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

1805 lines
51 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<55>;
.reg .f32 %f<1026>;
.reg .b32 %r<342>;
.reg .f64 %fd<5>;
.reg .b64 %rd<102>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r51), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r52), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u32 %r3, [params+540];
ld.const.u64 %rd19, [params+400];
cvta.to.global.u64 %rd20, %rd19;
ld.const.u32 %r57, [params+392];
mad.lo.s32 %r58, %r57, %r52, %r51;
mul.wide.u32 %rd21, %r58, 4;
add.s64 %rd2, %rd20, %rd21;
ld.global.v2.u8 {%rs7, %rs54}, [%rd2];
or.b16 %rs9, %rs7, %rs54;
and.b16 %rs10, %rs9, 255;
setp.eq.s16 %p7, %rs10, 0;
@%p7 bra $L__BB0_2;
ld.global.u8 %rs53, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs53, [%rd2+2];
setp.eq.s16 %p8, %rs53, 0;
mov.f32 %f975, 0f00000000;
mov.u16 %rs54, 0;
mov.f32 %f976, %f975;
mov.f32 %f977, %f975;
@%p8 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f181, %rs7;
div.rn.f32 %f182, %f181, 0f437F0000;
fma.rn.f32 %f183, %f182, 0f40000000, 0fBF800000;
and.b16 %rs13, %rs54, 255;
cvt.rn.f32.u16 %f184, %rs13;
div.rn.f32 %f185, %f184, 0f437F0000;
fma.rn.f32 %f186, %f185, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f187, %rs53;
div.rn.f32 %f188, %f187, 0f437F0000;
fma.rn.f32 %f189, %f188, 0f40000000, 0fBF800000;
mul.f32 %f190, %f186, %f186;
fma.rn.f32 %f191, %f183, %f183, %f190;
fma.rn.f32 %f192, %f189, %f189, %f191;
sqrt.rn.f32 %f193, %f192;
rcp.rn.f32 %f194, %f193;
mul.f32 %f977, %f194, %f189;
mul.f32 %f976, %f194, %f186;
mul.f32 %f975, %f183, %f194;
$L__BB0_4:
ld.const.v2.u32 {%r59, %r60}, [params];
add.s32 %r4, %r59, %r51;
add.s32 %r5, %r60, %r52;
setp.eq.f32 %p9, %f975, 0f00000000;
setp.eq.f32 %p10, %f976, 0f00000000;
and.pred %p11, %p9, %p10;
setp.eq.f32 %p12, %f977, 0f00000000;
and.pred %p13, %p12, %p11;
@%p13 bra $L__BB0_123;
bra.uni $L__BB0_5;
$L__BB0_123:
ld.const.u32 %r50, [params+104];
and.b32 %r314, %r50, 1;
setp.eq.b32 %p147, %r314, 1;
mov.pred %p148, 0;
xor.pred %p149, %p147, %p148;
not.pred %p150, %p149;
@%p150 bra $L__BB0_125;
ld.const.u64 %rd78, [params+144];
cvta.to.global.u64 %rd79, %rd78;
ld.const.u32 %r315, [params+136];
mad.lo.s32 %r316, %r315, %r5, %r4;
mul.wide.u32 %rd80, %r316, 4;
add.s64 %rd81, %rd79, %rd80;
mov.u16 %rs43, 0;
st.global.v4.u8 [%rd81], {%rs43, %rs43, %rs43, %rs43};
$L__BB0_125:
and.b32 %r317, %r50, 4;
setp.eq.s32 %p151, %r317, 0;
@%p151 bra $L__BB0_127;
ld.const.u64 %rd82, [params+224];
cvta.to.global.u64 %rd83, %rd82;
ld.const.u32 %r318, [params+216];
mad.lo.s32 %r319, %r318, %r5, %r4;
mov.f32 %f943, 0f00000000;
mul.wide.u32 %rd84, %r319, 8;
add.s64 %rd85, %rd83, %rd84;
// begin inline asm
{ cvt.rn.f16.f32 %rs46, %f943;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f943;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f943;}
// end inline asm
mov.u16 %rs47, 0;
st.global.v4.u16 [%rd85], {%rs44, %rs45, %rs46, %rs47};
$L__BB0_127:
and.b32 %r320, %r50, 16;
setp.eq.s32 %p152, %r320, 0;
@%p152 bra $L__BB0_129;
ld.const.u64 %rd86, [params+240];
cvta.to.global.u64 %rd87, %rd86;
ld.const.u32 %r321, [params+232];
mad.lo.s32 %r322, %r321, %r5, %r4;
mov.f32 %f946, 0f00000000;
mul.wide.u32 %rd88, %r322, 8;
add.s64 %rd89, %rd87, %rd88;
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f946;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs49, %f946;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs48, %f946;}
// end inline asm
mov.u16 %rs51, 0;
st.global.v4.u16 [%rd89], {%rs48, %rs49, %rs50, %rs51};
$L__BB0_129:
and.b32 %r323, %r50, 64;
setp.eq.s32 %p153, %r323, 0;
@%p153 bra $L__BB0_131;
ld.const.u64 %rd90, [params+208];
cvta.to.global.u64 %rd91, %rd90;
ld.const.u32 %r324, [params+200];
mad.lo.s32 %r325, %r324, %r5, %r4;
mul.wide.u32 %rd92, %r325, 4;
add.s64 %rd93, %rd91, %rd92;
mov.u16 %rs52, 0;
st.global.v4.u8 [%rd93], {%rs52, %rs52, %rs52, %rs52};
bra.uni $L__BB0_131;
$L__BB0_5:
ld.const.u64 %rd22, [params+496];
cvta.to.global.u64 %rd23, %rd22;
ld.const.u32 %r63, [params+488];
mad.lo.s32 %r64, %r63, %r52, %r51;
mul.wide.u32 %rd24, %r64, 8;
add.s64 %rd25, %rd23, %rd24;
ld.global.v4.u16 {%rs19, %rs20, %rs21, %rs22}, [%rd25];
// begin inline asm
{ cvt.f32.f16 %f195, %rs19;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f196, %rs20;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f197, %rs21;}
// end inline asm
ld.const.u64 %rd26, [params+432];
cvta.to.global.u64 %rd27, %rd26;
ld.const.u32 %r65, [params+424];
mad.lo.s32 %r66, %r65, %r52, %r51;
mul.wide.u32 %rd28, %r66, 12;
add.s64 %rd29, %rd27, %rd28;
ld.global.f32 %f205, [%rd29];
mul.f32 %f206, %f205, 0f3456BF95;
ld.global.f32 %f207, [%rd29+4];
mul.f32 %f208, %f207, 0f3456BF95;
ld.global.f32 %f209, [%rd29+8];
mul.f32 %f210, %f209, 0f3456BF95;
abs.f32 %f211, %f975;
div.rn.f32 %f212, %f206, %f211;
abs.f32 %f213, %f976;
div.rn.f32 %f214, %f208, %f213;
abs.f32 %f215, %f977;
div.rn.f32 %f216, %f210, %f215;
abs.f32 %f217, %f212;
abs.f32 %f218, %f214;
abs.f32 %f219, %f216;
mov.f32 %f220, 0f38D1B717;
max.f32 %f221, %f217, %f220;
max.f32 %f222, %f218, %f220;
max.f32 %f223, %f219, %f220;
fma.rn.f32 %f13, %f975, %f221, %f205;
fma.rn.f32 %f14, %f976, %f222, %f207;
fma.rn.f32 %f15, %f977, %f223, %f209;
setp.gt.f32 %p14, %f211, %f215;
neg.f32 %f224, %f976;
selp.f32 %f225, %f224, 0f00000000, %p14;
mov.f32 %f989, 0f00000000;
neg.f32 %f226, %f977;
selp.f32 %f227, %f975, %f226, %p14;
selp.f32 %f228, 0f00000000, %f976, %p14;
mul.f32 %f229, %f227, %f227;
fma.rn.f32 %f230, %f225, %f225, %f229;
fma.rn.f32 %f231, %f228, %f228, %f230;
sqrt.rn.f32 %f232, %f231;
rcp.rn.f32 %f233, %f232;
mul.f32 %f16, %f225, %f233;
mul.f32 %f17, %f227, %f233;
mul.f32 %f18, %f228, %f233;
ld.const.u64 %rd30, [params+128];
cvta.to.global.u64 %rd31, %rd30;
ld.const.u32 %r67, [params+120];
mad.lo.s32 %r68, %r67, %r52, %r51;
mul.wide.u32 %rd32, %r68, 4;
add.s64 %rd3, %rd31, %rd32;
setp.lt.s32 %p15, %r3, 1;
mov.f32 %f988, %f989;
mov.f32 %f987, %f989;
mov.f32 %f986, %f989;
mov.f32 %f985, %f989;
mov.f32 %f984, %f989;
@%p15 bra $L__BB0_34;
cvt.rn.f32.s32 %f240, %r3;
rcp.rn.f32 %f19, %f240;
ld.global.u32 %r334, [%rd3];
mul.f32 %f20, %f13, 0f3456BF95;
mul.f32 %f21, %f14, 0f3456BF95;
mul.f32 %f22, %f15, 0f3456BF95;
ld.const.u64 %rd4, [params+96];
mul.f32 %f241, %f975, %f17;
mul.f32 %f242, %f976, %f16;
sub.f32 %f23, %f242, %f241;
mul.f32 %f243, %f977, %f16;
mul.f32 %f244, %f975, %f18;
sub.f32 %f24, %f244, %f243;
mul.f32 %f245, %f976, %f18;
mul.f32 %f246, %f977, %f17;
sub.f32 %f25, %f246, %f245;
mov.u32 %r69, 0;
add.s64 %rd5, %rd1, 24;
mov.u64 %rd33, __cudart_i2opi_f;
abs.f32 %f309, %f21;
abs.f32 %f310, %f20;
max.f32 %f311, %f310, %f309;
abs.f32 %f312, %f22;
max.f32 %f313, %f311, %f312;
mov.u32 %r331, %r69;
$L__BB0_7:
mov.u32 %r333, %r69;
$L__BB0_8:
cvt.rn.f32.s32 %f959, %r331;
mad.lo.s32 %r71, %r334, 1664525, 1013904223;
and.b32 %r72, %r71, 16777215;
cvt.rn.f32.u32 %f247, %r72;
fma.rn.f32 %f248, %f247, 0f33800000, %f959;
mul.f32 %f39, %f19, %f248;
mad.lo.s32 %r334, %r71, 1664525, 1013904223;
and.b32 %r73, %r334, 16777215;
cvt.rn.f32.u32 %f249, %r73;
cvt.rn.f32.s32 %f250, %r333;
fma.rn.f32 %f251, %f249, 0f33800000, %f250;
mul.f32 %f252, %f19, %f251;
mul.f32 %f253, %f39, %f39;
mov.f32 %f254, 0f3F800000;
sub.f32 %f255, %f254, %f253;
mov.f32 %f256, 0f00000000;
max.f32 %f257, %f256, %f255;
sqrt.rn.f32 %f40, %f257;
mul.f32 %f41, %f252, 0f40C90FDB;
mul.f32 %f258, %f41, 0f3F22F983;
cvt.rni.s32.f32 %r341, %f258;
cvt.rn.f32.s32 %f259, %r341;
mov.f32 %f260, 0fBFC90FDA;
fma.rn.f32 %f261, %f259, %f260, %f41;
mov.f32 %f262, 0fB3A22168;
fma.rn.f32 %f263, %f259, %f262, %f261;
mov.f32 %f264, 0fA7C234C5;
fma.rn.f32 %f993, %f259, %f264, %f263;
abs.f32 %f43, %f41;
setp.ltu.f32 %p16, %f43, 0f47CE4780;
mov.u32 %r338, %r341;
mov.f32 %f990, %f993;
@%p16 bra $L__BB0_16;
setp.eq.f32 %p17, %f43, 0f7F800000;
@%p17 bra $L__BB0_15;
bra.uni $L__BB0_10;
$L__BB0_15:
mov.f32 %f267, 0f00000000;
mul.rn.f32 %f990, %f41, %f267;
mov.u32 %r338, 0;
bra.uni $L__BB0_16;
$L__BB0_10:
mov.b32 %r13, %f41;
bfe.u32 %r75, %r13, 23, 8;
add.s32 %r14, %r75, -128;
shl.b32 %r76, %r13, 8;
or.b32 %r15, %r76, -2147483648;
shr.u32 %r16, %r14, 5;
mov.u64 %rd99, 0;
mov.u32 %r335, 0;
mov.u64 %rd97, %rd1;
mov.u64 %rd98, %rd33;
$L__BB0_11:
.pragma "nounroll";
ld.global.nc.u32 %r77, [%rd98];
mad.wide.u32 %rd35, %r77, %r15, %rd99;
shr.u64 %rd99, %rd35, 32;
st.local.u32 [%rd97], %rd35;
add.s64 %rd98, %rd98, 4;
add.s64 %rd97, %rd97, 4;
add.s32 %r335, %r335, 1;
setp.ne.s32 %p18, %r335, 6;
@%p18 bra $L__BB0_11;
st.local.u32 [%rd5], %rd99;
mov.u32 %r78, 4;
sub.s32 %r19, %r78, %r16;
mov.u32 %r79, 6;
sub.s32 %r80, %r79, %r16;
mul.wide.s32 %rd36, %r80, 4;
add.s64 %rd37, %rd1, %rd36;
ld.local.u32 %r336, [%rd37];
ld.local.u32 %r337, [%rd37+-4];
and.b32 %r22, %r14, 31;
setp.eq.s32 %p19, %r22, 0;
@%p19 bra $L__BB0_14;
mov.u32 %r81, 32;
sub.s32 %r82, %r81, %r22;
shr.u32 %r83, %r337, %r82;
shl.b32 %r84, %r336, %r22;
add.s32 %r336, %r83, %r84;
mul.wide.s32 %rd38, %r19, 4;
add.s64 %rd39, %rd1, %rd38;
ld.local.u32 %r85, [%rd39];
shr.u32 %r86, %r85, %r82;
shl.b32 %r87, %r337, %r22;
add.s32 %r337, %r86, %r87;
$L__BB0_14:
and.b32 %r88, %r13, -2147483648;
shr.u32 %r89, %r337, 30;
shl.b32 %r90, %r336, 2;
or.b32 %r91, %r89, %r90;
shr.u32 %r92, %r91, 31;
shr.u32 %r93, %r336, 30;
add.s32 %r94, %r92, %r93;
neg.s32 %r95, %r94;
setp.eq.s32 %p20, %r88, 0;
selp.b32 %r338, %r94, %r95, %p20;
setp.ne.s32 %p21, %r92, 0;
xor.b32 %r96, %r88, -2147483648;
selp.b32 %r97, %r96, %r88, %p21;
selp.b32 %r98, -1, 0, %p21;
xor.b32 %r99, %r91, %r98;
shl.b32 %r100, %r337, 2;
xor.b32 %r101, %r100, %r98;
cvt.u64.u32 %rd40, %r99;
cvt.u64.u32 %rd41, %r101;
bfi.b64 %rd42, %rd40, %rd41, 32, 32;
cvt.rn.f64.s64 %fd1, %rd42;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f265, %fd2;
setp.eq.s32 %p22, %r97, 0;
neg.f32 %f266, %f265;
selp.f32 %f990, %f265, %f266, %p22;
$L__BB0_16:
add.s32 %r29, %r338, 1;
and.b32 %r30, %r29, 1;
setp.eq.s32 %p23, %r30, 0;
selp.f32 %f47, %f990, 0f3F800000, %p23;
mul.rn.f32 %f48, %f990, %f990;
mov.f32 %f991, 0fB94D4153;
@%p23 bra $L__BB0_18;
mov.f32 %f269, 0fBAB607ED;
mov.f32 %f270, 0f37CBAC00;
fma.rn.f32 %f991, %f270, %f48, %f269;
$L__BB0_18:
selp.f32 %f271, 0f3C0885E4, 0f3D2AAABB, %p23;
fma.rn.f32 %f272, %f991, %f48, %f271;
selp.f32 %f273, 0fBE2AAAA8, 0fBEFFFFFF, %p23;
fma.rn.f32 %f274, %f272, %f48, %f273;
mov.f32 %f275, 0f00000000;
fma.rn.f32 %f276, %f48, %f47, %f275;
fma.rn.f32 %f992, %f274, %f276, %f47;
and.b32 %r103, %r29, 2;
setp.eq.s32 %p25, %r103, 0;
@%p25 bra $L__BB0_20;
mov.f32 %f278, 0fBF800000;
fma.rn.f32 %f992, %f992, %f278, %f275;
$L__BB0_20:
@%p16 bra $L__BB0_28;
setp.eq.f32 %p27, %f43, 0f7F800000;
@%p27 bra $L__BB0_27;
bra.uni $L__BB0_22;
$L__BB0_27:
mov.f32 %f281, 0f00000000;
mul.rn.f32 %f993, %f41, %f281;
mov.u32 %r341, 0;
bra.uni $L__BB0_28;
$L__BB0_22:
mov.b32 %r31, %f41;
bfe.u32 %r104, %r31, 23, 8;
add.s32 %r32, %r104, -128;
shl.b32 %r105, %r31, 8;
or.b32 %r33, %r105, -2147483648;
shr.u32 %r34, %r32, 5;
mov.u64 %rd100, 0;
mov.u64 %rd101, %rd100;
$L__BB0_23:
.pragma "nounroll";
shl.b64 %rd45, %rd100, 2;
mov.u64 %rd46, __cudart_i2opi_f;
add.s64 %rd47, %rd46, %rd45;
ld.global.nc.u32 %r106, [%rd47];
mad.wide.u32 %rd48, %r106, %r33, %rd101;
shr.u64 %rd101, %rd48, 32;
add.s64 %rd49, %rd1, %rd45;
st.local.u32 [%rd49], %rd48;
cvt.u32.u64 %r107, %rd100;
add.s32 %r108, %r107, 1;
cvt.s64.s32 %rd100, %r108;
setp.ne.s32 %p28, %r108, 6;
@%p28 bra $L__BB0_23;
st.local.u32 [%rd5], %rd101;
mov.u32 %r109, 4;
sub.s32 %r35, %r109, %r34;
mov.u32 %r110, 6;
sub.s32 %r111, %r110, %r34;
mul.wide.s32 %rd50, %r111, 4;
add.s64 %rd51, %rd1, %rd50;
ld.local.u32 %r339, [%rd51];
ld.local.u32 %r340, [%rd51+-4];
and.b32 %r38, %r32, 31;
setp.eq.s32 %p29, %r38, 0;
@%p29 bra $L__BB0_26;
mov.u32 %r112, 32;
sub.s32 %r113, %r112, %r38;
shr.u32 %r114, %r340, %r113;
shl.b32 %r115, %r339, %r38;
add.s32 %r339, %r114, %r115;
mul.wide.s32 %rd52, %r35, 4;
add.s64 %rd53, %rd1, %rd52;
ld.local.u32 %r116, [%rd53];
shr.u32 %r117, %r116, %r113;
shl.b32 %r118, %r340, %r38;
add.s32 %r340, %r117, %r118;
$L__BB0_26:
and.b32 %r119, %r31, -2147483648;
shr.u32 %r120, %r340, 30;
shl.b32 %r121, %r339, 2;
or.b32 %r122, %r120, %r121;
shr.u32 %r123, %r122, 31;
shr.u32 %r124, %r339, 30;
add.s32 %r125, %r123, %r124;
neg.s32 %r126, %r125;
setp.eq.s32 %p30, %r119, 0;
selp.b32 %r341, %r125, %r126, %p30;
setp.ne.s32 %p31, %r123, 0;
xor.b32 %r127, %r119, -2147483648;
selp.b32 %r128, %r127, %r119, %p31;
selp.b32 %r129, -1, 0, %p31;
xor.b32 %r130, %r122, %r129;
shl.b32 %r131, %r340, 2;
xor.b32 %r132, %r131, %r129;
cvt.u64.u32 %rd54, %r130;
cvt.u64.u32 %rd55, %r132;
bfi.b64 %rd56, %rd54, %rd55, 32, 32;
cvt.rn.f64.s64 %fd3, %rd56;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f279, %fd4;
setp.eq.s32 %p32, %r128, 0;
neg.f32 %f280, %f279;
selp.f32 %f993, %f279, %f280, %p32;
$L__BB0_28:
mul.f32 %f57, %f40, %f992;
and.b32 %r45, %r341, 1;
setp.eq.s32 %p33, %r45, 0;
selp.f32 %f58, %f993, 0f3F800000, %p33;
mul.rn.f32 %f59, %f993, %f993;
mov.f32 %f994, 0fB94D4153;
@%p33 bra $L__BB0_30;
mov.f32 %f283, 0fBAB607ED;
mov.f32 %f284, 0f37CBAC00;
fma.rn.f32 %f994, %f284, %f59, %f283;
$L__BB0_30:
selp.f32 %f285, 0f3C0885E4, 0f3D2AAABB, %p33;
fma.rn.f32 %f286, %f994, %f59, %f285;
selp.f32 %f287, 0fBE2AAAA8, 0fBEFFFFFF, %p33;
fma.rn.f32 %f288, %f286, %f59, %f287;
mov.f32 %f289, 0f00000000;
fma.rn.f32 %f290, %f59, %f58, %f289;
fma.rn.f32 %f995, %f288, %f290, %f58;
and.b32 %r134, %r341, 2;
setp.eq.s32 %p35, %r134, 0;
@%p35 bra $L__BB0_32;
mov.f32 %f292, 0fBF800000;
fma.rn.f32 %f995, %f995, %f292, %f289;
$L__BB0_32:
mul.f32 %f302, %f40, %f995;
mul.f32 %f303, %f16, %f302;
mul.f32 %f304, %f17, %f302;
mul.f32 %f305, %f18, %f302;
fma.rn.f32 %f306, %f25, %f57, %f303;
fma.rn.f32 %f307, %f24, %f57, %f304;
fma.rn.f32 %f308, %f23, %f57, %f305;
fma.rn.f32 %f296, %f975, %f39, %f306;
fma.rn.f32 %f297, %f976, %f39, %f307;
fma.rn.f32 %f298, %f977, %f39, %f308;
mov.f32 %f314, 0f38D1B717;
max.f32 %f299, %f313, %f314;
mov.f32 %f300, 0f6C4ECB8F;
mov.u32 %r168, 1;
mov.u32 %r171, 2;
mov.u32 %r173, 3;
mov.u32 %r205, 0;
// begin inline asm
call(%r135,%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166),_optix_trace_typed_32,(%r205,%rd4,%f13,%f14,%f15,%f296,%f297,%f298,%f299,%f300,%f289,%r168,%r205,%r205,%r171,%r205,%r173,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205,%r205);
// end inline asm
mov.b32 %f315, %r135;
mov.b32 %f316, %r136;
mov.b32 %f317, %r137;
mul.f32 %f318, %f316, 0f3F372474;
fma.rn.f32 %f319, %f315, 0f3E59999A, %f318;
fma.rn.f32 %f320, %f317, 0f3D93A92A, %f319;
fma.rn.f32 %f989, %f296, %f320, %f989;
fma.rn.f32 %f988, %f297, %f320, %f988;
fma.rn.f32 %f987, %f298, %f320, %f987;
mul.f32 %f321, %f976, %f297;
fma.rn.f32 %f322, %f975, %f296, %f321;
fma.rn.f32 %f323, %f977, %f298, %f322;
cvt.sat.f32.f32 %f324, %f323;
fma.rn.f32 %f986, %f324, %f315, %f986;
fma.rn.f32 %f985, %f324, %f316, %f985;
fma.rn.f32 %f984, %f324, %f317, %f984;
add.s32 %r333, %r333, 1;
setp.lt.s32 %p36, %r333, %r3;
@%p36 bra $L__BB0_8;
add.s32 %r331, %r331, 1;
setp.lt.s32 %p37, %r331, %r3;
@%p37 bra $L__BB0_7;
$L__BB0_34:
mov.f32 %f1002, 0f00000000;
mul.lo.s32 %r206, %r3, %r3;
cvt.rn.f32.s32 %f328, %r206;
rcp.rn.f32 %f329, %f328;
mul.f32 %f330, %f329, %f986;
mul.f32 %f331, %f329, %f985;
mul.f32 %f332, %f329, %f984;
fma.rn.f32 %f333, %f329, %f986, %f330;
fma.rn.f32 %f334, %f329, %f985, %f331;
fma.rn.f32 %f335, %f329, %f984, %f332;
ld.const.u64 %rd58, [params+504];
cvt.rn.f32.u32 %f336, %r5;
cvt.rn.f32.u32 %f337, %r4;
tex.2d.v4.f32.f32 {%f338, %f339, %f340, %f341}, [%rd58, {%f337, %f336}];
mul.f32 %f77, %f333, %f338;
mul.f32 %f78, %f334, %f339;
mul.f32 %f79, %f335, %f340;
add.f32 %f80, %f195, %f77;
add.f32 %f81, %f196, %f78;
add.f32 %f82, %f197, %f79;
abs.f32 %f342, %f80;
setp.gtu.f32 %p38, %f342, 0f7F800000;
mov.f32 %f1003, %f1002;
mov.f32 %f1004, %f1002;
@%p38 bra $L__BB0_38;
mov.f32 %f1002, 0f00000000;
abs.f32 %f346, %f81;
setp.gtu.f32 %p39, %f346, 0f7F800000;
@%p39 bra $L__BB0_38;
mov.f32 %f1002, 0f00000000;
abs.f32 %f350, %f82;
setp.gtu.f32 %p40, %f350, 0f7F800000;
mov.f32 %f1003, %f1002;
mov.f32 %f1004, %f1002;
@%p40 bra $L__BB0_38;
mov.f32 %f1002, %f80;
mov.f32 %f1003, %f81;
mov.f32 %f1004, %f82;
$L__BB0_38:
ld.const.u32 %r48, [params+104];
and.b32 %r207, %r48, 1;
setp.eq.b32 %p41, %r207, 1;
mov.pred %p42, 0;
xor.pred %p43, %p41, %p42;
not.pred %p44, %p43;
@%p44 bra $L__BB0_112;
mov.f32 %f355, 0f3EE66666;
abs.f32 %f93, %f1002;
setp.lt.f32 %p45, %f93, 0f00800000;
mul.f32 %f357, %f93, 0f4B800000;
selp.f32 %f358, %f357, %f93, %p45;
selp.f32 %f359, 0fC3170000, 0fC2FE0000, %p45;
mov.b32 %r208, %f358;
and.b32 %r209, %r208, 8388607;
or.b32 %r210, %r209, 1065353216;
mov.b32 %f360, %r210;
shr.u32 %r211, %r208, 23;
cvt.rn.f32.u32 %f361, %r211;
add.f32 %f362, %f359, %f361;
setp.gt.f32 %p46, %f360, 0f3FB504F3;
mul.f32 %f363, %f360, 0f3F000000;
add.f32 %f364, %f362, 0f3F800000;
selp.f32 %f365, %f364, %f362, %p46;
selp.f32 %f366, %f363, %f360, %p46;
add.f32 %f367, %f366, 0fBF800000;
add.f32 %f368, %f366, 0f3F800000;
rcp.approx.ftz.f32 %f369, %f368;
add.f32 %f370, %f367, %f367;
mul.f32 %f371, %f370, %f369;
mul.f32 %f372, %f371, %f371;
mov.f32 %f373, 0f3C4CAF63;
mov.f32 %f374, 0f3B18F0FE;
fma.rn.f32 %f375, %f374, %f372, %f373;
mov.f32 %f376, 0f3DAAAABD;
fma.rn.f32 %f377, %f375, %f372, %f376;
mul.rn.f32 %f378, %f377, %f372;
mul.rn.f32 %f379, %f378, %f371;
sub.f32 %f380, %f367, %f371;
add.f32 %f381, %f380, %f380;
neg.f32 %f382, %f371;
fma.rn.f32 %f383, %f382, %f367, %f381;
mul.rn.f32 %f384, %f369, %f383;
add.f32 %f385, %f379, %f371;
sub.f32 %f386, %f371, %f385;
add.f32 %f387, %f379, %f386;
add.f32 %f388, %f384, %f387;
add.f32 %f389, %f385, %f388;
sub.f32 %f390, %f385, %f389;
add.f32 %f391, %f388, %f390;
mov.f32 %f392, 0f3F317200;
mul.rn.f32 %f393, %f365, %f392;
mov.f32 %f394, 0f35BFBE8E;
mul.rn.f32 %f395, %f365, %f394;
add.f32 %f396, %f393, %f389;
sub.f32 %f397, %f393, %f396;
add.f32 %f398, %f389, %f397;
add.f32 %f399, %f391, %f398;
add.f32 %f400, %f395, %f399;
add.f32 %f401, %f396, %f400;
sub.f32 %f402, %f396, %f401;
add.f32 %f403, %f400, %f402;
mul.rn.f32 %f404, %f355, %f401;
neg.f32 %f405, %f404;
fma.rn.f32 %f406, %f355, %f401, %f405;
fma.rn.f32 %f407, %f355, %f403, %f406;
mov.f32 %f408, 0f00000000;
fma.rn.f32 %f409, %f408, %f401, %f407;
add.rn.f32 %f410, %f404, %f409;
neg.f32 %f411, %f410;
add.rn.f32 %f412, %f404, %f411;
add.rn.f32 %f413, %f412, %f409;
mov.b32 %r212, %f410;
setp.eq.s32 %p47, %r212, 1118925336;
add.s32 %r213, %r212, -1;
mov.b32 %f414, %r213;
add.f32 %f415, %f413, 0f37000000;
selp.f32 %f94, %f415, %f413, %p47;
selp.f32 %f416, %f414, %f410, %p47;
mov.f32 %f417, 0f3FB8AA3B;
mul.rn.f32 %f418, %f416, %f417;
cvt.rzi.f32.f32 %f419, %f418;
abs.f32 %f420, %f419;
setp.gt.f32 %p48, %f420, 0f42FC0000;
mov.b32 %r214, %f419;
and.b32 %r215, %r214, -2147483648;
or.b32 %r216, %r215, 1123811328;
mov.b32 %f421, %r216;
selp.f32 %f422, %f421, %f419, %p48;
mov.f32 %f423, 0fBF317218;
fma.rn.f32 %f424, %f422, %f423, %f416;
mov.f32 %f425, 0f3102E308;
fma.rn.f32 %f426, %f422, %f425, %f424;
mul.f32 %f427, %f426, 0f3FB8AA3B;
add.f32 %f428, %f422, 0f4B40007F;
mov.b32 %r217, %f428;
shl.b32 %r218, %r217, 23;
mov.b32 %f429, %r218;
ex2.approx.ftz.f32 %f430, %f427;
mul.f32 %f95, %f430, %f429;
setp.eq.f32 %p49, %f95, 0f7F800000;
mov.f32 %f1005, 0f7F800000;
@%p49 bra $L__BB0_41;
fma.rn.f32 %f1005, %f95, %f94, %f95;
$L__BB0_41:
mov.f32 %f965, 0f3E666666;
cvt.rzi.f32.f32 %f964, %f965;
add.f32 %f963, %f964, %f964;
mov.f32 %f962, 0f3EE66666;
sub.f32 %f961, %f962, %f963;
abs.f32 %f960, %f961;
setp.lt.f32 %p50, %f1002, 0f00000000;
setp.eq.f32 %p51, %f960, 0f3F800000;
and.pred %p1, %p50, %p51;
setp.eq.f32 %p52, %f1002, 0f00000000;
@%p52 bra $L__BB0_45;
bra.uni $L__BB0_42;
$L__BB0_45:
add.f32 %f435, %f1002, %f1002;
selp.f32 %f1007, %f435, 0f00000000, %p51;
bra.uni $L__BB0_46;
$L__BB0_42:
mov.b32 %r219, %f1005;
xor.b32 %r220, %r219, -2147483648;
mov.b32 %f431, %r220;
selp.f32 %f1007, %f431, %f1005, %p1;
setp.geu.f32 %p53, %f1002, 0f00000000;
@%p53 bra $L__BB0_46;
mov.f32 %f432, 0f3EE66666;
cvt.rzi.f32.f32 %f433, %f432;
setp.eq.f32 %p54, %f433, 0f3EE66666;
@%p54 bra $L__BB0_46;
mov.f32 %f1007, 0f7FFFFFFF;
$L__BB0_46:
abs.f32 %f966, %f1002;
add.f32 %f436, %f966, 0f3EE66666;
mov.b32 %r221, %f436;
setp.lt.s32 %p56, %r221, 2139095040;
@%p56 bra $L__BB0_51;
abs.f32 %f967, %f1002;
setp.gtu.f32 %p57, %f967, 0f7F800000;
@%p57 bra $L__BB0_50;
bra.uni $L__BB0_48;
$L__BB0_50:
add.f32 %f1007, %f1002, 0f3EE66666;
bra.uni $L__BB0_51;
$L__BB0_48:
abs.f32 %f968, %f1002;
setp.neu.f32 %p58, %f968, 0f7F800000;
@%p58 bra $L__BB0_51;
selp.f32 %f1007, 0fFF800000, 0f7F800000, %p1;
$L__BB0_51:
setp.eq.f32 %p59, %f1002, 0f3F800000;
selp.f32 %f104, 0f3F800000, %f1007, %p59;
abs.f32 %f105, %f1003;
setp.lt.f32 %p60, %f105, 0f00800000;
mul.f32 %f438, %f105, 0f4B800000;
selp.f32 %f439, %f438, %f105, %p60;
selp.f32 %f440, 0fC3170000, 0fC2FE0000, %p60;
mov.b32 %r222, %f439;
and.b32 %r223, %r222, 8388607;
or.b32 %r224, %r223, 1065353216;
mov.b32 %f441, %r224;
shr.u32 %r225, %r222, 23;
cvt.rn.f32.u32 %f442, %r225;
add.f32 %f443, %f440, %f442;
setp.gt.f32 %p61, %f441, 0f3FB504F3;
mul.f32 %f444, %f441, 0f3F000000;
add.f32 %f445, %f443, 0f3F800000;
selp.f32 %f446, %f445, %f443, %p61;
selp.f32 %f447, %f444, %f441, %p61;
add.f32 %f448, %f447, 0fBF800000;
add.f32 %f449, %f447, 0f3F800000;
rcp.approx.ftz.f32 %f450, %f449;
add.f32 %f451, %f448, %f448;
mul.f32 %f452, %f451, %f450;
mul.f32 %f453, %f452, %f452;
mov.f32 %f454, 0f3C4CAF63;
mov.f32 %f455, 0f3B18F0FE;
fma.rn.f32 %f456, %f455, %f453, %f454;
mov.f32 %f457, 0f3DAAAABD;
fma.rn.f32 %f458, %f456, %f453, %f457;
mul.rn.f32 %f459, %f458, %f453;
mul.rn.f32 %f460, %f459, %f452;
sub.f32 %f461, %f448, %f452;
add.f32 %f462, %f461, %f461;
neg.f32 %f463, %f452;
fma.rn.f32 %f464, %f463, %f448, %f462;
mul.rn.f32 %f465, %f450, %f464;
add.f32 %f466, %f460, %f452;
sub.f32 %f467, %f452, %f466;
add.f32 %f468, %f460, %f467;
add.f32 %f469, %f465, %f468;
add.f32 %f470, %f466, %f469;
sub.f32 %f471, %f466, %f470;
add.f32 %f472, %f469, %f471;
mov.f32 %f473, 0f3F317200;
mul.rn.f32 %f474, %f446, %f473;
mov.f32 %f475, 0f35BFBE8E;
mul.rn.f32 %f476, %f446, %f475;
add.f32 %f477, %f474, %f470;
sub.f32 %f478, %f474, %f477;
add.f32 %f479, %f470, %f478;
add.f32 %f480, %f472, %f479;
add.f32 %f481, %f476, %f480;
add.f32 %f482, %f477, %f481;
sub.f32 %f483, %f477, %f482;
add.f32 %f484, %f481, %f483;
mov.f32 %f485, 0f3EE66666;
mul.rn.f32 %f486, %f485, %f482;
neg.f32 %f487, %f486;
fma.rn.f32 %f488, %f485, %f482, %f487;
fma.rn.f32 %f489, %f485, %f484, %f488;
mov.f32 %f490, 0f00000000;
fma.rn.f32 %f491, %f490, %f482, %f489;
add.rn.f32 %f492, %f486, %f491;
neg.f32 %f493, %f492;
add.rn.f32 %f494, %f486, %f493;
add.rn.f32 %f495, %f494, %f491;
mov.b32 %r226, %f492;
setp.eq.s32 %p62, %r226, 1118925336;
add.s32 %r227, %r226, -1;
mov.b32 %f496, %r227;
add.f32 %f497, %f495, 0f37000000;
selp.f32 %f106, %f497, %f495, %p62;
selp.f32 %f498, %f496, %f492, %p62;
mov.f32 %f499, 0f3FB8AA3B;
mul.rn.f32 %f500, %f498, %f499;
cvt.rzi.f32.f32 %f501, %f500;
abs.f32 %f502, %f501;
setp.gt.f32 %p63, %f502, 0f42FC0000;
mov.b32 %r228, %f501;
and.b32 %r229, %r228, -2147483648;
or.b32 %r230, %r229, 1123811328;
mov.b32 %f503, %r230;
selp.f32 %f504, %f503, %f501, %p63;
mov.f32 %f505, 0fBF317218;
fma.rn.f32 %f506, %f504, %f505, %f498;
mov.f32 %f507, 0f3102E308;
fma.rn.f32 %f508, %f504, %f507, %f506;
mul.f32 %f509, %f508, 0f3FB8AA3B;
add.f32 %f510, %f504, 0f4B40007F;
mov.b32 %r231, %f510;
shl.b32 %r232, %r231, 23;
mov.b32 %f511, %r232;
ex2.approx.ftz.f32 %f512, %f509;
mul.f32 %f107, %f512, %f511;
setp.eq.f32 %p64, %f107, 0f7F800000;
mov.f32 %f1008, 0f7F800000;
@%p64 bra $L__BB0_53;
fma.rn.f32 %f1008, %f107, %f106, %f107;
$L__BB0_53:
setp.lt.f32 %p65, %f1003, 0f00000000;
and.pred %p2, %p65, %p51;
setp.eq.f32 %p67, %f1003, 0f00000000;
@%p67 bra $L__BB0_57;
bra.uni $L__BB0_54;
$L__BB0_57:
add.f32 %f517, %f1003, %f1003;
selp.f32 %f1010, %f517, 0f00000000, %p51;
bra.uni $L__BB0_58;
$L__BB0_54:
mov.b32 %r233, %f1008;
xor.b32 %r234, %r233, -2147483648;
mov.b32 %f513, %r234;
selp.f32 %f1010, %f513, %f1008, %p2;
setp.geu.f32 %p68, %f1003, 0f00000000;
@%p68 bra $L__BB0_58;
mov.f32 %f514, 0f3EE66666;
cvt.rzi.f32.f32 %f515, %f514;
setp.eq.f32 %p69, %f515, 0f3EE66666;
@%p69 bra $L__BB0_58;
mov.f32 %f1010, 0f7FFFFFFF;
$L__BB0_58:
abs.f32 %f969, %f1003;
add.f32 %f518, %f969, 0f3EE66666;
mov.b32 %r235, %f518;
setp.lt.s32 %p71, %r235, 2139095040;
@%p71 bra $L__BB0_63;
abs.f32 %f970, %f1003;
setp.gtu.f32 %p72, %f970, 0f7F800000;
@%p72 bra $L__BB0_62;
bra.uni $L__BB0_60;
$L__BB0_62:
add.f32 %f1010, %f1003, 0f3EE66666;
bra.uni $L__BB0_63;
$L__BB0_60:
abs.f32 %f971, %f1003;
setp.neu.f32 %p73, %f971, 0f7F800000;
@%p73 bra $L__BB0_63;
selp.f32 %f1010, 0fFF800000, 0f7F800000, %p2;
$L__BB0_63:
setp.eq.f32 %p74, %f1003, 0f3F800000;
selp.f32 %f116, 0f3F800000, %f1010, %p74;
abs.f32 %f117, %f1004;
setp.lt.f32 %p75, %f117, 0f00800000;
mul.f32 %f520, %f117, 0f4B800000;
selp.f32 %f521, %f520, %f117, %p75;
selp.f32 %f522, 0fC3170000, 0fC2FE0000, %p75;
mov.b32 %r236, %f521;
and.b32 %r237, %r236, 8388607;
or.b32 %r238, %r237, 1065353216;
mov.b32 %f523, %r238;
shr.u32 %r239, %r236, 23;
cvt.rn.f32.u32 %f524, %r239;
add.f32 %f525, %f522, %f524;
setp.gt.f32 %p76, %f523, 0f3FB504F3;
mul.f32 %f526, %f523, 0f3F000000;
add.f32 %f527, %f525, 0f3F800000;
selp.f32 %f528, %f527, %f525, %p76;
selp.f32 %f529, %f526, %f523, %p76;
add.f32 %f530, %f529, 0fBF800000;
add.f32 %f531, %f529, 0f3F800000;
rcp.approx.ftz.f32 %f532, %f531;
add.f32 %f533, %f530, %f530;
mul.f32 %f534, %f533, %f532;
mul.f32 %f535, %f534, %f534;
mov.f32 %f536, 0f3C4CAF63;
mov.f32 %f537, 0f3B18F0FE;
fma.rn.f32 %f538, %f537, %f535, %f536;
mov.f32 %f539, 0f3DAAAABD;
fma.rn.f32 %f540, %f538, %f535, %f539;
mul.rn.f32 %f541, %f540, %f535;
mul.rn.f32 %f542, %f541, %f534;
sub.f32 %f543, %f530, %f534;
add.f32 %f544, %f543, %f543;
neg.f32 %f545, %f534;
fma.rn.f32 %f546, %f545, %f530, %f544;
mul.rn.f32 %f547, %f532, %f546;
add.f32 %f548, %f542, %f534;
sub.f32 %f549, %f534, %f548;
add.f32 %f550, %f542, %f549;
add.f32 %f551, %f547, %f550;
add.f32 %f552, %f548, %f551;
sub.f32 %f553, %f548, %f552;
add.f32 %f554, %f551, %f553;
mov.f32 %f555, 0f3F317200;
mul.rn.f32 %f556, %f528, %f555;
mov.f32 %f557, 0f35BFBE8E;
mul.rn.f32 %f558, %f528, %f557;
add.f32 %f559, %f556, %f552;
sub.f32 %f560, %f556, %f559;
add.f32 %f561, %f552, %f560;
add.f32 %f562, %f554, %f561;
add.f32 %f563, %f558, %f562;
add.f32 %f564, %f559, %f563;
sub.f32 %f565, %f559, %f564;
add.f32 %f566, %f563, %f565;
mov.f32 %f567, 0f3EE66666;
mul.rn.f32 %f568, %f567, %f564;
neg.f32 %f569, %f568;
fma.rn.f32 %f570, %f567, %f564, %f569;
fma.rn.f32 %f571, %f567, %f566, %f570;
mov.f32 %f572, 0f00000000;
fma.rn.f32 %f573, %f572, %f564, %f571;
add.rn.f32 %f574, %f568, %f573;
neg.f32 %f575, %f574;
add.rn.f32 %f576, %f568, %f575;
add.rn.f32 %f577, %f576, %f573;
mov.b32 %r240, %f574;
setp.eq.s32 %p77, %r240, 1118925336;
add.s32 %r241, %r240, -1;
mov.b32 %f578, %r241;
add.f32 %f579, %f577, 0f37000000;
selp.f32 %f118, %f579, %f577, %p77;
selp.f32 %f580, %f578, %f574, %p77;
mov.f32 %f581, 0f3FB8AA3B;
mul.rn.f32 %f582, %f580, %f581;
cvt.rzi.f32.f32 %f583, %f582;
abs.f32 %f584, %f583;
setp.gt.f32 %p78, %f584, 0f42FC0000;
mov.b32 %r242, %f583;
and.b32 %r243, %r242, -2147483648;
or.b32 %r244, %r243, 1123811328;
mov.b32 %f585, %r244;
selp.f32 %f586, %f585, %f583, %p78;
mov.f32 %f587, 0fBF317218;
fma.rn.f32 %f588, %f586, %f587, %f580;
mov.f32 %f589, 0f3102E308;
fma.rn.f32 %f590, %f586, %f589, %f588;
mul.f32 %f591, %f590, 0f3FB8AA3B;
add.f32 %f592, %f586, 0f4B40007F;
mov.b32 %r245, %f592;
shl.b32 %r246, %r245, 23;
mov.b32 %f593, %r246;
ex2.approx.ftz.f32 %f594, %f591;
mul.f32 %f119, %f594, %f593;
setp.eq.f32 %p79, %f119, 0f7F800000;
mov.f32 %f1011, 0f7F800000;
@%p79 bra $L__BB0_65;
fma.rn.f32 %f1011, %f119, %f118, %f119;
$L__BB0_65:
setp.lt.f32 %p80, %f1004, 0f00000000;
and.pred %p3, %p80, %p51;
setp.eq.f32 %p82, %f1004, 0f00000000;
@%p82 bra $L__BB0_69;
bra.uni $L__BB0_66;
$L__BB0_69:
add.f32 %f599, %f1004, %f1004;
selp.f32 %f1013, %f599, 0f00000000, %p51;
bra.uni $L__BB0_70;
$L__BB0_66:
mov.b32 %r247, %f1011;
xor.b32 %r248, %r247, -2147483648;
mov.b32 %f595, %r248;
selp.f32 %f1013, %f595, %f1011, %p3;
setp.geu.f32 %p83, %f1004, 0f00000000;
@%p83 bra $L__BB0_70;
mov.f32 %f596, 0f3EE66666;
cvt.rzi.f32.f32 %f597, %f596;
setp.eq.f32 %p84, %f597, 0f3EE66666;
@%p84 bra $L__BB0_70;
mov.f32 %f1013, 0f7FFFFFFF;
$L__BB0_70:
abs.f32 %f972, %f1004;
add.f32 %f600, %f972, 0f3EE66666;
mov.b32 %r249, %f600;
setp.lt.s32 %p86, %r249, 2139095040;
@%p86 bra $L__BB0_75;
abs.f32 %f973, %f1004;
setp.gtu.f32 %p87, %f973, 0f7F800000;
@%p87 bra $L__BB0_74;
bra.uni $L__BB0_72;
$L__BB0_74:
add.f32 %f1013, %f1004, 0f3EE66666;
bra.uni $L__BB0_75;
$L__BB0_72:
abs.f32 %f974, %f1004;
setp.neu.f32 %p88, %f974, 0f7F800000;
@%p88 bra $L__BB0_75;
selp.f32 %f1013, 0fFF800000, 0f7F800000, %p3;
$L__BB0_75:
setp.eq.f32 %p89, %f1004, 0f3F800000;
mov.f32 %f602, 0f3F800000;
selp.f32 %f603, 0f3F800000, %f1013, %p89;
min.f32 %f604, %f104, %f602;
mov.f32 %f605, 0f00000000;
max.f32 %f128, %f605, %f604;
min.f32 %f606, %f116, %f602;
max.f32 %f129, %f605, %f606;
min.f32 %f607, %f603, %f602;
max.f32 %f130, %f605, %f607;
mov.f32 %f611, 0f3ED55555;
abs.f32 %f132, %f128;
setp.lt.f32 %p90, %f132, 0f00800000;
mul.f32 %f613, %f132, 0f4B800000;
selp.f32 %f614, %f613, %f132, %p90;
selp.f32 %f615, 0fC3170000, 0fC2FE0000, %p90;
mov.b32 %r252, %f614;
and.b32 %r253, %r252, 8388607;
or.b32 %r254, %r253, 1065353216;
mov.b32 %f616, %r254;
shr.u32 %r255, %r252, 23;
cvt.rn.f32.u32 %f617, %r255;
add.f32 %f618, %f615, %f617;
setp.gt.f32 %p91, %f616, 0f3FB504F3;
mul.f32 %f619, %f616, 0f3F000000;
add.f32 %f620, %f618, 0f3F800000;
selp.f32 %f621, %f620, %f618, %p91;
selp.f32 %f622, %f619, %f616, %p91;
add.f32 %f623, %f622, 0fBF800000;
add.f32 %f624, %f622, 0f3F800000;
rcp.approx.ftz.f32 %f625, %f624;
add.f32 %f626, %f623, %f623;
mul.f32 %f627, %f626, %f625;
mul.f32 %f628, %f627, %f627;
mov.f32 %f629, 0f3C4CAF63;
mov.f32 %f630, 0f3B18F0FE;
fma.rn.f32 %f631, %f630, %f628, %f629;
mov.f32 %f632, 0f3DAAAABD;
fma.rn.f32 %f633, %f631, %f628, %f632;
mul.rn.f32 %f634, %f633, %f628;
mul.rn.f32 %f635, %f634, %f627;
sub.f32 %f636, %f623, %f627;
add.f32 %f637, %f636, %f636;
neg.f32 %f638, %f627;
fma.rn.f32 %f639, %f638, %f623, %f637;
mul.rn.f32 %f640, %f625, %f639;
add.f32 %f641, %f635, %f627;
sub.f32 %f642, %f627, %f641;
add.f32 %f643, %f635, %f642;
add.f32 %f644, %f640, %f643;
add.f32 %f645, %f641, %f644;
sub.f32 %f646, %f641, %f645;
add.f32 %f647, %f644, %f646;
mov.f32 %f648, 0f3F317200;
mul.rn.f32 %f649, %f621, %f648;
mov.f32 %f650, 0f35BFBE8E;
mul.rn.f32 %f651, %f621, %f650;
add.f32 %f652, %f649, %f645;
sub.f32 %f653, %f649, %f652;
add.f32 %f654, %f645, %f653;
add.f32 %f655, %f647, %f654;
add.f32 %f656, %f651, %f655;
add.f32 %f657, %f652, %f656;
sub.f32 %f658, %f652, %f657;
add.f32 %f659, %f656, %f658;
mul.rn.f32 %f660, %f611, %f657;
neg.f32 %f661, %f660;
fma.rn.f32 %f662, %f611, %f657, %f661;
fma.rn.f32 %f663, %f611, %f659, %f662;
fma.rn.f32 %f664, %f605, %f657, %f663;
add.rn.f32 %f665, %f660, %f664;
neg.f32 %f666, %f665;
add.rn.f32 %f667, %f660, %f666;
add.rn.f32 %f668, %f667, %f664;
mov.b32 %r256, %f665;
setp.eq.s32 %p92, %r256, 1118925336;
add.s32 %r257, %r256, -1;
mov.b32 %f669, %r257;
add.f32 %f670, %f668, 0f37000000;
selp.f32 %f133, %f670, %f668, %p92;
selp.f32 %f671, %f669, %f665, %p92;
mov.f32 %f672, 0f3FB8AA3B;
mul.rn.f32 %f673, %f671, %f672;
cvt.rzi.f32.f32 %f674, %f673;
abs.f32 %f675, %f674;
setp.gt.f32 %p93, %f675, 0f42FC0000;
mov.b32 %r258, %f674;
and.b32 %r259, %r258, -2147483648;
or.b32 %r260, %r259, 1123811328;
mov.b32 %f676, %r260;
selp.f32 %f677, %f676, %f674, %p93;
mov.f32 %f678, 0fBF317218;
fma.rn.f32 %f679, %f677, %f678, %f671;
mov.f32 %f680, 0f3102E308;
fma.rn.f32 %f681, %f677, %f680, %f679;
mul.f32 %f682, %f681, 0f3FB8AA3B;
add.f32 %f683, %f677, 0f4B40007F;
mov.b32 %r261, %f683;
shl.b32 %r262, %r261, 23;
mov.b32 %f684, %r262;
ex2.approx.ftz.f32 %f685, %f682;
mul.f32 %f134, %f685, %f684;
setp.eq.f32 %p94, %f134, 0f7F800000;
mov.f32 %f1014, 0f7F800000;
@%p94 bra $L__BB0_77;
fma.rn.f32 %f1014, %f134, %f133, %f134;
$L__BB0_77:
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, %f128, 0f00000000;
setp.eq.f32 %p96, %f950, 0f3F800000;
and.pred %p4, %p95, %p96;
setp.eq.f32 %p97, %f128, 0f00000000;
@%p97 bra $L__BB0_81;
bra.uni $L__BB0_78;
$L__BB0_81:
add.f32 %f690, %f128, %f128;
selp.f32 %f1016, %f690, 0f00000000, %p96;
bra.uni $L__BB0_82;
$L__BB0_78:
mov.b32 %r263, %f1014;
xor.b32 %r264, %r263, -2147483648;
mov.b32 %f686, %r264;
selp.f32 %f1016, %f686, %f1014, %p4;
setp.geu.f32 %p98, %f128, 0f00000000;
@%p98 bra $L__BB0_82;
mov.f32 %f687, 0f3ED55555;
cvt.rzi.f32.f32 %f688, %f687;
setp.eq.f32 %p99, %f688, 0f3ED55555;
@%p99 bra $L__BB0_82;
mov.f32 %f1016, 0f7FFFFFFF;
$L__BB0_82:
add.f32 %f691, %f132, 0f3ED55555;
mov.b32 %r265, %f691;
setp.lt.s32 %p101, %r265, 2139095040;
@%p101 bra $L__BB0_87;
setp.gtu.f32 %p102, %f132, 0f7F800000;
@%p102 bra $L__BB0_86;
bra.uni $L__BB0_84;
$L__BB0_86:
add.f32 %f1016, %f128, 0f3ED55555;
bra.uni $L__BB0_87;
$L__BB0_84:
setp.neu.f32 %p103, %f132, 0f7F800000;
@%p103 bra $L__BB0_87;
selp.f32 %f1016, 0fFF800000, 0f7F800000, %p4;
$L__BB0_87:
abs.f32 %f143, %f129;
setp.lt.f32 %p104, %f143, 0f00800000;
mul.f32 %f693, %f143, 0f4B800000;
selp.f32 %f694, %f693, %f143, %p104;
selp.f32 %f695, 0fC3170000, 0fC2FE0000, %p104;
mov.b32 %r266, %f694;
and.b32 %r267, %r266, 8388607;
or.b32 %r268, %r267, 1065353216;
mov.b32 %f696, %r268;
shr.u32 %r269, %r266, 23;
cvt.rn.f32.u32 %f697, %r269;
add.f32 %f698, %f695, %f697;
setp.gt.f32 %p105, %f696, 0f3FB504F3;
mul.f32 %f699, %f696, 0f3F000000;
add.f32 %f700, %f698, 0f3F800000;
selp.f32 %f701, %f700, %f698, %p105;
selp.f32 %f702, %f699, %f696, %p105;
add.f32 %f703, %f702, 0fBF800000;
add.f32 %f704, %f702, 0f3F800000;
rcp.approx.ftz.f32 %f705, %f704;
add.f32 %f706, %f703, %f703;
mul.f32 %f707, %f706, %f705;
mul.f32 %f708, %f707, %f707;
mov.f32 %f709, 0f3C4CAF63;
mov.f32 %f710, 0f3B18F0FE;
fma.rn.f32 %f711, %f710, %f708, %f709;
mov.f32 %f712, 0f3DAAAABD;
fma.rn.f32 %f713, %f711, %f708, %f712;
mul.rn.f32 %f714, %f713, %f708;
mul.rn.f32 %f715, %f714, %f707;
sub.f32 %f716, %f703, %f707;
add.f32 %f717, %f716, %f716;
neg.f32 %f718, %f707;
fma.rn.f32 %f719, %f718, %f703, %f717;
mul.rn.f32 %f720, %f705, %f719;
add.f32 %f721, %f715, %f707;
sub.f32 %f722, %f707, %f721;
add.f32 %f723, %f715, %f722;
add.f32 %f724, %f720, %f723;
add.f32 %f725, %f721, %f724;
sub.f32 %f726, %f721, %f725;
add.f32 %f727, %f724, %f726;
mov.f32 %f728, 0f3F317200;
mul.rn.f32 %f729, %f701, %f728;
mov.f32 %f730, 0f35BFBE8E;
mul.rn.f32 %f731, %f701, %f730;
add.f32 %f732, %f729, %f725;
sub.f32 %f733, %f729, %f732;
add.f32 %f734, %f725, %f733;
add.f32 %f735, %f727, %f734;
add.f32 %f736, %f731, %f735;
add.f32 %f737, %f732, %f736;
sub.f32 %f738, %f732, %f737;
add.f32 %f739, %f736, %f738;
mov.f32 %f740, 0f3ED55555;
mul.rn.f32 %f741, %f740, %f737;
neg.f32 %f742, %f741;
fma.rn.f32 %f743, %f740, %f737, %f742;
fma.rn.f32 %f744, %f740, %f739, %f743;
mov.f32 %f745, 0f00000000;
fma.rn.f32 %f746, %f745, %f737, %f744;
add.rn.f32 %f747, %f741, %f746;
neg.f32 %f748, %f747;
add.rn.f32 %f749, %f741, %f748;
add.rn.f32 %f750, %f749, %f746;
mov.b32 %r270, %f747;
setp.eq.s32 %p106, %r270, 1118925336;
add.s32 %r271, %r270, -1;
mov.b32 %f751, %r271;
add.f32 %f752, %f750, 0f37000000;
selp.f32 %f144, %f752, %f750, %p106;
selp.f32 %f753, %f751, %f747, %p106;
mov.f32 %f754, 0f3FB8AA3B;
mul.rn.f32 %f755, %f753, %f754;
cvt.rzi.f32.f32 %f756, %f755;
abs.f32 %f757, %f756;
setp.gt.f32 %p107, %f757, 0f42FC0000;
mov.b32 %r272, %f756;
and.b32 %r273, %r272, -2147483648;
or.b32 %r274, %r273, 1123811328;
mov.b32 %f758, %r274;
selp.f32 %f759, %f758, %f756, %p107;
mov.f32 %f760, 0fBF317218;
fma.rn.f32 %f761, %f759, %f760, %f753;
mov.f32 %f762, 0f3102E308;
fma.rn.f32 %f763, %f759, %f762, %f761;
mul.f32 %f764, %f763, 0f3FB8AA3B;
add.f32 %f765, %f759, 0f4B40007F;
mov.b32 %r275, %f765;
shl.b32 %r276, %r275, 23;
mov.b32 %f766, %r276;
ex2.approx.ftz.f32 %f767, %f764;
mul.f32 %f145, %f767, %f766;
setp.eq.f32 %p108, %f145, 0f7F800000;
mov.f32 %f1017, 0f7F800000;
@%p108 bra $L__BB0_89;
fma.rn.f32 %f1017, %f145, %f144, %f145;
$L__BB0_89:
setp.lt.f32 %p109, %f129, 0f00000000;
and.pred %p5, %p109, %p96;
setp.eq.f32 %p111, %f129, 0f00000000;
@%p111 bra $L__BB0_93;
bra.uni $L__BB0_90;
$L__BB0_93:
add.f32 %f772, %f129, %f129;
selp.f32 %f1019, %f772, 0f00000000, %p96;
bra.uni $L__BB0_94;
$L__BB0_90:
mov.b32 %r277, %f1017;
xor.b32 %r278, %r277, -2147483648;
mov.b32 %f768, %r278;
selp.f32 %f1019, %f768, %f1017, %p5;
setp.geu.f32 %p112, %f129, 0f00000000;
@%p112 bra $L__BB0_94;
mov.f32 %f769, 0f3ED55555;
cvt.rzi.f32.f32 %f770, %f769;
setp.eq.f32 %p113, %f770, 0f3ED55555;
@%p113 bra $L__BB0_94;
mov.f32 %f1019, 0f7FFFFFFF;
$L__BB0_94:
add.f32 %f773, %f143, 0f3ED55555;
mov.b32 %r279, %f773;
setp.lt.s32 %p115, %r279, 2139095040;
@%p115 bra $L__BB0_99;
setp.gtu.f32 %p116, %f143, 0f7F800000;
@%p116 bra $L__BB0_98;
bra.uni $L__BB0_96;
$L__BB0_98:
add.f32 %f1019, %f129, 0f3ED55555;
bra.uni $L__BB0_99;
$L__BB0_96:
setp.neu.f32 %p117, %f143, 0f7F800000;
@%p117 bra $L__BB0_99;
selp.f32 %f1019, 0fFF800000, 0f7F800000, %p5;
$L__BB0_99:
abs.f32 %f154, %f130;
setp.lt.f32 %p118, %f154, 0f00800000;
mul.f32 %f775, %f154, 0f4B800000;
selp.f32 %f776, %f775, %f154, %p118;
selp.f32 %f777, 0fC3170000, 0fC2FE0000, %p118;
mov.b32 %r280, %f776;
and.b32 %r281, %r280, 8388607;
or.b32 %r282, %r281, 1065353216;
mov.b32 %f778, %r282;
shr.u32 %r283, %r280, 23;
cvt.rn.f32.u32 %f779, %r283;
add.f32 %f780, %f777, %f779;
setp.gt.f32 %p119, %f778, 0f3FB504F3;
mul.f32 %f781, %f778, 0f3F000000;
add.f32 %f782, %f780, 0f3F800000;
selp.f32 %f783, %f782, %f780, %p119;
selp.f32 %f784, %f781, %f778, %p119;
add.f32 %f785, %f784, 0fBF800000;
add.f32 %f786, %f784, 0f3F800000;
rcp.approx.ftz.f32 %f787, %f786;
add.f32 %f788, %f785, %f785;
mul.f32 %f789, %f788, %f787;
mul.f32 %f790, %f789, %f789;
mov.f32 %f791, 0f3C4CAF63;
mov.f32 %f792, 0f3B18F0FE;
fma.rn.f32 %f793, %f792, %f790, %f791;
mov.f32 %f794, 0f3DAAAABD;
fma.rn.f32 %f795, %f793, %f790, %f794;
mul.rn.f32 %f796, %f795, %f790;
mul.rn.f32 %f797, %f796, %f789;
sub.f32 %f798, %f785, %f789;
add.f32 %f799, %f798, %f798;
neg.f32 %f800, %f789;
fma.rn.f32 %f801, %f800, %f785, %f799;
mul.rn.f32 %f802, %f787, %f801;
add.f32 %f803, %f797, %f789;
sub.f32 %f804, %f789, %f803;
add.f32 %f805, %f797, %f804;
add.f32 %f806, %f802, %f805;
add.f32 %f807, %f803, %f806;
sub.f32 %f808, %f803, %f807;
add.f32 %f809, %f806, %f808;
mov.f32 %f810, 0f3F317200;
mul.rn.f32 %f811, %f783, %f810;
mov.f32 %f812, 0f35BFBE8E;
mul.rn.f32 %f813, %f783, %f812;
add.f32 %f814, %f811, %f807;
sub.f32 %f815, %f811, %f814;
add.f32 %f816, %f807, %f815;
add.f32 %f817, %f809, %f816;
add.f32 %f818, %f813, %f817;
add.f32 %f819, %f814, %f818;
sub.f32 %f820, %f814, %f819;
add.f32 %f821, %f818, %f820;
mov.f32 %f822, 0f3ED55555;
mul.rn.f32 %f823, %f822, %f819;
neg.f32 %f824, %f823;
fma.rn.f32 %f825, %f822, %f819, %f824;
fma.rn.f32 %f826, %f822, %f821, %f825;
mov.f32 %f827, 0f00000000;
fma.rn.f32 %f828, %f827, %f819, %f826;
add.rn.f32 %f829, %f823, %f828;
neg.f32 %f830, %f829;
add.rn.f32 %f831, %f823, %f830;
add.rn.f32 %f832, %f831, %f828;
mov.b32 %r284, %f829;
setp.eq.s32 %p120, %r284, 1118925336;
add.s32 %r285, %r284, -1;
mov.b32 %f833, %r285;
add.f32 %f834, %f832, 0f37000000;
selp.f32 %f155, %f834, %f832, %p120;
selp.f32 %f835, %f833, %f829, %p120;
mov.f32 %f836, 0f3FB8AA3B;
mul.rn.f32 %f837, %f835, %f836;
cvt.rzi.f32.f32 %f838, %f837;
abs.f32 %f839, %f838;
setp.gt.f32 %p121, %f839, 0f42FC0000;
mov.b32 %r286, %f838;
and.b32 %r287, %r286, -2147483648;
or.b32 %r288, %r287, 1123811328;
mov.b32 %f840, %r288;
selp.f32 %f841, %f840, %f838, %p121;
mov.f32 %f842, 0fBF317218;
fma.rn.f32 %f843, %f841, %f842, %f835;
mov.f32 %f844, 0f3102E308;
fma.rn.f32 %f845, %f841, %f844, %f843;
mul.f32 %f846, %f845, 0f3FB8AA3B;
add.f32 %f847, %f841, 0f4B40007F;
mov.b32 %r289, %f847;
shl.b32 %r290, %r289, 23;
mov.b32 %f848, %r290;
ex2.approx.ftz.f32 %f849, %f846;
mul.f32 %f156, %f849, %f848;
setp.eq.f32 %p122, %f156, 0f7F800000;
mov.f32 %f1020, 0f7F800000;
@%p122 bra $L__BB0_101;
fma.rn.f32 %f1020, %f156, %f155, %f156;
$L__BB0_101:
setp.lt.f32 %p123, %f130, 0f00000000;
and.pred %p6, %p123, %p96;
setp.eq.f32 %p125, %f130, 0f00000000;
@%p125 bra $L__BB0_105;
bra.uni $L__BB0_102;
$L__BB0_105:
add.f32 %f854, %f130, %f130;
selp.f32 %f1022, %f854, 0f00000000, %p96;
bra.uni $L__BB0_106;
$L__BB0_102:
mov.b32 %r291, %f1020;
xor.b32 %r292, %r291, -2147483648;
mov.b32 %f850, %r292;
selp.f32 %f1022, %f850, %f1020, %p6;
setp.geu.f32 %p126, %f130, 0f00000000;
@%p126 bra $L__BB0_106;
mov.f32 %f851, 0f3ED55555;
cvt.rzi.f32.f32 %f852, %f851;
setp.eq.f32 %p127, %f852, 0f3ED55555;
@%p127 bra $L__BB0_106;
mov.f32 %f1022, 0f7FFFFFFF;
$L__BB0_106:
add.f32 %f855, %f154, 0f3ED55555;
mov.b32 %r293, %f855;
setp.lt.s32 %p129, %r293, 2139095040;
@%p129 bra $L__BB0_111;
setp.gtu.f32 %p130, %f154, 0f7F800000;
@%p130 bra $L__BB0_110;
bra.uni $L__BB0_108;
$L__BB0_110:
add.f32 %f1022, %f130, 0f3ED55555;
bra.uni $L__BB0_111;
$L__BB0_108:
setp.neu.f32 %p131, %f154, 0f7F800000;
@%p131 bra $L__BB0_111;
selp.f32 %f1022, 0fFF800000, 0f7F800000, %p6;
$L__BB0_111:
ld.const.u64 %rd96, [params+144];
cvta.to.global.u64 %rd95, %rd96;
ld.const.u32 %r327, [params+136];
mad.lo.s32 %r326, %r327, %r5, %r4;
cvt.u64.u32 %rd94, %r326;
fma.rn.f32 %f856, %f1016, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p132, %f128, 0f3F800000;
mov.f32 %f857, 0f3F800000;
selp.f32 %f858, 0f3F7FFFFF, %f856, %p132;
mul.f32 %f859, %f128, 0f414EB852;
setp.lt.f32 %p133, %f128, 0f3B4D2E1C;
selp.f32 %f860, %f859, %f858, %p133;
fma.rn.f32 %f861, %f1019, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p134, %f129, 0f3F800000;
selp.f32 %f862, 0f3F7FFFFF, %f861, %p134;
mul.f32 %f863, %f129, 0f414EB852;
setp.lt.f32 %p135, %f129, 0f3B4D2E1C;
selp.f32 %f864, %f863, %f862, %p135;
fma.rn.f32 %f865, %f1022, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p136, %f130, 0f3F800000;
selp.f32 %f866, 0f3F7FFFFF, %f865, %p136;
mul.f32 %f867, %f130, 0f414EB852;
setp.lt.f32 %p137, %f130, 0f3B4D2E1C;
selp.f32 %f868, %f867, %f866, %p137;
min.f32 %f869, %f860, %f857;
mov.f32 %f870, 0f00000000;
max.f32 %f871, %f870, %f869;
mul.f32 %f872, %f871, 0f43800000;
cvt.rzi.u32.f32 %r294, %f872;
min.u32 %r295, %r294, 255;
min.f32 %f873, %f864, %f857;
max.f32 %f874, %f870, %f873;
mul.f32 %f875, %f874, 0f43800000;
cvt.rzi.u32.f32 %r296, %f875;
min.u32 %r297, %r296, 255;
min.f32 %f876, %f868, %f857;
max.f32 %f877, %f870, %f876;
mul.f32 %f878, %f877, 0f43800000;
cvt.rzi.u32.f32 %r298, %f878;
min.u32 %r299, %r298, 255;
shl.b64 %rd60, %rd94, 2;
add.s64 %rd61, %rd95, %rd60;
cvt.u16.u32 %rs23, %r299;
cvt.u16.u32 %rs24, %r297;
cvt.u16.u32 %rs25, %r295;
mov.u16 %rs26, 255;
st.global.v4.u8 [%rd61], {%rs25, %rs24, %rs23, %rs26};
$L__BB0_112:
ld.const.u32 %r328, [params+104];
and.b32 %r300, %r328, 4;
setp.eq.s32 %p138, %r300, 0;
@%p138 bra $L__BB0_114;
ld.const.u64 %rd62, [params+224];
cvta.to.global.u64 %rd63, %rd62;
ld.const.u32 %r301, [params+216];
mad.lo.s32 %r302, %r301, %r5, %r4;
mov.f32 %f882, 0f3F800000;
mul.wide.u32 %rd64, %r302, 8;
add.s64 %rd65, %rd63, %rd64;
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f882;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f1004;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1003;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1002;}
// end inline asm
st.global.v4.u16 [%rd65], {%rs27, %rs28, %rs29, %rs30};
$L__BB0_114:
ld.const.u32 %r329, [params+104];
and.b32 %r49, %r329, 16;
setp.eq.s32 %p139, %r49, 0;
@%p139 bra $L__BB0_116;
ld.const.u64 %rd66, [params+240];
cvta.to.global.u64 %rd67, %rd66;
ld.const.u32 %r303, [params+232];
mad.lo.s32 %r304, %r303, %r5, %r4;
mov.f32 %f886, 0f3F800000;
mul.wide.u32 %rd68, %r304, 8;
add.s64 %rd69, %rd67, %rd68;
// begin inline asm
{ cvt.rn.f16.f32 %rs34, %f886;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs33, %f79;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs32, %f78;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs31, %f77;}
// end inline asm
st.global.v4.u16 [%rd69], {%rs31, %rs32, %rs33, %rs34};
$L__BB0_116:
ld.const.u32 %r330, [params+104];
and.b32 %r305, %r330, 64;
setp.eq.s32 %p140, %r305, 0;
@%p140 bra $L__BB0_131;
mul.f32 %f887, %f196, 0f3F372474;
fma.rn.f32 %f888, %f195, 0f3E59999A, %f887;
fma.rn.f32 %f165, %f197, 0f3D93A92A, %f888;
ld.const.u64 %rd70, [params+680];
cvta.to.global.u64 %rd71, %rd70;
ld.const.u32 %r306, [params+672];
mad.lo.s32 %r307, %r306, %r5, %r4;
mul.wide.u32 %rd72, %r307, 4;
add.s64 %rd73, %rd71, %rd72;
ld.global.v4.u8 {%rs35, %rs36, %rs37, %rs38}, [%rd73];
cvt.rn.f32.u16 %f889, %rs35;
div.rn.f32 %f890, %f889, 0f437F0000;
fma.rn.f32 %f891, %f890, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f892, %rs36;
div.rn.f32 %f893, %f892, 0f437F0000;
fma.rn.f32 %f894, %f893, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f895, %rs37;
div.rn.f32 %f896, %f895, 0f437F0000;
fma.rn.f32 %f897, %f896, 0f40000000, 0fBF800000;
mul.f32 %f898, %f894, %f894;
fma.rn.f32 %f899, %f891, %f891, %f898;
fma.rn.f32 %f900, %f897, %f897, %f899;
sqrt.rn.f32 %f901, %f900;
rcp.rn.f32 %f902, %f901;
mul.f32 %f166, %f902, %f891;
mul.f32 %f167, %f902, %f894;
mul.f32 %f168, %f902, %f897;
abs.f32 %f903, %f989;
setp.gt.f32 %p141, %f903, 0f00000000;
@%p141 bra $L__BB0_120;
abs.f32 %f904, %f988;
setp.gt.f32 %p142, %f904, 0f00000000;
@%p142 bra $L__BB0_120;
bra.uni $L__BB0_119;
$L__BB0_120:
setp.ne.s32 %p146, %r49, 0;
mov.f32 %f1023, %f989;
mov.f32 %f1024, %f988;
mov.f32 %f1025, %f987;
@%p146 bra $L__BB0_122;
$L__BB0_121:
mul.f32 %f906, %f988, %f988;
fma.rn.f32 %f907, %f989, %f989, %f906;
fma.rn.f32 %f908, %f987, %f987, %f907;
sqrt.rn.f32 %f909, %f908;
rcp.rn.f32 %f910, %f909;
mul.f32 %f911, %f989, %f910;
mul.f32 %f912, %f988, %f910;
mul.f32 %f913, %f987, %f910;
mul.f32 %f914, %f78, 0f3F372474;
fma.rn.f32 %f915, %f77, 0f3E59999A, %f914;
fma.rn.f32 %f916, %f79, 0f3D93A92A, %f915;
mul.f32 %f917, %f916, %f911;
mul.f32 %f918, %f916, %f912;
mul.f32 %f919, %f916, %f913;
fma.rn.f32 %f1023, %f165, %f166, %f917;
fma.rn.f32 %f1024, %f165, %f167, %f918;
fma.rn.f32 %f1025, %f165, %f168, %f919;
$L__BB0_122:
mul.f32 %f920, %f1024, %f1024;
fma.rn.f32 %f921, %f1023, %f1023, %f920;
fma.rn.f32 %f922, %f1025, %f1025, %f921;
sqrt.rn.f32 %f923, %f922;
rcp.rn.f32 %f924, %f923;
mul.f32 %f925, %f1023, %f924;
mul.f32 %f926, %f1024, %f924;
mul.f32 %f927, %f1025, %f924;
mul.f32 %f928, %f976, %f926;
fma.rn.f32 %f929, %f975, %f925, %f928;
fma.rn.f32 %f930, %f977, %f927, %f929;
fma.rn.f32 %f931, %f930, 0f3F000000, 0f3F000000;
mov.f32 %f932, 0f3B808081;
max.f32 %f933, %f931, %f932;
mul.f32 %f934, %f933, 0f437F0000;
cvt.rzi.s32.f32 %r308, %f934;
ld.const.u64 %rd74, [params+208];
cvta.to.global.u64 %rd75, %rd74;
ld.const.u32 %r309, [params+200];
mad.lo.s32 %r310, %r309, %r5, %r4;
fma.rn.f32 %f935, %f925, 0f3F000000, 0f3F000000;
mul.f32 %f936, %f935, 0f437F0000;
cvt.rzi.u32.f32 %r311, %f936;
fma.rn.f32 %f937, %f926, 0f3F000000, 0f3F000000;
mul.f32 %f938, %f937, 0f437F0000;
cvt.rzi.u32.f32 %r312, %f938;
fma.rn.f32 %f939, %f927, 0f3F000000, 0f3F000000;
mul.f32 %f940, %f939, 0f437F0000;
cvt.rzi.u32.f32 %r313, %f940;
mul.wide.u32 %rd76, %r310, 4;
add.s64 %rd77, %rd75, %rd76;
cvt.u16.u32 %rs39, %r308;
cvt.u16.u32 %rs40, %r313;
cvt.u16.u32 %rs41, %r312;
cvt.u16.u32 %rs42, %r311;
st.global.v4.u8 [%rd77], {%rs42, %rs41, %rs40, %rs39};
$L__BB0_131:
ret;
$L__BB0_119:
abs.f32 %f905, %f987;
setp.leu.f32 %p143, %f905, 0f00000000;
setp.ne.s32 %p144, %r49, 0;
or.pred %p145, %p143, %p144;
selp.f32 %f1023, %f166, %f989, %p143;
selp.f32 %f1024, %f167, %f988, %p143;
selp.f32 %f1025, %f168, %f987, %p143;
@%p145 bra $L__BB0_122;
bra.uni $L__BB0_121;
}