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

2196 lines
58 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<154>;
.reg .b16 %rs<172>;
.reg .f32 %f<1055>;
.reg .b32 %r<345>;
.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 %f1008, 0f00000000;
mov.u16 %rs171, 0;
mov.f32 %f1009, %f1008;
mov.f32 %f1010, %f1008;
@%p8 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f171, %rs7;
div.rn.f32 %f172, %f171, 0f437F0000;
fma.rn.f32 %f173, %f172, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs171, 255;
cvt.rn.f32.u16 %f174, %rs13;
div.rn.f32 %f175, %f174, 0f437F0000;
fma.rn.f32 %f176, %f175, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f177, %rs170;
div.rn.f32 %f178, %f177, 0f437F0000;
fma.rn.f32 %f179, %f178, 0f40000000, 0fBF800000;
mul.f32 %f180, %f176, %f176;
fma.rn.f32 %f181, %f173, %f173, %f180;
fma.rn.f32 %f182, %f179, %f179, %f181;
sqrt.rn.f32 %f183, %f182;
rcp.rn.f32 %f184, %f183;
mul.f32 %f1010, %f184, %f179;
mul.f32 %f1009, %f184, %f176;
mul.f32 %f1008, %f173, %f184;
$L__BB0_4:
ld.const.v2.u32 {%r63, %r64}, [params];
add.s32 %r4, %r63, %r55;
add.s32 %r5, %r64, %r56;
setp.eq.f32 %p9, %f1008, 0f00000000;
setp.eq.f32 %p10, %f1009, 0f00000000;
and.pred %p11, %p9, %p10;
setp.eq.f32 %p12, %f1010, 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 %r344, [params+108];
@%p148 bra $L__BB0_132;
setp.eq.s32 %p149, %r344, 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 %f905, %rs101;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f906, %rs102;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f907, %rs103;}
// end inline asm
add.f32 %f908, %f905, 0f00000000;
add.f32 %f909, %f906, 0f00000000;
add.f32 %f910, %f907, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs100, %f910;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs99, %f909;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs98, %f908;}
// 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 %f190, [%rd35];
mul.f32 %f191, %f190, 0f3456BF95;
ld.global.f32 %f192, [%rd35+4];
mul.f32 %f193, %f192, 0f3456BF95;
ld.global.f32 %f194, [%rd35+8];
mul.f32 %f195, %f194, 0f3456BF95;
abs.f32 %f196, %f1008;
div.rn.f32 %f197, %f191, %f196;
abs.f32 %f198, %f1009;
div.rn.f32 %f199, %f193, %f198;
abs.f32 %f200, %f1010;
div.rn.f32 %f201, %f195, %f200;
abs.f32 %f202, %f197;
abs.f32 %f203, %f199;
abs.f32 %f204, %f201;
mov.f32 %f205, 0f38D1B717;
max.f32 %f206, %f202, %f205;
max.f32 %f207, %f203, %f205;
max.f32 %f208, %f204, %f205;
fma.rn.f32 %f10, %f1008, %f206, %f190;
fma.rn.f32 %f11, %f1009, %f207, %f192;
fma.rn.f32 %f12, %f1010, %f208, %f194;
setp.gt.f32 %p14, %f196, %f200;
neg.f32 %f209, %f1009;
selp.f32 %f210, %f209, 0f00000000, %p14;
mov.f32 %f1027, 0f00000000;
neg.f32 %f211, %f1010;
selp.f32 %f212, %f1008, %f211, %p14;
selp.f32 %f213, 0f00000000, %f1009, %p14;
mul.f32 %f214, %f212, %f212;
fma.rn.f32 %f215, %f210, %f210, %f214;
fma.rn.f32 %f216, %f213, %f213, %f215;
sqrt.rn.f32 %f217, %f216;
rcp.rn.f32 %f218, %f217;
mul.f32 %f13, %f210, %f218;
mul.f32 %f14, %f212, %f218;
mul.f32 %f15, %f213, %f218;
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 %f1028, %f1027;
mov.f32 %f1029, %f1027;
mov.f32 %f1030, %f1027;
mov.f32 %f1031, %f1027;
@%p15 bra $L__BB0_36;
ld.const.u32 %r6, [params+588];
cvt.rn.f32.s32 %f224, %r3;
rcp.rn.f32 %f16, %f224;
ld.global.u32 %r335, [%rd3];
mul.f32 %f17, %f10, 0f3456BF95;
mul.f32 %f18, %f11, 0f3456BF95;
mul.f32 %f19, %f12, 0f3456BF95;
ld.const.u64 %rd4, [params+96];
mul.f32 %f225, %f1008, %f14;
mul.f32 %f226, %f1009, %f13;
sub.f32 %f20, %f226, %f225;
mul.f32 %f227, %f1010, %f13;
mul.f32 %f228, %f1008, %f15;
sub.f32 %f21, %f228, %f227;
mul.f32 %f229, %f1009, %f15;
mul.f32 %f230, %f1010, %f14;
sub.f32 %f22, %f230, %f229;
mov.u32 %r71, 0;
add.s64 %rd5, %rd1, 24;
mov.u64 %rd39, __cudart_i2opi_f;
abs.f32 %f293, %f17;
abs.f32 %f294, %f18;
max.f32 %f295, %f293, %f294;
abs.f32 %f296, %f19;
max.f32 %f297, %f295, %f296;
mov.u32 %r332, %r71;
$L__BB0_7:
mov.u32 %r334, %r71;
$L__BB0_8:
cvt.rn.f32.s32 %f979, %r332;
mad.lo.s32 %r73, %r335, 1664525, 1013904223;
and.b32 %r74, %r73, 16777215;
cvt.rn.f32.u32 %f231, %r74;
fma.rn.f32 %f232, %f231, 0f33800000, %f979;
mul.f32 %f34, %f16, %f232;
mad.lo.s32 %r335, %r73, 1664525, 1013904223;
and.b32 %r75, %r335, 16777215;
cvt.rn.f32.u32 %f233, %r75;
cvt.rn.f32.s32 %f234, %r334;
fma.rn.f32 %f235, %f233, 0f33800000, %f234;
mul.f32 %f236, %f16, %f235;
mul.f32 %f237, %f34, %f34;
mov.f32 %f238, 0f3F800000;
sub.f32 %f239, %f238, %f237;
mov.f32 %f240, 0f00000000;
max.f32 %f241, %f240, %f239;
sqrt.rn.f32 %f35, %f241;
mul.f32 %f36, %f236, 0f40C90FDB;
mul.f32 %f242, %f36, 0f3F22F983;
cvt.rni.s32.f32 %r342, %f242;
cvt.rn.f32.s32 %f243, %r342;
mov.f32 %f244, 0fBFC90FDA;
fma.rn.f32 %f245, %f243, %f244, %f36;
mov.f32 %f246, 0fB3A22168;
fma.rn.f32 %f247, %f243, %f246, %f245;
mov.f32 %f248, 0fA7C234C5;
fma.rn.f32 %f1024, %f243, %f248, %f247;
abs.f32 %f38, %f36;
setp.ltu.f32 %p16, %f38, 0f47CE4780;
mov.u32 %r339, %r342;
mov.f32 %f1021, %f1024;
@%p16 bra $L__BB0_16;
setp.eq.f32 %p17, %f38, 0f7F800000;
@%p17 bra $L__BB0_15;
bra.uni $L__BB0_10;
$L__BB0_15:
mov.f32 %f251, 0f00000000;
mul.rn.f32 %f1021, %f36, %f251;
mov.u32 %r339, 0;
bra.uni $L__BB0_16;
$L__BB0_10:
mov.b32 %r14, %f36;
bfe.u32 %r77, %r14, 23, 8;
add.s32 %r15, %r77, -128;
shl.b32 %r78, %r14, 8;
or.b32 %r16, %r78, -2147483648;
shr.u32 %r17, %r15, 5;
mov.u64 %rd106, 0;
mov.u32 %r336, 0;
mov.u64 %rd104, %rd1;
mov.u64 %rd105, %rd39;
$L__BB0_11:
.pragma "nounroll";
ld.global.nc.u32 %r79, [%rd105];
mad.wide.u32 %rd41, %r79, %r16, %rd106;
shr.u64 %rd106, %rd41, 32;
st.local.u32 [%rd104], %rd41;
add.s64 %rd105, %rd105, 4;
add.s64 %rd104, %rd104, 4;
add.s32 %r336, %r336, 1;
setp.ne.s32 %p18, %r336, 6;
@%p18 bra $L__BB0_11;
st.local.u32 [%rd5], %rd106;
mov.u32 %r80, 4;
sub.s32 %r20, %r80, %r17;
mov.u32 %r81, 6;
sub.s32 %r82, %r81, %r17;
mul.wide.s32 %rd42, %r82, 4;
add.s64 %rd43, %rd1, %rd42;
ld.local.u32 %r337, [%rd43];
ld.local.u32 %r338, [%rd43+-4];
and.b32 %r23, %r15, 31;
setp.eq.s32 %p19, %r23, 0;
@%p19 bra $L__BB0_14;
mov.u32 %r83, 32;
sub.s32 %r84, %r83, %r23;
shr.u32 %r85, %r338, %r84;
shl.b32 %r86, %r337, %r23;
add.s32 %r337, %r85, %r86;
mul.wide.s32 %rd44, %r20, 4;
add.s64 %rd45, %rd1, %rd44;
ld.local.u32 %r87, [%rd45];
shr.u32 %r88, %r87, %r84;
shl.b32 %r89, %r338, %r23;
add.s32 %r338, %r88, %r89;
$L__BB0_14:
and.b32 %r90, %r14, -2147483648;
shr.u32 %r91, %r338, 30;
shl.b32 %r92, %r337, 2;
or.b32 %r93, %r91, %r92;
shr.u32 %r94, %r93, 31;
shr.u32 %r95, %r337, 30;
add.s32 %r96, %r94, %r95;
neg.s32 %r97, %r96;
setp.eq.s32 %p20, %r90, 0;
selp.b32 %r339, %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, %r338, 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 %f249, %fd2;
setp.eq.s32 %p22, %r99, 0;
neg.f32 %f250, %f249;
selp.f32 %f1021, %f249, %f250, %p22;
$L__BB0_16:
add.s32 %r30, %r339, 1;
and.b32 %r31, %r30, 1;
setp.eq.s32 %p23, %r31, 0;
selp.f32 %f42, %f1021, 0f3F800000, %p23;
mul.rn.f32 %f43, %f1021, %f1021;
mov.f32 %f1022, 0fB94D4153;
@%p23 bra $L__BB0_18;
mov.f32 %f253, 0fBAB607ED;
mov.f32 %f254, 0f37CBAC00;
fma.rn.f32 %f1022, %f254, %f43, %f253;
$L__BB0_18:
selp.f32 %f255, 0f3C0885E4, 0f3D2AAABB, %p23;
fma.rn.f32 %f256, %f1022, %f43, %f255;
selp.f32 %f257, 0fBE2AAAA8, 0fBEFFFFFF, %p23;
fma.rn.f32 %f258, %f256, %f43, %f257;
mov.f32 %f259, 0f00000000;
fma.rn.f32 %f260, %f43, %f42, %f259;
fma.rn.f32 %f1023, %f258, %f260, %f42;
and.b32 %r105, %r30, 2;
setp.eq.s32 %p25, %r105, 0;
@%p25 bra $L__BB0_20;
mov.f32 %f262, 0fBF800000;
fma.rn.f32 %f1023, %f1023, %f262, %f259;
$L__BB0_20:
@%p16 bra $L__BB0_28;
setp.eq.f32 %p27, %f38, 0f7F800000;
@%p27 bra $L__BB0_27;
bra.uni $L__BB0_22;
$L__BB0_27:
mov.f32 %f265, 0f00000000;
mul.rn.f32 %f1024, %f36, %f265;
mov.u32 %r342, 0;
bra.uni $L__BB0_28;
$L__BB0_22:
mov.b32 %r32, %f36;
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 %r340, [%rd57];
ld.local.u32 %r341, [%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, %r341, %r115;
shl.b32 %r117, %r340, %r39;
add.s32 %r340, %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, %r341, %r39;
add.s32 %r341, %r119, %r120;
$L__BB0_26:
and.b32 %r121, %r32, -2147483648;
shr.u32 %r122, %r341, 30;
shl.b32 %r123, %r340, 2;
or.b32 %r124, %r122, %r123;
shr.u32 %r125, %r124, 31;
shr.u32 %r126, %r340, 30;
add.s32 %r127, %r125, %r126;
neg.s32 %r128, %r127;
setp.eq.s32 %p30, %r121, 0;
selp.b32 %r342, %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, %r341, 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 %f263, %fd4;
setp.eq.s32 %p32, %r130, 0;
neg.f32 %f264, %f263;
selp.f32 %f1024, %f263, %f264, %p32;
$L__BB0_28:
mul.f32 %f52, %f35, %f1023;
and.b32 %r46, %r342, 1;
setp.eq.s32 %p33, %r46, 0;
selp.f32 %f53, %f1024, 0f3F800000, %p33;
mul.rn.f32 %f54, %f1024, %f1024;
mov.f32 %f1025, 0fB94D4153;
@%p33 bra $L__BB0_30;
mov.f32 %f267, 0fBAB607ED;
mov.f32 %f268, 0f37CBAC00;
fma.rn.f32 %f1025, %f268, %f54, %f267;
$L__BB0_30:
selp.f32 %f269, 0f3C0885E4, 0f3D2AAABB, %p33;
fma.rn.f32 %f270, %f1025, %f54, %f269;
selp.f32 %f271, 0fBE2AAAA8, 0fBEFFFFFF, %p33;
fma.rn.f32 %f272, %f270, %f54, %f271;
mov.f32 %f273, 0f00000000;
fma.rn.f32 %f274, %f54, %f53, %f273;
fma.rn.f32 %f1026, %f272, %f274, %f53;
and.b32 %r136, %r342, 2;
setp.eq.s32 %p35, %r136, 0;
@%p35 bra $L__BB0_32;
mov.f32 %f276, 0fBF800000;
fma.rn.f32 %f1026, %f1026, %f276, %f273;
$L__BB0_32:
mul.f32 %f277, %f35, %f1026;
mul.f32 %f278, %f13, %f277;
mul.f32 %f279, %f14, %f277;
mul.f32 %f280, %f15, %f277;
fma.rn.f32 %f281, %f22, %f52, %f278;
fma.rn.f32 %f282, %f21, %f52, %f279;
fma.rn.f32 %f283, %f20, %f52, %f280;
fma.rn.f32 %f60, %f1008, %f34, %f281;
fma.rn.f32 %f61, %f1009, %f34, %f282;
fma.rn.f32 %f62, %f1010, %f34, %f283;
setp.leu.f32 %p36, %f61, 0f00000000;
setp.ne.s32 %p37, %r6, 0;
and.pred %p38, %p37, %p36;
@%p38 bra $L__BB0_34;
mov.f32 %f298, 0f38D1B717;
max.f32 %f290, %f297, %f298;
mov.f32 %f291, 0f6C4ECB8F;
mov.f32 %f292, 0f00000000;
mov.u32 %r173, 2;
mov.u32 %r175, 1;
mov.u32 %r176, 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,%f60,%f61,%f62,%f290,%f291,%f292,%r175,%r207,%r175,%r173,%r175,%r175,%r176,%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,%r207,%r207);
// end inline asm
mov.b32 %f299, %r137;
mul.f32 %f300, %f1009, %f61;
fma.rn.f32 %f301, %f1008, %f60, %f300;
fma.rn.f32 %f302, %f1010, %f62, %f301;
mul.f32 %f303, %f302, 0f40800000;
cvt.sat.f32.f32 %f304, %f303;
mul.f32 %f305, %f304, %f299;
fma.rn.f32 %f1029, %f60, %f305, %f1029;
fma.rn.f32 %f1030, %f61, %f305, %f1030;
fma.rn.f32 %f1031, %f62, %f305, %f1031;
add.f32 %f1027, %f1027, %f305;
cvt.sat.f32.f32 %f306, %f302;
fma.rn.f32 %f1028, %f306, %f299, %f1028;
$L__BB0_34:
add.s32 %r334, %r334, 1;
setp.lt.s32 %p39, %r334, %r3;
@%p39 bra $L__BB0_8;
add.s32 %r332, %r332, 1;
setp.lt.s32 %p40, %r332, %r3;
@%p40 bra $L__BB0_7;
$L__BB0_36:
mul.lo.s32 %r208, %r3, %r3;
cvt.rn.f32.s32 %f307, %r208;
div.rn.f32 %f308, %f1028, %f307;
div.rn.f32 %f78, %f1027, %f307;
div.rn.f32 %f79, %f1029, %f307;
div.rn.f32 %f80, %f1030, %f307;
div.rn.f32 %f81, %f1031, %f307;
add.f32 %f309, %f308, %f308;
ld.const.v4.f32 {%f310, %f311, %f312, %f313}, [params+576];
mul.f32 %f85, %f309, %f310;
mul.f32 %f86, %f309, %f311;
mul.f32 %f87, %f309, %f312;
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 %f318, 0f3EE66666;
abs.f32 %f89, %f85;
setp.lt.f32 %p45, %f89, 0f00800000;
mul.f32 %f320, %f89, 0f4B800000;
selp.f32 %f321, %f320, %f89, %p45;
selp.f32 %f322, 0fC3170000, 0fC2FE0000, %p45;
mov.b32 %r210, %f321;
and.b32 %r211, %r210, 8388607;
or.b32 %r212, %r211, 1065353216;
mov.b32 %f323, %r212;
shr.u32 %r213, %r210, 23;
cvt.rn.f32.u32 %f324, %r213;
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 %r214, %f373;
setp.eq.s32 %p47, %r214, 1118925336;
add.s32 %r215, %r214, -1;
mov.b32 %f377, %r215;
add.f32 %f378, %f376, 0f37000000;
selp.f32 %f90, %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 %r216, %f382;
and.b32 %r217, %r216, -2147483648;
or.b32 %r218, %r217, 1123811328;
mov.b32 %f384, %r218;
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 %r219, %f391;
shl.b32 %r220, %r219, 23;
mov.b32 %f392, %r220;
ex2.approx.ftz.f32 %f393, %f390;
mul.f32 %f91, %f393, %f392;
setp.eq.f32 %p49, %f91, 0f7F800000;
mov.f32 %f1037, 0f7F800000;
@%p49 bra $L__BB0_39;
fma.rn.f32 %f1037, %f91, %f90, %f91;
$L__BB0_39:
mov.f32 %f985, 0f3E666666;
cvt.rzi.f32.f32 %f984, %f985;
add.f32 %f983, %f984, %f984;
mov.f32 %f982, 0f3EE66666;
sub.f32 %f981, %f982, %f983;
abs.f32 %f980, %f981;
setp.lt.f32 %p50, %f85, 0f00000000;
setp.eq.f32 %p51, %f980, 0f3F800000;
and.pred %p1, %p50, %p51;
setp.eq.f32 %p52, %f85, 0f00000000;
@%p52 bra $L__BB0_43;
bra.uni $L__BB0_40;
$L__BB0_43:
add.f32 %f398, %f85, %f85;
selp.f32 %f1039, %f398, 0f00000000, %p51;
bra.uni $L__BB0_44;
$L__BB0_131:
mov.f32 %f913, 0f00000000;
mov.u32 %r344, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs108, %f913;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs107, %f913;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs106, %f913;}
// 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, %r344, 0;
@%p150 bra $L__BB0_134;
ld.global.v4.u16 {%rs116, %rs117, %rs118, %rs119}, [%rd24];
// begin inline asm
{ cvt.f32.f16 %f914, %rs116;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f915, %rs117;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f916, %rs118;}
// end inline asm
add.f32 %f917, %f914, 0f00000000;
add.f32 %f918, %f915, 0f00000000;
add.f32 %f919, %f916, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs115, %f919;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs114, %f918;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs113, %f917;}
// 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 %f922, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs123, %f922;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs122, %f922;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs121, %f922;}
// 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 %f923, %rs131;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f924, %rs132;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f925, %rs133;}
// end inline asm
add.f32 %f926, %f923, 0f00000000;
add.f32 %f927, %f924, 0f00000000;
add.f32 %f928, %f925, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs130, %f928;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs129, %f927;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f926;}
// 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 %f931, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs138, %f931;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs137, %f931;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs136, %f931;}
// 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 %f932, %rs146;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f933, %rs147;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f934, %rs148;}
// end inline asm
add.f32 %f935, %f932, 0f00000000;
add.f32 %f936, %f933, 0f00000000;
add.f32 %f937, %f934, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs145, %f937;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs144, %f936;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f935;}
// 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 %f940, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs153, %f940;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs152, %f940;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs151, %f940;}
// 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 %f941, %rs161;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f942, %rs162;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f943, %rs163;}
// end inline asm
add.f32 %f944, %f941, 0f00000000;
add.f32 %f945, %f942, 0f00000000;
add.f32 %f946, %f943, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs160, %f946;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs159, %f945;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f944;}
// 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 %f949, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs168, %f949;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs167, %f949;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs166, %f949;}
// 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, %f1037;
xor.b32 %r222, %r221, -2147483648;
mov.b32 %f394, %r222;
selp.f32 %f1039, %f394, %f1037, %p1;
setp.geu.f32 %p53, %f85, 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 %f1039, 0f7FFFFFFF;
$L__BB0_44:
add.f32 %f399, %f89, 0f3EE66666;
mov.b32 %r223, %f399;
setp.lt.s32 %p56, %r223, 2139095040;
@%p56 bra $L__BB0_49;
setp.gtu.f32 %p57, %f89, 0f7F800000;
@%p57 bra $L__BB0_48;
bra.uni $L__BB0_46;
$L__BB0_48:
add.f32 %f1039, %f85, 0f3EE66666;
bra.uni $L__BB0_49;
$L__BB0_46:
setp.neu.f32 %p58, %f89, 0f7F800000;
@%p58 bra $L__BB0_49;
selp.f32 %f1039, 0fFF800000, 0f7F800000, %p1;
$L__BB0_49:
setp.eq.f32 %p59, %f85, 0f3F800000;
selp.f32 %f100, 0f3F800000, %f1039, %p59;
abs.f32 %f101, %f86;
setp.lt.f32 %p60, %f101, 0f00800000;
mul.f32 %f401, %f101, 0f4B800000;
selp.f32 %f402, %f401, %f101, %p60;
selp.f32 %f403, 0fC3170000, 0fC2FE0000, %p60;
mov.b32 %r224, %f402;
and.b32 %r225, %r224, 8388607;
or.b32 %r226, %r225, 1065353216;
mov.b32 %f404, %r226;
shr.u32 %r227, %r224, 23;
cvt.rn.f32.u32 %f405, %r227;
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 %r228, %f455;
setp.eq.s32 %p62, %r228, 1118925336;
add.s32 %r229, %r228, -1;
mov.b32 %f459, %r229;
add.f32 %f460, %f458, 0f37000000;
selp.f32 %f102, %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 %r230, %f464;
and.b32 %r231, %r230, -2147483648;
or.b32 %r232, %r231, 1123811328;
mov.b32 %f466, %r232;
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 %r233, %f473;
shl.b32 %r234, %r233, 23;
mov.b32 %f474, %r234;
ex2.approx.ftz.f32 %f475, %f472;
mul.f32 %f103, %f475, %f474;
setp.eq.f32 %p64, %f103, 0f7F800000;
mov.f32 %f1040, 0f7F800000;
@%p64 bra $L__BB0_51;
fma.rn.f32 %f1040, %f103, %f102, %f103;
$L__BB0_51:
setp.lt.f32 %p65, %f86, 0f00000000;
and.pred %p2, %p65, %p51;
setp.eq.f32 %p67, %f86, 0f00000000;
@%p67 bra $L__BB0_55;
bra.uni $L__BB0_52;
$L__BB0_55:
add.f32 %f480, %f86, %f86;
selp.f32 %f1042, %f480, 0f00000000, %p51;
bra.uni $L__BB0_56;
$L__BB0_52:
mov.b32 %r235, %f1040;
xor.b32 %r236, %r235, -2147483648;
mov.b32 %f476, %r236;
selp.f32 %f1042, %f476, %f1040, %p2;
setp.geu.f32 %p68, %f86, 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 %f1042, 0f7FFFFFFF;
$L__BB0_56:
add.f32 %f481, %f101, 0f3EE66666;
mov.b32 %r237, %f481;
setp.lt.s32 %p71, %r237, 2139095040;
@%p71 bra $L__BB0_61;
setp.gtu.f32 %p72, %f101, 0f7F800000;
@%p72 bra $L__BB0_60;
bra.uni $L__BB0_58;
$L__BB0_60:
add.f32 %f1042, %f86, 0f3EE66666;
bra.uni $L__BB0_61;
$L__BB0_58:
setp.neu.f32 %p73, %f101, 0f7F800000;
@%p73 bra $L__BB0_61;
selp.f32 %f1042, 0fFF800000, 0f7F800000, %p2;
$L__BB0_61:
setp.eq.f32 %p74, %f86, 0f3F800000;
selp.f32 %f112, 0f3F800000, %f1042, %p74;
abs.f32 %f113, %f87;
setp.lt.f32 %p75, %f113, 0f00800000;
mul.f32 %f483, %f113, 0f4B800000;
selp.f32 %f484, %f483, %f113, %p75;
selp.f32 %f485, 0fC3170000, 0fC2FE0000, %p75;
mov.b32 %r238, %f484;
and.b32 %r239, %r238, 8388607;
or.b32 %r240, %r239, 1065353216;
mov.b32 %f486, %r240;
shr.u32 %r241, %r238, 23;
cvt.rn.f32.u32 %f487, %r241;
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 %r242, %f537;
setp.eq.s32 %p77, %r242, 1118925336;
add.s32 %r243, %r242, -1;
mov.b32 %f541, %r243;
add.f32 %f542, %f540, 0f37000000;
selp.f32 %f114, %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 %r244, %f546;
and.b32 %r245, %r244, -2147483648;
or.b32 %r246, %r245, 1123811328;
mov.b32 %f548, %r246;
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 %r247, %f555;
shl.b32 %r248, %r247, 23;
mov.b32 %f556, %r248;
ex2.approx.ftz.f32 %f557, %f554;
mul.f32 %f115, %f557, %f556;
setp.eq.f32 %p79, %f115, 0f7F800000;
mov.f32 %f1043, 0f7F800000;
@%p79 bra $L__BB0_63;
fma.rn.f32 %f1043, %f115, %f114, %f115;
$L__BB0_63:
setp.lt.f32 %p80, %f87, 0f00000000;
and.pred %p3, %p80, %p51;
setp.eq.f32 %p82, %f87, 0f00000000;
@%p82 bra $L__BB0_67;
bra.uni $L__BB0_64;
$L__BB0_67:
add.f32 %f562, %f87, %f87;
selp.f32 %f1045, %f562, 0f00000000, %p51;
bra.uni $L__BB0_68;
$L__BB0_64:
mov.b32 %r249, %f1043;
xor.b32 %r250, %r249, -2147483648;
mov.b32 %f558, %r250;
selp.f32 %f1045, %f558, %f1043, %p3;
setp.geu.f32 %p83, %f87, 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 %f1045, 0f7FFFFFFF;
$L__BB0_68:
add.f32 %f563, %f113, 0f3EE66666;
mov.b32 %r251, %f563;
setp.lt.s32 %p86, %r251, 2139095040;
@%p86 bra $L__BB0_73;
setp.gtu.f32 %p87, %f113, 0f7F800000;
@%p87 bra $L__BB0_72;
bra.uni $L__BB0_70;
$L__BB0_72:
add.f32 %f1045, %f87, 0f3EE66666;
bra.uni $L__BB0_73;
$L__BB0_70:
setp.neu.f32 %p88, %f113, 0f7F800000;
@%p88 bra $L__BB0_73;
selp.f32 %f1045, 0fFF800000, 0f7F800000, %p3;
$L__BB0_73:
setp.eq.f32 %p89, %f87, 0f3F800000;
mov.f32 %f565, 0f3F800000;
selp.f32 %f566, 0f3F800000, %f1045, %p89;
min.f32 %f567, %f100, %f565;
mov.f32 %f568, 0f00000000;
max.f32 %f124, %f568, %f567;
min.f32 %f569, %f112, %f565;
max.f32 %f125, %f568, %f569;
min.f32 %f570, %f566, %f565;
max.f32 %f126, %f568, %f570;
mov.f32 %f574, 0f3ED55555;
abs.f32 %f128, %f124;
setp.lt.f32 %p90, %f128, 0f00800000;
mul.f32 %f576, %f128, 0f4B800000;
selp.f32 %f577, %f576, %f128, %p90;
selp.f32 %f578, 0fC3170000, 0fC2FE0000, %p90;
mov.b32 %r254, %f577;
and.b32 %r255, %r254, 8388607;
or.b32 %r256, %r255, 1065353216;
mov.b32 %f579, %r256;
shr.u32 %r257, %r254, 23;
cvt.rn.f32.u32 %f580, %r257;
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 %r258, %f628;
setp.eq.s32 %p92, %r258, 1118925336;
add.s32 %r259, %r258, -1;
mov.b32 %f632, %r259;
add.f32 %f633, %f631, 0f37000000;
selp.f32 %f129, %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 %r260, %f637;
and.b32 %r261, %r260, -2147483648;
or.b32 %r262, %r261, 1123811328;
mov.b32 %f639, %r262;
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 %r263, %f646;
shl.b32 %r264, %r263, 23;
mov.b32 %f647, %r264;
ex2.approx.ftz.f32 %f648, %f645;
mul.f32 %f130, %f648, %f647;
setp.eq.f32 %p94, %f130, 0f7F800000;
mov.f32 %f1046, 0f7F800000;
@%p94 bra $L__BB0_75;
fma.rn.f32 %f1046, %f130, %f129, %f130;
$L__BB0_75:
mov.f32 %f957, 0f00000000;
max.f32 %f956, %f957, %f567;
mov.f32 %f955, 0f3E555555;
cvt.rzi.f32.f32 %f954, %f955;
add.f32 %f953, %f954, %f954;
mov.f32 %f952, 0f3ED55555;
sub.f32 %f951, %f952, %f953;
abs.f32 %f950, %f951;
setp.lt.f32 %p95, %f956, 0f00000000;
setp.eq.f32 %p96, %f950, 0f3F800000;
and.pred %p4, %p95, %p96;
setp.eq.f32 %p97, %f956, 0f00000000;
@%p97 bra $L__BB0_79;
bra.uni $L__BB0_76;
$L__BB0_79:
mov.f32 %f978, 0f00000000;
max.f32 %f977, %f978, %f567;
add.f32 %f653, %f977, %f977;
selp.f32 %f1048, %f653, 0f00000000, %p96;
bra.uni $L__BB0_80;
$L__BB0_76:
mov.f32 %f959, 0f00000000;
max.f32 %f958, %f959, %f567;
mov.b32 %r265, %f1046;
xor.b32 %r266, %r265, -2147483648;
mov.b32 %f649, %r266;
selp.f32 %f1048, %f649, %f1046, %p4;
setp.geu.f32 %p98, %f958, 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 %f1048, 0f7FFFFFFF;
$L__BB0_80:
mov.f32 %f962, 0f00000000;
max.f32 %f961, %f962, %f567;
abs.f32 %f960, %f961;
add.f32 %f654, %f960, 0f3ED55555;
mov.b32 %r267, %f654;
setp.lt.s32 %p101, %r267, 2139095040;
@%p101 bra $L__BB0_85;
mov.f32 %f971, 0f00000000;
max.f32 %f970, %f971, %f567;
abs.f32 %f969, %f970;
setp.gtu.f32 %p102, %f969, 0f7F800000;
@%p102 bra $L__BB0_84;
bra.uni $L__BB0_82;
$L__BB0_84:
mov.f32 %f976, 0f00000000;
max.f32 %f975, %f976, %f567;
add.f32 %f1048, %f975, 0f3ED55555;
bra.uni $L__BB0_85;
$L__BB0_82:
mov.f32 %f974, 0f00000000;
max.f32 %f973, %f974, %f567;
abs.f32 %f972, %f973;
setp.neu.f32 %p103, %f972, 0f7F800000;
@%p103 bra $L__BB0_85;
selp.f32 %f1048, 0fFF800000, 0f7F800000, %p4;
$L__BB0_85:
abs.f32 %f139, %f125;
setp.lt.f32 %p104, %f139, 0f00800000;
mul.f32 %f656, %f139, 0f4B800000;
selp.f32 %f657, %f656, %f139, %p104;
selp.f32 %f658, 0fC3170000, 0fC2FE0000, %p104;
mov.b32 %r268, %f657;
and.b32 %r269, %r268, 8388607;
or.b32 %r270, %r269, 1065353216;
mov.b32 %f659, %r270;
shr.u32 %r271, %r268, 23;
cvt.rn.f32.u32 %f660, %r271;
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 %r272, %f710;
setp.eq.s32 %p106, %r272, 1118925336;
add.s32 %r273, %r272, -1;
mov.b32 %f714, %r273;
add.f32 %f715, %f713, 0f37000000;
selp.f32 %f140, %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 %r274, %f719;
and.b32 %r275, %r274, -2147483648;
or.b32 %r276, %r275, 1123811328;
mov.b32 %f721, %r276;
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 %r277, %f728;
shl.b32 %r278, %r277, 23;
mov.b32 %f729, %r278;
ex2.approx.ftz.f32 %f730, %f727;
mul.f32 %f141, %f730, %f729;
setp.eq.f32 %p108, %f141, 0f7F800000;
mov.f32 %f1049, 0f7F800000;
@%p108 bra $L__BB0_87;
fma.rn.f32 %f1049, %f141, %f140, %f141;
$L__BB0_87:
setp.lt.f32 %p109, %f125, 0f00000000;
and.pred %p5, %p109, %p96;
setp.eq.f32 %p111, %f125, 0f00000000;
@%p111 bra $L__BB0_91;
bra.uni $L__BB0_88;
$L__BB0_91:
add.f32 %f735, %f125, %f125;
selp.f32 %f1051, %f735, 0f00000000, %p96;
bra.uni $L__BB0_92;
$L__BB0_88:
mov.b32 %r279, %f1049;
xor.b32 %r280, %r279, -2147483648;
mov.b32 %f731, %r280;
selp.f32 %f1051, %f731, %f1049, %p5;
setp.geu.f32 %p112, %f125, 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 %f1051, 0f7FFFFFFF;
$L__BB0_92:
mov.f32 %f992, 0f00000000;
max.f32 %f991, %f992, %f569;
abs.f32 %f990, %f991;
add.f32 %f736, %f990, 0f3ED55555;
mov.b32 %r281, %f736;
setp.lt.s32 %p115, %r281, 2139095040;
@%p115 bra $L__BB0_97;
mov.f32 %f995, 0f00000000;
max.f32 %f994, %f995, %f569;
abs.f32 %f993, %f994;
setp.gtu.f32 %p116, %f993, 0f7F800000;
@%p116 bra $L__BB0_96;
bra.uni $L__BB0_94;
$L__BB0_96:
add.f32 %f1051, %f125, 0f3ED55555;
bra.uni $L__BB0_97;
$L__BB0_94:
mov.f32 %f998, 0f00000000;
max.f32 %f997, %f998, %f569;
abs.f32 %f996, %f997;
setp.neu.f32 %p117, %f996, 0f7F800000;
@%p117 bra $L__BB0_97;
selp.f32 %f1051, 0fFF800000, 0f7F800000, %p5;
$L__BB0_97:
abs.f32 %f150, %f126;
setp.lt.f32 %p118, %f150, 0f00800000;
mul.f32 %f738, %f150, 0f4B800000;
selp.f32 %f739, %f738, %f150, %p118;
selp.f32 %f740, 0fC3170000, 0fC2FE0000, %p118;
mov.b32 %r282, %f739;
and.b32 %r283, %r282, 8388607;
or.b32 %r284, %r283, 1065353216;
mov.b32 %f741, %r284;
shr.u32 %r285, %r282, 23;
cvt.rn.f32.u32 %f742, %r285;
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 %r286, %f792;
setp.eq.s32 %p120, %r286, 1118925336;
add.s32 %r287, %r286, -1;
mov.b32 %f796, %r287;
add.f32 %f797, %f795, 0f37000000;
selp.f32 %f151, %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 %r288, %f801;
and.b32 %r289, %r288, -2147483648;
or.b32 %r290, %r289, 1123811328;
mov.b32 %f803, %r290;
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 %r291, %f810;
shl.b32 %r292, %r291, 23;
mov.b32 %f811, %r292;
ex2.approx.ftz.f32 %f812, %f809;
mul.f32 %f152, %f812, %f811;
setp.eq.f32 %p122, %f152, 0f7F800000;
mov.f32 %f1052, 0f7F800000;
@%p122 bra $L__BB0_99;
fma.rn.f32 %f1052, %f152, %f151, %f152;
$L__BB0_99:
setp.lt.f32 %p123, %f126, 0f00000000;
and.pred %p6, %p123, %p96;
setp.eq.f32 %p125, %f126, 0f00000000;
@%p125 bra $L__BB0_103;
bra.uni $L__BB0_100;
$L__BB0_103:
add.f32 %f817, %f126, %f126;
selp.f32 %f1054, %f817, 0f00000000, %p96;
bra.uni $L__BB0_104;
$L__BB0_100:
mov.b32 %r293, %f1052;
xor.b32 %r294, %r293, -2147483648;
mov.b32 %f813, %r294;
selp.f32 %f1054, %f813, %f1052, %p6;
setp.geu.f32 %p126, %f126, 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 %f1054, 0f7FFFFFFF;
$L__BB0_104:
mov.f32 %f1001, 0f00000000;
max.f32 %f1000, %f1001, %f570;
abs.f32 %f999, %f1000;
add.f32 %f818, %f999, 0f3ED55555;
mov.b32 %r295, %f818;
setp.lt.s32 %p129, %r295, 2139095040;
@%p129 bra $L__BB0_109;
mov.f32 %f1004, 0f00000000;
max.f32 %f1003, %f1004, %f570;
abs.f32 %f1002, %f1003;
setp.gtu.f32 %p130, %f1002, 0f7F800000;
@%p130 bra $L__BB0_108;
bra.uni $L__BB0_106;
$L__BB0_108:
add.f32 %f1054, %f126, 0f3ED55555;
bra.uni $L__BB0_109;
$L__BB0_106:
mov.f32 %f1007, 0f00000000;
max.f32 %f1006, %f1007, %f570;
abs.f32 %f1005, %f1006;
setp.neu.f32 %p131, %f1005, 0f7F800000;
@%p131 bra $L__BB0_109;
selp.f32 %f1054, 0fFF800000, 0f7F800000, %p6;
$L__BB0_109:
mov.f32 %f964, 0f00000000;
max.f32 %f963, %f964, %f567;
ld.const.u64 %rd103, [params+144];
cvta.to.global.u64 %rd102, %rd103;
ld.const.u32 %r330, [params+136];
mad.lo.s32 %r329, %r330, %r5, %r4;
cvt.u64.u32 %rd101, %r329;
fma.rn.f32 %f819, %f1048, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p132, %f963, 0f3F800000;
mov.f32 %f820, 0f3F800000;
selp.f32 %f821, 0f3F7FFFFF, %f819, %p132;
mul.f32 %f822, %f963, 0f414EB852;
setp.lt.f32 %p133, %f963, 0f3B4D2E1C;
selp.f32 %f823, %f822, %f821, %p133;
fma.rn.f32 %f824, %f1051, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p134, %f125, 0f3F800000;
selp.f32 %f825, 0f3F7FFFFF, %f824, %p134;
mul.f32 %f826, %f125, 0f414EB852;
setp.lt.f32 %p135, %f125, 0f3B4D2E1C;
selp.f32 %f827, %f826, %f825, %p135;
fma.rn.f32 %f828, %f1054, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p136, %f126, 0f3F800000;
selp.f32 %f829, 0f3F7FFFFF, %f828, %p136;
mul.f32 %f830, %f126, 0f414EB852;
setp.lt.f32 %p137, %f126, 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 %r296, %f835;
min.u32 %r297, %r296, 255;
min.f32 %f836, %f827, %f820;
max.f32 %f837, %f833, %f836;
mul.f32 %f838, %f837, 0f43800000;
cvt.rzi.u32.f32 %r298, %f838;
min.u32 %r299, %r298, 255;
min.f32 %f839, %f831, %f820;
max.f32 %f840, %f833, %f839;
mul.f32 %f841, %f840, 0f43800000;
cvt.rzi.u32.f32 %r300, %f841;
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 %r331, [params+104];
and.b32 %r302, %r331, 4;
setp.eq.s32 %p138, %r302, 0;
ld.const.u32 %r343, [params+108];
@%p138 bra $L__BB0_114;
setp.eq.s32 %p139, %r343, 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;
add.f32 %f987, %f308, %f308;
mul.f32 %f986, %f987, %f310;
add.f32 %f966, %f308, %f308;
mul.f32 %f965, %f966, %f311;
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, %f986, %f842;
add.f32 %f846, %f965, %f843;
add.f32 %f847, %f87, %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:
add.f32 %f989, %f308, %f308;
mul.f32 %f988, %f989, %f310;
add.f32 %f968, %f308, %f308;
mul.f32 %f967, %f968, %f311;
mov.f32 %f852, 0f3F800000;
mov.u32 %r343, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs33, %f852;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs32, %f87;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs31, %f967;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f988;}
// end inline asm
st.global.v4.u16 [%rd18], {%rs30, %rs31, %rs32, %rs33};
$L__BB0_114:
mul.f32 %f161, %f78, 0f3F000000;
mul.f32 %f162, %f161, %f310;
mul.f32 %f163, %f161, %f311;
mul.f32 %f164, %f161, %f312;
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, %r343, 0;
@%p140 bra $L__BB0_116;
ld.global.v4.u16 {%rs41, %rs42, %rs43, %rs44}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f853, %rs41;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f854, %rs42;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f855, %rs43;}
// end inline asm
add.f32 %f856, %f162, %f853;
add.f32 %f857, %f163, %f854;
add.f32 %f858, %f164, %f855;
mov.f32 %f859, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs39, %f858;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f857;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f856;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs40, %f859;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs37, %rs38, %rs39, %rs40};
bra.uni $L__BB0_117;
$L__BB0_116:
mov.f32 %f863, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs48, %f863;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs47, %f164;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs46, %f163;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f162;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs45, %rs46, %rs47, %rs48};
$L__BB0_117:
mov.f32 %f864, 0f34000000;
max.f32 %f865, %f161, %f864;
mul.f32 %f866, %f79, 0f3F000000;
div.rn.f32 %f867, %f866, %f865;
fma.rn.f32 %f165, %f867, 0f3F000000, 0f3F000000;
mul.f32 %f868, %f80, 0f3F000000;
div.rn.f32 %f869, %f868, %f865;
fma.rn.f32 %f166, %f869, 0f3F000000, 0f3F000000;
mul.f32 %f870, %f81, 0f3F000000;
div.rn.f32 %f871, %f870, %f865;
fma.rn.f32 %f167, %f871, 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 %f872, %rs56;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f873, %rs57;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f874, %rs58;}
// end inline asm
add.f32 %f875, %f165, %f872;
add.f32 %f876, %f165, %f873;
add.f32 %f877, %f165, %f874;
mov.f32 %f878, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs54, %f877;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs53, %f876;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs52, %f875;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs55, %f878;}
// end inline asm
st.global.v4.u16 [%rd20], {%rs52, %rs53, %rs54, %rs55};
bra.uni $L__BB0_120;
$L__BB0_119:
mov.f32 %f882, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs63, %f882;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs62, %f165;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs61, %f165;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs60, %f165;}
// 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 %f883, %rs71;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f884, %rs72;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f885, %rs73;}
// end inline asm
add.f32 %f886, %f166, %f883;
add.f32 %f887, %f166, %f884;
add.f32 %f888, %f166, %f885;
mov.f32 %f889, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs69, %f888;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs68, %f887;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs67, %f886;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs70, %f889;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs67, %rs68, %rs69, %rs70};
bra.uni $L__BB0_123;
$L__BB0_122:
mov.f32 %f893, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs78, %f893;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs77, %f166;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs76, %f166;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs75, %f166;}
// 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 %f894, %rs86;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f895, %rs87;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f896, %rs88;}
// end inline asm
add.f32 %f897, %f167, %f894;
add.f32 %f898, %f167, %f895;
add.f32 %f899, %f167, %f896;
mov.f32 %f900, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs84, %f899;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs83, %f898;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs82, %f897;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs85, %f900;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs82, %rs83, %rs84, %rs85};
bra.uni $L__BB0_144;
$L__BB0_125:
mov.f32 %f904, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs93, %f904;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs92, %f167;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs91, %f167;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs90, %f167;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs90, %rs91, %rs92, %rs93};
$L__BB0_144:
ret;
}