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

1058 lines
28 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<89>;
.reg .b16 %rs<51>;
.reg .f32 %f<476>;
.reg .b32 %r<269>;
.reg .f64 %fd<5>;
.reg .b64 %rd<74>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r50), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r51), _optix_get_launch_index_y, ();
// end inline asm
ld.const.f32 %f1, [params+700];
ld.const.u64 %rd21, [params+400];
cvta.to.global.u64 %rd22, %rd21;
ld.const.u32 %r56, [params+392];
mad.lo.s32 %r57, %r56, %r51, %r50;
mul.wide.u32 %rd23, %r57, 4;
add.s64 %rd2, %rd22, %rd23;
ld.global.v2.u8 {%rs7, %rs50}, [%rd2];
or.b16 %rs9, %rs7, %rs50;
and.b16 %rs10, %rs9, 255;
setp.eq.s16 %p3, %rs10, 0;
@%p3 bra $L__BB0_2;
ld.global.u8 %rs49, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs49, [%rd2+2];
setp.eq.s16 %p4, %rs49, 0;
mov.f32 %f447, 0f00000000;
mov.u16 %rs50, 0;
mov.f32 %f448, %f447;
mov.f32 %f449, %f447;
@%p4 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f106, %rs7;
div.rn.f32 %f107, %f106, 0f437F0000;
fma.rn.f32 %f108, %f107, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs50, 255;
cvt.rn.f32.u16 %f109, %rs13;
div.rn.f32 %f110, %f109, 0f437F0000;
fma.rn.f32 %f111, %f110, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f112, %rs49;
div.rn.f32 %f113, %f112, 0f437F0000;
fma.rn.f32 %f114, %f113, 0f40000000, 0fBF800000;
mul.f32 %f115, %f111, %f111;
fma.rn.f32 %f116, %f108, %f108, %f115;
fma.rn.f32 %f117, %f114, %f114, %f116;
sqrt.rn.f32 %f118, %f117;
rcp.rn.f32 %f119, %f118;
mul.f32 %f449, %f119, %f114;
mul.f32 %f448, %f119, %f111;
mul.f32 %f447, %f108, %f119;
$L__BB0_4:
ld.const.v2.u32 {%r58, %r59}, [params];
add.s32 %r3, %r58, %r50;
add.s32 %r4, %r59, %r51;
setp.eq.f32 %p5, %f447, 0f00000000;
setp.eq.f32 %p6, %f448, 0f00000000;
and.pred %p7, %p5, %p6;
setp.eq.f32 %p8, %f449, 0f00000000;
and.pred %p9, %p8, %p7;
@%p9 bra $L__BB0_74;
bra.uni $L__BB0_5;
$L__BB0_74:
ld.const.u32 %r49, [params+104];
and.b32 %r250, %r49, 4;
setp.eq.s32 %p86, %r250, 0;
@%p86 bra $L__BB0_78;
ld.const.u32 %r251, [params+108];
setp.eq.s32 %p87, %r251, 0;
ld.const.u64 %rd62, [params+224];
cvta.to.global.u64 %rd63, %rd62;
ld.const.u32 %r252, [params+216];
mad.lo.s32 %r253, %r252, %r4, %r3;
mul.wide.u32 %rd64, %r253, 8;
add.s64 %rd19, %rd63, %rd64;
@%p87 bra $L__BB0_77;
ld.global.v4.u16 {%rs38, %rs39, %rs40, %rs41}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f434, %rs38;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f435, %rs39;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f436, %rs40;}
// end inline asm
add.f32 %f437, %f434, 0f00000000;
add.f32 %f438, %f435, 0f00000000;
add.f32 %f439, %f436, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f439;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f438;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f437;}
// end inline asm
mov.u16 %rs42, 0;
st.global.v4.u16 [%rd19], {%rs35, %rs36, %rs37, %rs42};
bra.uni $L__BB0_78;
$L__BB0_5:
ld.const.u64 %rd24, [params+432];
cvta.to.global.u64 %rd25, %rd24;
ld.const.u32 %r62, [params+424];
mad.lo.s32 %r63, %r62, %r51, %r50;
mul.wide.u32 %rd26, %r63, 12;
add.s64 %rd27, %rd25, %rd26;
ld.global.f32 %f121, [%rd27];
mul.f32 %f122, %f121, 0f3456BF95;
ld.global.f32 %f123, [%rd27+4];
mul.f32 %f124, %f123, 0f3456BF95;
ld.global.f32 %f125, [%rd27+8];
mul.f32 %f126, %f125, 0f3456BF95;
abs.f32 %f127, %f447;
div.rn.f32 %f128, %f122, %f127;
abs.f32 %f129, %f448;
div.rn.f32 %f130, %f124, %f129;
abs.f32 %f131, %f449;
div.rn.f32 %f132, %f126, %f131;
abs.f32 %f133, %f128;
abs.f32 %f134, %f130;
abs.f32 %f135, %f132;
mov.f32 %f136, 0f38D1B717;
max.f32 %f137, %f133, %f136;
max.f32 %f138, %f134, %f136;
max.f32 %f139, %f135, %f136;
fma.rn.f32 %f465, %f447, %f137, %f121;
fma.rn.f32 %f466, %f448, %f138, %f123;
fma.rn.f32 %f467, %f449, %f139, %f125;
setp.gt.f32 %p10, %f127, %f131;
neg.f32 %f140, %f448;
selp.f32 %f141, %f140, 0f00000000, %p10;
mov.f32 %f457, 0f00000000;
neg.f32 %f142, %f449;
selp.f32 %f143, %f447, %f142, %p10;
selp.f32 %f144, 0f00000000, %f448, %p10;
mul.f32 %f145, %f143, %f143;
fma.rn.f32 %f146, %f141, %f141, %f145;
fma.rn.f32 %f147, %f144, %f144, %f146;
sqrt.rn.f32 %f148, %f147;
rcp.rn.f32 %f149, %f148;
mul.f32 %f14, %f141, %f149;
mul.f32 %f15, %f143, %f149;
mul.f32 %f16, %f144, %f149;
ld.const.u64 %rd28, [params+128];
cvta.to.global.u64 %rd29, %rd28;
ld.const.u32 %r64, [params+120];
mad.lo.s32 %r65, %r64, %r51, %r50;
mul.wide.u32 %rd30, %r65, 4;
add.s64 %rd3, %rd29, %rd30;
ld.const.u32 %r5, [params+540];
setp.lt.s32 %p11, %r5, 1;
@%p11 bra $L__BB0_37;
cvt.rn.f32.s32 %f151, %r5;
rcp.rn.f32 %f17, %f151;
ld.global.u32 %r261, [%rd3];
ld.const.u64 %rd4, [params+96];
mul.f32 %f152, %f447, %f15;
mul.f32 %f153, %f448, %f14;
sub.f32 %f18, %f153, %f152;
mul.f32 %f154, %f449, %f14;
mul.f32 %f155, %f447, %f16;
sub.f32 %f19, %f155, %f154;
mul.f32 %f156, %f448, %f16;
mul.f32 %f157, %f449, %f15;
sub.f32 %f20, %f157, %f156;
mov.u32 %r66, 0;
add.s64 %rd5, %rd1, 24;
mov.u64 %rd31, __cudart_i2opi_f;
mov.u32 %r258, %r66;
$L__BB0_7:
mov.u32 %r260, %r66;
$L__BB0_8:
cvt.rn.f32.s32 %f444, %r258;
mad.lo.s32 %r68, %r261, 1664525, 1013904223;
and.b32 %r69, %r68, 16777215;
cvt.rn.f32.u32 %f158, %r69;
fma.rn.f32 %f159, %f158, 0f33800000, %f444;
mul.f32 %f160, %f17, %f159;
mad.lo.s32 %r261, %r68, 1664525, 1013904223;
and.b32 %r70, %r261, 16777215;
cvt.rn.f32.u32 %f161, %r70;
cvt.rn.f32.s32 %f162, %r260;
fma.rn.f32 %f163, %f161, 0f33800000, %f162;
mul.f32 %f164, %f17, %f163;
sqrt.rn.f32 %f30, %f160;
mul.f32 %f31, %f164, 0f40C90FDB;
mul.f32 %f165, %f31, 0f3F22F983;
cvt.rni.s32.f32 %r268, %f165;
cvt.rn.f32.s32 %f166, %r268;
mov.f32 %f167, 0fBFC90FDA;
fma.rn.f32 %f168, %f166, %f167, %f31;
mov.f32 %f169, 0fB3A22168;
fma.rn.f32 %f170, %f166, %f169, %f168;
mov.f32 %f171, 0fA7C234C5;
fma.rn.f32 %f461, %f166, %f171, %f170;
abs.f32 %f33, %f31;
setp.ltu.f32 %p12, %f33, 0f47CE4780;
mov.u32 %r265, %r268;
mov.f32 %f458, %f461;
@%p12 bra $L__BB0_16;
setp.eq.f32 %p13, %f33, 0f7F800000;
@%p13 bra $L__BB0_15;
bra.uni $L__BB0_10;
$L__BB0_15:
mov.f32 %f174, 0f00000000;
mul.rn.f32 %f458, %f31, %f174;
mov.u32 %r265, 0;
bra.uni $L__BB0_16;
$L__BB0_10:
mov.b32 %r13, %f31;
bfe.u32 %r72, %r13, 23, 8;
add.s32 %r14, %r72, -128;
shl.b32 %r73, %r13, 8;
or.b32 %r15, %r73, -2147483648;
shr.u32 %r16, %r14, 5;
mov.u64 %rd71, 0;
mov.u32 %r262, 0;
mov.u64 %rd69, %rd1;
mov.u64 %rd70, %rd31;
$L__BB0_11:
.pragma "nounroll";
ld.global.nc.u32 %r74, [%rd70];
mad.wide.u32 %rd33, %r74, %r15, %rd71;
shr.u64 %rd71, %rd33, 32;
st.local.u32 [%rd69], %rd33;
add.s64 %rd70, %rd70, 4;
add.s64 %rd69, %rd69, 4;
add.s32 %r262, %r262, 1;
setp.ne.s32 %p14, %r262, 6;
@%p14 bra $L__BB0_11;
st.local.u32 [%rd5], %rd71;
mov.u32 %r75, 4;
sub.s32 %r19, %r75, %r16;
mov.u32 %r76, 6;
sub.s32 %r77, %r76, %r16;
mul.wide.s32 %rd34, %r77, 4;
add.s64 %rd35, %rd1, %rd34;
ld.local.u32 %r263, [%rd35];
ld.local.u32 %r264, [%rd35+-4];
and.b32 %r22, %r14, 31;
setp.eq.s32 %p15, %r22, 0;
@%p15 bra $L__BB0_14;
mov.u32 %r78, 32;
sub.s32 %r79, %r78, %r22;
shr.u32 %r80, %r264, %r79;
shl.b32 %r81, %r263, %r22;
add.s32 %r263, %r80, %r81;
mul.wide.s32 %rd36, %r19, 4;
add.s64 %rd37, %rd1, %rd36;
ld.local.u32 %r82, [%rd37];
shr.u32 %r83, %r82, %r79;
shl.b32 %r84, %r264, %r22;
add.s32 %r264, %r83, %r84;
$L__BB0_14:
and.b32 %r85, %r13, -2147483648;
shr.u32 %r86, %r264, 30;
shl.b32 %r87, %r263, 2;
or.b32 %r88, %r86, %r87;
shr.u32 %r89, %r88, 31;
shr.u32 %r90, %r263, 30;
add.s32 %r91, %r89, %r90;
neg.s32 %r92, %r91;
setp.eq.s32 %p16, %r85, 0;
selp.b32 %r265, %r91, %r92, %p16;
setp.ne.s32 %p17, %r89, 0;
xor.b32 %r93, %r85, -2147483648;
selp.b32 %r94, %r93, %r85, %p17;
selp.b32 %r95, -1, 0, %p17;
xor.b32 %r96, %r88, %r95;
shl.b32 %r97, %r264, 2;
xor.b32 %r98, %r97, %r95;
cvt.u64.u32 %rd38, %r96;
cvt.u64.u32 %rd39, %r98;
bfi.b64 %rd40, %rd38, %rd39, 32, 32;
cvt.rn.f64.s64 %fd1, %rd40;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f172, %fd2;
setp.eq.s32 %p18, %r94, 0;
neg.f32 %f173, %f172;
selp.f32 %f458, %f172, %f173, %p18;
$L__BB0_16:
add.s32 %r29, %r265, 1;
and.b32 %r30, %r29, 1;
setp.eq.s32 %p19, %r30, 0;
selp.f32 %f37, %f458, 0f3F800000, %p19;
mul.rn.f32 %f38, %f458, %f458;
mov.f32 %f459, 0fB94D4153;
@%p19 bra $L__BB0_18;
mov.f32 %f176, 0fBAB607ED;
mov.f32 %f177, 0f37CBAC00;
fma.rn.f32 %f459, %f177, %f38, %f176;
$L__BB0_18:
selp.f32 %f178, 0f3C0885E4, 0f3D2AAABB, %p19;
fma.rn.f32 %f179, %f459, %f38, %f178;
selp.f32 %f180, 0fBE2AAAA8, 0fBEFFFFFF, %p19;
fma.rn.f32 %f181, %f179, %f38, %f180;
mov.f32 %f182, 0f00000000;
fma.rn.f32 %f183, %f38, %f37, %f182;
fma.rn.f32 %f460, %f181, %f183, %f37;
and.b32 %r100, %r29, 2;
setp.eq.s32 %p21, %r100, 0;
@%p21 bra $L__BB0_20;
mov.f32 %f185, 0fBF800000;
fma.rn.f32 %f460, %f460, %f185, %f182;
$L__BB0_20:
@%p12 bra $L__BB0_28;
setp.eq.f32 %p23, %f33, 0f7F800000;
@%p23 bra $L__BB0_27;
bra.uni $L__BB0_22;
$L__BB0_27:
mov.f32 %f188, 0f00000000;
mul.rn.f32 %f461, %f31, %f188;
mov.u32 %r268, 0;
bra.uni $L__BB0_28;
$L__BB0_22:
mov.b32 %r31, %f31;
bfe.u32 %r101, %r31, 23, 8;
add.s32 %r32, %r101, -128;
shl.b32 %r102, %r31, 8;
or.b32 %r33, %r102, -2147483648;
shr.u32 %r34, %r32, 5;
mov.u64 %rd72, 0;
mov.u64 %rd73, %rd72;
$L__BB0_23:
.pragma "nounroll";
shl.b64 %rd43, %rd72, 2;
mov.u64 %rd44, __cudart_i2opi_f;
add.s64 %rd45, %rd44, %rd43;
ld.global.nc.u32 %r103, [%rd45];
mad.wide.u32 %rd46, %r103, %r33, %rd73;
shr.u64 %rd73, %rd46, 32;
add.s64 %rd47, %rd1, %rd43;
st.local.u32 [%rd47], %rd46;
cvt.u32.u64 %r104, %rd72;
add.s32 %r105, %r104, 1;
cvt.s64.s32 %rd72, %r105;
setp.ne.s32 %p24, %r105, 6;
@%p24 bra $L__BB0_23;
st.local.u32 [%rd5], %rd73;
mov.u32 %r106, 4;
sub.s32 %r35, %r106, %r34;
mov.u32 %r107, 6;
sub.s32 %r108, %r107, %r34;
mul.wide.s32 %rd48, %r108, 4;
add.s64 %rd49, %rd1, %rd48;
ld.local.u32 %r266, [%rd49];
ld.local.u32 %r267, [%rd49+-4];
and.b32 %r38, %r32, 31;
setp.eq.s32 %p25, %r38, 0;
@%p25 bra $L__BB0_26;
mov.u32 %r109, 32;
sub.s32 %r110, %r109, %r38;
shr.u32 %r111, %r267, %r110;
shl.b32 %r112, %r266, %r38;
add.s32 %r266, %r111, %r112;
mul.wide.s32 %rd50, %r35, 4;
add.s64 %rd51, %rd1, %rd50;
ld.local.u32 %r113, [%rd51];
shr.u32 %r114, %r113, %r110;
shl.b32 %r115, %r267, %r38;
add.s32 %r267, %r114, %r115;
$L__BB0_26:
and.b32 %r116, %r31, -2147483648;
shr.u32 %r117, %r267, 30;
shl.b32 %r118, %r266, 2;
or.b32 %r119, %r117, %r118;
shr.u32 %r120, %r119, 31;
shr.u32 %r121, %r266, 30;
add.s32 %r122, %r120, %r121;
neg.s32 %r123, %r122;
setp.eq.s32 %p26, %r116, 0;
selp.b32 %r268, %r122, %r123, %p26;
setp.ne.s32 %p27, %r120, 0;
xor.b32 %r124, %r116, -2147483648;
selp.b32 %r125, %r124, %r116, %p27;
selp.b32 %r126, -1, 0, %p27;
xor.b32 %r127, %r119, %r126;
shl.b32 %r128, %r267, 2;
xor.b32 %r129, %r128, %r126;
cvt.u64.u32 %rd52, %r127;
cvt.u64.u32 %rd53, %r129;
bfi.b64 %rd54, %rd52, %rd53, 32, 32;
cvt.rn.f64.s64 %fd3, %rd54;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f186, %fd4;
setp.eq.s32 %p28, %r125, 0;
neg.f32 %f187, %f186;
selp.f32 %f461, %f186, %f187, %p28;
$L__BB0_28:
mul.f32 %f47, %f30, %f460;
and.b32 %r45, %r268, 1;
setp.eq.s32 %p29, %r45, 0;
selp.f32 %f48, %f461, 0f3F800000, %p29;
mul.rn.f32 %f49, %f461, %f461;
mov.f32 %f462, 0fB94D4153;
@%p29 bra $L__BB0_30;
mov.f32 %f190, 0fBAB607ED;
mov.f32 %f191, 0f37CBAC00;
fma.rn.f32 %f462, %f191, %f49, %f190;
$L__BB0_30:
selp.f32 %f192, 0f3C0885E4, 0f3D2AAABB, %p29;
fma.rn.f32 %f193, %f462, %f49, %f192;
selp.f32 %f194, 0fBE2AAAA8, 0fBEFFFFFF, %p29;
fma.rn.f32 %f195, %f193, %f49, %f194;
mov.f32 %f196, 0f00000000;
fma.rn.f32 %f197, %f49, %f48, %f196;
fma.rn.f32 %f463, %f195, %f197, %f48;
and.b32 %r131, %r268, 2;
setp.eq.s32 %p31, %r131, 0;
@%p31 bra $L__BB0_32;
mov.f32 %f198, 0f00000000;
mov.f32 %f199, 0fBF800000;
fma.rn.f32 %f463, %f463, %f199, %f198;
$L__BB0_32:
mov.f32 %f464, 0f00000000;
mul.f32 %f210, %f47, %f47;
mov.f32 %f211, 0f3F800000;
sub.f32 %f212, %f211, %f210;
mul.f32 %f213, %f30, %f463;
mul.f32 %f214, %f213, %f213;
sub.f32 %f215, %f212, %f214;
max.f32 %f216, %f464, %f215;
sqrt.rn.f32 %f217, %f216;
mul.f32 %f218, %f14, %f213;
mul.f32 %f219, %f15, %f213;
mul.f32 %f220, %f16, %f213;
fma.rn.f32 %f221, %f20, %f47, %f218;
fma.rn.f32 %f222, %f19, %f47, %f219;
fma.rn.f32 %f223, %f18, %f47, %f220;
fma.rn.f32 %f55, %f447, %f217, %f221;
fma.rn.f32 %f56, %f448, %f217, %f222;
fma.rn.f32 %f57, %f449, %f217, %f223;
mul.f32 %f224, %f465, 0f3456BF95;
abs.f32 %f225, %f224;
mul.f32 %f226, %f466, 0f3456BF95;
abs.f32 %f227, %f226;
mul.f32 %f228, %f467, 0f3456BF95;
abs.f32 %f229, %f228;
max.f32 %f230, %f225, %f227;
max.f32 %f231, %f230, %f229;
mov.f32 %f232, 0f38D1B717;
max.f32 %f206, %f231, %f232;
mov.u32 %r165, 1;
mov.u32 %r168, 2;
mov.u32 %r170, 3;
mov.u32 %r173, -1027080192;
mov.u32 %r202, 0;
// begin inline asm
call(%r132,%r133,%r134,%r135,%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163),_optix_trace_typed_32,(%r202,%rd4,%f465,%f466,%f467,%f55,%f56,%f57,%f206,%f1,%f464,%r165,%r202,%r202,%r168,%r202,%r170,%r173,%r173,%r173,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202);
// end inline asm
mov.b32 %f58, %r132;
mov.b32 %f59, %r133;
mov.b32 %f60, %r134;
add.f32 %f233, %f58, %f59;
add.f32 %f234, %f233, %f60;
setp.eq.f32 %p32, %f234, 0f00000000;
@%p32 bra $L__BB0_35;
setp.eq.f32 %p33, %f58, 0fC2C80000;
setp.eq.f32 %p34, %f59, 0fC2C80000;
or.pred %p35, %p33, %p34;
setp.eq.f32 %p36, %f60, 0fC2C80000;
or.pred %p37, %p36, %p35;
mov.f32 %f464, %f211;
@%p37 bra $L__BB0_35;
mov.f32 %f464, 0f00000000;
mul.f32 %f237, %f59, %f59;
fma.rn.f32 %f238, %f58, %f58, %f237;
fma.rn.f32 %f239, %f60, %f60, %f238;
sqrt.rn.f32 %f240, %f239;
add.f32 %f241, %f240, 0f3A83126F;
fma.rn.f32 %f465, %f55, %f241, %f465;
fma.rn.f32 %f466, %f56, %f241, %f466;
fma.rn.f32 %f467, %f57, %f241, %f467;
$L__BB0_35:
add.f32 %f457, %f457, %f464;
add.s32 %r260, %r260, 1;
setp.lt.s32 %p38, %r260, %r5;
@%p38 bra $L__BB0_8;
add.s32 %r258, %r258, 1;
setp.lt.s32 %p39, %r258, %r5;
@%p39 bra $L__BB0_7;
$L__BB0_37:
mul.lo.s32 %r203, %r5, %r5;
cvt.rn.f32.s32 %f242, %r203;
div.rn.f32 %f70, %f457, %f242;
ld.const.f32 %f71, [params+704];
setp.lt.f32 %p40, %f71, 0f00000000;
@%p40 bra $L__BB0_39;
bra.uni $L__BB0_38;
$L__BB0_39:
neg.f32 %f73, %f71;
setp.lt.f32 %p41, %f71, 0fBF800000;
@%p41 bra $L__BB0_41;
bra.uni $L__BB0_40;
$L__BB0_41:
mul.f32 %f248, %f71, 0fBF000000;
cvt.rzi.f32.f32 %f249, %f248;
add.f32 %f250, %f249, %f249;
sub.f32 %f251, %f73, %f250;
abs.f32 %f75, %f251;
abs.f32 %f76, %f70;
setp.lt.f32 %p42, %f76, 0f00800000;
mul.f32 %f252, %f76, 0f4B800000;
selp.f32 %f253, %f252, %f76, %p42;
selp.f32 %f254, 0fC3170000, 0fC2FE0000, %p42;
mov.b32 %r204, %f253;
and.b32 %r205, %r204, 8388607;
or.b32 %r206, %r205, 1065353216;
mov.b32 %f255, %r206;
shr.u32 %r207, %r204, 23;
cvt.rn.f32.u32 %f256, %r207;
add.f32 %f257, %f254, %f256;
setp.gt.f32 %p43, %f255, 0f3FB504F3;
mul.f32 %f258, %f255, 0f3F000000;
add.f32 %f259, %f257, 0f3F800000;
selp.f32 %f260, %f259, %f257, %p43;
selp.f32 %f261, %f258, %f255, %p43;
add.f32 %f262, %f261, 0fBF800000;
add.f32 %f263, %f261, 0f3F800000;
rcp.approx.ftz.f32 %f264, %f263;
add.f32 %f265, %f262, %f262;
mul.f32 %f266, %f265, %f264;
mul.f32 %f267, %f266, %f266;
mov.f32 %f268, 0f3C4CAF63;
mov.f32 %f269, 0f3B18F0FE;
fma.rn.f32 %f270, %f269, %f267, %f268;
mov.f32 %f271, 0f3DAAAABD;
fma.rn.f32 %f272, %f270, %f267, %f271;
mul.rn.f32 %f273, %f272, %f267;
mul.rn.f32 %f274, %f273, %f266;
sub.f32 %f275, %f262, %f266;
add.f32 %f276, %f275, %f275;
neg.f32 %f277, %f266;
fma.rn.f32 %f278, %f277, %f262, %f276;
mul.rn.f32 %f279, %f264, %f278;
add.f32 %f280, %f274, %f266;
sub.f32 %f281, %f266, %f280;
add.f32 %f282, %f274, %f281;
add.f32 %f283, %f279, %f282;
add.f32 %f284, %f280, %f283;
sub.f32 %f285, %f280, %f284;
add.f32 %f286, %f283, %f285;
mov.f32 %f287, 0f3F317200;
mul.rn.f32 %f288, %f260, %f287;
mov.f32 %f289, 0f35BFBE8E;
mul.rn.f32 %f290, %f260, %f289;
add.f32 %f291, %f288, %f284;
sub.f32 %f292, %f288, %f291;
add.f32 %f293, %f284, %f292;
add.f32 %f294, %f286, %f293;
add.f32 %f295, %f290, %f294;
add.f32 %f296, %f291, %f295;
sub.f32 %f297, %f291, %f296;
add.f32 %f298, %f295, %f297;
abs.f32 %f77, %f73;
setp.gt.f32 %p44, %f77, 0f77F684DF;
mul.f32 %f299, %f71, 0fB9000000;
selp.f32 %f300, %f299, %f73, %p44;
mul.rn.f32 %f301, %f300, %f296;
neg.f32 %f302, %f301;
fma.rn.f32 %f303, %f300, %f296, %f302;
fma.rn.f32 %f304, %f300, %f298, %f303;
mov.f32 %f305, 0f00000000;
fma.rn.f32 %f306, %f305, %f296, %f304;
add.rn.f32 %f307, %f301, %f306;
neg.f32 %f308, %f307;
add.rn.f32 %f309, %f301, %f308;
add.rn.f32 %f310, %f309, %f306;
mov.b32 %r208, %f307;
setp.eq.s32 %p45, %r208, 1118925336;
add.s32 %r209, %r208, -1;
mov.b32 %f311, %r209;
add.f32 %f312, %f310, 0f37000000;
selp.f32 %f78, %f312, %f310, %p45;
selp.f32 %f313, %f311, %f307, %p45;
mov.f32 %f314, 0f3FB8AA3B;
mul.rn.f32 %f315, %f313, %f314;
cvt.rzi.f32.f32 %f316, %f315;
abs.f32 %f317, %f316;
setp.gt.f32 %p46, %f317, 0f42FC0000;
mov.b32 %r210, %f316;
and.b32 %r211, %r210, -2147483648;
or.b32 %r212, %r211, 1123811328;
mov.b32 %f318, %r212;
selp.f32 %f319, %f318, %f316, %p46;
mov.f32 %f320, 0fBF317218;
fma.rn.f32 %f321, %f319, %f320, %f313;
mov.f32 %f322, 0f3102E308;
fma.rn.f32 %f323, %f319, %f322, %f321;
mul.f32 %f324, %f323, 0f3FB8AA3B;
add.f32 %f325, %f319, 0f4B40007F;
mov.b32 %r213, %f325;
shl.b32 %r214, %r213, 23;
mov.b32 %f326, %r214;
ex2.approx.ftz.f32 %f327, %f324;
mul.f32 %f79, %f327, %f326;
setp.eq.f32 %p47, %f79, 0f7F800000;
mov.f32 %f469, 0f7F800000;
@%p47 bra $L__BB0_43;
fma.rn.f32 %f469, %f79, %f78, %f79;
$L__BB0_43:
setp.lt.f32 %p48, %f70, 0f00000000;
setp.eq.f32 %p49, %f75, 0f3F800000;
and.pred %p1, %p48, %p49;
setp.eq.f32 %p50, %f70, 0f00000000;
@%p50 bra $L__BB0_47;
bra.uni $L__BB0_44;
$L__BB0_47:
add.f32 %f331, %f70, %f70;
mov.b32 %r217, %f331;
selp.b32 %r218, %r217, 0, %p49;
or.b32 %r219, %r218, 2139095040;
setp.gt.f32 %p54, %f71, 0f80000000;
selp.b32 %r220, %r219, %r218, %p54;
mov.b32 %f471, %r220;
bra.uni $L__BB0_48;
$L__BB0_38:
add.f32 %f243, %f70, 0fBF800000;
fma.rn.f32 %f244, %f71, %f243, 0f3F800000;
cvt.sat.f32.f32 %f472, %f244;
bra.uni $L__BB0_56;
$L__BB0_40:
add.f32 %f245, %f70, 0fBF800000;
fma.rn.f32 %f246, %f245, %f73, 0f3F800000;
cvt.sat.f32.f32 %f472, %f246;
bra.uni $L__BB0_56;
$L__BB0_77:
mov.f32 %f442, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f442;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f442;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f442;}
// end inline asm
mov.u16 %rs46, 0;
st.global.v4.u16 [%rd19], {%rs43, %rs44, %rs45, %rs46};
$L__BB0_78:
and.b32 %r254, %r49, 8;
setp.eq.s32 %p88, %r254, 0;
@%p88 bra $L__BB0_80;
ld.const.u64 %rd65, [params+192];
cvta.to.global.u64 %rd66, %rd65;
ld.const.u32 %r255, [params+184];
mad.lo.s32 %r256, %r255, %r4, %r3;
mov.f32 %f443, 0f00000000;
cvt.rzi.u32.f32 %r257, %f443;
mul.wide.u32 %rd67, %r256, 2;
add.s64 %rd68, %rd66, %rd67;
mov.u16 %rs47, 0;
cvt.u16.u32 %rs48, %r257;
st.global.v2.u8 [%rd68], {%rs48, %rs47};
bra.uni $L__BB0_80;
$L__BB0_44:
mov.b32 %r215, %f469;
xor.b32 %r216, %r215, -2147483648;
mov.b32 %f328, %r216;
selp.f32 %f471, %f328, %f469, %p1;
setp.geu.f32 %p51, %f70, 0f00000000;
@%p51 bra $L__BB0_48;
cvt.rzi.f32.f32 %f329, %f73;
setp.eq.f32 %p52, %f329, %f73;
@%p52 bra $L__BB0_48;
mov.f32 %f471, 0f7FFFFFFF;
$L__BB0_48:
add.f32 %f332, %f76, %f77;
mov.b32 %r221, %f332;
setp.lt.s32 %p55, %r221, 2139095040;
@%p55 bra $L__BB0_55;
setp.gtu.f32 %p56, %f76, 0f7F800000;
setp.gtu.f32 %p57, %f77, 0f7F800000;
or.pred %p58, %p56, %p57;
@%p58 bra $L__BB0_54;
bra.uni $L__BB0_50;
$L__BB0_54:
sub.f32 %f471, %f70, %f71;
bra.uni $L__BB0_55;
$L__BB0_50:
setp.eq.f32 %p59, %f77, 0f7F800000;
@%p59 bra $L__BB0_53;
bra.uni $L__BB0_51;
$L__BB0_53:
setp.gt.f32 %p62, %f76, 0f3F800000;
selp.b32 %r225, 2139095040, 0, %p62;
xor.b32 %r226, %r225, 2139095040;
setp.gt.f32 %p63, %f71, 0f80000000;
selp.b32 %r227, %r226, %r225, %p63;
mov.b32 %f333, %r227;
setp.eq.f32 %p64, %f70, 0fBF800000;
selp.f32 %f471, 0f3F800000, %f333, %p64;
bra.uni $L__BB0_55;
$L__BB0_51:
setp.neu.f32 %p60, %f76, 0f7F800000;
@%p60 bra $L__BB0_55;
setp.le.f32 %p61, %f71, 0f80000000;
selp.b32 %r222, 2139095040, 0, %p61;
or.b32 %r223, %r222, -2147483648;
selp.b32 %r224, %r223, %r222, %p1;
mov.b32 %f471, %r224;
$L__BB0_55:
setp.eq.f32 %p65, %f71, 0f80000000;
setp.eq.f32 %p66, %f70, 0f3F800000;
or.pred %p67, %p65, %p66;
selp.f32 %f334, 0f3F800000, %f471, %p67;
cvt.sat.f32.f32 %f472, %f334;
$L__BB0_56:
ld.const.u32 %r48, [params+104];
and.b32 %r228, %r48, 4;
setp.eq.s32 %p68, %r228, 0;
@%p68 bra $L__BB0_60;
ld.const.u32 %r229, [params+108];
setp.eq.s32 %p69, %r229, 0;
ld.const.u64 %rd56, [params+224];
cvta.to.global.u64 %rd57, %rd56;
ld.const.u32 %r230, [params+216];
mad.lo.s32 %r231, %r230, %r4, %r3;
mul.wide.u32 %rd58, %r231, 8;
add.s64 %rd16, %rd57, %rd58;
@%p69 bra $L__BB0_59;
ld.global.v4.u16 {%rs22, %rs23, %rs24, %rs25}, [%rd16];
// begin inline asm
{ cvt.f32.f16 %f335, %rs22;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f336, %rs23;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f337, %rs24;}
// end inline asm
add.f32 %f338, %f472, %f335;
add.f32 %f339, %f472, %f336;
add.f32 %f340, %f472, %f337;
mov.f32 %f341, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs20, %f340;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs19, %f339;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs18, %f338;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs21, %f341;}
// end inline asm
st.global.v4.u16 [%rd16], {%rs18, %rs19, %rs20, %rs21};
bra.uni $L__BB0_60;
$L__BB0_59:
mov.f32 %f345, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f345;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f472;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f472;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs26, %f472;}
// end inline asm
st.global.v4.u16 [%rd16], {%rs26, %rs27, %rs28, %rs29};
$L__BB0_60:
and.b32 %r232, %r48, 8;
setp.eq.s32 %p70, %r232, 0;
@%p70 bra $L__BB0_80;
ld.const.u64 %rd59, [params+192];
cvta.to.global.u64 %rd17, %rd59;
ld.const.u32 %r233, [params+184];
mad.lo.s32 %r234, %r233, %r4, %r3;
cvt.u64.u32 %rd18, %r234;
mov.f32 %f347, 0f3E68BA2E;
cvt.rzi.f32.f32 %f348, %f347;
add.f32 %f349, %f348, %f348;
mov.f32 %f350, 0f3EE8BA2E;
sub.f32 %f351, %f350, %f349;
abs.f32 %f91, %f351;
abs.f32 %f92, %f472;
setp.lt.f32 %p71, %f92, 0f00800000;
mul.f32 %f352, %f92, 0f4B800000;
selp.f32 %f353, %f352, %f92, %p71;
selp.f32 %f354, 0fC3170000, 0fC2FE0000, %p71;
mov.b32 %r235, %f353;
and.b32 %r236, %r235, 8388607;
or.b32 %r237, %r236, 1065353216;
mov.b32 %f355, %r237;
shr.u32 %r238, %r235, 23;
cvt.rn.f32.u32 %f356, %r238;
add.f32 %f357, %f354, %f356;
setp.gt.f32 %p72, %f355, 0f3FB504F3;
mul.f32 %f358, %f355, 0f3F000000;
add.f32 %f359, %f357, 0f3F800000;
selp.f32 %f360, %f359, %f357, %p72;
selp.f32 %f361, %f358, %f355, %p72;
add.f32 %f362, %f361, 0fBF800000;
add.f32 %f363, %f361, 0f3F800000;
rcp.approx.ftz.f32 %f364, %f363;
add.f32 %f365, %f362, %f362;
mul.f32 %f366, %f365, %f364;
mul.f32 %f367, %f366, %f366;
mov.f32 %f368, 0f3C4CAF63;
mov.f32 %f369, 0f3B18F0FE;
fma.rn.f32 %f370, %f369, %f367, %f368;
mov.f32 %f371, 0f3DAAAABD;
fma.rn.f32 %f372, %f370, %f367, %f371;
mul.rn.f32 %f373, %f372, %f367;
mul.rn.f32 %f374, %f373, %f366;
sub.f32 %f375, %f362, %f366;
add.f32 %f376, %f375, %f375;
neg.f32 %f377, %f366;
fma.rn.f32 %f378, %f377, %f362, %f376;
mul.rn.f32 %f379, %f364, %f378;
add.f32 %f380, %f374, %f366;
sub.f32 %f381, %f366, %f380;
add.f32 %f382, %f374, %f381;
add.f32 %f383, %f379, %f382;
add.f32 %f384, %f380, %f383;
sub.f32 %f385, %f380, %f384;
add.f32 %f386, %f383, %f385;
mov.f32 %f387, 0f3F317200;
mul.rn.f32 %f388, %f360, %f387;
mov.f32 %f389, 0f35BFBE8E;
mul.rn.f32 %f390, %f360, %f389;
add.f32 %f391, %f388, %f384;
sub.f32 %f392, %f388, %f391;
add.f32 %f393, %f384, %f392;
add.f32 %f394, %f386, %f393;
add.f32 %f395, %f390, %f394;
add.f32 %f396, %f391, %f395;
sub.f32 %f397, %f391, %f396;
add.f32 %f398, %f395, %f397;
mul.rn.f32 %f399, %f350, %f396;
neg.f32 %f400, %f399;
fma.rn.f32 %f401, %f350, %f396, %f400;
fma.rn.f32 %f402, %f350, %f398, %f401;
mov.f32 %f403, 0f00000000;
fma.rn.f32 %f404, %f403, %f396, %f402;
add.rn.f32 %f405, %f399, %f404;
neg.f32 %f406, %f405;
add.rn.f32 %f407, %f399, %f406;
add.rn.f32 %f408, %f407, %f404;
mov.b32 %r239, %f405;
setp.eq.s32 %p73, %r239, 1118925336;
add.s32 %r240, %r239, -1;
mov.b32 %f409, %r240;
add.f32 %f410, %f408, 0f37000000;
selp.f32 %f93, %f410, %f408, %p73;
selp.f32 %f411, %f409, %f405, %p73;
mov.f32 %f412, 0f3FB8AA3B;
mul.rn.f32 %f413, %f411, %f412;
cvt.rzi.f32.f32 %f414, %f413;
abs.f32 %f415, %f414;
setp.gt.f32 %p74, %f415, 0f42FC0000;
mov.b32 %r241, %f414;
and.b32 %r242, %r241, -2147483648;
or.b32 %r243, %r242, 1123811328;
mov.b32 %f416, %r243;
selp.f32 %f417, %f416, %f414, %p74;
mov.f32 %f418, 0fBF317218;
fma.rn.f32 %f419, %f417, %f418, %f411;
mov.f32 %f420, 0f3102E308;
fma.rn.f32 %f421, %f417, %f420, %f419;
mul.f32 %f422, %f421, 0f3FB8AA3B;
add.f32 %f423, %f417, 0f4B40007F;
mov.b32 %r244, %f423;
shl.b32 %r245, %r244, 23;
mov.b32 %f424, %r245;
ex2.approx.ftz.f32 %f425, %f422;
mul.f32 %f94, %f425, %f424;
setp.eq.f32 %p75, %f94, 0f7F800000;
mov.f32 %f473, 0f7F800000;
@%p75 bra $L__BB0_63;
fma.rn.f32 %f473, %f94, %f93, %f94;
$L__BB0_63:
setp.lt.f32 %p76, %f472, 0f00000000;
setp.eq.f32 %p77, %f91, 0f3F800000;
and.pred %p2, %p76, %p77;
setp.eq.f32 %p78, %f472, 0f00000000;
@%p78 bra $L__BB0_67;
bra.uni $L__BB0_64;
$L__BB0_67:
add.f32 %f430, %f472, %f472;
selp.f32 %f475, %f430, 0f00000000, %p77;
bra.uni $L__BB0_68;
$L__BB0_64:
mov.b32 %r246, %f473;
xor.b32 %r247, %r246, -2147483648;
mov.b32 %f426, %r247;
selp.f32 %f475, %f426, %f473, %p2;
setp.geu.f32 %p79, %f472, 0f00000000;
@%p79 bra $L__BB0_68;
mov.f32 %f427, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f428, %f427;
setp.eq.f32 %p80, %f428, 0f3EE8BA2E;
@%p80 bra $L__BB0_68;
mov.f32 %f475, 0f7FFFFFFF;
$L__BB0_68:
add.f32 %f431, %f92, 0f3EE8BA2E;
mov.b32 %r248, %f431;
setp.lt.s32 %p82, %r248, 2139095040;
@%p82 bra $L__BB0_73;
setp.gtu.f32 %p83, %f92, 0f7F800000;
@%p83 bra $L__BB0_72;
bra.uni $L__BB0_70;
$L__BB0_72:
add.f32 %f475, %f472, 0f3EE8BA2E;
bra.uni $L__BB0_73;
$L__BB0_70:
setp.neu.f32 %p84, %f92, 0f7F800000;
@%p84 bra $L__BB0_73;
selp.f32 %f475, 0fFF800000, 0f7F800000, %p2;
$L__BB0_73:
mul.f32 %f432, %f475, 0f437F0000;
setp.eq.f32 %p85, %f472, 0f3F800000;
selp.f32 %f433, 0f437F0000, %f432, %p85;
cvt.rzi.u32.f32 %r249, %f433;
shl.b64 %rd60, %rd18, 1;
add.s64 %rd61, %rd17, %rd60;
cvt.u16.u32 %rs30, %r249;
mov.u16 %rs31, 255;
st.global.v2.u8 [%rd61], {%rs30, %rs31};
$L__BB0_80:
ret;
}