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

2690 lines
74 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<193>;
.reg .b16 %rs<181>;
.reg .f32 %f<1482>;
.reg .b32 %r<375>;
.reg .f64 %fd<5>;
.reg .b64 %rd<116>;
mov.u64 %SPL, __local_depot0;
add.u64 %rd1, %SPL, 0;
// begin inline asm
call (%r53), _optix_get_launch_index_x, ();
// end inline asm
// begin inline asm
call (%r54), _optix_get_launch_index_y, ();
// end inline asm
ld.const.u64 %rd29, [params+400];
cvta.to.global.u64 %rd30, %rd29;
ld.const.u32 %r59, [params+392];
mad.lo.s32 %r60, %r59, %r54, %r53;
mul.wide.u32 %rd31, %r60, 4;
add.s64 %rd2, %rd30, %rd31;
ld.global.v2.u8 {%rs16, %rs176}, [%rd2];
or.b16 %rs18, %rs16, %rs176;
and.b16 %rs19, %rs18, 255;
setp.eq.s16 %p9, %rs19, 0;
@%p9 bra $L__BB0_2;
ld.global.u8 %rs175, [%rd2+2];
bra.uni $L__BB0_3;
$L__BB0_2:
ld.global.u8 %rs175, [%rd2+2];
setp.eq.s16 %p10, %rs175, 0;
mov.f32 %f1428, 0f00000000;
mov.u16 %rs176, 0;
mov.f32 %f1429, %f1428;
mov.f32 %f1430, %f1428;
@%p10 bra $L__BB0_4;
$L__BB0_3:
cvt.rn.f32.u16 %f224, %rs16;
div.rn.f32 %f225, %f224, 0f437F0000;
fma.rn.f32 %f226, %f225, 0f40000000, 0fBF800000;
and.b16 %rs22, %rs176, 255;
cvt.rn.f32.u16 %f227, %rs22;
div.rn.f32 %f228, %f227, 0f437F0000;
fma.rn.f32 %f229, %f228, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f230, %rs175;
div.rn.f32 %f231, %f230, 0f437F0000;
fma.rn.f32 %f232, %f231, 0f40000000, 0fBF800000;
mul.f32 %f233, %f229, %f229;
fma.rn.f32 %f234, %f226, %f226, %f233;
fma.rn.f32 %f235, %f232, %f232, %f234;
sqrt.rn.f32 %f236, %f235;
rcp.rn.f32 %f237, %f236;
mul.f32 %f1430, %f237, %f232;
mul.f32 %f1429, %f237, %f229;
mul.f32 %f1428, %f226, %f237;
$L__BB0_4:
ld.const.v2.u32 {%r61, %r62}, [params];
add.s32 %r3, %r61, %r53;
add.s32 %r4, %r62, %r54;
setp.eq.f32 %p11, %f1428, 0f00000000;
setp.eq.f32 %p12, %f1429, 0f00000000;
and.pred %p13, %p11, %p12;
setp.eq.f32 %p14, %f1430, 0f00000000;
and.pred %p15, %p14, %p13;
@%p15 bra $L__BB0_157;
bra.uni $L__BB0_5;
$L__BB0_157:
ld.const.u32 %r50, [params+104];
and.b32 %r344, %r50, 1;
setp.eq.b32 %p183, %r344, 1;
mov.pred %p184, 0;
xor.pred %p185, %p183, %p184;
not.pred %p186, %p185;
@%p186 bra $L__BB0_159;
ld.const.u64 %rd87, [params+144];
cvta.to.global.u64 %rd88, %rd87;
ld.const.u32 %r345, [params+136];
mad.lo.s32 %r346, %r345, %r4, %r3;
mul.wide.u32 %rd89, %r346, 4;
add.s64 %rd90, %rd88, %rd89;
mov.u16 %rs112, 0;
st.global.v4.u8 [%rd90], {%rs112, %rs112, %rs112, %rs112};
$L__BB0_159:
and.b32 %r347, %r50, 8;
setp.eq.s32 %p187, %r347, 0;
@%p187 bra $L__BB0_161;
ld.const.u64 %rd91, [params+192];
cvta.to.global.u64 %rd92, %rd91;
ld.const.u32 %r348, [params+184];
mad.lo.s32 %r349, %r348, %r4, %r3;
mov.f32 %f1324, 0f00000000;
cvt.rzi.u32.f32 %r350, %f1324;
mul.wide.u32 %rd93, %r349, 2;
add.s64 %rd94, %rd92, %rd93;
mov.u16 %rs113, 0;
cvt.u16.u32 %rs114, %r350;
st.global.v2.u8 [%rd94], {%rs114, %rs113};
$L__BB0_161:
and.b32 %r351, %r50, 4;
setp.eq.s32 %p188, %r351, 0;
ld.const.u32 %r374, [params+108];
@%p188 bra $L__BB0_165;
setp.eq.s32 %p189, %r374, 0;
ld.const.u64 %rd95, [params+224];
cvta.to.global.u64 %rd96, %rd95;
ld.const.u32 %r352, [params+216];
mad.lo.s32 %r353, %r352, %r4, %r3;
mul.wide.u32 %rd97, %r353, 8;
add.s64 %rd24, %rd96, %rd97;
@%p189 bra $L__BB0_164;
ld.global.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd24];
// begin inline asm
{ cvt.f32.f16 %f1325, %rs121;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1326, %rs122;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1327, %rs123;}
// end inline asm
add.f32 %f1328, %f1325, 0f00000000;
add.f32 %f1329, %f1326, 0f00000000;
add.f32 %f1330, %f1327, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs120, %f1330;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs119, %f1329;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs118, %f1328;}
// end inline asm
mov.u16 %rs125, 0;
st.global.v4.u16 [%rd24], {%rs118, %rs119, %rs120, %rs125};
bra.uni $L__BB0_165;
$L__BB0_5:
ld.const.v4.f32 {%f241, %f242, %f243, %f244}, [params+512];
neg.f32 %f245, %f242;
neg.f32 %f246, %f243;
mul.f32 %f247, %f1428, %f241;
mul.f32 %f248, %f1429, %f245;
sub.f32 %f249, %f248, %f247;
mul.f32 %f250, %f1430, %f243;
sub.f32 %f13, %f249, %f250;
ld.const.u64 %rd32, [params+432];
cvta.to.global.u64 %rd33, %rd32;
ld.const.u32 %r65, [params+424];
mad.lo.s32 %r66, %r65, %r54, %r53;
mul.wide.u32 %rd34, %r66, 12;
add.s64 %rd35, %rd33, %rd34;
ld.global.f32 %f14, [%rd35];
mul.f32 %f251, %f14, 0f3456BF95;
ld.global.f32 %f15, [%rd35+4];
mul.f32 %f252, %f15, 0f3456BF95;
ld.global.f32 %f16, [%rd35+8];
mul.f32 %f253, %f16, 0f3456BF95;
abs.f32 %f254, %f1428;
div.rn.f32 %f255, %f251, %f254;
abs.f32 %f256, %f1429;
div.rn.f32 %f257, %f252, %f256;
abs.f32 %f258, %f1430;
div.rn.f32 %f259, %f253, %f258;
abs.f32 %f260, %f255;
abs.f32 %f261, %f257;
abs.f32 %f262, %f259;
mov.f32 %f263, 0f38D1B717;
max.f32 %f264, %f260, %f263;
max.f32 %f265, %f261, %f263;
max.f32 %f266, %f262, %f263;
fma.rn.f32 %f17, %f1428, %f264, %f14;
fma.rn.f32 %f18, %f1429, %f265, %f15;
fma.rn.f32 %f19, %f1430, %f266, %f16;
abs.f32 %f267, %f241;
abs.f32 %f268, %f243;
setp.gt.f32 %p16, %f267, %f268;
selp.f32 %f269, %f245, 0f00000000, %p16;
mov.f32 %f1436, 0f00000000;
selp.f32 %f270, %f241, %f246, %p16;
selp.f32 %f271, 0f00000000, %f242, %p16;
mul.f32 %f272, %f270, %f270;
fma.rn.f32 %f273, %f269, %f269, %f272;
fma.rn.f32 %f274, %f271, %f271, %f273;
sqrt.rn.f32 %f275, %f274;
rcp.rn.f32 %f276, %f275;
mul.f32 %f20, %f269, %f276;
mul.f32 %f21, %f270, %f276;
mul.f32 %f22, %f271, %f276;
mul.f32 %f277, %f243, %f21;
mul.f32 %f278, %f242, %f22;
sub.f32 %f23, %f277, %f278;
mul.f32 %f279, %f241, %f22;
mul.f32 %f280, %f243, %f20;
sub.f32 %f24, %f279, %f280;
mul.f32 %f281, %f242, %f20;
mul.f32 %f282, %f241, %f21;
sub.f32 %f25, %f281, %f282;
ld.const.u32 %r5, [params+540];
setp.lt.s32 %p17, %r5, 1;
mov.f32 %f1435, %f1436;
mov.f32 %f1434, %f1436;
@%p17 bra $L__BB0_34;
ld.const.u64 %rd36, [params+128];
ld.const.u32 %r68, [params+120];
cvt.rn.f32.s32 %f286, %r5;
rcp.rn.f32 %f26, %f286;
mad.lo.s32 %r69, %r68, %r54, %r53;
cvta.to.global.u64 %rd37, %rd36;
mul.wide.u32 %rd38, %r69, 4;
add.s64 %rd39, %rd37, %rd38;
ld.global.u32 %r366, [%rd39];
ld.const.f32 %f27, [params+536];
ld.const.u8 %rs24, [params+104];
and.b16 %rs25, %rs24, 32;
setp.eq.s16 %p18, %rs25, 0;
mov.u32 %r67, 0;
selp.f32 %f28, 0f3F800000, 0f41200000, %p18;
mul.f32 %f29, %f17, 0f3456BF95;
mul.f32 %f30, %f18, 0f3456BF95;
mul.f32 %f31, %f19, 0f3456BF95;
ld.const.u64 %rd3, [params+96];
add.s64 %rd4, %rd1, 24;
mov.u64 %rd40, __cudart_i2opi_f;
abs.f32 %f369, %f30;
abs.f32 %f370, %f29;
max.f32 %f371, %f370, %f369;
abs.f32 %f372, %f31;
max.f32 %f373, %f371, %f372;
mov.u32 %r363, %r67;
$L__BB0_7:
mov.u32 %r365, %r67;
$L__BB0_8:
cvt.rn.f32.s32 %f1376, %r363;
mad.lo.s32 %r71, %r366, 1664525, 1013904223;
and.b32 %r72, %r71, 16777215;
cvt.rn.f32.u32 %f287, %r72;
fma.rn.f32 %f288, %f287, 0f33800000, %f1376;
mul.f32 %f289, %f26, %f288;
mad.lo.s32 %r366, %r71, 1664525, 1013904223;
and.b32 %r73, %r366, 16777215;
cvt.rn.f32.u32 %f290, %r73;
cvt.rn.f32.s32 %f291, %r365;
fma.rn.f32 %f292, %f290, 0f33800000, %f291;
mul.f32 %f293, %f26, %f292;
sqrt.rn.f32 %f39, %f289;
mul.f32 %f40, %f293, 0f40C90FDB;
mul.f32 %f294, %f40, 0f3F22F983;
cvt.rni.s32.f32 %r373, %f294;
cvt.rn.f32.s32 %f295, %r373;
mov.f32 %f296, 0fBFC90FDA;
fma.rn.f32 %f297, %f295, %f296, %f40;
mov.f32 %f298, 0fB3A22168;
fma.rn.f32 %f299, %f295, %f298, %f297;
mov.f32 %f300, 0fA7C234C5;
fma.rn.f32 %f1440, %f295, %f300, %f299;
abs.f32 %f42, %f40;
setp.ltu.f32 %p19, %f42, 0f47CE4780;
mov.u32 %r370, %r373;
mov.f32 %f1437, %f1440;
@%p19 bra $L__BB0_16;
setp.eq.f32 %p20, %f42, 0f7F800000;
@%p20 bra $L__BB0_15;
bra.uni $L__BB0_10;
$L__BB0_15:
mov.f32 %f303, 0f00000000;
mul.rn.f32 %f1437, %f40, %f303;
mov.u32 %r370, 0;
bra.uni $L__BB0_16;
$L__BB0_10:
mov.b32 %r13, %f40;
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 %rd113, 0;
mov.u32 %r367, 0;
mov.u64 %rd111, %rd1;
mov.u64 %rd112, %rd40;
$L__BB0_11:
.pragma "nounroll";
ld.global.nc.u32 %r77, [%rd112];
mad.wide.u32 %rd42, %r77, %r15, %rd113;
shr.u64 %rd113, %rd42, 32;
st.local.u32 [%rd111], %rd42;
add.s64 %rd112, %rd112, 4;
add.s64 %rd111, %rd111, 4;
add.s32 %r367, %r367, 1;
setp.ne.s32 %p21, %r367, 6;
@%p21 bra $L__BB0_11;
st.local.u32 [%rd4], %rd113;
mov.u32 %r78, 4;
sub.s32 %r19, %r78, %r16;
mov.u32 %r79, 6;
sub.s32 %r80, %r79, %r16;
mul.wide.s32 %rd43, %r80, 4;
add.s64 %rd44, %rd1, %rd43;
ld.local.u32 %r368, [%rd44];
ld.local.u32 %r369, [%rd44+-4];
and.b32 %r22, %r14, 31;
setp.eq.s32 %p22, %r22, 0;
@%p22 bra $L__BB0_14;
mov.u32 %r81, 32;
sub.s32 %r82, %r81, %r22;
shr.u32 %r83, %r369, %r82;
shl.b32 %r84, %r368, %r22;
add.s32 %r368, %r83, %r84;
mul.wide.s32 %rd45, %r19, 4;
add.s64 %rd46, %rd1, %rd45;
ld.local.u32 %r85, [%rd46];
shr.u32 %r86, %r85, %r82;
shl.b32 %r87, %r369, %r22;
add.s32 %r369, %r86, %r87;
$L__BB0_14:
and.b32 %r88, %r13, -2147483648;
shr.u32 %r89, %r369, 30;
shl.b32 %r90, %r368, 2;
or.b32 %r91, %r89, %r90;
shr.u32 %r92, %r91, 31;
shr.u32 %r93, %r368, 30;
add.s32 %r94, %r92, %r93;
neg.s32 %r95, %r94;
setp.eq.s32 %p23, %r88, 0;
selp.b32 %r370, %r94, %r95, %p23;
setp.ne.s32 %p24, %r92, 0;
xor.b32 %r96, %r88, -2147483648;
selp.b32 %r97, %r96, %r88, %p24;
selp.b32 %r98, -1, 0, %p24;
xor.b32 %r99, %r91, %r98;
shl.b32 %r100, %r369, 2;
xor.b32 %r101, %r100, %r98;
cvt.u64.u32 %rd47, %r99;
cvt.u64.u32 %rd48, %r101;
bfi.b64 %rd49, %rd47, %rd48, 32, 32;
cvt.rn.f64.s64 %fd1, %rd49;
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f301, %fd2;
setp.eq.s32 %p25, %r97, 0;
neg.f32 %f302, %f301;
selp.f32 %f1437, %f301, %f302, %p25;
$L__BB0_16:
add.s32 %r29, %r370, 1;
and.b32 %r30, %r29, 1;
setp.eq.s32 %p26, %r30, 0;
selp.f32 %f46, %f1437, 0f3F800000, %p26;
mul.rn.f32 %f47, %f1437, %f1437;
mov.f32 %f1438, 0fB94D4153;
@%p26 bra $L__BB0_18;
mov.f32 %f305, 0fBAB607ED;
mov.f32 %f306, 0f37CBAC00;
fma.rn.f32 %f1438, %f306, %f47, %f305;
$L__BB0_18:
selp.f32 %f307, 0f3C0885E4, 0f3D2AAABB, %p26;
fma.rn.f32 %f308, %f1438, %f47, %f307;
selp.f32 %f309, 0fBE2AAAA8, 0fBEFFFFFF, %p26;
fma.rn.f32 %f310, %f308, %f47, %f309;
mov.f32 %f311, 0f00000000;
fma.rn.f32 %f312, %f47, %f46, %f311;
fma.rn.f32 %f1439, %f310, %f312, %f46;
and.b32 %r103, %r29, 2;
setp.eq.s32 %p28, %r103, 0;
@%p28 bra $L__BB0_20;
mov.f32 %f314, 0fBF800000;
fma.rn.f32 %f1439, %f1439, %f314, %f311;
$L__BB0_20:
@%p19 bra $L__BB0_28;
setp.eq.f32 %p30, %f42, 0f7F800000;
@%p30 bra $L__BB0_27;
bra.uni $L__BB0_22;
$L__BB0_27:
mov.f32 %f317, 0f00000000;
mul.rn.f32 %f1440, %f40, %f317;
mov.u32 %r373, 0;
bra.uni $L__BB0_28;
$L__BB0_22:
mov.b32 %r31, %f40;
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 %rd114, 0;
mov.u64 %rd115, %rd114;
$L__BB0_23:
.pragma "nounroll";
shl.b64 %rd52, %rd114, 2;
mov.u64 %rd53, __cudart_i2opi_f;
add.s64 %rd54, %rd53, %rd52;
ld.global.nc.u32 %r106, [%rd54];
mad.wide.u32 %rd55, %r106, %r33, %rd115;
shr.u64 %rd115, %rd55, 32;
add.s64 %rd56, %rd1, %rd52;
st.local.u32 [%rd56], %rd55;
cvt.u32.u64 %r107, %rd114;
add.s32 %r108, %r107, 1;
cvt.s64.s32 %rd114, %r108;
setp.ne.s32 %p31, %r108, 6;
@%p31 bra $L__BB0_23;
st.local.u32 [%rd4], %rd115;
mov.u32 %r109, 4;
sub.s32 %r35, %r109, %r34;
mov.u32 %r110, 6;
sub.s32 %r111, %r110, %r34;
mul.wide.s32 %rd57, %r111, 4;
add.s64 %rd58, %rd1, %rd57;
ld.local.u32 %r371, [%rd58];
ld.local.u32 %r372, [%rd58+-4];
and.b32 %r38, %r32, 31;
setp.eq.s32 %p32, %r38, 0;
@%p32 bra $L__BB0_26;
mov.u32 %r112, 32;
sub.s32 %r113, %r112, %r38;
shr.u32 %r114, %r372, %r113;
shl.b32 %r115, %r371, %r38;
add.s32 %r371, %r114, %r115;
mul.wide.s32 %rd59, %r35, 4;
add.s64 %rd60, %rd1, %rd59;
ld.local.u32 %r116, [%rd60];
shr.u32 %r117, %r116, %r113;
shl.b32 %r118, %r372, %r38;
add.s32 %r372, %r117, %r118;
$L__BB0_26:
and.b32 %r119, %r31, -2147483648;
shr.u32 %r120, %r372, 30;
shl.b32 %r121, %r371, 2;
or.b32 %r122, %r120, %r121;
shr.u32 %r123, %r122, 31;
shr.u32 %r124, %r371, 30;
add.s32 %r125, %r123, %r124;
neg.s32 %r126, %r125;
setp.eq.s32 %p33, %r119, 0;
selp.b32 %r373, %r125, %r126, %p33;
setp.ne.s32 %p34, %r123, 0;
xor.b32 %r127, %r119, -2147483648;
selp.b32 %r128, %r127, %r119, %p34;
selp.b32 %r129, -1, 0, %p34;
xor.b32 %r130, %r122, %r129;
shl.b32 %r131, %r372, 2;
xor.b32 %r132, %r131, %r129;
cvt.u64.u32 %rd61, %r130;
cvt.u64.u32 %rd62, %r132;
bfi.b64 %rd63, %rd61, %rd62, 32, 32;
cvt.rn.f64.s64 %fd3, %rd63;
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
cvt.rn.f32.f64 %f315, %fd4;
setp.eq.s32 %p35, %r128, 0;
neg.f32 %f316, %f315;
selp.f32 %f1440, %f315, %f316, %p35;
$L__BB0_28:
mul.f32 %f56, %f39, %f1439;
and.b32 %r45, %r373, 1;
setp.eq.s32 %p36, %r45, 0;
selp.f32 %f57, %f1440, 0f3F800000, %p36;
mul.rn.f32 %f58, %f1440, %f1440;
mov.f32 %f1441, 0fB94D4153;
@%p36 bra $L__BB0_30;
mov.f32 %f319, 0fBAB607ED;
mov.f32 %f320, 0f37CBAC00;
fma.rn.f32 %f1441, %f320, %f58, %f319;
$L__BB0_30:
selp.f32 %f321, 0f3C0885E4, 0f3D2AAABB, %p36;
fma.rn.f32 %f322, %f1441, %f58, %f321;
selp.f32 %f323, 0fBE2AAAA8, 0fBEFFFFFF, %p36;
fma.rn.f32 %f324, %f322, %f58, %f323;
mov.f32 %f325, 0f00000000;
fma.rn.f32 %f326, %f58, %f57, %f325;
fma.rn.f32 %f1442, %f324, %f326, %f57;
and.b32 %r134, %r373, 2;
setp.eq.s32 %p38, %r134, 0;
@%p38 bra $L__BB0_32;
mov.f32 %f328, 0fBF800000;
fma.rn.f32 %f1442, %f1442, %f328, %f325;
$L__BB0_32:
mul.f32 %f338, %f56, %f56;
mov.f32 %f339, 0f3F800000;
sub.f32 %f340, %f339, %f338;
mul.f32 %f341, %f39, %f1442;
mul.f32 %f342, %f341, %f341;
sub.f32 %f343, %f340, %f342;
max.f32 %f344, %f325, %f343;
sqrt.rn.f32 %f345, %f344;
mul.f32 %f346, %f20, %f341;
mul.f32 %f347, %f21, %f341;
mul.f32 %f348, %f22, %f341;
fma.rn.f32 %f349, %f23, %f56, %f346;
fma.rn.f32 %f350, %f24, %f56, %f347;
fma.rn.f32 %f351, %f25, %f56, %f348;
fma.rn.f32 %f352, %f241, %f345, %f349;
fma.rn.f32 %f353, %f242, %f345, %f350;
fma.rn.f32 %f354, %f243, %f345, %f351;
add.f32 %f355, %f241, %f352;
add.f32 %f356, %f242, %f353;
add.f32 %f357, %f243, %f354;
mul.f32 %f358, %f27, %f355;
mul.f32 %f359, %f27, %f356;
mul.f32 %f360, %f27, %f357;
sub.f32 %f361, %f358, %f241;
sub.f32 %f362, %f359, %f242;
sub.f32 %f363, %f360, %f243;
mul.f32 %f364, %f362, %f362;
fma.rn.f32 %f365, %f361, %f361, %f364;
fma.rn.f32 %f366, %f363, %f363, %f365;
sqrt.rn.f32 %f367, %f366;
rcp.rn.f32 %f368, %f367;
mul.f32 %f332, %f368, %f361;
mul.f32 %f333, %f368, %f362;
mul.f32 %f334, %f368, %f363;
mov.f32 %f374, 0f38D1B717;
max.f32 %f375, %f373, %f374;
mul.f32 %f335, %f28, %f375;
mov.f32 %f336, 0f6C4ECB8F;
mov.u32 %r171, 2;
mov.u32 %r172, 1;
mov.u32 %r173, 3;
mov.u32 %r176, 1065353216;
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,%rd3,%f17,%f18,%f19,%f332,%f333,%f334,%f335,%f336,%f325,%r172,%r205,%r172,%r171,%r172,%r173,%r176,%r176,%r176,%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 %f376, %r135;
mov.b32 %f377, %r136;
mov.b32 %f378, %r137;
add.f32 %f1436, %f1436, %f376;
add.f32 %f1435, %f1435, %f377;
add.f32 %f1434, %f1434, %f378;
add.s32 %r365, %r365, 1;
setp.lt.s32 %p39, %r365, %r5;
@%p39 bra $L__BB0_8;
add.s32 %r363, %r363, 1;
setp.lt.s32 %p40, %r363, %r5;
@%p40 bra $L__BB0_7;
$L__BB0_34:
setp.eq.s32 %p41, %r5, 0;
mov.f32 %f1446, 0f3F800000;
mov.f32 %f1447, %f1446;
mov.f32 %f1448, %f1446;
@%p41 bra $L__BB0_36;
mul.lo.s32 %r206, %r5, %r5;
cvt.rn.f32.s32 %f382, %r206;
rcp.rn.f32 %f383, %f382;
mul.f32 %f1446, %f1436, %f383;
mul.f32 %f1447, %f1435, %f383;
mul.f32 %f1448, %f1434, %f383;
$L__BB0_36:
mul.f32 %f385, %f15, %f24;
fma.rn.f32 %f386, %f14, %f23, %f385;
fma.rn.f32 %f387, %f16, %f25, %f386;
ld.const.v4.f32 {%f388, %f389, %f390, %f391}, [params+560];
fma.rn.f32 %f396, %f387, %f388, %f390;
mul.f32 %f397, %f15, %f21;
fma.rn.f32 %f398, %f14, %f20, %f397;
fma.rn.f32 %f399, %f16, %f22, %f398;
fma.rn.f32 %f400, %f399, %f389, %f391;
ld.const.u64 %rd65, [params+552];
tex.2d.v4.f32.f32 {%f401, %f402, %f403, %f404}, [%rd65, {%f396, %f400}];
min.f32 %f405, %f401, %f402;
min.f32 %f406, %f405, %f403;
mul.f32 %f76, %f1446, %f406;
ld.const.f32 %f407, [params+524];
mul.f32 %f408, %f1446, %f407;
ld.const.v2.f32 {%f409, %f410}, [params+528];
mul.f32 %f413, %f1447, %f409;
mul.f32 %f414, %f1448, %f410;
mul.f32 %f77, %f401, %f408;
mul.f32 %f78, %f402, %f413;
mul.f32 %f79, %f403, %f414;
cvt.sat.f32.f32 %f415, %f13;
mul.f32 %f80, %f77, %f415;
mul.f32 %f81, %f78, %f415;
mul.f32 %f82, %f415, %f79;
fma.rn.f32 %f416, %f13, 0f3F000000, 0f3F000000;
cvt.sat.f32.f32 %f417, %f416;
add.f32 %f83, %f417, %f417;
mov.f32 %f421, 0f41A00000;
abs.f32 %f85, %f83;
setp.lt.f32 %p42, %f85, 0f00800000;
mul.f32 %f423, %f85, 0f4B800000;
selp.f32 %f424, %f423, %f85, %p42;
selp.f32 %f425, 0fC3170000, 0fC2FE0000, %p42;
mov.b32 %r207, %f424;
and.b32 %r208, %r207, 8388607;
or.b32 %r209, %r208, 1065353216;
mov.b32 %f426, %r209;
shr.u32 %r210, %r207, 23;
cvt.rn.f32.u32 %f427, %r210;
add.f32 %f428, %f425, %f427;
setp.gt.f32 %p43, %f426, 0f3FB504F3;
mul.f32 %f429, %f426, 0f3F000000;
add.f32 %f430, %f428, 0f3F800000;
selp.f32 %f431, %f430, %f428, %p43;
selp.f32 %f432, %f429, %f426, %p43;
add.f32 %f433, %f432, 0fBF800000;
add.f32 %f434, %f432, 0f3F800000;
rcp.approx.ftz.f32 %f435, %f434;
add.f32 %f436, %f433, %f433;
mul.f32 %f437, %f436, %f435;
mul.f32 %f438, %f437, %f437;
mov.f32 %f439, 0f3C4CAF63;
mov.f32 %f440, 0f3B18F0FE;
fma.rn.f32 %f441, %f440, %f438, %f439;
mov.f32 %f442, 0f3DAAAABD;
fma.rn.f32 %f443, %f441, %f438, %f442;
mul.rn.f32 %f444, %f443, %f438;
mul.rn.f32 %f445, %f444, %f437;
sub.f32 %f446, %f433, %f437;
add.f32 %f447, %f446, %f446;
neg.f32 %f448, %f437;
fma.rn.f32 %f449, %f448, %f433, %f447;
mul.rn.f32 %f450, %f435, %f449;
add.f32 %f451, %f445, %f437;
sub.f32 %f452, %f437, %f451;
add.f32 %f453, %f445, %f452;
add.f32 %f454, %f450, %f453;
add.f32 %f455, %f451, %f454;
sub.f32 %f456, %f451, %f455;
add.f32 %f457, %f454, %f456;
mov.f32 %f458, 0f3F317200;
mul.rn.f32 %f459, %f431, %f458;
mov.f32 %f460, 0f35BFBE8E;
mul.rn.f32 %f461, %f431, %f460;
add.f32 %f462, %f459, %f455;
sub.f32 %f463, %f459, %f462;
add.f32 %f464, %f455, %f463;
add.f32 %f465, %f457, %f464;
add.f32 %f466, %f461, %f465;
add.f32 %f467, %f462, %f466;
sub.f32 %f468, %f462, %f467;
add.f32 %f469, %f466, %f468;
mul.rn.f32 %f470, %f421, %f467;
neg.f32 %f471, %f470;
fma.rn.f32 %f472, %f421, %f467, %f471;
fma.rn.f32 %f473, %f421, %f469, %f472;
mov.f32 %f474, 0f00000000;
fma.rn.f32 %f475, %f474, %f467, %f473;
add.rn.f32 %f476, %f470, %f475;
neg.f32 %f477, %f476;
add.rn.f32 %f478, %f470, %f477;
add.rn.f32 %f479, %f478, %f475;
mov.b32 %r211, %f476;
setp.eq.s32 %p44, %r211, 1118925336;
add.s32 %r212, %r211, -1;
mov.b32 %f480, %r212;
add.f32 %f481, %f479, 0f37000000;
selp.f32 %f86, %f481, %f479, %p44;
selp.f32 %f482, %f480, %f476, %p44;
mov.f32 %f483, 0f3FB8AA3B;
mul.rn.f32 %f484, %f482, %f483;
cvt.rzi.f32.f32 %f485, %f484;
abs.f32 %f486, %f485;
setp.gt.f32 %p45, %f486, 0f42FC0000;
mov.b32 %r213, %f485;
and.b32 %r214, %r213, -2147483648;
or.b32 %r215, %r214, 1123811328;
mov.b32 %f487, %r215;
selp.f32 %f488, %f487, %f485, %p45;
mov.f32 %f489, 0fBF317218;
fma.rn.f32 %f490, %f488, %f489, %f482;
mov.f32 %f491, 0f3102E308;
fma.rn.f32 %f492, %f488, %f491, %f490;
mul.f32 %f493, %f492, 0f3FB8AA3B;
add.f32 %f494, %f488, 0f4B40007F;
mov.b32 %r216, %f494;
shl.b32 %r217, %r216, 23;
mov.b32 %f495, %r217;
ex2.approx.ftz.f32 %f496, %f493;
mul.f32 %f87, %f496, %f495;
setp.eq.f32 %p46, %f87, 0f7F800000;
mov.f32 %f1449, 0f7F800000;
@%p46 bra $L__BB0_38;
fma.rn.f32 %f1449, %f87, %f86, %f87;
$L__BB0_38:
mov.f32 %f1382, 0f41200000;
cvt.rzi.f32.f32 %f1381, %f1382;
add.f32 %f1380, %f1381, %f1381;
mov.f32 %f1379, 0f41A00000;
sub.f32 %f1378, %f1379, %f1380;
abs.f32 %f1377, %f1378;
setp.lt.f32 %p47, %f83, 0f00000000;
setp.eq.f32 %p48, %f1377, 0f3F800000;
and.pred %p1, %p47, %p48;
setp.eq.f32 %p49, %f83, 0f00000000;
@%p49 bra $L__BB0_42;
bra.uni $L__BB0_39;
$L__BB0_42:
add.f32 %f501, %f83, %f83;
selp.f32 %f1451, %f501, 0f00000000, %p48;
bra.uni $L__BB0_43;
$L__BB0_39:
mov.b32 %r218, %f1449;
xor.b32 %r219, %r218, -2147483648;
mov.b32 %f497, %r219;
selp.f32 %f1451, %f497, %f1449, %p1;
setp.geu.f32 %p50, %f83, 0f00000000;
@%p50 bra $L__BB0_43;
mov.f32 %f498, 0f41A00000;
cvt.rzi.f32.f32 %f499, %f498;
setp.eq.f32 %p51, %f499, 0f41A00000;
@%p51 bra $L__BB0_43;
mov.f32 %f1451, 0f7FFFFFFF;
$L__BB0_43:
abs.f32 %f1383, %f83;
add.f32 %f502, %f1383, 0f41A00000;
mov.b32 %r220, %f502;
setp.lt.s32 %p53, %r220, 2139095040;
@%p53 bra $L__BB0_48;
abs.f32 %f1384, %f83;
setp.gtu.f32 %p54, %f1384, 0f7F800000;
@%p54 bra $L__BB0_47;
bra.uni $L__BB0_45;
$L__BB0_47:
add.f32 %f1451, %f83, 0f41A00000;
bra.uni $L__BB0_48;
$L__BB0_45:
abs.f32 %f1385, %f83;
setp.neu.f32 %p55, %f1385, 0f7F800000;
@%p55 bra $L__BB0_48;
selp.f32 %f1451, 0fFF800000, 0f7F800000, %p1;
$L__BB0_48:
setp.eq.f32 %p56, %f83, 0f3F800000;
selp.f32 %f503, 0f3F800000, %f1451, %p56;
cvt.sat.f32.f32 %f504, %f503;
mul.f32 %f96, %f77, %f504;
mul.f32 %f97, %f78, %f504;
mul.f32 %f98, %f79, %f504;
ld.const.u32 %r48, [params+104];
and.b32 %r221, %r48, 8;
setp.eq.s32 %p57, %r221, 0;
@%p57 bra $L__BB0_62;
ld.const.u32 %r222, [params+184];
mad.lo.s32 %r223, %r222, %r4, %r3;
cvt.u64.u32 %rd16, %r223;
mov.f32 %f509, 0f3EE8BA2E;
abs.f32 %f100, %f76;
setp.lt.f32 %p58, %f100, 0f00800000;
mul.f32 %f511, %f100, 0f4B800000;
selp.f32 %f512, %f511, %f100, %p58;
selp.f32 %f513, 0fC3170000, 0fC2FE0000, %p58;
mov.b32 %r224, %f512;
and.b32 %r225, %r224, 8388607;
or.b32 %r226, %r225, 1065353216;
mov.b32 %f514, %r226;
shr.u32 %r227, %r224, 23;
cvt.rn.f32.u32 %f515, %r227;
add.f32 %f516, %f513, %f515;
setp.gt.f32 %p59, %f514, 0f3FB504F3;
mul.f32 %f517, %f514, 0f3F000000;
add.f32 %f518, %f516, 0f3F800000;
selp.f32 %f519, %f518, %f516, %p59;
selp.f32 %f520, %f517, %f514, %p59;
add.f32 %f521, %f520, 0fBF800000;
add.f32 %f522, %f520, 0f3F800000;
rcp.approx.ftz.f32 %f523, %f522;
add.f32 %f524, %f521, %f521;
mul.f32 %f525, %f524, %f523;
mul.f32 %f526, %f525, %f525;
mov.f32 %f527, 0f3C4CAF63;
mov.f32 %f528, 0f3B18F0FE;
fma.rn.f32 %f529, %f528, %f526, %f527;
mov.f32 %f530, 0f3DAAAABD;
fma.rn.f32 %f531, %f529, %f526, %f530;
mul.rn.f32 %f532, %f531, %f526;
mul.rn.f32 %f533, %f532, %f525;
sub.f32 %f534, %f521, %f525;
add.f32 %f535, %f534, %f534;
neg.f32 %f536, %f525;
fma.rn.f32 %f537, %f536, %f521, %f535;
mul.rn.f32 %f538, %f523, %f537;
add.f32 %f539, %f533, %f525;
sub.f32 %f540, %f525, %f539;
add.f32 %f541, %f533, %f540;
add.f32 %f542, %f538, %f541;
add.f32 %f543, %f539, %f542;
sub.f32 %f544, %f539, %f543;
add.f32 %f545, %f542, %f544;
mov.f32 %f546, 0f3F317200;
mul.rn.f32 %f547, %f519, %f546;
mov.f32 %f548, 0f35BFBE8E;
mul.rn.f32 %f549, %f519, %f548;
add.f32 %f550, %f547, %f543;
sub.f32 %f551, %f547, %f550;
add.f32 %f552, %f543, %f551;
add.f32 %f553, %f545, %f552;
add.f32 %f554, %f549, %f553;
add.f32 %f555, %f550, %f554;
sub.f32 %f556, %f550, %f555;
add.f32 %f557, %f554, %f556;
mul.rn.f32 %f558, %f509, %f555;
neg.f32 %f559, %f558;
fma.rn.f32 %f560, %f509, %f555, %f559;
fma.rn.f32 %f561, %f509, %f557, %f560;
mov.f32 %f562, 0f00000000;
fma.rn.f32 %f563, %f562, %f555, %f561;
add.rn.f32 %f564, %f558, %f563;
neg.f32 %f565, %f564;
add.rn.f32 %f566, %f558, %f565;
add.rn.f32 %f567, %f566, %f563;
mov.b32 %r228, %f564;
setp.eq.s32 %p60, %r228, 1118925336;
add.s32 %r229, %r228, -1;
mov.b32 %f568, %r229;
add.f32 %f569, %f567, 0f37000000;
selp.f32 %f101, %f569, %f567, %p60;
selp.f32 %f570, %f568, %f564, %p60;
mov.f32 %f571, 0f3FB8AA3B;
mul.rn.f32 %f572, %f570, %f571;
cvt.rzi.f32.f32 %f573, %f572;
abs.f32 %f574, %f573;
setp.gt.f32 %p61, %f574, 0f42FC0000;
mov.b32 %r230, %f573;
and.b32 %r231, %r230, -2147483648;
or.b32 %r232, %r231, 1123811328;
mov.b32 %f575, %r232;
selp.f32 %f576, %f575, %f573, %p61;
mov.f32 %f577, 0fBF317218;
fma.rn.f32 %f578, %f576, %f577, %f570;
mov.f32 %f579, 0f3102E308;
fma.rn.f32 %f580, %f576, %f579, %f578;
mul.f32 %f581, %f580, 0f3FB8AA3B;
add.f32 %f582, %f576, 0f4B40007F;
mov.b32 %r233, %f582;
shl.b32 %r234, %r233, 23;
mov.b32 %f583, %r234;
ex2.approx.ftz.f32 %f584, %f581;
mul.f32 %f102, %f584, %f583;
setp.eq.f32 %p62, %f102, 0f7F800000;
mov.f32 %f1452, 0f7F800000;
@%p62 bra $L__BB0_51;
fma.rn.f32 %f1452, %f102, %f101, %f102;
$L__BB0_51:
mov.f32 %f1427, 0f3E68BA2E;
cvt.rzi.f32.f32 %f1426, %f1427;
add.f32 %f1425, %f1426, %f1426;
mov.f32 %f1424, 0f3EE8BA2E;
sub.f32 %f1423, %f1424, %f1425;
abs.f32 %f1422, %f1423;
setp.lt.f32 %p63, %f76, 0f00000000;
setp.eq.f32 %p64, %f1422, 0f3F800000;
and.pred %p2, %p63, %p64;
setp.eq.f32 %p65, %f76, 0f00000000;
@%p65 bra $L__BB0_55;
bra.uni $L__BB0_52;
$L__BB0_55:
add.f32 %f589, %f76, %f76;
selp.f32 %f1454, %f589, 0f00000000, %p64;
bra.uni $L__BB0_56;
$L__BB0_164:
mov.f32 %f1333, 0f00000000;
mov.u32 %r374, 0;
// begin inline asm
{ cvt.rn.f16.f32 %rs128, %f1333;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs127, %f1333;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs126, %f1333;}
// end inline asm
mov.u16 %rs129, 0;
st.global.v4.u16 [%rd24], {%rs126, %rs127, %rs128, %rs129};
$L__BB0_165:
ld.const.u64 %rd98, [params+256];
cvta.to.global.u64 %rd99, %rd98;
ld.const.u32 %r355, [params+248];
mad.lo.s32 %r356, %r355, %r4, %r3;
mul.wide.u32 %rd100, %r356, 8;
add.s64 %rd25, %rd99, %rd100;
setp.eq.s32 %p190, %r374, 0;
@%p190 bra $L__BB0_167;
ld.global.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd25];
// begin inline asm
{ cvt.f32.f16 %f1334, %rs136;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1335, %rs137;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1336, %rs138;}
// end inline asm
add.f32 %f1337, %f1334, 0f00000000;
add.f32 %f1338, %f1335, 0f00000000;
add.f32 %f1339, %f1336, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs135, %f1339;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs134, %f1338;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs133, %f1337;}
// end inline asm
mov.u16 %rs140, 0;
st.global.v4.u16 [%rd25], {%rs133, %rs134, %rs135, %rs140};
bra.uni $L__BB0_168;
$L__BB0_167:
mov.f32 %f1342, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs143, %f1342;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs142, %f1342;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs141, %f1342;}
// end inline asm
mov.u16 %rs144, 0;
st.global.v4.u16 [%rd25], {%rs141, %rs142, %rs143, %rs144};
$L__BB0_168:
ld.const.u64 %rd101, [params+272];
cvta.to.global.u64 %rd102, %rd101;
ld.const.u32 %r357, [params+264];
mad.lo.s32 %r358, %r357, %r4, %r3;
mul.wide.u32 %rd103, %r358, 8;
add.s64 %rd26, %rd102, %rd103;
@%p190 bra $L__BB0_170;
ld.global.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd26];
// begin inline asm
{ cvt.f32.f16 %f1343, %rs151;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1344, %rs152;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1345, %rs153;}
// end inline asm
add.f32 %f1346, %f1343, 0f00000000;
add.f32 %f1347, %f1344, 0f00000000;
add.f32 %f1348, %f1345, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs150, %f1348;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs149, %f1347;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs148, %f1346;}
// end inline asm
mov.u16 %rs155, 0;
st.global.v4.u16 [%rd26], {%rs148, %rs149, %rs150, %rs155};
bra.uni $L__BB0_171;
$L__BB0_170:
mov.f32 %f1351, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs158, %f1351;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs157, %f1351;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs156, %f1351;}
// end inline asm
mov.u16 %rs159, 0;
st.global.v4.u16 [%rd26], {%rs156, %rs157, %rs158, %rs159};
$L__BB0_171:
ld.const.u64 %rd104, [params+288];
cvta.to.global.u64 %rd105, %rd104;
ld.const.u32 %r359, [params+280];
mad.lo.s32 %r360, %r359, %r4, %r3;
mul.wide.u32 %rd106, %r360, 8;
add.s64 %rd27, %rd105, %rd106;
@%p190 bra $L__BB0_173;
ld.global.v4.u16 {%rs166, %rs167, %rs168, %rs169}, [%rd27];
// begin inline asm
{ cvt.f32.f16 %f1352, %rs166;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1353, %rs167;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1354, %rs168;}
// end inline asm
add.f32 %f1355, %f1352, 0f00000000;
add.f32 %f1356, %f1353, 0f00000000;
add.f32 %f1357, %f1354, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs165, %f1357;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs164, %f1356;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs163, %f1355;}
// end inline asm
mov.u16 %rs170, 0;
st.global.v4.u16 [%rd27], {%rs163, %rs164, %rs165, %rs170};
bra.uni $L__BB0_174;
$L__BB0_173:
mov.f32 %f1360, 0f00000000;
// begin inline asm
{ cvt.rn.f16.f32 %rs173, %f1360;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs172, %f1360;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs171, %f1360;}
// end inline asm
mov.u16 %rs174, 0;
st.global.v4.u16 [%rd27], {%rs171, %rs172, %rs173, %rs174};
bra.uni $L__BB0_174;
$L__BB0_52:
mov.b32 %r235, %f1452;
xor.b32 %r236, %r235, -2147483648;
mov.b32 %f585, %r236;
selp.f32 %f1454, %f585, %f1452, %p2;
setp.geu.f32 %p66, %f76, 0f00000000;
@%p66 bra $L__BB0_56;
mov.f32 %f586, 0f3EE8BA2E;
cvt.rzi.f32.f32 %f587, %f586;
setp.eq.f32 %p67, %f587, 0f3EE8BA2E;
@%p67 bra $L__BB0_56;
mov.f32 %f1454, 0f7FFFFFFF;
$L__BB0_56:
abs.f32 %f1386, %f76;
add.f32 %f590, %f1386, 0f3EE8BA2E;
mov.b32 %r237, %f590;
setp.lt.s32 %p69, %r237, 2139095040;
@%p69 bra $L__BB0_61;
abs.f32 %f1387, %f76;
setp.gtu.f32 %p70, %f1387, 0f7F800000;
@%p70 bra $L__BB0_60;
bra.uni $L__BB0_58;
$L__BB0_60:
add.f32 %f1454, %f76, 0f3EE8BA2E;
bra.uni $L__BB0_61;
$L__BB0_58:
abs.f32 %f1388, %f76;
setp.neu.f32 %p71, %f1388, 0f7F800000;
@%p71 bra $L__BB0_61;
selp.f32 %f1454, 0fFF800000, 0f7F800000, %p2;
$L__BB0_61:
ld.const.u64 %rd110, [params+192];
cvta.to.global.u64 %rd109, %rd110;
mul.f32 %f591, %f1454, 0f437F0000;
setp.eq.f32 %p72, %f76, 0f3F800000;
selp.f32 %f592, 0f437F0000, %f591, %p72;
cvt.rzi.u32.f32 %r238, %f592;
shl.b64 %rd67, %rd16, 1;
add.s64 %rd68, %rd109, %rd67;
cvt.u16.u32 %rs26, %r238;
mov.u16 %rs27, 255;
st.global.v2.u8 [%rd68], {%rs26, %rs27};
$L__BB0_62:
ld.const.u32 %r362, [params+104];
and.b32 %r239, %r362, 1;
setp.eq.b32 %p73, %r239, 1;
mov.pred %p74, 0;
xor.pred %p75, %p73, %p74;
not.pred %p76, %p75;
@%p76 bra $L__BB0_136;
mov.f32 %f597, 0f3EE66666;
abs.f32 %f112, %f80;
setp.lt.f32 %p77, %f112, 0f00800000;
mul.f32 %f599, %f112, 0f4B800000;
selp.f32 %f600, %f599, %f112, %p77;
selp.f32 %f601, 0fC3170000, 0fC2FE0000, %p77;
mov.b32 %r240, %f600;
and.b32 %r241, %r240, 8388607;
or.b32 %r242, %r241, 1065353216;
mov.b32 %f602, %r242;
shr.u32 %r243, %r240, 23;
cvt.rn.f32.u32 %f603, %r243;
add.f32 %f604, %f601, %f603;
setp.gt.f32 %p78, %f602, 0f3FB504F3;
mul.f32 %f605, %f602, 0f3F000000;
add.f32 %f606, %f604, 0f3F800000;
selp.f32 %f607, %f606, %f604, %p78;
selp.f32 %f608, %f605, %f602, %p78;
add.f32 %f609, %f608, 0fBF800000;
add.f32 %f610, %f608, 0f3F800000;
rcp.approx.ftz.f32 %f611, %f610;
add.f32 %f612, %f609, %f609;
mul.f32 %f613, %f612, %f611;
mul.f32 %f614, %f613, %f613;
mov.f32 %f615, 0f3C4CAF63;
mov.f32 %f616, 0f3B18F0FE;
fma.rn.f32 %f617, %f616, %f614, %f615;
mov.f32 %f618, 0f3DAAAABD;
fma.rn.f32 %f619, %f617, %f614, %f618;
mul.rn.f32 %f620, %f619, %f614;
mul.rn.f32 %f621, %f620, %f613;
sub.f32 %f622, %f609, %f613;
add.f32 %f623, %f622, %f622;
neg.f32 %f624, %f613;
fma.rn.f32 %f625, %f624, %f609, %f623;
mul.rn.f32 %f626, %f611, %f625;
add.f32 %f627, %f621, %f613;
sub.f32 %f628, %f613, %f627;
add.f32 %f629, %f621, %f628;
add.f32 %f630, %f626, %f629;
add.f32 %f631, %f627, %f630;
sub.f32 %f632, %f627, %f631;
add.f32 %f633, %f630, %f632;
mov.f32 %f634, 0f3F317200;
mul.rn.f32 %f635, %f607, %f634;
mov.f32 %f636, 0f35BFBE8E;
mul.rn.f32 %f637, %f607, %f636;
add.f32 %f638, %f635, %f631;
sub.f32 %f639, %f635, %f638;
add.f32 %f640, %f631, %f639;
add.f32 %f641, %f633, %f640;
add.f32 %f642, %f637, %f641;
add.f32 %f643, %f638, %f642;
sub.f32 %f644, %f638, %f643;
add.f32 %f645, %f642, %f644;
mul.rn.f32 %f646, %f597, %f643;
neg.f32 %f647, %f646;
fma.rn.f32 %f648, %f597, %f643, %f647;
fma.rn.f32 %f649, %f597, %f645, %f648;
mov.f32 %f650, 0f00000000;
fma.rn.f32 %f651, %f650, %f643, %f649;
add.rn.f32 %f652, %f646, %f651;
neg.f32 %f653, %f652;
add.rn.f32 %f654, %f646, %f653;
add.rn.f32 %f655, %f654, %f651;
mov.b32 %r244, %f652;
setp.eq.s32 %p79, %r244, 1118925336;
add.s32 %r245, %r244, -1;
mov.b32 %f656, %r245;
add.f32 %f657, %f655, 0f37000000;
selp.f32 %f113, %f657, %f655, %p79;
selp.f32 %f658, %f656, %f652, %p79;
mov.f32 %f659, 0f3FB8AA3B;
mul.rn.f32 %f660, %f658, %f659;
cvt.rzi.f32.f32 %f661, %f660;
abs.f32 %f662, %f661;
setp.gt.f32 %p80, %f662, 0f42FC0000;
mov.b32 %r246, %f661;
and.b32 %r247, %r246, -2147483648;
or.b32 %r248, %r247, 1123811328;
mov.b32 %f663, %r248;
selp.f32 %f664, %f663, %f661, %p80;
mov.f32 %f665, 0fBF317218;
fma.rn.f32 %f666, %f664, %f665, %f658;
mov.f32 %f667, 0f3102E308;
fma.rn.f32 %f668, %f664, %f667, %f666;
mul.f32 %f669, %f668, 0f3FB8AA3B;
add.f32 %f670, %f664, 0f4B40007F;
mov.b32 %r249, %f670;
shl.b32 %r250, %r249, 23;
mov.b32 %f671, %r250;
ex2.approx.ftz.f32 %f672, %f669;
mul.f32 %f114, %f672, %f671;
setp.eq.f32 %p81, %f114, 0f7F800000;
mov.f32 %f1455, 0f7F800000;
@%p81 bra $L__BB0_65;
fma.rn.f32 %f1455, %f114, %f113, %f114;
$L__BB0_65:
mov.f32 %f1394, 0f3E666666;
cvt.rzi.f32.f32 %f1393, %f1394;
add.f32 %f1392, %f1393, %f1393;
mov.f32 %f1391, 0f3EE66666;
sub.f32 %f1390, %f1391, %f1392;
abs.f32 %f1389, %f1390;
setp.lt.f32 %p82, %f80, 0f00000000;
setp.eq.f32 %p83, %f1389, 0f3F800000;
and.pred %p3, %p82, %p83;
setp.eq.f32 %p84, %f80, 0f00000000;
@%p84 bra $L__BB0_69;
bra.uni $L__BB0_66;
$L__BB0_69:
add.f32 %f677, %f80, %f80;
selp.f32 %f1457, %f677, 0f00000000, %p83;
bra.uni $L__BB0_70;
$L__BB0_66:
mov.b32 %r251, %f1455;
xor.b32 %r252, %r251, -2147483648;
mov.b32 %f673, %r252;
selp.f32 %f1457, %f673, %f1455, %p3;
setp.geu.f32 %p85, %f80, 0f00000000;
@%p85 bra $L__BB0_70;
mov.f32 %f674, 0f3EE66666;
cvt.rzi.f32.f32 %f675, %f674;
setp.eq.f32 %p86, %f675, 0f3EE66666;
@%p86 bra $L__BB0_70;
mov.f32 %f1457, 0f7FFFFFFF;
$L__BB0_70:
abs.f32 %f1395, %f80;
add.f32 %f678, %f1395, 0f3EE66666;
mov.b32 %r253, %f678;
setp.lt.s32 %p88, %r253, 2139095040;
@%p88 bra $L__BB0_75;
abs.f32 %f1396, %f80;
setp.gtu.f32 %p89, %f1396, 0f7F800000;
@%p89 bra $L__BB0_74;
bra.uni $L__BB0_72;
$L__BB0_74:
add.f32 %f1457, %f80, 0f3EE66666;
bra.uni $L__BB0_75;
$L__BB0_72:
abs.f32 %f1397, %f80;
setp.neu.f32 %p90, %f1397, 0f7F800000;
@%p90 bra $L__BB0_75;
selp.f32 %f1457, 0fFF800000, 0f7F800000, %p3;
$L__BB0_75:
setp.eq.f32 %p91, %f80, 0f3F800000;
selp.f32 %f123, 0f3F800000, %f1457, %p91;
abs.f32 %f124, %f81;
setp.lt.f32 %p92, %f124, 0f00800000;
mul.f32 %f680, %f124, 0f4B800000;
selp.f32 %f681, %f680, %f124, %p92;
selp.f32 %f682, 0fC3170000, 0fC2FE0000, %p92;
mov.b32 %r254, %f681;
and.b32 %r255, %r254, 8388607;
or.b32 %r256, %r255, 1065353216;
mov.b32 %f683, %r256;
shr.u32 %r257, %r254, 23;
cvt.rn.f32.u32 %f684, %r257;
add.f32 %f685, %f682, %f684;
setp.gt.f32 %p93, %f683, 0f3FB504F3;
mul.f32 %f686, %f683, 0f3F000000;
add.f32 %f687, %f685, 0f3F800000;
selp.f32 %f688, %f687, %f685, %p93;
selp.f32 %f689, %f686, %f683, %p93;
add.f32 %f690, %f689, 0fBF800000;
add.f32 %f691, %f689, 0f3F800000;
rcp.approx.ftz.f32 %f692, %f691;
add.f32 %f693, %f690, %f690;
mul.f32 %f694, %f693, %f692;
mul.f32 %f695, %f694, %f694;
mov.f32 %f696, 0f3C4CAF63;
mov.f32 %f697, 0f3B18F0FE;
fma.rn.f32 %f698, %f697, %f695, %f696;
mov.f32 %f699, 0f3DAAAABD;
fma.rn.f32 %f700, %f698, %f695, %f699;
mul.rn.f32 %f701, %f700, %f695;
mul.rn.f32 %f702, %f701, %f694;
sub.f32 %f703, %f690, %f694;
add.f32 %f704, %f703, %f703;
neg.f32 %f705, %f694;
fma.rn.f32 %f706, %f705, %f690, %f704;
mul.rn.f32 %f707, %f692, %f706;
add.f32 %f708, %f702, %f694;
sub.f32 %f709, %f694, %f708;
add.f32 %f710, %f702, %f709;
add.f32 %f711, %f707, %f710;
add.f32 %f712, %f708, %f711;
sub.f32 %f713, %f708, %f712;
add.f32 %f714, %f711, %f713;
mov.f32 %f715, 0f3F317200;
mul.rn.f32 %f716, %f688, %f715;
mov.f32 %f717, 0f35BFBE8E;
mul.rn.f32 %f718, %f688, %f717;
add.f32 %f719, %f716, %f712;
sub.f32 %f720, %f716, %f719;
add.f32 %f721, %f712, %f720;
add.f32 %f722, %f714, %f721;
add.f32 %f723, %f718, %f722;
add.f32 %f724, %f719, %f723;
sub.f32 %f725, %f719, %f724;
add.f32 %f726, %f723, %f725;
mov.f32 %f727, 0f3EE66666;
mul.rn.f32 %f728, %f727, %f724;
neg.f32 %f729, %f728;
fma.rn.f32 %f730, %f727, %f724, %f729;
fma.rn.f32 %f731, %f727, %f726, %f730;
mov.f32 %f732, 0f00000000;
fma.rn.f32 %f733, %f732, %f724, %f731;
add.rn.f32 %f734, %f728, %f733;
neg.f32 %f735, %f734;
add.rn.f32 %f736, %f728, %f735;
add.rn.f32 %f737, %f736, %f733;
mov.b32 %r258, %f734;
setp.eq.s32 %p94, %r258, 1118925336;
add.s32 %r259, %r258, -1;
mov.b32 %f738, %r259;
add.f32 %f739, %f737, 0f37000000;
selp.f32 %f125, %f739, %f737, %p94;
selp.f32 %f740, %f738, %f734, %p94;
mov.f32 %f741, 0f3FB8AA3B;
mul.rn.f32 %f742, %f740, %f741;
cvt.rzi.f32.f32 %f743, %f742;
abs.f32 %f744, %f743;
setp.gt.f32 %p95, %f744, 0f42FC0000;
mov.b32 %r260, %f743;
and.b32 %r261, %r260, -2147483648;
or.b32 %r262, %r261, 1123811328;
mov.b32 %f745, %r262;
selp.f32 %f746, %f745, %f743, %p95;
mov.f32 %f747, 0fBF317218;
fma.rn.f32 %f748, %f746, %f747, %f740;
mov.f32 %f749, 0f3102E308;
fma.rn.f32 %f750, %f746, %f749, %f748;
mul.f32 %f751, %f750, 0f3FB8AA3B;
add.f32 %f752, %f746, 0f4B40007F;
mov.b32 %r263, %f752;
shl.b32 %r264, %r263, 23;
mov.b32 %f753, %r264;
ex2.approx.ftz.f32 %f754, %f751;
mul.f32 %f126, %f754, %f753;
setp.eq.f32 %p96, %f126, 0f7F800000;
mov.f32 %f1458, 0f7F800000;
@%p96 bra $L__BB0_77;
fma.rn.f32 %f1458, %f126, %f125, %f126;
$L__BB0_77:
setp.lt.f32 %p97, %f81, 0f00000000;
and.pred %p4, %p97, %p83;
setp.eq.f32 %p99, %f81, 0f00000000;
@%p99 bra $L__BB0_81;
bra.uni $L__BB0_78;
$L__BB0_81:
add.f32 %f759, %f81, %f81;
selp.f32 %f1460, %f759, 0f00000000, %p83;
bra.uni $L__BB0_82;
$L__BB0_78:
mov.b32 %r265, %f1458;
xor.b32 %r266, %r265, -2147483648;
mov.b32 %f755, %r266;
selp.f32 %f1460, %f755, %f1458, %p4;
setp.geu.f32 %p100, %f81, 0f00000000;
@%p100 bra $L__BB0_82;
mov.f32 %f756, 0f3EE66666;
cvt.rzi.f32.f32 %f757, %f756;
setp.eq.f32 %p101, %f757, 0f3EE66666;
@%p101 bra $L__BB0_82;
mov.f32 %f1460, 0f7FFFFFFF;
$L__BB0_82:
abs.f32 %f1398, %f81;
add.f32 %f760, %f1398, 0f3EE66666;
mov.b32 %r267, %f760;
setp.lt.s32 %p103, %r267, 2139095040;
@%p103 bra $L__BB0_87;
abs.f32 %f1399, %f81;
setp.gtu.f32 %p104, %f1399, 0f7F800000;
@%p104 bra $L__BB0_86;
bra.uni $L__BB0_84;
$L__BB0_86:
add.f32 %f1460, %f81, 0f3EE66666;
bra.uni $L__BB0_87;
$L__BB0_84:
abs.f32 %f1400, %f81;
setp.neu.f32 %p105, %f1400, 0f7F800000;
@%p105 bra $L__BB0_87;
selp.f32 %f1460, 0fFF800000, 0f7F800000, %p4;
$L__BB0_87:
setp.eq.f32 %p106, %f81, 0f3F800000;
selp.f32 %f135, 0f3F800000, %f1460, %p106;
abs.f32 %f136, %f82;
setp.lt.f32 %p107, %f136, 0f00800000;
mul.f32 %f762, %f136, 0f4B800000;
selp.f32 %f763, %f762, %f136, %p107;
selp.f32 %f764, 0fC3170000, 0fC2FE0000, %p107;
mov.b32 %r268, %f763;
and.b32 %r269, %r268, 8388607;
or.b32 %r270, %r269, 1065353216;
mov.b32 %f765, %r270;
shr.u32 %r271, %r268, 23;
cvt.rn.f32.u32 %f766, %r271;
add.f32 %f767, %f764, %f766;
setp.gt.f32 %p108, %f765, 0f3FB504F3;
mul.f32 %f768, %f765, 0f3F000000;
add.f32 %f769, %f767, 0f3F800000;
selp.f32 %f770, %f769, %f767, %p108;
selp.f32 %f771, %f768, %f765, %p108;
add.f32 %f772, %f771, 0fBF800000;
add.f32 %f773, %f771, 0f3F800000;
rcp.approx.ftz.f32 %f774, %f773;
add.f32 %f775, %f772, %f772;
mul.f32 %f776, %f775, %f774;
mul.f32 %f777, %f776, %f776;
mov.f32 %f778, 0f3C4CAF63;
mov.f32 %f779, 0f3B18F0FE;
fma.rn.f32 %f780, %f779, %f777, %f778;
mov.f32 %f781, 0f3DAAAABD;
fma.rn.f32 %f782, %f780, %f777, %f781;
mul.rn.f32 %f783, %f782, %f777;
mul.rn.f32 %f784, %f783, %f776;
sub.f32 %f785, %f772, %f776;
add.f32 %f786, %f785, %f785;
neg.f32 %f787, %f776;
fma.rn.f32 %f788, %f787, %f772, %f786;
mul.rn.f32 %f789, %f774, %f788;
add.f32 %f790, %f784, %f776;
sub.f32 %f791, %f776, %f790;
add.f32 %f792, %f784, %f791;
add.f32 %f793, %f789, %f792;
add.f32 %f794, %f790, %f793;
sub.f32 %f795, %f790, %f794;
add.f32 %f796, %f793, %f795;
mov.f32 %f797, 0f3F317200;
mul.rn.f32 %f798, %f770, %f797;
mov.f32 %f799, 0f35BFBE8E;
mul.rn.f32 %f800, %f770, %f799;
add.f32 %f801, %f798, %f794;
sub.f32 %f802, %f798, %f801;
add.f32 %f803, %f794, %f802;
add.f32 %f804, %f796, %f803;
add.f32 %f805, %f800, %f804;
add.f32 %f806, %f801, %f805;
sub.f32 %f807, %f801, %f806;
add.f32 %f808, %f805, %f807;
mov.f32 %f809, 0f3EE66666;
mul.rn.f32 %f810, %f809, %f806;
neg.f32 %f811, %f810;
fma.rn.f32 %f812, %f809, %f806, %f811;
fma.rn.f32 %f813, %f809, %f808, %f812;
mov.f32 %f814, 0f00000000;
fma.rn.f32 %f815, %f814, %f806, %f813;
add.rn.f32 %f816, %f810, %f815;
neg.f32 %f817, %f816;
add.rn.f32 %f818, %f810, %f817;
add.rn.f32 %f819, %f818, %f815;
mov.b32 %r272, %f816;
setp.eq.s32 %p109, %r272, 1118925336;
add.s32 %r273, %r272, -1;
mov.b32 %f820, %r273;
add.f32 %f821, %f819, 0f37000000;
selp.f32 %f137, %f821, %f819, %p109;
selp.f32 %f822, %f820, %f816, %p109;
mov.f32 %f823, 0f3FB8AA3B;
mul.rn.f32 %f824, %f822, %f823;
cvt.rzi.f32.f32 %f825, %f824;
abs.f32 %f826, %f825;
setp.gt.f32 %p110, %f826, 0f42FC0000;
mov.b32 %r274, %f825;
and.b32 %r275, %r274, -2147483648;
or.b32 %r276, %r275, 1123811328;
mov.b32 %f827, %r276;
selp.f32 %f828, %f827, %f825, %p110;
mov.f32 %f829, 0fBF317218;
fma.rn.f32 %f830, %f828, %f829, %f822;
mov.f32 %f831, 0f3102E308;
fma.rn.f32 %f832, %f828, %f831, %f830;
mul.f32 %f833, %f832, 0f3FB8AA3B;
add.f32 %f834, %f828, 0f4B40007F;
mov.b32 %r277, %f834;
shl.b32 %r278, %r277, 23;
mov.b32 %f835, %r278;
ex2.approx.ftz.f32 %f836, %f833;
mul.f32 %f138, %f836, %f835;
setp.eq.f32 %p111, %f138, 0f7F800000;
mov.f32 %f1461, 0f7F800000;
@%p111 bra $L__BB0_89;
fma.rn.f32 %f1461, %f138, %f137, %f138;
$L__BB0_89:
setp.lt.f32 %p112, %f82, 0f00000000;
and.pred %p5, %p112, %p83;
setp.eq.f32 %p114, %f82, 0f00000000;
@%p114 bra $L__BB0_93;
bra.uni $L__BB0_90;
$L__BB0_93:
add.f32 %f841, %f82, %f82;
selp.f32 %f1463, %f841, 0f00000000, %p83;
bra.uni $L__BB0_94;
$L__BB0_90:
mov.b32 %r279, %f1461;
xor.b32 %r280, %r279, -2147483648;
mov.b32 %f837, %r280;
selp.f32 %f1463, %f837, %f1461, %p5;
setp.geu.f32 %p115, %f82, 0f00000000;
@%p115 bra $L__BB0_94;
mov.f32 %f838, 0f3EE66666;
cvt.rzi.f32.f32 %f839, %f838;
setp.eq.f32 %p116, %f839, 0f3EE66666;
@%p116 bra $L__BB0_94;
mov.f32 %f1463, 0f7FFFFFFF;
$L__BB0_94:
abs.f32 %f1401, %f82;
add.f32 %f842, %f1401, 0f3EE66666;
mov.b32 %r281, %f842;
setp.lt.s32 %p118, %r281, 2139095040;
@%p118 bra $L__BB0_99;
abs.f32 %f1402, %f82;
setp.gtu.f32 %p119, %f1402, 0f7F800000;
@%p119 bra $L__BB0_98;
bra.uni $L__BB0_96;
$L__BB0_98:
add.f32 %f1463, %f82, 0f3EE66666;
bra.uni $L__BB0_99;
$L__BB0_96:
abs.f32 %f1403, %f82;
setp.neu.f32 %p120, %f1403, 0f7F800000;
@%p120 bra $L__BB0_99;
selp.f32 %f1463, 0fFF800000, 0f7F800000, %p5;
$L__BB0_99:
setp.eq.f32 %p121, %f82, 0f3F800000;
mov.f32 %f844, 0f3F800000;
selp.f32 %f845, 0f3F800000, %f1463, %p121;
ld.const.u32 %r282, [params+136];
mad.lo.s32 %r283, %r282, %r4, %r3;
cvt.u64.u32 %rd18, %r283;
min.f32 %f846, %f123, %f844;
mov.f32 %f847, 0f00000000;
max.f32 %f147, %f847, %f846;
min.f32 %f848, %f135, %f844;
max.f32 %f148, %f847, %f848;
min.f32 %f849, %f845, %f844;
max.f32 %f149, %f847, %f849;
mov.f32 %f853, 0f3ED55555;
abs.f32 %f151, %f147;
setp.lt.f32 %p122, %f151, 0f00800000;
mul.f32 %f855, %f151, 0f4B800000;
selp.f32 %f856, %f855, %f151, %p122;
selp.f32 %f857, 0fC3170000, 0fC2FE0000, %p122;
mov.b32 %r284, %f856;
and.b32 %r285, %r284, 8388607;
or.b32 %r286, %r285, 1065353216;
mov.b32 %f858, %r286;
shr.u32 %r287, %r284, 23;
cvt.rn.f32.u32 %f859, %r287;
add.f32 %f860, %f857, %f859;
setp.gt.f32 %p123, %f858, 0f3FB504F3;
mul.f32 %f861, %f858, 0f3F000000;
add.f32 %f862, %f860, 0f3F800000;
selp.f32 %f863, %f862, %f860, %p123;
selp.f32 %f864, %f861, %f858, %p123;
add.f32 %f865, %f864, 0fBF800000;
add.f32 %f866, %f864, 0f3F800000;
rcp.approx.ftz.f32 %f867, %f866;
add.f32 %f868, %f865, %f865;
mul.f32 %f869, %f868, %f867;
mul.f32 %f870, %f869, %f869;
mov.f32 %f871, 0f3C4CAF63;
mov.f32 %f872, 0f3B18F0FE;
fma.rn.f32 %f873, %f872, %f870, %f871;
mov.f32 %f874, 0f3DAAAABD;
fma.rn.f32 %f875, %f873, %f870, %f874;
mul.rn.f32 %f876, %f875, %f870;
mul.rn.f32 %f877, %f876, %f869;
sub.f32 %f878, %f865, %f869;
add.f32 %f879, %f878, %f878;
neg.f32 %f880, %f869;
fma.rn.f32 %f881, %f880, %f865, %f879;
mul.rn.f32 %f882, %f867, %f881;
add.f32 %f883, %f877, %f869;
sub.f32 %f884, %f869, %f883;
add.f32 %f885, %f877, %f884;
add.f32 %f886, %f882, %f885;
add.f32 %f887, %f883, %f886;
sub.f32 %f888, %f883, %f887;
add.f32 %f889, %f886, %f888;
mov.f32 %f890, 0f3F317200;
mul.rn.f32 %f891, %f863, %f890;
mov.f32 %f892, 0f35BFBE8E;
mul.rn.f32 %f893, %f863, %f892;
add.f32 %f894, %f891, %f887;
sub.f32 %f895, %f891, %f894;
add.f32 %f896, %f887, %f895;
add.f32 %f897, %f889, %f896;
add.f32 %f898, %f893, %f897;
add.f32 %f899, %f894, %f898;
sub.f32 %f900, %f894, %f899;
add.f32 %f901, %f898, %f900;
mul.rn.f32 %f902, %f853, %f899;
neg.f32 %f903, %f902;
fma.rn.f32 %f904, %f853, %f899, %f903;
fma.rn.f32 %f905, %f853, %f901, %f904;
fma.rn.f32 %f906, %f847, %f899, %f905;
add.rn.f32 %f907, %f902, %f906;
neg.f32 %f908, %f907;
add.rn.f32 %f909, %f902, %f908;
add.rn.f32 %f910, %f909, %f906;
mov.b32 %r288, %f907;
setp.eq.s32 %p124, %r288, 1118925336;
add.s32 %r289, %r288, -1;
mov.b32 %f911, %r289;
add.f32 %f912, %f910, 0f37000000;
selp.f32 %f152, %f912, %f910, %p124;
selp.f32 %f913, %f911, %f907, %p124;
mov.f32 %f914, 0f3FB8AA3B;
mul.rn.f32 %f915, %f913, %f914;
cvt.rzi.f32.f32 %f916, %f915;
abs.f32 %f917, %f916;
setp.gt.f32 %p125, %f917, 0f42FC0000;
mov.b32 %r290, %f916;
and.b32 %r291, %r290, -2147483648;
or.b32 %r292, %r291, 1123811328;
mov.b32 %f918, %r292;
selp.f32 %f919, %f918, %f916, %p125;
mov.f32 %f920, 0fBF317218;
fma.rn.f32 %f921, %f919, %f920, %f913;
mov.f32 %f922, 0f3102E308;
fma.rn.f32 %f923, %f919, %f922, %f921;
mul.f32 %f924, %f923, 0f3FB8AA3B;
add.f32 %f925, %f919, 0f4B40007F;
mov.b32 %r293, %f925;
shl.b32 %r294, %r293, 23;
mov.b32 %f926, %r294;
ex2.approx.ftz.f32 %f927, %f924;
mul.f32 %f153, %f927, %f926;
setp.eq.f32 %p126, %f153, 0f7F800000;
mov.f32 %f1464, 0f7F800000;
@%p126 bra $L__BB0_101;
fma.rn.f32 %f1464, %f153, %f152, %f153;
$L__BB0_101:
mov.f32 %f1366, 0f3E555555;
cvt.rzi.f32.f32 %f1365, %f1366;
add.f32 %f1364, %f1365, %f1365;
mov.f32 %f1363, 0f3ED55555;
sub.f32 %f1362, %f1363, %f1364;
abs.f32 %f1361, %f1362;
setp.lt.f32 %p127, %f147, 0f00000000;
setp.eq.f32 %p128, %f1361, 0f3F800000;
and.pred %p6, %p127, %p128;
setp.eq.f32 %p129, %f147, 0f00000000;
@%p129 bra $L__BB0_105;
bra.uni $L__BB0_102;
$L__BB0_105:
add.f32 %f932, %f147, %f147;
selp.f32 %f1466, %f932, 0f00000000, %p128;
bra.uni $L__BB0_106;
$L__BB0_102:
mov.b32 %r295, %f1464;
xor.b32 %r296, %r295, -2147483648;
mov.b32 %f928, %r296;
selp.f32 %f1466, %f928, %f1464, %p6;
setp.geu.f32 %p130, %f147, 0f00000000;
@%p130 bra $L__BB0_106;
mov.f32 %f929, 0f3ED55555;
cvt.rzi.f32.f32 %f930, %f929;
setp.eq.f32 %p131, %f930, 0f3ED55555;
@%p131 bra $L__BB0_106;
mov.f32 %f1466, 0f7FFFFFFF;
$L__BB0_106:
mov.f32 %f1369, 0f00000000;
max.f32 %f1368, %f1369, %f846;
abs.f32 %f1367, %f1368;
add.f32 %f933, %f1367, 0f3ED55555;
mov.b32 %r297, %f933;
setp.lt.s32 %p133, %r297, 2139095040;
@%p133 bra $L__BB0_111;
mov.f32 %f1372, 0f00000000;
max.f32 %f1371, %f1372, %f846;
abs.f32 %f1370, %f1371;
setp.gtu.f32 %p134, %f1370, 0f7F800000;
@%p134 bra $L__BB0_110;
bra.uni $L__BB0_108;
$L__BB0_110:
add.f32 %f1466, %f147, 0f3ED55555;
bra.uni $L__BB0_111;
$L__BB0_108:
mov.f32 %f1375, 0f00000000;
max.f32 %f1374, %f1375, %f846;
abs.f32 %f1373, %f1374;
setp.neu.f32 %p135, %f1373, 0f7F800000;
@%p135 bra $L__BB0_111;
selp.f32 %f1466, 0fFF800000, 0f7F800000, %p6;
$L__BB0_111:
abs.f32 %f162, %f148;
setp.lt.f32 %p136, %f162, 0f00800000;
mul.f32 %f935, %f162, 0f4B800000;
selp.f32 %f936, %f935, %f162, %p136;
selp.f32 %f937, 0fC3170000, 0fC2FE0000, %p136;
mov.b32 %r298, %f936;
and.b32 %r299, %r298, 8388607;
or.b32 %r300, %r299, 1065353216;
mov.b32 %f938, %r300;
shr.u32 %r301, %r298, 23;
cvt.rn.f32.u32 %f939, %r301;
add.f32 %f940, %f937, %f939;
setp.gt.f32 %p137, %f938, 0f3FB504F3;
mul.f32 %f941, %f938, 0f3F000000;
add.f32 %f942, %f940, 0f3F800000;
selp.f32 %f943, %f942, %f940, %p137;
selp.f32 %f944, %f941, %f938, %p137;
add.f32 %f945, %f944, 0fBF800000;
add.f32 %f946, %f944, 0f3F800000;
rcp.approx.ftz.f32 %f947, %f946;
add.f32 %f948, %f945, %f945;
mul.f32 %f949, %f948, %f947;
mul.f32 %f950, %f949, %f949;
mov.f32 %f951, 0f3C4CAF63;
mov.f32 %f952, 0f3B18F0FE;
fma.rn.f32 %f953, %f952, %f950, %f951;
mov.f32 %f954, 0f3DAAAABD;
fma.rn.f32 %f955, %f953, %f950, %f954;
mul.rn.f32 %f956, %f955, %f950;
mul.rn.f32 %f957, %f956, %f949;
sub.f32 %f958, %f945, %f949;
add.f32 %f959, %f958, %f958;
neg.f32 %f960, %f949;
fma.rn.f32 %f961, %f960, %f945, %f959;
mul.rn.f32 %f962, %f947, %f961;
add.f32 %f963, %f957, %f949;
sub.f32 %f964, %f949, %f963;
add.f32 %f965, %f957, %f964;
add.f32 %f966, %f962, %f965;
add.f32 %f967, %f963, %f966;
sub.f32 %f968, %f963, %f967;
add.f32 %f969, %f966, %f968;
mov.f32 %f970, 0f3F317200;
mul.rn.f32 %f971, %f943, %f970;
mov.f32 %f972, 0f35BFBE8E;
mul.rn.f32 %f973, %f943, %f972;
add.f32 %f974, %f971, %f967;
sub.f32 %f975, %f971, %f974;
add.f32 %f976, %f967, %f975;
add.f32 %f977, %f969, %f976;
add.f32 %f978, %f973, %f977;
add.f32 %f979, %f974, %f978;
sub.f32 %f980, %f974, %f979;
add.f32 %f981, %f978, %f980;
mov.f32 %f982, 0f3ED55555;
mul.rn.f32 %f983, %f982, %f979;
neg.f32 %f984, %f983;
fma.rn.f32 %f985, %f982, %f979, %f984;
fma.rn.f32 %f986, %f982, %f981, %f985;
mov.f32 %f987, 0f00000000;
fma.rn.f32 %f988, %f987, %f979, %f986;
add.rn.f32 %f989, %f983, %f988;
neg.f32 %f990, %f989;
add.rn.f32 %f991, %f983, %f990;
add.rn.f32 %f992, %f991, %f988;
mov.b32 %r302, %f989;
setp.eq.s32 %p138, %r302, 1118925336;
add.s32 %r303, %r302, -1;
mov.b32 %f993, %r303;
add.f32 %f994, %f992, 0f37000000;
selp.f32 %f163, %f994, %f992, %p138;
selp.f32 %f995, %f993, %f989, %p138;
mov.f32 %f996, 0f3FB8AA3B;
mul.rn.f32 %f997, %f995, %f996;
cvt.rzi.f32.f32 %f998, %f997;
abs.f32 %f999, %f998;
setp.gt.f32 %p139, %f999, 0f42FC0000;
mov.b32 %r304, %f998;
and.b32 %r305, %r304, -2147483648;
or.b32 %r306, %r305, 1123811328;
mov.b32 %f1000, %r306;
selp.f32 %f1001, %f1000, %f998, %p139;
mov.f32 %f1002, 0fBF317218;
fma.rn.f32 %f1003, %f1001, %f1002, %f995;
mov.f32 %f1004, 0f3102E308;
fma.rn.f32 %f1005, %f1001, %f1004, %f1003;
mul.f32 %f1006, %f1005, 0f3FB8AA3B;
add.f32 %f1007, %f1001, 0f4B40007F;
mov.b32 %r307, %f1007;
shl.b32 %r308, %r307, 23;
mov.b32 %f1008, %r308;
ex2.approx.ftz.f32 %f1009, %f1006;
mul.f32 %f164, %f1009, %f1008;
setp.eq.f32 %p140, %f164, 0f7F800000;
mov.f32 %f1467, 0f7F800000;
@%p140 bra $L__BB0_113;
fma.rn.f32 %f1467, %f164, %f163, %f164;
$L__BB0_113:
setp.lt.f32 %p141, %f148, 0f00000000;
and.pred %p7, %p141, %p128;
setp.eq.f32 %p143, %f148, 0f00000000;
@%p143 bra $L__BB0_117;
bra.uni $L__BB0_114;
$L__BB0_117:
add.f32 %f1014, %f148, %f148;
selp.f32 %f1469, %f1014, 0f00000000, %p128;
bra.uni $L__BB0_118;
$L__BB0_114:
mov.b32 %r309, %f1467;
xor.b32 %r310, %r309, -2147483648;
mov.b32 %f1010, %r310;
selp.f32 %f1469, %f1010, %f1467, %p7;
setp.geu.f32 %p144, %f148, 0f00000000;
@%p144 bra $L__BB0_118;
mov.f32 %f1011, 0f3ED55555;
cvt.rzi.f32.f32 %f1012, %f1011;
setp.eq.f32 %p145, %f1012, 0f3ED55555;
@%p145 bra $L__BB0_118;
mov.f32 %f1469, 0f7FFFFFFF;
$L__BB0_118:
mov.f32 %f1406, 0f00000000;
max.f32 %f1405, %f1406, %f848;
abs.f32 %f1404, %f1405;
add.f32 %f1015, %f1404, 0f3ED55555;
mov.b32 %r311, %f1015;
setp.lt.s32 %p147, %r311, 2139095040;
@%p147 bra $L__BB0_123;
mov.f32 %f1409, 0f00000000;
max.f32 %f1408, %f1409, %f848;
abs.f32 %f1407, %f1408;
setp.gtu.f32 %p148, %f1407, 0f7F800000;
@%p148 bra $L__BB0_122;
bra.uni $L__BB0_120;
$L__BB0_122:
add.f32 %f1469, %f148, 0f3ED55555;
bra.uni $L__BB0_123;
$L__BB0_120:
mov.f32 %f1412, 0f00000000;
max.f32 %f1411, %f1412, %f848;
abs.f32 %f1410, %f1411;
setp.neu.f32 %p149, %f1410, 0f7F800000;
@%p149 bra $L__BB0_123;
selp.f32 %f1469, 0fFF800000, 0f7F800000, %p7;
$L__BB0_123:
abs.f32 %f173, %f149;
setp.lt.f32 %p150, %f173, 0f00800000;
mul.f32 %f1017, %f173, 0f4B800000;
selp.f32 %f1018, %f1017, %f173, %p150;
selp.f32 %f1019, 0fC3170000, 0fC2FE0000, %p150;
mov.b32 %r312, %f1018;
and.b32 %r313, %r312, 8388607;
or.b32 %r314, %r313, 1065353216;
mov.b32 %f1020, %r314;
shr.u32 %r315, %r312, 23;
cvt.rn.f32.u32 %f1021, %r315;
add.f32 %f1022, %f1019, %f1021;
setp.gt.f32 %p151, %f1020, 0f3FB504F3;
mul.f32 %f1023, %f1020, 0f3F000000;
add.f32 %f1024, %f1022, 0f3F800000;
selp.f32 %f1025, %f1024, %f1022, %p151;
selp.f32 %f1026, %f1023, %f1020, %p151;
add.f32 %f1027, %f1026, 0fBF800000;
add.f32 %f1028, %f1026, 0f3F800000;
rcp.approx.ftz.f32 %f1029, %f1028;
add.f32 %f1030, %f1027, %f1027;
mul.f32 %f1031, %f1030, %f1029;
mul.f32 %f1032, %f1031, %f1031;
mov.f32 %f1033, 0f3C4CAF63;
mov.f32 %f1034, 0f3B18F0FE;
fma.rn.f32 %f1035, %f1034, %f1032, %f1033;
mov.f32 %f1036, 0f3DAAAABD;
fma.rn.f32 %f1037, %f1035, %f1032, %f1036;
mul.rn.f32 %f1038, %f1037, %f1032;
mul.rn.f32 %f1039, %f1038, %f1031;
sub.f32 %f1040, %f1027, %f1031;
add.f32 %f1041, %f1040, %f1040;
neg.f32 %f1042, %f1031;
fma.rn.f32 %f1043, %f1042, %f1027, %f1041;
mul.rn.f32 %f1044, %f1029, %f1043;
add.f32 %f1045, %f1039, %f1031;
sub.f32 %f1046, %f1031, %f1045;
add.f32 %f1047, %f1039, %f1046;
add.f32 %f1048, %f1044, %f1047;
add.f32 %f1049, %f1045, %f1048;
sub.f32 %f1050, %f1045, %f1049;
add.f32 %f1051, %f1048, %f1050;
mov.f32 %f1052, 0f3F317200;
mul.rn.f32 %f1053, %f1025, %f1052;
mov.f32 %f1054, 0f35BFBE8E;
mul.rn.f32 %f1055, %f1025, %f1054;
add.f32 %f1056, %f1053, %f1049;
sub.f32 %f1057, %f1053, %f1056;
add.f32 %f1058, %f1049, %f1057;
add.f32 %f1059, %f1051, %f1058;
add.f32 %f1060, %f1055, %f1059;
add.f32 %f1061, %f1056, %f1060;
sub.f32 %f1062, %f1056, %f1061;
add.f32 %f1063, %f1060, %f1062;
mov.f32 %f1064, 0f3ED55555;
mul.rn.f32 %f1065, %f1064, %f1061;
neg.f32 %f1066, %f1065;
fma.rn.f32 %f1067, %f1064, %f1061, %f1066;
fma.rn.f32 %f1068, %f1064, %f1063, %f1067;
mov.f32 %f1069, 0f00000000;
fma.rn.f32 %f1070, %f1069, %f1061, %f1068;
add.rn.f32 %f1071, %f1065, %f1070;
neg.f32 %f1072, %f1071;
add.rn.f32 %f1073, %f1065, %f1072;
add.rn.f32 %f1074, %f1073, %f1070;
mov.b32 %r316, %f1071;
setp.eq.s32 %p152, %r316, 1118925336;
add.s32 %r317, %r316, -1;
mov.b32 %f1075, %r317;
add.f32 %f1076, %f1074, 0f37000000;
selp.f32 %f174, %f1076, %f1074, %p152;
selp.f32 %f1077, %f1075, %f1071, %p152;
mov.f32 %f1078, 0f3FB8AA3B;
mul.rn.f32 %f1079, %f1077, %f1078;
cvt.rzi.f32.f32 %f1080, %f1079;
abs.f32 %f1081, %f1080;
setp.gt.f32 %p153, %f1081, 0f42FC0000;
mov.b32 %r318, %f1080;
and.b32 %r319, %r318, -2147483648;
or.b32 %r320, %r319, 1123811328;
mov.b32 %f1082, %r320;
selp.f32 %f1083, %f1082, %f1080, %p153;
mov.f32 %f1084, 0fBF317218;
fma.rn.f32 %f1085, %f1083, %f1084, %f1077;
mov.f32 %f1086, 0f3102E308;
fma.rn.f32 %f1087, %f1083, %f1086, %f1085;
mul.f32 %f1088, %f1087, 0f3FB8AA3B;
add.f32 %f1089, %f1083, 0f4B40007F;
mov.b32 %r321, %f1089;
shl.b32 %r322, %r321, 23;
mov.b32 %f1090, %r322;
ex2.approx.ftz.f32 %f1091, %f1088;
mul.f32 %f175, %f1091, %f1090;
setp.eq.f32 %p154, %f175, 0f7F800000;
mov.f32 %f1470, 0f7F800000;
@%p154 bra $L__BB0_125;
fma.rn.f32 %f1470, %f175, %f174, %f175;
$L__BB0_125:
setp.lt.f32 %p155, %f149, 0f00000000;
and.pred %p8, %p155, %p128;
setp.eq.f32 %p157, %f149, 0f00000000;
@%p157 bra $L__BB0_129;
bra.uni $L__BB0_126;
$L__BB0_129:
add.f32 %f1096, %f149, %f149;
selp.f32 %f1472, %f1096, 0f00000000, %p128;
bra.uni $L__BB0_130;
$L__BB0_126:
mov.b32 %r323, %f1470;
xor.b32 %r324, %r323, -2147483648;
mov.b32 %f1092, %r324;
selp.f32 %f1472, %f1092, %f1470, %p8;
setp.geu.f32 %p158, %f149, 0f00000000;
@%p158 bra $L__BB0_130;
mov.f32 %f1093, 0f3ED55555;
cvt.rzi.f32.f32 %f1094, %f1093;
setp.eq.f32 %p159, %f1094, 0f3ED55555;
@%p159 bra $L__BB0_130;
mov.f32 %f1472, 0f7FFFFFFF;
$L__BB0_130:
mov.f32 %f1415, 0f00000000;
max.f32 %f1414, %f1415, %f849;
abs.f32 %f1413, %f1414;
add.f32 %f1097, %f1413, 0f3ED55555;
mov.b32 %r325, %f1097;
setp.lt.s32 %p161, %r325, 2139095040;
@%p161 bra $L__BB0_135;
mov.f32 %f1418, 0f00000000;
max.f32 %f1417, %f1418, %f849;
abs.f32 %f1416, %f1417;
setp.gtu.f32 %p162, %f1416, 0f7F800000;
@%p162 bra $L__BB0_134;
bra.uni $L__BB0_132;
$L__BB0_134:
add.f32 %f1472, %f149, 0f3ED55555;
bra.uni $L__BB0_135;
$L__BB0_132:
mov.f32 %f1421, 0f00000000;
max.f32 %f1420, %f1421, %f849;
abs.f32 %f1419, %f1420;
setp.neu.f32 %p163, %f1419, 0f7F800000;
@%p163 bra $L__BB0_135;
selp.f32 %f1472, 0fFF800000, 0f7F800000, %p8;
$L__BB0_135:
ld.const.u64 %rd108, [params+144];
cvta.to.global.u64 %rd107, %rd108;
fma.rn.f32 %f1098, %f1466, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p164, %f147, 0f3F800000;
mov.f32 %f1099, 0f3F800000;
selp.f32 %f1100, 0f3F7FFFFF, %f1098, %p164;
mul.f32 %f1101, %f147, 0f414EB852;
setp.lt.f32 %p165, %f147, 0f3B4D2E1C;
selp.f32 %f1102, %f1101, %f1100, %p165;
fma.rn.f32 %f1103, %f1469, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p166, %f148, 0f3F800000;
selp.f32 %f1104, 0f3F7FFFFF, %f1103, %p166;
mul.f32 %f1105, %f148, 0f414EB852;
setp.lt.f32 %p167, %f148, 0f3B4D2E1C;
selp.f32 %f1106, %f1105, %f1104, %p167;
fma.rn.f32 %f1107, %f1472, 0f3F870A3D, 0fBD6147AE;
setp.eq.f32 %p168, %f149, 0f3F800000;
selp.f32 %f1108, 0f3F7FFFFF, %f1107, %p168;
mul.f32 %f1109, %f149, 0f414EB852;
setp.lt.f32 %p169, %f149, 0f3B4D2E1C;
selp.f32 %f1110, %f1109, %f1108, %p169;
min.f32 %f1111, %f1102, %f1099;
mov.f32 %f1112, 0f00000000;
max.f32 %f1113, %f1112, %f1111;
mul.f32 %f1114, %f1113, 0f43800000;
cvt.rzi.u32.f32 %r326, %f1114;
min.u32 %r327, %r326, 255;
min.f32 %f1115, %f1106, %f1099;
max.f32 %f1116, %f1112, %f1115;
mul.f32 %f1117, %f1116, 0f43800000;
cvt.rzi.u32.f32 %r328, %f1117;
min.u32 %r329, %r328, 255;
min.f32 %f1118, %f1110, %f1099;
max.f32 %f1119, %f1112, %f1118;
mul.f32 %f1120, %f1119, 0f43800000;
cvt.rzi.u32.f32 %r330, %f1120;
min.u32 %r331, %r330, 255;
shl.b64 %rd70, %rd18, 2;
add.s64 %rd71, %rd107, %rd70;
cvt.u16.u32 %rs28, %r331;
cvt.u16.u32 %rs29, %r329;
cvt.u16.u32 %rs30, %r327;
mov.u16 %rs31, 255;
st.global.v4.u8 [%rd71], {%rs30, %rs29, %rs28, %rs31};
$L__BB0_136:
ld.const.u32 %r361, [params+104];
and.b32 %r332, %r361, 4;
setp.eq.s32 %p170, %r332, 0;
@%p170 bra $L__BB0_140;
ld.const.u32 %r333, [params+108];
setp.eq.s32 %p171, %r333, 0;
ld.const.u64 %rd72, [params+224];
cvta.to.global.u64 %rd73, %rd72;
ld.const.u32 %r334, [params+216];
mad.lo.s32 %r335, %r334, %r4, %r3;
mul.wide.u32 %rd74, %r335, 8;
add.s64 %rd19, %rd73, %rd74;
@%p171 bra $L__BB0_139;
ld.global.v4.u16 {%rs39, %rs40, %rs41, %rs42}, [%rd19];
// begin inline asm
{ cvt.f32.f16 %f1121, %rs39;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1122, %rs40;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1123, %rs41;}
// end inline asm
add.f32 %f1124, %f80, %f1121;
add.f32 %f1125, %f81, %f1122;
add.f32 %f1126, %f82, %f1123;
mov.f32 %f1127, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs37, %f1126;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs36, %f1125;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs35, %f1124;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs38, %f1127;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs35, %rs36, %rs37, %rs38};
bra.uni $L__BB0_140;
$L__BB0_139:
mov.f32 %f1131, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs46, %f1131;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs45, %f82;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs44, %f81;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs43, %f80;}
// end inline asm
st.global.v4.u16 [%rd19], {%rs43, %rs44, %rs45, %rs46};
$L__BB0_140:
ld.const.u64 %rd75, [params+464];
cvta.to.global.u64 %rd76, %rd75;
ld.const.u32 %r336, [params+456];
mad.lo.s32 %r337, %r336, %r54, %r53;
mul.wide.u32 %rd77, %r337, 4;
add.s64 %rd20, %rd76, %rd77;
ld.global.v2.u8 {%rs47, %rs179}, [%rd20];
or.b16 %rs49, %rs47, %rs179;
and.b16 %rs50, %rs49, 255;
setp.eq.s16 %p172, %rs50, 0;
@%p172 bra $L__BB0_142;
ld.global.u8 %rs177, [%rd20+2];
bra.uni $L__BB0_143;
$L__BB0_142:
ld.global.u8 %rs177, [%rd20+2];
setp.eq.s16 %p173, %rs177, 0;
mov.f32 %f1473, 0f00000000;
mov.u16 %rs179, 0;
mov.u16 %rs180, %rs179;
mov.f32 %f1474, %f1473;
mov.f32 %f1475, %f1473;
@%p173 bra $L__BB0_144;
$L__BB0_143:
mov.u16 %rs180, %rs179;
cvt.rn.f32.u16 %f1135, %rs47;
div.rn.f32 %f1136, %f1135, 0f437F0000;
fma.rn.f32 %f1137, %f1136, 0f40000000, 0fBF800000;
and.b16 %rs55, %rs180, 255;
cvt.rn.f32.u16 %f1138, %rs55;
div.rn.f32 %f1139, %f1138, 0f437F0000;
fma.rn.f32 %f1140, %f1139, 0f40000000, 0fBF800000;
cvt.rn.f32.u16 %f1141, %rs177;
div.rn.f32 %f1142, %f1141, 0f437F0000;
fma.rn.f32 %f1143, %f1142, 0f40000000, 0fBF800000;
mul.f32 %f1144, %f1140, %f1140;
fma.rn.f32 %f1145, %f1137, %f1137, %f1144;
fma.rn.f32 %f1146, %f1143, %f1143, %f1145;
sqrt.rn.f32 %f1147, %f1146;
rcp.rn.f32 %f1148, %f1147;
mul.f32 %f1475, %f1148, %f1143;
mul.f32 %f1474, %f1148, %f1140;
mul.f32 %f1473, %f1137, %f1148;
mov.u16 %rs179, %rs177;
$L__BB0_144:
mul.f32 %f1152, %f1429, %f1475;
mul.f32 %f1153, %f1430, %f1474;
sub.f32 %f1154, %f1152, %f1153;
mul.f32 %f1155, %f1430, %f1473;
mul.f32 %f1156, %f1428, %f1475;
sub.f32 %f1157, %f1155, %f1156;
mul.f32 %f1158, %f1428, %f1474;
mul.f32 %f1159, %f1429, %f1473;
sub.f32 %f1160, %f1158, %f1159;
ld.global.u8 %rs57, [%rd20+3];
setp.eq.s16 %p174, %rs57, 0;
selp.f32 %f190, 0fBF800000, 0f3F800000, %p174;
mul.f32 %f1161, %f1154, %f190;
mul.f32 %f1162, %f1157, %f190;
mul.f32 %f1163, %f1160, %f190;
mul.f32 %f1164, %f1161, 0f00000000;
mul.f32 %f1165, %f1162, 0f00000000;
mul.f32 %f1166, %f1163, 0f00000000;
fma.rn.f32 %f1167, %f1473, 0f3F5105EC, %f1164;
fma.rn.f32 %f1168, %f1474, 0f3F5105EC, %f1165;
fma.rn.f32 %f1169, %f1475, 0f3F5105EC, %f1166;
mul.f32 %f191, %f1428, 0f3F13CD3A;
add.f32 %f192, %f191, %f1167;
mul.f32 %f193, %f1429, 0f3F13CD3A;
add.f32 %f194, %f193, %f1168;
mul.f32 %f195, %f1430, 0f3F13CD3A;
add.f32 %f196, %f195, %f1169;
or.b16 %rs58, %rs47, %rs180;
or.b16 %rs15, %rs58, %rs179;
and.b16 %rs59, %rs15, 255;
setp.eq.s16 %p175, %rs59, 0;
mov.f32 %f1479, 0f00000000;
mov.f32 %f1476, %f1479;
mov.f32 %f1477, %f1479;
mov.f32 %f1478, %f1479;
@%p175 bra $L__BB0_146;
cvt.rn.f32.u16 %f1170, %rs47;
div.rn.f32 %f1171, %f1170, 0f437F0000;
fma.rn.f32 %f1172, %f1171, 0f40000000, 0fBF800000;
and.b16 %rs61, %rs180, 255;
cvt.rn.f32.u16 %f1173, %rs61;
div.rn.f32 %f1174, %f1173, 0f437F0000;
fma.rn.f32 %f1175, %f1174, 0f40000000, 0fBF800000;
and.b16 %rs62, %rs179, 255;
cvt.rn.f32.u16 %f1176, %rs62;
div.rn.f32 %f1177, %f1176, 0f437F0000;
fma.rn.f32 %f1178, %f1177, 0f40000000, 0fBF800000;
mul.f32 %f1179, %f1175, %f1175;
fma.rn.f32 %f1180, %f1172, %f1172, %f1179;
fma.rn.f32 %f1181, %f1178, %f1178, %f1180;
sqrt.rn.f32 %f1182, %f1181;
rcp.rn.f32 %f1183, %f1182;
mul.f32 %f1478, %f1183, %f1178;
mul.f32 %f1477, %f1183, %f1175;
mul.f32 %f1476, %f1172, %f1183;
$L__BB0_146:
mul.f32 %f1187, %f1429, %f1478;
mul.f32 %f1188, %f1430, %f1477;
sub.f32 %f1189, %f1187, %f1188;
mul.f32 %f1190, %f1430, %f1476;
mul.f32 %f1191, %f1428, %f1478;
sub.f32 %f1192, %f1190, %f1191;
mul.f32 %f1193, %f1428, %f1477;
mul.f32 %f1194, %f1429, %f1476;
sub.f32 %f1195, %f1193, %f1194;
mul.f32 %f1196, %f1189, %f190;
mul.f32 %f1197, %f1192, %f190;
mul.f32 %f1198, %f1195, %f190;
mul.f32 %f1199, %f1476, 0f3ED105EC;
mul.f32 %f1200, %f1477, 0f3ED105EC;
mul.f32 %f1201, %f1478, 0f3ED105EC;
mul.f32 %f1202, %f1196, 0f3F3504F3;
mul.f32 %f1203, %f1197, 0f3F3504F3;
mul.f32 %f1204, %f1198, 0f3F3504F3;
sub.f32 %f1205, %f1202, %f1199;
sub.f32 %f1206, %f1203, %f1200;
sub.f32 %f1207, %f1204, %f1201;
add.f32 %f203, %f191, %f1205;
add.f32 %f204, %f193, %f1206;
add.f32 %f205, %f195, %f1207;
mov.f32 %f1480, %f1479;
mov.f32 %f1481, %f1479;
@%p175 bra $L__BB0_148;
cvt.rn.f32.u16 %f1208, %rs47;
div.rn.f32 %f1209, %f1208, 0f437F0000;
fma.rn.f32 %f1210, %f1209, 0f40000000, 0fBF800000;
and.b16 %rs65, %rs180, 255;
cvt.rn.f32.u16 %f1211, %rs65;
div.rn.f32 %f1212, %f1211, 0f437F0000;
fma.rn.f32 %f1213, %f1212, 0f40000000, 0fBF800000;
and.b16 %rs66, %rs179, 255;
cvt.rn.f32.u16 %f1214, %rs66;
div.rn.f32 %f1215, %f1214, 0f437F0000;
fma.rn.f32 %f1216, %f1215, 0f40000000, 0fBF800000;
mul.f32 %f1217, %f1213, %f1213;
fma.rn.f32 %f1218, %f1210, %f1210, %f1217;
fma.rn.f32 %f1219, %f1216, %f1216, %f1218;
sqrt.rn.f32 %f1220, %f1219;
rcp.rn.f32 %f1221, %f1220;
mul.f32 %f1481, %f1221, %f1216;
mul.f32 %f1480, %f1221, %f1213;
mul.f32 %f1479, %f1210, %f1221;
$L__BB0_148:
mul.f32 %f1222, %f1429, %f1481;
mul.f32 %f1223, %f1430, %f1480;
sub.f32 %f1224, %f1222, %f1223;
mul.f32 %f1225, %f1430, %f1479;
mul.f32 %f1226, %f1428, %f1481;
sub.f32 %f1227, %f1225, %f1226;
mul.f32 %f1228, %f1428, %f1480;
mul.f32 %f1229, %f1429, %f1479;
sub.f32 %f1230, %f1228, %f1229;
mul.f32 %f1231, %f1224, %f190;
mul.f32 %f1232, %f1227, %f190;
mul.f32 %f1233, %f1230, %f190;
mul.f32 %f1234, %f1479, 0f3ED105EC;
mul.f32 %f1235, %f1480, 0f3ED105EC;
mul.f32 %f1236, %f1481, 0f3ED105EC;
mul.f32 %f1237, %f1231, 0fBF3504F3;
mul.f32 %f1238, %f1232, 0fBF3504F3;
mul.f32 %f1239, %f1233, 0fBF3504F3;
sub.f32 %f1240, %f1237, %f1234;
sub.f32 %f1241, %f1238, %f1235;
sub.f32 %f1242, %f1239, %f1236;
add.f32 %f1243, %f191, %f1240;
add.f32 %f1244, %f193, %f1241;
add.f32 %f1245, %f195, %f1242;
mul.f32 %f1246, %f194, %f242;
neg.f32 %f1247, %f1246;
mul.f32 %f1248, %f192, %f241;
sub.f32 %f1249, %f1247, %f1248;
mul.f32 %f1250, %f196, %f243;
sub.f32 %f1251, %f1249, %f1250;
cvt.sat.f32.f32 %f1252, %f1251;
mul.f32 %f1253, %f96, %f1252;
mul.f32 %f1254, %f97, %f1252;
mul.f32 %f1255, %f98, %f1252;
mul.f32 %f1256, %f204, %f242;
neg.f32 %f1257, %f1256;
mul.f32 %f1258, %f203, %f241;
sub.f32 %f1259, %f1257, %f1258;
mul.f32 %f1260, %f205, %f243;
sub.f32 %f1261, %f1259, %f1260;
cvt.sat.f32.f32 %f1262, %f1261;
mul.f32 %f1263, %f96, %f1262;
mul.f32 %f1264, %f97, %f1262;
mul.f32 %f1265, %f98, %f1262;
mul.f32 %f1266, %f1243, %f241;
mul.f32 %f1267, %f1244, %f242;
neg.f32 %f1268, %f1267;
sub.f32 %f1269, %f1268, %f1266;
mul.f32 %f1270, %f1245, %f243;
sub.f32 %f1271, %f1269, %f1270;
cvt.sat.f32.f32 %f1272, %f1271;
mul.f32 %f1273, %f96, %f1272;
mul.f32 %f1274, %f97, %f1272;
mul.f32 %f1275, %f98, %f1272;
add.f32 %f1276, %f1253, %f1263;
add.f32 %f1277, %f1254, %f1264;
add.f32 %f1278, %f1255, %f1265;
add.f32 %f1279, %f1276, %f1273;
add.f32 %f1280, %f1277, %f1274;
add.f32 %f1281, %f1278, %f1275;
mul.f32 %f1282, %f1279, 0f3F13CD3A;
mul.f32 %f1283, %f1280, 0f3F13CD3A;
mul.f32 %f1284, %f1281, 0f3F13CD3A;
div.rn.f32 %f1285, %f80, %f1282;
div.rn.f32 %f1286, %f81, %f1283;
div.rn.f32 %f1287, %f82, %f1284;
setp.eq.f32 %p177, %f80, 0f00000000;
selp.f32 %f1288, 0f00000000, %f1285, %p177;
setp.eq.f32 %p178, %f81, 0f00000000;
selp.f32 %f1289, 0f00000000, %f1286, %p178;
setp.eq.f32 %p179, %f82, 0f00000000;
selp.f32 %f1290, 0f00000000, %f1287, %p179;
mul.f32 %f212, %f1253, %f1288;
mul.f32 %f213, %f1254, %f1289;
mul.f32 %f214, %f1255, %f1290;
mul.f32 %f215, %f1263, %f1288;
mul.f32 %f216, %f1264, %f1289;
mul.f32 %f217, %f1265, %f1290;
mul.f32 %f218, %f1273, %f1288;
mul.f32 %f219, %f1274, %f1289;
mul.f32 %f220, %f1275, %f1290;
ld.const.u32 %r49, [params+108];
setp.eq.s32 %p180, %r49, 0;
ld.const.u64 %rd78, [params+256];
cvta.to.global.u64 %rd79, %rd78;
ld.const.u32 %r338, [params+248];
mad.lo.s32 %r339, %r338, %r4, %r3;
mul.wide.u32 %rd80, %r339, 8;
add.s64 %rd21, %rd79, %rd80;
@%p180 bra $L__BB0_150;
ld.global.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd21];
// begin inline asm
{ cvt.f32.f16 %f1291, %rs74;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1292, %rs75;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1293, %rs76;}
// end inline asm
add.f32 %f1294, %f212, %f1291;
add.f32 %f1295, %f213, %f1292;
add.f32 %f1296, %f214, %f1293;
mov.f32 %f1297, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs72, %f1296;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs71, %f1295;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs70, %f1294;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs73, %f1297;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs70, %rs71, %rs72, %rs73};
bra.uni $L__BB0_151;
$L__BB0_150:
mov.f32 %f1301, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs81, %f1301;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs80, %f214;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs79, %f213;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs78, %f212;}
// end inline asm
st.global.v4.u16 [%rd21], {%rs78, %rs79, %rs80, %rs81};
$L__BB0_151:
ld.const.u64 %rd81, [params+272];
cvta.to.global.u64 %rd82, %rd81;
ld.const.u32 %r340, [params+264];
mad.lo.s32 %r341, %r340, %r4, %r3;
mul.wide.u32 %rd83, %r341, 8;
add.s64 %rd22, %rd82, %rd83;
@%p180 bra $L__BB0_153;
ld.global.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd22];
// begin inline asm
{ cvt.f32.f16 %f1302, %rs89;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1303, %rs90;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1304, %rs91;}
// end inline asm
add.f32 %f1305, %f215, %f1302;
add.f32 %f1306, %f216, %f1303;
add.f32 %f1307, %f217, %f1304;
mov.f32 %f1308, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs87, %f1307;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs86, %f1306;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs85, %f1305;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs88, %f1308;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs85, %rs86, %rs87, %rs88};
bra.uni $L__BB0_154;
$L__BB0_153:
mov.f32 %f1312, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs96, %f1312;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs95, %f217;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs94, %f216;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs93, %f215;}
// end inline asm
st.global.v4.u16 [%rd22], {%rs93, %rs94, %rs95, %rs96};
$L__BB0_154:
ld.const.u64 %rd84, [params+288];
cvta.to.global.u64 %rd85, %rd84;
ld.const.u32 %r342, [params+280];
mad.lo.s32 %r343, %r342, %r4, %r3;
mul.wide.u32 %rd86, %r343, 8;
add.s64 %rd23, %rd85, %rd86;
@%p180 bra $L__BB0_156;
ld.global.v4.u16 {%rs104, %rs105, %rs106, %rs107}, [%rd23];
// begin inline asm
{ cvt.f32.f16 %f1313, %rs104;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1314, %rs105;}
// end inline asm
// begin inline asm
{ cvt.f32.f16 %f1315, %rs106;}
// end inline asm
add.f32 %f1316, %f218, %f1313;
add.f32 %f1317, %f219, %f1314;
add.f32 %f1318, %f220, %f1315;
mov.f32 %f1319, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs102, %f1318;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs101, %f1317;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs100, %f1316;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs103, %f1319;}
// end inline asm
st.global.v4.u16 [%rd23], {%rs100, %rs101, %rs102, %rs103};
bra.uni $L__BB0_174;
$L__BB0_156:
mov.f32 %f1323, 0f3F800000;
// begin inline asm
{ cvt.rn.f16.f32 %rs111, %f1323;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs110, %f220;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs109, %f219;}
// end inline asm
// begin inline asm
{ cvt.rn.f16.f32 %rs108, %f218;}
// end inline asm
st.global.v4.u16 [%rd23], {%rs108, %rs109, %rs110, %rs111};
$L__BB0_174:
ret;
}