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

1738 lines
48 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<148>;
.reg .b16 %rs<57>;
.reg .f32 %f<986>;
.reg .b32 %r<333>;
.reg .f64 %fd<5>;
.reg .b64 %rd<85>;
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, %rs56}, [%rd2];
or.b16 %rs9, %rs7, %rs56;
and.b16 %rs10, %rs9, 255;
setp.eq.s16 %p7, %rs10, 0;
@%p7 bra $L__BB0_2;
ld.global.u8 %rs55, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs55, [%rd2+2];
setp.eq.s16 %p8, %rs55, 0;
mov.f32 %f935, 0f00000000;
mov.u16 %rs56, 0;
mov.f32 %f936, %f935;
mov.f32 %f937, %f935;
@%p8 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f162, %rs7;
div.rn.f32 %f163, %f162, 0f437F0000;
fma.rn.f32 %f164, %f163, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs56, 255;
cvt.rn.f32.u16 %f165, %rs13;
div.rn.f32 %f166, %f165, 0f437F0000;
fma.rn.f32 %f167, %f166, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f168, %rs55;
div.rn.f32 %f169, %f168, 0f437F0000;
fma.rn.f32 %f170, %f169, 0f40000000, 0fBF800000;
mul.f32 %f171, %f167, %f167;
fma.rn.f32 %f172, %f164, %f164, %f171;
fma.rn.f32 %f173, %f170, %f170, %f172;
sqrt.rn.f32 %f174, %f173;
rcp.rn.f32 %f175, %f174;
mul.f32 %f937, %f175, %f170;
mul.f32 %f936, %f175, %f167;
mul.f32 %f935, %f164, %f175;
$L__BB0_4:
ld.const.v2.u32 {%r59, %r60}, [params];
add.s32 %r4, %r59, %r51;
add.s32 %r5, %r60, %r52;
setp.eq.f32 %p9, %f935, 0f00000000;
setp.eq.f32 %p10, %f936, 0f00000000;
and.pred %p11, %p9, %p10;
setp.eq.f32 %p12, %f937, 0f00000000;
and.pred %p13, %p12, %p11;
@%p13 bra $L__BB0_116;
bra.uni $L__BB0_5;
$L__BB0_116:
ld.const.u32 %r50, [params+104];
and.b32 %r308, %r50, 1;
setp.eq.b32 %p141, %r308, 1;
mov.pred %p142, 0;
xor.pred %p143, %p141, %p142;
not.pred %p144, %p143;
@%p144 bra $L__BB0_118;
ld.const.u64 %rd66, [params+144];
cvta.to.global.u64 %rd67, %rd66;
ld.const.u32 %r309, [params+136];
mad.lo.s32 %r310, %r309, %r5, %r4;
mul.wide.u32 %rd68, %r310, 4;
add.s64 %rd69, %rd67, %rd68;
mov.u16 %rs38, 0;
st.global.v4.u8 [%rd69], {%rs38, %rs38, %rs38, %rs38};
$L__BB0_118:
and.b32 %r311, %r50, 4;
setp.eq.s32 %p145, %r311, 0;
@%p145 bra $L__BB0_122;
ld.const.u32 %r312, [params+108];
setp.eq.s32 %p146, %r312, 0;
ld.const.u64 %rd70, [params+224];
cvta.to.global.u64 %rd71, %rd70;
ld.const.u32 %r313, [params+216];
mad.lo.s32 %r314, %r313, %r5, %r4;
mul.wide.u32 %rd72, %r314, 8;
add.s64 %rd19, %rd71, %rd72;
@%p146 bra $L__BB0_121;
ld.global.v4.u16 {%rs45, %rs46, %rs47, %rs48}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f867, %rs45;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f868, %rs46;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f869, %rs47;}
// end inline asm
add.f32 %f870, %f867, 0f00000000;
add.f32 %f871, %f868, 0f00000000;
add.f32 %f872, %f869, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f872;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f871;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs42, %f870;}
// end inline asm
mov.u16 %rs49, 0;
st.global.v4.u16 [%rd19], {%rs42, %rs43, %rs44, %rs49};
bra.uni $L__BB0_122;
$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 %f182, [%rd27];
mul.f32 %f183, %f182, 0f3456BF95;
ld.global.f32 %f184, [%rd27+4];
mul.f32 %f185, %f184, 0f3456BF95;
ld.global.f32 %f186, [%rd27+8];
mul.f32 %f187, %f186, 0f3456BF95;
abs.f32 %f188, %f935;
div.rn.f32 %f189, %f183, %f188;
abs.f32 %f190, %f936;
div.rn.f32 %f191, %f185, %f190;
abs.f32 %f192, %f937;
div.rn.f32 %f193, %f187, %f192;
abs.f32 %f194, %f189;
abs.f32 %f195, %f191;
abs.f32 %f196, %f193;
mov.f32 %f197, 0f38D1B717;
max.f32 %f198, %f194, %f197;
max.f32 %f199, %f195, %f197;
max.f32 %f200, %f196, %f197;
fma.rn.f32 %f10, %f935, %f198, %f182;
fma.rn.f32 %f11, %f936, %f199, %f184;
fma.rn.f32 %f12, %f937, %f200, %f186;
setp.gt.f32 %p14, %f188, %f192;
neg.f32 %f201, %f936;
selp.f32 %f202, %f201, 0f00000000, %p14;
mov.f32 %f956, 0f00000000;
neg.f32 %f203, %f937;
selp.f32 %f204, %f935, %f203, %p14;
selp.f32 %f205, 0f00000000, %f936, %p14;
mul.f32 %f206, %f204, %f204;
fma.rn.f32 %f207, %f202, %f202, %f206;
fma.rn.f32 %f208, %f205, %f205, %f207;
sqrt.rn.f32 %f209, %f208;
rcp.rn.f32 %f210, %f209;
mul.f32 %f13, %f202, %f210;
mul.f32 %f14, %f204, %f210;
mul.f32 %f15, %f205, %f210;
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;
mov.f32 %f957, %f956;
mov.f32 %f958, %f956;
mov.f32 %f959, %f956;
mov.f32 %f960, %f956;
mov.f32 %f961, %f956;
@%p15 bra $L__BB0_36;
ld.const.u32 %r6, [params+588];
cvt.rn.f32.s32 %f217, %r3;
rcp.rn.f32 %f16, %f217;
ld.global.u32 %r325, [%rd3];
mul.f32 %f17, %f10, 0f3456BF95;
mul.f32 %f18, %f11, 0f3456BF95;
mul.f32 %f19, %f12, 0f3456BF95;
ld.const.u64 %rd4, [params+96];
mul.f32 %f218, %f935, %f14;
mul.f32 %f219, %f936, %f13;
sub.f32 %f20, %f219, %f218;
mul.f32 %f220, %f937, %f13;
mul.f32 %f221, %f935, %f15;
sub.f32 %f21, %f221, %f220;
mul.f32 %f222, %f936, %f15;
mul.f32 %f223, %f937, %f14;
sub.f32 %f22, %f223, %f222;
mov.u32 %r67, 0;
add.s64 %rd5, %rd1, 24;
mov.u64 %rd31, __cudart_i2opi_f;
abs.f32 %f286, %f17;
abs.f32 %f287, %f18;
max.f32 %f288, %f286, %f287;
abs.f32 %f289, %f19;
max.f32 %f290, %f288, %f289;
mov.u32 %r322, %r67;
$L__BB0_7:
mov.u32 %r324, %r67;
$L__BB0_8:
cvt.rn.f32.s32 %f882, %r322;
mad.lo.s32 %r69, %r325, 1664525, 1013904223;
and.b32 %r70, %r69, 16777215;
cvt.rn.f32.u32 %f224, %r70;
fma.rn.f32 %f225, %f224, 0f33800000, %f882;
mul.f32 %f36, %f16, %f225;
mad.lo.s32 %r325, %r69, 1664525, 1013904223;
and.b32 %r71, %r325, 16777215;
cvt.rn.f32.u32 %f226, %r71;
cvt.rn.f32.s32 %f227, %r324;
fma.rn.f32 %f228, %f226, 0f33800000, %f227;
mul.f32 %f229, %f16, %f228;
mul.f32 %f230, %f36, %f36;
mov.f32 %f231, 0f3F800000;
sub.f32 %f232, %f231, %f230;
mov.f32 %f233, 0f00000000;
max.f32 %f234, %f233, %f232;
sqrt.rn.f32 %f37, %f234;
mul.f32 %f38, %f229, 0f40C90FDB;
mul.f32 %f235, %f38, 0f3F22F983;
cvt.rni.s32.f32 %r332, %f235;
cvt.rn.f32.s32 %f236, %r332;
mov.f32 %f237, 0fBFC90FDA;
fma.rn.f32 %f238, %f236, %f237, %f38;
mov.f32 %f239, 0fB3A22168;
fma.rn.f32 %f240, %f236, %f239, %f238;
mov.f32 %f241, 0fA7C234C5;
fma.rn.f32 %f953, %f236, %f241, %f240;
abs.f32 %f40, %f38;
setp.ltu.f32 %p16, %f40, 0f47CE4780;
mov.u32 %r329, %r332;
mov.f32 %f950, %f953;
@%p16 bra $L__BB0_16;
setp.eq.f32 %p17, %f40, 0f7F800000;
@%p17 bra $L__BB0_15;
bra.uni $L__BB0_10;
$L__BB0_15:
mov.f32 %f244, 0f00000000;
mul.rn.f32 %f950, %f38, %f244;
mov.u32 %r329, 0;
bra.uni $L__BB0_16;
$L__BB0_10:
mov.b32 %r14, %f38;
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 %rd82, 0;
mov.u32 %r326, 0;
mov.u64 %rd80, %rd1;
mov.u64 %rd81, %rd31;
$L__BB0_11:
.pragma "nounroll";
ld.global.nc.u32 %r75, [%rd81];
mad.wide.u32 %rd33, %r75, %r16, %rd82;
shr.u64 %rd82, %rd33, 32;
st.local.u32 [%rd80], %rd33;
add.s64 %rd81, %rd81, 4;
add.s64 %rd80, %rd80, 4;
add.s32 %r326, %r326, 1;
setp.ne.s32 %p18, %r326, 6;
@%p18 bra $L__BB0_11;
st.local.u32 [%rd5], %rd82;
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 %r327, [%rd35];
ld.local.u32 %r328, [%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, %r328, %r80;
shl.b32 %r82, %r327, %r23;
add.s32 %r327, %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, %r328, %r23;
add.s32 %r328, %r84, %r85;
$L__BB0_14:
and.b32 %r86, %r14, -2147483648;
shr.u32 %r87, %r328, 30;
shl.b32 %r88, %r327, 2;
or.b32 %r89, %r87, %r88;
shr.u32 %r90, %r89, 31;
shr.u32 %r91, %r327, 30;
add.s32 %r92, %r90, %r91;
neg.s32 %r93, %r92;
setp.eq.s32 %p20, %r86, 0;
selp.b32 %r329, %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, %r328, 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 %f242, %fd2;
setp.eq.s32 %p22, %r95, 0;
neg.f32 %f243, %f242;
selp.f32 %f950, %f242, %f243, %p22;
$L__BB0_16:
add.s32 %r30, %r329, 1;
and.b32 %r31, %r30, 1;
setp.eq.s32 %p23, %r31, 0;
selp.f32 %f44, %f950, 0f3F800000, %p23;
mul.rn.f32 %f45, %f950, %f950;
mov.f32 %f951, 0fB94D4153;
@%p23 bra $L__BB0_18;
mov.f32 %f246, 0fBAB607ED;
mov.f32 %f247, 0f37CBAC00;
fma.rn.f32 %f951, %f247, %f45, %f246;
$L__BB0_18:
selp.f32 %f248, 0f3C0885E4, 0f3D2AAABB, %p23;
fma.rn.f32 %f249, %f951, %f45, %f248;
selp.f32 %f250, 0fBE2AAAA8, 0fBEFFFFFF, %p23;
fma.rn.f32 %f251, %f249, %f45, %f250;
mov.f32 %f252, 0f00000000;
fma.rn.f32 %f253, %f45, %f44, %f252;
fma.rn.f32 %f952, %f251, %f253, %f44;
and.b32 %r101, %r30, 2;
setp.eq.s32 %p25, %r101, 0;
@%p25 bra $L__BB0_20;
mov.f32 %f255, 0fBF800000;
fma.rn.f32 %f952, %f952, %f255, %f252;
$L__BB0_20:
@%p16 bra $L__BB0_28;
setp.eq.f32 %p27, %f40, 0f7F800000;
@%p27 bra $L__BB0_27;
bra.uni $L__BB0_22;
$L__BB0_27:
mov.f32 %f258, 0f00000000;
mul.rn.f32 %f953, %f38, %f258;
mov.u32 %r332, 0;
bra.uni $L__BB0_28;
$L__BB0_22:
mov.b32 %r32, %f38;
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 %rd83, 0;
mov.u64 %rd84, %rd83;
$L__BB0_23:
.pragma "nounroll";
shl.b64 %rd43, %rd83, 2;
mov.u64 %rd44, __cudart_i2opi_f;
add.s64 %rd45, %rd44, %rd43;
ld.global.nc.u32 %r104, [%rd45];
mad.wide.u32 %rd46, %r104, %r34, %rd84;
shr.u64 %rd84, %rd46, 32;
add.s64 %rd47, %rd1, %rd43;
st.local.u32 [%rd47], %rd46;
cvt.u32.u64 %r105, %rd83;
add.s32 %r106, %r105, 1;
cvt.s64.s32 %rd83, %r106;
setp.ne.s32 %p28, %r106, 6;
@%p28 bra $L__BB0_23;
st.local.u32 [%rd5], %rd84;
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 %r330, [%rd49];
ld.local.u32 %r331, [%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, %r331, %r111;
shl.b32 %r113, %r330, %r39;
add.s32 %r330, %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, %r331, %r39;
add.s32 %r331, %r115, %r116;
$L__BB0_26:
and.b32 %r117, %r32, -2147483648;
shr.u32 %r118, %r331, 30;
shl.b32 %r119, %r330, 2;
or.b32 %r120, %r118, %r119;
shr.u32 %r121, %r120, 31;
shr.u32 %r122, %r330, 30;
add.s32 %r123, %r121, %r122;
neg.s32 %r124, %r123;
setp.eq.s32 %p30, %r117, 0;
selp.b32 %r332, %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, %r331, 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 %f256, %fd4;
setp.eq.s32 %p32, %r126, 0;
neg.f32 %f257, %f256;
selp.f32 %f953, %f256, %f257, %p32;
$L__BB0_28:
mul.f32 %f54, %f37, %f952;
and.b32 %r46, %r332, 1;
setp.eq.s32 %p33, %r46, 0;
selp.f32 %f55, %f953, 0f3F800000, %p33;
mul.rn.f32 %f56, %f953, %f953;
mov.f32 %f954, 0fB94D4153;
@%p33 bra $L__BB0_30;
mov.f32 %f260, 0fBAB607ED;
mov.f32 %f261, 0f37CBAC00;
fma.rn.f32 %f954, %f261, %f56, %f260;
$L__BB0_30:
selp.f32 %f262, 0f3C0885E4, 0f3D2AAABB, %p33;
fma.rn.f32 %f263, %f954, %f56, %f262;
selp.f32 %f264, 0fBE2AAAA8, 0fBEFFFFFF, %p33;
fma.rn.f32 %f265, %f263, %f56, %f264;
mov.f32 %f266, 0f00000000;
fma.rn.f32 %f267, %f56, %f55, %f266;
fma.rn.f32 %f955, %f265, %f267, %f55;
and.b32 %r132, %r332, 2;
setp.eq.s32 %p35, %r132, 0;
@%p35 bra $L__BB0_32;
mov.f32 %f269, 0fBF800000;
fma.rn.f32 %f955, %f955, %f269, %f266;
$L__BB0_32:
mul.f32 %f270, %f37, %f955;
mul.f32 %f271, %f13, %f270;
mul.f32 %f272, %f14, %f270;
mul.f32 %f273, %f15, %f270;
fma.rn.f32 %f274, %f22, %f54, %f271;
fma.rn.f32 %f275, %f21, %f54, %f272;
fma.rn.f32 %f276, %f20, %f54, %f273;
fma.rn.f32 %f62, %f935, %f36, %f274;
fma.rn.f32 %f63, %f936, %f36, %f275;
fma.rn.f32 %f64, %f937, %f36, %f276;
setp.leu.f32 %p36, %f63, 0f00000000;
setp.ne.s32 %p37, %r6, 0;
and.pred %p38, %p37, %p36;
@%p38 bra $L__BB0_34;
mov.f32 %f291, 0f38D1B717;
max.f32 %f283, %f290, %f291;
mov.f32 %f284, 0f6C4ECB8F;
mov.f32 %f285, 0f00000000;
mov.u32 %r169, 2;
mov.u32 %r170, 1;
mov.u32 %r171, 3;
mov.u32 %r174, 1065353216;
mov.u32 %r203, 0;
// begin inline asm
call(%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,%f62,%f63,%f64,%f283,%f284,%f285,%r170,%r203,%r170,%r169,%r170,%r171,%r174,%r174,%r174,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203);
// end inline asm
mov.b32 %f292, %r133;
mov.b32 %f293, %r134;
mov.b32 %f294, %r135;
fma.rn.f32 %f956, %f62, %f292, %f956;
fma.rn.f32 %f957, %f63, %f293, %f957;
fma.rn.f32 %f958, %f64, %f294, %f958;
mul.f32 %f295, %f936, %f63;
fma.rn.f32 %f296, %f935, %f62, %f295;
fma.rn.f32 %f297, %f937, %f64, %f296;
cvt.sat.f32.f32 %f298, %f297;
fma.rn.f32 %f959, %f298, %f292, %f959;
fma.rn.f32 %f960, %f298, %f293, %f960;
fma.rn.f32 %f961, %f298, %f294, %f961;
$L__BB0_34:
add.s32 %r324, %r324, 1;
setp.lt.s32 %p39, %r324, %r3;
@%p39 bra $L__BB0_8;
add.s32 %r322, %r322, 1;
setp.lt.s32 %p40, %r322, %r3;
@%p40 bra $L__BB0_7;
$L__BB0_36:
mul.lo.s32 %r204, %r3, %r3;
cvt.rn.f32.s32 %f299, %r204;
rcp.rn.f32 %f300, %f299;
mul.f32 %f301, %f300, %f959;
mul.f32 %f302, %f300, %f960;
mul.f32 %f303, %f300, %f961;
fma.rn.f32 %f304, %f300, %f959, %f301;
fma.rn.f32 %f305, %f300, %f960, %f302;
fma.rn.f32 %f306, %f300, %f961, %f303;
ld.const.v4.f32 {%f307, %f308, %f309, %f310}, [params+576];
mul.f32 %f83, %f304, %f307;
mul.f32 %f84, %f305, %f308;
mul.f32 %f85, %f306, %f309;
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 %f318, 0f3EE66666;
abs.f32 %f87, %f83;
setp.lt.f32 %p45, %f87, 0f00800000;
mul.f32 %f320, %f87, 0f4B800000;
selp.f32 %f321, %f320, %f87, %p45;
selp.f32 %f322, 0fC3170000, 0fC2FE0000, %p45;
mov.b32 %r206, %f321;
and.b32 %r207, %r206, 8388607;
or.b32 %r208, %r207, 1065353216;
mov.b32 %f323, %r208;
shr.u32 %r209, %r206, 23;
cvt.rn.f32.u32 %f324, %r209;
add.f32 %f325, %f322, %f324;
setp.gt.f32 %p46, %f323, 0f3FB504F3;
mul.f32 %f326, %f323, 0f3F000000;
add.f32 %f327, %f325, 0f3F800000;
selp.f32 %f328, %f327, %f325, %p46;
selp.f32 %f329, %f326, %f323, %p46;
add.f32 %f330, %f329, 0fBF800000;
add.f32 %f331, %f329, 0f3F800000;
rcp.approx.ftz.f32 %f332, %f331;
add.f32 %f333, %f330, %f330;
mul.f32 %f334, %f333, %f332;
mul.f32 %f335, %f334, %f334;
mov.f32 %f336, 0f3C4CAF63;
mov.f32 %f337, 0f3B18F0FE;
fma.rn.f32 %f338, %f337, %f335, %f336;
mov.f32 %f339, 0f3DAAAABD;
fma.rn.f32 %f340, %f338, %f335, %f339;
mul.rn.f32 %f341, %f340, %f335;
mul.rn.f32 %f342, %f341, %f334;
sub.f32 %f343, %f330, %f334;
add.f32 %f344, %f343, %f343;
neg.f32 %f345, %f334;
fma.rn.f32 %f346, %f345, %f330, %f344;
mul.rn.f32 %f347, %f332, %f346;
add.f32 %f348, %f342, %f334;
sub.f32 %f349, %f334, %f348;
add.f32 %f350, %f342, %f349;
add.f32 %f351, %f347, %f350;
add.f32 %f352, %f348, %f351;
sub.f32 %f353, %f348, %f352;
add.f32 %f354, %f351, %f353;
mov.f32 %f355, 0f3F317200;
mul.rn.f32 %f356, %f328, %f355;
mov.f32 %f357, 0f35BFBE8E;
mul.rn.f32 %f358, %f328, %f357;
add.f32 %f359, %f356, %f352;
sub.f32 %f360, %f356, %f359;
add.f32 %f361, %f352, %f360;
add.f32 %f362, %f354, %f361;
add.f32 %f363, %f358, %f362;
add.f32 %f364, %f359, %f363;
sub.f32 %f365, %f359, %f364;
add.f32 %f366, %f363, %f365;
mul.rn.f32 %f367, %f318, %f364;
neg.f32 %f368, %f367;
fma.rn.f32 %f369, %f318, %f364, %f368;
fma.rn.f32 %f370, %f318, %f366, %f369;
mov.f32 %f371, 0f00000000;
fma.rn.f32 %f372, %f371, %f364, %f370;
add.rn.f32 %f373, %f367, %f372;
neg.f32 %f374, %f373;
add.rn.f32 %f375, %f367, %f374;
add.rn.f32 %f376, %f375, %f372;
mov.b32 %r210, %f373;
setp.eq.s32 %p47, %r210, 1118925336;
add.s32 %r211, %r210, -1;
mov.b32 %f377, %r211;
add.f32 %f378, %f376, 0f37000000;
selp.f32 %f88, %f378, %f376, %p47;
selp.f32 %f379, %f377, %f373, %p47;
mov.f32 %f380, 0f3FB8AA3B;
mul.rn.f32 %f381, %f379, %f380;
cvt.rzi.f32.f32 %f382, %f381;
abs.f32 %f383, %f382;
setp.gt.f32 %p48, %f383, 0f42FC0000;
mov.b32 %r212, %f382;
and.b32 %r213, %r212, -2147483648;
or.b32 %r214, %r213, 1123811328;
mov.b32 %f384, %r214;
selp.f32 %f385, %f384, %f382, %p48;
mov.f32 %f386, 0fBF317218;
fma.rn.f32 %f387, %f385, %f386, %f379;
mov.f32 %f388, 0f3102E308;
fma.rn.f32 %f389, %f385, %f388, %f387;
mul.f32 %f390, %f389, 0f3FB8AA3B;
add.f32 %f391, %f385, 0f4B40007F;
mov.b32 %r215, %f391;
shl.b32 %r216, %r215, 23;
mov.b32 %f392, %r216;
ex2.approx.ftz.f32 %f393, %f390;
mul.f32 %f89, %f393, %f392;
setp.eq.f32 %p49, %f89, 0f7F800000;
mov.f32 %f968, 0f7F800000;
@%p49 bra $L__BB0_39;
fma.rn.f32 %f968, %f89, %f88, %f89;
$L__BB0_39:
mov.f32 %f888, 0f3E666666;
cvt.rzi.f32.f32 %f887, %f888;
add.f32 %f886, %f887, %f887;
mov.f32 %f885, 0f3EE66666;
sub.f32 %f884, %f885, %f886;
abs.f32 %f883, %f884;
setp.lt.f32 %p50, %f83, 0f00000000;
setp.eq.f32 %p51, %f883, 0f3F800000;
and.pred %p1, %p50, %p51;
setp.eq.f32 %p52, %f83, 0f00000000;
@%p52 bra $L__BB0_43;
bra.uni $L__BB0_40;
$L__BB0_43:
add.f32 %f398, %f83, %f83;
selp.f32 %f970, %f398, 0f00000000, %p51;
bra.uni $L__BB0_44;
$L__BB0_121:
mov.f32 %f875, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs52, %f875;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs51, %f875;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f875;}
// end inline asm
mov.u16 %rs53, 0;
st.global.v4.u16 [%rd19], {%rs50, %rs51, %rs52, %rs53};
$L__BB0_122:
and.b32 %r315, %r50, 64;
setp.eq.s32 %p147, %r315, 0;
@%p147 bra $L__BB0_124;
ld.const.u64 %rd73, [params+208];
cvta.to.global.u64 %rd74, %rd73;
ld.const.u32 %r316, [params+200];
mad.lo.s32 %r317, %r316, %r5, %r4;
mul.wide.u32 %rd75, %r317, 4;
add.s64 %rd76, %rd74, %rd75;
mov.u16 %rs54, 0;
st.global.v4.u8 [%rd76], {%rs54, %rs54, %rs54, %rs54};
bra.uni $L__BB0_124;
$L__BB0_40:
mov.b32 %r217, %f968;
xor.b32 %r218, %r217, -2147483648;
mov.b32 %f394, %r218;
selp.f32 %f970, %f394, %f968, %p1;
setp.geu.f32 %p53, %f83, 0f00000000;
@%p53 bra $L__BB0_44;
mov.f32 %f395, 0f3EE66666;
cvt.rzi.f32.f32 %f396, %f395;
setp.eq.f32 %p54, %f396, 0f3EE66666;
@%p54 bra $L__BB0_44;
mov.f32 %f970, 0f7FFFFFFF;
$L__BB0_44:
abs.f32 %f889, %f83;
add.f32 %f399, %f889, 0f3EE66666;
mov.b32 %r219, %f399;
setp.lt.s32 %p56, %r219, 2139095040;
@%p56 bra $L__BB0_49;
abs.f32 %f890, %f83;
setp.gtu.f32 %p57, %f890, 0f7F800000;
@%p57 bra $L__BB0_48;
bra.uni $L__BB0_46;
$L__BB0_48:
add.f32 %f970, %f83, 0f3EE66666;
bra.uni $L__BB0_49;
$L__BB0_46:
abs.f32 %f891, %f83;
setp.neu.f32 %p58, %f891, 0f7F800000;
@%p58 bra $L__BB0_49;
selp.f32 %f970, 0fFF800000, 0f7F800000, %p1;
$L__BB0_49:
setp.eq.f32 %p59, %f83, 0f3F800000;
selp.f32 %f98, 0f3F800000, %f970, %p59;
abs.f32 %f99, %f84;
setp.lt.f32 %p60, %f99, 0f00800000;
mul.f32 %f401, %f99, 0f4B800000;
selp.f32 %f402, %f401, %f99, %p60;
selp.f32 %f403, 0fC3170000, 0fC2FE0000, %p60;
mov.b32 %r220, %f402;
and.b32 %r221, %r220, 8388607;
or.b32 %r222, %r221, 1065353216;
mov.b32 %f404, %r222;
shr.u32 %r223, %r220, 23;
cvt.rn.f32.u32 %f405, %r223;
add.f32 %f406, %f403, %f405;
setp.gt.f32 %p61, %f404, 0f3FB504F3;
mul.f32 %f407, %f404, 0f3F000000;
add.f32 %f408, %f406, 0f3F800000;
selp.f32 %f409, %f408, %f406, %p61;
selp.f32 %f410, %f407, %f404, %p61;
add.f32 %f411, %f410, 0fBF800000;
add.f32 %f412, %f410, 0f3F800000;
rcp.approx.ftz.f32 %f413, %f412;
add.f32 %f414, %f411, %f411;
mul.f32 %f415, %f414, %f413;
mul.f32 %f416, %f415, %f415;
mov.f32 %f417, 0f3C4CAF63;
mov.f32 %f418, 0f3B18F0FE;
fma.rn.f32 %f419, %f418, %f416, %f417;
mov.f32 %f420, 0f3DAAAABD;
fma.rn.f32 %f421, %f419, %f416, %f420;
mul.rn.f32 %f422, %f421, %f416;
mul.rn.f32 %f423, %f422, %f415;
sub.f32 %f424, %f411, %f415;
add.f32 %f425, %f424, %f424;
neg.f32 %f426, %f415;
fma.rn.f32 %f427, %f426, %f411, %f425;
mul.rn.f32 %f428, %f413, %f427;
add.f32 %f429, %f423, %f415;
sub.f32 %f430, %f415, %f429;
add.f32 %f431, %f423, %f430;
add.f32 %f432, %f428, %f431;
add.f32 %f433, %f429, %f432;
sub.f32 %f434, %f429, %f433;
add.f32 %f435, %f432, %f434;
mov.f32 %f436, 0f3F317200;
mul.rn.f32 %f437, %f409, %f436;
mov.f32 %f438, 0f35BFBE8E;
mul.rn.f32 %f439, %f409, %f438;
add.f32 %f440, %f437, %f433;
sub.f32 %f441, %f437, %f440;
add.f32 %f442, %f433, %f441;
add.f32 %f443, %f435, %f442;
add.f32 %f444, %f439, %f443;
add.f32 %f445, %f440, %f444;
sub.f32 %f446, %f440, %f445;
add.f32 %f447, %f444, %f446;
mov.f32 %f448, 0f3EE66666;
mul.rn.f32 %f449, %f448, %f445;
neg.f32 %f450, %f449;
fma.rn.f32 %f451, %f448, %f445, %f450;
fma.rn.f32 %f452, %f448, %f447, %f451;
mov.f32 %f453, 0f00000000;
fma.rn.f32 %f454, %f453, %f445, %f452;
add.rn.f32 %f455, %f449, %f454;
neg.f32 %f456, %f455;
add.rn.f32 %f457, %f449, %f456;
add.rn.f32 %f458, %f457, %f454;
mov.b32 %r224, %f455;
setp.eq.s32 %p62, %r224, 1118925336;
add.s32 %r225, %r224, -1;
mov.b32 %f459, %r225;
add.f32 %f460, %f458, 0f37000000;
selp.f32 %f100, %f460, %f458, %p62;
selp.f32 %f461, %f459, %f455, %p62;
mov.f32 %f462, 0f3FB8AA3B;
mul.rn.f32 %f463, %f461, %f462;
cvt.rzi.f32.f32 %f464, %f463;
abs.f32 %f465, %f464;
setp.gt.f32 %p63, %f465, 0f42FC0000;
mov.b32 %r226, %f464;
and.b32 %r227, %r226, -2147483648;
or.b32 %r228, %r227, 1123811328;
mov.b32 %f466, %r228;
selp.f32 %f467, %f466, %f464, %p63;
mov.f32 %f468, 0fBF317218;
fma.rn.f32 %f469, %f467, %f468, %f461;
mov.f32 %f470, 0f3102E308;
fma.rn.f32 %f471, %f467, %f470, %f469;
mul.f32 %f472, %f471, 0f3FB8AA3B;
add.f32 %f473, %f467, 0f4B40007F;
mov.b32 %r229, %f473;
shl.b32 %r230, %r229, 23;
mov.b32 %f474, %r230;
ex2.approx.ftz.f32 %f475, %f472;
mul.f32 %f101, %f475, %f474;
setp.eq.f32 %p64, %f101, 0f7F800000;
mov.f32 %f971, 0f7F800000;
@%p64 bra $L__BB0_51;
fma.rn.f32 %f971, %f101, %f100, %f101;
$L__BB0_51:
setp.lt.f32 %p65, %f84, 0f00000000;
and.pred %p2, %p65, %p51;
setp.eq.f32 %p67, %f84, 0f00000000;
@%p67 bra $L__BB0_55;
bra.uni $L__BB0_52;
$L__BB0_55:
add.f32 %f480, %f84, %f84;
selp.f32 %f973, %f480, 0f00000000, %p51;
bra.uni $L__BB0_56;
$L__BB0_52:
mov.b32 %r231, %f971;
xor.b32 %r232, %r231, -2147483648;
mov.b32 %f476, %r232;
selp.f32 %f973, %f476, %f971, %p2;
setp.geu.f32 %p68, %f84, 0f00000000;
@%p68 bra $L__BB0_56;
mov.f32 %f477, 0f3EE66666;
cvt.rzi.f32.f32 %f478, %f477;
setp.eq.f32 %p69, %f478, 0f3EE66666;
@%p69 bra $L__BB0_56;
mov.f32 %f973, 0f7FFFFFFF;
$L__BB0_56:
abs.f32 %f892, %f84;
add.f32 %f481, %f892, 0f3EE66666;
mov.b32 %r233, %f481;
setp.lt.s32 %p71, %r233, 2139095040;
@%p71 bra $L__BB0_61;
abs.f32 %f893, %f84;
setp.gtu.f32 %p72, %f893, 0f7F800000;
@%p72 bra $L__BB0_60;
bra.uni $L__BB0_58;
$L__BB0_60:
add.f32 %f973, %f84, 0f3EE66666;
bra.uni $L__BB0_61;
$L__BB0_58:
abs.f32 %f894, %f84;
setp.neu.f32 %p73, %f894, 0f7F800000;
@%p73 bra $L__BB0_61;
selp.f32 %f973, 0fFF800000, 0f7F800000, %p2;
$L__BB0_61:
setp.eq.f32 %p74, %f84, 0f3F800000;
selp.f32 %f110, 0f3F800000, %f973, %p74;
abs.f32 %f111, %f85;
setp.lt.f32 %p75, %f111, 0f00800000;
mul.f32 %f483, %f111, 0f4B800000;
selp.f32 %f484, %f483, %f111, %p75;
selp.f32 %f485, 0fC3170000, 0fC2FE0000, %p75;
mov.b32 %r234, %f484;
and.b32 %r235, %r234, 8388607;
or.b32 %r236, %r235, 1065353216;
mov.b32 %f486, %r236;
shr.u32 %r237, %r234, 23;
cvt.rn.f32.u32 %f487, %r237;
add.f32 %f488, %f485, %f487;
setp.gt.f32 %p76, %f486, 0f3FB504F3;
mul.f32 %f489, %f486, 0f3F000000;
add.f32 %f490, %f488, 0f3F800000;
selp.f32 %f491, %f490, %f488, %p76;
selp.f32 %f492, %f489, %f486, %p76;
add.f32 %f493, %f492, 0fBF800000;
add.f32 %f494, %f492, 0f3F800000;
rcp.approx.ftz.f32 %f495, %f494;
add.f32 %f496, %f493, %f493;
mul.f32 %f497, %f496, %f495;
mul.f32 %f498, %f497, %f497;
mov.f32 %f499, 0f3C4CAF63;
mov.f32 %f500, 0f3B18F0FE;
fma.rn.f32 %f501, %f500, %f498, %f499;
mov.f32 %f502, 0f3DAAAABD;
fma.rn.f32 %f503, %f501, %f498, %f502;
mul.rn.f32 %f504, %f503, %f498;
mul.rn.f32 %f505, %f504, %f497;
sub.f32 %f506, %f493, %f497;
add.f32 %f507, %f506, %f506;
neg.f32 %f508, %f497;
fma.rn.f32 %f509, %f508, %f493, %f507;
mul.rn.f32 %f510, %f495, %f509;
add.f32 %f511, %f505, %f497;
sub.f32 %f512, %f497, %f511;
add.f32 %f513, %f505, %f512;
add.f32 %f514, %f510, %f513;
add.f32 %f515, %f511, %f514;
sub.f32 %f516, %f511, %f515;
add.f32 %f517, %f514, %f516;
mov.f32 %f518, 0f3F317200;
mul.rn.f32 %f519, %f491, %f518;
mov.f32 %f520, 0f35BFBE8E;
mul.rn.f32 %f521, %f491, %f520;
add.f32 %f522, %f519, %f515;
sub.f32 %f523, %f519, %f522;
add.f32 %f524, %f515, %f523;
add.f32 %f525, %f517, %f524;
add.f32 %f526, %f521, %f525;
add.f32 %f527, %f522, %f526;
sub.f32 %f528, %f522, %f527;
add.f32 %f529, %f526, %f528;
mov.f32 %f530, 0f3EE66666;
mul.rn.f32 %f531, %f530, %f527;
neg.f32 %f532, %f531;
fma.rn.f32 %f533, %f530, %f527, %f532;
fma.rn.f32 %f534, %f530, %f529, %f533;
mov.f32 %f535, 0f00000000;
fma.rn.f32 %f536, %f535, %f527, %f534;
add.rn.f32 %f537, %f531, %f536;
neg.f32 %f538, %f537;
add.rn.f32 %f539, %f531, %f538;
add.rn.f32 %f540, %f539, %f536;
mov.b32 %r238, %f537;
setp.eq.s32 %p77, %r238, 1118925336;
add.s32 %r239, %r238, -1;
mov.b32 %f541, %r239;
add.f32 %f542, %f540, 0f37000000;
selp.f32 %f112, %f542, %f540, %p77;
selp.f32 %f543, %f541, %f537, %p77;
mov.f32 %f544, 0f3FB8AA3B;
mul.rn.f32 %f545, %f543, %f544;
cvt.rzi.f32.f32 %f546, %f545;
abs.f32 %f547, %f546;
setp.gt.f32 %p78, %f547, 0f42FC0000;
mov.b32 %r240, %f546;
and.b32 %r241, %r240, -2147483648;
or.b32 %r242, %r241, 1123811328;
mov.b32 %f548, %r242;
selp.f32 %f549, %f548, %f546, %p78;
mov.f32 %f550, 0fBF317218;
fma.rn.f32 %f551, %f549, %f550, %f543;
mov.f32 %f552, 0f3102E308;
fma.rn.f32 %f553, %f549, %f552, %f551;
mul.f32 %f554, %f553, 0f3FB8AA3B;
add.f32 %f555, %f549, 0f4B40007F;
mov.b32 %r243, %f555;
shl.b32 %r244, %r243, 23;
mov.b32 %f556, %r244;
ex2.approx.ftz.f32 %f557, %f554;
mul.f32 %f113, %f557, %f556;
setp.eq.f32 %p79, %f113, 0f7F800000;
mov.f32 %f974, 0f7F800000;
@%p79 bra $L__BB0_63;
fma.rn.f32 %f974, %f113, %f112, %f113;
$L__BB0_63:
setp.lt.f32 %p80, %f85, 0f00000000;
and.pred %p3, %p80, %p51;
setp.eq.f32 %p82, %f85, 0f00000000;
@%p82 bra $L__BB0_67;
bra.uni $L__BB0_64;
$L__BB0_67:
add.f32 %f562, %f85, %f85;
selp.f32 %f976, %f562, 0f00000000, %p51;
bra.uni $L__BB0_68;
$L__BB0_64:
mov.b32 %r245, %f974;
xor.b32 %r246, %r245, -2147483648;
mov.b32 %f558, %r246;
selp.f32 %f976, %f558, %f974, %p3;
setp.geu.f32 %p83, %f85, 0f00000000;
@%p83 bra $L__BB0_68;
mov.f32 %f559, 0f3EE66666;
cvt.rzi.f32.f32 %f560, %f559;
setp.eq.f32 %p84, %f560, 0f3EE66666;
@%p84 bra $L__BB0_68;
mov.f32 %f976, 0f7FFFFFFF;
$L__BB0_68:
abs.f32 %f895, %f85;
add.f32 %f563, %f895, 0f3EE66666;
mov.b32 %r247, %f563;
setp.lt.s32 %p86, %r247, 2139095040;
@%p86 bra $L__BB0_73;
abs.f32 %f896, %f85;
setp.gtu.f32 %p87, %f896, 0f7F800000;
@%p87 bra $L__BB0_72;
bra.uni $L__BB0_70;
$L__BB0_72:
add.f32 %f976, %f85, 0f3EE66666;
bra.uni $L__BB0_73;
$L__BB0_70:
abs.f32 %f897, %f85;
setp.neu.f32 %p88, %f897, 0f7F800000;
@%p88 bra $L__BB0_73;
selp.f32 %f976, 0fFF800000, 0f7F800000, %p3;
$L__BB0_73:
setp.eq.f32 %p89, %f85, 0f3F800000;
mov.f32 %f565, 0f3F800000;
selp.f32 %f566, 0f3F800000, %f976, %p89;
min.f32 %f567, %f98, %f565;
mov.f32 %f568, 0f00000000;
max.f32 %f122, %f568, %f567;
min.f32 %f569, %f110, %f565;
max.f32 %f123, %f568, %f569;
min.f32 %f570, %f566, %f565;
max.f32 %f124, %f568, %f570;
mov.f32 %f574, 0f3ED55555;
abs.f32 %f126, %f122;
setp.lt.f32 %p90, %f126, 0f00800000;
mul.f32 %f576, %f126, 0f4B800000;
selp.f32 %f577, %f576, %f126, %p90;
selp.f32 %f578, 0fC3170000, 0fC2FE0000, %p90;
mov.b32 %r250, %f577;
and.b32 %r251, %r250, 8388607;
or.b32 %r252, %r251, 1065353216;
mov.b32 %f579, %r252;
shr.u32 %r253, %r250, 23;
cvt.rn.f32.u32 %f580, %r253;
add.f32 %f581, %f578, %f580;
setp.gt.f32 %p91, %f579, 0f3FB504F3;
mul.f32 %f582, %f579, 0f3F000000;
add.f32 %f583, %f581, 0f3F800000;
selp.f32 %f584, %f583, %f581, %p91;
selp.f32 %f585, %f582, %f579, %p91;
add.f32 %f586, %f585, 0fBF800000;
add.f32 %f587, %f585, 0f3F800000;
rcp.approx.ftz.f32 %f588, %f587;
add.f32 %f589, %f586, %f586;
mul.f32 %f590, %f589, %f588;
mul.f32 %f591, %f590, %f590;
mov.f32 %f592, 0f3C4CAF63;
mov.f32 %f593, 0f3B18F0FE;
fma.rn.f32 %f594, %f593, %f591, %f592;
mov.f32 %f595, 0f3DAAAABD;
fma.rn.f32 %f596, %f594, %f591, %f595;
mul.rn.f32 %f597, %f596, %f591;
mul.rn.f32 %f598, %f597, %f590;
sub.f32 %f599, %f586, %f590;
add.f32 %f600, %f599, %f599;
neg.f32 %f601, %f590;
fma.rn.f32 %f602, %f601, %f586, %f600;
mul.rn.f32 %f603, %f588, %f602;
add.f32 %f604, %f598, %f590;
sub.f32 %f605, %f590, %f604;
add.f32 %f606, %f598, %f605;
add.f32 %f607, %f603, %f606;
add.f32 %f608, %f604, %f607;
sub.f32 %f609, %f604, %f608;
add.f32 %f610, %f607, %f609;
mov.f32 %f611, 0f3F317200;
mul.rn.f32 %f612, %f584, %f611;
mov.f32 %f613, 0f35BFBE8E;
mul.rn.f32 %f614, %f584, %f613;
add.f32 %f615, %f612, %f608;
sub.f32 %f616, %f612, %f615;
add.f32 %f617, %f608, %f616;
add.f32 %f618, %f610, %f617;
add.f32 %f619, %f614, %f618;
add.f32 %f620, %f615, %f619;
sub.f32 %f621, %f615, %f620;
add.f32 %f622, %f619, %f621;
mul.rn.f32 %f623, %f574, %f620;
neg.f32 %f624, %f623;
fma.rn.f32 %f625, %f574, %f620, %f624;
fma.rn.f32 %f626, %f574, %f622, %f625;
fma.rn.f32 %f627, %f568, %f620, %f626;
add.rn.f32 %f628, %f623, %f627;
neg.f32 %f629, %f628;
add.rn.f32 %f630, %f623, %f629;
add.rn.f32 %f631, %f630, %f627;
mov.b32 %r254, %f628;
setp.eq.s32 %p92, %r254, 1118925336;
add.s32 %r255, %r254, -1;
mov.b32 %f632, %r255;
add.f32 %f633, %f631, 0f37000000;
selp.f32 %f127, %f633, %f631, %p92;
selp.f32 %f634, %f632, %f628, %p92;
mov.f32 %f635, 0f3FB8AA3B;
mul.rn.f32 %f636, %f634, %f635;
cvt.rzi.f32.f32 %f637, %f636;
abs.f32 %f638, %f637;
setp.gt.f32 %p93, %f638, 0f42FC0000;
mov.b32 %r256, %f637;
and.b32 %r257, %r256, -2147483648;
or.b32 %r258, %r257, 1123811328;
mov.b32 %f639, %r258;
selp.f32 %f640, %f639, %f637, %p93;
mov.f32 %f641, 0fBF317218;
fma.rn.f32 %f642, %f640, %f641, %f634;
mov.f32 %f643, 0f3102E308;
fma.rn.f32 %f644, %f640, %f643, %f642;
mul.f32 %f645, %f644, 0f3FB8AA3B;
add.f32 %f646, %f640, 0f4B40007F;
mov.b32 %r259, %f646;
shl.b32 %r260, %r259, 23;
mov.b32 %f647, %r260;
ex2.approx.ftz.f32 %f648, %f645;
mul.f32 %f128, %f648, %f647;
setp.eq.f32 %p94, %f128, 0f7F800000;
mov.f32 %f977, 0f7F800000;
@%p94 bra $L__BB0_75;
fma.rn.f32 %f977, %f128, %f127, %f128;
$L__BB0_75:
mov.f32 %f917, 0f00000000;
max.f32 %f916, %f917, %f567;
mov.f32 %f881, 0f3E555555;
cvt.rzi.f32.f32 %f880, %f881;
add.f32 %f879, %f880, %f880;
mov.f32 %f878, 0f3ED55555;
sub.f32 %f877, %f878, %f879;
abs.f32 %f876, %f877;
setp.lt.f32 %p95, %f916, 0f00000000;
setp.eq.f32 %p96, %f876, 0f3F800000;
and.pred %p4, %p95, %p96;
setp.eq.f32 %p97, %f916, 0f00000000;
@%p97 bra $L__BB0_79;
bra.uni $L__BB0_76;
$L__BB0_79:
mov.f32 %f934, 0f00000000;
max.f32 %f933, %f934, %f567;
add.f32 %f653, %f933, %f933;
selp.f32 %f979, %f653, 0f00000000, %p96;
bra.uni $L__BB0_80;
$L__BB0_76:
mov.f32 %f919, 0f00000000;
max.f32 %f918, %f919, %f567;
mov.b32 %r261, %f977;
xor.b32 %r262, %r261, -2147483648;
mov.b32 %f649, %r262;
selp.f32 %f979, %f649, %f977, %p4;
setp.geu.f32 %p98, %f918, 0f00000000;
@%p98 bra $L__BB0_80;
mov.f32 %f650, 0f3ED55555;
cvt.rzi.f32.f32 %f651, %f650;
setp.eq.f32 %p99, %f651, 0f3ED55555;
@%p99 bra $L__BB0_80;
mov.f32 %f979, 0f7FFFFFFF;
$L__BB0_80:
mov.f32 %f922, 0f00000000;
max.f32 %f921, %f922, %f567;
abs.f32 %f920, %f921;
add.f32 %f654, %f920, 0f3ED55555;
mov.b32 %r263, %f654;
setp.lt.s32 %p101, %r263, 2139095040;
@%p101 bra $L__BB0_85;
mov.f32 %f927, 0f00000000;
max.f32 %f926, %f927, %f567;
abs.f32 %f925, %f926;
setp.gtu.f32 %p102, %f925, 0f7F800000;
@%p102 bra $L__BB0_84;
bra.uni $L__BB0_82;
$L__BB0_84:
mov.f32 %f932, 0f00000000;
max.f32 %f931, %f932, %f567;
add.f32 %f979, %f931, 0f3ED55555;
bra.uni $L__BB0_85;
$L__BB0_82:
mov.f32 %f930, 0f00000000;
max.f32 %f929, %f930, %f567;
abs.f32 %f928, %f929;
setp.neu.f32 %p103, %f928, 0f7F800000;
@%p103 bra $L__BB0_85;
selp.f32 %f979, 0fFF800000, 0f7F800000, %p4;
$L__BB0_85:
abs.f32 %f137, %f123;
setp.lt.f32 %p104, %f137, 0f00800000;
mul.f32 %f656, %f137, 0f4B800000;
selp.f32 %f657, %f656, %f137, %p104;
selp.f32 %f658, 0fC3170000, 0fC2FE0000, %p104;
mov.b32 %r264, %f657;
and.b32 %r265, %r264, 8388607;
or.b32 %r266, %r265, 1065353216;
mov.b32 %f659, %r266;
shr.u32 %r267, %r264, 23;
cvt.rn.f32.u32 %f660, %r267;
add.f32 %f661, %f658, %f660;
setp.gt.f32 %p105, %f659, 0f3FB504F3;
mul.f32 %f662, %f659, 0f3F000000;
add.f32 %f663, %f661, 0f3F800000;
selp.f32 %f664, %f663, %f661, %p105;
selp.f32 %f665, %f662, %f659, %p105;
add.f32 %f666, %f665, 0fBF800000;
add.f32 %f667, %f665, 0f3F800000;
rcp.approx.ftz.f32 %f668, %f667;
add.f32 %f669, %f666, %f666;
mul.f32 %f670, %f669, %f668;
mul.f32 %f671, %f670, %f670;
mov.f32 %f672, 0f3C4CAF63;
mov.f32 %f673, 0f3B18F0FE;
fma.rn.f32 %f674, %f673, %f671, %f672;
mov.f32 %f675, 0f3DAAAABD;
fma.rn.f32 %f676, %f674, %f671, %f675;
mul.rn.f32 %f677, %f676, %f671;
mul.rn.f32 %f678, %f677, %f670;
sub.f32 %f679, %f666, %f670;
add.f32 %f680, %f679, %f679;
neg.f32 %f681, %f670;
fma.rn.f32 %f682, %f681, %f666, %f680;
mul.rn.f32 %f683, %f668, %f682;
add.f32 %f684, %f678, %f670;
sub.f32 %f685, %f670, %f684;
add.f32 %f686, %f678, %f685;
add.f32 %f687, %f683, %f686;
add.f32 %f688, %f684, %f687;
sub.f32 %f689, %f684, %f688;
add.f32 %f690, %f687, %f689;
mov.f32 %f691, 0f3F317200;
mul.rn.f32 %f692, %f664, %f691;
mov.f32 %f693, 0f35BFBE8E;
mul.rn.f32 %f694, %f664, %f693;
add.f32 %f695, %f692, %f688;
sub.f32 %f696, %f692, %f695;
add.f32 %f697, %f688, %f696;
add.f32 %f698, %f690, %f697;
add.f32 %f699, %f694, %f698;
add.f32 %f700, %f695, %f699;
sub.f32 %f701, %f695, %f700;
add.f32 %f702, %f699, %f701;
mov.f32 %f703, 0f3ED55555;
mul.rn.f32 %f704, %f703, %f700;
neg.f32 %f705, %f704;
fma.rn.f32 %f706, %f703, %f700, %f705;
fma.rn.f32 %f707, %f703, %f702, %f706;
mov.f32 %f708, 0f00000000;
fma.rn.f32 %f709, %f708, %f700, %f707;
add.rn.f32 %f710, %f704, %f709;
neg.f32 %f711, %f710;
add.rn.f32 %f712, %f704, %f711;
add.rn.f32 %f713, %f712, %f709;
mov.b32 %r268, %f710;
setp.eq.s32 %p106, %r268, 1118925336;
add.s32 %r269, %r268, -1;
mov.b32 %f714, %r269;
add.f32 %f715, %f713, 0f37000000;
selp.f32 %f138, %f715, %f713, %p106;
selp.f32 %f716, %f714, %f710, %p106;
mov.f32 %f717, 0f3FB8AA3B;
mul.rn.f32 %f718, %f716, %f717;
cvt.rzi.f32.f32 %f719, %f718;
abs.f32 %f720, %f719;
setp.gt.f32 %p107, %f720, 0f42FC0000;
mov.b32 %r270, %f719;
and.b32 %r271, %r270, -2147483648;
or.b32 %r272, %r271, 1123811328;
mov.b32 %f721, %r272;
selp.f32 %f722, %f721, %f719, %p107;
mov.f32 %f723, 0fBF317218;
fma.rn.f32 %f724, %f722, %f723, %f716;
mov.f32 %f725, 0f3102E308;
fma.rn.f32 %f726, %f722, %f725, %f724;
mul.f32 %f727, %f726, 0f3FB8AA3B;
add.f32 %f728, %f722, 0f4B40007F;
mov.b32 %r273, %f728;
shl.b32 %r274, %r273, 23;
mov.b32 %f729, %r274;
ex2.approx.ftz.f32 %f730, %f727;
mul.f32 %f139, %f730, %f729;
setp.eq.f32 %p108, %f139, 0f7F800000;
mov.f32 %f980, 0f7F800000;
@%p108 bra $L__BB0_87;
fma.rn.f32 %f980, %f139, %f138, %f139;
$L__BB0_87:
setp.lt.f32 %p109, %f123, 0f00000000;
and.pred %p5, %p109, %p96;
setp.eq.f32 %p111, %f123, 0f00000000;
@%p111 bra $L__BB0_91;
bra.uni $L__BB0_88;
$L__BB0_91:
add.f32 %f735, %f123, %f123;
selp.f32 %f982, %f735, 0f00000000, %p96;
bra.uni $L__BB0_92;
$L__BB0_88:
mov.b32 %r275, %f980;
xor.b32 %r276, %r275, -2147483648;
mov.b32 %f731, %r276;
selp.f32 %f982, %f731, %f980, %p5;
setp.geu.f32 %p112, %f123, 0f00000000;
@%p112 bra $L__BB0_92;
mov.f32 %f732, 0f3ED55555;
cvt.rzi.f32.f32 %f733, %f732;
setp.eq.f32 %p113, %f733, 0f3ED55555;
@%p113 bra $L__BB0_92;
mov.f32 %f982, 0f7FFFFFFF;
$L__BB0_92:
mov.f32 %f900, 0f00000000;
max.f32 %f899, %f900, %f569;
abs.f32 %f898, %f899;
add.f32 %f736, %f898, 0f3ED55555;
mov.b32 %r277, %f736;
setp.lt.s32 %p115, %r277, 2139095040;
@%p115 bra $L__BB0_97;
mov.f32 %f903, 0f00000000;
max.f32 %f902, %f903, %f569;
abs.f32 %f901, %f902;
setp.gtu.f32 %p116, %f901, 0f7F800000;
@%p116 bra $L__BB0_96;
bra.uni $L__BB0_94;
$L__BB0_96:
add.f32 %f982, %f123, 0f3ED55555;
bra.uni $L__BB0_97;
$L__BB0_94:
mov.f32 %f906, 0f00000000;
max.f32 %f905, %f906, %f569;
abs.f32 %f904, %f905;
setp.neu.f32 %p117, %f904, 0f7F800000;
@%p117 bra $L__BB0_97;
selp.f32 %f982, 0fFF800000, 0f7F800000, %p5;
$L__BB0_97:
abs.f32 %f148, %f124;
setp.lt.f32 %p118, %f148, 0f00800000;
mul.f32 %f738, %f148, 0f4B800000;
selp.f32 %f739, %f738, %f148, %p118;
selp.f32 %f740, 0fC3170000, 0fC2FE0000, %p118;
mov.b32 %r278, %f739;
and.b32 %r279, %r278, 8388607;
or.b32 %r280, %r279, 1065353216;
mov.b32 %f741, %r280;
shr.u32 %r281, %r278, 23;
cvt.rn.f32.u32 %f742, %r281;
add.f32 %f743, %f740, %f742;
setp.gt.f32 %p119, %f741, 0f3FB504F3;
mul.f32 %f744, %f741, 0f3F000000;
add.f32 %f745, %f743, 0f3F800000;
selp.f32 %f746, %f745, %f743, %p119;
selp.f32 %f747, %f744, %f741, %p119;
add.f32 %f748, %f747, 0fBF800000;
add.f32 %f749, %f747, 0f3F800000;
rcp.approx.ftz.f32 %f750, %f749;
add.f32 %f751, %f748, %f748;
mul.f32 %f752, %f751, %f750;
mul.f32 %f753, %f752, %f752;
mov.f32 %f754, 0f3C4CAF63;
mov.f32 %f755, 0f3B18F0FE;
fma.rn.f32 %f756, %f755, %f753, %f754;
mov.f32 %f757, 0f3DAAAABD;
fma.rn.f32 %f758, %f756, %f753, %f757;
mul.rn.f32 %f759, %f758, %f753;
mul.rn.f32 %f760, %f759, %f752;
sub.f32 %f761, %f748, %f752;
add.f32 %f762, %f761, %f761;
neg.f32 %f763, %f752;
fma.rn.f32 %f764, %f763, %f748, %f762;
mul.rn.f32 %f765, %f750, %f764;
add.f32 %f766, %f760, %f752;
sub.f32 %f767, %f752, %f766;
add.f32 %f768, %f760, %f767;
add.f32 %f769, %f765, %f768;
add.f32 %f770, %f766, %f769;
sub.f32 %f771, %f766, %f770;
add.f32 %f772, %f769, %f771;
mov.f32 %f773, 0f3F317200;
mul.rn.f32 %f774, %f746, %f773;
mov.f32 %f775, 0f35BFBE8E;
mul.rn.f32 %f776, %f746, %f775;
add.f32 %f777, %f774, %f770;
sub.f32 %f778, %f774, %f777;
add.f32 %f779, %f770, %f778;
add.f32 %f780, %f772, %f779;
add.f32 %f781, %f776, %f780;
add.f32 %f782, %f777, %f781;
sub.f32 %f783, %f777, %f782;
add.f32 %f784, %f781, %f783;
mov.f32 %f785, 0f3ED55555;
mul.rn.f32 %f786, %f785, %f782;
neg.f32 %f787, %f786;
fma.rn.f32 %f788, %f785, %f782, %f787;
fma.rn.f32 %f789, %f785, %f784, %f788;
mov.f32 %f790, 0f00000000;
fma.rn.f32 %f791, %f790, %f782, %f789;
add.rn.f32 %f792, %f786, %f791;
neg.f32 %f793, %f792;
add.rn.f32 %f794, %f786, %f793;
add.rn.f32 %f795, %f794, %f791;
mov.b32 %r282, %f792;
setp.eq.s32 %p120, %r282, 1118925336;
add.s32 %r283, %r282, -1;
mov.b32 %f796, %r283;
add.f32 %f797, %f795, 0f37000000;
selp.f32 %f149, %f797, %f795, %p120;
selp.f32 %f798, %f796, %f792, %p120;
mov.f32 %f799, 0f3FB8AA3B;
mul.rn.f32 %f800, %f798, %f799;
cvt.rzi.f32.f32 %f801, %f800;
abs.f32 %f802, %f801;
setp.gt.f32 %p121, %f802, 0f42FC0000;
mov.b32 %r284, %f801;
and.b32 %r285, %r284, -2147483648;
or.b32 %r286, %r285, 1123811328;
mov.b32 %f803, %r286;
selp.f32 %f804, %f803, %f801, %p121;
mov.f32 %f805, 0fBF317218;
fma.rn.f32 %f806, %f804, %f805, %f798;
mov.f32 %f807, 0f3102E308;
fma.rn.f32 %f808, %f804, %f807, %f806;
mul.f32 %f809, %f808, 0f3FB8AA3B;
add.f32 %f810, %f804, 0f4B40007F;
mov.b32 %r287, %f810;
shl.b32 %r288, %r287, 23;
mov.b32 %f811, %r288;
ex2.approx.ftz.f32 %f812, %f809;
mul.f32 %f150, %f812, %f811;
setp.eq.f32 %p122, %f150, 0f7F800000;
mov.f32 %f983, 0f7F800000;
@%p122 bra $L__BB0_99;
fma.rn.f32 %f983, %f150, %f149, %f150;
$L__BB0_99:
setp.lt.f32 %p123, %f124, 0f00000000;
and.pred %p6, %p123, %p96;
setp.eq.f32 %p125, %f124, 0f00000000;
@%p125 bra $L__BB0_103;
bra.uni $L__BB0_100;
$L__BB0_103:
add.f32 %f817, %f124, %f124;
selp.f32 %f985, %f817, 0f00000000, %p96;
bra.uni $L__BB0_104;
$L__BB0_100:
mov.b32 %r289, %f983;
xor.b32 %r290, %r289, -2147483648;
mov.b32 %f813, %r290;
selp.f32 %f985, %f813, %f983, %p6;
setp.geu.f32 %p126, %f124, 0f00000000;
@%p126 bra $L__BB0_104;
mov.f32 %f814, 0f3ED55555;
cvt.rzi.f32.f32 %f815, %f814;
setp.eq.f32 %p127, %f815, 0f3ED55555;
@%p127 bra $L__BB0_104;
mov.f32 %f985, 0f7FFFFFFF;
$L__BB0_104:
mov.f32 %f909, 0f00000000;
max.f32 %f908, %f909, %f570;
abs.f32 %f907, %f908;
add.f32 %f818, %f907, 0f3ED55555;
mov.b32 %r291, %f818;
setp.lt.s32 %p129, %r291, 2139095040;
@%p129 bra $L__BB0_109;
mov.f32 %f912, 0f00000000;
max.f32 %f911, %f912, %f570;
abs.f32 %f910, %f911;
setp.gtu.f32 %p130, %f910, 0f7F800000;
@%p130 bra $L__BB0_108;
bra.uni $L__BB0_106;
$L__BB0_108:
add.f32 %f985, %f124, 0f3ED55555;
bra.uni $L__BB0_109;
$L__BB0_106:
mov.f32 %f915, 0f00000000;
max.f32 %f914, %f915, %f570;
abs.f32 %f913, %f914;
setp.neu.f32 %p131, %f913, 0f7F800000;
@%p131 bra $L__BB0_109;
selp.f32 %f985, 0fFF800000, 0f7F800000, %p6;
$L__BB0_109:
mov.f32 %f924, 0f00000000;
max.f32 %f923, %f924, %f567;
ld.const.u64 %rd79, [params+144];
cvta.to.global.u64 %rd78, %rd79;
ld.const.u32 %r319, [params+136];
mad.lo.s32 %r318, %r319, %r5, %r4;
cvt.u64.u32 %rd77, %r318;
fma.rn.f32 %f819, %f979, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p132, %f923, 0f3F800000;
mov.f32 %f820, 0f3F800000;
selp.f32 %f821, 0f3F7FFFFF, %f819, %p132;
mul.f32 %f822, %f923, 0f414EB852;
setp.lt.f32 %p133, %f923, 0f3B4D2E1C;
selp.f32 %f823, %f822, %f821, %p133;
fma.rn.f32 %f824, %f982, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p134, %f123, 0f3F800000;
selp.f32 %f825, 0f3F7FFFFF, %f824, %p134;
mul.f32 %f826, %f123, 0f414EB852;
setp.lt.f32 %p135, %f123, 0f3B4D2E1C;
selp.f32 %f827, %f826, %f825, %p135;
fma.rn.f32 %f828, %f985, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p136, %f124, 0f3F800000;
selp.f32 %f829, 0f3F7FFFFF, %f828, %p136;
mul.f32 %f830, %f124, 0f414EB852;
setp.lt.f32 %p137, %f124, 0f3B4D2E1C;
selp.f32 %f831, %f830, %f829, %p137;
min.f32 %f832, %f823, %f820;
mov.f32 %f833, 0f00000000;
max.f32 %f834, %f833, %f832;
mul.f32 %f835, %f834, 0f43800000;
cvt.rzi.u32.f32 %r292, %f835;
min.u32 %r293, %r292, 255;
min.f32 %f836, %f827, %f820;
max.f32 %f837, %f833, %f836;
mul.f32 %f838, %f837, 0f43800000;
cvt.rzi.u32.f32 %r294, %f838;
min.u32 %r295, %r294, 255;
min.f32 %f839, %f831, %f820;
max.f32 %f840, %f833, %f839;
mul.f32 %f841, %f840, 0f43800000;
cvt.rzi.u32.f32 %r296, %f841;
min.u32 %r297, %r296, 255;
shl.b64 %rd57, %rd77, 2;
add.s64 %rd58, %rd78, %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 %r320, [params+104];
and.b32 %r298, %r320, 4;
setp.eq.s32 %p138, %r298, 0;
@%p138 bra $L__BB0_114;
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 %f842, %rs26;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f843, %rs27;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f844, %rs28;}
// end inline asm
add.f32 %f845, %f83, %f842;
add.f32 %f846, %f84, %f843;
add.f32 %f847, %f85, %f844;
mov.f32 %f848, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs24, %f847;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs23, %f846;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs22, %f845;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs25, %f848;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs22, %rs23, %rs24, %rs25};
bra.uni $L__BB0_114;
$L__BB0_113:
mov.f32 %f852, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs33, %f852;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs32, %f85;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs31, %f84;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f83;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs30, %rs31, %rs32, %rs33};
$L__BB0_114:
ld.const.u32 %r321, [params+104];
and.b32 %r302, %r321, 64;
setp.eq.s32 %p140, %r302, 0;
@%p140 bra $L__BB0_124;
mul.f32 %f853, %f957, %f957;
fma.rn.f32 %f854, %f956, %f956, %f853;
fma.rn.f32 %f855, %f958, %f958, %f854;
sqrt.rn.f32 %f856, %f855;
rcp.rn.f32 %f857, %f856;
mul.f32 %f858, %f956, %f857;
mul.f32 %f859, %f957, %f857;
mul.f32 %f860, %f958, %f857;
ld.const.u64 %rd62, [params+208];
cvta.to.global.u64 %rd63, %rd62;
ld.const.u32 %r303, [params+200];
mad.lo.s32 %r304, %r303, %r5, %r4;
fma.rn.f32 %f861, %f858, 0f3F000000, 0f3F000000;
mul.f32 %f862, %f861, 0f437F0000;
cvt.rzi.u32.f32 %r305, %f862;
fma.rn.f32 %f863, %f859, 0f3F000000, 0f3F000000;
mul.f32 %f864, %f863, 0f437F0000;
cvt.rzi.u32.f32 %r306, %f864;
fma.rn.f32 %f865, %f860, 0f3F000000, 0f3F000000;
mul.f32 %f866, %f865, 0f437F0000;
cvt.rzi.u32.f32 %r307, %f866;
mul.wide.u32 %rd64, %r304, 4;
add.s64 %rd65, %rd63, %rd64;
cvt.u16.u32 %rs34, %r307;
cvt.u16.u32 %rs35, %r306;
cvt.u16.u32 %rs36, %r305;
mov.u16 %rs37, 255;
st.global.v4.u8 [%rd65], {%rs36, %rs35, %rs34, %rs37};
$L__BB0_124:
ret;
}