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

1812 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<155>;
.reg .b16 %rs<55>;
.reg .f32 %f<1030>;
.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 %f979, 0f00000000;
mov.u16 %rs54, 0;
mov.f32 %f980, %f979;
mov.f32 %f981, %f979;
@%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 %f981, %f194, %f189;
mul.f32 %f980, %f194, %f186;
mul.f32 %f979, %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, %f979, 0f00000000;
setp.eq.f32 %p10, %f980, 0f00000000;
and.pred %p11, %p9, %p10;
setp.eq.f32 %p12, %f981, 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 %p148, %r314, 1;
mov.pred %p149, 0;
xor.pred %p150, %p148, %p149;
not.pred %p151, %p150;
@%p151 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 %p152, %r317, 0;
@%p152 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 %f947, 0f00000000;
mul.wide.u32 %rd84, %r319, 8;
add.s64 %rd85, %rd83, %rd84;
// begin inline asm
{ cvt.rn.f16.f32 %rs46, %f947;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f947;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f947;}
// 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 %p153, %r320, 0;
@%p153 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 %f950, 0f00000000;
mul.wide.u32 %rd88, %r322, 8;
add.s64 %rd89, %rd87, %rd88;
// begin inline asm
{ cvt.rn.f16.f32 %rs50, %f950;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs49, %f950;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs48, %f950;}
// 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 %p154, %r323, 0;
@%p154 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, %f979;
div.rn.f32 %f212, %f206, %f211;
abs.f32 %f213, %f980;
div.rn.f32 %f214, %f208, %f213;
abs.f32 %f215, %f981;
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, %f979, %f221, %f205;
fma.rn.f32 %f14, %f980, %f222, %f207;
fma.rn.f32 %f15, %f981, %f223, %f209;
setp.gt.f32 %p14, %f211, %f215;
neg.f32 %f224, %f980;
selp.f32 %f225, %f224, 0f00000000, %p14;
mov.f32 %f993, 0f00000000;
neg.f32 %f226, %f981;
selp.f32 %f227, %f979, %f226, %p14;
selp.f32 %f228, 0f00000000, %f980, %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 %f992, %f993;
mov.f32 %f991, %f993;
mov.f32 %f990, %f993;
mov.f32 %f989, %f993;
mov.f32 %f988, %f993;
@%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, %f979, %f17;
mul.f32 %f242, %f980, %f16;
sub.f32 %f23, %f242, %f241;
mul.f32 %f243, %f981, %f16;
mul.f32 %f244, %f979, %f18;
sub.f32 %f24, %f244, %f243;
mul.f32 %f245, %f980, %f18;
mul.f32 %f246, %f981, %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 %f963, %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, %f963;
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 %f997, %f259, %f264, %f263;
abs.f32 %f43, %f41;
setp.ltu.f32 %p16, %f43, 0f47CE4780;
mov.u32 %r338, %r341;
mov.f32 %f994, %f997;
@%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 %f994, %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 %f994, %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, %f994, 0f3F800000, %p23;
mul.rn.f32 %f48, %f994, %f994;
mov.f32 %f995, 0fB94D4153;
@%p23 bra $L__BB0_18;
mov.f32 %f269, 0fBAB607ED;
mov.f32 %f270, 0f37CBAC00;
fma.rn.f32 %f995, %f270, %f48, %f269;
$L__BB0_18:
selp.f32 %f271, 0f3C0885E4, 0f3D2AAABB, %p23;
fma.rn.f32 %f272, %f995, %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 %f996, %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 %f996, %f996, %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 %f997, %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 %f997, %f279, %f280, %p32;
$L__BB0_28:
mul.f32 %f57, %f40, %f996;
and.b32 %r45, %r341, 1;
setp.eq.s32 %p33, %r45, 0;
selp.f32 %f58, %f997, 0f3F800000, %p33;
mul.rn.f32 %f59, %f997, %f997;
mov.f32 %f998, 0fB94D4153;
@%p33 bra $L__BB0_30;
mov.f32 %f283, 0fBAB607ED;
mov.f32 %f284, 0f37CBAC00;
fma.rn.f32 %f998, %f284, %f59, %f283;
$L__BB0_30:
selp.f32 %f285, 0f3C0885E4, 0f3D2AAABB, %p33;
fma.rn.f32 %f286, %f998, %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 %f999, %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 %f999, %f999, %f292, %f289;
$L__BB0_32:
mul.f32 %f302, %f40, %f999;
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, %f979, %f39, %f306;
fma.rn.f32 %f297, %f980, %f39, %f307;
fma.rn.f32 %f298, %f981, %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, 4;
mov.u32 %r176, 1065353216;
mov.u32 %r177, 2139095039;
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,%r176,%r176,%r176,%r177,%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;
mov.b32 %f318, %r138;
setp.ltu.f32 %p36, %f318, 0f00000000;
selp.f32 %f319, %f315, 0f00000000, %p36;
selp.f32 %f320, %f316, 0f00000000, %p36;
selp.f32 %f321, %f317, 0f00000000, %p36;
mul.f32 %f322, %f320, 0f3F372474;
fma.rn.f32 %f323, %f319, 0f3E59999A, %f322;
fma.rn.f32 %f324, %f321, 0f3D93A92A, %f323;
fma.rn.f32 %f993, %f296, %f324, %f993;
fma.rn.f32 %f992, %f297, %f324, %f992;
fma.rn.f32 %f991, %f298, %f324, %f991;
mul.f32 %f325, %f980, %f297;
fma.rn.f32 %f326, %f979, %f296, %f325;
fma.rn.f32 %f327, %f981, %f298, %f326;
cvt.sat.f32.f32 %f328, %f327;
fma.rn.f32 %f990, %f328, %f319, %f990;
fma.rn.f32 %f989, %f328, %f320, %f989;
fma.rn.f32 %f988, %f328, %f321, %f988;
add.s32 %r333, %r333, 1;
setp.lt.s32 %p37, %r333, %r3;
@%p37 bra $L__BB0_8;
add.s32 %r331, %r331, 1;
setp.lt.s32 %p38, %r331, %r3;
@%p38 bra $L__BB0_7;
$L__BB0_34:
mov.f32 %f1006, 0f00000000;
mul.lo.s32 %r206, %r3, %r3;
cvt.rn.f32.s32 %f332, %r206;
rcp.rn.f32 %f333, %f332;
mul.f32 %f334, %f333, %f990;
mul.f32 %f335, %f333, %f989;
mul.f32 %f336, %f333, %f988;
fma.rn.f32 %f337, %f333, %f990, %f334;
fma.rn.f32 %f338, %f333, %f989, %f335;
fma.rn.f32 %f339, %f333, %f988, %f336;
ld.const.u64 %rd58, [params+504];
cvt.rn.f32.u32 %f340, %r5;
cvt.rn.f32.u32 %f341, %r4;
tex.2d.v4.f32.f32 {%f342, %f343, %f344, %f345}, [%rd58, {%f341, %f340}];
mul.f32 %f77, %f337, %f342;
mul.f32 %f78, %f338, %f343;
mul.f32 %f79, %f339, %f344;
add.f32 %f80, %f195, %f77;
add.f32 %f81, %f196, %f78;
add.f32 %f82, %f197, %f79;
abs.f32 %f346, %f80;
setp.gtu.f32 %p39, %f346, 0f7F800000;
mov.f32 %f1007, %f1006;
mov.f32 %f1008, %f1006;
@%p39 bra $L__BB0_38;
mov.f32 %f1006, 0f00000000;
abs.f32 %f350, %f81;
setp.gtu.f32 %p40, %f350, 0f7F800000;
@%p40 bra $L__BB0_38;
mov.f32 %f1006, 0f00000000;
abs.f32 %f354, %f82;
setp.gtu.f32 %p41, %f354, 0f7F800000;
mov.f32 %f1007, %f1006;
mov.f32 %f1008, %f1006;
@%p41 bra $L__BB0_38;
mov.f32 %f1006, %f80;
mov.f32 %f1007, %f81;
mov.f32 %f1008, %f82;
$L__BB0_38:
ld.const.u32 %r48, [params+104];
and.b32 %r207, %r48, 1;
setp.eq.b32 %p42, %r207, 1;
mov.pred %p43, 0;
xor.pred %p44, %p42, %p43;
not.pred %p45, %p44;
@%p45 bra $L__BB0_112;
mov.f32 %f359, 0f3EE66666;
abs.f32 %f93, %f1006;
setp.lt.f32 %p46, %f93, 0f00800000;
mul.f32 %f361, %f93, 0f4B800000;
selp.f32 %f362, %f361, %f93, %p46;
selp.f32 %f363, 0fC3170000, 0fC2FE0000, %p46;
mov.b32 %r208, %f362;
and.b32 %r209, %r208, 8388607;
or.b32 %r210, %r209, 1065353216;
mov.b32 %f364, %r210;
shr.u32 %r211, %r208, 23;
cvt.rn.f32.u32 %f365, %r211;
add.f32 %f366, %f363, %f365;
setp.gt.f32 %p47, %f364, 0f3FB504F3;
mul.f32 %f367, %f364, 0f3F000000;
add.f32 %f368, %f366, 0f3F800000;
selp.f32 %f369, %f368, %f366, %p47;
selp.f32 %f370, %f367, %f364, %p47;
add.f32 %f371, %f370, 0fBF800000;
add.f32 %f372, %f370, 0f3F800000;
rcp.approx.ftz.f32 %f373, %f372;
add.f32 %f374, %f371, %f371;
mul.f32 %f375, %f374, %f373;
mul.f32 %f376, %f375, %f375;
mov.f32 %f377, 0f3C4CAF63;
mov.f32 %f378, 0f3B18F0FE;
fma.rn.f32 %f379, %f378, %f376, %f377;
mov.f32 %f380, 0f3DAAAABD;
fma.rn.f32 %f381, %f379, %f376, %f380;
mul.rn.f32 %f382, %f381, %f376;
mul.rn.f32 %f383, %f382, %f375;
sub.f32 %f384, %f371, %f375;
add.f32 %f385, %f384, %f384;
neg.f32 %f386, %f375;
fma.rn.f32 %f387, %f386, %f371, %f385;
mul.rn.f32 %f388, %f373, %f387;
add.f32 %f389, %f383, %f375;
sub.f32 %f390, %f375, %f389;
add.f32 %f391, %f383, %f390;
add.f32 %f392, %f388, %f391;
add.f32 %f393, %f389, %f392;
sub.f32 %f394, %f389, %f393;
add.f32 %f395, %f392, %f394;
mov.f32 %f396, 0f3F317200;
mul.rn.f32 %f397, %f369, %f396;
mov.f32 %f398, 0f35BFBE8E;
mul.rn.f32 %f399, %f369, %f398;
add.f32 %f400, %f397, %f393;
sub.f32 %f401, %f397, %f400;
add.f32 %f402, %f393, %f401;
add.f32 %f403, %f395, %f402;
add.f32 %f404, %f399, %f403;
add.f32 %f405, %f400, %f404;
sub.f32 %f406, %f400, %f405;
add.f32 %f407, %f404, %f406;
mul.rn.f32 %f408, %f359, %f405;
neg.f32 %f409, %f408;
fma.rn.f32 %f410, %f359, %f405, %f409;
fma.rn.f32 %f411, %f359, %f407, %f410;
mov.f32 %f412, 0f00000000;
fma.rn.f32 %f413, %f412, %f405, %f411;
add.rn.f32 %f414, %f408, %f413;
neg.f32 %f415, %f414;
add.rn.f32 %f416, %f408, %f415;
add.rn.f32 %f417, %f416, %f413;
mov.b32 %r212, %f414;
setp.eq.s32 %p48, %r212, 1118925336;
add.s32 %r213, %r212, -1;
mov.b32 %f418, %r213;
add.f32 %f419, %f417, 0f37000000;
selp.f32 %f94, %f419, %f417, %p48;
selp.f32 %f420, %f418, %f414, %p48;
mov.f32 %f421, 0f3FB8AA3B;
mul.rn.f32 %f422, %f420, %f421;
cvt.rzi.f32.f32 %f423, %f422;
abs.f32 %f424, %f423;
setp.gt.f32 %p49, %f424, 0f42FC0000;
mov.b32 %r214, %f423;
and.b32 %r215, %r214, -2147483648;
or.b32 %r216, %r215, 1123811328;
mov.b32 %f425, %r216;
selp.f32 %f426, %f425, %f423, %p49;
mov.f32 %f427, 0fBF317218;
fma.rn.f32 %f428, %f426, %f427, %f420;
mov.f32 %f429, 0f3102E308;
fma.rn.f32 %f430, %f426, %f429, %f428;
mul.f32 %f431, %f430, 0f3FB8AA3B;
add.f32 %f432, %f426, 0f4B40007F;
mov.b32 %r217, %f432;
shl.b32 %r218, %r217, 23;
mov.b32 %f433, %r218;
ex2.approx.ftz.f32 %f434, %f431;
mul.f32 %f95, %f434, %f433;
setp.eq.f32 %p50, %f95, 0f7F800000;
mov.f32 %f1009, 0f7F800000;
@%p50 bra $L__BB0_41;
fma.rn.f32 %f1009, %f95, %f94, %f95;
$L__BB0_41:
mov.f32 %f969, 0f3E666666;
cvt.rzi.f32.f32 %f968, %f969;
add.f32 %f967, %f968, %f968;
mov.f32 %f966, 0f3EE66666;
sub.f32 %f965, %f966, %f967;
abs.f32 %f964, %f965;
setp.lt.f32 %p51, %f1006, 0f00000000;
setp.eq.f32 %p52, %f964, 0f3F800000;
and.pred %p1, %p51, %p52;
setp.eq.f32 %p53, %f1006, 0f00000000;
@%p53 bra $L__BB0_45;
bra.uni $L__BB0_42;
$L__BB0_45:
add.f32 %f439, %f1006, %f1006;
selp.f32 %f1011, %f439, 0f00000000, %p52;
bra.uni $L__BB0_46;
$L__BB0_42:
mov.b32 %r219, %f1009;
xor.b32 %r220, %r219, -2147483648;
mov.b32 %f435, %r220;
selp.f32 %f1011, %f435, %f1009, %p1;
setp.geu.f32 %p54, %f1006, 0f00000000;
@%p54 bra $L__BB0_46;
mov.f32 %f436, 0f3EE66666;
cvt.rzi.f32.f32 %f437, %f436;
setp.eq.f32 %p55, %f437, 0f3EE66666;
@%p55 bra $L__BB0_46;
mov.f32 %f1011, 0f7FFFFFFF;
$L__BB0_46:
abs.f32 %f970, %f1006;
add.f32 %f440, %f970, 0f3EE66666;
mov.b32 %r221, %f440;
setp.lt.s32 %p57, %r221, 2139095040;
@%p57 bra $L__BB0_51;
abs.f32 %f971, %f1006;
setp.gtu.f32 %p58, %f971, 0f7F800000;
@%p58 bra $L__BB0_50;
bra.uni $L__BB0_48;
$L__BB0_50:
add.f32 %f1011, %f1006, 0f3EE66666;
bra.uni $L__BB0_51;
$L__BB0_48:
abs.f32 %f972, %f1006;
setp.neu.f32 %p59, %f972, 0f7F800000;
@%p59 bra $L__BB0_51;
selp.f32 %f1011, 0fFF800000, 0f7F800000, %p1;
$L__BB0_51:
setp.eq.f32 %p60, %f1006, 0f3F800000;
selp.f32 %f104, 0f3F800000, %f1011, %p60;
abs.f32 %f105, %f1007;
setp.lt.f32 %p61, %f105, 0f00800000;
mul.f32 %f442, %f105, 0f4B800000;
selp.f32 %f443, %f442, %f105, %p61;
selp.f32 %f444, 0fC3170000, 0fC2FE0000, %p61;
mov.b32 %r222, %f443;
and.b32 %r223, %r222, 8388607;
or.b32 %r224, %r223, 1065353216;
mov.b32 %f445, %r224;
shr.u32 %r225, %r222, 23;
cvt.rn.f32.u32 %f446, %r225;
add.f32 %f447, %f444, %f446;
setp.gt.f32 %p62, %f445, 0f3FB504F3;
mul.f32 %f448, %f445, 0f3F000000;
add.f32 %f449, %f447, 0f3F800000;
selp.f32 %f450, %f449, %f447, %p62;
selp.f32 %f451, %f448, %f445, %p62;
add.f32 %f452, %f451, 0fBF800000;
add.f32 %f453, %f451, 0f3F800000;
rcp.approx.ftz.f32 %f454, %f453;
add.f32 %f455, %f452, %f452;
mul.f32 %f456, %f455, %f454;
mul.f32 %f457, %f456, %f456;
mov.f32 %f458, 0f3C4CAF63;
mov.f32 %f459, 0f3B18F0FE;
fma.rn.f32 %f460, %f459, %f457, %f458;
mov.f32 %f461, 0f3DAAAABD;
fma.rn.f32 %f462, %f460, %f457, %f461;
mul.rn.f32 %f463, %f462, %f457;
mul.rn.f32 %f464, %f463, %f456;
sub.f32 %f465, %f452, %f456;
add.f32 %f466, %f465, %f465;
neg.f32 %f467, %f456;
fma.rn.f32 %f468, %f467, %f452, %f466;
mul.rn.f32 %f469, %f454, %f468;
add.f32 %f470, %f464, %f456;
sub.f32 %f471, %f456, %f470;
add.f32 %f472, %f464, %f471;
add.f32 %f473, %f469, %f472;
add.f32 %f474, %f470, %f473;
sub.f32 %f475, %f470, %f474;
add.f32 %f476, %f473, %f475;
mov.f32 %f477, 0f3F317200;
mul.rn.f32 %f478, %f450, %f477;
mov.f32 %f479, 0f35BFBE8E;
mul.rn.f32 %f480, %f450, %f479;
add.f32 %f481, %f478, %f474;
sub.f32 %f482, %f478, %f481;
add.f32 %f483, %f474, %f482;
add.f32 %f484, %f476, %f483;
add.f32 %f485, %f480, %f484;
add.f32 %f486, %f481, %f485;
sub.f32 %f487, %f481, %f486;
add.f32 %f488, %f485, %f487;
mov.f32 %f489, 0f3EE66666;
mul.rn.f32 %f490, %f489, %f486;
neg.f32 %f491, %f490;
fma.rn.f32 %f492, %f489, %f486, %f491;
fma.rn.f32 %f493, %f489, %f488, %f492;
mov.f32 %f494, 0f00000000;
fma.rn.f32 %f495, %f494, %f486, %f493;
add.rn.f32 %f496, %f490, %f495;
neg.f32 %f497, %f496;
add.rn.f32 %f498, %f490, %f497;
add.rn.f32 %f499, %f498, %f495;
mov.b32 %r226, %f496;
setp.eq.s32 %p63, %r226, 1118925336;
add.s32 %r227, %r226, -1;
mov.b32 %f500, %r227;
add.f32 %f501, %f499, 0f37000000;
selp.f32 %f106, %f501, %f499, %p63;
selp.f32 %f502, %f500, %f496, %p63;
mov.f32 %f503, 0f3FB8AA3B;
mul.rn.f32 %f504, %f502, %f503;
cvt.rzi.f32.f32 %f505, %f504;
abs.f32 %f506, %f505;
setp.gt.f32 %p64, %f506, 0f42FC0000;
mov.b32 %r228, %f505;
and.b32 %r229, %r228, -2147483648;
or.b32 %r230, %r229, 1123811328;
mov.b32 %f507, %r230;
selp.f32 %f508, %f507, %f505, %p64;
mov.f32 %f509, 0fBF317218;
fma.rn.f32 %f510, %f508, %f509, %f502;
mov.f32 %f511, 0f3102E308;
fma.rn.f32 %f512, %f508, %f511, %f510;
mul.f32 %f513, %f512, 0f3FB8AA3B;
add.f32 %f514, %f508, 0f4B40007F;
mov.b32 %r231, %f514;
shl.b32 %r232, %r231, 23;
mov.b32 %f515, %r232;
ex2.approx.ftz.f32 %f516, %f513;
mul.f32 %f107, %f516, %f515;
setp.eq.f32 %p65, %f107, 0f7F800000;
mov.f32 %f1012, 0f7F800000;
@%p65 bra $L__BB0_53;
fma.rn.f32 %f1012, %f107, %f106, %f107;
$L__BB0_53:
setp.lt.f32 %p66, %f1007, 0f00000000;
and.pred %p2, %p66, %p52;
setp.eq.f32 %p68, %f1007, 0f00000000;
@%p68 bra $L__BB0_57;
bra.uni $L__BB0_54;
$L__BB0_57:
add.f32 %f521, %f1007, %f1007;
selp.f32 %f1014, %f521, 0f00000000, %p52;
bra.uni $L__BB0_58;
$L__BB0_54:
mov.b32 %r233, %f1012;
xor.b32 %r234, %r233, -2147483648;
mov.b32 %f517, %r234;
selp.f32 %f1014, %f517, %f1012, %p2;
setp.geu.f32 %p69, %f1007, 0f00000000;
@%p69 bra $L__BB0_58;
mov.f32 %f518, 0f3EE66666;
cvt.rzi.f32.f32 %f519, %f518;
setp.eq.f32 %p70, %f519, 0f3EE66666;
@%p70 bra $L__BB0_58;
mov.f32 %f1014, 0f7FFFFFFF;
$L__BB0_58:
abs.f32 %f973, %f1007;
add.f32 %f522, %f973, 0f3EE66666;
mov.b32 %r235, %f522;
setp.lt.s32 %p72, %r235, 2139095040;
@%p72 bra $L__BB0_63;
abs.f32 %f974, %f1007;
setp.gtu.f32 %p73, %f974, 0f7F800000;
@%p73 bra $L__BB0_62;
bra.uni $L__BB0_60;
$L__BB0_62:
add.f32 %f1014, %f1007, 0f3EE66666;
bra.uni $L__BB0_63;
$L__BB0_60:
abs.f32 %f975, %f1007;
setp.neu.f32 %p74, %f975, 0f7F800000;
@%p74 bra $L__BB0_63;
selp.f32 %f1014, 0fFF800000, 0f7F800000, %p2;
$L__BB0_63:
setp.eq.f32 %p75, %f1007, 0f3F800000;
selp.f32 %f116, 0f3F800000, %f1014, %p75;
abs.f32 %f117, %f1008;
setp.lt.f32 %p76, %f117, 0f00800000;
mul.f32 %f524, %f117, 0f4B800000;
selp.f32 %f525, %f524, %f117, %p76;
selp.f32 %f526, 0fC3170000, 0fC2FE0000, %p76;
mov.b32 %r236, %f525;
and.b32 %r237, %r236, 8388607;
or.b32 %r238, %r237, 1065353216;
mov.b32 %f527, %r238;
shr.u32 %r239, %r236, 23;
cvt.rn.f32.u32 %f528, %r239;
add.f32 %f529, %f526, %f528;
setp.gt.f32 %p77, %f527, 0f3FB504F3;
mul.f32 %f530, %f527, 0f3F000000;
add.f32 %f531, %f529, 0f3F800000;
selp.f32 %f532, %f531, %f529, %p77;
selp.f32 %f533, %f530, %f527, %p77;
add.f32 %f534, %f533, 0fBF800000;
add.f32 %f535, %f533, 0f3F800000;
rcp.approx.ftz.f32 %f536, %f535;
add.f32 %f537, %f534, %f534;
mul.f32 %f538, %f537, %f536;
mul.f32 %f539, %f538, %f538;
mov.f32 %f540, 0f3C4CAF63;
mov.f32 %f541, 0f3B18F0FE;
fma.rn.f32 %f542, %f541, %f539, %f540;
mov.f32 %f543, 0f3DAAAABD;
fma.rn.f32 %f544, %f542, %f539, %f543;
mul.rn.f32 %f545, %f544, %f539;
mul.rn.f32 %f546, %f545, %f538;
sub.f32 %f547, %f534, %f538;
add.f32 %f548, %f547, %f547;
neg.f32 %f549, %f538;
fma.rn.f32 %f550, %f549, %f534, %f548;
mul.rn.f32 %f551, %f536, %f550;
add.f32 %f552, %f546, %f538;
sub.f32 %f553, %f538, %f552;
add.f32 %f554, %f546, %f553;
add.f32 %f555, %f551, %f554;
add.f32 %f556, %f552, %f555;
sub.f32 %f557, %f552, %f556;
add.f32 %f558, %f555, %f557;
mov.f32 %f559, 0f3F317200;
mul.rn.f32 %f560, %f532, %f559;
mov.f32 %f561, 0f35BFBE8E;
mul.rn.f32 %f562, %f532, %f561;
add.f32 %f563, %f560, %f556;
sub.f32 %f564, %f560, %f563;
add.f32 %f565, %f556, %f564;
add.f32 %f566, %f558, %f565;
add.f32 %f567, %f562, %f566;
add.f32 %f568, %f563, %f567;
sub.f32 %f569, %f563, %f568;
add.f32 %f570, %f567, %f569;
mov.f32 %f571, 0f3EE66666;
mul.rn.f32 %f572, %f571, %f568;
neg.f32 %f573, %f572;
fma.rn.f32 %f574, %f571, %f568, %f573;
fma.rn.f32 %f575, %f571, %f570, %f574;
mov.f32 %f576, 0f00000000;
fma.rn.f32 %f577, %f576, %f568, %f575;
add.rn.f32 %f578, %f572, %f577;
neg.f32 %f579, %f578;
add.rn.f32 %f580, %f572, %f579;
add.rn.f32 %f581, %f580, %f577;
mov.b32 %r240, %f578;
setp.eq.s32 %p78, %r240, 1118925336;
add.s32 %r241, %r240, -1;
mov.b32 %f582, %r241;
add.f32 %f583, %f581, 0f37000000;
selp.f32 %f118, %f583, %f581, %p78;
selp.f32 %f584, %f582, %f578, %p78;
mov.f32 %f585, 0f3FB8AA3B;
mul.rn.f32 %f586, %f584, %f585;
cvt.rzi.f32.f32 %f587, %f586;
abs.f32 %f588, %f587;
setp.gt.f32 %p79, %f588, 0f42FC0000;
mov.b32 %r242, %f587;
and.b32 %r243, %r242, -2147483648;
or.b32 %r244, %r243, 1123811328;
mov.b32 %f589, %r244;
selp.f32 %f590, %f589, %f587, %p79;
mov.f32 %f591, 0fBF317218;
fma.rn.f32 %f592, %f590, %f591, %f584;
mov.f32 %f593, 0f3102E308;
fma.rn.f32 %f594, %f590, %f593, %f592;
mul.f32 %f595, %f594, 0f3FB8AA3B;
add.f32 %f596, %f590, 0f4B40007F;
mov.b32 %r245, %f596;
shl.b32 %r246, %r245, 23;
mov.b32 %f597, %r246;
ex2.approx.ftz.f32 %f598, %f595;
mul.f32 %f119, %f598, %f597;
setp.eq.f32 %p80, %f119, 0f7F800000;
mov.f32 %f1015, 0f7F800000;
@%p80 bra $L__BB0_65;
fma.rn.f32 %f1015, %f119, %f118, %f119;
$L__BB0_65:
setp.lt.f32 %p81, %f1008, 0f00000000;
and.pred %p3, %p81, %p52;
setp.eq.f32 %p83, %f1008, 0f00000000;
@%p83 bra $L__BB0_69;
bra.uni $L__BB0_66;
$L__BB0_69:
add.f32 %f603, %f1008, %f1008;
selp.f32 %f1017, %f603, 0f00000000, %p52;
bra.uni $L__BB0_70;
$L__BB0_66:
mov.b32 %r247, %f1015;
xor.b32 %r248, %r247, -2147483648;
mov.b32 %f599, %r248;
selp.f32 %f1017, %f599, %f1015, %p3;
setp.geu.f32 %p84, %f1008, 0f00000000;
@%p84 bra $L__BB0_70;
mov.f32 %f600, 0f3EE66666;
cvt.rzi.f32.f32 %f601, %f600;
setp.eq.f32 %p85, %f601, 0f3EE66666;
@%p85 bra $L__BB0_70;
mov.f32 %f1017, 0f7FFFFFFF;
$L__BB0_70:
abs.f32 %f976, %f1008;
add.f32 %f604, %f976, 0f3EE66666;
mov.b32 %r249, %f604;
setp.lt.s32 %p87, %r249, 2139095040;
@%p87 bra $L__BB0_75;
abs.f32 %f977, %f1008;
setp.gtu.f32 %p88, %f977, 0f7F800000;
@%p88 bra $L__BB0_74;
bra.uni $L__BB0_72;
$L__BB0_74:
add.f32 %f1017, %f1008, 0f3EE66666;
bra.uni $L__BB0_75;
$L__BB0_72:
abs.f32 %f978, %f1008;
setp.neu.f32 %p89, %f978, 0f7F800000;
@%p89 bra $L__BB0_75;
selp.f32 %f1017, 0fFF800000, 0f7F800000, %p3;
$L__BB0_75:
setp.eq.f32 %p90, %f1008, 0f3F800000;
mov.f32 %f606, 0f3F800000;
selp.f32 %f607, 0f3F800000, %f1017, %p90;
min.f32 %f608, %f104, %f606;
mov.f32 %f609, 0f00000000;
max.f32 %f128, %f609, %f608;
min.f32 %f610, %f116, %f606;
max.f32 %f129, %f609, %f610;
min.f32 %f611, %f607, %f606;
max.f32 %f130, %f609, %f611;
mov.f32 %f615, 0f3ED55555;
abs.f32 %f132, %f128;
setp.lt.f32 %p91, %f132, 0f00800000;
mul.f32 %f617, %f132, 0f4B800000;
selp.f32 %f618, %f617, %f132, %p91;
selp.f32 %f619, 0fC3170000, 0fC2FE0000, %p91;
mov.b32 %r252, %f618;
and.b32 %r253, %r252, 8388607;
or.b32 %r254, %r253, 1065353216;
mov.b32 %f620, %r254;
shr.u32 %r255, %r252, 23;
cvt.rn.f32.u32 %f621, %r255;
add.f32 %f622, %f619, %f621;
setp.gt.f32 %p92, %f620, 0f3FB504F3;
mul.f32 %f623, %f620, 0f3F000000;
add.f32 %f624, %f622, 0f3F800000;
selp.f32 %f625, %f624, %f622, %p92;
selp.f32 %f626, %f623, %f620, %p92;
add.f32 %f627, %f626, 0fBF800000;
add.f32 %f628, %f626, 0f3F800000;
rcp.approx.ftz.f32 %f629, %f628;
add.f32 %f630, %f627, %f627;
mul.f32 %f631, %f630, %f629;
mul.f32 %f632, %f631, %f631;
mov.f32 %f633, 0f3C4CAF63;
mov.f32 %f634, 0f3B18F0FE;
fma.rn.f32 %f635, %f634, %f632, %f633;
mov.f32 %f636, 0f3DAAAABD;
fma.rn.f32 %f637, %f635, %f632, %f636;
mul.rn.f32 %f638, %f637, %f632;
mul.rn.f32 %f639, %f638, %f631;
sub.f32 %f640, %f627, %f631;
add.f32 %f641, %f640, %f640;
neg.f32 %f642, %f631;
fma.rn.f32 %f643, %f642, %f627, %f641;
mul.rn.f32 %f644, %f629, %f643;
add.f32 %f645, %f639, %f631;
sub.f32 %f646, %f631, %f645;
add.f32 %f647, %f639, %f646;
add.f32 %f648, %f644, %f647;
add.f32 %f649, %f645, %f648;
sub.f32 %f650, %f645, %f649;
add.f32 %f651, %f648, %f650;
mov.f32 %f652, 0f3F317200;
mul.rn.f32 %f653, %f625, %f652;
mov.f32 %f654, 0f35BFBE8E;
mul.rn.f32 %f655, %f625, %f654;
add.f32 %f656, %f653, %f649;
sub.f32 %f657, %f653, %f656;
add.f32 %f658, %f649, %f657;
add.f32 %f659, %f651, %f658;
add.f32 %f660, %f655, %f659;
add.f32 %f661, %f656, %f660;
sub.f32 %f662, %f656, %f661;
add.f32 %f663, %f660, %f662;
mul.rn.f32 %f664, %f615, %f661;
neg.f32 %f665, %f664;
fma.rn.f32 %f666, %f615, %f661, %f665;
fma.rn.f32 %f667, %f615, %f663, %f666;
fma.rn.f32 %f668, %f609, %f661, %f667;
add.rn.f32 %f669, %f664, %f668;
neg.f32 %f670, %f669;
add.rn.f32 %f671, %f664, %f670;
add.rn.f32 %f672, %f671, %f668;
mov.b32 %r256, %f669;
setp.eq.s32 %p93, %r256, 1118925336;
add.s32 %r257, %r256, -1;
mov.b32 %f673, %r257;
add.f32 %f674, %f672, 0f37000000;
selp.f32 %f133, %f674, %f672, %p93;
selp.f32 %f675, %f673, %f669, %p93;
mov.f32 %f676, 0f3FB8AA3B;
mul.rn.f32 %f677, %f675, %f676;
cvt.rzi.f32.f32 %f678, %f677;
abs.f32 %f679, %f678;
setp.gt.f32 %p94, %f679, 0f42FC0000;
mov.b32 %r258, %f678;
and.b32 %r259, %r258, -2147483648;
or.b32 %r260, %r259, 1123811328;
mov.b32 %f680, %r260;
selp.f32 %f681, %f680, %f678, %p94;
mov.f32 %f682, 0fBF317218;
fma.rn.f32 %f683, %f681, %f682, %f675;
mov.f32 %f684, 0f3102E308;
fma.rn.f32 %f685, %f681, %f684, %f683;
mul.f32 %f686, %f685, 0f3FB8AA3B;
add.f32 %f687, %f681, 0f4B40007F;
mov.b32 %r261, %f687;
shl.b32 %r262, %r261, 23;
mov.b32 %f688, %r262;
ex2.approx.ftz.f32 %f689, %f686;
mul.f32 %f134, %f689, %f688;
setp.eq.f32 %p95, %f134, 0f7F800000;
mov.f32 %f1018, 0f7F800000;
@%p95 bra $L__BB0_77;
fma.rn.f32 %f1018, %f134, %f133, %f134;
$L__BB0_77:
mov.f32 %f959, 0f3E555555;
cvt.rzi.f32.f32 %f958, %f959;
add.f32 %f957, %f958, %f958;
mov.f32 %f956, 0f3ED55555;
sub.f32 %f955, %f956, %f957;
abs.f32 %f954, %f955;
setp.lt.f32 %p96, %f128, 0f00000000;
setp.eq.f32 %p97, %f954, 0f3F800000;
and.pred %p4, %p96, %p97;
setp.eq.f32 %p98, %f128, 0f00000000;
@%p98 bra $L__BB0_81;
bra.uni $L__BB0_78;
$L__BB0_81:
add.f32 %f694, %f128, %f128;
selp.f32 %f1020, %f694, 0f00000000, %p97;
bra.uni $L__BB0_82;
$L__BB0_78:
mov.b32 %r263, %f1018;
xor.b32 %r264, %r263, -2147483648;
mov.b32 %f690, %r264;
selp.f32 %f1020, %f690, %f1018, %p4;
setp.geu.f32 %p99, %f128, 0f00000000;
@%p99 bra $L__BB0_82;
mov.f32 %f691, 0f3ED55555;
cvt.rzi.f32.f32 %f692, %f691;
setp.eq.f32 %p100, %f692, 0f3ED55555;
@%p100 bra $L__BB0_82;
mov.f32 %f1020, 0f7FFFFFFF;
$L__BB0_82:
add.f32 %f695, %f132, 0f3ED55555;
mov.b32 %r265, %f695;
setp.lt.s32 %p102, %r265, 2139095040;
@%p102 bra $L__BB0_87;
setp.gtu.f32 %p103, %f132, 0f7F800000;
@%p103 bra $L__BB0_86;
bra.uni $L__BB0_84;
$L__BB0_86:
add.f32 %f1020, %f128, 0f3ED55555;
bra.uni $L__BB0_87;
$L__BB0_84:
setp.neu.f32 %p104, %f132, 0f7F800000;
@%p104 bra $L__BB0_87;
selp.f32 %f1020, 0fFF800000, 0f7F800000, %p4;
$L__BB0_87:
abs.f32 %f143, %f129;
setp.lt.f32 %p105, %f143, 0f00800000;
mul.f32 %f697, %f143, 0f4B800000;
selp.f32 %f698, %f697, %f143, %p105;
selp.f32 %f699, 0fC3170000, 0fC2FE0000, %p105;
mov.b32 %r266, %f698;
and.b32 %r267, %r266, 8388607;
or.b32 %r268, %r267, 1065353216;
mov.b32 %f700, %r268;
shr.u32 %r269, %r266, 23;
cvt.rn.f32.u32 %f701, %r269;
add.f32 %f702, %f699, %f701;
setp.gt.f32 %p106, %f700, 0f3FB504F3;
mul.f32 %f703, %f700, 0f3F000000;
add.f32 %f704, %f702, 0f3F800000;
selp.f32 %f705, %f704, %f702, %p106;
selp.f32 %f706, %f703, %f700, %p106;
add.f32 %f707, %f706, 0fBF800000;
add.f32 %f708, %f706, 0f3F800000;
rcp.approx.ftz.f32 %f709, %f708;
add.f32 %f710, %f707, %f707;
mul.f32 %f711, %f710, %f709;
mul.f32 %f712, %f711, %f711;
mov.f32 %f713, 0f3C4CAF63;
mov.f32 %f714, 0f3B18F0FE;
fma.rn.f32 %f715, %f714, %f712, %f713;
mov.f32 %f716, 0f3DAAAABD;
fma.rn.f32 %f717, %f715, %f712, %f716;
mul.rn.f32 %f718, %f717, %f712;
mul.rn.f32 %f719, %f718, %f711;
sub.f32 %f720, %f707, %f711;
add.f32 %f721, %f720, %f720;
neg.f32 %f722, %f711;
fma.rn.f32 %f723, %f722, %f707, %f721;
mul.rn.f32 %f724, %f709, %f723;
add.f32 %f725, %f719, %f711;
sub.f32 %f726, %f711, %f725;
add.f32 %f727, %f719, %f726;
add.f32 %f728, %f724, %f727;
add.f32 %f729, %f725, %f728;
sub.f32 %f730, %f725, %f729;
add.f32 %f731, %f728, %f730;
mov.f32 %f732, 0f3F317200;
mul.rn.f32 %f733, %f705, %f732;
mov.f32 %f734, 0f35BFBE8E;
mul.rn.f32 %f735, %f705, %f734;
add.f32 %f736, %f733, %f729;
sub.f32 %f737, %f733, %f736;
add.f32 %f738, %f729, %f737;
add.f32 %f739, %f731, %f738;
add.f32 %f740, %f735, %f739;
add.f32 %f741, %f736, %f740;
sub.f32 %f742, %f736, %f741;
add.f32 %f743, %f740, %f742;
mov.f32 %f744, 0f3ED55555;
mul.rn.f32 %f745, %f744, %f741;
neg.f32 %f746, %f745;
fma.rn.f32 %f747, %f744, %f741, %f746;
fma.rn.f32 %f748, %f744, %f743, %f747;
mov.f32 %f749, 0f00000000;
fma.rn.f32 %f750, %f749, %f741, %f748;
add.rn.f32 %f751, %f745, %f750;
neg.f32 %f752, %f751;
add.rn.f32 %f753, %f745, %f752;
add.rn.f32 %f754, %f753, %f750;
mov.b32 %r270, %f751;
setp.eq.s32 %p107, %r270, 1118925336;
add.s32 %r271, %r270, -1;
mov.b32 %f755, %r271;
add.f32 %f756, %f754, 0f37000000;
selp.f32 %f144, %f756, %f754, %p107;
selp.f32 %f757, %f755, %f751, %p107;
mov.f32 %f758, 0f3FB8AA3B;
mul.rn.f32 %f759, %f757, %f758;
cvt.rzi.f32.f32 %f760, %f759;
abs.f32 %f761, %f760;
setp.gt.f32 %p108, %f761, 0f42FC0000;
mov.b32 %r272, %f760;
and.b32 %r273, %r272, -2147483648;
or.b32 %r274, %r273, 1123811328;
mov.b32 %f762, %r274;
selp.f32 %f763, %f762, %f760, %p108;
mov.f32 %f764, 0fBF317218;
fma.rn.f32 %f765, %f763, %f764, %f757;
mov.f32 %f766, 0f3102E308;
fma.rn.f32 %f767, %f763, %f766, %f765;
mul.f32 %f768, %f767, 0f3FB8AA3B;
add.f32 %f769, %f763, 0f4B40007F;
mov.b32 %r275, %f769;
shl.b32 %r276, %r275, 23;
mov.b32 %f770, %r276;
ex2.approx.ftz.f32 %f771, %f768;
mul.f32 %f145, %f771, %f770;
setp.eq.f32 %p109, %f145, 0f7F800000;
mov.f32 %f1021, 0f7F800000;
@%p109 bra $L__BB0_89;
fma.rn.f32 %f1021, %f145, %f144, %f145;
$L__BB0_89:
setp.lt.f32 %p110, %f129, 0f00000000;
and.pred %p5, %p110, %p97;
setp.eq.f32 %p112, %f129, 0f00000000;
@%p112 bra $L__BB0_93;
bra.uni $L__BB0_90;
$L__BB0_93:
add.f32 %f776, %f129, %f129;
selp.f32 %f1023, %f776, 0f00000000, %p97;
bra.uni $L__BB0_94;
$L__BB0_90:
mov.b32 %r277, %f1021;
xor.b32 %r278, %r277, -2147483648;
mov.b32 %f772, %r278;
selp.f32 %f1023, %f772, %f1021, %p5;
setp.geu.f32 %p113, %f129, 0f00000000;
@%p113 bra $L__BB0_94;
mov.f32 %f773, 0f3ED55555;
cvt.rzi.f32.f32 %f774, %f773;
setp.eq.f32 %p114, %f774, 0f3ED55555;
@%p114 bra $L__BB0_94;
mov.f32 %f1023, 0f7FFFFFFF;
$L__BB0_94:
add.f32 %f777, %f143, 0f3ED55555;
mov.b32 %r279, %f777;
setp.lt.s32 %p116, %r279, 2139095040;
@%p116 bra $L__BB0_99;
setp.gtu.f32 %p117, %f143, 0f7F800000;
@%p117 bra $L__BB0_98;
bra.uni $L__BB0_96;
$L__BB0_98:
add.f32 %f1023, %f129, 0f3ED55555;
bra.uni $L__BB0_99;
$L__BB0_96:
setp.neu.f32 %p118, %f143, 0f7F800000;
@%p118 bra $L__BB0_99;
selp.f32 %f1023, 0fFF800000, 0f7F800000, %p5;
$L__BB0_99:
abs.f32 %f154, %f130;
setp.lt.f32 %p119, %f154, 0f00800000;
mul.f32 %f779, %f154, 0f4B800000;
selp.f32 %f780, %f779, %f154, %p119;
selp.f32 %f781, 0fC3170000, 0fC2FE0000, %p119;
mov.b32 %r280, %f780;
and.b32 %r281, %r280, 8388607;
or.b32 %r282, %r281, 1065353216;
mov.b32 %f782, %r282;
shr.u32 %r283, %r280, 23;
cvt.rn.f32.u32 %f783, %r283;
add.f32 %f784, %f781, %f783;
setp.gt.f32 %p120, %f782, 0f3FB504F3;
mul.f32 %f785, %f782, 0f3F000000;
add.f32 %f786, %f784, 0f3F800000;
selp.f32 %f787, %f786, %f784, %p120;
selp.f32 %f788, %f785, %f782, %p120;
add.f32 %f789, %f788, 0fBF800000;
add.f32 %f790, %f788, 0f3F800000;
rcp.approx.ftz.f32 %f791, %f790;
add.f32 %f792, %f789, %f789;
mul.f32 %f793, %f792, %f791;
mul.f32 %f794, %f793, %f793;
mov.f32 %f795, 0f3C4CAF63;
mov.f32 %f796, 0f3B18F0FE;
fma.rn.f32 %f797, %f796, %f794, %f795;
mov.f32 %f798, 0f3DAAAABD;
fma.rn.f32 %f799, %f797, %f794, %f798;
mul.rn.f32 %f800, %f799, %f794;
mul.rn.f32 %f801, %f800, %f793;
sub.f32 %f802, %f789, %f793;
add.f32 %f803, %f802, %f802;
neg.f32 %f804, %f793;
fma.rn.f32 %f805, %f804, %f789, %f803;
mul.rn.f32 %f806, %f791, %f805;
add.f32 %f807, %f801, %f793;
sub.f32 %f808, %f793, %f807;
add.f32 %f809, %f801, %f808;
add.f32 %f810, %f806, %f809;
add.f32 %f811, %f807, %f810;
sub.f32 %f812, %f807, %f811;
add.f32 %f813, %f810, %f812;
mov.f32 %f814, 0f3F317200;
mul.rn.f32 %f815, %f787, %f814;
mov.f32 %f816, 0f35BFBE8E;
mul.rn.f32 %f817, %f787, %f816;
add.f32 %f818, %f815, %f811;
sub.f32 %f819, %f815, %f818;
add.f32 %f820, %f811, %f819;
add.f32 %f821, %f813, %f820;
add.f32 %f822, %f817, %f821;
add.f32 %f823, %f818, %f822;
sub.f32 %f824, %f818, %f823;
add.f32 %f825, %f822, %f824;
mov.f32 %f826, 0f3ED55555;
mul.rn.f32 %f827, %f826, %f823;
neg.f32 %f828, %f827;
fma.rn.f32 %f829, %f826, %f823, %f828;
fma.rn.f32 %f830, %f826, %f825, %f829;
mov.f32 %f831, 0f00000000;
fma.rn.f32 %f832, %f831, %f823, %f830;
add.rn.f32 %f833, %f827, %f832;
neg.f32 %f834, %f833;
add.rn.f32 %f835, %f827, %f834;
add.rn.f32 %f836, %f835, %f832;
mov.b32 %r284, %f833;
setp.eq.s32 %p121, %r284, 1118925336;
add.s32 %r285, %r284, -1;
mov.b32 %f837, %r285;
add.f32 %f838, %f836, 0f37000000;
selp.f32 %f155, %f838, %f836, %p121;
selp.f32 %f839, %f837, %f833, %p121;
mov.f32 %f840, 0f3FB8AA3B;
mul.rn.f32 %f841, %f839, %f840;
cvt.rzi.f32.f32 %f842, %f841;
abs.f32 %f843, %f842;
setp.gt.f32 %p122, %f843, 0f42FC0000;
mov.b32 %r286, %f842;
and.b32 %r287, %r286, -2147483648;
or.b32 %r288, %r287, 1123811328;
mov.b32 %f844, %r288;
selp.f32 %f845, %f844, %f842, %p122;
mov.f32 %f846, 0fBF317218;
fma.rn.f32 %f847, %f845, %f846, %f839;
mov.f32 %f848, 0f3102E308;
fma.rn.f32 %f849, %f845, %f848, %f847;
mul.f32 %f850, %f849, 0f3FB8AA3B;
add.f32 %f851, %f845, 0f4B40007F;
mov.b32 %r289, %f851;
shl.b32 %r290, %r289, 23;
mov.b32 %f852, %r290;
ex2.approx.ftz.f32 %f853, %f850;
mul.f32 %f156, %f853, %f852;
setp.eq.f32 %p123, %f156, 0f7F800000;
mov.f32 %f1024, 0f7F800000;
@%p123 bra $L__BB0_101;
fma.rn.f32 %f1024, %f156, %f155, %f156;
$L__BB0_101:
setp.lt.f32 %p124, %f130, 0f00000000;
and.pred %p6, %p124, %p97;
setp.eq.f32 %p126, %f130, 0f00000000;
@%p126 bra $L__BB0_105;
bra.uni $L__BB0_102;
$L__BB0_105:
add.f32 %f858, %f130, %f130;
selp.f32 %f1026, %f858, 0f00000000, %p97;
bra.uni $L__BB0_106;
$L__BB0_102:
mov.b32 %r291, %f1024;
xor.b32 %r292, %r291, -2147483648;
mov.b32 %f854, %r292;
selp.f32 %f1026, %f854, %f1024, %p6;
setp.geu.f32 %p127, %f130, 0f00000000;
@%p127 bra $L__BB0_106;
mov.f32 %f855, 0f3ED55555;
cvt.rzi.f32.f32 %f856, %f855;
setp.eq.f32 %p128, %f856, 0f3ED55555;
@%p128 bra $L__BB0_106;
mov.f32 %f1026, 0f7FFFFFFF;
$L__BB0_106:
add.f32 %f859, %f154, 0f3ED55555;
mov.b32 %r293, %f859;
setp.lt.s32 %p130, %r293, 2139095040;
@%p130 bra $L__BB0_111;
setp.gtu.f32 %p131, %f154, 0f7F800000;
@%p131 bra $L__BB0_110;
bra.uni $L__BB0_108;
$L__BB0_110:
add.f32 %f1026, %f130, 0f3ED55555;
bra.uni $L__BB0_111;
$L__BB0_108:
setp.neu.f32 %p132, %f154, 0f7F800000;
@%p132 bra $L__BB0_111;
selp.f32 %f1026, 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 %f860, %f1020, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p133, %f128, 0f3F800000;
mov.f32 %f861, 0f3F800000;
selp.f32 %f862, 0f3F7FFFFF, %f860, %p133;
mul.f32 %f863, %f128, 0f414EB852;
setp.lt.f32 %p134, %f128, 0f3B4D2E1C;
selp.f32 %f864, %f863, %f862, %p134;
fma.rn.f32 %f865, %f1023, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p135, %f129, 0f3F800000;
selp.f32 %f866, 0f3F7FFFFF, %f865, %p135;
mul.f32 %f867, %f129, 0f414EB852;
setp.lt.f32 %p136, %f129, 0f3B4D2E1C;
selp.f32 %f868, %f867, %f866, %p136;
fma.rn.f32 %f869, %f1026, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p137, %f130, 0f3F800000;
selp.f32 %f870, 0f3F7FFFFF, %f869, %p137;
mul.f32 %f871, %f130, 0f414EB852;
setp.lt.f32 %p138, %f130, 0f3B4D2E1C;
selp.f32 %f872, %f871, %f870, %p138;
min.f32 %f873, %f864, %f861;
mov.f32 %f874, 0f00000000;
max.f32 %f875, %f874, %f873;
mul.f32 %f876, %f875, 0f43800000;
cvt.rzi.u32.f32 %r294, %f876;
min.u32 %r295, %r294, 255;
min.f32 %f877, %f868, %f861;
max.f32 %f878, %f874, %f877;
mul.f32 %f879, %f878, 0f43800000;
cvt.rzi.u32.f32 %r296, %f879;
min.u32 %r297, %r296, 255;
min.f32 %f880, %f872, %f861;
max.f32 %f881, %f874, %f880;
mul.f32 %f882, %f881, 0f43800000;
cvt.rzi.u32.f32 %r298, %f882;
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 %p139, %r300, 0;
@%p139 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 %f886, 0f3F800000;
mul.wide.u32 %rd64, %r302, 8;
add.s64 %rd65, %rd63, %rd64;
// begin inline asm
{ cvt.rn.f16.f32 %rs30, %f886;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs29, %f1008;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs28, %f1007;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs27, %f1006;}
// 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 %p140, %r49, 0;
@%p140 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 %f890, 0f3F800000;
mul.wide.u32 %rd68, %r304, 8;
add.s64 %rd69, %rd67, %rd68;
// begin inline asm
{ cvt.rn.f16.f32 %rs34, %f890;}
// 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 %p141, %r305, 0;
@%p141 bra $L__BB0_131;
mul.f32 %f891, %f196, 0f3F372474;
fma.rn.f32 %f892, %f195, 0f3E59999A, %f891;
fma.rn.f32 %f165, %f197, 0f3D93A92A, %f892;
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 %f893, %rs35;
div.rn.f32 %f894, %f893, 0f437F0000;
fma.rn.f32 %f895, %f894, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f896, %rs36;
div.rn.f32 %f897, %f896, 0f437F0000;
fma.rn.f32 %f898, %f897, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f899, %rs37;
div.rn.f32 %f900, %f899, 0f437F0000;
fma.rn.f32 %f901, %f900, 0f40000000, 0fBF800000;
mul.f32 %f902, %f898, %f898;
fma.rn.f32 %f903, %f895, %f895, %f902;
fma.rn.f32 %f904, %f901, %f901, %f903;
sqrt.rn.f32 %f905, %f904;
rcp.rn.f32 %f906, %f905;
mul.f32 %f166, %f906, %f895;
mul.f32 %f167, %f906, %f898;
mul.f32 %f168, %f906, %f901;
abs.f32 %f907, %f993;
setp.gt.f32 %p142, %f907, 0f00000000;
@%p142 bra $L__BB0_120;
abs.f32 %f908, %f992;
setp.gt.f32 %p143, %f908, 0f00000000;
@%p143 bra $L__BB0_120;
bra.uni $L__BB0_119;
$L__BB0_120:
setp.ne.s32 %p147, %r49, 0;
mov.f32 %f1027, %f993;
mov.f32 %f1028, %f992;
mov.f32 %f1029, %f991;
@%p147 bra $L__BB0_122;
$L__BB0_121:
mul.f32 %f910, %f992, %f992;
fma.rn.f32 %f911, %f993, %f993, %f910;
fma.rn.f32 %f912, %f991, %f991, %f911;
sqrt.rn.f32 %f913, %f912;
rcp.rn.f32 %f914, %f913;
mul.f32 %f915, %f993, %f914;
mul.f32 %f916, %f992, %f914;
mul.f32 %f917, %f991, %f914;
mul.f32 %f918, %f78, 0f3F372474;
fma.rn.f32 %f919, %f77, 0f3E59999A, %f918;
fma.rn.f32 %f920, %f79, 0f3D93A92A, %f919;
mul.f32 %f921, %f920, %f915;
mul.f32 %f922, %f920, %f916;
mul.f32 %f923, %f920, %f917;
fma.rn.f32 %f1027, %f165, %f166, %f921;
fma.rn.f32 %f1028, %f165, %f167, %f922;
fma.rn.f32 %f1029, %f165, %f168, %f923;
$L__BB0_122:
mul.f32 %f924, %f1028, %f1028;
fma.rn.f32 %f925, %f1027, %f1027, %f924;
fma.rn.f32 %f926, %f1029, %f1029, %f925;
sqrt.rn.f32 %f927, %f926;
rcp.rn.f32 %f928, %f927;
mul.f32 %f929, %f1027, %f928;
mul.f32 %f930, %f1028, %f928;
mul.f32 %f931, %f1029, %f928;
mul.f32 %f932, %f980, %f930;
fma.rn.f32 %f933, %f979, %f929, %f932;
fma.rn.f32 %f934, %f981, %f931, %f933;
fma.rn.f32 %f935, %f934, 0f3F000000, 0f3F000000;
mov.f32 %f936, 0f3B808081;
max.f32 %f937, %f935, %f936;
mul.f32 %f938, %f937, 0f437F0000;
cvt.rzi.s32.f32 %r308, %f938;
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 %f939, %f929, 0f3F000000, 0f3F000000;
mul.f32 %f940, %f939, 0f437F0000;
cvt.rzi.u32.f32 %r311, %f940;
fma.rn.f32 %f941, %f930, 0f3F000000, 0f3F000000;
mul.f32 %f942, %f941, 0f437F0000;
cvt.rzi.u32.f32 %r312, %f942;
fma.rn.f32 %f943, %f931, 0f3F000000, 0f3F000000;
mul.f32 %f944, %f943, 0f437F0000;
cvt.rzi.u32.f32 %r313, %f944;
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 %f909, %f991;
setp.leu.f32 %p144, %f909, 0f00000000;
setp.ne.s32 %p145, %r49, 0;
or.pred %p146, %p144, %p145;
selp.f32 %f1027, %f166, %f993, %p144;
selp.f32 %f1028, %f167, %f992, %p144;
selp.f32 %f1029, %f168, %f991, %p144;
@%p146 bra $L__BB0_122;
bra.uni $L__BB0_121;
}