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

2258 lines
60 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<155>;
.reg .b16 %rs<172>;
.reg .f32 %f<1192>;
.reg .b32 %r<361>;
.reg .f64 %fd<5>;
.reg .b64 %rd<109>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r55), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r56), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u32 %r3, [params+540];
ld.const.u64 %rd29, [params+400];
cvta.to.global.u64 %rd30, %rd29;
ld.const.u32 %r61, [params+392];
mad.lo.s32 %r62, %r61, %r56, %r55;
mul.wide.u32 %rd31, %r62, 4;
add.s64 %rd2, %rd30, %rd31;
ld.global.v2.u8 {%rs7, %rs171}, [%rd2];
or.b16 %rs9, %rs7, %rs171;
and.b16 %rs10, %rs9, 255;
setp.eq.s16 %p7, %rs10, 0;
@%p7 bra $L__BB0_2;
ld.global.u8 %rs170, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs170, [%rd2+2];
setp.eq.s16 %p8, %rs170, 0;
mov.f32 %f1105, 0f00000000;
mov.u16 %rs171, 0;
mov.f32 %f1106, %f1105;
mov.f32 %f1107, %f1105;
@%p8 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f237, %rs7;
div.rn.f32 %f238, %f237, 0f437F0000;
fma.rn.f32 %f239, %f238, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs171, 255;
cvt.rn.f32.u16 %f240, %rs13;
div.rn.f32 %f241, %f240, 0f437F0000;
fma.rn.f32 %f242, %f241, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f243, %rs170;
div.rn.f32 %f244, %f243, 0f437F0000;
fma.rn.f32 %f245, %f244, 0f40000000, 0fBF800000;
mul.f32 %f246, %f242, %f242;
fma.rn.f32 %f247, %f239, %f239, %f246;
fma.rn.f32 %f248, %f245, %f245, %f247;
sqrt.rn.f32 %f249, %f248;
rcp.rn.f32 %f250, %f249;
mul.f32 %f1107, %f250, %f245;
mul.f32 %f1106, %f250, %f242;
mul.f32 %f1105, %f239, %f250;
$L__BB0_4:
ld.const.v2.u32 {%r63, %r64}, [params];
add.s32 %r4, %r63, %r55;
add.s32 %r5, %r64, %r56;
setp.eq.f32 %p9, %f1105, 0f00000000;
setp.eq.f32 %p10, %f1106, 0f00000000;
and.pred %p11, %p9, %p10;
setp.eq.f32 %p12, %f1107, 0f00000000;
and.pred %p13, %p12, %p11;
@%p13 bra $L__BB0_126;
bra.uni $L__BB0_5;
$L__BB0_126:
ld.const.u32 %r52, [params+104];
and.b32 %r314, %r52, 1;
setp.eq.b32 %p144, %r314, 1;
mov.pred %p145, 0;
xor.pred %p146, %p144, %p145;
not.pred %p147, %p146;
@%p147 bra $L__BB0_128;
ld.const.u64 %rd82, [params+144];
cvta.to.global.u64 %rd83, %rd82;
ld.const.u32 %r315, [params+136];
mad.lo.s32 %r316, %r315, %r5, %r4;
mul.wide.u32 %rd84, %r316, 4;
add.s64 %rd85, %rd83, %rd84;
mov.u16 %rs94, 0;
st.global.v4.u8 [%rd85], {%rs94, %rs94, %rs94, %rs94};
$L__BB0_128:
and.b32 %r317, %r52, 4;
setp.eq.s32 %p148, %r317, 0;
ld.const.u32 %r360, [params+108];
@%p148 bra $L__BB0_132;
setp.eq.s32 %p149, %r360, 0;
ld.const.u64 %rd86, [params+224];
cvta.to.global.u64 %rd87, %rd86;
ld.const.u32 %r318, [params+216];
mad.lo.s32 %r319, %r318, %r5, %r4;
mul.wide.u32 %rd88, %r319, 8;
add.s64 %rd23, %rd87, %rd88;
@%p149 bra $L__BB0_131;
ld.global.v4.u16 {%rs101, %rs102, %rs103, %rs104}, [%rd23];
// begin inline asm
{ cvt.f32.f16 %f1014, %rs101;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1015, %rs102;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1016, %rs103;}
// end inline asm
add.f32 %f1017, %f1014, 0f00000000;
add.f32 %f1018, %f1015, 0f00000000;
add.f32 %f1019, %f1016, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs100, %f1019;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs99, %f1018;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs98, %f1017;}
// end inline asm
mov.u16 %rs105, 0;
st.global.v4.u16 [%rd23], {%rs98, %rs99, %rs100, %rs105};
bra.uni $L__BB0_132;
$L__BB0_5:
ld.const.u64 %rd32, [params+432];
cvta.to.global.u64 %rd33, %rd32;
ld.const.u32 %r67, [params+424];
mad.lo.s32 %r68, %r67, %r56, %r55;
mul.wide.u32 %rd34, %r68, 12;
add.s64 %rd35, %rd33, %rd34;
ld.global.f32 %f266, [%rd35];
mul.f32 %f267, %f266, 0f3456BF95;
ld.global.f32 %f268, [%rd35+4];
mul.f32 %f269, %f268, 0f3456BF95;
ld.global.f32 %f270, [%rd35+8];
mul.f32 %f271, %f270, 0f3456BF95;
abs.f32 %f272, %f1105;
div.rn.f32 %f273, %f267, %f272;
abs.f32 %f274, %f1106;
div.rn.f32 %f275, %f269, %f274;
abs.f32 %f276, %f1107;
div.rn.f32 %f277, %f271, %f276;
abs.f32 %f278, %f273;
abs.f32 %f279, %f275;
abs.f32 %f280, %f277;
mov.f32 %f281, 0f38D1B717;
max.f32 %f282, %f278, %f281;
max.f32 %f283, %f279, %f281;
max.f32 %f284, %f280, %f281;
fma.rn.f32 %f10, %f1105, %f282, %f266;
fma.rn.f32 %f11, %f1106, %f283, %f268;
fma.rn.f32 %f12, %f1107, %f284, %f270;
setp.gt.f32 %p14, %f272, %f276;
neg.f32 %f285, %f1106;
selp.f32 %f286, %f285, 0f00000000, %p14;
mov.f32 %f1144, 0f00000000;
neg.f32 %f287, %f1107;
selp.f32 %f288, %f1105, %f287, %p14;
selp.f32 %f289, 0f00000000, %f1106, %p14;
mul.f32 %f290, %f288, %f288;
fma.rn.f32 %f291, %f286, %f286, %f290;
fma.rn.f32 %f292, %f289, %f289, %f291;
sqrt.rn.f32 %f293, %f292;
rcp.rn.f32 %f294, %f293;
mul.f32 %f13, %f286, %f294;
mul.f32 %f14, %f288, %f294;
mul.f32 %f15, %f289, %f294;
ld.const.u64 %rd36, [params+128];
cvta.to.global.u64 %rd37, %rd36;
ld.const.u32 %r69, [params+120];
mad.lo.s32 %r70, %r69, %r56, %r55;
mul.wide.u32 %rd38, %r70, 4;
add.s64 %rd3, %rd37, %rd38;
setp.lt.s32 %p15, %r3, 1;
mov.f32 %f1145, %f1144;
mov.f32 %f1146, %f1144;
mov.f32 %f1147, %f1144;
mov.f32 %f1148, %f1144;
mov.f32 %f1149, %f1144;
mov.f32 %f1150, %f1144;
mov.f32 %f1151, %f1144;
mov.f32 %f1152, %f1144;
mov.f32 %f1153, %f1144;
mov.f32 %f1154, %f1144;
mov.f32 %f1155, %f1144;
mov.f32 %f1156, %f1144;
mov.f32 %f1157, %f1144;
mov.f32 %f1158, %f1144;
@%p15 bra $L__BB0_36;
ld.const.u32 %r6, [params+588];
cvt.rn.f32.s32 %f310, %r3;
rcp.rn.f32 %f16, %f310;
ld.global.u32 %r351, [%rd3];
mul.f32 %f17, %f10, 0f3456BF95;
mul.f32 %f18, %f11, 0f3456BF95;
mul.f32 %f19, %f12, 0f3456BF95;
ld.const.u64 %rd4, [params+96];
mul.f32 %f311, %f1105, %f14;
mul.f32 %f312, %f1106, %f13;
sub.f32 %f20, %f312, %f311;
mul.f32 %f313, %f1107, %f13;
mul.f32 %f314, %f1105, %f15;
sub.f32 %f21, %f314, %f313;
mul.f32 %f315, %f1106, %f15;
mul.f32 %f316, %f1107, %f14;
sub.f32 %f22, %f316, %f315;
mov.u32 %r71, 0;
add.s64 %rd5, %rd1, 24;
mov.u64 %rd39, __cudart_i2opi_f;
abs.f32 %f379, %f17;
abs.f32 %f380, %f18;
max.f32 %f381, %f379, %f380;
abs.f32 %f382, %f19;
max.f32 %f383, %f381, %f382;
mov.u32 %r348, %r71;
$L__BB0_7:
mov.u32 %r350, %r71;
$L__BB0_8:
cvt.rn.f32.s32 %f1060, %r348;
mad.lo.s32 %r73, %r351, 1664525, 1013904223;
and.b32 %r74, %r73, 16777215;
cvt.rn.f32.u32 %f317, %r74;
fma.rn.f32 %f318, %f317, 0f33800000, %f1060;
mul.f32 %f54, %f16, %f318;
mad.lo.s32 %r351, %r73, 1664525, 1013904223;
and.b32 %r75, %r351, 16777215;
cvt.rn.f32.u32 %f319, %r75;
cvt.rn.f32.s32 %f320, %r350;
fma.rn.f32 %f321, %f319, 0f33800000, %f320;
mul.f32 %f322, %f16, %f321;
mul.f32 %f323, %f54, %f54;
mov.f32 %f324, 0f3F800000;
sub.f32 %f325, %f324, %f323;
mov.f32 %f326, 0f00000000;
max.f32 %f327, %f326, %f325;
sqrt.rn.f32 %f55, %f327;
mul.f32 %f56, %f322, 0f40C90FDB;
mul.f32 %f328, %f56, 0f3F22F983;
cvt.rni.s32.f32 %r355, %f328;
cvt.rn.f32.s32 %f329, %r355;
mov.f32 %f330, 0fBFC90FDA;
fma.rn.f32 %f331, %f329, %f330, %f56;
mov.f32 %f332, 0fB3A22168;
fma.rn.f32 %f333, %f329, %f332, %f331;
mov.f32 %f334, 0fA7C234C5;
fma.rn.f32 %f1141, %f329, %f334, %f333;
abs.f32 %f58, %f56;
setp.ltu.f32 %p16, %f58, 0f47CE4780;
mov.f32 %f1138, %f1141;
@%p16 bra $L__BB0_16;
setp.eq.f32 %p17, %f58, 0f7F800000;
@%p17 bra $L__BB0_15;
bra.uni $L__BB0_10;
$L__BB0_15:
mov.f32 %f337, 0f00000000;
mul.rn.f32 %f1138, %f56, %f337;
mov.u32 %r355, 0;
bra.uni $L__BB0_16;
$L__BB0_10:
mov.b32 %r14, %f56;
bfe.u32 %r77, %r14, 23, 8;
mov.u64 %rd106, 0;
mov.u32 %r352, 0;
mov.u64 %rd104, %rd1;
mov.u64 %rd105, %rd39;
$L__BB0_11:
.pragma "nounroll";
mov.b32 %r331, %f56;
shl.b32 %r330, %r331, 8;
or.b32 %r329, %r330, -2147483648;
ld.global.nc.u32 %r79, [%rd105];
mad.wide.u32 %rd41, %r79, %r329, %rd106;
shr.u64 %rd106, %rd41, 32;
st.local.u32 [%rd104], %rd41;
add.s64 %rd105, %rd105, 4;
add.s64 %rd104, %rd104, 4;
add.s32 %r352, %r352, 1;
setp.ne.s32 %p18, %r352, 6;
@%p18 bra $L__BB0_11;
add.s32 %r336, %r77, -128;
mov.b32 %r335, %f56;
bfe.u32 %r334, %r335, 23, 8;
add.s32 %r333, %r334, -128;
shr.u32 %r332, %r333, 5;
st.local.u32 [%rd5], %rd106;
mov.u32 %r81, 6;
sub.s32 %r82, %r81, %r332;
mul.wide.s32 %rd42, %r82, 4;
add.s64 %rd43, %rd1, %rd42;
ld.local.u32 %r353, [%rd43];
ld.local.u32 %r354, [%rd43+-4];
and.b32 %r23, %r333, 31;
setp.eq.s32 %p19, %r23, 0;
@%p19 bra $L__BB0_14;
mov.b32 %r344, %f56;
bfe.u32 %r343, %r344, 23, 8;
add.s32 %r342, %r343, -128;
shr.u32 %r341, %r342, 5;
mov.u32 %r340, 4;
sub.s32 %r339, %r340, %r341;
mov.u32 %r83, 32;
sub.s32 %r84, %r83, %r23;
shr.u32 %r85, %r354, %r84;
shl.b32 %r86, %r353, %r23;
add.s32 %r353, %r85, %r86;
mul.wide.s32 %rd44, %r339, 4;
add.s64 %rd45, %rd1, %rd44;
ld.local.u32 %r87, [%rd45];
shr.u32 %r88, %r87, %r84;
shl.b32 %r89, %r354, %r23;
add.s32 %r354, %r88, %r89;
$L__BB0_14:
mov.b32 %r337, %f56;
and.b32 %r90, %r337, -2147483648;
shr.u32 %r91, %r354, 30;
shl.b32 %r92, %r353, 2;
or.b32 %r93, %r91, %r92;
shr.u32 %r94, %r93, 31;
shr.u32 %r95, %r353, 30;
add.s32 %r96, %r94, %r95;
neg.s32 %r97, %r96;
setp.eq.s32 %p20, %r90, 0;
selp.b32 %r355, %r96, %r97, %p20;
setp.ne.s32 %p21, %r94, 0;
xor.b32 %r98, %r90, -2147483648;
selp.b32 %r99, %r98, %r90, %p21;
selp.b32 %r100, -1, 0, %p21;
xor.b32 %r101, %r93, %r100;
shl.b32 %r102, %r354, 2;
xor.b32 %r103, %r102, %r100;
cvt.u64.u32 %rd46, %r101;
cvt.u64.u32 %rd47, %r103;
bfi.b64 %rd48, %rd46, %rd47, 32, 32;
cvt.rn.f64.s64 %fd1, %rd48;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f335, %fd2;
setp.eq.s32 %p22, %r99, 0;
neg.f32 %f336, %f335;
selp.f32 %f1138, %f335, %f336, %p22;
$L__BB0_16:
add.s32 %r30, %r355, 1;
and.b32 %r31, %r30, 1;
setp.eq.s32 %p23, %r31, 0;
selp.f32 %f62, %f1138, 0f3F800000, %p23;
mul.rn.f32 %f63, %f1138, %f1138;
mov.f32 %f1139, 0fB94D4153;
@%p23 bra $L__BB0_18;
mov.f32 %f339, 0fBAB607ED;
mov.f32 %f340, 0f37CBAC00;
fma.rn.f32 %f1139, %f340, %f63, %f339;
$L__BB0_18:
selp.f32 %f341, 0f3C0885E4, 0f3D2AAABB, %p23;
fma.rn.f32 %f342, %f1139, %f63, %f341;
selp.f32 %f343, 0fBE2AAAA8, 0fBEFFFFFF, %p23;
fma.rn.f32 %f344, %f342, %f63, %f343;
mov.f32 %f345, 0f00000000;
fma.rn.f32 %f346, %f63, %f62, %f345;
fma.rn.f32 %f1140, %f344, %f346, %f62;
and.b32 %r105, %r30, 2;
setp.eq.s32 %p25, %r105, 0;
@%p25 bra $L__BB0_20;
mov.f32 %f348, 0fBF800000;
fma.rn.f32 %f1140, %f1140, %f348, %f345;
$L__BB0_20:
mul.f32 %f1059, %f56, 0f3F22F983;
cvt.rni.s32.f32 %r358, %f1059;
setp.ltu.f32 %p154, %f58, 0f47CE4780;
@%p154 bra $L__BB0_28;
setp.eq.f32 %p27, %f58, 0f7F800000;
@%p27 bra $L__BB0_27;
bra.uni $L__BB0_22;
$L__BB0_27:
mov.f32 %f351, 0f00000000;
mul.rn.f32 %f1141, %f56, %f351;
mov.u32 %r358, 0;
bra.uni $L__BB0_28;
$L__BB0_22:
mov.b32 %r32, %f56;
bfe.u32 %r106, %r32, 23, 8;
add.s32 %r33, %r106, -128;
shl.b32 %r107, %r32, 8;
or.b32 %r34, %r107, -2147483648;
shr.u32 %r35, %r33, 5;
mov.u64 %rd107, 0;
mov.u64 %rd108, %rd107;
$L__BB0_23:
.pragma "nounroll";
shl.b64 %rd51, %rd107, 2;
mov.u64 %rd52, __cudart_i2opi_f;
add.s64 %rd53, %rd52, %rd51;
ld.global.nc.u32 %r108, [%rd53];
mad.wide.u32 %rd54, %r108, %r34, %rd108;
shr.u64 %rd108, %rd54, 32;
add.s64 %rd55, %rd1, %rd51;
st.local.u32 [%rd55], %rd54;
cvt.u32.u64 %r109, %rd107;
add.s32 %r110, %r109, 1;
cvt.s64.s32 %rd107, %r110;
setp.ne.s32 %p28, %r110, 6;
@%p28 bra $L__BB0_23;
st.local.u32 [%rd5], %rd108;
mov.u32 %r111, 4;
sub.s32 %r36, %r111, %r35;
mov.u32 %r112, 6;
sub.s32 %r113, %r112, %r35;
mul.wide.s32 %rd56, %r113, 4;
add.s64 %rd57, %rd1, %rd56;
ld.local.u32 %r356, [%rd57];
ld.local.u32 %r357, [%rd57+-4];
and.b32 %r39, %r33, 31;
setp.eq.s32 %p29, %r39, 0;
@%p29 bra $L__BB0_26;
mov.u32 %r114, 32;
sub.s32 %r115, %r114, %r39;
shr.u32 %r116, %r357, %r115;
shl.b32 %r117, %r356, %r39;
add.s32 %r356, %r116, %r117;
mul.wide.s32 %rd58, %r36, 4;
add.s64 %rd59, %rd1, %rd58;
ld.local.u32 %r118, [%rd59];
shr.u32 %r119, %r118, %r115;
shl.b32 %r120, %r357, %r39;
add.s32 %r357, %r119, %r120;
$L__BB0_26:
and.b32 %r121, %r32, -2147483648;
shr.u32 %r122, %r357, 30;
shl.b32 %r123, %r356, 2;
or.b32 %r124, %r122, %r123;
shr.u32 %r125, %r124, 31;
shr.u32 %r126, %r356, 30;
add.s32 %r127, %r125, %r126;
neg.s32 %r128, %r127;
setp.eq.s32 %p30, %r121, 0;
selp.b32 %r358, %r127, %r128, %p30;
setp.ne.s32 %p31, %r125, 0;
xor.b32 %r129, %r121, -2147483648;
selp.b32 %r130, %r129, %r121, %p31;
selp.b32 %r131, -1, 0, %p31;
xor.b32 %r132, %r124, %r131;
shl.b32 %r133, %r357, 2;
xor.b32 %r134, %r133, %r131;
cvt.u64.u32 %rd60, %r132;
cvt.u64.u32 %rd61, %r134;
bfi.b64 %rd62, %rd60, %rd61, 32, 32;
cvt.rn.f64.s64 %fd3, %rd62;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f349, %fd4;
setp.eq.s32 %p32, %r130, 0;
neg.f32 %f350, %f349;
selp.f32 %f1141, %f349, %f350, %p32;
$L__BB0_28:
mul.f32 %f72, %f55, %f1140;
and.b32 %r46, %r358, 1;
setp.eq.s32 %p33, %r46, 0;
mul.rn.f32 %f74, %f1141, %f1141;
mov.f32 %f1142, 0fB94D4153;
@%p33 bra $L__BB0_30;
mov.f32 %f353, 0fBAB607ED;
mov.f32 %f354, 0f37CBAC00;
fma.rn.f32 %f1142, %f354, %f74, %f353;
$L__BB0_30:
selp.f32 %f1061, %f1141, 0f3F800000, %p33;
selp.f32 %f355, 0f3C0885E4, 0f3D2AAABB, %p33;
fma.rn.f32 %f356, %f1142, %f74, %f355;
selp.f32 %f357, 0fBE2AAAA8, 0fBEFFFFFF, %p33;
fma.rn.f32 %f358, %f356, %f74, %f357;
mov.f32 %f359, 0f00000000;
fma.rn.f32 %f360, %f74, %f1061, %f359;
fma.rn.f32 %f1143, %f358, %f360, %f1061;
and.b32 %r136, %r358, 2;
setp.eq.s32 %p35, %r136, 0;
@%p35 bra $L__BB0_32;
mov.f32 %f362, 0fBF800000;
fma.rn.f32 %f1143, %f1143, %f362, %f359;
$L__BB0_32:
mul.f32 %f363, %f55, %f1143;
mul.f32 %f364, %f13, %f363;
mul.f32 %f365, %f14, %f363;
mul.f32 %f366, %f15, %f363;
fma.rn.f32 %f367, %f22, %f72, %f364;
fma.rn.f32 %f368, %f21, %f72, %f365;
fma.rn.f32 %f369, %f20, %f72, %f366;
fma.rn.f32 %f80, %f1105, %f54, %f367;
fma.rn.f32 %f81, %f1106, %f54, %f368;
fma.rn.f32 %f82, %f1107, %f54, %f369;
setp.leu.f32 %p36, %f81, 0f00000000;
setp.ne.s32 %p37, %r6, 0;
and.pred %p38, %p37, %p36;
@%p38 bra $L__BB0_34;
mov.f32 %f384, 0f38D1B717;
max.f32 %f376, %f383, %f384;
mov.f32 %f377, 0f6C4ECB8F;
mov.f32 %f378, 0f00000000;
mov.u32 %r173, 2;
mov.u32 %r174, 1;
mov.u32 %r175, 3;
mov.u32 %r178, 1065353216;
mov.u32 %r207, 0;
// begin inline asm
call(%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167,%r168),_optix_trace_typed_32,(%r207,%rd4,%f10,%f11,%f12,%f80,%f81,%f82,%f376,%f377,%f378,%r174,%r207,%r174,%r173,%r174,%r175,%r178,%r178,%r178,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207);
// end inline asm
mov.b32 %f385, %r137;
mov.b32 %f386, %r138;
mov.b32 %f387, %r139;
mul.f32 %f388, %f1106, %f81;
fma.rn.f32 %f389, %f1105, %f80, %f388;
fma.rn.f32 %f390, %f1107, %f82, %f389;
mul.f32 %f391, %f390, 0f40800000;
cvt.sat.f32.f32 %f392, %f391;
mul.f32 %f393, %f392, %f385;
mul.f32 %f394, %f392, %f386;
mul.f32 %f395, %f392, %f387;
fma.rn.f32 %f1150, %f80, %f393, %f1150;
fma.rn.f32 %f1151, %f80, %f394, %f1151;
fma.rn.f32 %f1152, %f80, %f395, %f1152;
fma.rn.f32 %f1147, %f81, %f393, %f1147;
fma.rn.f32 %f1148, %f81, %f394, %f1148;
fma.rn.f32 %f1149, %f81, %f395, %f1149;
fma.rn.f32 %f1144, %f82, %f393, %f1144;
fma.rn.f32 %f1145, %f82, %f394, %f1145;
fma.rn.f32 %f1146, %f82, %f395, %f1146;
add.f32 %f1153, %f1153, %f393;
add.f32 %f1154, %f1154, %f394;
add.f32 %f1155, %f1155, %f395;
cvt.sat.f32.f32 %f396, %f390;
fma.rn.f32 %f1156, %f396, %f385, %f1156;
fma.rn.f32 %f1157, %f396, %f386, %f1157;
fma.rn.f32 %f1158, %f396, %f387, %f1158;
$L__BB0_34:
add.s32 %r350, %r350, 1;
setp.lt.s32 %p39, %r350, %r3;
@%p39 bra $L__BB0_8;
add.s32 %r348, %r348, 1;
setp.lt.s32 %p40, %r348, %r3;
@%p40 bra $L__BB0_7;
$L__BB0_36:
mul.lo.s32 %r208, %r3, %r3;
cvt.rn.f32.s32 %f397, %r208;
rcp.rn.f32 %f398, %f397;
mul.f32 %f399, %f398, %f1156;
mul.f32 %f400, %f398, %f1157;
mul.f32 %f401, %f398, %f1158;
mul.f32 %f128, %f398, %f1153;
mul.f32 %f129, %f398, %f1154;
mul.f32 %f130, %f398, %f1155;
mul.f32 %f131, %f398, %f1150;
mul.f32 %f132, %f398, %f1151;
mul.f32 %f133, %f398, %f1152;
mul.f32 %f134, %f398, %f1147;
mul.f32 %f135, %f398, %f1148;
mul.f32 %f136, %f398, %f1149;
mul.f32 %f137, %f398, %f1144;
mul.f32 %f138, %f398, %f1145;
mul.f32 %f139, %f398, %f1146;
fma.rn.f32 %f402, %f398, %f1156, %f399;
fma.rn.f32 %f403, %f398, %f1157, %f400;
fma.rn.f32 %f404, %f398, %f1158, %f401;
ld.const.v4.f32 {%f405, %f406, %f407, %f408}, [params+576];
mul.f32 %f143, %f402, %f405;
mul.f32 %f144, %f403, %f406;
mul.f32 %f145, %f404, %f407;
ld.const.u32 %r49, [params+104];
and.b32 %r209, %r49, 1;
setp.eq.b32 %p41, %r209, 1;
mov.pred %p42, 0;
xor.pred %p43, %p41, %p42;
not.pred %p44, %p43;
@%p44 bra $L__BB0_110;
mov.f32 %f410, 0f3E666666;
cvt.rzi.f32.f32 %f411, %f410;
add.f32 %f412, %f411, %f411;
mov.f32 %f413, 0f3EE66666;
sub.f32 %f414, %f413, %f412;
abs.f32 %f146, %f414;
abs.f32 %f147, %f143;
setp.lt.f32 %p45, %f147, 0f00800000;
mul.f32 %f415, %f147, 0f4B800000;
selp.f32 %f416, %f415, %f147, %p45;
selp.f32 %f417, 0fC3170000, 0fC2FE0000, %p45;
mov.b32 %r210, %f416;
and.b32 %r211, %r210, 8388607;
or.b32 %r212, %r211, 1065353216;
mov.b32 %f418, %r212;
shr.u32 %r213, %r210, 23;
cvt.rn.f32.u32 %f419, %r213;
add.f32 %f420, %f417, %f419;
setp.gt.f32 %p46, %f418, 0f3FB504F3;
mul.f32 %f421, %f418, 0f3F000000;
add.f32 %f422, %f420, 0f3F800000;
selp.f32 %f423, %f422, %f420, %p46;
selp.f32 %f424, %f421, %f418, %p46;
add.f32 %f425, %f424, 0fBF800000;
add.f32 %f426, %f424, 0f3F800000;
rcp.approx.ftz.f32 %f427, %f426;
add.f32 %f428, %f425, %f425;
mul.f32 %f429, %f428, %f427;
mul.f32 %f430, %f429, %f429;
mov.f32 %f431, 0f3C4CAF63;
mov.f32 %f432, 0f3B18F0FE;
fma.rn.f32 %f433, %f432, %f430, %f431;
mov.f32 %f434, 0f3DAAAABD;
fma.rn.f32 %f435, %f433, %f430, %f434;
mul.rn.f32 %f436, %f435, %f430;
mul.rn.f32 %f437, %f436, %f429;
sub.f32 %f438, %f425, %f429;
add.f32 %f439, %f438, %f438;
neg.f32 %f440, %f429;
fma.rn.f32 %f441, %f440, %f425, %f439;
mul.rn.f32 %f442, %f427, %f441;
add.f32 %f443, %f437, %f429;
sub.f32 %f444, %f429, %f443;
add.f32 %f445, %f437, %f444;
add.f32 %f446, %f442, %f445;
add.f32 %f447, %f443, %f446;
sub.f32 %f448, %f443, %f447;
add.f32 %f449, %f446, %f448;
mov.f32 %f450, 0f3F317200;
mul.rn.f32 %f451, %f423, %f450;
mov.f32 %f452, 0f35BFBE8E;
mul.rn.f32 %f453, %f423, %f452;
add.f32 %f454, %f451, %f447;
sub.f32 %f455, %f451, %f454;
add.f32 %f456, %f447, %f455;
add.f32 %f457, %f449, %f456;
add.f32 %f458, %f453, %f457;
add.f32 %f459, %f454, %f458;
sub.f32 %f460, %f454, %f459;
add.f32 %f461, %f458, %f460;
mul.rn.f32 %f462, %f413, %f459;
neg.f32 %f463, %f462;
fma.rn.f32 %f464, %f413, %f459, %f463;
fma.rn.f32 %f465, %f413, %f461, %f464;
mov.f32 %f466, 0f00000000;
fma.rn.f32 %f467, %f466, %f459, %f465;
add.rn.f32 %f468, %f462, %f467;
neg.f32 %f469, %f468;
add.rn.f32 %f470, %f462, %f469;
add.rn.f32 %f471, %f470, %f467;
mov.b32 %r214, %f468;
setp.eq.s32 %p47, %r214, 1118925336;
add.s32 %r215, %r214, -1;
mov.b32 %f472, %r215;
add.f32 %f473, %f471, 0f37000000;
selp.f32 %f148, %f473, %f471, %p47;
selp.f32 %f474, %f472, %f468, %p47;
mov.f32 %f475, 0f3FB8AA3B;
mul.rn.f32 %f476, %f474, %f475;
cvt.rzi.f32.f32 %f477, %f476;
abs.f32 %f478, %f477;
setp.gt.f32 %p48, %f478, 0f42FC0000;
mov.b32 %r216, %f477;
and.b32 %r217, %r216, -2147483648;
or.b32 %r218, %r217, 1123811328;
mov.b32 %f479, %r218;
selp.f32 %f480, %f479, %f477, %p48;
mov.f32 %f481, 0fBF317218;
fma.rn.f32 %f482, %f480, %f481, %f474;
mov.f32 %f483, 0f3102E308;
fma.rn.f32 %f484, %f480, %f483, %f482;
mul.f32 %f485, %f484, 0f3FB8AA3B;
add.f32 %f486, %f480, 0f4B40007F;
mov.b32 %r219, %f486;
shl.b32 %r220, %r219, 23;
mov.b32 %f487, %r220;
ex2.approx.ftz.f32 %f488, %f485;
mul.f32 %f149, %f488, %f487;
setp.eq.f32 %p49, %f149, 0f7F800000;
mov.f32 %f1174, 0f7F800000;
@%p49 bra $L__BB0_39;
fma.rn.f32 %f1174, %f149, %f148, %f149;
$L__BB0_39:
setp.lt.f32 %p50, %f143, 0f00000000;
setp.eq.f32 %p51, %f146, 0f3F800000;
and.pred %p1, %p50, %p51;
setp.eq.f32 %p52, %f143, 0f00000000;
@%p52 bra $L__BB0_43;
bra.uni $L__BB0_40;
$L__BB0_43:
add.f32 %f493, %f143, %f143;
selp.f32 %f1176, %f493, 0f00000000, %p51;
bra.uni $L__BB0_44;
$L__BB0_131:
mov.f32 %f1022, 0f00000000;
mov.u32 %r360, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs108, %f1022;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs107, %f1022;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs106, %f1022;}
// end inline asm
mov.u16 %rs109, 0;
st.global.v4.u16 [%rd23], {%rs106, %rs107, %rs108, %rs109};
$L__BB0_132:
ld.const.u64 %rd89, [params+256];
cvta.to.global.u64 %rd90, %rd89;
ld.const.u32 %r321, [params+248];
mad.lo.s32 %r322, %r321, %r5, %r4;
mul.wide.u32 %rd91, %r322, 8;
add.s64 %rd24, %rd90, %rd91;
setp.eq.s32 %p150, %r360, 0;
@%p150 bra $L__BB0_134;
ld.global.v4.u16 {%rs116, %rs117, %rs118, %rs119}, [%rd24];
// begin inline asm
{ cvt.f32.f16 %f1023, %rs116;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1024, %rs117;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1025, %rs118;}
// end inline asm
add.f32 %f1026, %f1023, 0f00000000;
add.f32 %f1027, %f1024, 0f00000000;
add.f32 %f1028, %f1025, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs115, %f1028;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs114, %f1027;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f1026;}
// end inline asm
mov.u16 %rs120, 0;
st.global.v4.u16 [%rd24], {%rs113, %rs114, %rs115, %rs120};
bra.uni $L__BB0_135;
$L__BB0_134:
mov.f32 %f1031, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs123, %f1031;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs122, %f1031;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs121, %f1031;}
// end inline asm
mov.u16 %rs124, 0;
st.global.v4.u16 [%rd24], {%rs121, %rs122, %rs123, %rs124};
$L__BB0_135:
ld.const.u64 %rd92, [params+272];
cvta.to.global.u64 %rd93, %rd92;
ld.const.u32 %r323, [params+264];
mad.lo.s32 %r324, %r323, %r5, %r4;
mul.wide.u32 %rd94, %r324, 8;
add.s64 %rd25, %rd93, %rd94;
@%p150 bra $L__BB0_137;
ld.global.v4.u16 {%rs131, %rs132, %rs133, %rs134}, [%rd25];
// begin inline asm
{ cvt.f32.f16 %f1032, %rs131;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1033, %rs132;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1034, %rs133;}
// end inline asm
add.f32 %f1035, %f1032, 0f00000000;
add.f32 %f1036, %f1033, 0f00000000;
add.f32 %f1037, %f1034, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs130, %f1037;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs129, %f1036;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1035;}
// end inline asm
mov.u16 %rs135, 0;
st.global.v4.u16 [%rd25], {%rs128, %rs129, %rs130, %rs135};
bra.uni $L__BB0_138;
$L__BB0_137:
mov.f32 %f1040, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs138, %f1040;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs137, %f1040;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs136, %f1040;}
// end inline asm
mov.u16 %rs139, 0;
st.global.v4.u16 [%rd25], {%rs136, %rs137, %rs138, %rs139};
$L__BB0_138:
ld.const.u64 %rd95, [params+288];
cvta.to.global.u64 %rd96, %rd95;
ld.const.u32 %r325, [params+280];
mad.lo.s32 %r326, %r325, %r5, %r4;
mul.wide.u32 %rd97, %r326, 8;
add.s64 %rd26, %rd96, %rd97;
@%p150 bra $L__BB0_140;
ld.global.v4.u16 {%rs146, %rs147, %rs148, %rs149}, [%rd26];
// begin inline asm
{ cvt.f32.f16 %f1041, %rs146;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1042, %rs147;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1043, %rs148;}
// end inline asm
add.f32 %f1044, %f1041, 0f00000000;
add.f32 %f1045, %f1042, 0f00000000;
add.f32 %f1046, %f1043, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs145, %f1046;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs144, %f1045;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1044;}
// end inline asm
mov.u16 %rs150, 0;
st.global.v4.u16 [%rd26], {%rs143, %rs144, %rs145, %rs150};
bra.uni $L__BB0_141;
$L__BB0_140:
mov.f32 %f1049, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs153, %f1049;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs152, %f1049;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs151, %f1049;}
// end inline asm
mov.u16 %rs154, 0;
st.global.v4.u16 [%rd26], {%rs151, %rs152, %rs153, %rs154};
$L__BB0_141:
ld.const.u64 %rd98, [params+304];
cvta.to.global.u64 %rd99, %rd98;
ld.const.u32 %r327, [params+296];
mad.lo.s32 %r328, %r327, %r5, %r4;
mul.wide.u32 %rd100, %r328, 8;
add.s64 %rd27, %rd99, %rd100;
@%p150 bra $L__BB0_143;
ld.global.v4.u16 {%rs161, %rs162, %rs163, %rs164}, [%rd27];
// begin inline asm
{ cvt.f32.f16 %f1050, %rs161;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1051, %rs162;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1052, %rs163;}
// end inline asm
add.f32 %f1053, %f1050, 0f00000000;
add.f32 %f1054, %f1051, 0f00000000;
add.f32 %f1055, %f1052, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs160, %f1055;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs159, %f1054;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1053;}
// end inline asm
mov.u16 %rs165, 0;
st.global.v4.u16 [%rd27], {%rs158, %rs159, %rs160, %rs165};
bra.uni $L__BB0_144;
$L__BB0_143:
mov.f32 %f1058, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs168, %f1058;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs167, %f1058;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs166, %f1058;}
// end inline asm
mov.u16 %rs169, 0;
st.global.v4.u16 [%rd27], {%rs166, %rs167, %rs168, %rs169};
bra.uni $L__BB0_144;
$L__BB0_40:
mov.b32 %r221, %f1174;
xor.b32 %r222, %r221, -2147483648;
mov.b32 %f489, %r222;
selp.f32 %f1176, %f489, %f1174, %p1;
setp.geu.f32 %p53, %f143, 0f00000000;
@%p53 bra $L__BB0_44;
mov.f32 %f490, 0f3EE66666;
cvt.rzi.f32.f32 %f491, %f490;
setp.eq.f32 %p54, %f491, 0f3EE66666;
@%p54 bra $L__BB0_44;
mov.f32 %f1176, 0f7FFFFFFF;
$L__BB0_44:
add.f32 %f494, %f147, 0f3EE66666;
mov.b32 %r223, %f494;
setp.lt.s32 %p56, %r223, 2139095040;
@%p56 bra $L__BB0_49;
setp.gtu.f32 %p57, %f147, 0f7F800000;
@%p57 bra $L__BB0_48;
bra.uni $L__BB0_46;
$L__BB0_48:
add.f32 %f1176, %f143, 0f3EE66666;
bra.uni $L__BB0_49;
$L__BB0_46:
setp.neu.f32 %p58, %f147, 0f7F800000;
@%p58 bra $L__BB0_49;
selp.f32 %f1176, 0fFF800000, 0f7F800000, %p1;
$L__BB0_49:
setp.eq.f32 %p59, %f143, 0f3F800000;
selp.f32 %f158, 0f3F800000, %f1176, %p59;
abs.f32 %f159, %f144;
setp.lt.f32 %p60, %f159, 0f00800000;
mul.f32 %f496, %f159, 0f4B800000;
selp.f32 %f497, %f496, %f159, %p60;
selp.f32 %f498, 0fC3170000, 0fC2FE0000, %p60;
mov.b32 %r224, %f497;
and.b32 %r225, %r224, 8388607;
or.b32 %r226, %r225, 1065353216;
mov.b32 %f499, %r226;
shr.u32 %r227, %r224, 23;
cvt.rn.f32.u32 %f500, %r227;
add.f32 %f501, %f498, %f500;
setp.gt.f32 %p61, %f499, 0f3FB504F3;
mul.f32 %f502, %f499, 0f3F000000;
add.f32 %f503, %f501, 0f3F800000;
selp.f32 %f504, %f503, %f501, %p61;
selp.f32 %f505, %f502, %f499, %p61;
add.f32 %f506, %f505, 0fBF800000;
add.f32 %f507, %f505, 0f3F800000;
rcp.approx.ftz.f32 %f508, %f507;
add.f32 %f509, %f506, %f506;
mul.f32 %f510, %f509, %f508;
mul.f32 %f511, %f510, %f510;
mov.f32 %f512, 0f3C4CAF63;
mov.f32 %f513, 0f3B18F0FE;
fma.rn.f32 %f514, %f513, %f511, %f512;
mov.f32 %f515, 0f3DAAAABD;
fma.rn.f32 %f516, %f514, %f511, %f515;
mul.rn.f32 %f517, %f516, %f511;
mul.rn.f32 %f518, %f517, %f510;
sub.f32 %f519, %f506, %f510;
add.f32 %f520, %f519, %f519;
neg.f32 %f521, %f510;
fma.rn.f32 %f522, %f521, %f506, %f520;
mul.rn.f32 %f523, %f508, %f522;
add.f32 %f524, %f518, %f510;
sub.f32 %f525, %f510, %f524;
add.f32 %f526, %f518, %f525;
add.f32 %f527, %f523, %f526;
add.f32 %f528, %f524, %f527;
sub.f32 %f529, %f524, %f528;
add.f32 %f530, %f527, %f529;
mov.f32 %f531, 0f3F317200;
mul.rn.f32 %f532, %f504, %f531;
mov.f32 %f533, 0f35BFBE8E;
mul.rn.f32 %f534, %f504, %f533;
add.f32 %f535, %f532, %f528;
sub.f32 %f536, %f532, %f535;
add.f32 %f537, %f528, %f536;
add.f32 %f538, %f530, %f537;
add.f32 %f539, %f534, %f538;
add.f32 %f540, %f535, %f539;
sub.f32 %f541, %f535, %f540;
add.f32 %f542, %f539, %f541;
mov.f32 %f543, 0f3EE66666;
mul.rn.f32 %f544, %f543, %f540;
neg.f32 %f545, %f544;
fma.rn.f32 %f546, %f543, %f540, %f545;
fma.rn.f32 %f547, %f543, %f542, %f546;
mov.f32 %f548, 0f00000000;
fma.rn.f32 %f549, %f548, %f540, %f547;
add.rn.f32 %f550, %f544, %f549;
neg.f32 %f551, %f550;
add.rn.f32 %f552, %f544, %f551;
add.rn.f32 %f553, %f552, %f549;
mov.b32 %r228, %f550;
setp.eq.s32 %p62, %r228, 1118925336;
add.s32 %r229, %r228, -1;
mov.b32 %f554, %r229;
add.f32 %f555, %f553, 0f37000000;
selp.f32 %f160, %f555, %f553, %p62;
selp.f32 %f556, %f554, %f550, %p62;
mov.f32 %f557, 0f3FB8AA3B;
mul.rn.f32 %f558, %f556, %f557;
cvt.rzi.f32.f32 %f559, %f558;
abs.f32 %f560, %f559;
setp.gt.f32 %p63, %f560, 0f42FC0000;
mov.b32 %r230, %f559;
and.b32 %r231, %r230, -2147483648;
or.b32 %r232, %r231, 1123811328;
mov.b32 %f561, %r232;
selp.f32 %f562, %f561, %f559, %p63;
mov.f32 %f563, 0fBF317218;
fma.rn.f32 %f564, %f562, %f563, %f556;
mov.f32 %f565, 0f3102E308;
fma.rn.f32 %f566, %f562, %f565, %f564;
mul.f32 %f567, %f566, 0f3FB8AA3B;
add.f32 %f568, %f562, 0f4B40007F;
mov.b32 %r233, %f568;
shl.b32 %r234, %r233, 23;
mov.b32 %f569, %r234;
ex2.approx.ftz.f32 %f570, %f567;
mul.f32 %f161, %f570, %f569;
setp.eq.f32 %p64, %f161, 0f7F800000;
mov.f32 %f1177, 0f7F800000;
@%p64 bra $L__BB0_51;
fma.rn.f32 %f1177, %f161, %f160, %f161;
$L__BB0_51:
setp.lt.f32 %p65, %f144, 0f00000000;
and.pred %p2, %p65, %p51;
setp.eq.f32 %p67, %f144, 0f00000000;
@%p67 bra $L__BB0_55;
bra.uni $L__BB0_52;
$L__BB0_55:
add.f32 %f575, %f144, %f144;
selp.f32 %f1179, %f575, 0f00000000, %p51;
bra.uni $L__BB0_56;
$L__BB0_52:
mov.b32 %r235, %f1177;
xor.b32 %r236, %r235, -2147483648;
mov.b32 %f571, %r236;
selp.f32 %f1179, %f571, %f1177, %p2;
setp.geu.f32 %p68, %f144, 0f00000000;
@%p68 bra $L__BB0_56;
mov.f32 %f572, 0f3EE66666;
cvt.rzi.f32.f32 %f573, %f572;
setp.eq.f32 %p69, %f573, 0f3EE66666;
@%p69 bra $L__BB0_56;
mov.f32 %f1179, 0f7FFFFFFF;
$L__BB0_56:
add.f32 %f576, %f159, 0f3EE66666;
mov.b32 %r237, %f576;
setp.lt.s32 %p71, %r237, 2139095040;
@%p71 bra $L__BB0_61;
setp.gtu.f32 %p72, %f159, 0f7F800000;
@%p72 bra $L__BB0_60;
bra.uni $L__BB0_58;
$L__BB0_60:
add.f32 %f1179, %f144, 0f3EE66666;
bra.uni $L__BB0_61;
$L__BB0_58:
setp.neu.f32 %p73, %f159, 0f7F800000;
@%p73 bra $L__BB0_61;
selp.f32 %f1179, 0fFF800000, 0f7F800000, %p2;
$L__BB0_61:
setp.eq.f32 %p74, %f144, 0f3F800000;
selp.f32 %f170, 0f3F800000, %f1179, %p74;
abs.f32 %f171, %f145;
setp.lt.f32 %p75, %f171, 0f00800000;
mul.f32 %f578, %f171, 0f4B800000;
selp.f32 %f579, %f578, %f171, %p75;
selp.f32 %f580, 0fC3170000, 0fC2FE0000, %p75;
mov.b32 %r238, %f579;
and.b32 %r239, %r238, 8388607;
or.b32 %r240, %r239, 1065353216;
mov.b32 %f581, %r240;
shr.u32 %r241, %r238, 23;
cvt.rn.f32.u32 %f582, %r241;
add.f32 %f583, %f580, %f582;
setp.gt.f32 %p76, %f581, 0f3FB504F3;
mul.f32 %f584, %f581, 0f3F000000;
add.f32 %f585, %f583, 0f3F800000;
selp.f32 %f586, %f585, %f583, %p76;
selp.f32 %f587, %f584, %f581, %p76;
add.f32 %f588, %f587, 0fBF800000;
add.f32 %f589, %f587, 0f3F800000;
rcp.approx.ftz.f32 %f590, %f589;
add.f32 %f591, %f588, %f588;
mul.f32 %f592, %f591, %f590;
mul.f32 %f593, %f592, %f592;
mov.f32 %f594, 0f3C4CAF63;
mov.f32 %f595, 0f3B18F0FE;
fma.rn.f32 %f596, %f595, %f593, %f594;
mov.f32 %f597, 0f3DAAAABD;
fma.rn.f32 %f598, %f596, %f593, %f597;
mul.rn.f32 %f599, %f598, %f593;
mul.rn.f32 %f600, %f599, %f592;
sub.f32 %f601, %f588, %f592;
add.f32 %f602, %f601, %f601;
neg.f32 %f603, %f592;
fma.rn.f32 %f604, %f603, %f588, %f602;
mul.rn.f32 %f605, %f590, %f604;
add.f32 %f606, %f600, %f592;
sub.f32 %f607, %f592, %f606;
add.f32 %f608, %f600, %f607;
add.f32 %f609, %f605, %f608;
add.f32 %f610, %f606, %f609;
sub.f32 %f611, %f606, %f610;
add.f32 %f612, %f609, %f611;
mov.f32 %f613, 0f3F317200;
mul.rn.f32 %f614, %f586, %f613;
mov.f32 %f615, 0f35BFBE8E;
mul.rn.f32 %f616, %f586, %f615;
add.f32 %f617, %f614, %f610;
sub.f32 %f618, %f614, %f617;
add.f32 %f619, %f610, %f618;
add.f32 %f620, %f612, %f619;
add.f32 %f621, %f616, %f620;
add.f32 %f622, %f617, %f621;
sub.f32 %f623, %f617, %f622;
add.f32 %f624, %f621, %f623;
mov.f32 %f625, 0f3EE66666;
mul.rn.f32 %f626, %f625, %f622;
neg.f32 %f627, %f626;
fma.rn.f32 %f628, %f625, %f622, %f627;
fma.rn.f32 %f629, %f625, %f624, %f628;
mov.f32 %f630, 0f00000000;
fma.rn.f32 %f631, %f630, %f622, %f629;
add.rn.f32 %f632, %f626, %f631;
neg.f32 %f633, %f632;
add.rn.f32 %f634, %f626, %f633;
add.rn.f32 %f635, %f634, %f631;
mov.b32 %r242, %f632;
setp.eq.s32 %p77, %r242, 1118925336;
add.s32 %r243, %r242, -1;
mov.b32 %f636, %r243;
add.f32 %f637, %f635, 0f37000000;
selp.f32 %f172, %f637, %f635, %p77;
selp.f32 %f638, %f636, %f632, %p77;
mov.f32 %f639, 0f3FB8AA3B;
mul.rn.f32 %f640, %f638, %f639;
cvt.rzi.f32.f32 %f641, %f640;
abs.f32 %f642, %f641;
setp.gt.f32 %p78, %f642, 0f42FC0000;
mov.b32 %r244, %f641;
and.b32 %r245, %r244, -2147483648;
or.b32 %r246, %r245, 1123811328;
mov.b32 %f643, %r246;
selp.f32 %f644, %f643, %f641, %p78;
mov.f32 %f645, 0fBF317218;
fma.rn.f32 %f646, %f644, %f645, %f638;
mov.f32 %f647, 0f3102E308;
fma.rn.f32 %f648, %f644, %f647, %f646;
mul.f32 %f649, %f648, 0f3FB8AA3B;
add.f32 %f650, %f644, 0f4B40007F;
mov.b32 %r247, %f650;
shl.b32 %r248, %r247, 23;
mov.b32 %f651, %r248;
ex2.approx.ftz.f32 %f652, %f649;
mul.f32 %f173, %f652, %f651;
setp.eq.f32 %p79, %f173, 0f7F800000;
mov.f32 %f1180, 0f7F800000;
@%p79 bra $L__BB0_63;
fma.rn.f32 %f1180, %f173, %f172, %f173;
$L__BB0_63:
setp.lt.f32 %p80, %f145, 0f00000000;
and.pred %p3, %p80, %p51;
setp.eq.f32 %p82, %f145, 0f00000000;
@%p82 bra $L__BB0_67;
bra.uni $L__BB0_64;
$L__BB0_67:
add.f32 %f657, %f145, %f145;
selp.f32 %f1182, %f657, 0f00000000, %p51;
bra.uni $L__BB0_68;
$L__BB0_64:
mov.b32 %r249, %f1180;
xor.b32 %r250, %r249, -2147483648;
mov.b32 %f653, %r250;
selp.f32 %f1182, %f653, %f1180, %p3;
setp.geu.f32 %p83, %f145, 0f00000000;
@%p83 bra $L__BB0_68;
mov.f32 %f654, 0f3EE66666;
cvt.rzi.f32.f32 %f655, %f654;
setp.eq.f32 %p84, %f655, 0f3EE66666;
@%p84 bra $L__BB0_68;
mov.f32 %f1182, 0f7FFFFFFF;
$L__BB0_68:
add.f32 %f658, %f171, 0f3EE66666;
mov.b32 %r251, %f658;
setp.lt.s32 %p86, %r251, 2139095040;
@%p86 bra $L__BB0_73;
setp.gtu.f32 %p87, %f171, 0f7F800000;
@%p87 bra $L__BB0_72;
bra.uni $L__BB0_70;
$L__BB0_72:
add.f32 %f1182, %f145, 0f3EE66666;
bra.uni $L__BB0_73;
$L__BB0_70:
setp.neu.f32 %p88, %f171, 0f7F800000;
@%p88 bra $L__BB0_73;
selp.f32 %f1182, 0fFF800000, 0f7F800000, %p3;
$L__BB0_73:
setp.eq.f32 %p89, %f145, 0f3F800000;
mov.f32 %f660, 0f3F800000;
selp.f32 %f661, 0f3F800000, %f1182, %p89;
min.f32 %f662, %f158, %f660;
mov.f32 %f663, 0f00000000;
max.f32 %f182, %f663, %f662;
min.f32 %f664, %f170, %f660;
max.f32 %f183, %f663, %f664;
min.f32 %f665, %f661, %f660;
max.f32 %f184, %f663, %f665;
mov.f32 %f669, 0f3ED55555;
abs.f32 %f186, %f182;
setp.lt.f32 %p90, %f186, 0f00800000;
mul.f32 %f671, %f186, 0f4B800000;
selp.f32 %f672, %f671, %f186, %p90;
selp.f32 %f673, 0fC3170000, 0fC2FE0000, %p90;
mov.b32 %r254, %f672;
and.b32 %r255, %r254, 8388607;
or.b32 %r256, %r255, 1065353216;
mov.b32 %f674, %r256;
shr.u32 %r257, %r254, 23;
cvt.rn.f32.u32 %f675, %r257;
add.f32 %f676, %f673, %f675;
setp.gt.f32 %p91, %f674, 0f3FB504F3;
mul.f32 %f677, %f674, 0f3F000000;
add.f32 %f678, %f676, 0f3F800000;
selp.f32 %f679, %f678, %f676, %p91;
selp.f32 %f680, %f677, %f674, %p91;
add.f32 %f681, %f680, 0fBF800000;
add.f32 %f682, %f680, 0f3F800000;
rcp.approx.ftz.f32 %f683, %f682;
add.f32 %f684, %f681, %f681;
mul.f32 %f685, %f684, %f683;
mul.f32 %f686, %f685, %f685;
mov.f32 %f687, 0f3C4CAF63;
mov.f32 %f688, 0f3B18F0FE;
fma.rn.f32 %f689, %f688, %f686, %f687;
mov.f32 %f690, 0f3DAAAABD;
fma.rn.f32 %f691, %f689, %f686, %f690;
mul.rn.f32 %f692, %f691, %f686;
mul.rn.f32 %f693, %f692, %f685;
sub.f32 %f694, %f681, %f685;
add.f32 %f695, %f694, %f694;
neg.f32 %f696, %f685;
fma.rn.f32 %f697, %f696, %f681, %f695;
mul.rn.f32 %f698, %f683, %f697;
add.f32 %f699, %f693, %f685;
sub.f32 %f700, %f685, %f699;
add.f32 %f701, %f693, %f700;
add.f32 %f702, %f698, %f701;
add.f32 %f703, %f699, %f702;
sub.f32 %f704, %f699, %f703;
add.f32 %f705, %f702, %f704;
mov.f32 %f706, 0f3F317200;
mul.rn.f32 %f707, %f679, %f706;
mov.f32 %f708, 0f35BFBE8E;
mul.rn.f32 %f709, %f679, %f708;
add.f32 %f710, %f707, %f703;
sub.f32 %f711, %f707, %f710;
add.f32 %f712, %f703, %f711;
add.f32 %f713, %f705, %f712;
add.f32 %f714, %f709, %f713;
add.f32 %f715, %f710, %f714;
sub.f32 %f716, %f710, %f715;
add.f32 %f717, %f714, %f716;
mul.rn.f32 %f718, %f669, %f715;
neg.f32 %f719, %f718;
fma.rn.f32 %f720, %f669, %f715, %f719;
fma.rn.f32 %f721, %f669, %f717, %f720;
fma.rn.f32 %f722, %f663, %f715, %f721;
add.rn.f32 %f723, %f718, %f722;
neg.f32 %f724, %f723;
add.rn.f32 %f725, %f718, %f724;
add.rn.f32 %f726, %f725, %f722;
mov.b32 %r258, %f723;
setp.eq.s32 %p92, %r258, 1118925336;
add.s32 %r259, %r258, -1;
mov.b32 %f727, %r259;
add.f32 %f728, %f726, 0f37000000;
selp.f32 %f187, %f728, %f726, %p92;
selp.f32 %f729, %f727, %f723, %p92;
mov.f32 %f730, 0f3FB8AA3B;
mul.rn.f32 %f731, %f729, %f730;
cvt.rzi.f32.f32 %f732, %f731;
abs.f32 %f733, %f732;
setp.gt.f32 %p93, %f733, 0f42FC0000;
mov.b32 %r260, %f732;
and.b32 %r261, %r260, -2147483648;
or.b32 %r262, %r261, 1123811328;
mov.b32 %f734, %r262;
selp.f32 %f735, %f734, %f732, %p93;
mov.f32 %f736, 0fBF317218;
fma.rn.f32 %f737, %f735, %f736, %f729;
mov.f32 %f738, 0f3102E308;
fma.rn.f32 %f739, %f735, %f738, %f737;
mul.f32 %f740, %f739, 0f3FB8AA3B;
add.f32 %f741, %f735, 0f4B40007F;
mov.b32 %r263, %f741;
shl.b32 %r264, %r263, 23;
mov.b32 %f742, %r264;
ex2.approx.ftz.f32 %f743, %f740;
mul.f32 %f188, %f743, %f742;
setp.eq.f32 %p94, %f188, 0f7F800000;
mov.f32 %f1183, 0f7F800000;
@%p94 bra $L__BB0_75;
fma.rn.f32 %f1183, %f188, %f187, %f188;
$L__BB0_75:
mov.f32 %f1087, 0f00000000;
max.f32 %f1086, %f1087, %f662;
mov.f32 %f1067, 0f3E555555;
cvt.rzi.f32.f32 %f1066, %f1067;
add.f32 %f1065, %f1066, %f1066;
mov.f32 %f1064, 0f3ED55555;
sub.f32 %f1063, %f1064, %f1065;
abs.f32 %f1062, %f1063;
setp.lt.f32 %p95, %f1086, 0f00000000;
setp.eq.f32 %p96, %f1062, 0f3F800000;
and.pred %p4, %p95, %p96;
setp.eq.f32 %p97, %f1086, 0f00000000;
@%p97 bra $L__BB0_79;
bra.uni $L__BB0_76;
$L__BB0_79:
mov.f32 %f1104, 0f00000000;
max.f32 %f1103, %f1104, %f662;
add.f32 %f748, %f1103, %f1103;
selp.f32 %f1185, %f748, 0f00000000, %p96;
bra.uni $L__BB0_80;
$L__BB0_76:
mov.f32 %f1089, 0f00000000;
max.f32 %f1088, %f1089, %f662;
mov.b32 %r265, %f1183;
xor.b32 %r266, %r265, -2147483648;
mov.b32 %f744, %r266;
selp.f32 %f1185, %f744, %f1183, %p4;
setp.geu.f32 %p98, %f1088, 0f00000000;
@%p98 bra $L__BB0_80;
mov.f32 %f745, 0f3ED55555;
cvt.rzi.f32.f32 %f746, %f745;
setp.eq.f32 %p99, %f746, 0f3ED55555;
@%p99 bra $L__BB0_80;
mov.f32 %f1185, 0f7FFFFFFF;
$L__BB0_80:
mov.f32 %f1092, 0f00000000;
max.f32 %f1091, %f1092, %f662;
abs.f32 %f1090, %f1091;
add.f32 %f749, %f1090, 0f3ED55555;
mov.b32 %r267, %f749;
setp.lt.s32 %p101, %r267, 2139095040;
@%p101 bra $L__BB0_85;
mov.f32 %f1097, 0f00000000;
max.f32 %f1096, %f1097, %f662;
abs.f32 %f1095, %f1096;
setp.gtu.f32 %p102, %f1095, 0f7F800000;
@%p102 bra $L__BB0_84;
bra.uni $L__BB0_82;
$L__BB0_84:
mov.f32 %f1102, 0f00000000;
max.f32 %f1101, %f1102, %f662;
add.f32 %f1185, %f1101, 0f3ED55555;
bra.uni $L__BB0_85;
$L__BB0_82:
mov.f32 %f1100, 0f00000000;
max.f32 %f1099, %f1100, %f662;
abs.f32 %f1098, %f1099;
setp.neu.f32 %p103, %f1098, 0f7F800000;
@%p103 bra $L__BB0_85;
selp.f32 %f1185, 0fFF800000, 0f7F800000, %p4;
$L__BB0_85:
abs.f32 %f197, %f183;
setp.lt.f32 %p104, %f197, 0f00800000;
mul.f32 %f751, %f197, 0f4B800000;
selp.f32 %f752, %f751, %f197, %p104;
selp.f32 %f753, 0fC3170000, 0fC2FE0000, %p104;
mov.b32 %r268, %f752;
and.b32 %r269, %r268, 8388607;
or.b32 %r270, %r269, 1065353216;
mov.b32 %f754, %r270;
shr.u32 %r271, %r268, 23;
cvt.rn.f32.u32 %f755, %r271;
add.f32 %f756, %f753, %f755;
setp.gt.f32 %p105, %f754, 0f3FB504F3;
mul.f32 %f757, %f754, 0f3F000000;
add.f32 %f758, %f756, 0f3F800000;
selp.f32 %f759, %f758, %f756, %p105;
selp.f32 %f760, %f757, %f754, %p105;
add.f32 %f761, %f760, 0fBF800000;
add.f32 %f762, %f760, 0f3F800000;
rcp.approx.ftz.f32 %f763, %f762;
add.f32 %f764, %f761, %f761;
mul.f32 %f765, %f764, %f763;
mul.f32 %f766, %f765, %f765;
mov.f32 %f767, 0f3C4CAF63;
mov.f32 %f768, 0f3B18F0FE;
fma.rn.f32 %f769, %f768, %f766, %f767;
mov.f32 %f770, 0f3DAAAABD;
fma.rn.f32 %f771, %f769, %f766, %f770;
mul.rn.f32 %f772, %f771, %f766;
mul.rn.f32 %f773, %f772, %f765;
sub.f32 %f774, %f761, %f765;
add.f32 %f775, %f774, %f774;
neg.f32 %f776, %f765;
fma.rn.f32 %f777, %f776, %f761, %f775;
mul.rn.f32 %f778, %f763, %f777;
add.f32 %f779, %f773, %f765;
sub.f32 %f780, %f765, %f779;
add.f32 %f781, %f773, %f780;
add.f32 %f782, %f778, %f781;
add.f32 %f783, %f779, %f782;
sub.f32 %f784, %f779, %f783;
add.f32 %f785, %f782, %f784;
mov.f32 %f786, 0f3F317200;
mul.rn.f32 %f787, %f759, %f786;
mov.f32 %f788, 0f35BFBE8E;
mul.rn.f32 %f789, %f759, %f788;
add.f32 %f790, %f787, %f783;
sub.f32 %f791, %f787, %f790;
add.f32 %f792, %f783, %f791;
add.f32 %f793, %f785, %f792;
add.f32 %f794, %f789, %f793;
add.f32 %f795, %f790, %f794;
sub.f32 %f796, %f790, %f795;
add.f32 %f797, %f794, %f796;
mov.f32 %f798, 0f3ED55555;
mul.rn.f32 %f799, %f798, %f795;
neg.f32 %f800, %f799;
fma.rn.f32 %f801, %f798, %f795, %f800;
fma.rn.f32 %f802, %f798, %f797, %f801;
mov.f32 %f803, 0f00000000;
fma.rn.f32 %f804, %f803, %f795, %f802;
add.rn.f32 %f805, %f799, %f804;
neg.f32 %f806, %f805;
add.rn.f32 %f807, %f799, %f806;
add.rn.f32 %f808, %f807, %f804;
mov.b32 %r272, %f805;
setp.eq.s32 %p106, %r272, 1118925336;
add.s32 %r273, %r272, -1;
mov.b32 %f809, %r273;
add.f32 %f810, %f808, 0f37000000;
selp.f32 %f198, %f810, %f808, %p106;
selp.f32 %f811, %f809, %f805, %p106;
mov.f32 %f812, 0f3FB8AA3B;
mul.rn.f32 %f813, %f811, %f812;
cvt.rzi.f32.f32 %f814, %f813;
abs.f32 %f815, %f814;
setp.gt.f32 %p107, %f815, 0f42FC0000;
mov.b32 %r274, %f814;
and.b32 %r275, %r274, -2147483648;
or.b32 %r276, %r275, 1123811328;
mov.b32 %f816, %r276;
selp.f32 %f817, %f816, %f814, %p107;
mov.f32 %f818, 0fBF317218;
fma.rn.f32 %f819, %f817, %f818, %f811;
mov.f32 %f820, 0f3102E308;
fma.rn.f32 %f821, %f817, %f820, %f819;
mul.f32 %f822, %f821, 0f3FB8AA3B;
add.f32 %f823, %f817, 0f4B40007F;
mov.b32 %r277, %f823;
shl.b32 %r278, %r277, 23;
mov.b32 %f824, %r278;
ex2.approx.ftz.f32 %f825, %f822;
mul.f32 %f199, %f825, %f824;
setp.eq.f32 %p108, %f199, 0f7F800000;
mov.f32 %f1186, 0f7F800000;
@%p108 bra $L__BB0_87;
fma.rn.f32 %f1186, %f199, %f198, %f199;
$L__BB0_87:
setp.lt.f32 %p109, %f183, 0f00000000;
and.pred %p5, %p109, %p96;
setp.eq.f32 %p111, %f183, 0f00000000;
@%p111 bra $L__BB0_91;
bra.uni $L__BB0_88;
$L__BB0_91:
add.f32 %f830, %f183, %f183;
selp.f32 %f1188, %f830, 0f00000000, %p96;
bra.uni $L__BB0_92;
$L__BB0_88:
mov.b32 %r279, %f1186;
xor.b32 %r280, %r279, -2147483648;
mov.b32 %f826, %r280;
selp.f32 %f1188, %f826, %f1186, %p5;
setp.geu.f32 %p112, %f183, 0f00000000;
@%p112 bra $L__BB0_92;
mov.f32 %f827, 0f3ED55555;
cvt.rzi.f32.f32 %f828, %f827;
setp.eq.f32 %p113, %f828, 0f3ED55555;
@%p113 bra $L__BB0_92;
mov.f32 %f1188, 0f7FFFFFFF;
$L__BB0_92:
mov.f32 %f1070, 0f00000000;
max.f32 %f1069, %f1070, %f664;
abs.f32 %f1068, %f1069;
add.f32 %f831, %f1068, 0f3ED55555;
mov.b32 %r281, %f831;
setp.lt.s32 %p115, %r281, 2139095040;
@%p115 bra $L__BB0_97;
mov.f32 %f1073, 0f00000000;
max.f32 %f1072, %f1073, %f664;
abs.f32 %f1071, %f1072;
setp.gtu.f32 %p116, %f1071, 0f7F800000;
@%p116 bra $L__BB0_96;
bra.uni $L__BB0_94;
$L__BB0_96:
add.f32 %f1188, %f183, 0f3ED55555;
bra.uni $L__BB0_97;
$L__BB0_94:
mov.f32 %f1076, 0f00000000;
max.f32 %f1075, %f1076, %f664;
abs.f32 %f1074, %f1075;
setp.neu.f32 %p117, %f1074, 0f7F800000;
@%p117 bra $L__BB0_97;
selp.f32 %f1188, 0fFF800000, 0f7F800000, %p5;
$L__BB0_97:
abs.f32 %f208, %f184;
setp.lt.f32 %p118, %f208, 0f00800000;
mul.f32 %f833, %f208, 0f4B800000;
selp.f32 %f834, %f833, %f208, %p118;
selp.f32 %f835, 0fC3170000, 0fC2FE0000, %p118;
mov.b32 %r282, %f834;
and.b32 %r283, %r282, 8388607;
or.b32 %r284, %r283, 1065353216;
mov.b32 %f836, %r284;
shr.u32 %r285, %r282, 23;
cvt.rn.f32.u32 %f837, %r285;
add.f32 %f838, %f835, %f837;
setp.gt.f32 %p119, %f836, 0f3FB504F3;
mul.f32 %f839, %f836, 0f3F000000;
add.f32 %f840, %f838, 0f3F800000;
selp.f32 %f841, %f840, %f838, %p119;
selp.f32 %f842, %f839, %f836, %p119;
add.f32 %f843, %f842, 0fBF800000;
add.f32 %f844, %f842, 0f3F800000;
rcp.approx.ftz.f32 %f845, %f844;
add.f32 %f846, %f843, %f843;
mul.f32 %f847, %f846, %f845;
mul.f32 %f848, %f847, %f847;
mov.f32 %f849, 0f3C4CAF63;
mov.f32 %f850, 0f3B18F0FE;
fma.rn.f32 %f851, %f850, %f848, %f849;
mov.f32 %f852, 0f3DAAAABD;
fma.rn.f32 %f853, %f851, %f848, %f852;
mul.rn.f32 %f854, %f853, %f848;
mul.rn.f32 %f855, %f854, %f847;
sub.f32 %f856, %f843, %f847;
add.f32 %f857, %f856, %f856;
neg.f32 %f858, %f847;
fma.rn.f32 %f859, %f858, %f843, %f857;
mul.rn.f32 %f860, %f845, %f859;
add.f32 %f861, %f855, %f847;
sub.f32 %f862, %f847, %f861;
add.f32 %f863, %f855, %f862;
add.f32 %f864, %f860, %f863;
add.f32 %f865, %f861, %f864;
sub.f32 %f866, %f861, %f865;
add.f32 %f867, %f864, %f866;
mov.f32 %f868, 0f3F317200;
mul.rn.f32 %f869, %f841, %f868;
mov.f32 %f870, 0f35BFBE8E;
mul.rn.f32 %f871, %f841, %f870;
add.f32 %f872, %f869, %f865;
sub.f32 %f873, %f869, %f872;
add.f32 %f874, %f865, %f873;
add.f32 %f875, %f867, %f874;
add.f32 %f876, %f871, %f875;
add.f32 %f877, %f872, %f876;
sub.f32 %f878, %f872, %f877;
add.f32 %f879, %f876, %f878;
mov.f32 %f880, 0f3ED55555;
mul.rn.f32 %f881, %f880, %f877;
neg.f32 %f882, %f881;
fma.rn.f32 %f883, %f880, %f877, %f882;
fma.rn.f32 %f884, %f880, %f879, %f883;
mov.f32 %f885, 0f00000000;
fma.rn.f32 %f886, %f885, %f877, %f884;
add.rn.f32 %f887, %f881, %f886;
neg.f32 %f888, %f887;
add.rn.f32 %f889, %f881, %f888;
add.rn.f32 %f890, %f889, %f886;
mov.b32 %r286, %f887;
setp.eq.s32 %p120, %r286, 1118925336;
add.s32 %r287, %r286, -1;
mov.b32 %f891, %r287;
add.f32 %f892, %f890, 0f37000000;
selp.f32 %f209, %f892, %f890, %p120;
selp.f32 %f893, %f891, %f887, %p120;
mov.f32 %f894, 0f3FB8AA3B;
mul.rn.f32 %f895, %f893, %f894;
cvt.rzi.f32.f32 %f896, %f895;
abs.f32 %f897, %f896;
setp.gt.f32 %p121, %f897, 0f42FC0000;
mov.b32 %r288, %f896;
and.b32 %r289, %r288, -2147483648;
or.b32 %r290, %r289, 1123811328;
mov.b32 %f898, %r290;
selp.f32 %f899, %f898, %f896, %p121;
mov.f32 %f900, 0fBF317218;
fma.rn.f32 %f901, %f899, %f900, %f893;
mov.f32 %f902, 0f3102E308;
fma.rn.f32 %f903, %f899, %f902, %f901;
mul.f32 %f904, %f903, 0f3FB8AA3B;
add.f32 %f905, %f899, 0f4B40007F;
mov.b32 %r291, %f905;
shl.b32 %r292, %r291, 23;
mov.b32 %f906, %r292;
ex2.approx.ftz.f32 %f907, %f904;
mul.f32 %f210, %f907, %f906;
setp.eq.f32 %p122, %f210, 0f7F800000;
mov.f32 %f1189, 0f7F800000;
@%p122 bra $L__BB0_99;
fma.rn.f32 %f1189, %f210, %f209, %f210;
$L__BB0_99:
setp.lt.f32 %p123, %f184, 0f00000000;
and.pred %p6, %p123, %p96;
setp.eq.f32 %p125, %f184, 0f00000000;
@%p125 bra $L__BB0_103;
bra.uni $L__BB0_100;
$L__BB0_103:
add.f32 %f912, %f184, %f184;
selp.f32 %f1191, %f912, 0f00000000, %p96;
bra.uni $L__BB0_104;
$L__BB0_100:
mov.b32 %r293, %f1189;
xor.b32 %r294, %r293, -2147483648;
mov.b32 %f908, %r294;
selp.f32 %f1191, %f908, %f1189, %p6;
setp.geu.f32 %p126, %f184, 0f00000000;
@%p126 bra $L__BB0_104;
mov.f32 %f909, 0f3ED55555;
cvt.rzi.f32.f32 %f910, %f909;
setp.eq.f32 %p127, %f910, 0f3ED55555;
@%p127 bra $L__BB0_104;
mov.f32 %f1191, 0f7FFFFFFF;
$L__BB0_104:
mov.f32 %f1079, 0f00000000;
max.f32 %f1078, %f1079, %f665;
abs.f32 %f1077, %f1078;
add.f32 %f913, %f1077, 0f3ED55555;
mov.b32 %r295, %f913;
setp.lt.s32 %p129, %r295, 2139095040;
@%p129 bra $L__BB0_109;
mov.f32 %f1082, 0f00000000;
max.f32 %f1081, %f1082, %f665;
abs.f32 %f1080, %f1081;
setp.gtu.f32 %p130, %f1080, 0f7F800000;
@%p130 bra $L__BB0_108;
bra.uni $L__BB0_106;
$L__BB0_108:
add.f32 %f1191, %f184, 0f3ED55555;
bra.uni $L__BB0_109;
$L__BB0_106:
mov.f32 %f1085, 0f00000000;
max.f32 %f1084, %f1085, %f665;
abs.f32 %f1083, %f1084;
setp.neu.f32 %p131, %f1083, 0f7F800000;
@%p131 bra $L__BB0_109;
selp.f32 %f1191, 0fFF800000, 0f7F800000, %p6;
$L__BB0_109:
mov.f32 %f1094, 0f00000000;
max.f32 %f1093, %f1094, %f662;
ld.const.u64 %rd103, [params+144];
cvta.to.global.u64 %rd102, %rd103;
ld.const.u32 %r346, [params+136];
mad.lo.s32 %r345, %r346, %r5, %r4;
cvt.u64.u32 %rd101, %r345;
fma.rn.f32 %f914, %f1185, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p132, %f1093, 0f3F800000;
mov.f32 %f915, 0f3F800000;
selp.f32 %f916, 0f3F7FFFFF, %f914, %p132;
mul.f32 %f917, %f1093, 0f414EB852;
setp.lt.f32 %p133, %f1093, 0f3B4D2E1C;
selp.f32 %f918, %f917, %f916, %p133;
fma.rn.f32 %f919, %f1188, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p134, %f183, 0f3F800000;
selp.f32 %f920, 0f3F7FFFFF, %f919, %p134;
mul.f32 %f921, %f183, 0f414EB852;
setp.lt.f32 %p135, %f183, 0f3B4D2E1C;
selp.f32 %f922, %f921, %f920, %p135;
fma.rn.f32 %f923, %f1191, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p136, %f184, 0f3F800000;
selp.f32 %f924, 0f3F7FFFFF, %f923, %p136;
mul.f32 %f925, %f184, 0f414EB852;
setp.lt.f32 %p137, %f184, 0f3B4D2E1C;
selp.f32 %f926, %f925, %f924, %p137;
min.f32 %f927, %f918, %f915;
mov.f32 %f928, 0f00000000;
max.f32 %f929, %f928, %f927;
mul.f32 %f930, %f929, 0f43800000;
cvt.rzi.u32.f32 %r296, %f930;
min.u32 %r297, %r296, 255;
min.f32 %f931, %f922, %f915;
max.f32 %f932, %f928, %f931;
mul.f32 %f933, %f932, 0f43800000;
cvt.rzi.u32.f32 %r298, %f933;
min.u32 %r299, %r298, 255;
min.f32 %f934, %f926, %f915;
max.f32 %f935, %f928, %f934;
mul.f32 %f936, %f935, 0f43800000;
cvt.rzi.u32.f32 %r300, %f936;
min.u32 %r301, %r300, 255;
shl.b64 %rd65, %rd101, 2;
add.s64 %rd66, %rd102, %rd65;
cvt.u16.u32 %rs15, %r301;
cvt.u16.u32 %rs16, %r299;
cvt.u16.u32 %rs17, %r297;
mov.u16 %rs18, 255;
st.global.v4.u8 [%rd66], {%rs17, %rs16, %rs15, %rs18};
$L__BB0_110:
ld.const.u32 %r347, [params+104];
and.b32 %r302, %r347, 4;
setp.eq.s32 %p138, %r302, 0;
ld.const.u32 %r359, [params+108];
@%p138 bra $L__BB0_114;
setp.eq.s32 %p139, %r359, 0;
ld.const.u64 %rd67, [params+224];
cvta.to.global.u64 %rd68, %rd67;
ld.const.u32 %r303, [params+216];
mad.lo.s32 %r304, %r303, %r5, %r4;
mul.wide.u32 %rd69, %r304, 8;
add.s64 %rd18, %rd68, %rd69;
@%p139 bra $L__BB0_113;
ld.global.v4.u16 {%rs26, %rs27, %rs28, %rs29}, [%rd18];
// begin inline asm
{ cvt.f32.f16 %f937, %rs26;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f938, %rs27;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f939, %rs28;}
// end inline asm
add.f32 %f940, %f143, %f937;
add.f32 %f941, %f144, %f938;
add.f32 %f942, %f145, %f939;
mov.f32 %f943, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs24, %f942;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs23, %f941;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs22, %f940;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs25, %f943;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs22, %rs23, %rs24, %rs25};
bra.uni $L__BB0_114;
$L__BB0_113:
mov.f32 %f947, 0f3F800000;
mov.u32 %r359, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs33, %f947;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs32, %f145;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs31, %f144;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f143;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs30, %rs31, %rs32, %rs33};
$L__BB0_114:
mul.f32 %f219, %f128, 0f3F000000;
mul.f32 %f220, %f219, %f405;
mul.f32 %f221, %f129, 0f3F000000;
mul.f32 %f222, %f221, %f406;
mul.f32 %f223, %f130, 0f3F000000;
mul.f32 %f224, %f223, %f407;
ld.const.u64 %rd70, [params+256];
cvta.to.global.u64 %rd71, %rd70;
ld.const.u32 %r306, [params+248];
mad.lo.s32 %r307, %r306, %r5, %r4;
mul.wide.u32 %rd72, %r307, 8;
add.s64 %rd19, %rd71, %rd72;
setp.eq.s32 %p140, %r359, 0;
@%p140 bra $L__BB0_116;
ld.global.v4.u16 {%rs41, %rs42, %rs43, %rs44}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f948, %rs41;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f949, %rs42;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f950, %rs43;}
// end inline asm
add.f32 %f951, %f220, %f948;
add.f32 %f952, %f222, %f949;
add.f32 %f953, %f224, %f950;
mov.f32 %f954, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs39, %f953;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f952;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f951;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs40, %f954;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs37, %rs38, %rs39, %rs40};
bra.uni $L__BB0_117;
$L__BB0_116:
mov.f32 %f958, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs48, %f958;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs47, %f224;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs46, %f222;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f220;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs45, %rs46, %rs47, %rs48};
$L__BB0_117:
mov.f32 %f959, 0f34000000;
max.f32 %f960, %f219, %f959;
mul.f32 %f961, %f131, 0f3F000000;
div.rn.f32 %f962, %f961, %f960;
max.f32 %f963, %f221, %f959;
mul.f32 %f964, %f132, 0f3F000000;
div.rn.f32 %f965, %f964, %f963;
max.f32 %f966, %f223, %f959;
mul.f32 %f967, %f133, 0f3F000000;
div.rn.f32 %f968, %f967, %f966;
fma.rn.f32 %f225, %f962, 0f3F000000, 0f3F000000;
fma.rn.f32 %f226, %f965, 0f3F000000, 0f3F000000;
fma.rn.f32 %f227, %f968, 0f3F000000, 0f3F000000;
mul.f32 %f969, %f134, 0f3F000000;
div.rn.f32 %f970, %f969, %f960;
mul.f32 %f971, %f135, 0f3F000000;
div.rn.f32 %f972, %f971, %f963;
mul.f32 %f973, %f136, 0f3F000000;
div.rn.f32 %f974, %f973, %f966;
fma.rn.f32 %f228, %f970, 0f3F000000, 0f3F000000;
fma.rn.f32 %f229, %f972, 0f3F000000, 0f3F000000;
fma.rn.f32 %f230, %f974, 0f3F000000, 0f3F000000;
mul.f32 %f975, %f137, 0f3F000000;
div.rn.f32 %f976, %f975, %f960;
mul.f32 %f977, %f138, 0f3F000000;
div.rn.f32 %f978, %f977, %f963;
mul.f32 %f979, %f139, 0f3F000000;
div.rn.f32 %f980, %f979, %f966;
fma.rn.f32 %f231, %f976, 0f3F000000, 0f3F000000;
fma.rn.f32 %f232, %f978, 0f3F000000, 0f3F000000;
fma.rn.f32 %f233, %f980, 0f3F000000, 0f3F000000;
ld.const.u64 %rd73, [params+272];
cvta.to.global.u64 %rd74, %rd73;
ld.const.u32 %r308, [params+264];
mad.lo.s32 %r309, %r308, %r5, %r4;
mul.wide.u32 %rd75, %r309, 8;
add.s64 %rd20, %rd74, %rd75;
@%p140 bra $L__BB0_119;
ld.global.v4.u16 {%rs56, %rs57, %rs58, %rs59}, [%rd20];
// begin inline asm
{ cvt.f32.f16 %f981, %rs56;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f982, %rs57;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f983, %rs58;}
// end inline asm
add.f32 %f984, %f225, %f981;
add.f32 %f985, %f226, %f982;
add.f32 %f986, %f227, %f983;
mov.f32 %f987, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs54, %f986;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs53, %f985;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs52, %f984;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs55, %f987;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs52, %rs53, %rs54, %rs55};
bra.uni $L__BB0_120;
$L__BB0_119:
mov.f32 %f991, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs63, %f991;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs62, %f227;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs61, %f226;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs60, %f225;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs60, %rs61, %rs62, %rs63};
$L__BB0_120:
ld.const.u64 %rd76, [params+288];
cvta.to.global.u64 %rd77, %rd76;
ld.const.u32 %r310, [params+280];
mad.lo.s32 %r311, %r310, %r5, %r4;
mul.wide.u32 %rd78, %r311, 8;
add.s64 %rd21, %rd77, %rd78;
@%p140 bra $L__BB0_122;
ld.global.v4.u16 {%rs71, %rs72, %rs73, %rs74}, [%rd21];
// begin inline asm
{ cvt.f32.f16 %f992, %rs71;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f993, %rs72;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f994, %rs73;}
// end inline asm
add.f32 %f995, %f228, %f992;
add.f32 %f996, %f229, %f993;
add.f32 %f997, %f230, %f994;
mov.f32 %f998, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs69, %f997;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs68, %f996;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs67, %f995;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs70, %f998;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs67, %rs68, %rs69, %rs70};
bra.uni $L__BB0_123;
$L__BB0_122:
mov.f32 %f1002, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs78, %f1002;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs77, %f230;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs76, %f229;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs75, %f228;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs75, %rs76, %rs77, %rs78};
$L__BB0_123:
ld.const.u64 %rd79, [params+304];
cvta.to.global.u64 %rd80, %rd79;
ld.const.u32 %r312, [params+296];
mad.lo.s32 %r313, %r312, %r5, %r4;
mul.wide.u32 %rd81, %r313, 8;
add.s64 %rd22, %rd80, %rd81;
@%p140 bra $L__BB0_125;
ld.global.v4.u16 {%rs86, %rs87, %rs88, %rs89}, [%rd22];
// begin inline asm
{ cvt.f32.f16 %f1003, %rs86;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1004, %rs87;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1005, %rs88;}
// end inline asm
add.f32 %f1006, %f231, %f1003;
add.f32 %f1007, %f232, %f1004;
add.f32 %f1008, %f233, %f1005;
mov.f32 %f1009, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs84, %f1008;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs83, %f1007;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs82, %f1006;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs85, %f1009;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs82, %rs83, %rs84, %rs85};
bra.uni $L__BB0_144;
$L__BB0_125:
mov.f32 %f1013, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs93, %f1013;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs92, %f233;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs91, %f232;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs90, %f231;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs90, %rs91, %rs92, %rs93};
$L__BB0_144:
ret;
}