];
shr.u32 %r118, %r117, %r114;
shl.b32 %r119, %r365, %r38;
add.s32 %r365, %r118, %r119;
$L__BB0_26:
and.b32 %r120, %r31, -2147483648;
shr.u32 %r121, %r365, 30;
shl.b32 %r122, %r364, 2;
or.b32 %r123, %r121, %r122;
shr.u32 %r124, %r123, 31;
shr.u32 %r125, %r364, 30;
add.s32 %r126, %r124, %r125;
neg.s32 %r127, %r126;
setp.eq.s32 %p32, %r120, 0;
selp.b32 %r366, %r126, %r127, %p32;
setp.ne.s32 %p33, %r124, 0;
xor.b32 %r128, %r120, -2147483648;
selp.b32 %r129, %r128, %r120, %p33;
selp.b32 %r130, -1, 0, %p33;
xor.b32 %r131, %r123, %r130;
shl.b32 %r132, %r365, 2;
xor.b32 %r133, %r132, %r130;
cvt.u64.u32 %rd62, %r131;
cvt.u64.u32 %rd63, %r133;
bfi.b64 %rd64, %rd62, %rd63, 32, 32;
cvt.rn.f64.s64 %fd3, %rd64;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f264, %fd4;
setp.eq.s32 %p34, %r129, 0;
neg.f32 %f265, %f264;
selp.f32 %f1129, %f264, %f265, %p34;
$L__BB0_28:
mul.f32 %f53, %f36, %f1128;
and.b32 %r45, %r366, 1;
setp.eq.s32 %p35, %r45, 0;
selp.f32 %f54, %f1129, 0f3F800000, %p35;
mul.rn.f32 %f55, %f1129, %f1129;
mov.f32 %f1130, 0fB94D4153;
@%p35 bra $L__BB0_30;
mov.f32 %f268, 0fBAB607ED;
mov.f32 %f269, 0f37CBAC00;
fma.rn.f32 %f1130, %f269, %f55, %f268;
$L__BB0_30:
selp.f32 %f270, 0f3C0885E4, 0f3D2AAABB, %p35;
fma.rn.f32 %f271, %f1130, %f55, %f270;
selp.f32 %f272, 0fBE2AAAA8, 0fBEFFFFFF, %p35;
fma.rn.f32 %f273, %f271, %f55, %f272;
mov.f32 %f274, 0f00000000;
fma.rn.f32 %f275, %f55, %f54, %f274;
fma.rn.f32 %f1131, %f273, %f275, %f54;
and.b32 %r135, %r366, 2;
setp.eq.s32 %p37, %r135, 0;
@%p37 bra $L__BB0_32;
mov.f32 %f277, 0fBF800000;
fma.rn.f32 %f1131, %f1131, %f277, %f274;
$L__BB0_32:
mul.f32 %f287, %f53, %f53;
mov.f32 %f288, 0f3F800000;
sub.f32 %f289, %f288, %f287;
mul.f32 %f290, %f36, %f1131;
mul.f32 %f291, %f290, %f290;
sub.f32 %f292, %f289, %f291;
max.f32 %f293, %f274, %f292;
sqrt.rn.f32 %f294, %f293;
mul.f32 %f295, %f17, %f290;
mul.f32 %f296, %f18, %f290;
mul.f32 %f297, %f19, %f290;
fma.rn.f32 %f298, %f28, %f53, %f295;
fma.rn.f32 %f299, %f27, %f53, %f296;
fma.rn.f32 %f300, %f26, %f53, %f297;
fma.rn.f32 %f301, %f187, %f294, %f298;
fma.rn.f32 %f302, %f188, %f294, %f299;
fma.rn.f32 %f303, %f189, %f294, %f300;
add.f32 %f304, %f187, %f301;
add.f32 %f305, %f188, %f302;
add.f32 %f306, %f189, %f303;
mul.f32 %f307, %f21, %f304;
mul.f32 %f308, %f21, %f305;
mul.f32 %f309, %f21, %f306;
sub.f32 %f310, %f307, %f187;
sub.f32 %f311, %f308, %f188;
sub.f32 %f312, %f309, %f189;
mul.f32 %f313, %f311, %f311;
fma.rn.f32 %f314, %f310, %f310, %f313;
fma.rn.f32 %f315, %f312, %f312, %f314;
sqrt.rn.f32 %f316, %f315;
rcp.rn.f32 %f317, %f316;
mul.f32 %f281, %f317, %f310;
mul.f32 %f282, %f317, %f311;
mul.f32 %f283, %f317, %f312;
mov.f32 %f323, 0f38D1B717;
max.f32 %f324, %f322, %f323;
mul.f32 %f284, %f22, %f324;
mov.f32 %f285, 0f6C4ECB8F;
mov.u32 %r172, 2;
mov.u32 %r173, 1;
mov.u32 %r174, 3;
mov.u32 %r177, 1065353216;
mov.u32 %r206, 0;
// begin inline asm
call(%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167),_optix_trace_typed_32,(%r206,%rd4,%f14,%f15,%f16,%f281,%f282,%f283,%f284,%f285,%f274,%r173,%r206,%r173,%r172,%r173,%r174,%r177,%r177,%r177,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206);
// end inline asm
mov.b32 %f325, %r136;
mov.b32 %f326, %r137;
mov.b32 %f327, %r138;
add.f32 %f1125, %f1125, %f325;
add.f32 %f1124, %f1124, %f326;
add.f32 %f1123, %f1123, %f327;
add.s32 %r358, %r358, 1;
setp.lt.s32 %p38, %r358, %r5;
@%p38 bra $L__BB0_8;
add.s32 %r356, %r356, 1;
setp.lt.s32 %p39, %r356, %r5;
@%p39 bra $L__BB0_7;
$L__BB0_34:
setp.eq.s32 %p40, %r5, 0;
mov.f32 %f1135, 0f3F800000;
mov.f32 %f1136, %f1135;
mov.f32 %f1137, %f1135;
@%p40 bra $L__BB0_36;
mul.lo.s32 %r207, %r5, %r5;
cvt.rn.f32.s32 %f331, %r207;
rcp.rn.f32 %f332, %f331;
mul.f32 %f1135, %f1125, %f332;
mul.f32 %f1136, %f1124, %f332;
mul.f32 %f1137, %f1123, %f332;
$L__BB0_36:
ld.const.f32 %f333, [params+524];
mul.f32 %f334, %f1135, %f333;
ld.const.v2.f32 {%f335, %f336}, [params+528];
mul.f32 %f339, %f1136, %f335;
mul.f32 %f340, %f1137, %f336;
cvt.sat.f32.f32 %f341, %f13;
mul.f32 %f73, %f334, %f341;
mul.f32 %f74, %f339, %f341;
mul.f32 %f75, %f340, %f341;
mul.f32 %f342, %f13, 0f40800000;
cvt.sat.f32.f32 %f343, %f342;
mul.f32 %f344, %f334, %f343;
mul.f32 %f345, %f339, %f343;
mul.f32 %f346, %f340, %f343;
mul.f32 %f76, %f344, 0f3E800000;
mul.f32 %f77, %f345, 0f3E800000;
mul.f32 %f78, %f346, 0f3E800000;
ld.const.u32 %r48, [params+104];
and.b32 %r208, %r48, 8;
setp.eq.s32 %p41, %r208, 0;
@%p41 bra $L__BB0_50;
mov.f32 %f351, 0f3EE8BA2E;
abs.f32 %f80, %f1135;
setp.lt.f32 %p42, %f80, 0f00800000;
mul.f32 %f353, %f80, 0f4B800000;
selp.f32 %f354, %f353, %f80, %p42;
selp.f32 %f355, 0fC3170000, 0fC2FE0000, %p42;
mov.b32 %r211, %f354;
and.b32 %r212, %r211, 8388607;
or.b32 %r213, %r212, 1065353216;
mov.b32 %f356, %r213;
shr.u32 %r214, %r211, 23;
cvt.rn.f32.u32 %f357, %r214;
add.f32 %f358, %f355, %f357;
setp.gt.f32 %p43, %f356, 0f3FB504F3;
mul.f32 %f359, %f356, 0f3F000000;
add.f32 %f360, %f358, 0f3F800000;
selp.f32 %f361, %f360, %f358, %p43;
selp.f32 %f362, %f359, %f356, %p43;
add.f32 %f363, %f362, 0fBF800000;
add.f32 %f364, %f362, 0f3F800000;
rcp.approx.ftz.f32 %f365, %f364;
add.f32 %f366, %f363, %f363;
mul.f32 %f367, %f366, %f365;
mul.f32 %f368, %f367, %f367;
mov.f32 %f369, 0f3C4CAF63;
mov.f32 %f370, 0f3B18F0FE;
fma.rn.f32 %f371, %f370, %f368, %f369;
mov.f32 %f372, 0f3DAAAABD;
fma.rn.f32 %f373, %f371, %f368, %f372;
mul.rn.f32 %f374, %f373, %f368;
mul.rn.f32 %f375, %f374, %f367;
sub.f32 %f376, %f363, %f367;
add.f32 %f377, %f376, %f376;
neg.f32 %f378, %f367;
fma.rn.f32 %f379, %f378, %f363, %f377;
mul.rn.f32 %f380, %f365, %f379;
add.f32 %f381, %f375, %f367;
sub.f32 %f382, %f367, %f381;
add.f32 %f383, %f375, %f382;
add.f32 %f384, %f380, %f383;
add.f32 %f385, %f381, %f384;
sub.f32 %f386, %f381, %f385;
add.f32 %f387, %f384, %f386;
mov.f32 %f388, 0f3F317200;
mul.rn.f32 %f389, %f361, %f388;
mov.f32 %f390, 0f35BFBE8E;
mul.rn.f32 %f391, %f361, %f390;
add.f32 %f392, %f389, %f385;
sub.f32 %f393, %f389, %f392;
add.f32 %f394, %f385, %f393;
add.f32 %f395, %f387, %f394;
add.f32 %f396, %f391, %f395;
add.f32 %f397, %f392, %f396;
sub.f32 %f398, %f392, %f397;
add.f32 %f399, %f396, %f398;
mul.rn.f32 %f400, %f351, %f397;
neg.f32 %f401, %f400;
fma.rn.f32 %f402, %f351, %f397, %f401;
fma.rn.f32 %f403, %f351, %f399, %f402;
mov.f32 %f404, 0f00000000;
fma.rn.f32 %f405, %f404, %f397, %f403;
add.rn.f32 %f406, %f400, %f405;
neg.f32 %f407, %f406;
add.rn.f32 %f408, %f400, %f407;
add.rn.f32 %f409, %f408, %f405;
mov.b32 %r215, %f406;
setp.eq.s32 %p44, %r215, 1118925336;
add.s32 %r216, %r215, -1;
mov.b32 %f410, %r216;
add.f32 %f411, %f409, 0f37000000;
selp.f32 %f81, %f411, %f409, %p44;
selp.f32 %f412, %f410, %f406, %p44;
mov.f32 %f413, 0f3FB8AA3B;
mul.rn.f32 %f414, %f412, %f413;
cvt.rzi.f32.f32 %f415, %f414;
abs.f32 %f416, %f415;
setp.gt.f32 %p45, %f416, 0f42FC0000;
mov.b32 %r217, %f415;
and.b32 %r218, %r217, -2147483648;
or.b32 %r219, %r218, 1123811328;
mov.b32 %f417, %r219;
selp.f32 %f418, %f417, %f415, %p45;
mov.f32 %f419, 0fBF317218;
fma.rn.f32 %f420, %f418, %f419, %f412;
mov.f32 %f421, 0f3102E308;
fma.rn.f32 %f422, %f418, %f421, %f420;
mul.f32 %f423, %f422, 0f3FB8AA3B;
add.f32 %f424, %f418, 0f4B40007F;
mov.b32 %r220, %f424;
shl.b32 %r221, %r220, 23;
mov.b32 %f425, %r221;
ex2.approx.ftz.f32 %f426, %f423;
mul.f32 %f82, %f426, %f425;
setp.eq.f32 %p46, %f82, 0f7F800000;
mov.f32 %f1138, 0f7F800000;
@%p46 bra $L__BB0_39;
fma.rn.f32 %f1138, %f82, %f81, %f82;
$L__BB0_39:
mov.f32 %f1080, 0f3E68BA2E;
cvt.rzi.f32.f32 %f1079, %f1080;
add.f32 %f1078, %f1079, %f1079;
mov.f32 %f1077, 0f3EE8BA2E;
sub.f32 %f1076, %f1077, %f1078;
abs.f32 %f1075, %f1076;
setp.lt.f32 %p47, %f1135, 0f00000000;
setp.eq.f32 %p48, %f1075, 0f3F800000;
and.pred %p1, %p47, %p48;
setp.eq.f32 %p49, %f1135, 0f00000000;
@%p49 bra $L__BB0_43;
bra.uni $L__BB0_40;
$L__BB0_43:
add.f32 %f431, %f1135, %f1135;
selp.f32 %f1140, %f431, 0f00000000, %p48;
bra.uni $L__BB0_44;
$L__BB0_147:
mov.f32 %f1031, 0f00000000;
mov.u32 %r368, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs114, %f1031;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f1031;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs112, %f1031;}
// end inline asm
mov.u16 %rs115, 0;
st.global.v4.u16 [%rd25], {%rs112, %rs113, %rs114, %rs115};
$L__BB0_148:
ld.const.u64 %rd98, [params+256];
cvta.to.global.u64 %rd99, %rd98;
ld.const.u32 %r342, [params+248];
mad.lo.s32 %r343, %r342, %r4, %r3;
mul.wide.u32 %rd100, %r343, 8;
add.s64 %rd26, %rd99, %rd100;
setp.eq.s32 %p167, %r368, 0;
@%p167 bra $L__BB0_150;
ld.global.v4.u16 {%rs122, %rs123, %rs124, %rs125}, [%rd26];
// begin inline asm
{ cvt.f32.f16 %f1032, %rs122;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1033, %rs123;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1034, %rs124;}
// end inline asm
add.f32 %f1035, %f1032, 0f00000000;
add.f32 %f1036, %f1033, 0f00000000;
add.f32 %f1037, %f1034, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs121, %f1037;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs120, %f1036;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs119, %f1035;}
// end inline asm
mov.u16 %rs126, 0;
st.global.v4.u16 [%rd26], {%rs119, %rs120, %rs121, %rs126};
bra.uni $L__BB0_151;
$L__BB0_150:
mov.f32 %f1040, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs129, %f1040;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1040;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs127, %f1040;}
// end inline asm
mov.u16 %rs130, 0;
st.global.v4.u16 [%rd26], {%rs127, %rs128, %rs129, %rs130};
$L__BB0_151:
ld.const.u64 %rd101, [params+272];
cvta.to.global.u64 %rd102, %rd101;
ld.const.u32 %r344, [params+264];
mad.lo.s32 %r345, %r344, %r4, %r3;
mul.wide.u32 %rd103, %r345, 8;
add.s64 %rd27, %rd102, %rd103;
@%p167 bra $L__BB0_153;
ld.global.v4.u16 {%rs137, %rs138, %rs139, %rs140}, [%rd27];
// begin inline asm
{ cvt.f32.f16 %f1041, %rs137;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1042, %rs138;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1043, %rs139;}
// end inline asm
add.f32 %f1044, %f1041, 0f00000000;
add.f32 %f1045, %f1042, 0f00000000;
add.f32 %f1046, %f1043, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs136, %f1046;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs135, %f1045;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs134, %f1044;}
// end inline asm
mov.u16 %rs141, 0;
st.global.v4.u16 [%rd27], {%rs134, %rs135, %rs136, %rs141};
bra.uni $L__BB0_154;
$L__BB0_153:
mov.f32 %f1049, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs144, %f1049;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1049;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs142, %f1049;}
// end inline asm
mov.u16 %rs145, 0;
st.global.v4.u16 [%rd27], {%rs142, %rs143, %rs144, %rs145};
$L__BB0_154:
ld.const.u64 %rd104, [params+288];
cvta.to.global.u64 %rd105, %rd104;
ld.const.u32 %r346, [params+280];
mad.lo.s32 %r347, %r346, %r4, %r3;
mul.wide.u32 %rd106, %r347, 8;
add.s64 %rd28, %rd105, %rd106;
@%p167 bra $L__BB0_156;
ld.global.v4.u16 {%rs152, %rs153, %rs154, %rs155}, [%rd28];
// begin inline asm
{ cvt.f32.f16 %f1050, %rs152;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1051, %rs153;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1052, %rs154;}
// end inline asm
add.f32 %f1053, %f1050, 0f00000000;
add.f32 %f1054, %f1051, 0f00000000;
add.f32 %f1055, %f1052, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs151, %f1055;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs150, %f1054;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs149, %f1053;}
// end inline asm
mov.u16 %rs156, 0;
st.global.v4.u16 [%rd28], {%rs149, %rs150, %rs151, %rs156};
bra.uni $L__BB0_157;
$L__BB0_156:
mov.f32 %f1058, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs159, %f1058;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1058;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs157, %f1058;}
// end inline asm
mov.u16 %rs160, 0;
st.global.v4.u16 [%rd28], {%rs157, %rs158, %rs159, %rs160};
$L__BB0_157:
ld.const.u64 %rd107, [params+304];
cvta.to.global.u64 %rd108, %rd107;
ld.const.u32 %r348, [params+296];
mad.lo.s32 %r349, %r348, %r4, %r3;
mul.wide.u32 %rd109, %r349, 8;
add.s64 %rd29, %rd108, %rd109;
@%p167 bra $L__BB0_159;
ld.global.v4.u16 {%rs167, %rs168, %rs169, %rs170}, [%rd29];
// begin inline asm
{ cvt.f32.f16 %f1059, %rs167;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1060, %rs168;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1061, %rs169;}
// end inline asm
add.f32 %f1062, %f1059, 0f00000000;
add.f32 %f1063, %f1060, 0f00000000;
add.f32 %f1064, %f1061, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs166, %f1064;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs165, %f1063;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs164, %f1062;}
// end inline asm
mov.u16 %rs171, 0;
st.global.v4.u16 [%rd29], {%rs164, %rs165, %rs166, %rs171};
bra.uni $L__BB0_160;
$L__BB0_159:
mov.f32 %f1067, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs174, %f1067;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs173, %f1067;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs172, %f1067;}
// end inline asm
mov.u16 %rs175, 0;
st.global.v4.u16 [%rd29], {%rs172, %rs173, %rs174, %rs175};
bra.uni $L__BB0_160;
$L__BB0_40:
mov.b32 %r222, %f1138;
xor.b32 %r223, %r222, -2147483648;
mov.b32 %f427, %r223;
selp.f32 %f1140, %f427, %f1138, %p1;
setp.geu.f32 %p50, %f1135, 0f00000000;
@%p50 bra $L__BB0_44;
mov.f32 %f428, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f429, %f428;
setp.eq.f32 %p51, %f429, 0f3EE8BA2E;
@%p51 bra $L__BB0_44;
mov.f32 %f1140, 0f7FFFFFFF;
$L__BB0_44:
abs.f32 %f1081, %f1135;
add.f32 %f432, %f1081, 0f3EE8BA2E;
mov.b32 %r224, %f432;
setp.lt.s32 %p53, %r224, 2139095040;
@%p53 bra $L__BB0_49;
abs.f32 %f1082, %f1135;
setp.gtu.f32 %p54, %f1082, 0f7F800000;
@%p54 bra $L__BB0_48;
bra.uni $L__BB0_46;
$L__BB0_48:
add.f32 %f1140, %f1135, 0f3EE8BA2E;
bra.uni $L__BB0_49;
$L__BB0_46:
abs.f32 %f1083, %f1135;
setp.neu.f32 %p55, %f1083, 0f7F800000;
@%p55 bra $L__BB0_49;
selp.f32 %f1140, 0fFF800000, 0f7F800000, %p1;
$L__BB0_49:
ld.const.u64 %rd115, [params+192];
cvta.to.global.u64 %rd114, %rd115;
ld.const.u32 %r354, [params+184];
mad.lo.s32 %r353, %r354, %r4, %r3;
cvt.u64.u32 %rd113, %r353;
mul.f32 %f433, %f1140, 0f437F0000;
setp.eq.f32 %p56, %f1135, 0f3F800000;
selp.f32 %f434, 0f437F0000, %f433, %p56;
cvt.rzi.u32.f32 %r225, %f434;
shl.b64 %rd67, %rd113, 1;
add.s64 %rd68, %rd114, %rd67;
cvt.u16.u32 %rs17, %r225;
mov.u16 %rs18, 255;
st.global.v2.u8 [%rd68], {%rs17, %rs18};
$L__BB0_50:
ld.const.u32 %r355, [params+104];
and.b32 %r226, %r355, 1;
setp.eq.b32 %p57, %r226, 1;
mov.pred %p58, 0;
xor.pred %p59, %p57, %p58;
not.pred %p60, %p59;
@%p60 bra $L__BB0_124;
mov.f32 %f439, 0f3EE66666;
abs.f32 %f92, %f73;
setp.lt.f32 %p61, %f92, 0f00800000;
mul.f32 %f441, %f92, 0f4B800000;
selp.f32 %f442, %f441, %f92, %p61;
selp.f32 %f443, 0fC3170000, 0fC2FE0000, %p61;
mov.b32 %r227, %f442;
and.b32 %r228, %r227, 8388607;
or.b32 %r229, %r228, 1065353216;
mov.b32 %f444, %r229;
shr.u32 %r230, %r227, 23;
cvt.rn.f32.u32 %f445, %r230;
add.f32 %f446, %f443, %f445;
setp.gt.f32 %p62, %f444, 0f3FB504F3;
mul.f32 %f447, %f444, 0f3F000000;
add.f32 %f448, %f446, 0f3F800000;
selp.f32 %f449, %f448, %f446, %p62;
selp.f32 %f450, %f447, %f444, %p62;
add.f32 %f451, %f450, 0fBF800000;
add.f32 %f452, %f450, 0f3F800000;
rcp.approx.ftz.f32 %f453, %f452;
add.f32 %f454, %f451, %f451;
mul.f32 %f455, %f454, %f453;
mul.f32 %f456, %f455, %f455;
mov.f32 %f457, 0f3C4CAF63;
mov.f32 %f458, 0f3B18F0FE;
fma.rn.f32 %f459, %f458, %f456, %f457;
mov.f32 %f460, 0f3DAAAABD;
fma.rn.f32 %f461, %f459, %f456, %f460;
mul.rn.f32 %f462, %f461, %f456;
mul.rn.f32 %f463, %f462, %f455;
sub.f32 %f464, %f451, %f455;
add.f32 %f465, %f464, %f464;
neg.f32 %f466, %f455;
fma.rn.f32 %f467, %f466, %f451, %f465;
mul.rn.f32 %f468, %f453, %f467;
add.f32 %f469, %f463, %f455;
sub.f32 %f470, %f455, %f469;
add.f32 %f471, %f463, %f470;
add.f32 %f472, %f468, %f471;
add.f32 %f473, %f469, %f472;
sub.f32 %f474, %f469, %f473;
add.f32 %f475, %f472, %f474;
mov.f32 %f476, 0f3F317200;
mul.rn.f32 %f477, %f449, %f476;
mov.f32 %f478, 0f35BFBE8E;
mul.rn.f32 %f479, %f449, %f478;
add.f32 %f480, %f477, %f473;
sub.f32 %f481, %f477, %f480;
add.f32 %f482, %f473, %f481;
add.f32 %f483, %f475, %f482;
add.f32 %f484, %f479, %f483;
add.f32 %f485, %f480, %f484;
sub.f32 %f486, %f480, %f485;
add.f32 %f487, %f484, %f486;
mul.rn.f32 %f488, %f439, %f485;
neg.f32 %f489, %f488;
fma.rn.f32 %f490, %f439, %f485, %f489;
fma.rn.f32 %f491, %f439, %f487, %f490;
mov.f32 %f492, 0f00000000;
fma.rn.f32 %f493, %f492, %f485, %f491;
add.rn.f32 %f494, %f488, %f493;
neg.f32 %f495, %f494;
add.rn.f32 %f496, %f488, %f495;
add.rn.f32 %f497, %f496, %f493;
mov.b32 %r231, %f494;
setp.eq.s32 %p63, %r231, 1118925336;
add.s32 %r232, %r231, -1;
mov.b32 %f498, %r232;
add.f32 %f499, %f497, 0f37000000;
selp.f32 %f93, %f499, %f497, %p63;
selp.f32 %f500, %f498, %f494, %p63;
mov.f32 %f501, 0f3FB8AA3B;
mul.rn.f32 %f502, %f500, %f501;
cvt.rzi.f32.f32 %f503, %f502;
abs.f32 %f504, %f503;
setp.gt.f32 %p64, %f504, 0f42FC0000;
mov.b32 %r233, %f503;
and.b32 %r234, %r233, -2147483648;
or.b32 %r235, %r234, 1123811328;
mov.b32 %f505, %r235;
selp.f32 %f506, %f505, %f503, %p64;
mov.f32 %f507, 0fBF317218;
fma.rn.f32 %f508, %f506, %f507, %f500;
mov.f32 %f509, 0f3102E308;
fma.rn.f32 %f510, %f506, %f509, %f508;
mul.f32 %f511, %f510, 0f3FB8AA3B;
add.f32 %f512, %f506, 0f4B40007F;
mov.b32 %r236, %f512;
shl.b32 %r237, %r236, 23;
mov.b32 %f513, %r237;
ex2.approx.ftz.f32 %f514, %f511;
mul.f32 %f94, %f514, %f513;
setp.eq.f32 %p65, %f94, 0f7F800000;
mov.f32 %f1141, 0f7F800000;
@%p65 bra $L__BB0_53;
fma.rn.f32 %f1141, %f94, %f93, %f94;
$L__BB0_53:
mov.f32 %f1089, 0f3E666666;
cvt.rzi.f32.f32 %f1088, %f1089;
add.f32 %f1087, %f1088, %f1088;
mov.f32 %f1086, 0f3EE66666;
sub.f32 %f1085, %f1086, %f1087;
abs.f32 %f1084, %f1085;
setp.lt.f32 %p66, %f73, 0f00000000;
setp.eq.f32 %p67, %f1084, 0f3F800000;
and.pred %p2, %p66, %p67;
setp.eq.f32 %p68, %f73, 0f00000000;
@%p68 bra $L__BB0_57;
bra.uni $L__BB0_54;
$L__BB0_57:
add.f32 %f519, %f73, %f73;
selp.f32 %f1143, %f519, 0f00000000, %p67;
bra.uni $L__BB0_58;
$L__BB0_54:
mov.b32 %r238, %f1141;
xor.b32 %r239, %r238, -2147483648;
mov.b32 %f515, %r239;
selp.f32 %f1143, %f515, %f1141, %p2;
setp.geu.f32 %p69, %f73, 0f00000000;
@%p69 bra $L__BB0_58;
mov.f32 %f516, 0f3EE66666;
cvt.rzi.f32.f32 %f517, %f516;
setp.eq.f32 %p70, %f517, 0f3EE66666;
@%p70 bra $L__BB0_58;
mov.f32 %f1143, 0f7FFFFFFF;
$L__BB0_58:
abs.f32 %f1090, %f73;
add.f32 %f520, %f1090, 0f3EE66666;
mov.b32 %r240, %f520;
setp.lt.s32 %p72, %r240, 2139095040;
@%p72 bra $L__BB0_63;
abs.f32 %f1091, %f73;
setp.gtu.f32 %p73, %f1091, 0f7F800000;
@%p73 bra $L__BB0_62;
bra.uni $L__BB0_60;
$L__BB0_62:
add.f32 %f1143, %f73, 0f3EE66666;
bra.uni $L__BB0_63;
$L__BB0_60:
abs.f32 %f1092, %f73;
setp.neu.f32 %p74, %f1092, 0f7F800000;
@%p74 bra $L__BB0_63;
selp.f32 %f1143, 0fFF800000, 0f7F800000, %p2;
$L__BB0_63:
setp.eq.f32 %p75, %f73, 0f3F800000;
selp.f32 %f103, 0f3F800000, %f1143, %p75;
abs.f32 %f104, %f74;
setp.lt.f32 %p76, %f104, 0f00800000;
mul.f32 %f522, %f104, 0f4B800000;
selp.f32 %f523, %f522, %f104, %p76;
selp.f32 %f524, 0fC3170000, 0fC2FE0000, %p76;
mov.b32 %r241, %f523;
and.b32 %r242, %r241, 8388607;
or.b32 %r243, %r242, 1065353216;
mov.b32 %f525, %r243;
shr.u32 %r244, %r241, 23;
cvt.rn.f32.u32 %f526, %r244;
add.f32 %f527, %f524, %f526;
setp.gt.f32 %p77, %f525, 0f3FB504F3;
mul.f32 %f528, %f525, 0f3F000000;
add.f32 %f529, %f527, 0f3F800000;
selp.f32 %f530, %f529, %f527, %p77;
selp.f32 %f531, %f528, %f525, %p77;
add.f32 %f532, %f531, 0fBF800000;
add.f32 %f533, %f531, 0f3F800000;
rcp.approx.ftz.f32 %f534, %f533;
add.f32 %f535, %f532, %f532;
mul.f32 %f536, %f535, %f534;
mul.f32 %f537, %f536, %f536;
mov.f32 %f538, 0f3C4CAF63;
mov.f32 %f539, 0f3B18F0FE;
fma.rn.f32 %f540, %f539, %f537, %f538;
mov.f32 %f541, 0f3DAAAABD;
fma.rn.f32 %f542, %f540, %f537, %f541;
mul.rn.f32 %f543, %f542, %f537;
mul.rn.f32 %f544, %f543, %f536;
sub.f32 %f545, %f532, %f536;
add.f32 %f546, %f545, %f545;
neg.f32 %f547, %f536;
fma.rn.f32 %f548, %f547, %f532, %f546;
mul.rn.f32 %f549, %f534, %f548;
add.f32 %f550, %f544, %f536;
sub.f32 %f551, %f536, %f550;
add.f32 %f552, %f544, %f551;
add.f32 %f553, %f549, %f552;
add.f32 %f554, %f550, %f553;
sub.f32 %f555, %f550, %f554;
add.f32 %f556, %f553, %f555;
mov.f32 %f557, 0f3F317200;
mul.rn.f32 %f558, %f530, %f557;
mov.f32 %f559, 0f35BFBE8E;
mul.rn.f32 %f560, %f530, %f559;
add.f32 %f561, %f558, %f554;
sub.f32 %f562, %f558, %f561;
add.f32 %f563, %f554, %f562;
add.f32 %f564, %f556, %f563;
add.f32 %f565, %f560, %f564;
add.f32 %f566, %f561, %f565;
sub.f32 %f567, %f561, %f566;
add.f32 %f568, %f565, %f567;
mov.f32 %f569, 0f3EE66666;
mul.rn.f32 %f570, %f569, %f566;
neg.f32 %f571, %f570;
fma.rn.f32 %f572, %f569, %f566, %f571;
fma.rn.f32 %f573, %f569, %f568, %f572;
mov.f32 %f574, 0f00000000;
fma.rn.f32 %f575, %f574, %f566, %f573;
add.rn.f32 %f576, %f570, %f575;
neg.f32 %f577, %f576;
add.rn.f32 %f578, %f570, %f577;
add.rn.f32 %f579, %f578, %f575;
mov.b32 %r245, %f576;
setp.eq.s32 %p78, %r245, 1118925336;
add.s32 %r246, %r245, -1;
mov.b32 %f580, %r246;
add.f32 %f581, %f579, 0f37000000;
selp.f32 %f105, %f581, %f579, %p78;
selp.f32 %f582, %f580, %f576, %p78;
mov.f32 %f583, 0f3FB8AA3B;
mul.rn.f32 %f584, %f582, %f583;
cvt.rzi.f32.f32 %f585, %f584;
abs.f32 %f586, %f585;
setp.gt.f32 %p79, %f586, 0f42FC0000;
mov.b32 %r247, %f585;
and.b32 %r248, %r247, -2147483648;
or.b32 %r249, %r248, 1123811328;
mov.b32 %f587, %r249;
selp.f32 %f588, %f587, %f585, %p79;
mov.f32 %f589, 0fBF317218;
fma.rn.f32 %f590, %f588, %f589, %f582;
mov.f32 %f591, 0f3102E308;
fma.rn.f32 %f592, %f588, %f591, %f590;
mul.f32 %f593, %f592, 0f3FB8AA3B;
add.f32 %f594, %f588, 0f4B40007F;
mov.b32 %r250, %f594;
shl.b32 %r251, %r250, 23;
mov.b32 %f595, %r251;
ex2.approx.ftz.f32 %f596, %f593;
mul.f32 %f106, %f596, %f595;
setp.eq.f32 %p80, %f106, 0f7F800000;
mov.f32 %f1144, 0f7F800000;
@%p80 bra $L__BB0_65;
fma.rn.f32 %f1144, %f106, %f105, %f106;
$L__BB0_65:
setp.lt.f32 %p81, %f74, 0f00000000;
and.pred %p3, %p81, %p67;
setp.eq.f32 %p83, %f74, 0f00000000;
@%p83 bra $L__BB0_69;
bra.uni $L__BB0_66;
$L__BB0_69:
add.f32 %f601, %f74, %f74;
selp.f32 %f1146, %f601, 0f00000000, %p67;
bra.uni $L__BB0_70;
$L__BB0_66:
mov.b32 %r252, %f1144;
xor.b32 %r253, %r252, -2147483648;
mov.b32 %f597, %r253;
selp.f32 %f1146, %f597, %f1144, %p3;
setp.geu.f32 %p84, %f74, 0f00000000;
@%p84 bra $L__BB0_70;
mov.f32 %f598, 0f3EE66666;
cvt.rzi.f32.f32 %f599, %f598;
setp.eq.f32 %p85, %f599, 0f3EE66666;
@%p85 bra $L__BB0_70;
mov.f32 %f1146, 0f7FFFFFFF;
$L__BB0_70:
abs.f32 %f1093, %f74;
add.f32 %f602, %f1093, 0f3EE66666;
mov.b32 %r254, %f602;
setp.lt.s32 %p87, %r254, 2139095040;
@%p87 bra $L__BB0_75;
abs.f32 %f1094, %f74;
setp.gtu.f32 %p88, %f1094, 0f7F800000;
@%p88 bra $L__BB0_74;
bra.uni $L__BB0_72;
$L__BB0_74:
add.f32 %f1146, %f74, 0f3EE66666;
bra.uni $L__BB0_75;
$L__BB0_72:
abs.f32 %f1095, %f74;
setp.neu.f32 %p89, %f1095, 0f7F800000;
@%p89 bra $L__BB0_75;
selp.f32 %f1146, 0fFF800000, 0f7F800000, %p3;
$L__BB0_75:
setp.eq.f32 %p90, %f74, 0f3F800000;
selp.f32 %f115, 0f3F800000, %f1146, %p90;
abs.f32 %f116, %f75;
setp.lt.f32 %p91, %f116, 0f00800000;
mul.f32 %f604, %f116, 0f4B800000;
selp.f32 %f605, %f604, %f116, %p91;
selp.f32 %f606, 0fC3170000, 0fC2FE0000, %p91;
mov.b32 %r255, %f605;
and.b32 %r256, %r255, 8388607;
or.b32 %r257, %r256, 1065353216;
mov.b32 %f607, %r257;
shr.u32 %r258, %r255, 23;
cvt.rn.f32.u32 %f608, %r258;
add.f32 %f609, %f606, %f608;
setp.gt.f32 %p92, %f607, 0f3FB504F3;
mul.f32 %f610, %f607, 0f3F000000;
add.f32 %f611, %f609, 0f3F800000;
selp.f32 %f612, %f611, %f609, %p92;
selp.f32 %f613, %f610, %f607, %p92;
add.f32 %f614, %f613, 0fBF800000;
add.f32 %f615, %f613, 0f3F800000;
rcp.approx.ftz.f32 %f616, %f615;
add.f32 %f617, %f614, %f614;
mul.f32 %f618, %f617, %f616;
mul.f32 %f619, %f618, %f618;
mov.f32 %f620, 0f3C4CAF63;
mov.f32 %f621, 0f3B18F0FE;
fma.rn.f32 %f622, %f621, %f619, %f620;
mov.f32 %f623, 0f3DAAAABD;
fma.rn.f32 %f624, %f622, %f619, %f623;
mul.rn.f32 %f625, %f624, %f619;
mul.rn.f32 %f626, %f625, %f618;
sub.f32 %f627, %f614, %f618;
add.f32 %f628, %f627, %f627;
neg.f32 %f629, %f618;
fma.rn.f32 %f630, %f629, %f614, %f628;
mul.rn.f32 %f631, %f616, %f630;
add.f32 %f632, %f626, %f618;
sub.f32 %f633, %f618, %f632;
add.f32 %f634, %f626, %f633;
add.f32 %f635, %f631, %f634;
add.f32 %f636, %f632, %f635;
sub.f32 %f637, %f632, %f636;
add.f32 %f638, %f635, %f637;
mov.f32 %f639, 0f3F317200;
mul.rn.f32 %f640, %f612, %f639;
mov.f32 %f641, 0f35BFBE8E;
mul.rn.f32 %f642, %f612, %f641;
add.f32 %f643, %f640, %f636;
sub.f32 %f644, %f640, %f643;
add.f32 %f645, %f636, %f644;
add.f32 %f646, %f638, %f645;
add.f32 %f647, %f642, %f646;
add.f32 %f648, %f643, %f647;
sub.f32 %f649, %f643, %f648;
add.f32 %f650, %f647, %f649;
mov.f32 %f651, 0f3EE66666;
mul.rn.f32 %f652, %f651, %f648;
neg.f32 %f653, %f652;
fma.rn.f32 %f654, %f651, %f648, %f653;
fma.rn.f32 %f655, %f651, %f650, %f654;
mov.f32 %f656, 0f00000000;
fma.rn.f32 %f657, %f656, %f648, %f655;
add.rn.f32 %f658, %f652, %f657;
neg.f32 %f659, %f658;
add.rn.f32 %f660, %f652, %f659;
add.rn.f32 %f661, %f660, %f657;
mov.b32 %r259, %f658;
setp.eq.s32 %p93, %r259, 1118925336;
add.s32 %r260, %r259, -1;
mov.b32 %f662, %r260;
add.f32 %f663, %f661, 0f37000000;
selp.f32 %f117, %f663, %f661, %p93;
selp.f32 %f664, %f662, %f658, %p93;
mov.f32 %f665, 0f3FB8AA3B;
mul.rn.f32 %f666, %f664, %f665;
cvt.rzi.f32.f32 %f667, %f666;
abs.f32 %f668, %f667;
setp.gt.f32 %p94, %f668, 0f42FC0000;
mov.b32 %r261, %f667;
and.b32 %r262, %r261, -2147483648;
or.b32 %r263, %r262, 1123811328;
mov.b32 %f669, %r263;
selp.f32 %f670, %f669, %f667, %p94;
mov.f32 %f671, 0fBF317218;
fma.rn.f32 %f672, %f670, %f671, %f664;
mov.f32 %f673, 0f3102E308;
fma.rn.f32 %f674, %f670, %f673, %f672;
mul.f32 %f675, %f674, 0f3FB8AA3B;
add.f32 %f676, %f670, 0f4B40007F;
mov.b32 %r264, %f676;
shl.b32 %r265, %r264, 23;
mov.b32 %f677, %r265;
ex2.approx.ftz.f32 %f678, %f675;
mul.f32 %f118, %f678, %f677;
setp.eq.f32 %p95, %f118, 0f7F800000;
mov.f32 %f1147, 0f7F800000;
@%p95 bra $L__BB0_77;
fma.rn.f32 %f1147, %f118, %f117, %f118;
$L__BB0_77:
setp.lt.f32 %p96, %f75, 0f00000000;
and.pred %p4, %p96, %p67;
setp.eq.f32 %p98, %f75, 0f00000000;
@%p98 bra $L__BB0_81;
bra.uni $L__BB0_78;
$L__BB0_81:
add.f32 %f683, %f75, %f75;
selp.f32 %f1149, %f683, 0f00000000, %p67;
bra.uni $L__BB0_82;
$L__BB0_78:
mov.b32 %r266, %f1147;
xor.b32 %r267, %r266, -2147483648;
mov.b32 %f679, %r267;
selp.f32 %f1149, %f679, %f1147, %p4;
setp.geu.f32 %p99, %f75, 0f00000000;
@%p99 bra $L__BB0_82;
mov.f32 %f680, 0f3EE66666;
cvt.rzi.f32.f32 %f681, %f680;
setp.eq.f32 %p100, %f681, 0f3EE66666;
@%p100 bra $L__BB0_82;
mov.f32 %f1149, 0f7FFFFFFF;
$L__BB0_82:
abs.f32 %f1096, %f75;
add.f32 %f684, %f1096, 0f3EE66666;
mov.b32 %r268, %f684;
setp.lt.s32 %p102, %r268, 2139095040;
@%p102 bra $L__BB0_87;
abs.f32 %f1097, %f75;
setp.gtu.f32 %p103, %f1097, 0f7F800000;
@%p103 bra $L__BB0_86;
bra.uni $L__BB0_84;
$L__BB0_86:
add.f32 %f1149, %f75, 0f3EE66666;
bra.uni $L__BB0_87;
$L__BB0_84:
abs.f32 %f1098, %f75;
setp.neu.f32 %p104, %f1098, 0f7F800000;
@%p104 bra $L__BB0_87;
selp.f32 %f1149, 0fFF800000, 0f7F800000, %p4;
$L__BB0_87:
setp.eq.f32 %p105, %f75, 0f3F800000;
mov.f32 %f686, 0f3F800000;
selp.f32 %f687, 0f3F800000, %f1149, %p105;
min.f32 %f688, %f103, %f686;
mov.f32 %f689, 0f00000000;
max.f32 %f127, %f689, %f688;
min.f32 %f690, %f115, %f686;
max.f32 %f128, %f689, %f690;
min.f32 %f691, %f687, %f686;
max.f32 %f129, %f689, %f691;
mov.f32 %f695, 0f3ED55555;
abs.f32 %f131, %f127;
setp.lt.f32 %p106, %f131, 0f00800000;
mul.f32 %f697, %f131, 0f4B800000;
selp.f32 %f698, %f697, %f131, %p106;
selp.f32 %f699, 0fC3170000, 0fC2FE0000, %p106;
mov.b32 %r271, %f698;
and.b32 %r272, %r271, 8388607;
or.b32 %r273, %r272, 1065353216;
mov.b32 %f700, %r273;
shr.u32 %r274, %r271, 23;
cvt.rn.f32.u32 %f701, %r274;
add.f32 %f702, %f699, %f701;
setp.gt.f32 %p107, %f700, 0f3FB504F3;
mul.f32 %f703, %f700, 0f3F000000;
add.f32 %f704, %f702, 0f3F800000;
selp.f32 %f705, %f704, %f702, %p107;
selp.f32 %f706, %f703, %f700, %p107;
add.f32 %f707, %f706, 0fBF800000;
add.f32 %f708, %f706, 0f3F800000;
rcp.approx.ftz.f32 %f709, %f708;
add.f32 %f710, %f707, %f707;
mul.f32 %f711, %f710, %f709;
mul.f32 %f712, %f711, %f711;
mov.f32 %f713, 0f3C4CAF63;
mov.f32 %f714, 0f3B18F0FE;
fma.rn.f32 %f715, %f714, %f712, %f713;
mov.f32 %f716, 0f3DAAAABD;
fma.rn.f32 %f717, %f715, %f712, %f716;
mul.rn.f32 %f718, %f717, %f712;
mul.rn.f32 %f719, %f718, %f711;
sub.f32 %f720, %f707, %f711;
add.f32 %f721, %f720, %f720;
neg.f32 %f722, %f711;
fma.rn.f32 %f723, %f722, %f707, %f721;
mul.rn.f32 %f724, %f709, %f723;
add.f32 %f725, %f719, %f711;
sub.f32 %f726, %f711, %f725;
add.f32 %f727, %f719, %f726;
add.f32 %f728, %f724, %f727;
add.f32 %f729, %f725, %f728;
sub.f32 %f730, %f725, %f729;
add.f32 %f731, %f728, %f730;
mov.f32 %f732, 0f3F317200;
mul.rn.f32 %f733, %f705, %f732;
mov.f32 %f734, 0f35BFBE8E;
mul.rn.f32 %f735, %f705, %f734;
add.f32 %f736, %f733, %f729;
sub.f32 %f737, %f733, %f736;
add.f32 %f738, %f729, %f737;
add.f32 %f739, %f731, %f738;
add.f32 %f740, %f735, %f739;
add.f32 %f741, %f736, %f740;
sub.f32 %f742, %f736, %f741;
add.f32 %f743, %f740, %f742;
mul.rn.f32 %f744, %f695, %f741;
neg.f32 %f745, %f744;
fma.rn.f32 %f746, %f695, %f741, %f745;
fma.rn.f32 %f747, %f695, %f743, %f746;
fma.rn.f32 %f748, %f689, %f741, %f747;
add.rn.f32 %f749, %f744, %f748;
neg.f32 %f750, %f749;
add.rn.f32 %f751, %f744, %f750;
add.rn.f32 %f752, %f751, %f748;
mov.b32 %r275, %f749;
setp.eq.s32 %p108, %r275, 1118925336;
add.s32 %r276, %r275, -1;
mov.b32 %f753, %r276;
add.f32 %f754, %f752, 0f37000000;
selp.f32 %f132, %f754, %f752, %p108;
selp.f32 %f755, %f753, %f749, %p108;
mov.f32 %f756, 0f3FB8AA3B;
mul.rn.f32 %f757, %f755, %f756;
cvt.rzi.f32.f32 %f758, %f757;
abs.f32 %f759, %f758;
setp.gt.f32 %p109, %f759, 0f42FC0000;
mov.b32 %r277, %f758;
and.b32 %r278, %r277, -2147483648;
or.b32 %r279, %r278, 1123811328;
mov.b32 %f760, %r279;
selp.f32 %f761, %f760, %f758, %p109;
mov.f32 %f762, 0fBF317218;
fma.rn.f32 %f763, %f761, %f762, %f755;
mov.f32 %f764, 0f3102E308;
fma.rn.f32 %f765, %f761, %f764, %f763;
mul.f32 %f766, %f765, 0f3FB8AA3B;
add.f32 %f767, %f761, 0f4B40007F;
mov.b32 %r280, %f767;
shl.b32 %r281, %r280, 23;
mov.b32 %f768, %r281;
ex2.approx.ftz.f32 %f769, %f766;
mul.f32 %f133, %f769, %f768;
setp.eq.f32 %p110, %f133, 0f7F800000;
mov.f32 %f1150, 0f7F800000;
@%p110 bra $L__BB0_89;
fma.rn.f32 %f1150, %f133, %f132, %f133;
$L__BB0_89:
mov.f32 %f1073, 0f3E555555;
cvt.rzi.f32.f32 %f1072, %f1073;
add.f32 %f1071, %f1072, %f1072;
mov.f32 %f1070, 0f3ED55555;
sub.f32 %f1069, %f1070, %f1071;
abs.f32 %f1068, %f1069;
setp.lt.f32 %p111, %f127, 0f00000000;
setp.eq.f32 %p112, %f1068, 0f3F800000;
and.pred %p5, %p111, %p112;
setp.eq.f32 %p113, %f127, 0f00000000;
@%p113 bra $L__BB0_93;
bra.uni $L__BB0_90;
$L__BB0_93:
add.f32 %f774, %f127, %f127;
selp.f32 %f1152, %f774, 0f00000000, %p112;
bra.uni $L__BB0_94;
$L__BB0_90:
mov.b32 %r282, %f1150;
xor.b32 %r283, %r282, -2147483648;
mov.b32 %f770, %r283;
selp.f32 %f1152, %f770, %f1150, %p5;
setp.geu.f32 %p114, %f127, 0f00000000;
@%p114 bra $L__BB0_94;
mov.f32 %f771, 0f3ED55555;
cvt.rzi.f32.f32 %f772, %f771;
setp.eq.f32 %p115, %f772, 0f3ED55555;
@%p115 bra $L__BB0_94;
mov.f32 %f1152, 0f7FFFFFFF;
$L__BB0_94:
add.f32 %f775, %f131, 0f3ED55555;
mov.b32 %r284, %f775;
setp.lt.s32 %p117, %r284, 2139095040;
@%p117 bra $L__BB0_99;
setp.gtu.f32 %p118, %f131, 0f7F800000;
@%p118 bra $L__BB0_98;
bra.uni $L__BB0_96;
$L__BB0_98:
add.f32 %f1152, %f127, 0f3ED55555;
bra.uni $L__BB0_99;
$L__BB0_96:
setp.neu.f32 %p119, %f131, 0f7F800000;
@%p119 bra $L__BB0_99;
selp.f32 %f1152, 0fFF800000, 0f7F800000, %p5;
$L__BB0_99:
abs.f32 %f142, %f128;
setp.lt.f32 %p120, %f142, 0f00800000;
mul.f32 %f777, %f142, 0f4B800000;
selp.f32 %f778, %f777, %f142, %p120;
selp.f32 %f779, 0fC3170000, 0fC2FE0000, %p120;
mov.b32 %r285, %f778;
and.b32 %r286, %r285, 8388607;
or.b32 %r287, %r286, 1065353216;
mov.b32 %f780, %r287;
shr.u32 %r288, %r285, 23;
cvt.rn.f32.u32 %f781, %r288;
add.f32 %f782, %f779, %f781;
setp.gt.f32 %p121, %f780, 0f3FB504F3;
mul.f32 %f783, %f780, 0f3F000000;
add.f32 %f784, %f782, 0f3F800000;
selp.f32 %f785, %f784, %f782, %p121;
selp.f32 %f786, %f783, %f780, %p121;
add.f32 %f787, %f786, 0fBF800000;
add.f32 %f788, %f786, 0f3F800000;
rcp.approx.ftz.f32 %f789, %f788;
add.f32 %f790, %f787, %f787;
mul.f32 %f791, %f790, %f789;
mul.f32 %f792, %f791, %f791;
mov.f32 %f793, 0f3C4CAF63;
mov.f32 %f794, 0f3B18F0FE;
fma.rn.f32 %f795, %f794, %f792, %f793;
mov.f32 %f796, 0f3DAAAABD;
fma.rn.f32 %f797, %f795, %f792, %f796;
mul.rn.f32 %f798, %f797, %f792;
mul.rn.f32 %f799, %f798, %f791;
sub.f32 %f800, %f787, %f791;
add.f32 %f801, %f800, %f800;
neg.f32 %f802, %f791;
fma.rn.f32 %f803, %f802, %f787, %f801;
mul.rn.f32 %f804, %f789, %f803;
add.f32 %f805, %f799, %f791;
sub.f32 %f806, %f791, %f805;
add.f32 %f807, %f799, %f806;
add.f32 %f808, %f804, %f807;
add.f32 %f809, %f805, %f808;
sub.f32 %f810, %f805, %f809;
add.f32 %f811, %f808, %f810;
mov.f32 %f812, 0f3F317200;
mul.rn.f32 %f813, %f785, %f812;
mov.f32 %f814, 0f35BFBE8E;
mul.rn.f32 %f815, %f785, %f814;
add.f32 %f816, %f813, %f809;
sub.f32 %f817, %f813, %f816;
add.f32 %f818, %f809, %f817;
add.f32 %f819, %f811, %f818;
add.f32 %f820, %f815, %f819;
add.f32 %f821, %f816, %f820;
sub.f32 %f822, %f816, %f821;
add.f32 %f823, %f820, %f822;
mov.f32 %f824, 0f3ED55555;
mul.rn.f32 %f825, %f824, %f821;
neg.f32 %f826, %f825;
fma.rn.f32 %f827, %f824, %f821, %f826;
fma.rn.f32 %f828, %f824, %f823, %f827;
mov.f32 %f829, 0f00000000;
fma.rn.f32 %f830, %f829, %f821, %f828;
add.rn.f32 %f831, %f825, %f830;
neg.f32 %f832, %f831;
add.rn.f32 %f833, %f825, %f832;
add.rn.f32 %f834, %f833, %f830;
mov.b32 %r289, %f831;
setp.eq.s32 %p122, %r289, 1118925336;
add.s32 %r290, %r289, -1;
mov.b32 %f835, %r290;
add.f32 %f836, %f834, 0f37000000;
selp.f32 %f143, %f836, %f834, %p122;
selp.f32 %f837, %f835, %f831, %p122;
mov.f32 %f838, 0f3FB8AA3B;
mul.rn.f32 %f839, %f837, %f838;
cvt.rzi.f32.f32 %f840, %f839;
abs.f32 %f841, %f840;
setp.gt.f32 %p123, %f841, 0f42FC0000;
mov.b32 %r291, %f840;
and.b32 %r292, %r291, -2147483648;
or.b32 %r293, %r292, 1123811328;
mov.b32 %f842, %r293;
selp.f32 %f843, %f842, %f840, %p123;
mov.f32 %f844, 0fBF317218;
fma.rn.f32 %f845, %f843, %f844, %f837;
mov.f32 %f846, 0f3102E308;
fma.rn.f32 %f847, %f843, %f846, %f845;
mul.f32 %f848, %f847, 0f3FB8AA3B;
add.f32 %f849, %f843, 0f4B40007F;
mov.b32 %r294, %f849;
shl.b32 %r295, %r294, 23;
mov.b32 %f850, %r295;
ex2.approx.ftz.f32 %f851, %f848;
mul.f32 %f144, %f851, %f850;
setp.eq.f32 %p124, %f144, 0f7F800000;
mov.f32 %f1153, 0f7F800000;
@%p124 bra $L__BB0_101;
fma.rn.f32 %f1153, %f144, %f143, %f144;
$L__BB0_101:
setp.lt.f32 %p125, %f128, 0f00000000;
and.pred %p6, %p125, %p112;
setp.eq.f32 %p127, %f128, 0f00000000;
@%p127 bra $L__BB0_105;
bra.uni $L__BB0_102;
$L__BB0_105:
add.f32 %f856, %f128, %f128;
selp.f32 %f1155, %f856, 0f00000000, %p112;
bra.uni $L__BB0_106;
$L__BB0_102:
mov.b32 %r296, %f1153;
xor.b32 %r297, %r296, -2147483648;
mov.b32 %f852, %r297;
selp.f32 %f1155, %f852, %f1153, %p6;
setp.geu.f32 %p128, %f128, 0f00000000;
@%p128 bra $L__BB0_106;
mov.f32 %f853, 0f3ED55555;
cvt.rzi.f32.f32 %f854, %f853;
setp.eq.f32 %p129, %f854, 0f3ED55555;
@%p129 bra $L__BB0_106;
mov.f32 %f1155, 0f7FFFFFFF;
$L__BB0_106:
mov.f32 %f1101, 0f00000000;
max.f32 %f1100, %f1101, %f690;
abs.f32 %f1099, %f1100;
add.f32 %f857, %f1099, 0f3ED55555;
mov.b32 %r298, %f857;
setp.lt.s32 %p131, %r298, 2139095040;
@%p131 bra $L__BB0_111;
mov.f32 %f1104, 0f00000000;
max.f32 %f1103, %f1104, %f690;
abs.f32 %f1102, %f1103;
setp.gtu.f32 %p132, %f1102, 0f7F800000;
@%p132 bra $L__BB0_110;
bra.uni $L__BB0_108;
$L__BB0_110:
add.f32 %f1155, %f128, 0f3ED55555;
bra.uni $L__BB0_111;
$L__BB0_108:
mov.f32 %f1107, 0f00000000;
max.f32 %f1106, %f1107, %f690;
abs.f32 %f1105, %f1106;
setp.neu.f32 %p133, %f1105, 0f7F800000;
@%p133 bra $L__BB0_111;
selp.f32 %f1155, 0fFF800000, 0f7F800000, %p6;
$L__BB0_111:
abs.f32 %f153, %f129;
setp.lt.f32 %p134, %f153, 0f00800000;
mul.f32 %f859, %f153, 0f4B800000;
selp.f32 %f860, %f859, %f153, %p134;
selp.f32 %f861, 0fC3170000, 0fC2FE0000, %p134;
mov.b32 %r299, %f860;
and.b32 %r300, %r299, 8388607;
or.b32 %r301, %r300, 1065353216;
mov.b32 %f862, %r301;
shr.u32 %r302, %r299, 23;
cvt.rn.f32.u32 %f863, %r302;
add.f32 %f864, %f861, %f863;
setp.gt.f32 %p135, %f862, 0f3FB504F3;
mul.f32 %f865, %f862, 0f3F000000;
add.f32 %f866, %f864, 0f3F800000;
selp.f32 %f867, %f866, %f864, %p135;
selp.f32 %f868, %f865, %f862, %p135;
add.f32 %f869, %f868, 0fBF800000;
add.f32 %f870, %f868, 0f3F800000;
rcp.approx.ftz.f32 %f871, %f870;
add.f32 %f872, %f869, %f869;
mul.f32 %f873, %f872, %f871;
mul.f32 %f874, %f873, %f873;
mov.f32 %f875, 0f3C4CAF63;
mov.f32 %f876, 0f3B18F0FE;
fma.rn.f32 %f877, %f876, %f874, %f875;
mov.f32 %f878, 0f3DAAAABD;
fma.rn.f32 %f879, %f877, %f874, %f878;
mul.rn.f32 %f880, %f879, %f874;
mul.rn.f32 %f881, %f880, %f873;
sub.f32 %f882, %f869, %f873;
add.f32 %f883, %f882, %f882;
neg.f32 %f884, %f873;
fma.rn.f32 %f885, %f884, %f869, %f883;
mul.rn.f32 %f886, %f871, %f885;
add.f32 %f887, %f881, %f873;
sub.f32 %f888, %f873, %f887;
add.f32 %f889, %f881, %f888;
add.f32 %f890, %f886, %f889;
add.f32 %f891, %f887, %f890;
sub.f32 %f892, %f887, %f891;
add.f32 %f893, %f890, %f892;
mov.f32 %f894, 0f3F317200;
mul.rn.f32 %f895, %f867, %f894;
mov.f32 %f896, 0f35BFBE8E;
mul.rn.f32 %f897, %f867, %f896;
add.f32 %f898, %f895, %f891;
sub.f32 %f899, %f895, %f898;
add.f32 %f900, %f891, %f899;
add.f32 %f901, %f893, %f900;
add.f32 %f902, %f897, %f901;
add.f32 %f903, %f898, %f902;
sub.f32 %f904, %f898, %f903;
add.f32 %f905, %f902, %f904;
mov.f32 %f906, 0f3ED55555;
mul.rn.f32 %f907, %f906, %f903;
neg.f32 %f908, %f907;
fma.rn.f32 %f909, %f906, %f903, %f908;
fma.rn.f32 %f910, %f906, %f905, %f909;
mov.f32 %f911, 0f00000000;
fma.rn.f32 %f912, %f911, %f903, %f910;
add.rn.f32 %f913, %f907, %f912;
neg.f32 %f914, %f913;
add.rn.f32 %f915, %f907, %f914;
add.rn.f32 %f916, %f915, %f912;
mov.b32 %r303, %f913;
setp.eq.s32 %p136, %r303, 1118925336;
add.s32 %r304, %r303, -1;
mov.b32 %f917, %r304;
add.f32 %f918, %f916, 0f37000000;
selp.f32 %f154, %f918, %f916, %p136;
selp.f32 %f919, %f917, %f913, %p136;
mov.f32 %f920, 0f3FB8AA3B;
mul.rn.f32 %f921, %f919, %f920;
cvt.rzi.f32.f32 %f922, %f921;
abs.f32 %f923, %f922;
setp.gt.f32 %p137, %f923, 0f42FC0000;
mov.b32 %r305, %f922;
and.b32 %r306, %r305, -2147483648;
or.b32 %r307, %r306, 1123811328;
mov.b32 %f924, %r307;
selp.f32 %f925, %f924, %f922, %p137;
mov.f32 %f926, 0fBF317218;
fma.rn.f32 %f927, %f925, %f926, %f919;
mov.f32 %f928, 0f3102E308;
fma.rn.f32 %f929, %f925, %f928, %f927;
mul.f32 %f930, %f929, 0f3FB8AA3B;
add.f32 %f931, %f925, 0f4B40007F;
mov.b32 %r308, %f931;
shl.b32 %r309, %r308, 23;
mov.b32 %f932, %r309;
ex2.approx.ftz.f32 %f933, %f930;
mul.f32 %f155, %f933, %f932;
setp.eq.f32 %p138, %f155, 0f7F800000;
mov.f32 %f1156, 0f7F800000;
@%p138 bra $L__BB0_113;
fma.rn.f32 %f1156, %f155, %f154, %f155;
$L__BB0_113:
setp.lt.f32 %p139, %f129, 0f00000000;
and.pred %p7, %p139, %p112;
setp.eq.f32 %p141, %f129, 0f00000000;
@%p141 bra $L__BB0_117;
bra.uni $L__BB0_114;
$L__BB0_117:
add.f32 %f938, %f129, %f129;
selp.f32 %f1158, %f938, 0f00000000, %p112;
bra.uni $L__BB0_118;
$L__BB0_114:
mov.b32 %r310, %f1156;
xor.b32 %r311, %r310, -2147483648;
mov.b32 %f934, %r311;
selp.f32 %f1158, %f934, %f1156, %p7;
setp.geu.f32 %p142, %f129, 0f00000000;
@%p142 bra $L__BB0_118;
mov.f32 %f935, 0f3ED55555;
cvt.rzi.f32.f32 %f936, %f935;
setp.eq.f32 %p143, %f936, 0f3ED55555;
@%p143 bra $L__BB0_118;
mov.f32 %f1158, 0f7FFFFFFF;
$L__BB0_118:
mov.f32 %f1110, 0f00000000;
max.f32 %f1109, %f1110, %f691;
abs.f32 %f1108, %f1109;
add.f32 %f939, %f1108, 0f3ED55555;
mov.b32 %r312, %f939;
setp.lt.s32 %p145, %r312, 2139095040;
@%p145 bra $L__BB0_123;
mov.f32 %f1113, 0f00000000;
max.f32 %f1112, %f1113, %f691;
abs.f32 %f1111, %f1112;
setp.gtu.f32 %p146, %f1111, 0f7F800000;
@%p146 bra $L__BB0_122;
bra.uni $L__BB0_120;
$L__BB0_122:
add.f32 %f1158, %f129, 0f3ED55555;
bra.uni $L__BB0_123;
$L__BB0_120:
mov.f32 %f1116, 0f00000000;
max.f32 %f1115, %f1116, %f691;
abs.f32 %f1114, %f1115;
setp.neu.f32 %p147, %f1114, 0f7F800000;
@%p147 bra $L__BB0_123;
selp.f32 %f1158, 0fFF800000, 0f7F800000, %p7;
$L__BB0_123:
ld.const.u64 %rd112, [params+144];
cvta.to.global.u64 %rd111, %rd112;
ld.const.u32 %r351, [params+136];
mad.lo.s32 %r350, %r351, %r4, %r3;
cvt.u64.u32 %rd110, %r350;
fma.rn.f32 %f940, %f1152, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p148, %f127, 0f3F800000;
mov.f32 %f941, 0f3F800000;
selp.f32 %f942, 0f3F7FFFFF, %f940, %p148;
mul.f32 %f943, %f127, 0f414EB852;
setp.lt.f32 %p149, %f127, 0f3B4D2E1C;
selp.f32 %f944, %f943, %f942, %p149;
fma.rn.f32 %f945, %f1155, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p150, %f128, 0f3F800000;
selp.f32 %f946, 0f3F7FFFFF, %f945, %p150;
mul.f32 %f947, %f128, 0f414EB852;
setp.lt.f32 %p151, %f128, 0f3B4D2E1C;
selp.f32 %f948, %f947, %f946, %p151;
fma.rn.f32 %f949, %f1158, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p152, %f129, 0f3F800000;
selp.f32 %f950, 0f3F7FFFFF, %f949, %p152;
mul.f32 %f951, %f129, 0f414EB852;
setp.lt.f32 %p153, %f129, 0f3B4D2E1C;
selp.f32 %f952, %f951, %f950, %p153;
min.f32 %f953, %f944, %f941;
mov.f32 %f954, 0f00000000;
max.f32 %f955, %f954, %f953;
mul.f32 %f956, %f955, 0f43800000;
cvt.rzi.u32.f32 %r313, %f956;
min.u32 %r314, %r313, 255;
min.f32 %f957, %f948, %f941;
max.f32 %f958, %f954, %f957;
mul.f32 %f959, %f958, 0f43800000;
cvt.rzi.u32.f32 %r315, %f959;
min.u32 %r316, %r315, 255;
min.f32 %f960, %f952, %f941;
max.f32 %f961, %f954, %f960;
mul.f32 %f962, %f961, 0f43800000;
cvt.rzi.u32.f32 %r317, %f962;
min.u32 %r318, %r317, 255;
shl.b64 %rd70, %rd110, 2;
add.s64 %rd71, %rd111, %rd70;
cvt.u16.u32 %rs19, %r318;
cvt.u16.u32 %rs20, %r316;
cvt.u16.u32 %rs21, %r314;
mov.u16 %rs22, 255;
st.global.v4.u8 [%rd71], {%rs21, %rs20, %rs19, %rs22};
$L__BB0_124:
ld.const.u32 %r352, [params+104];
and.b32 %r319, %r352, 4;
setp.eq.s32 %p154, %r319, 0;
ld.const.u32 %r367, [params+108];
@%p154 bra $L__BB0_128;
setp.eq.s32 %p155, %r367, 0;
ld.const.u64 %rd72, [params+224];
cvta.to.global.u64 %rd73, %rd72;
ld.const.u32 %r320, [params+216];
mad.lo.s32 %r321, %r320, %r4, %r3;
mul.wide.u32 %rd74, %r321, 8;
add.s64 %rd20, %rd73, %rd74;
@%p155 bra $L__BB0_127;
ld.global.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd20];
// begin inline asm
{ cvt.f32.f16 %f963, %rs30;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f964, %rs31;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f965, %rs32;}
// end inline asm
add.f32 %f966, %f73, %f963;
add.f32 %f967, %f74, %f964;
add.f32 %f968, %f75, %f965;
mov.f32 %f969, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f968;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f967;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs26, %f966;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f969;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs26, %rs27, %rs28, %rs29};
bra.uni $L__BB0_128;
$L__BB0_127:
mov.f32 %f973, 0f3F800000;
mov.u32 %r367, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f973;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f75;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f74;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs34, %f73;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs34, %rs35, %rs36, %rs37};
$L__BB0_128:
ld.const.u64 %rd75, [params+256];
cvta.to.global.u64 %rd76, %rd75;
ld.const.u32 %r323, [params+248];
mad.lo.s32 %r324, %r323, %r4, %r3;
mul.wide.u32 %rd77, %r324, 8;
add.s64 %rd21, %rd76, %rd77;
setp.eq.s32 %p156, %r367, 0;
@%p156 bra $L__BB0_130;
ld.global.v4.u16 {%rs45, %rs46, %rs47, %rs48}, [%rd21];
// begin inline asm
{ cvt.f32.f16 %f974, %rs45;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f975, %rs46;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f976, %rs47;}
// end inline asm
add.f32 %f977, %f76, %f974;
add.f32 %f978, %f77, %f975;
add.f32 %f979, %f78, %f976;
mov.f32 %f980, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f979;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs42, %f978;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs41, %f977;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f980;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs41, %rs42, %rs43, %rs44};
bra.uni $L__BB0_131;
$L__BB0_130:
mov.f32 %f984, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs52, %f984;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f78;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f77;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs49, %f76;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs49, %rs50, %rs51, %rs52};
$L__BB0_131:
mul.f32 %f985, %f187, 0f3F000000;
mov.f32 %f986, 0f3F000000;
sub.f32 %f164, %f986, %f985;
mul.f32 %f987, %f188, 0f3F000000;
sub.f32 %f165, %f986, %f987;
mul.f32 %f988, %f189, 0f3F000000;
sub.f32 %f166, %f986, %f988;
ld.const.u64 %rd78, [params+272];
cvta.to.global.u64 %rd79, %rd78;
ld.const.u32 %r325, [params+264];
mad.lo.s32 %r326, %r325, %r4, %r3;
mul.wide.u32 %rd80, %r326, 8;
add.s64 %rd22, %rd79, %rd80;
@%p156 bra $L__BB0_133;
ld.global.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd22];
// begin inline asm
{ cvt.f32.f16 %f989, %rs60;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f990, %rs61;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f991, %rs62;}
// end inline asm
add.f32 %f992, %f164, %f989;
add.f32 %f993, %f164, %f990;
add.f32 %f994, %f164, %f991;
mov.f32 %f995, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs58, %f994;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs57, %f993;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs56, %f992;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs59, %f995;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs56, %rs57, %rs58, %rs59};
bra.uni $L__BB0_134;
$L__BB0_133:
mov.f32 %f999, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs67, %f999;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs66, %f164;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs65, %f164;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs64, %f164;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs64, %rs65, %rs66, %rs67};
$L__BB0_134:
ld.const.u64 %rd81, [params+288];
cvta.to.global.u64 %rd82, %rd81;
ld.const.u32 %r327, [params+280];
mad.lo.s32 %r328, %r327, %r4, %r3;
mul.wide.u32 %rd83, %r328, 8;
add.s64 %rd23, %rd82, %rd83;
@%p156 bra $L__BB0_136;
ld.global.v4.u16 {%rs75, %rs76, %rs77, %rs78}, [%rd23];
// begin inline asm
{ cvt.f32.f16 %f1000, %rs75;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1001, %rs76;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1002, %rs77;}
// end inline asm
add.f32 %f1003, %f165, %f1000;
add.f32 %f1004, %f165, %f1001;
add.f32 %f1005, %f165, %f1002;
mov.f32 %f1006, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs73, %f1005;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs72, %f1004;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs71, %f1003;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs74, %f1006;}
// end inline asm
st.global.v4.u16 [%rd23], {%rs71, %rs72, %rs73, %rs74};
bra.uni $L__BB0_137;
$L__BB0_136:
mov.f32 %f1010, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs82, %f1010;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs81, %f165;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs80, %f165;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs79, %f165;}
// end inline asm
st.global.v4.u16 [%rd23], {%rs79, %rs80, %rs81, %rs82};
$L__BB0_137:
ld.const.u64 %rd84, [params+304];
cvta.to.global.u64 %rd85, %rd84;
ld.const.u32 %r329, [params+296];
mad.lo.s32 %r330, %r329, %r4, %r3;
mul.wide.u32 %rd86, %r330, 8;
add.s64 %rd24, %rd85, %rd86;
@%p156 bra $L__BB0_139;
ld.global.v4.u16 {%rs90, %rs91, %rs92, %rs93}, [%rd24];
// begin inline asm
{ cvt.f32.f16 %f1011, %rs90;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1012, %rs91;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1013, %rs92;}
// end inline asm
add.f32 %f1014, %f166, %f1011;
add.f32 %f1015, %f166, %f1012;
add.f32 %f1016, %f166, %f1013;
mov.f32 %f1017, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs88, %f1016;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs87, %f1015;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs86, %f1014;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs89, %f1017;}
// end inline asm
st.global.v4.u16 [%rd24], {%rs86, %rs87, %rs88, %rs89};
bra.uni $L__BB0_160;
$L__BB0_139:
mov.f32 %f1021, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs97, %f1021;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs96, %f166;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs95, %f166;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs94, %f166;}
// end inline asm
st.global.v4.u16 [%rd24], {%rs94, %rs95, %rs96, %rs97};
$L__BB0_160:
ret;
}