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

1621 lines
45 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<146>;
.reg .b16 %rs<52>;
.reg .f32 %f<855>;
.reg .b32 %r<323>;
.reg .f64 %fd<5>;
.reg .b64 %rd<77>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r51), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r52), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u32 %r3, [params+540];
ld.const.u64 %rd21, [params+400];
cvta.to.global.u64 %rd22, %rd21;
ld.const.u32 %r57, [params+392];
mad.lo.s32 %r58, %r57, %r52, %r51;
mul.wide.u32 %rd23, %r58, 4;
add.s64 %rd2, %rd22, %rd23;
ld.global.v2.u8 {%rs7, %rs51}, [%rd2];
or.b16 %rs9, %rs7, %rs51;
and.b16 %rs10, %rs9, 255;
setp.eq.s16 %p7, %rs10, 0;
@%p7 bra $L__BB0_2;
ld.global.u8 %rs50, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs50, [%rd2+2];
setp.eq.s16 %p8, %rs50, 0;
mov.f32 %f824, 0f00000000;
mov.u16 %rs51, 0;
mov.f32 %f825, %f824;
mov.f32 %f826, %f824;
@%p8 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f136, %rs7;
div.rn.f32 %f137, %f136, 0f437F0000;
fma.rn.f32 %f138, %f137, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs51, 255;
cvt.rn.f32.u16 %f139, %rs13;
div.rn.f32 %f140, %f139, 0f437F0000;
fma.rn.f32 %f141, %f140, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f142, %rs50;
div.rn.f32 %f143, %f142, 0f437F0000;
fma.rn.f32 %f144, %f143, 0f40000000, 0fBF800000;
mul.f32 %f145, %f141, %f141;
fma.rn.f32 %f146, %f138, %f138, %f145;
fma.rn.f32 %f147, %f144, %f144, %f146;
sqrt.rn.f32 %f148, %f147;
rcp.rn.f32 %f149, %f148;
mul.f32 %f826, %f149, %f144;
mul.f32 %f825, %f149, %f141;
mul.f32 %f824, %f138, %f149;
$L__BB0_4:
ld.const.v2.u32 {%r59, %r60}, [params];
add.s32 %r4, %r59, %r51;
add.s32 %r5, %r60, %r52;
setp.eq.f32 %p9, %f824, 0f00000000;
setp.eq.f32 %p10, %f825, 0f00000000;
and.pred %p11, %p9, %p10;
setp.eq.f32 %p12, %f826, 0f00000000;
and.pred %p13, %p12, %p11;
@%p13 bra $L__BB0_114;
bra.uni $L__BB0_5;
$L__BB0_114:
ld.const.u32 %r50, [params+104];
and.b32 %r302, %r50, 1;
setp.eq.b32 %p140, %r302, 1;
mov.pred %p141, 0;
xor.pred %p142, %p140, %p141;
not.pred %p143, %p142;
@%p143 bra $L__BB0_116;
ld.const.u64 %rd62, [params+144];
cvta.to.global.u64 %rd63, %rd62;
ld.const.u32 %r303, [params+136];
mad.lo.s32 %r304, %r303, %r5, %r4;
mul.wide.u32 %rd64, %r304, 4;
add.s64 %rd65, %rd63, %rd64;
mov.u16 %rs34, 0;
st.global.v4.u8 [%rd65], {%rs34, %rs34, %rs34, %rs34};
$L__BB0_116:
and.b32 %r305, %r50, 4;
setp.eq.s32 %p144, %r305, 0;
@%p144 bra $L__BB0_120;
ld.const.u32 %r306, [params+108];
setp.eq.s32 %p145, %r306, 0;
ld.const.u64 %rd66, [params+224];
cvta.to.global.u64 %rd67, %rd66;
ld.const.u32 %r307, [params+216];
mad.lo.s32 %r308, %r307, %r5, %r4;
mul.wide.u32 %rd68, %r308, 8;
add.s64 %rd19, %rd67, %rd68;
@%p145 bra $L__BB0_119;
ld.global.v4.u16 {%rs41, %rs42, %rs43, %rs44}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f809, %rs41;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f810, %rs42;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f811, %rs43;}
// end inline asm
add.f32 %f812, %f809, 0f00000000;
add.f32 %f813, %f810, 0f00000000;
add.f32 %f814, %f811, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs40, %f814;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs39, %f813;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f812;}
// end inline asm
mov.u16 %rs45, 0;
st.global.v4.u16 [%rd19], {%rs38, %rs39, %rs40, %rs45};
bra.uni $L__BB0_120;
$L__BB0_5:
ld.const.u64 %rd24, [params+432];
cvta.to.global.u64 %rd25, %rd24;
ld.const.u32 %r63, [params+424];
mad.lo.s32 %r64, %r63, %r52, %r51;
mul.wide.u32 %rd26, %r64, 12;
add.s64 %rd27, %rd25, %rd26;
ld.global.f32 %f151, [%rd27];
mul.f32 %f152, %f151, 0f3456BF95;
ld.global.f32 %f153, [%rd27+4];
mul.f32 %f154, %f153, 0f3456BF95;
ld.global.f32 %f155, [%rd27+8];
mul.f32 %f156, %f155, 0f3456BF95;
abs.f32 %f157, %f824;
div.rn.f32 %f158, %f152, %f157;
abs.f32 %f159, %f825;
div.rn.f32 %f160, %f154, %f159;
abs.f32 %f161, %f826;
div.rn.f32 %f162, %f156, %f161;
abs.f32 %f163, %f158;
abs.f32 %f164, %f160;
abs.f32 %f165, %f162;
mov.f32 %f166, 0f38D1B717;
max.f32 %f167, %f163, %f166;
max.f32 %f168, %f164, %f166;
max.f32 %f169, %f165, %f166;
fma.rn.f32 %f10, %f824, %f167, %f151;
fma.rn.f32 %f11, %f825, %f168, %f153;
fma.rn.f32 %f12, %f826, %f169, %f155;
setp.gt.f32 %p14, %f157, %f161;
neg.f32 %f170, %f825;
selp.f32 %f171, %f170, 0f00000000, %p14;
mov.f32 %f835, 0f00000000;
neg.f32 %f172, %f826;
selp.f32 %f173, %f824, %f172, %p14;
selp.f32 %f174, 0f00000000, %f825, %p14;
mul.f32 %f175, %f173, %f173;
fma.rn.f32 %f176, %f171, %f171, %f175;
fma.rn.f32 %f177, %f174, %f174, %f176;
sqrt.rn.f32 %f178, %f177;
rcp.rn.f32 %f179, %f178;
mul.f32 %f13, %f171, %f179;
mul.f32 %f14, %f173, %f179;
mul.f32 %f15, %f174, %f179;
ld.const.u64 %rd28, [params+128];
cvta.to.global.u64 %rd29, %rd28;
ld.const.u32 %r65, [params+120];
mad.lo.s32 %r66, %r65, %r52, %r51;
mul.wide.u32 %rd30, %r66, 4;
add.s64 %rd3, %rd29, %rd30;
setp.lt.s32 %p15, %r3, 1;
@%p15 bra $L__BB0_36;
ld.const.u32 %r6, [params+588];
cvt.rn.f32.s32 %f181, %r3;
rcp.rn.f32 %f16, %f181;
ld.global.u32 %r315, [%rd3];
mul.f32 %f17, %f10, 0f3456BF95;
mul.f32 %f18, %f11, 0f3456BF95;
mul.f32 %f19, %f12, 0f3456BF95;
ld.const.u64 %rd4, [params+96];
mul.f32 %f182, %f824, %f14;
mul.f32 %f183, %f825, %f13;
sub.f32 %f20, %f183, %f182;
mul.f32 %f184, %f826, %f13;
mul.f32 %f185, %f824, %f15;
sub.f32 %f21, %f185, %f184;
mul.f32 %f186, %f825, %f15;
mul.f32 %f187, %f826, %f14;
sub.f32 %f22, %f187, %f186;
mov.u32 %r67, 0;
add.s64 %rd5, %rd1, 24;
mov.u64 %rd31, __cudart_i2opi_f;
abs.f32 %f254, %f17;
abs.f32 %f255, %f18;
max.f32 %f256, %f254, %f255;
abs.f32 %f257, %f19;
max.f32 %f258, %f256, %f257;
mov.u32 %r312, %r67;
$L__BB0_7:
cvt.rn.f32.s32 %f24, %r312;
mov.u32 %r314, %r67;
$L__BB0_8:
mad.lo.s32 %r69, %r315, 1664525, 1013904223;
and.b32 %r70, %r69, 16777215;
cvt.rn.f32.u32 %f188, %r70;
fma.rn.f32 %f189, %f188, 0f33800000, %f24;
mul.f32 %f190, %f16, %f189;
mad.lo.s32 %r315, %r69, 1664525, 1013904223;
and.b32 %r71, %r315, 16777215;
cvt.rn.f32.u32 %f191, %r71;
cvt.rn.f32.s32 %f192, %r314;
fma.rn.f32 %f193, %f191, 0f33800000, %f192;
mul.f32 %f194, %f16, %f193;
sqrt.rn.f32 %f26, %f190;
mul.f32 %f27, %f194, 0f40C90FDB;
mul.f32 %f195, %f27, 0f3F22F983;
cvt.rni.s32.f32 %r322, %f195;
cvt.rn.f32.s32 %f196, %r322;
mov.f32 %f197, 0fBFC90FDA;
fma.rn.f32 %f198, %f196, %f197, %f27;
mov.f32 %f199, 0fB3A22168;
fma.rn.f32 %f200, %f196, %f199, %f198;
mov.f32 %f201, 0fA7C234C5;
fma.rn.f32 %f832, %f196, %f201, %f200;
abs.f32 %f29, %f27;
setp.ltu.f32 %p16, %f29, 0f47CE4780;
mov.u32 %r319, %r322;
mov.f32 %f829, %f832;
@%p16 bra $L__BB0_16;
setp.eq.f32 %p17, %f29, 0f7F800000;
@%p17 bra $L__BB0_15;
bra.uni $L__BB0_10;
$L__BB0_15:
mov.f32 %f204, 0f00000000;
mul.rn.f32 %f829, %f27, %f204;
mov.u32 %r319, 0;
bra.uni $L__BB0_16;
$L__BB0_10:
mov.b32 %r14, %f27;
bfe.u32 %r73, %r14, 23, 8;
add.s32 %r15, %r73, -128;
shl.b32 %r74, %r14, 8;
or.b32 %r16, %r74, -2147483648;
shr.u32 %r17, %r15, 5;
mov.u64 %rd74, 0;
mov.u32 %r316, 0;
mov.u64 %rd72, %rd1;
mov.u64 %rd73, %rd31;
$L__BB0_11:
.pragma "nounroll";
ld.global.nc.u32 %r75, [%rd73];
mad.wide.u32 %rd33, %r75, %r16, %rd74;
shr.u64 %rd74, %rd33, 32;
st.local.u32 [%rd72], %rd33;
add.s64 %rd73, %rd73, 4;
add.s64 %rd72, %rd72, 4;
add.s32 %r316, %r316, 1;
setp.ne.s32 %p18, %r316, 6;
@%p18 bra $L__BB0_11;
st.local.u32 [%rd5], %rd74;
mov.u32 %r76, 4;
sub.s32 %r20, %r76, %r17;
mov.u32 %r77, 6;
sub.s32 %r78, %r77, %r17;
mul.wide.s32 %rd34, %r78, 4;
add.s64 %rd35, %rd1, %rd34;
ld.local.u32 %r317, [%rd35];
ld.local.u32 %r318, [%rd35+-4];
and.b32 %r23, %r15, 31;
setp.eq.s32 %p19, %r23, 0;
@%p19 bra $L__BB0_14;
mov.u32 %r79, 32;
sub.s32 %r80, %r79, %r23;
shr.u32 %r81, %r318, %r80;
shl.b32 %r82, %r317, %r23;
add.s32 %r317, %r81, %r82;
mul.wide.s32 %rd36, %r20, 4;
add.s64 %rd37, %rd1, %rd36;
ld.local.u32 %r83, [%rd37];
shr.u32 %r84, %r83, %r80;
shl.b32 %r85, %r318, %r23;
add.s32 %r318, %r84, %r85;
$L__BB0_14:
and.b32 %r86, %r14, -2147483648;
shr.u32 %r87, %r318, 30;
shl.b32 %r88, %r317, 2;
or.b32 %r89, %r87, %r88;
shr.u32 %r90, %r89, 31;
shr.u32 %r91, %r317, 30;
add.s32 %r92, %r90, %r91;
neg.s32 %r93, %r92;
setp.eq.s32 %p20, %r86, 0;
selp.b32 %r319, %r92, %r93, %p20;
setp.ne.s32 %p21, %r90, 0;
xor.b32 %r94, %r86, -2147483648;
selp.b32 %r95, %r94, %r86, %p21;
selp.b32 %r96, -1, 0, %p21;
xor.b32 %r97, %r89, %r96;
shl.b32 %r98, %r318, 2;
xor.b32 %r99, %r98, %r96;
cvt.u64.u32 %rd38, %r97;
cvt.u64.u32 %rd39, %r99;
bfi.b64 %rd40, %rd38, %rd39, 32, 32;
cvt.rn.f64.s64 %fd1, %rd40;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f202, %fd2;
setp.eq.s32 %p22, %r95, 0;
neg.f32 %f203, %f202;
selp.f32 %f829, %f202, %f203, %p22;
$L__BB0_16:
add.s32 %r30, %r319, 1;
and.b32 %r31, %r30, 1;
setp.eq.s32 %p23, %r31, 0;
selp.f32 %f33, %f829, 0f3F800000, %p23;
mul.rn.f32 %f34, %f829, %f829;
mov.f32 %f830, 0fB94D4153;
@%p23 bra $L__BB0_18;
mov.f32 %f206, 0fBAB607ED;
mov.f32 %f207, 0f37CBAC00;
fma.rn.f32 %f830, %f207, %f34, %f206;
$L__BB0_18:
selp.f32 %f208, 0f3C0885E4, 0f3D2AAABB, %p23;
fma.rn.f32 %f209, %f830, %f34, %f208;
selp.f32 %f210, 0fBE2AAAA8, 0fBEFFFFFF, %p23;
fma.rn.f32 %f211, %f209, %f34, %f210;
mov.f32 %f212, 0f00000000;
fma.rn.f32 %f213, %f34, %f33, %f212;
fma.rn.f32 %f831, %f211, %f213, %f33;
and.b32 %r101, %r30, 2;
setp.eq.s32 %p25, %r101, 0;
@%p25 bra $L__BB0_20;
mov.f32 %f215, 0fBF800000;
fma.rn.f32 %f831, %f831, %f215, %f212;
$L__BB0_20:
@%p16 bra $L__BB0_28;
setp.eq.f32 %p27, %f29, 0f7F800000;
@%p27 bra $L__BB0_27;
bra.uni $L__BB0_22;
$L__BB0_27:
mov.f32 %f218, 0f00000000;
mul.rn.f32 %f832, %f27, %f218;
mov.u32 %r322, 0;
bra.uni $L__BB0_28;
$L__BB0_22:
mov.b32 %r32, %f27;
bfe.u32 %r102, %r32, 23, 8;
add.s32 %r33, %r102, -128;
shl.b32 %r103, %r32, 8;
or.b32 %r34, %r103, -2147483648;
shr.u32 %r35, %r33, 5;
mov.u64 %rd75, 0;
mov.u64 %rd76, %rd75;
$L__BB0_23:
.pragma "nounroll";
shl.b64 %rd43, %rd75, 2;
mov.u64 %rd44, __cudart_i2opi_f;
add.s64 %rd45, %rd44, %rd43;
ld.global.nc.u32 %r104, [%rd45];
mad.wide.u32 %rd46, %r104, %r34, %rd76;
shr.u64 %rd76, %rd46, 32;
add.s64 %rd47, %rd1, %rd43;
st.local.u32 [%rd47], %rd46;
cvt.u32.u64 %r105, %rd75;
add.s32 %r106, %r105, 1;
cvt.s64.s32 %rd75, %r106;
setp.ne.s32 %p28, %r106, 6;
@%p28 bra $L__BB0_23;
st.local.u32 [%rd5], %rd76;
mov.u32 %r107, 4;
sub.s32 %r36, %r107, %r35;
mov.u32 %r108, 6;
sub.s32 %r109, %r108, %r35;
mul.wide.s32 %rd48, %r109, 4;
add.s64 %rd49, %rd1, %rd48;
ld.local.u32 %r320, [%rd49];
ld.local.u32 %r321, [%rd49+-4];
and.b32 %r39, %r33, 31;
setp.eq.s32 %p29, %r39, 0;
@%p29 bra $L__BB0_26;
mov.u32 %r110, 32;
sub.s32 %r111, %r110, %r39;
shr.u32 %r112, %r321, %r111;
shl.b32 %r113, %r320, %r39;
add.s32 %r320, %r112, %r113;
mul.wide.s32 %rd50, %r36, 4;
add.s64 %rd51, %rd1, %rd50;
ld.local.u32 %r114, [%rd51];
shr.u32 %r115, %r114, %r111;
shl.b32 %r116, %r321, %r39;
add.s32 %r321, %r115, %r116;
$L__BB0_26:
and.b32 %r117, %r32, -2147483648;
shr.u32 %r118, %r321, 30;
shl.b32 %r119, %r320, 2;
or.b32 %r120, %r118, %r119;
shr.u32 %r121, %r120, 31;
shr.u32 %r122, %r320, 30;
add.s32 %r123, %r121, %r122;
neg.s32 %r124, %r123;
setp.eq.s32 %p30, %r117, 0;
selp.b32 %r322, %r123, %r124, %p30;
setp.ne.s32 %p31, %r121, 0;
xor.b32 %r125, %r117, -2147483648;
selp.b32 %r126, %r125, %r117, %p31;
selp.b32 %r127, -1, 0, %p31;
xor.b32 %r128, %r120, %r127;
shl.b32 %r129, %r321, 2;
xor.b32 %r130, %r129, %r127;
cvt.u64.u32 %rd52, %r128;
cvt.u64.u32 %rd53, %r130;
bfi.b64 %rd54, %rd52, %rd53, 32, 32;
cvt.rn.f64.s64 %fd3, %rd54;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f216, %fd4;
setp.eq.s32 %p32, %r126, 0;
neg.f32 %f217, %f216;
selp.f32 %f832, %f216, %f217, %p32;
$L__BB0_28:
mul.f32 %f43, %f26, %f831;
and.b32 %r46, %r322, 1;
setp.eq.s32 %p33, %r46, 0;
selp.f32 %f44, %f832, 0f3F800000, %p33;
mul.rn.f32 %f45, %f832, %f832;
mov.f32 %f833, 0fB94D4153;
@%p33 bra $L__BB0_30;
mov.f32 %f220, 0fBAB607ED;
mov.f32 %f221, 0f37CBAC00;
fma.rn.f32 %f833, %f221, %f45, %f220;
$L__BB0_30:
selp.f32 %f222, 0f3C0885E4, 0f3D2AAABB, %p33;
fma.rn.f32 %f223, %f833, %f45, %f222;
selp.f32 %f224, 0fBE2AAAA8, 0fBEFFFFFF, %p33;
fma.rn.f32 %f225, %f223, %f45, %f224;
mov.f32 %f226, 0f00000000;
fma.rn.f32 %f227, %f45, %f44, %f226;
fma.rn.f32 %f834, %f225, %f227, %f44;
and.b32 %r132, %r322, 2;
setp.eq.s32 %p35, %r132, 0;
@%p35 bra $L__BB0_32;
mov.f32 %f229, 0fBF800000;
fma.rn.f32 %f834, %f834, %f229, %f226;
$L__BB0_32:
mul.f32 %f230, %f43, %f43;
mov.f32 %f231, 0f3F800000;
sub.f32 %f232, %f231, %f230;
mul.f32 %f233, %f26, %f834;
mul.f32 %f234, %f233, %f233;
sub.f32 %f235, %f232, %f234;
max.f32 %f237, %f226, %f235;
sqrt.rn.f32 %f238, %f237;
mul.f32 %f239, %f13, %f233;
mul.f32 %f240, %f14, %f233;
mul.f32 %f241, %f15, %f233;
fma.rn.f32 %f242, %f22, %f43, %f239;
fma.rn.f32 %f243, %f21, %f43, %f240;
fma.rn.f32 %f244, %f20, %f43, %f241;
fma.rn.f32 %f51, %f824, %f238, %f242;
fma.rn.f32 %f52, %f825, %f238, %f243;
fma.rn.f32 %f53, %f826, %f238, %f244;
setp.leu.f32 %p36, %f52, 0f00000000;
setp.ne.s32 %p37, %r6, 0;
and.pred %p38, %p37, %p36;
@%p38 bra $L__BB0_34;
mov.f32 %f259, 0f38D1B717;
max.f32 %f251, %f258, %f259;
mov.f32 %f252, 0f6C4ECB8F;
mov.f32 %f253, 0f00000000;
mov.u32 %r169, 2;
mov.u32 %r171, 1;
mov.u32 %r172, 1065353216;
mov.u32 %r203, 0;
// begin inline asm
call(%r133,%r134,%r135,%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164),_optix_trace_typed_32,(%r203,%rd4,%f10,%f11,%f12,%f51,%f52,%f53,%f251,%f252,%f253,%r171,%r203,%r171,%r169,%r171,%r171,%r172,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203);
// end inline asm
mov.b32 %f260, %r133;
add.f32 %f835, %f835, %f260;
$L__BB0_34:
add.s32 %r314, %r314, 1;
setp.lt.s32 %p39, %r314, %r3;
@%p39 bra $L__BB0_8;
add.s32 %r312, %r312, 1;
setp.lt.s32 %p40, %r312, %r3;
@%p40 bra $L__BB0_7;
$L__BB0_36:
mul.lo.s32 %r204, %r3, %r3;
cvt.rn.f32.s32 %f261, %r204;
div.rn.f32 %f262, %f835, %f261;
ld.const.v4.f32 {%f263, %f264, %f265, %f266}, [params+576];
mul.f32 %f57, %f262, %f263;
mul.f32 %f58, %f262, %f264;
mul.f32 %f59, %f262, %f265;
ld.const.u32 %r49, [params+104];
and.b32 %r205, %r49, 1;
setp.eq.b32 %p41, %r205, 1;
mov.pred %p42, 0;
xor.pred %p43, %p41, %p42;
not.pred %p44, %p43;
@%p44 bra $L__BB0_110;
mov.f32 %f271, 0f3E666666;
cvt.rzi.f32.f32 %f272, %f271;
add.f32 %f273, %f272, %f272;
mov.f32 %f274, 0f3EE66666;
sub.f32 %f275, %f274, %f273;
abs.f32 %f60, %f275;
abs.f32 %f61, %f57;
setp.lt.f32 %p45, %f61, 0f00800000;
mul.f32 %f276, %f61, 0f4B800000;
selp.f32 %f277, %f276, %f61, %p45;
selp.f32 %f278, 0fC3170000, 0fC2FE0000, %p45;
mov.b32 %r206, %f277;
and.b32 %r207, %r206, 8388607;
or.b32 %r208, %r207, 1065353216;
mov.b32 %f279, %r208;
shr.u32 %r209, %r206, 23;
cvt.rn.f32.u32 %f280, %r209;
add.f32 %f281, %f278, %f280;
setp.gt.f32 %p46, %f279, 0f3FB504F3;
mul.f32 %f282, %f279, 0f3F000000;
add.f32 %f283, %f281, 0f3F800000;
selp.f32 %f284, %f283, %f281, %p46;
selp.f32 %f285, %f282, %f279, %p46;
add.f32 %f286, %f285, 0fBF800000;
add.f32 %f287, %f285, 0f3F800000;
rcp.approx.ftz.f32 %f288, %f287;
add.f32 %f289, %f286, %f286;
mul.f32 %f290, %f289, %f288;
mul.f32 %f291, %f290, %f290;
mov.f32 %f292, 0f3C4CAF63;
mov.f32 %f293, 0f3B18F0FE;
fma.rn.f32 %f294, %f293, %f291, %f292;
mov.f32 %f295, 0f3DAAAABD;
fma.rn.f32 %f296, %f294, %f291, %f295;
mul.rn.f32 %f297, %f296, %f291;
mul.rn.f32 %f298, %f297, %f290;
sub.f32 %f299, %f286, %f290;
add.f32 %f300, %f299, %f299;
neg.f32 %f301, %f290;
fma.rn.f32 %f302, %f301, %f286, %f300;
mul.rn.f32 %f303, %f288, %f302;
add.f32 %f304, %f298, %f290;
sub.f32 %f305, %f290, %f304;
add.f32 %f306, %f298, %f305;
add.f32 %f307, %f303, %f306;
add.f32 %f308, %f304, %f307;
sub.f32 %f309, %f304, %f308;
add.f32 %f310, %f307, %f309;
mov.f32 %f311, 0f3F317200;
mul.rn.f32 %f312, %f284, %f311;
mov.f32 %f313, 0f35BFBE8E;
mul.rn.f32 %f314, %f284, %f313;
add.f32 %f315, %f312, %f308;
sub.f32 %f316, %f312, %f315;
add.f32 %f317, %f308, %f316;
add.f32 %f318, %f310, %f317;
add.f32 %f319, %f314, %f318;
add.f32 %f320, %f315, %f319;
sub.f32 %f321, %f315, %f320;
add.f32 %f322, %f319, %f321;
mul.rn.f32 %f323, %f274, %f320;
neg.f32 %f324, %f323;
fma.rn.f32 %f325, %f274, %f320, %f324;
fma.rn.f32 %f326, %f274, %f322, %f325;
mov.f32 %f327, 0f00000000;
fma.rn.f32 %f328, %f327, %f320, %f326;
add.rn.f32 %f329, %f323, %f328;
neg.f32 %f330, %f329;
add.rn.f32 %f331, %f323, %f330;
add.rn.f32 %f332, %f331, %f328;
mov.b32 %r210, %f329;
setp.eq.s32 %p47, %r210, 1118925336;
add.s32 %r211, %r210, -1;
mov.b32 %f333, %r211;
add.f32 %f334, %f332, 0f37000000;
selp.f32 %f62, %f334, %f332, %p47;
selp.f32 %f335, %f333, %f329, %p47;
mov.f32 %f336, 0f3FB8AA3B;
mul.rn.f32 %f337, %f335, %f336;
cvt.rzi.f32.f32 %f338, %f337;
abs.f32 %f339, %f338;
setp.gt.f32 %p48, %f339, 0f42FC0000;
mov.b32 %r212, %f338;
and.b32 %r213, %r212, -2147483648;
or.b32 %r214, %r213, 1123811328;
mov.b32 %f340, %r214;
selp.f32 %f341, %f340, %f338, %p48;
mov.f32 %f342, 0fBF317218;
fma.rn.f32 %f343, %f341, %f342, %f335;
mov.f32 %f344, 0f3102E308;
fma.rn.f32 %f345, %f341, %f344, %f343;
mul.f32 %f346, %f345, 0f3FB8AA3B;
add.f32 %f347, %f341, 0f4B40007F;
mov.b32 %r215, %f347;
shl.b32 %r216, %r215, 23;
mov.b32 %f348, %r216;
ex2.approx.ftz.f32 %f349, %f346;
mul.f32 %f63, %f349, %f348;
setp.eq.f32 %p49, %f63, 0f7F800000;
mov.f32 %f837, 0f7F800000;
@%p49 bra $L__BB0_39;
fma.rn.f32 %f837, %f63, %f62, %f63;
$L__BB0_39:
setp.lt.f32 %p50, %f57, 0f00000000;
setp.eq.f32 %p51, %f60, 0f3F800000;
and.pred %p1, %p50, %p51;
setp.eq.f32 %p52, %f57, 0f00000000;
@%p52 bra $L__BB0_43;
bra.uni $L__BB0_40;
$L__BB0_43:
add.f32 %f354, %f57, %f57;
selp.f32 %f839, %f354, 0f00000000, %p51;
bra.uni $L__BB0_44;
$L__BB0_119:
mov.f32 %f817, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs48, %f817;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs47, %f817;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs46, %f817;}
// end inline asm
mov.u16 %rs49, 0;
st.global.v4.u16 [%rd19], {%rs46, %rs47, %rs48, %rs49};
bra.uni $L__BB0_120;
$L__BB0_40:
mov.b32 %r217, %f837;
xor.b32 %r218, %r217, -2147483648;
mov.b32 %f350, %r218;
selp.f32 %f839, %f350, %f837, %p1;
setp.geu.f32 %p53, %f57, 0f00000000;
@%p53 bra $L__BB0_44;
mov.f32 %f351, 0f3EE66666;
cvt.rzi.f32.f32 %f352, %f351;
setp.eq.f32 %p54, %f352, 0f3EE66666;
@%p54 bra $L__BB0_44;
mov.f32 %f839, 0f7FFFFFFF;
$L__BB0_44:
add.f32 %f355, %f61, 0f3EE66666;
mov.b32 %r219, %f355;
setp.lt.s32 %p56, %r219, 2139095040;
@%p56 bra $L__BB0_49;
setp.gtu.f32 %p57, %f61, 0f7F800000;
@%p57 bra $L__BB0_48;
bra.uni $L__BB0_46;
$L__BB0_48:
add.f32 %f839, %f57, 0f3EE66666;
bra.uni $L__BB0_49;
$L__BB0_46:
setp.neu.f32 %p58, %f61, 0f7F800000;
@%p58 bra $L__BB0_49;
selp.f32 %f839, 0fFF800000, 0f7F800000, %p1;
$L__BB0_49:
setp.eq.f32 %p59, %f57, 0f3F800000;
selp.f32 %f72, 0f3F800000, %f839, %p59;
abs.f32 %f73, %f58;
setp.lt.f32 %p60, %f73, 0f00800000;
mul.f32 %f357, %f73, 0f4B800000;
selp.f32 %f358, %f357, %f73, %p60;
selp.f32 %f359, 0fC3170000, 0fC2FE0000, %p60;
mov.b32 %r220, %f358;
and.b32 %r221, %r220, 8388607;
or.b32 %r222, %r221, 1065353216;
mov.b32 %f360, %r222;
shr.u32 %r223, %r220, 23;
cvt.rn.f32.u32 %f361, %r223;
add.f32 %f362, %f359, %f361;
setp.gt.f32 %p61, %f360, 0f3FB504F3;
mul.f32 %f363, %f360, 0f3F000000;
add.f32 %f364, %f362, 0f3F800000;
selp.f32 %f365, %f364, %f362, %p61;
selp.f32 %f366, %f363, %f360, %p61;
add.f32 %f367, %f366, 0fBF800000;
add.f32 %f368, %f366, 0f3F800000;
rcp.approx.ftz.f32 %f369, %f368;
add.f32 %f370, %f367, %f367;
mul.f32 %f371, %f370, %f369;
mul.f32 %f372, %f371, %f371;
mov.f32 %f373, 0f3C4CAF63;
mov.f32 %f374, 0f3B18F0FE;
fma.rn.f32 %f375, %f374, %f372, %f373;
mov.f32 %f376, 0f3DAAAABD;
fma.rn.f32 %f377, %f375, %f372, %f376;
mul.rn.f32 %f378, %f377, %f372;
mul.rn.f32 %f379, %f378, %f371;
sub.f32 %f380, %f367, %f371;
add.f32 %f381, %f380, %f380;
neg.f32 %f382, %f371;
fma.rn.f32 %f383, %f382, %f367, %f381;
mul.rn.f32 %f384, %f369, %f383;
add.f32 %f385, %f379, %f371;
sub.f32 %f386, %f371, %f385;
add.f32 %f387, %f379, %f386;
add.f32 %f388, %f384, %f387;
add.f32 %f389, %f385, %f388;
sub.f32 %f390, %f385, %f389;
add.f32 %f391, %f388, %f390;
mov.f32 %f392, 0f3F317200;
mul.rn.f32 %f393, %f365, %f392;
mov.f32 %f394, 0f35BFBE8E;
mul.rn.f32 %f395, %f365, %f394;
add.f32 %f396, %f393, %f389;
sub.f32 %f397, %f393, %f396;
add.f32 %f398, %f389, %f397;
add.f32 %f399, %f391, %f398;
add.f32 %f400, %f395, %f399;
add.f32 %f401, %f396, %f400;
sub.f32 %f402, %f396, %f401;
add.f32 %f403, %f400, %f402;
mov.f32 %f404, 0f3EE66666;
mul.rn.f32 %f405, %f404, %f401;
neg.f32 %f406, %f405;
fma.rn.f32 %f407, %f404, %f401, %f406;
fma.rn.f32 %f408, %f404, %f403, %f407;
mov.f32 %f409, 0f00000000;
fma.rn.f32 %f410, %f409, %f401, %f408;
add.rn.f32 %f411, %f405, %f410;
neg.f32 %f412, %f411;
add.rn.f32 %f413, %f405, %f412;
add.rn.f32 %f414, %f413, %f410;
mov.b32 %r224, %f411;
setp.eq.s32 %p62, %r224, 1118925336;
add.s32 %r225, %r224, -1;
mov.b32 %f415, %r225;
add.f32 %f416, %f414, 0f37000000;
selp.f32 %f74, %f416, %f414, %p62;
selp.f32 %f417, %f415, %f411, %p62;
mov.f32 %f418, 0f3FB8AA3B;
mul.rn.f32 %f419, %f417, %f418;
cvt.rzi.f32.f32 %f420, %f419;
abs.f32 %f421, %f420;
setp.gt.f32 %p63, %f421, 0f42FC0000;
mov.b32 %r226, %f420;
and.b32 %r227, %r226, -2147483648;
or.b32 %r228, %r227, 1123811328;
mov.b32 %f422, %r228;
selp.f32 %f423, %f422, %f420, %p63;
mov.f32 %f424, 0fBF317218;
fma.rn.f32 %f425, %f423, %f424, %f417;
mov.f32 %f426, 0f3102E308;
fma.rn.f32 %f427, %f423, %f426, %f425;
mul.f32 %f428, %f427, 0f3FB8AA3B;
add.f32 %f429, %f423, 0f4B40007F;
mov.b32 %r229, %f429;
shl.b32 %r230, %r229, 23;
mov.b32 %f430, %r230;
ex2.approx.ftz.f32 %f431, %f428;
mul.f32 %f75, %f431, %f430;
setp.eq.f32 %p64, %f75, 0f7F800000;
mov.f32 %f840, 0f7F800000;
@%p64 bra $L__BB0_51;
fma.rn.f32 %f840, %f75, %f74, %f75;
$L__BB0_51:
setp.lt.f32 %p65, %f58, 0f00000000;
and.pred %p2, %p65, %p51;
setp.eq.f32 %p67, %f58, 0f00000000;
@%p67 bra $L__BB0_55;
bra.uni $L__BB0_52;
$L__BB0_55:
add.f32 %f436, %f58, %f58;
selp.f32 %f842, %f436, 0f00000000, %p51;
bra.uni $L__BB0_56;
$L__BB0_52:
mov.b32 %r231, %f840;
xor.b32 %r232, %r231, -2147483648;
mov.b32 %f432, %r232;
selp.f32 %f842, %f432, %f840, %p2;
setp.geu.f32 %p68, %f58, 0f00000000;
@%p68 bra $L__BB0_56;
mov.f32 %f433, 0f3EE66666;
cvt.rzi.f32.f32 %f434, %f433;
setp.eq.f32 %p69, %f434, 0f3EE66666;
@%p69 bra $L__BB0_56;
mov.f32 %f842, 0f7FFFFFFF;
$L__BB0_56:
add.f32 %f437, %f73, 0f3EE66666;
mov.b32 %r233, %f437;
setp.lt.s32 %p71, %r233, 2139095040;
@%p71 bra $L__BB0_61;
setp.gtu.f32 %p72, %f73, 0f7F800000;
@%p72 bra $L__BB0_60;
bra.uni $L__BB0_58;
$L__BB0_60:
add.f32 %f842, %f58, 0f3EE66666;
bra.uni $L__BB0_61;
$L__BB0_58:
setp.neu.f32 %p73, %f73, 0f7F800000;
@%p73 bra $L__BB0_61;
selp.f32 %f842, 0fFF800000, 0f7F800000, %p2;
$L__BB0_61:
setp.eq.f32 %p74, %f58, 0f3F800000;
selp.f32 %f84, 0f3F800000, %f842, %p74;
abs.f32 %f85, %f59;
setp.lt.f32 %p75, %f85, 0f00800000;
mul.f32 %f439, %f85, 0f4B800000;
selp.f32 %f440, %f439, %f85, %p75;
selp.f32 %f441, 0fC3170000, 0fC2FE0000, %p75;
mov.b32 %r234, %f440;
and.b32 %r235, %r234, 8388607;
or.b32 %r236, %r235, 1065353216;
mov.b32 %f442, %r236;
shr.u32 %r237, %r234, 23;
cvt.rn.f32.u32 %f443, %r237;
add.f32 %f444, %f441, %f443;
setp.gt.f32 %p76, %f442, 0f3FB504F3;
mul.f32 %f445, %f442, 0f3F000000;
add.f32 %f446, %f444, 0f3F800000;
selp.f32 %f447, %f446, %f444, %p76;
selp.f32 %f448, %f445, %f442, %p76;
add.f32 %f449, %f448, 0fBF800000;
add.f32 %f450, %f448, 0f3F800000;
rcp.approx.ftz.f32 %f451, %f450;
add.f32 %f452, %f449, %f449;
mul.f32 %f453, %f452, %f451;
mul.f32 %f454, %f453, %f453;
mov.f32 %f455, 0f3C4CAF63;
mov.f32 %f456, 0f3B18F0FE;
fma.rn.f32 %f457, %f456, %f454, %f455;
mov.f32 %f458, 0f3DAAAABD;
fma.rn.f32 %f459, %f457, %f454, %f458;
mul.rn.f32 %f460, %f459, %f454;
mul.rn.f32 %f461, %f460, %f453;
sub.f32 %f462, %f449, %f453;
add.f32 %f463, %f462, %f462;
neg.f32 %f464, %f453;
fma.rn.f32 %f465, %f464, %f449, %f463;
mul.rn.f32 %f466, %f451, %f465;
add.f32 %f467, %f461, %f453;
sub.f32 %f468, %f453, %f467;
add.f32 %f469, %f461, %f468;
add.f32 %f470, %f466, %f469;
add.f32 %f471, %f467, %f470;
sub.f32 %f472, %f467, %f471;
add.f32 %f473, %f470, %f472;
mov.f32 %f474, 0f3F317200;
mul.rn.f32 %f475, %f447, %f474;
mov.f32 %f476, 0f35BFBE8E;
mul.rn.f32 %f477, %f447, %f476;
add.f32 %f478, %f475, %f471;
sub.f32 %f479, %f475, %f478;
add.f32 %f480, %f471, %f479;
add.f32 %f481, %f473, %f480;
add.f32 %f482, %f477, %f481;
add.f32 %f483, %f478, %f482;
sub.f32 %f484, %f478, %f483;
add.f32 %f485, %f482, %f484;
mov.f32 %f486, 0f3EE66666;
mul.rn.f32 %f487, %f486, %f483;
neg.f32 %f488, %f487;
fma.rn.f32 %f489, %f486, %f483, %f488;
fma.rn.f32 %f490, %f486, %f485, %f489;
mov.f32 %f491, 0f00000000;
fma.rn.f32 %f492, %f491, %f483, %f490;
add.rn.f32 %f493, %f487, %f492;
neg.f32 %f494, %f493;
add.rn.f32 %f495, %f487, %f494;
add.rn.f32 %f496, %f495, %f492;
mov.b32 %r238, %f493;
setp.eq.s32 %p77, %r238, 1118925336;
add.s32 %r239, %r238, -1;
mov.b32 %f497, %r239;
add.f32 %f498, %f496, 0f37000000;
selp.f32 %f86, %f498, %f496, %p77;
selp.f32 %f499, %f497, %f493, %p77;
mov.f32 %f500, 0f3FB8AA3B;
mul.rn.f32 %f501, %f499, %f500;
cvt.rzi.f32.f32 %f502, %f501;
abs.f32 %f503, %f502;
setp.gt.f32 %p78, %f503, 0f42FC0000;
mov.b32 %r240, %f502;
and.b32 %r241, %r240, -2147483648;
or.b32 %r242, %r241, 1123811328;
mov.b32 %f504, %r242;
selp.f32 %f505, %f504, %f502, %p78;
mov.f32 %f506, 0fBF317218;
fma.rn.f32 %f507, %f505, %f506, %f499;
mov.f32 %f508, 0f3102E308;
fma.rn.f32 %f509, %f505, %f508, %f507;
mul.f32 %f510, %f509, 0f3FB8AA3B;
add.f32 %f511, %f505, 0f4B40007F;
mov.b32 %r243, %f511;
shl.b32 %r244, %r243, 23;
mov.b32 %f512, %r244;
ex2.approx.ftz.f32 %f513, %f510;
mul.f32 %f87, %f513, %f512;
setp.eq.f32 %p79, %f87, 0f7F800000;
mov.f32 %f843, 0f7F800000;
@%p79 bra $L__BB0_63;
fma.rn.f32 %f843, %f87, %f86, %f87;
$L__BB0_63:
setp.lt.f32 %p80, %f59, 0f00000000;
and.pred %p3, %p80, %p51;
setp.eq.f32 %p82, %f59, 0f00000000;
@%p82 bra $L__BB0_67;
bra.uni $L__BB0_64;
$L__BB0_67:
add.f32 %f518, %f59, %f59;
selp.f32 %f845, %f518, 0f00000000, %p51;
bra.uni $L__BB0_68;
$L__BB0_64:
mov.b32 %r245, %f843;
xor.b32 %r246, %r245, -2147483648;
mov.b32 %f514, %r246;
selp.f32 %f845, %f514, %f843, %p3;
setp.geu.f32 %p83, %f59, 0f00000000;
@%p83 bra $L__BB0_68;
mov.f32 %f515, 0f3EE66666;
cvt.rzi.f32.f32 %f516, %f515;
setp.eq.f32 %p84, %f516, 0f3EE66666;
@%p84 bra $L__BB0_68;
mov.f32 %f845, 0f7FFFFFFF;
$L__BB0_68:
add.f32 %f519, %f85, 0f3EE66666;
mov.b32 %r247, %f519;
setp.lt.s32 %p86, %r247, 2139095040;
@%p86 bra $L__BB0_73;
setp.gtu.f32 %p87, %f85, 0f7F800000;
@%p87 bra $L__BB0_72;
bra.uni $L__BB0_70;
$L__BB0_72:
add.f32 %f845, %f59, 0f3EE66666;
bra.uni $L__BB0_73;
$L__BB0_70:
setp.neu.f32 %p88, %f85, 0f7F800000;
@%p88 bra $L__BB0_73;
selp.f32 %f845, 0fFF800000, 0f7F800000, %p3;
$L__BB0_73:
setp.eq.f32 %p89, %f59, 0f3F800000;
mov.f32 %f521, 0f3F800000;
selp.f32 %f522, 0f3F800000, %f845, %p89;
min.f32 %f523, %f72, %f521;
mov.f32 %f524, 0f00000000;
max.f32 %f96, %f524, %f523;
min.f32 %f525, %f84, %f521;
max.f32 %f97, %f524, %f525;
min.f32 %f526, %f522, %f521;
max.f32 %f98, %f524, %f526;
mov.f32 %f530, 0f3ED55555;
abs.f32 %f100, %f96;
setp.lt.f32 %p90, %f100, 0f00800000;
mul.f32 %f532, %f100, 0f4B800000;
selp.f32 %f533, %f532, %f100, %p90;
selp.f32 %f534, 0fC3170000, 0fC2FE0000, %p90;
mov.b32 %r250, %f533;
and.b32 %r251, %r250, 8388607;
or.b32 %r252, %r251, 1065353216;
mov.b32 %f535, %r252;
shr.u32 %r253, %r250, 23;
cvt.rn.f32.u32 %f536, %r253;
add.f32 %f537, %f534, %f536;
setp.gt.f32 %p91, %f535, 0f3FB504F3;
mul.f32 %f538, %f535, 0f3F000000;
add.f32 %f539, %f537, 0f3F800000;
selp.f32 %f540, %f539, %f537, %p91;
selp.f32 %f541, %f538, %f535, %p91;
add.f32 %f542, %f541, 0fBF800000;
add.f32 %f543, %f541, 0f3F800000;
rcp.approx.ftz.f32 %f544, %f543;
add.f32 %f545, %f542, %f542;
mul.f32 %f546, %f545, %f544;
mul.f32 %f547, %f546, %f546;
mov.f32 %f548, 0f3C4CAF63;
mov.f32 %f549, 0f3B18F0FE;
fma.rn.f32 %f550, %f549, %f547, %f548;
mov.f32 %f551, 0f3DAAAABD;
fma.rn.f32 %f552, %f550, %f547, %f551;
mul.rn.f32 %f553, %f552, %f547;
mul.rn.f32 %f554, %f553, %f546;
sub.f32 %f555, %f542, %f546;
add.f32 %f556, %f555, %f555;
neg.f32 %f557, %f546;
fma.rn.f32 %f558, %f557, %f542, %f556;
mul.rn.f32 %f559, %f544, %f558;
add.f32 %f560, %f554, %f546;
sub.f32 %f561, %f546, %f560;
add.f32 %f562, %f554, %f561;
add.f32 %f563, %f559, %f562;
add.f32 %f564, %f560, %f563;
sub.f32 %f565, %f560, %f564;
add.f32 %f566, %f563, %f565;
mov.f32 %f567, 0f3F317200;
mul.rn.f32 %f568, %f540, %f567;
mov.f32 %f569, 0f35BFBE8E;
mul.rn.f32 %f570, %f540, %f569;
add.f32 %f571, %f568, %f564;
sub.f32 %f572, %f568, %f571;
add.f32 %f573, %f564, %f572;
add.f32 %f574, %f566, %f573;
add.f32 %f575, %f570, %f574;
add.f32 %f576, %f571, %f575;
sub.f32 %f577, %f571, %f576;
add.f32 %f578, %f575, %f577;
mul.rn.f32 %f579, %f530, %f576;
neg.f32 %f580, %f579;
fma.rn.f32 %f581, %f530, %f576, %f580;
fma.rn.f32 %f582, %f530, %f578, %f581;
fma.rn.f32 %f583, %f524, %f576, %f582;
add.rn.f32 %f584, %f579, %f583;
neg.f32 %f585, %f584;
add.rn.f32 %f586, %f579, %f585;
add.rn.f32 %f587, %f586, %f583;
mov.b32 %r254, %f584;
setp.eq.s32 %p92, %r254, 1118925336;
add.s32 %r255, %r254, -1;
mov.b32 %f588, %r255;
add.f32 %f589, %f587, 0f37000000;
selp.f32 %f101, %f589, %f587, %p92;
selp.f32 %f590, %f588, %f584, %p92;
mov.f32 %f591, 0f3FB8AA3B;
mul.rn.f32 %f592, %f590, %f591;
cvt.rzi.f32.f32 %f593, %f592;
abs.f32 %f594, %f593;
setp.gt.f32 %p93, %f594, 0f42FC0000;
mov.b32 %r256, %f593;
and.b32 %r257, %r256, -2147483648;
or.b32 %r258, %r257, 1123811328;
mov.b32 %f595, %r258;
selp.f32 %f596, %f595, %f593, %p93;
mov.f32 %f597, 0fBF317218;
fma.rn.f32 %f598, %f596, %f597, %f590;
mov.f32 %f599, 0f3102E308;
fma.rn.f32 %f600, %f596, %f599, %f598;
mul.f32 %f601, %f600, 0f3FB8AA3B;
add.f32 %f602, %f596, 0f4B40007F;
mov.b32 %r259, %f602;
shl.b32 %r260, %r259, 23;
mov.b32 %f603, %r260;
ex2.approx.ftz.f32 %f604, %f601;
mul.f32 %f102, %f604, %f603;
setp.eq.f32 %p94, %f102, 0f7F800000;
mov.f32 %f846, 0f7F800000;
@%p94 bra $L__BB0_75;
fma.rn.f32 %f846, %f102, %f101, %f102;
$L__BB0_75:
mov.f32 %f823, 0f3E555555;
cvt.rzi.f32.f32 %f822, %f823;
add.f32 %f821, %f822, %f822;
mov.f32 %f820, 0f3ED55555;
sub.f32 %f819, %f820, %f821;
abs.f32 %f818, %f819;
setp.lt.f32 %p95, %f96, 0f00000000;
setp.eq.f32 %p96, %f818, 0f3F800000;
and.pred %p4, %p95, %p96;
setp.eq.f32 %p97, %f96, 0f00000000;
@%p97 bra $L__BB0_79;
bra.uni $L__BB0_76;
$L__BB0_79:
add.f32 %f609, %f96, %f96;
selp.f32 %f848, %f609, 0f00000000, %p96;
bra.uni $L__BB0_80;
$L__BB0_76:
mov.b32 %r261, %f846;
xor.b32 %r262, %r261, -2147483648;
mov.b32 %f605, %r262;
selp.f32 %f848, %f605, %f846, %p4;
setp.geu.f32 %p98, %f96, 0f00000000;
@%p98 bra $L__BB0_80;
mov.f32 %f606, 0f3ED55555;
cvt.rzi.f32.f32 %f607, %f606;
setp.eq.f32 %p99, %f607, 0f3ED55555;
@%p99 bra $L__BB0_80;
mov.f32 %f848, 0f7FFFFFFF;
$L__BB0_80:
add.f32 %f610, %f100, 0f3ED55555;
mov.b32 %r263, %f610;
setp.lt.s32 %p101, %r263, 2139095040;
@%p101 bra $L__BB0_85;
setp.gtu.f32 %p102, %f100, 0f7F800000;
@%p102 bra $L__BB0_84;
bra.uni $L__BB0_82;
$L__BB0_84:
add.f32 %f848, %f96, 0f3ED55555;
bra.uni $L__BB0_85;
$L__BB0_82:
setp.neu.f32 %p103, %f100, 0f7F800000;
@%p103 bra $L__BB0_85;
selp.f32 %f848, 0fFF800000, 0f7F800000, %p4;
$L__BB0_85:
abs.f32 %f111, %f97;
setp.lt.f32 %p104, %f111, 0f00800000;
mul.f32 %f612, %f111, 0f4B800000;
selp.f32 %f613, %f612, %f111, %p104;
selp.f32 %f614, 0fC3170000, 0fC2FE0000, %p104;
mov.b32 %r264, %f613;
and.b32 %r265, %r264, 8388607;
or.b32 %r266, %r265, 1065353216;
mov.b32 %f615, %r266;
shr.u32 %r267, %r264, 23;
cvt.rn.f32.u32 %f616, %r267;
add.f32 %f617, %f614, %f616;
setp.gt.f32 %p105, %f615, 0f3FB504F3;
mul.f32 %f618, %f615, 0f3F000000;
add.f32 %f619, %f617, 0f3F800000;
selp.f32 %f620, %f619, %f617, %p105;
selp.f32 %f621, %f618, %f615, %p105;
add.f32 %f622, %f621, 0fBF800000;
add.f32 %f623, %f621, 0f3F800000;
rcp.approx.ftz.f32 %f624, %f623;
add.f32 %f625, %f622, %f622;
mul.f32 %f626, %f625, %f624;
mul.f32 %f627, %f626, %f626;
mov.f32 %f628, 0f3C4CAF63;
mov.f32 %f629, 0f3B18F0FE;
fma.rn.f32 %f630, %f629, %f627, %f628;
mov.f32 %f631, 0f3DAAAABD;
fma.rn.f32 %f632, %f630, %f627, %f631;
mul.rn.f32 %f633, %f632, %f627;
mul.rn.f32 %f634, %f633, %f626;
sub.f32 %f635, %f622, %f626;
add.f32 %f636, %f635, %f635;
neg.f32 %f637, %f626;
fma.rn.f32 %f638, %f637, %f622, %f636;
mul.rn.f32 %f639, %f624, %f638;
add.f32 %f640, %f634, %f626;
sub.f32 %f641, %f626, %f640;
add.f32 %f642, %f634, %f641;
add.f32 %f643, %f639, %f642;
add.f32 %f644, %f640, %f643;
sub.f32 %f645, %f640, %f644;
add.f32 %f646, %f643, %f645;
mov.f32 %f647, 0f3F317200;
mul.rn.f32 %f648, %f620, %f647;
mov.f32 %f649, 0f35BFBE8E;
mul.rn.f32 %f650, %f620, %f649;
add.f32 %f651, %f648, %f644;
sub.f32 %f652, %f648, %f651;
add.f32 %f653, %f644, %f652;
add.f32 %f654, %f646, %f653;
add.f32 %f655, %f650, %f654;
add.f32 %f656, %f651, %f655;
sub.f32 %f657, %f651, %f656;
add.f32 %f658, %f655, %f657;
mov.f32 %f659, 0f3ED55555;
mul.rn.f32 %f660, %f659, %f656;
neg.f32 %f661, %f660;
fma.rn.f32 %f662, %f659, %f656, %f661;
fma.rn.f32 %f663, %f659, %f658, %f662;
mov.f32 %f664, 0f00000000;
fma.rn.f32 %f665, %f664, %f656, %f663;
add.rn.f32 %f666, %f660, %f665;
neg.f32 %f667, %f666;
add.rn.f32 %f668, %f660, %f667;
add.rn.f32 %f669, %f668, %f665;
mov.b32 %r268, %f666;
setp.eq.s32 %p106, %r268, 1118925336;
add.s32 %r269, %r268, -1;
mov.b32 %f670, %r269;
add.f32 %f671, %f669, 0f37000000;
selp.f32 %f112, %f671, %f669, %p106;
selp.f32 %f672, %f670, %f666, %p106;
mov.f32 %f673, 0f3FB8AA3B;
mul.rn.f32 %f674, %f672, %f673;
cvt.rzi.f32.f32 %f675, %f674;
abs.f32 %f676, %f675;
setp.gt.f32 %p107, %f676, 0f42FC0000;
mov.b32 %r270, %f675;
and.b32 %r271, %r270, -2147483648;
or.b32 %r272, %r271, 1123811328;
mov.b32 %f677, %r272;
selp.f32 %f678, %f677, %f675, %p107;
mov.f32 %f679, 0fBF317218;
fma.rn.f32 %f680, %f678, %f679, %f672;
mov.f32 %f681, 0f3102E308;
fma.rn.f32 %f682, %f678, %f681, %f680;
mul.f32 %f683, %f682, 0f3FB8AA3B;
add.f32 %f684, %f678, 0f4B40007F;
mov.b32 %r273, %f684;
shl.b32 %r274, %r273, 23;
mov.b32 %f685, %r274;
ex2.approx.ftz.f32 %f686, %f683;
mul.f32 %f113, %f686, %f685;
setp.eq.f32 %p108, %f113, 0f7F800000;
mov.f32 %f849, 0f7F800000;
@%p108 bra $L__BB0_87;
fma.rn.f32 %f849, %f113, %f112, %f113;
$L__BB0_87:
setp.lt.f32 %p109, %f97, 0f00000000;
and.pred %p5, %p109, %p96;
setp.eq.f32 %p111, %f97, 0f00000000;
@%p111 bra $L__BB0_91;
bra.uni $L__BB0_88;
$L__BB0_91:
add.f32 %f691, %f97, %f97;
selp.f32 %f851, %f691, 0f00000000, %p96;
bra.uni $L__BB0_92;
$L__BB0_88:
mov.b32 %r275, %f849;
xor.b32 %r276, %r275, -2147483648;
mov.b32 %f687, %r276;
selp.f32 %f851, %f687, %f849, %p5;
setp.geu.f32 %p112, %f97, 0f00000000;
@%p112 bra $L__BB0_92;
mov.f32 %f688, 0f3ED55555;
cvt.rzi.f32.f32 %f689, %f688;
setp.eq.f32 %p113, %f689, 0f3ED55555;
@%p113 bra $L__BB0_92;
mov.f32 %f851, 0f7FFFFFFF;
$L__BB0_92:
add.f32 %f692, %f111, 0f3ED55555;
mov.b32 %r277, %f692;
setp.lt.s32 %p115, %r277, 2139095040;
@%p115 bra $L__BB0_97;
setp.gtu.f32 %p116, %f111, 0f7F800000;
@%p116 bra $L__BB0_96;
bra.uni $L__BB0_94;
$L__BB0_96:
add.f32 %f851, %f97, 0f3ED55555;
bra.uni $L__BB0_97;
$L__BB0_94:
setp.neu.f32 %p117, %f111, 0f7F800000;
@%p117 bra $L__BB0_97;
selp.f32 %f851, 0fFF800000, 0f7F800000, %p5;
$L__BB0_97:
abs.f32 %f122, %f98;
setp.lt.f32 %p118, %f122, 0f00800000;
mul.f32 %f694, %f122, 0f4B800000;
selp.f32 %f695, %f694, %f122, %p118;
selp.f32 %f696, 0fC3170000, 0fC2FE0000, %p118;
mov.b32 %r278, %f695;
and.b32 %r279, %r278, 8388607;
or.b32 %r280, %r279, 1065353216;
mov.b32 %f697, %r280;
shr.u32 %r281, %r278, 23;
cvt.rn.f32.u32 %f698, %r281;
add.f32 %f699, %f696, %f698;
setp.gt.f32 %p119, %f697, 0f3FB504F3;
mul.f32 %f700, %f697, 0f3F000000;
add.f32 %f701, %f699, 0f3F800000;
selp.f32 %f702, %f701, %f699, %p119;
selp.f32 %f703, %f700, %f697, %p119;
add.f32 %f704, %f703, 0fBF800000;
add.f32 %f705, %f703, 0f3F800000;
rcp.approx.ftz.f32 %f706, %f705;
add.f32 %f707, %f704, %f704;
mul.f32 %f708, %f707, %f706;
mul.f32 %f709, %f708, %f708;
mov.f32 %f710, 0f3C4CAF63;
mov.f32 %f711, 0f3B18F0FE;
fma.rn.f32 %f712, %f711, %f709, %f710;
mov.f32 %f713, 0f3DAAAABD;
fma.rn.f32 %f714, %f712, %f709, %f713;
mul.rn.f32 %f715, %f714, %f709;
mul.rn.f32 %f716, %f715, %f708;
sub.f32 %f717, %f704, %f708;
add.f32 %f718, %f717, %f717;
neg.f32 %f719, %f708;
fma.rn.f32 %f720, %f719, %f704, %f718;
mul.rn.f32 %f721, %f706, %f720;
add.f32 %f722, %f716, %f708;
sub.f32 %f723, %f708, %f722;
add.f32 %f724, %f716, %f723;
add.f32 %f725, %f721, %f724;
add.f32 %f726, %f722, %f725;
sub.f32 %f727, %f722, %f726;
add.f32 %f728, %f725, %f727;
mov.f32 %f729, 0f3F317200;
mul.rn.f32 %f730, %f702, %f729;
mov.f32 %f731, 0f35BFBE8E;
mul.rn.f32 %f732, %f702, %f731;
add.f32 %f733, %f730, %f726;
sub.f32 %f734, %f730, %f733;
add.f32 %f735, %f726, %f734;
add.f32 %f736, %f728, %f735;
add.f32 %f737, %f732, %f736;
add.f32 %f738, %f733, %f737;
sub.f32 %f739, %f733, %f738;
add.f32 %f740, %f737, %f739;
mov.f32 %f741, 0f3ED55555;
mul.rn.f32 %f742, %f741, %f738;
neg.f32 %f743, %f742;
fma.rn.f32 %f744, %f741, %f738, %f743;
fma.rn.f32 %f745, %f741, %f740, %f744;
mov.f32 %f746, 0f00000000;
fma.rn.f32 %f747, %f746, %f738, %f745;
add.rn.f32 %f748, %f742, %f747;
neg.f32 %f749, %f748;
add.rn.f32 %f750, %f742, %f749;
add.rn.f32 %f751, %f750, %f747;
mov.b32 %r282, %f748;
setp.eq.s32 %p120, %r282, 1118925336;
add.s32 %r283, %r282, -1;
mov.b32 %f752, %r283;
add.f32 %f753, %f751, 0f37000000;
selp.f32 %f123, %f753, %f751, %p120;
selp.f32 %f754, %f752, %f748, %p120;
mov.f32 %f755, 0f3FB8AA3B;
mul.rn.f32 %f756, %f754, %f755;
cvt.rzi.f32.f32 %f757, %f756;
abs.f32 %f758, %f757;
setp.gt.f32 %p121, %f758, 0f42FC0000;
mov.b32 %r284, %f757;
and.b32 %r285, %r284, -2147483648;
or.b32 %r286, %r285, 1123811328;
mov.b32 %f759, %r286;
selp.f32 %f760, %f759, %f757, %p121;
mov.f32 %f761, 0fBF317218;
fma.rn.f32 %f762, %f760, %f761, %f754;
mov.f32 %f763, 0f3102E308;
fma.rn.f32 %f764, %f760, %f763, %f762;
mul.f32 %f765, %f764, 0f3FB8AA3B;
add.f32 %f766, %f760, 0f4B40007F;
mov.b32 %r287, %f766;
shl.b32 %r288, %r287, 23;
mov.b32 %f767, %r288;
ex2.approx.ftz.f32 %f768, %f765;
mul.f32 %f124, %f768, %f767;
setp.eq.f32 %p122, %f124, 0f7F800000;
mov.f32 %f852, 0f7F800000;
@%p122 bra $L__BB0_99;
fma.rn.f32 %f852, %f124, %f123, %f124;
$L__BB0_99:
setp.lt.f32 %p123, %f98, 0f00000000;
and.pred %p6, %p123, %p96;
setp.eq.f32 %p125, %f98, 0f00000000;
@%p125 bra $L__BB0_103;
bra.uni $L__BB0_100;
$L__BB0_103:
add.f32 %f773, %f98, %f98;
selp.f32 %f854, %f773, 0f00000000, %p96;
bra.uni $L__BB0_104;
$L__BB0_100:
mov.b32 %r289, %f852;
xor.b32 %r290, %r289, -2147483648;
mov.b32 %f769, %r290;
selp.f32 %f854, %f769, %f852, %p6;
setp.geu.f32 %p126, %f98, 0f00000000;
@%p126 bra $L__BB0_104;
mov.f32 %f770, 0f3ED55555;
cvt.rzi.f32.f32 %f771, %f770;
setp.eq.f32 %p127, %f771, 0f3ED55555;
@%p127 bra $L__BB0_104;
mov.f32 %f854, 0f7FFFFFFF;
$L__BB0_104:
add.f32 %f774, %f122, 0f3ED55555;
mov.b32 %r291, %f774;
setp.lt.s32 %p129, %r291, 2139095040;
@%p129 bra $L__BB0_109;
setp.gtu.f32 %p130, %f122, 0f7F800000;
@%p130 bra $L__BB0_108;
bra.uni $L__BB0_106;
$L__BB0_108:
add.f32 %f854, %f98, 0f3ED55555;
bra.uni $L__BB0_109;
$L__BB0_106:
setp.neu.f32 %p131, %f122, 0f7F800000;
@%p131 bra $L__BB0_109;
selp.f32 %f854, 0fFF800000, 0f7F800000, %p6;
$L__BB0_109:
ld.const.u64 %rd71, [params+144];
cvta.to.global.u64 %rd70, %rd71;
ld.const.u32 %r310, [params+136];
mad.lo.s32 %r309, %r310, %r5, %r4;
cvt.u64.u32 %rd69, %r309;
fma.rn.f32 %f775, %f848, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p132, %f96, 0f3F800000;
mov.f32 %f776, 0f3F800000;
selp.f32 %f777, 0f3F7FFFFF, %f775, %p132;
mul.f32 %f778, %f96, 0f414EB852;
setp.lt.f32 %p133, %f96, 0f3B4D2E1C;
selp.f32 %f779, %f778, %f777, %p133;
fma.rn.f32 %f780, %f851, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p134, %f97, 0f3F800000;
selp.f32 %f781, 0f3F7FFFFF, %f780, %p134;
mul.f32 %f782, %f97, 0f414EB852;
setp.lt.f32 %p135, %f97, 0f3B4D2E1C;
selp.f32 %f783, %f782, %f781, %p135;
fma.rn.f32 %f784, %f854, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p136, %f98, 0f3F800000;
selp.f32 %f785, 0f3F7FFFFF, %f784, %p136;
mul.f32 %f786, %f98, 0f414EB852;
setp.lt.f32 %p137, %f98, 0f3B4D2E1C;
selp.f32 %f787, %f786, %f785, %p137;
min.f32 %f788, %f779, %f776;
mov.f32 %f789, 0f00000000;
max.f32 %f790, %f789, %f788;
mul.f32 %f791, %f790, 0f43800000;
cvt.rzi.u32.f32 %r292, %f791;
min.u32 %r293, %r292, 255;
min.f32 %f792, %f783, %f776;
max.f32 %f793, %f789, %f792;
mul.f32 %f794, %f793, 0f43800000;
cvt.rzi.u32.f32 %r294, %f794;
min.u32 %r295, %r294, 255;
min.f32 %f795, %f787, %f776;
max.f32 %f796, %f789, %f795;
mul.f32 %f797, %f796, 0f43800000;
cvt.rzi.u32.f32 %r296, %f797;
min.u32 %r297, %r296, 255;
shl.b64 %rd57, %rd69, 2;
add.s64 %rd58, %rd70, %rd57;
cvt.u16.u32 %rs15, %r297;
cvt.u16.u32 %rs16, %r295;
cvt.u16.u32 %rs17, %r293;
mov.u16 %rs18, 255;
st.global.v4.u8 [%rd58], {%rs17, %rs16, %rs15, %rs18};
$L__BB0_110:
ld.const.u32 %r311, [params+104];
and.b32 %r298, %r311, 4;
setp.eq.s32 %p138, %r298, 0;
@%p138 bra $L__BB0_120;
ld.const.u32 %r299, [params+108];
setp.eq.s32 %p139, %r299, 0;
ld.const.u64 %rd59, [params+224];
cvta.to.global.u64 %rd60, %rd59;
ld.const.u32 %r300, [params+216];
mad.lo.s32 %r301, %r300, %r5, %r4;
mul.wide.u32 %rd61, %r301, 8;
add.s64 %rd18, %rd60, %rd61;
@%p139 bra $L__BB0_113;
ld.global.v4.u16 {%rs26, %rs27, %rs28, %rs29}, [%rd18];
// begin inline asm
{ cvt.f32.f16 %f798, %rs26;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f799, %rs27;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f800, %rs28;}
// end inline asm
add.f32 %f801, %f57, %f798;
add.f32 %f802, %f58, %f799;
add.f32 %f803, %f59, %f800;
mov.f32 %f804, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs24, %f803;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs23, %f802;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs22, %f801;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs25, %f804;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs22, %rs23, %rs24, %rs25};
bra.uni $L__BB0_120;
$L__BB0_113:
mov.f32 %f808, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs33, %f808;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs32, %f59;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs31, %f58;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f57;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs30, %rs31, %rs32, %rs33};
$L__BB0_120:
ret;
}