2121 lines
62 KiB
Plaintext
2121 lines
62 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];
|
|
|
|
.visible .entry __raygen__oxMain()
|
|
{
|
|
.reg .pred %p<179>;
|
|
.reg .b16 %rs<65>;
|
|
.reg .f32 %f<1422>;
|
|
.reg .b32 %r<419>;
|
|
.reg .b64 %rd<54>;
|
|
|
|
|
|
// begin inline asm
|
|
call (%r22), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r23), _optix_get_launch_index_y, ();
|
|
// end inline asm
|
|
ld.const.u64 %rd13, [params+400];
|
|
cvta.to.global.u64 %rd14, %rd13;
|
|
ld.const.u32 %r28, [params+392];
|
|
mad.lo.s32 %r29, %r28, %r23, %r22;
|
|
mul.wide.u32 %rd15, %r29, 4;
|
|
add.s64 %rd1, %rd14, %rd15;
|
|
ld.global.v2.u8 {%rs9, %rs63}, [%rd1];
|
|
or.b16 %rs11, %rs9, %rs63;
|
|
and.b16 %rs12, %rs11, 255;
|
|
setp.eq.s16 %p10, %rs12, 0;
|
|
@%p10 bra $L__BB0_2;
|
|
|
|
ld.global.u8 %rs62, [%rd1+2];
|
|
bra.uni $L__BB0_3;
|
|
|
|
$L__BB0_2:
|
|
ld.global.u8 %rs62, [%rd1+2];
|
|
setp.eq.s16 %p11, %rs62, 0;
|
|
mov.f32 %f1341, 0f00000000;
|
|
mov.u16 %rs63, 0;
|
|
mov.f32 %f1342, %f1341;
|
|
mov.f32 %f1343, %f1341;
|
|
@%p11 bra $L__BB0_4;
|
|
|
|
$L__BB0_3:
|
|
cvt.rn.f32.u16 %f244, %rs9;
|
|
div.rn.f32 %f245, %f244, 0f437F0000;
|
|
fma.rn.f32 %f246, %f245, 0f40000000, 0fBF800000;
|
|
and.b16 %rs15, %rs63, 255;
|
|
cvt.rn.f32.u16 %f247, %rs15;
|
|
div.rn.f32 %f248, %f247, 0f437F0000;
|
|
fma.rn.f32 %f249, %f248, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f250, %rs62;
|
|
div.rn.f32 %f251, %f250, 0f437F0000;
|
|
fma.rn.f32 %f252, %f251, 0f40000000, 0fBF800000;
|
|
mul.f32 %f253, %f249, %f249;
|
|
fma.rn.f32 %f254, %f246, %f246, %f253;
|
|
fma.rn.f32 %f255, %f252, %f252, %f254;
|
|
sqrt.rn.f32 %f256, %f255;
|
|
rcp.rn.f32 %f257, %f256;
|
|
mul.f32 %f1343, %f257, %f252;
|
|
mul.f32 %f1342, %f257, %f249;
|
|
mul.f32 %f1341, %f246, %f257;
|
|
|
|
$L__BB0_4:
|
|
ld.const.v2.u32 {%r30, %r31}, [params];
|
|
add.s32 %r3, %r30, %r22;
|
|
add.s32 %r4, %r31, %r23;
|
|
setp.eq.f32 %p12, %f1341, 0f00000000;
|
|
setp.eq.f32 %p13, %f1342, 0f00000000;
|
|
and.pred %p14, %p12, %p13;
|
|
setp.eq.f32 %p15, %f1343, 0f00000000;
|
|
and.pred %p16, %p15, %p14;
|
|
@%p16 bra $L__BB0_141;
|
|
bra.uni $L__BB0_5;
|
|
|
|
$L__BB0_141:
|
|
ld.const.u32 %r21, [params+104];
|
|
and.b32 %r400, %r21, 1;
|
|
setp.eq.b32 %p170, %r400, 1;
|
|
mov.pred %p171, 0;
|
|
xor.pred %p172, %p170, %p171;
|
|
not.pred %p173, %p172;
|
|
@%p173 bra $L__BB0_143;
|
|
|
|
ld.const.u64 %rd38, [params+144];
|
|
cvta.to.global.u64 %rd39, %rd38;
|
|
ld.const.u32 %r401, [params+136];
|
|
mad.lo.s32 %r402, %r401, %r4, %r3;
|
|
mul.wide.u32 %rd40, %r402, 4;
|
|
add.s64 %rd41, %rd39, %rd40;
|
|
mov.u16 %rs43, 0;
|
|
st.global.v4.u8 [%rd41], {%rs43, %rs43, %rs43, %rs43};
|
|
|
|
$L__BB0_143:
|
|
and.b32 %r403, %r21, 8;
|
|
setp.eq.s32 %p174, %r403, 0;
|
|
@%p174 bra $L__BB0_145;
|
|
|
|
ld.const.u64 %rd42, [params+192];
|
|
cvta.to.global.u64 %rd43, %rd42;
|
|
ld.const.u32 %r404, [params+184];
|
|
mad.lo.s32 %r405, %r404, %r4, %r3;
|
|
mov.f32 %f1328, 0f00000000;
|
|
cvt.rzi.u32.f32 %r406, %f1328;
|
|
mul.wide.u32 %rd44, %r405, 2;
|
|
add.s64 %rd45, %rd43, %rd44;
|
|
mov.u16 %rs44, 0;
|
|
cvt.u16.u32 %rs45, %r406;
|
|
st.global.v2.u8 [%rd45], {%rs45, %rs44};
|
|
|
|
$L__BB0_145:
|
|
and.b32 %r407, %r21, 4;
|
|
setp.eq.s32 %p175, %r407, 0;
|
|
@%p175 bra $L__BB0_149;
|
|
|
|
ld.const.u32 %r408, [params+108];
|
|
setp.eq.s32 %p176, %r408, 0;
|
|
ld.const.u64 %rd46, [params+224];
|
|
cvta.to.global.u64 %rd47, %rd46;
|
|
ld.const.u32 %r409, [params+216];
|
|
mad.lo.s32 %r410, %r409, %r4, %r3;
|
|
mul.wide.u32 %rd48, %r410, 8;
|
|
add.s64 %rd12, %rd47, %rd48;
|
|
@%p176 bra $L__BB0_148;
|
|
|
|
ld.global.v4.u16 {%rs52, %rs53, %rs54, %rs55}, [%rd12];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1329, %rs52;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1330, %rs53;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1331, %rs54;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1332, %f1329, 0f00000000;
|
|
add.f32 %f1333, %f1330, 0f00000000;
|
|
add.f32 %f1334, %f1331, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f1334;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f1333;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f1332;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs56, 0;
|
|
st.global.v4.u16 [%rd12], {%rs49, %rs50, %rs51, %rs56};
|
|
bra.uni $L__BB0_149;
|
|
|
|
$L__BB0_5:
|
|
ld.const.u64 %rd16, [params+432];
|
|
cvta.to.global.u64 %rd17, %rd16;
|
|
ld.const.u32 %r34, [params+424];
|
|
mad.lo.s32 %r35, %r34, %r23, %r22;
|
|
mul.wide.u32 %rd18, %r35, 12;
|
|
add.s64 %rd19, %rd17, %rd18;
|
|
ld.global.f32 %f10, [%rd19];
|
|
mul.f32 %f258, %f10, 0f3456BF95;
|
|
ld.global.f32 %f11, [%rd19+4];
|
|
mul.f32 %f259, %f11, 0f3456BF95;
|
|
ld.global.f32 %f12, [%rd19+8];
|
|
mul.f32 %f260, %f12, 0f3456BF95;
|
|
abs.f32 %f261, %f1341;
|
|
div.rn.f32 %f262, %f258, %f261;
|
|
abs.f32 %f263, %f1342;
|
|
div.rn.f32 %f264, %f259, %f263;
|
|
abs.f32 %f265, %f1343;
|
|
div.rn.f32 %f266, %f260, %f265;
|
|
abs.f32 %f267, %f262;
|
|
abs.f32 %f268, %f264;
|
|
abs.f32 %f269, %f266;
|
|
mov.f32 %f270, 0f38D1B717;
|
|
max.f32 %f271, %f267, %f270;
|
|
max.f32 %f272, %f268, %f270;
|
|
max.f32 %f273, %f269, %f270;
|
|
fma.rn.f32 %f13, %f1341, %f271, %f10;
|
|
fma.rn.f32 %f14, %f1342, %f272, %f11;
|
|
fma.rn.f32 %f15, %f1343, %f273, %f12;
|
|
cvt.rn.f32.u32 %f16, %r22;
|
|
cvt.rn.f32.u32 %f17, %r23;
|
|
ld.const.u32 %r5, [params+600];
|
|
setp.eq.s32 %p17, %r5, 0;
|
|
@%p17 bra $L__BB0_46;
|
|
|
|
ld.const.u64 %rd20, [params+608];
|
|
cvta.to.global.u64 %rd2, %rd20;
|
|
mov.f32 %f281, 0f40000000;
|
|
cvt.rzi.f32.f32 %f282, %f281;
|
|
add.f32 %f283, %f282, %f282;
|
|
mov.f32 %f284, 0f40800000;
|
|
sub.f32 %f285, %f284, %f283;
|
|
abs.f32 %f18, %f285;
|
|
ld.const.u32 %r6, [params+544];
|
|
ld.const.u32 %r418, [params+104];
|
|
and.b32 %r37, %r418, 32;
|
|
setp.eq.s32 %p18, %r37, 0;
|
|
mov.u32 %r414, 0;
|
|
selp.f32 %f19, 0f3F800000, 0f41200000, %p18;
|
|
mul.f32 %f20, %f13, 0f3456BF95;
|
|
mul.f32 %f21, %f14, 0f3456BF95;
|
|
mul.f32 %f22, %f15, 0f3456BF95;
|
|
mov.f32 %f280, 0f00000000;
|
|
ld.const.u64 %rd3, [params+96];
|
|
abs.f32 %f476, %f20;
|
|
abs.f32 %f477, %f21;
|
|
max.f32 %f478, %f476, %f477;
|
|
abs.f32 %f479, %f22;
|
|
max.f32 %f480, %f478, %f479;
|
|
mov.f32 %f1393, %f280;
|
|
mov.f32 %f1392, %f280;
|
|
mov.f32 %f1391, %f280;
|
|
mov.f32 %f1390, %f280;
|
|
mov.f32 %f1389, %f280;
|
|
mov.f32 %f1388, %f280;
|
|
mov.f32 %f1387, %f280;
|
|
|
|
$L__BB0_7:
|
|
mul.wide.u32 %rd21, %r414, 112;
|
|
add.s64 %rd5, %rd2, %rd21;
|
|
ld.global.v4.f32 {%f287, %f288, %f289, %f290}, [%rd5];
|
|
ld.global.v4.f32 {%f294, %f295, %f296, %f297}, [%rd5+16];
|
|
ld.global.v4.f32 {%f300, %f1359, %f1360, %f303}, [%rd5+32];
|
|
ld.global.v4.f32 {%f304, %f305, %f306, %f1357}, [%rd5+48];
|
|
ld.global.v4.f32 {%f308, %f309, %f310, %f311}, [%rd5+64];
|
|
ld.global.v4.u32 {%r38, %r39, %r40, %r41}, [%rd5+80];
|
|
mov.b32 %f48, %r38;
|
|
mov.b32 %f49, %r39;
|
|
mov.b32 %f50, %r40;
|
|
ld.global.u64 %rd6, [%rd5+96];
|
|
sub.f32 %f51, %f288, %f10;
|
|
sub.f32 %f52, %f289, %f11;
|
|
sub.f32 %f53, %f290, %f12;
|
|
mul.f32 %f312, %f52, %f52;
|
|
fma.rn.f32 %f313, %f51, %f51, %f312;
|
|
fma.rn.f32 %f314, %f53, %f53, %f313;
|
|
sqrt.rn.f32 %f315, %f314;
|
|
rcp.rn.f32 %f316, %f315;
|
|
mul.f32 %f54, %f51, %f316;
|
|
mul.f32 %f55, %f52, %f316;
|
|
mul.f32 %f56, %f53, %f316;
|
|
mul.f32 %f57, %f315, %f295;
|
|
mul.f32 %f58, %f315, %f297;
|
|
abs.f32 %f59, %f58;
|
|
setp.lt.f32 %p19, %f59, 0f00800000;
|
|
mul.f32 %f317, %f59, 0f4B800000;
|
|
selp.f32 %f318, %f317, %f59, %p19;
|
|
selp.f32 %f319, 0fC3170000, 0fC2FE0000, %p19;
|
|
mov.b32 %r45, %f318;
|
|
and.b32 %r46, %r45, 8388607;
|
|
or.b32 %r47, %r46, 1065353216;
|
|
mov.b32 %f320, %r47;
|
|
shr.u32 %r48, %r45, 23;
|
|
cvt.rn.f32.u32 %f321, %r48;
|
|
add.f32 %f322, %f319, %f321;
|
|
setp.gt.f32 %p20, %f320, 0f3FB504F3;
|
|
mul.f32 %f323, %f320, 0f3F000000;
|
|
add.f32 %f324, %f322, 0f3F800000;
|
|
selp.f32 %f325, %f324, %f322, %p20;
|
|
selp.f32 %f326, %f323, %f320, %p20;
|
|
add.f32 %f327, %f326, 0fBF800000;
|
|
add.f32 %f328, %f326, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f329, %f328;
|
|
add.f32 %f330, %f327, %f327;
|
|
mul.f32 %f331, %f330, %f329;
|
|
mul.f32 %f332, %f331, %f331;
|
|
mov.f32 %f333, 0f3C4CAF63;
|
|
mov.f32 %f334, 0f3B18F0FE;
|
|
fma.rn.f32 %f335, %f334, %f332, %f333;
|
|
mov.f32 %f336, 0f3DAAAABD;
|
|
fma.rn.f32 %f337, %f335, %f332, %f336;
|
|
mul.rn.f32 %f338, %f337, %f332;
|
|
mul.rn.f32 %f339, %f338, %f331;
|
|
sub.f32 %f340, %f327, %f331;
|
|
add.f32 %f341, %f340, %f340;
|
|
neg.f32 %f342, %f331;
|
|
fma.rn.f32 %f343, %f342, %f327, %f341;
|
|
mul.rn.f32 %f344, %f329, %f343;
|
|
add.f32 %f345, %f339, %f331;
|
|
sub.f32 %f346, %f331, %f345;
|
|
add.f32 %f347, %f339, %f346;
|
|
add.f32 %f348, %f344, %f347;
|
|
add.f32 %f349, %f345, %f348;
|
|
sub.f32 %f350, %f345, %f349;
|
|
add.f32 %f351, %f348, %f350;
|
|
mov.f32 %f352, 0f3F317200;
|
|
mul.rn.f32 %f353, %f325, %f352;
|
|
mov.f32 %f354, 0f35BFBE8E;
|
|
mul.rn.f32 %f355, %f325, %f354;
|
|
add.f32 %f356, %f353, %f349;
|
|
sub.f32 %f357, %f353, %f356;
|
|
add.f32 %f358, %f349, %f357;
|
|
add.f32 %f359, %f351, %f358;
|
|
add.f32 %f360, %f355, %f359;
|
|
add.f32 %f361, %f356, %f360;
|
|
sub.f32 %f362, %f356, %f361;
|
|
add.f32 %f363, %f360, %f362;
|
|
mul.rn.f32 %f365, %f284, %f361;
|
|
neg.f32 %f366, %f365;
|
|
fma.rn.f32 %f367, %f284, %f361, %f366;
|
|
fma.rn.f32 %f368, %f284, %f363, %f367;
|
|
fma.rn.f32 %f370, %f280, %f361, %f368;
|
|
add.rn.f32 %f371, %f365, %f370;
|
|
neg.f32 %f372, %f371;
|
|
add.rn.f32 %f373, %f365, %f372;
|
|
add.rn.f32 %f374, %f373, %f370;
|
|
mov.b32 %r49, %f371;
|
|
setp.eq.s32 %p21, %r49, 1118925336;
|
|
add.s32 %r50, %r49, -1;
|
|
mov.b32 %f375, %r50;
|
|
add.f32 %f376, %f374, 0f37000000;
|
|
selp.f32 %f60, %f376, %f374, %p21;
|
|
selp.f32 %f377, %f375, %f371, %p21;
|
|
mov.f32 %f378, 0f3FB8AA3B;
|
|
mul.rn.f32 %f379, %f377, %f378;
|
|
cvt.rzi.f32.f32 %f380, %f379;
|
|
abs.f32 %f381, %f380;
|
|
setp.gt.f32 %p22, %f381, 0f42FC0000;
|
|
mov.b32 %r51, %f380;
|
|
and.b32 %r52, %r51, -2147483648;
|
|
or.b32 %r53, %r52, 1123811328;
|
|
mov.b32 %f382, %r53;
|
|
selp.f32 %f383, %f382, %f380, %p22;
|
|
mov.f32 %f384, 0fBF317218;
|
|
fma.rn.f32 %f385, %f383, %f384, %f377;
|
|
mov.f32 %f386, 0f3102E308;
|
|
fma.rn.f32 %f387, %f383, %f386, %f385;
|
|
mul.f32 %f388, %f387, 0f3FB8AA3B;
|
|
add.f32 %f389, %f383, 0f4B40007F;
|
|
mov.b32 %r54, %f389;
|
|
shl.b32 %r55, %r54, 23;
|
|
mov.b32 %f390, %r55;
|
|
ex2.approx.ftz.f32 %f391, %f388;
|
|
mul.f32 %f61, %f391, %f390;
|
|
setp.eq.f32 %p23, %f61, 0f7F800000;
|
|
mov.f32 %f1351, 0f7F800000;
|
|
@%p23 bra $L__BB0_9;
|
|
|
|
fma.rn.f32 %f1351, %f61, %f60, %f61;
|
|
|
|
$L__BB0_9:
|
|
setp.lt.f32 %p24, %f58, 0f00000000;
|
|
setp.eq.f32 %p25, %f18, 0f3F800000;
|
|
and.pred %p1, %p24, %p25;
|
|
setp.eq.f32 %p26, %f58, 0f00000000;
|
|
@%p26 bra $L__BB0_13;
|
|
bra.uni $L__BB0_10;
|
|
|
|
$L__BB0_13:
|
|
add.f32 %f396, %f58, %f58;
|
|
selp.f32 %f1353, %f396, 0f00000000, %p25;
|
|
bra.uni $L__BB0_14;
|
|
|
|
$L__BB0_10:
|
|
mov.b32 %r56, %f1351;
|
|
xor.b32 %r57, %r56, -2147483648;
|
|
mov.b32 %f392, %r57;
|
|
selp.f32 %f1353, %f392, %f1351, %p1;
|
|
setp.geu.f32 %p27, %f58, 0f00000000;
|
|
@%p27 bra $L__BB0_14;
|
|
|
|
mov.f32 %f393, 0f40800000;
|
|
cvt.rzi.f32.f32 %f394, %f393;
|
|
setp.eq.f32 %p28, %f394, 0f40800000;
|
|
@%p28 bra $L__BB0_14;
|
|
|
|
mov.f32 %f1353, 0f7FFFFFFF;
|
|
|
|
$L__BB0_14:
|
|
abs.f32 %f1338, %f58;
|
|
add.f32 %f397, %f1338, 0f40800000;
|
|
mov.b32 %r58, %f397;
|
|
setp.lt.s32 %p30, %r58, 2139095040;
|
|
@%p30 bra $L__BB0_19;
|
|
|
|
abs.f32 %f1339, %f58;
|
|
setp.gtu.f32 %p31, %f1339, 0f7F800000;
|
|
@%p31 bra $L__BB0_18;
|
|
bra.uni $L__BB0_16;
|
|
|
|
$L__BB0_18:
|
|
add.f32 %f1353, %f58, 0f40800000;
|
|
bra.uni $L__BB0_19;
|
|
|
|
$L__BB0_16:
|
|
abs.f32 %f1340, %f58;
|
|
setp.neu.f32 %p32, %f1340, 0f7F800000;
|
|
@%p32 bra $L__BB0_19;
|
|
|
|
selp.f32 %f1353, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
$L__BB0_19:
|
|
mov.f32 %f398, 0f3F800000;
|
|
sub.f32 %f399, %f398, %f1353;
|
|
setp.eq.f32 %p33, %f58, 0f3F800000;
|
|
selp.f32 %f400, 0f00000000, %f399, %p33;
|
|
cvt.sat.f32.f32 %f401, %f400;
|
|
fma.rn.f32 %f402, %f57, %f57, %f296;
|
|
div.rn.f32 %f70, %f401, %f402;
|
|
mul.f32 %f403, %f1342, %f55;
|
|
fma.rn.f32 %f404, %f1341, %f54, %f403;
|
|
fma.rn.f32 %f71, %f1343, %f56, %f404;
|
|
setp.eq.f32 %p34, %f287, 0f3F800000;
|
|
@%p34 bra $L__BB0_25;
|
|
bra.uni $L__BB0_20;
|
|
|
|
$L__BB0_25:
|
|
setp.eq.s64 %p38, %rd6, 0;
|
|
mov.f32 %f1358, %f300;
|
|
mov.f32 %f1361, %f70;
|
|
@%p38 bra $L__BB0_29;
|
|
|
|
setp.geu.f32 %p39, %f1357, 0f00000000;
|
|
mov.f32 %f1354, %f54;
|
|
mov.f32 %f1355, %f55;
|
|
mov.f32 %f1356, %f56;
|
|
@%p39 bra $L__BB0_28;
|
|
|
|
mul.f32 %f427, %f49, %f52;
|
|
fma.rn.f32 %f428, %f48, %f51, %f427;
|
|
fma.rn.f32 %f429, %f50, %f53, %f428;
|
|
rcp.rn.f32 %f430, %f429;
|
|
mul.f32 %f1354, %f51, %f430;
|
|
mul.f32 %f1355, %f52, %f430;
|
|
mul.f32 %f1356, %f53, %f430;
|
|
neg.f32 %f1357, %f1357;
|
|
|
|
$L__BB0_28:
|
|
mul.f32 %f431, %f50, %f56;
|
|
mul.f32 %f432, %f49, %f55;
|
|
neg.f32 %f433, %f432;
|
|
mul.f32 %f434, %f48, %f54;
|
|
sub.f32 %f435, %f433, %f434;
|
|
sub.f32 %f436, %f435, %f431;
|
|
setp.gt.f32 %p40, %f436, 0f00000000;
|
|
selp.f32 %f437, 0f3F800000, 0f00000000, %p40;
|
|
mov.f32 %f438, 0f3F800000;
|
|
mul.f32 %f439, %f305, %f1355;
|
|
fma.rn.f32 %f440, %f304, %f1354, %f439;
|
|
mul.f32 %f441, %f309, %f1355;
|
|
fma.rn.f32 %f442, %f308, %f1354, %f441;
|
|
fma.rn.f32 %f443, %f306, %f1356, %f440;
|
|
fma.rn.f32 %f444, %f310, %f1356, %f442;
|
|
fma.rn.f32 %f445, %f1357, %f443, 0f3F000000;
|
|
sub.f32 %f446, %f438, %f445;
|
|
fma.rn.f32 %f447, %f1357, %f444, 0f3F000000;
|
|
tex.2d.v4.f32.f32 {%f448, %f449, %f450, %f451}, [%rd6, {%f446, %f447}];
|
|
mul.f32 %f452, %f437, %f448;
|
|
mul.f32 %f453, %f437, %f449;
|
|
mul.f32 %f454, %f437, %f450;
|
|
mul.f32 %f1358, %f300, %f452;
|
|
mul.f32 %f1359, %f1359, %f453;
|
|
mul.f32 %f1360, %f1360, %f454;
|
|
mov.f32 %f1361, %f70;
|
|
bra.uni $L__BB0_29;
|
|
|
|
$L__BB0_20:
|
|
setp.eq.f32 %p35, %f287, 0f40000000;
|
|
@%p35 bra $L__BB0_23;
|
|
bra.uni $L__BB0_21;
|
|
|
|
$L__BB0_23:
|
|
setp.eq.s64 %p37, %rd6, 0;
|
|
mov.f32 %f1358, %f300;
|
|
mov.f32 %f1361, %f70;
|
|
@%p37 bra $L__BB0_29;
|
|
|
|
mul.f32 %f414, %f305, %f55;
|
|
fma.rn.f32 %f415, %f304, %f54, %f414;
|
|
mul.f32 %f416, %f309, %f55;
|
|
fma.rn.f32 %f417, %f308, %f54, %f416;
|
|
mul.f32 %f418, %f49, %f55;
|
|
fma.rn.f32 %f419, %f48, %f54, %f418;
|
|
fma.rn.f32 %f420, %f306, %f56, %f415;
|
|
fma.rn.f32 %f421, %f310, %f56, %f417;
|
|
fma.rn.f32 %f422, %f50, %f56, %f419;
|
|
tex.cube.v4.f32.f32 {%f423, %f424, %f425, %f426}, [%rd6, {%f420, %f421, %f422, %f422}];
|
|
mul.f32 %f1358, %f300, %f423;
|
|
mul.f32 %f1359, %f1359, %f424;
|
|
mul.f32 %f1360, %f1360, %f425;
|
|
mov.f32 %f1361, %f70;
|
|
bra.uni $L__BB0_29;
|
|
|
|
$L__BB0_21:
|
|
setp.neu.f32 %p36, %f287, 0f40800000;
|
|
mov.f32 %f1358, %f300;
|
|
mov.f32 %f1361, %f70;
|
|
@%p36 bra $L__BB0_29;
|
|
|
|
mul.f32 %f405, %f48, %f54;
|
|
mul.f32 %f406, %f49, %f55;
|
|
neg.f32 %f407, %f406;
|
|
sub.f32 %f408, %f407, %f405;
|
|
mul.f32 %f409, %f50, %f56;
|
|
sub.f32 %f410, %f408, %f409;
|
|
fma.rn.f32 %f411, %f303, %f410, %f1357;
|
|
cvt.sat.f32.f32 %f412, %f411;
|
|
mul.f32 %f413, %f412, %f412;
|
|
mul.f32 %f1361, %f70, %f413;
|
|
mov.f32 %f1358, %f300;
|
|
|
|
$L__BB0_29:
|
|
max.f32 %f461, %f1358, %f1359;
|
|
max.f32 %f462, %f461, %f1360;
|
|
setp.eq.s32 %p42, %r6, 0;
|
|
selp.f32 %f463, %f71, 0f3F800000, %p42;
|
|
cvt.sat.f32.f32 %f464, %f463;
|
|
mul.f32 %f91, %f464, %f1361;
|
|
mul.f32 %f465, %f91, %f462;
|
|
setp.lt.f32 %p43, %f465, 0f3727C5AC;
|
|
mov.pred %p178, -1;
|
|
mov.f32 %f1362, 0f00000000;
|
|
mov.f32 %f1363, %f1362;
|
|
mov.f32 %f1364, %f1362;
|
|
mov.f32 %f1365, %f1362;
|
|
mov.f32 %f1366, %f1362;
|
|
mov.f32 %f1367, %f1362;
|
|
@%p43 bra $L__BB0_31;
|
|
|
|
and.b32 %r59, %r418, 64;
|
|
setp.eq.s32 %p45, %r59, 0;
|
|
mul.f32 %f1362, %f1358, %f91;
|
|
mul.f32 %f1363, %f1359, %f91;
|
|
mul.f32 %f1364, %f1360, %f91;
|
|
selp.f32 %f1365, 0f00000000, %f54, %p45;
|
|
selp.f32 %f1366, 0f00000000, %f55, %p45;
|
|
selp.f32 %f1367, 0f00000000, %f56, %p45;
|
|
mov.pred %p178, 0;
|
|
|
|
$L__BB0_31:
|
|
@%p178 bra $L__BB0_45;
|
|
|
|
setp.eq.s32 %p46, %r41, 0;
|
|
mov.u16 %rs64, 0;
|
|
mov.f32 %f1386, 0f3F800000;
|
|
mov.f32 %f1381, %f1386;
|
|
mov.f32 %f1382, %f1386;
|
|
@%p46 bra $L__BB0_40;
|
|
|
|
abs.s32 %r10, %r41;
|
|
setp.lt.s32 %p47, %r10, 1;
|
|
mov.f32 %f1371, 0f00000000;
|
|
mov.f32 %f1372, %f1371;
|
|
mov.f32 %f1373, %f1371;
|
|
@%p47 bra $L__BB0_39;
|
|
|
|
mov.f32 %f481, 0f38D1B717;
|
|
max.f32 %f482, %f480, %f481;
|
|
mul.f32 %f104, %f19, %f482;
|
|
and.b32 %r11, %r10, 1;
|
|
setp.eq.s32 %p48, %r10, 1;
|
|
mov.f32 %f1373, 0f00000000;
|
|
mov.u32 %r417, 0;
|
|
mov.f32 %f1372, %f1373;
|
|
mov.f32 %f1371, %f1373;
|
|
@%p48 bra $L__BB0_37;
|
|
|
|
sub.s32 %r416, %r10, %r11;
|
|
|
|
$L__BB0_36:
|
|
cvt.rn.f32.s32 %f504, %r417;
|
|
add.f32 %f505, %f16, %f504;
|
|
sub.f32 %f506, %f17, %f504;
|
|
mul.f32 %f507, %f505, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f508, %f507;
|
|
sub.f32 %f509, %f507, %f508;
|
|
mul.f32 %f510, %f506, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f511, %f510;
|
|
sub.f32 %f512, %f510, %f511;
|
|
mul.f32 %f513, %f504, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f514, %f513;
|
|
sub.f32 %f515, %f513, %f514;
|
|
add.f32 %f516, %f512, 0f420551EC;
|
|
add.f32 %f517, %f509, 0f420551EC;
|
|
add.f32 %f518, %f515, 0f420551EC;
|
|
mul.f32 %f519, %f509, %f516;
|
|
fma.rn.f32 %f520, %f512, %f517, %f519;
|
|
fma.rn.f32 %f521, %f515, %f518, %f520;
|
|
add.f32 %f522, %f509, %f521;
|
|
add.f32 %f523, %f512, %f521;
|
|
add.f32 %f524, %f515, %f521;
|
|
add.f32 %f525, %f522, %f523;
|
|
mul.f32 %f526, %f524, %f525;
|
|
cvt.rmi.f32.f32 %f527, %f526;
|
|
sub.f32 %f528, %f526, %f527;
|
|
add.f32 %f529, %f522, %f522;
|
|
mul.f32 %f530, %f523, %f529;
|
|
cvt.rmi.f32.f32 %f531, %f530;
|
|
sub.f32 %f532, %f530, %f531;
|
|
mul.f32 %f533, %f522, %f525;
|
|
cvt.rmi.f32.f32 %f534, %f533;
|
|
sub.f32 %f535, %f533, %f534;
|
|
fma.rn.f32 %f536, %f528, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f537, %f532, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f538, %f535, 0f40000000, 0fBF800000;
|
|
ld.global.f32 %f539, [%rd5+16];
|
|
ld.global.f32 %f540, [%rd5+4];
|
|
fma.rn.f32 %f541, %f539, %f536, %f540;
|
|
ld.global.f32 %f542, [%rd5+8];
|
|
fma.rn.f32 %f543, %f539, %f537, %f542;
|
|
ld.global.f32 %f544, [%rd5+12];
|
|
fma.rn.f32 %f545, %f539, %f538, %f544;
|
|
sub.f32 %f546, %f541, %f10;
|
|
sub.f32 %f547, %f543, %f11;
|
|
sub.f32 %f548, %f545, %f12;
|
|
mul.f32 %f549, %f547, %f547;
|
|
fma.rn.f32 %f550, %f546, %f546, %f549;
|
|
fma.rn.f32 %f551, %f548, %f548, %f550;
|
|
sqrt.rn.f32 %f493, %f551;
|
|
rcp.rn.f32 %f552, %f493;
|
|
mul.f32 %f489, %f552, %f546;
|
|
mul.f32 %f490, %f552, %f547;
|
|
mul.f32 %f491, %f552, %f548;
|
|
mov.f32 %f503, 0f00000000;
|
|
mov.u32 %r169, 2;
|
|
mov.u32 %r170, 1;
|
|
mov.u32 %r171, 3;
|
|
mov.u32 %r174, 1065353216;
|
|
mov.u32 %r203, 0;
|
|
// begin inline asm
|
|
call(%r62,%r63,%r64,%r65,%r66,%r67,%r68,%r69,%r70,%r71,%r72,%r73,%r74,%r75,%r76,%r77,%r78,%r79,%r80,%r81,%r82,%r83,%r84,%r85,%r86,%r87,%r88,%r89,%r90,%r91,%r92,%r93),_optix_trace_typed_32,(%r203,%rd3,%f13,%f14,%f15,%f489,%f490,%f491,%f104,%f493,%f503,%r170,%r203,%r170,%r169,%r170,%r171,%r174,%r174,%r174,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203);
|
|
// end inline asm
|
|
mov.b32 %f553, %r62;
|
|
mov.b32 %f554, %r63;
|
|
mov.b32 %f555, %r64;
|
|
add.f32 %f556, %f1371, %f553;
|
|
add.f32 %f557, %f1372, %f554;
|
|
add.f32 %f558, %f1373, %f555;
|
|
add.s32 %r204, %r417, 1;
|
|
cvt.rn.f32.s32 %f559, %r204;
|
|
add.f32 %f560, %f16, %f559;
|
|
sub.f32 %f561, %f17, %f559;
|
|
mul.f32 %f562, %f560, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f563, %f562;
|
|
sub.f32 %f564, %f562, %f563;
|
|
mul.f32 %f565, %f561, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f566, %f565;
|
|
sub.f32 %f567, %f565, %f566;
|
|
mul.f32 %f568, %f559, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f569, %f568;
|
|
sub.f32 %f570, %f568, %f569;
|
|
add.f32 %f571, %f567, 0f420551EC;
|
|
add.f32 %f572, %f564, 0f420551EC;
|
|
add.f32 %f573, %f570, 0f420551EC;
|
|
mul.f32 %f574, %f564, %f571;
|
|
fma.rn.f32 %f575, %f567, %f572, %f574;
|
|
fma.rn.f32 %f576, %f570, %f573, %f575;
|
|
add.f32 %f577, %f564, %f576;
|
|
add.f32 %f578, %f567, %f576;
|
|
add.f32 %f579, %f570, %f576;
|
|
add.f32 %f580, %f577, %f578;
|
|
mul.f32 %f581, %f579, %f580;
|
|
cvt.rmi.f32.f32 %f582, %f581;
|
|
sub.f32 %f583, %f581, %f582;
|
|
add.f32 %f584, %f577, %f577;
|
|
mul.f32 %f585, %f578, %f584;
|
|
cvt.rmi.f32.f32 %f586, %f585;
|
|
sub.f32 %f587, %f585, %f586;
|
|
mul.f32 %f588, %f577, %f580;
|
|
cvt.rmi.f32.f32 %f589, %f588;
|
|
sub.f32 %f590, %f588, %f589;
|
|
fma.rn.f32 %f591, %f583, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f592, %f587, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f593, %f590, 0f40000000, 0fBF800000;
|
|
ld.global.f32 %f594, [%rd5+16];
|
|
ld.global.f32 %f595, [%rd5+4];
|
|
fma.rn.f32 %f596, %f594, %f591, %f595;
|
|
ld.global.f32 %f597, [%rd5+8];
|
|
fma.rn.f32 %f598, %f594, %f592, %f597;
|
|
ld.global.f32 %f599, [%rd5+12];
|
|
fma.rn.f32 %f600, %f594, %f593, %f599;
|
|
sub.f32 %f601, %f596, %f10;
|
|
sub.f32 %f602, %f598, %f11;
|
|
sub.f32 %f603, %f600, %f12;
|
|
mul.f32 %f604, %f602, %f602;
|
|
fma.rn.f32 %f605, %f601, %f601, %f604;
|
|
fma.rn.f32 %f606, %f603, %f603, %f605;
|
|
sqrt.rn.f32 %f502, %f606;
|
|
rcp.rn.f32 %f607, %f502;
|
|
mul.f32 %f498, %f607, %f601;
|
|
mul.f32 %f499, %f607, %f602;
|
|
mul.f32 %f500, %f607, %f603;
|
|
// begin inline asm
|
|
call(%r133,%r134,%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),_optix_trace_typed_32,(%r203,%rd3,%f13,%f14,%f15,%f498,%f499,%f500,%f104,%f502,%f503,%r170,%r203,%r170,%r169,%r170,%r171,%r174,%r174,%r174,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203);
|
|
// end inline asm
|
|
mov.b32 %f608, %r133;
|
|
mov.b32 %f609, %r134;
|
|
mov.b32 %f610, %r135;
|
|
add.f32 %f1371, %f556, %f608;
|
|
add.f32 %f1372, %f557, %f609;
|
|
add.f32 %f1373, %f558, %f610;
|
|
add.s32 %r417, %r417, 2;
|
|
add.s32 %r416, %r416, -2;
|
|
setp.ne.s32 %p49, %r416, 0;
|
|
@%p49 bra $L__BB0_36;
|
|
|
|
$L__BB0_37:
|
|
setp.eq.s32 %p50, %r11, 0;
|
|
@%p50 bra $L__BB0_39;
|
|
|
|
cvt.rn.f32.s32 %f620, %r417;
|
|
add.f32 %f621, %f16, %f620;
|
|
sub.f32 %f622, %f17, %f620;
|
|
mul.f32 %f623, %f621, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f624, %f623;
|
|
sub.f32 %f625, %f623, %f624;
|
|
mul.f32 %f626, %f622, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f627, %f626;
|
|
sub.f32 %f628, %f626, %f627;
|
|
mul.f32 %f629, %f620, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f630, %f629;
|
|
sub.f32 %f631, %f629, %f630;
|
|
add.f32 %f632, %f628, 0f420551EC;
|
|
add.f32 %f633, %f625, 0f420551EC;
|
|
add.f32 %f634, %f631, 0f420551EC;
|
|
mul.f32 %f635, %f625, %f632;
|
|
fma.rn.f32 %f636, %f628, %f633, %f635;
|
|
fma.rn.f32 %f637, %f631, %f634, %f636;
|
|
add.f32 %f638, %f625, %f637;
|
|
add.f32 %f639, %f628, %f637;
|
|
add.f32 %f640, %f631, %f637;
|
|
add.f32 %f641, %f638, %f639;
|
|
mul.f32 %f642, %f640, %f641;
|
|
cvt.rmi.f32.f32 %f643, %f642;
|
|
sub.f32 %f644, %f642, %f643;
|
|
add.f32 %f645, %f638, %f638;
|
|
mul.f32 %f646, %f639, %f645;
|
|
cvt.rmi.f32.f32 %f647, %f646;
|
|
sub.f32 %f648, %f646, %f647;
|
|
mul.f32 %f649, %f638, %f641;
|
|
cvt.rmi.f32.f32 %f650, %f649;
|
|
sub.f32 %f651, %f649, %f650;
|
|
fma.rn.f32 %f652, %f644, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f653, %f648, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f654, %f651, 0f40000000, 0fBF800000;
|
|
ld.global.f32 %f655, [%rd5+16];
|
|
ld.global.f32 %f656, [%rd5+4];
|
|
fma.rn.f32 %f657, %f655, %f652, %f656;
|
|
ld.global.f32 %f658, [%rd5+8];
|
|
fma.rn.f32 %f659, %f655, %f653, %f658;
|
|
ld.global.f32 %f660, [%rd5+12];
|
|
fma.rn.f32 %f661, %f655, %f654, %f660;
|
|
sub.f32 %f662, %f657, %f10;
|
|
sub.f32 %f663, %f659, %f11;
|
|
sub.f32 %f664, %f661, %f12;
|
|
mul.f32 %f665, %f663, %f663;
|
|
fma.rn.f32 %f666, %f662, %f662, %f665;
|
|
fma.rn.f32 %f667, %f664, %f664, %f666;
|
|
sqrt.rn.f32 %f618, %f667;
|
|
rcp.rn.f32 %f668, %f618;
|
|
mul.f32 %f614, %f668, %f662;
|
|
mul.f32 %f615, %f668, %f663;
|
|
mul.f32 %f616, %f668, %f664;
|
|
mov.f32 %f619, 0f00000000;
|
|
mov.u32 %r241, 2;
|
|
mov.u32 %r242, 1;
|
|
mov.u32 %r243, 3;
|
|
mov.u32 %r246, 1065353216;
|
|
mov.u32 %r275, 0;
|
|
// begin inline asm
|
|
call(%r205,%r206,%r207,%r208,%r209,%r210,%r211,%r212,%r213,%r214,%r215,%r216,%r217,%r218,%r219,%r220,%r221,%r222,%r223,%r224,%r225,%r226,%r227,%r228,%r229,%r230,%r231,%r232,%r233,%r234,%r235,%r236),_optix_trace_typed_32,(%r275,%rd3,%f13,%f14,%f15,%f614,%f615,%f616,%f104,%f618,%f619,%r242,%r275,%r242,%r241,%r242,%r243,%r246,%r246,%r246,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275,%r275);
|
|
// end inline asm
|
|
mov.b32 %f669, %r205;
|
|
mov.b32 %f670, %r206;
|
|
mov.b32 %f671, %r207;
|
|
add.f32 %f1371, %f1371, %f669;
|
|
add.f32 %f1372, %f1372, %f670;
|
|
add.f32 %f1373, %f1373, %f671;
|
|
|
|
$L__BB0_39:
|
|
cvt.rn.f32.s32 %f672, %r10;
|
|
rcp.rn.f32 %f673, %f672;
|
|
mul.f32 %f1386, %f673, %f1371;
|
|
mul.f32 %f1381, %f673, %f1372;
|
|
mul.f32 %f1382, %f673, %f1373;
|
|
shr.u32 %r276, %r41, 31;
|
|
cvt.u16.u32 %rs64, %r276;
|
|
|
|
$L__BB0_40:
|
|
and.b32 %r277, %r418, 64;
|
|
setp.eq.s32 %p51, %r277, 0;
|
|
fma.rn.f32 %f1387, %f1362, %f1386, %f1387;
|
|
fma.rn.f32 %f1388, %f1363, %f1381, %f1388;
|
|
fma.rn.f32 %f1389, %f1364, %f1382, %f1389;
|
|
@%p51 bra $L__BB0_42;
|
|
|
|
mul.f32 %f674, %f1363, 0f3F372474;
|
|
fma.rn.f32 %f675, %f1362, 0f3E59999A, %f674;
|
|
fma.rn.f32 %f676, %f1364, 0f3D93A92A, %f675;
|
|
fma.rn.f32 %f1390, %f1365, %f676, %f1390;
|
|
fma.rn.f32 %f1391, %f1366, %f676, %f1391;
|
|
fma.rn.f32 %f1392, %f676, %f1367, %f1392;
|
|
|
|
$L__BB0_42:
|
|
setp.eq.s16 %p52, %rs64, 0;
|
|
@%p52 bra $L__BB0_44;
|
|
|
|
div.rn.f32 %f677, %f1362, %f300;
|
|
div.rn.f32 %f678, %f677, %f70;
|
|
cvt.sat.f32.f32 %f679, %f678;
|
|
mul.f32 %f1386, %f1386, %f679;
|
|
|
|
$L__BB0_44:
|
|
add.f32 %f1393, %f1393, %f1386;
|
|
|
|
$L__BB0_45:
|
|
cvt.u64.u32 %rd53, %r414;
|
|
cvt.u32.u64 %r278, %rd53;
|
|
add.s32 %r414, %r278, 1;
|
|
setp.lt.u32 %p53, %r414, %r5;
|
|
@%p53 bra $L__BB0_7;
|
|
bra.uni $L__BB0_47;
|
|
|
|
$L__BB0_46:
|
|
mov.f32 %f1387, 0f00000000;
|
|
ld.const.u32 %r418, [params+104];
|
|
mov.f32 %f1388, %f1387;
|
|
mov.f32 %f1389, %f1387;
|
|
mov.f32 %f1390, %f1387;
|
|
mov.f32 %f1391, %f1387;
|
|
mov.f32 %f1392, %f1387;
|
|
mov.f32 %f1393, %f1387;
|
|
|
|
$L__BB0_47:
|
|
and.b32 %r279, %r418, 8;
|
|
setp.eq.s32 %p54, %r279, 0;
|
|
@%p54 bra $L__BB0_61;
|
|
|
|
cvt.sat.f32.f32 %f155, %f1393;
|
|
ld.const.u64 %rd25, [params+192];
|
|
cvta.to.global.u64 %rd7, %rd25;
|
|
ld.const.u32 %r280, [params+184];
|
|
mad.lo.s32 %r281, %r280, %r4, %r3;
|
|
cvt.u64.u32 %rd8, %r281;
|
|
mov.f32 %f688, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f689, %f688;
|
|
add.f32 %f690, %f689, %f689;
|
|
mov.f32 %f691, 0f3EE8BA2E;
|
|
sub.f32 %f692, %f691, %f690;
|
|
abs.f32 %f156, %f692;
|
|
abs.f32 %f157, %f155;
|
|
setp.lt.f32 %p55, %f157, 0f00800000;
|
|
mul.f32 %f693, %f157, 0f4B800000;
|
|
selp.f32 %f694, %f693, %f157, %p55;
|
|
selp.f32 %f695, 0fC3170000, 0fC2FE0000, %p55;
|
|
mov.b32 %r282, %f694;
|
|
and.b32 %r283, %r282, 8388607;
|
|
or.b32 %r284, %r283, 1065353216;
|
|
mov.b32 %f696, %r284;
|
|
shr.u32 %r285, %r282, 23;
|
|
cvt.rn.f32.u32 %f697, %r285;
|
|
add.f32 %f698, %f695, %f697;
|
|
setp.gt.f32 %p56, %f696, 0f3FB504F3;
|
|
mul.f32 %f699, %f696, 0f3F000000;
|
|
add.f32 %f700, %f698, 0f3F800000;
|
|
selp.f32 %f701, %f700, %f698, %p56;
|
|
selp.f32 %f702, %f699, %f696, %p56;
|
|
add.f32 %f703, %f702, 0fBF800000;
|
|
add.f32 %f704, %f702, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f705, %f704;
|
|
add.f32 %f706, %f703, %f703;
|
|
mul.f32 %f707, %f706, %f705;
|
|
mul.f32 %f708, %f707, %f707;
|
|
mov.f32 %f709, 0f3C4CAF63;
|
|
mov.f32 %f710, 0f3B18F0FE;
|
|
fma.rn.f32 %f711, %f710, %f708, %f709;
|
|
mov.f32 %f712, 0f3DAAAABD;
|
|
fma.rn.f32 %f713, %f711, %f708, %f712;
|
|
mul.rn.f32 %f714, %f713, %f708;
|
|
mul.rn.f32 %f715, %f714, %f707;
|
|
sub.f32 %f716, %f703, %f707;
|
|
add.f32 %f717, %f716, %f716;
|
|
neg.f32 %f718, %f707;
|
|
fma.rn.f32 %f719, %f718, %f703, %f717;
|
|
mul.rn.f32 %f720, %f705, %f719;
|
|
add.f32 %f721, %f715, %f707;
|
|
sub.f32 %f722, %f707, %f721;
|
|
add.f32 %f723, %f715, %f722;
|
|
add.f32 %f724, %f720, %f723;
|
|
add.f32 %f725, %f721, %f724;
|
|
sub.f32 %f726, %f721, %f725;
|
|
add.f32 %f727, %f724, %f726;
|
|
mov.f32 %f728, 0f3F317200;
|
|
mul.rn.f32 %f729, %f701, %f728;
|
|
mov.f32 %f730, 0f35BFBE8E;
|
|
mul.rn.f32 %f731, %f701, %f730;
|
|
add.f32 %f732, %f729, %f725;
|
|
sub.f32 %f733, %f729, %f732;
|
|
add.f32 %f734, %f725, %f733;
|
|
add.f32 %f735, %f727, %f734;
|
|
add.f32 %f736, %f731, %f735;
|
|
add.f32 %f737, %f732, %f736;
|
|
sub.f32 %f738, %f732, %f737;
|
|
add.f32 %f739, %f736, %f738;
|
|
mul.rn.f32 %f740, %f691, %f737;
|
|
neg.f32 %f741, %f740;
|
|
fma.rn.f32 %f742, %f691, %f737, %f741;
|
|
fma.rn.f32 %f743, %f691, %f739, %f742;
|
|
mov.f32 %f744, 0f00000000;
|
|
fma.rn.f32 %f745, %f744, %f737, %f743;
|
|
add.rn.f32 %f746, %f740, %f745;
|
|
neg.f32 %f747, %f746;
|
|
add.rn.f32 %f748, %f740, %f747;
|
|
add.rn.f32 %f749, %f748, %f745;
|
|
mov.b32 %r286, %f746;
|
|
setp.eq.s32 %p57, %r286, 1118925336;
|
|
add.s32 %r287, %r286, -1;
|
|
mov.b32 %f750, %r287;
|
|
add.f32 %f751, %f749, 0f37000000;
|
|
selp.f32 %f158, %f751, %f749, %p57;
|
|
selp.f32 %f752, %f750, %f746, %p57;
|
|
mov.f32 %f753, 0f3FB8AA3B;
|
|
mul.rn.f32 %f754, %f752, %f753;
|
|
cvt.rzi.f32.f32 %f755, %f754;
|
|
abs.f32 %f756, %f755;
|
|
setp.gt.f32 %p58, %f756, 0f42FC0000;
|
|
mov.b32 %r288, %f755;
|
|
and.b32 %r289, %r288, -2147483648;
|
|
or.b32 %r290, %r289, 1123811328;
|
|
mov.b32 %f757, %r290;
|
|
selp.f32 %f758, %f757, %f755, %p58;
|
|
mov.f32 %f759, 0fBF317218;
|
|
fma.rn.f32 %f760, %f758, %f759, %f752;
|
|
mov.f32 %f761, 0f3102E308;
|
|
fma.rn.f32 %f762, %f758, %f761, %f760;
|
|
mul.f32 %f763, %f762, 0f3FB8AA3B;
|
|
add.f32 %f764, %f758, 0f4B40007F;
|
|
mov.b32 %r291, %f764;
|
|
shl.b32 %r292, %r291, 23;
|
|
mov.b32 %f765, %r292;
|
|
ex2.approx.ftz.f32 %f766, %f763;
|
|
mul.f32 %f159, %f766, %f765;
|
|
setp.eq.f32 %p59, %f159, 0f7F800000;
|
|
mov.f32 %f1401, 0f7F800000;
|
|
@%p59 bra $L__BB0_50;
|
|
|
|
fma.rn.f32 %f1401, %f159, %f158, %f159;
|
|
|
|
$L__BB0_50:
|
|
setp.lt.f32 %p60, %f155, 0f00000000;
|
|
setp.eq.f32 %p61, %f156, 0f3F800000;
|
|
and.pred %p3, %p60, %p61;
|
|
setp.eq.f32 %p62, %f155, 0f00000000;
|
|
@%p62 bra $L__BB0_54;
|
|
bra.uni $L__BB0_51;
|
|
|
|
$L__BB0_54:
|
|
add.f32 %f771, %f155, %f155;
|
|
selp.f32 %f1403, %f771, 0f00000000, %p61;
|
|
bra.uni $L__BB0_55;
|
|
|
|
$L__BB0_148:
|
|
mov.f32 %f1337, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f1337;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f1337;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f1337;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs60, 0;
|
|
st.global.v4.u16 [%rd12], {%rs57, %rs58, %rs59, %rs60};
|
|
|
|
$L__BB0_149:
|
|
and.b32 %r411, %r21, 64;
|
|
setp.eq.s32 %p177, %r411, 0;
|
|
@%p177 bra $L__BB0_151;
|
|
|
|
ld.const.u64 %rd49, [params+208];
|
|
cvta.to.global.u64 %rd50, %rd49;
|
|
ld.const.u32 %r412, [params+200];
|
|
mad.lo.s32 %r413, %r412, %r4, %r3;
|
|
mul.wide.u32 %rd51, %r413, 4;
|
|
add.s64 %rd52, %rd50, %rd51;
|
|
mov.u16 %rs61, 0;
|
|
st.global.v4.u8 [%rd52], {%rs61, %rs61, %rs61, %rs61};
|
|
bra.uni $L__BB0_151;
|
|
|
|
$L__BB0_51:
|
|
mov.b32 %r293, %f1401;
|
|
xor.b32 %r294, %r293, -2147483648;
|
|
mov.b32 %f767, %r294;
|
|
selp.f32 %f1403, %f767, %f1401, %p3;
|
|
setp.geu.f32 %p63, %f155, 0f00000000;
|
|
@%p63 bra $L__BB0_55;
|
|
|
|
mov.f32 %f768, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f769, %f768;
|
|
setp.eq.f32 %p64, %f769, 0f3EE8BA2E;
|
|
@%p64 bra $L__BB0_55;
|
|
|
|
mov.f32 %f1403, 0f7FFFFFFF;
|
|
|
|
$L__BB0_55:
|
|
add.f32 %f772, %f157, 0f3EE8BA2E;
|
|
mov.b32 %r295, %f772;
|
|
setp.lt.s32 %p66, %r295, 2139095040;
|
|
@%p66 bra $L__BB0_60;
|
|
|
|
setp.gtu.f32 %p67, %f157, 0f7F800000;
|
|
@%p67 bra $L__BB0_59;
|
|
bra.uni $L__BB0_57;
|
|
|
|
$L__BB0_59:
|
|
add.f32 %f1403, %f155, 0f3EE8BA2E;
|
|
bra.uni $L__BB0_60;
|
|
|
|
$L__BB0_57:
|
|
setp.neu.f32 %p68, %f157, 0f7F800000;
|
|
@%p68 bra $L__BB0_60;
|
|
|
|
selp.f32 %f1403, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
$L__BB0_60:
|
|
mul.f32 %f773, %f1403, 0f437F0000;
|
|
setp.eq.f32 %p69, %f155, 0f3F800000;
|
|
selp.f32 %f774, 0f437F0000, %f773, %p69;
|
|
cvt.rzi.u32.f32 %r296, %f774;
|
|
shl.b64 %rd26, %rd8, 1;
|
|
add.s64 %rd27, %rd7, %rd26;
|
|
cvt.u16.u32 %rs18, %r296;
|
|
mov.u16 %rs19, 255;
|
|
st.global.v2.u8 [%rd27], {%rs18, %rs19};
|
|
|
|
$L__BB0_61:
|
|
and.b32 %r297, %r418, 1;
|
|
setp.eq.b32 %p70, %r297, 1;
|
|
mov.pred %p71, 0;
|
|
xor.pred %p72, %p70, %p71;
|
|
not.pred %p73, %p72;
|
|
@%p73 bra $L__BB0_135;
|
|
|
|
mov.f32 %f776, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f777, %f776;
|
|
add.f32 %f778, %f777, %f777;
|
|
mov.f32 %f779, 0f3EE66666;
|
|
sub.f32 %f780, %f779, %f778;
|
|
abs.f32 %f168, %f780;
|
|
abs.f32 %f169, %f1387;
|
|
setp.lt.f32 %p74, %f169, 0f00800000;
|
|
mul.f32 %f781, %f169, 0f4B800000;
|
|
selp.f32 %f782, %f781, %f169, %p74;
|
|
selp.f32 %f783, 0fC3170000, 0fC2FE0000, %p74;
|
|
mov.b32 %r298, %f782;
|
|
and.b32 %r299, %r298, 8388607;
|
|
or.b32 %r300, %r299, 1065353216;
|
|
mov.b32 %f784, %r300;
|
|
shr.u32 %r301, %r298, 23;
|
|
cvt.rn.f32.u32 %f785, %r301;
|
|
add.f32 %f786, %f783, %f785;
|
|
setp.gt.f32 %p75, %f784, 0f3FB504F3;
|
|
mul.f32 %f787, %f784, 0f3F000000;
|
|
add.f32 %f788, %f786, 0f3F800000;
|
|
selp.f32 %f789, %f788, %f786, %p75;
|
|
selp.f32 %f790, %f787, %f784, %p75;
|
|
add.f32 %f791, %f790, 0fBF800000;
|
|
add.f32 %f792, %f790, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f793, %f792;
|
|
add.f32 %f794, %f791, %f791;
|
|
mul.f32 %f795, %f794, %f793;
|
|
mul.f32 %f796, %f795, %f795;
|
|
mov.f32 %f797, 0f3C4CAF63;
|
|
mov.f32 %f798, 0f3B18F0FE;
|
|
fma.rn.f32 %f799, %f798, %f796, %f797;
|
|
mov.f32 %f800, 0f3DAAAABD;
|
|
fma.rn.f32 %f801, %f799, %f796, %f800;
|
|
mul.rn.f32 %f802, %f801, %f796;
|
|
mul.rn.f32 %f803, %f802, %f795;
|
|
sub.f32 %f804, %f791, %f795;
|
|
add.f32 %f805, %f804, %f804;
|
|
neg.f32 %f806, %f795;
|
|
fma.rn.f32 %f807, %f806, %f791, %f805;
|
|
mul.rn.f32 %f808, %f793, %f807;
|
|
add.f32 %f809, %f803, %f795;
|
|
sub.f32 %f810, %f795, %f809;
|
|
add.f32 %f811, %f803, %f810;
|
|
add.f32 %f812, %f808, %f811;
|
|
add.f32 %f813, %f809, %f812;
|
|
sub.f32 %f814, %f809, %f813;
|
|
add.f32 %f815, %f812, %f814;
|
|
mov.f32 %f816, 0f3F317200;
|
|
mul.rn.f32 %f817, %f789, %f816;
|
|
mov.f32 %f818, 0f35BFBE8E;
|
|
mul.rn.f32 %f819, %f789, %f818;
|
|
add.f32 %f820, %f817, %f813;
|
|
sub.f32 %f821, %f817, %f820;
|
|
add.f32 %f822, %f813, %f821;
|
|
add.f32 %f823, %f815, %f822;
|
|
add.f32 %f824, %f819, %f823;
|
|
add.f32 %f825, %f820, %f824;
|
|
sub.f32 %f826, %f820, %f825;
|
|
add.f32 %f827, %f824, %f826;
|
|
mul.rn.f32 %f828, %f779, %f825;
|
|
neg.f32 %f829, %f828;
|
|
fma.rn.f32 %f830, %f779, %f825, %f829;
|
|
fma.rn.f32 %f831, %f779, %f827, %f830;
|
|
mov.f32 %f832, 0f00000000;
|
|
fma.rn.f32 %f833, %f832, %f825, %f831;
|
|
add.rn.f32 %f834, %f828, %f833;
|
|
neg.f32 %f835, %f834;
|
|
add.rn.f32 %f836, %f828, %f835;
|
|
add.rn.f32 %f837, %f836, %f833;
|
|
mov.b32 %r302, %f834;
|
|
setp.eq.s32 %p76, %r302, 1118925336;
|
|
add.s32 %r303, %r302, -1;
|
|
mov.b32 %f838, %r303;
|
|
add.f32 %f839, %f837, 0f37000000;
|
|
selp.f32 %f170, %f839, %f837, %p76;
|
|
selp.f32 %f840, %f838, %f834, %p76;
|
|
mov.f32 %f841, 0f3FB8AA3B;
|
|
mul.rn.f32 %f842, %f840, %f841;
|
|
cvt.rzi.f32.f32 %f843, %f842;
|
|
abs.f32 %f844, %f843;
|
|
setp.gt.f32 %p77, %f844, 0f42FC0000;
|
|
mov.b32 %r304, %f843;
|
|
and.b32 %r305, %r304, -2147483648;
|
|
or.b32 %r306, %r305, 1123811328;
|
|
mov.b32 %f845, %r306;
|
|
selp.f32 %f846, %f845, %f843, %p77;
|
|
mov.f32 %f847, 0fBF317218;
|
|
fma.rn.f32 %f848, %f846, %f847, %f840;
|
|
mov.f32 %f849, 0f3102E308;
|
|
fma.rn.f32 %f850, %f846, %f849, %f848;
|
|
mul.f32 %f851, %f850, 0f3FB8AA3B;
|
|
add.f32 %f852, %f846, 0f4B40007F;
|
|
mov.b32 %r307, %f852;
|
|
shl.b32 %r308, %r307, 23;
|
|
mov.b32 %f853, %r308;
|
|
ex2.approx.ftz.f32 %f854, %f851;
|
|
mul.f32 %f171, %f854, %f853;
|
|
setp.eq.f32 %p78, %f171, 0f7F800000;
|
|
mov.f32 %f1404, 0f7F800000;
|
|
@%p78 bra $L__BB0_64;
|
|
|
|
fma.rn.f32 %f1404, %f171, %f170, %f171;
|
|
|
|
$L__BB0_64:
|
|
setp.lt.f32 %p79, %f1387, 0f00000000;
|
|
setp.eq.f32 %p80, %f168, 0f3F800000;
|
|
and.pred %p4, %p79, %p80;
|
|
setp.eq.f32 %p81, %f1387, 0f00000000;
|
|
@%p81 bra $L__BB0_68;
|
|
bra.uni $L__BB0_65;
|
|
|
|
$L__BB0_68:
|
|
add.f32 %f859, %f1387, %f1387;
|
|
selp.f32 %f1406, %f859, 0f00000000, %p80;
|
|
bra.uni $L__BB0_69;
|
|
|
|
$L__BB0_65:
|
|
mov.b32 %r309, %f1404;
|
|
xor.b32 %r310, %r309, -2147483648;
|
|
mov.b32 %f855, %r310;
|
|
selp.f32 %f1406, %f855, %f1404, %p4;
|
|
setp.geu.f32 %p82, %f1387, 0f00000000;
|
|
@%p82 bra $L__BB0_69;
|
|
|
|
mov.f32 %f856, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f857, %f856;
|
|
setp.eq.f32 %p83, %f857, 0f3EE66666;
|
|
@%p83 bra $L__BB0_69;
|
|
|
|
mov.f32 %f1406, 0f7FFFFFFF;
|
|
|
|
$L__BB0_69:
|
|
add.f32 %f860, %f169, 0f3EE66666;
|
|
mov.b32 %r311, %f860;
|
|
setp.lt.s32 %p85, %r311, 2139095040;
|
|
@%p85 bra $L__BB0_74;
|
|
|
|
setp.gtu.f32 %p86, %f169, 0f7F800000;
|
|
@%p86 bra $L__BB0_73;
|
|
bra.uni $L__BB0_71;
|
|
|
|
$L__BB0_73:
|
|
add.f32 %f1406, %f1387, 0f3EE66666;
|
|
bra.uni $L__BB0_74;
|
|
|
|
$L__BB0_71:
|
|
setp.neu.f32 %p87, %f169, 0f7F800000;
|
|
@%p87 bra $L__BB0_74;
|
|
|
|
selp.f32 %f1406, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
$L__BB0_74:
|
|
setp.eq.f32 %p88, %f1387, 0f3F800000;
|
|
selp.f32 %f180, 0f3F800000, %f1406, %p88;
|
|
abs.f32 %f181, %f1388;
|
|
setp.lt.f32 %p89, %f181, 0f00800000;
|
|
mul.f32 %f862, %f181, 0f4B800000;
|
|
selp.f32 %f863, %f862, %f181, %p89;
|
|
selp.f32 %f864, 0fC3170000, 0fC2FE0000, %p89;
|
|
mov.b32 %r312, %f863;
|
|
and.b32 %r313, %r312, 8388607;
|
|
or.b32 %r314, %r313, 1065353216;
|
|
mov.b32 %f865, %r314;
|
|
shr.u32 %r315, %r312, 23;
|
|
cvt.rn.f32.u32 %f866, %r315;
|
|
add.f32 %f867, %f864, %f866;
|
|
setp.gt.f32 %p90, %f865, 0f3FB504F3;
|
|
mul.f32 %f868, %f865, 0f3F000000;
|
|
add.f32 %f869, %f867, 0f3F800000;
|
|
selp.f32 %f870, %f869, %f867, %p90;
|
|
selp.f32 %f871, %f868, %f865, %p90;
|
|
add.f32 %f872, %f871, 0fBF800000;
|
|
add.f32 %f873, %f871, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f874, %f873;
|
|
add.f32 %f875, %f872, %f872;
|
|
mul.f32 %f876, %f875, %f874;
|
|
mul.f32 %f877, %f876, %f876;
|
|
mov.f32 %f878, 0f3C4CAF63;
|
|
mov.f32 %f879, 0f3B18F0FE;
|
|
fma.rn.f32 %f880, %f879, %f877, %f878;
|
|
mov.f32 %f881, 0f3DAAAABD;
|
|
fma.rn.f32 %f882, %f880, %f877, %f881;
|
|
mul.rn.f32 %f883, %f882, %f877;
|
|
mul.rn.f32 %f884, %f883, %f876;
|
|
sub.f32 %f885, %f872, %f876;
|
|
add.f32 %f886, %f885, %f885;
|
|
neg.f32 %f887, %f876;
|
|
fma.rn.f32 %f888, %f887, %f872, %f886;
|
|
mul.rn.f32 %f889, %f874, %f888;
|
|
add.f32 %f890, %f884, %f876;
|
|
sub.f32 %f891, %f876, %f890;
|
|
add.f32 %f892, %f884, %f891;
|
|
add.f32 %f893, %f889, %f892;
|
|
add.f32 %f894, %f890, %f893;
|
|
sub.f32 %f895, %f890, %f894;
|
|
add.f32 %f896, %f893, %f895;
|
|
mov.f32 %f897, 0f3F317200;
|
|
mul.rn.f32 %f898, %f870, %f897;
|
|
mov.f32 %f899, 0f35BFBE8E;
|
|
mul.rn.f32 %f900, %f870, %f899;
|
|
add.f32 %f901, %f898, %f894;
|
|
sub.f32 %f902, %f898, %f901;
|
|
add.f32 %f903, %f894, %f902;
|
|
add.f32 %f904, %f896, %f903;
|
|
add.f32 %f905, %f900, %f904;
|
|
add.f32 %f906, %f901, %f905;
|
|
sub.f32 %f907, %f901, %f906;
|
|
add.f32 %f908, %f905, %f907;
|
|
mov.f32 %f909, 0f3EE66666;
|
|
mul.rn.f32 %f910, %f909, %f906;
|
|
neg.f32 %f911, %f910;
|
|
fma.rn.f32 %f912, %f909, %f906, %f911;
|
|
fma.rn.f32 %f913, %f909, %f908, %f912;
|
|
mov.f32 %f914, 0f00000000;
|
|
fma.rn.f32 %f915, %f914, %f906, %f913;
|
|
add.rn.f32 %f916, %f910, %f915;
|
|
neg.f32 %f917, %f916;
|
|
add.rn.f32 %f918, %f910, %f917;
|
|
add.rn.f32 %f919, %f918, %f915;
|
|
mov.b32 %r316, %f916;
|
|
setp.eq.s32 %p91, %r316, 1118925336;
|
|
add.s32 %r317, %r316, -1;
|
|
mov.b32 %f920, %r317;
|
|
add.f32 %f921, %f919, 0f37000000;
|
|
selp.f32 %f182, %f921, %f919, %p91;
|
|
selp.f32 %f922, %f920, %f916, %p91;
|
|
mov.f32 %f923, 0f3FB8AA3B;
|
|
mul.rn.f32 %f924, %f922, %f923;
|
|
cvt.rzi.f32.f32 %f925, %f924;
|
|
abs.f32 %f926, %f925;
|
|
setp.gt.f32 %p92, %f926, 0f42FC0000;
|
|
mov.b32 %r318, %f925;
|
|
and.b32 %r319, %r318, -2147483648;
|
|
or.b32 %r320, %r319, 1123811328;
|
|
mov.b32 %f927, %r320;
|
|
selp.f32 %f928, %f927, %f925, %p92;
|
|
mov.f32 %f929, 0fBF317218;
|
|
fma.rn.f32 %f930, %f928, %f929, %f922;
|
|
mov.f32 %f931, 0f3102E308;
|
|
fma.rn.f32 %f932, %f928, %f931, %f930;
|
|
mul.f32 %f933, %f932, 0f3FB8AA3B;
|
|
add.f32 %f934, %f928, 0f4B40007F;
|
|
mov.b32 %r321, %f934;
|
|
shl.b32 %r322, %r321, 23;
|
|
mov.b32 %f935, %r322;
|
|
ex2.approx.ftz.f32 %f936, %f933;
|
|
mul.f32 %f183, %f936, %f935;
|
|
setp.eq.f32 %p93, %f183, 0f7F800000;
|
|
mov.f32 %f1407, 0f7F800000;
|
|
@%p93 bra $L__BB0_76;
|
|
|
|
fma.rn.f32 %f1407, %f183, %f182, %f183;
|
|
|
|
$L__BB0_76:
|
|
setp.lt.f32 %p94, %f1388, 0f00000000;
|
|
and.pred %p5, %p94, %p80;
|
|
setp.eq.f32 %p96, %f1388, 0f00000000;
|
|
@%p96 bra $L__BB0_80;
|
|
bra.uni $L__BB0_77;
|
|
|
|
$L__BB0_80:
|
|
add.f32 %f941, %f1388, %f1388;
|
|
selp.f32 %f1409, %f941, 0f00000000, %p80;
|
|
bra.uni $L__BB0_81;
|
|
|
|
$L__BB0_77:
|
|
mov.b32 %r323, %f1407;
|
|
xor.b32 %r324, %r323, -2147483648;
|
|
mov.b32 %f937, %r324;
|
|
selp.f32 %f1409, %f937, %f1407, %p5;
|
|
setp.geu.f32 %p97, %f1388, 0f00000000;
|
|
@%p97 bra $L__BB0_81;
|
|
|
|
mov.f32 %f938, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f939, %f938;
|
|
setp.eq.f32 %p98, %f939, 0f3EE66666;
|
|
@%p98 bra $L__BB0_81;
|
|
|
|
mov.f32 %f1409, 0f7FFFFFFF;
|
|
|
|
$L__BB0_81:
|
|
add.f32 %f942, %f181, 0f3EE66666;
|
|
mov.b32 %r325, %f942;
|
|
setp.lt.s32 %p100, %r325, 2139095040;
|
|
@%p100 bra $L__BB0_86;
|
|
|
|
setp.gtu.f32 %p101, %f181, 0f7F800000;
|
|
@%p101 bra $L__BB0_85;
|
|
bra.uni $L__BB0_83;
|
|
|
|
$L__BB0_85:
|
|
add.f32 %f1409, %f1388, 0f3EE66666;
|
|
bra.uni $L__BB0_86;
|
|
|
|
$L__BB0_83:
|
|
setp.neu.f32 %p102, %f181, 0f7F800000;
|
|
@%p102 bra $L__BB0_86;
|
|
|
|
selp.f32 %f1409, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
$L__BB0_86:
|
|
setp.eq.f32 %p103, %f1388, 0f3F800000;
|
|
selp.f32 %f192, 0f3F800000, %f1409, %p103;
|
|
abs.f32 %f193, %f1389;
|
|
setp.lt.f32 %p104, %f193, 0f00800000;
|
|
mul.f32 %f944, %f193, 0f4B800000;
|
|
selp.f32 %f945, %f944, %f193, %p104;
|
|
selp.f32 %f946, 0fC3170000, 0fC2FE0000, %p104;
|
|
mov.b32 %r326, %f945;
|
|
and.b32 %r327, %r326, 8388607;
|
|
or.b32 %r328, %r327, 1065353216;
|
|
mov.b32 %f947, %r328;
|
|
shr.u32 %r329, %r326, 23;
|
|
cvt.rn.f32.u32 %f948, %r329;
|
|
add.f32 %f949, %f946, %f948;
|
|
setp.gt.f32 %p105, %f947, 0f3FB504F3;
|
|
mul.f32 %f950, %f947, 0f3F000000;
|
|
add.f32 %f951, %f949, 0f3F800000;
|
|
selp.f32 %f952, %f951, %f949, %p105;
|
|
selp.f32 %f953, %f950, %f947, %p105;
|
|
add.f32 %f954, %f953, 0fBF800000;
|
|
add.f32 %f955, %f953, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f956, %f955;
|
|
add.f32 %f957, %f954, %f954;
|
|
mul.f32 %f958, %f957, %f956;
|
|
mul.f32 %f959, %f958, %f958;
|
|
mov.f32 %f960, 0f3C4CAF63;
|
|
mov.f32 %f961, 0f3B18F0FE;
|
|
fma.rn.f32 %f962, %f961, %f959, %f960;
|
|
mov.f32 %f963, 0f3DAAAABD;
|
|
fma.rn.f32 %f964, %f962, %f959, %f963;
|
|
mul.rn.f32 %f965, %f964, %f959;
|
|
mul.rn.f32 %f966, %f965, %f958;
|
|
sub.f32 %f967, %f954, %f958;
|
|
add.f32 %f968, %f967, %f967;
|
|
neg.f32 %f969, %f958;
|
|
fma.rn.f32 %f970, %f969, %f954, %f968;
|
|
mul.rn.f32 %f971, %f956, %f970;
|
|
add.f32 %f972, %f966, %f958;
|
|
sub.f32 %f973, %f958, %f972;
|
|
add.f32 %f974, %f966, %f973;
|
|
add.f32 %f975, %f971, %f974;
|
|
add.f32 %f976, %f972, %f975;
|
|
sub.f32 %f977, %f972, %f976;
|
|
add.f32 %f978, %f975, %f977;
|
|
mov.f32 %f979, 0f3F317200;
|
|
mul.rn.f32 %f980, %f952, %f979;
|
|
mov.f32 %f981, 0f35BFBE8E;
|
|
mul.rn.f32 %f982, %f952, %f981;
|
|
add.f32 %f983, %f980, %f976;
|
|
sub.f32 %f984, %f980, %f983;
|
|
add.f32 %f985, %f976, %f984;
|
|
add.f32 %f986, %f978, %f985;
|
|
add.f32 %f987, %f982, %f986;
|
|
add.f32 %f988, %f983, %f987;
|
|
sub.f32 %f989, %f983, %f988;
|
|
add.f32 %f990, %f987, %f989;
|
|
mov.f32 %f991, 0f3EE66666;
|
|
mul.rn.f32 %f992, %f991, %f988;
|
|
neg.f32 %f993, %f992;
|
|
fma.rn.f32 %f994, %f991, %f988, %f993;
|
|
fma.rn.f32 %f995, %f991, %f990, %f994;
|
|
mov.f32 %f996, 0f00000000;
|
|
fma.rn.f32 %f997, %f996, %f988, %f995;
|
|
add.rn.f32 %f998, %f992, %f997;
|
|
neg.f32 %f999, %f998;
|
|
add.rn.f32 %f1000, %f992, %f999;
|
|
add.rn.f32 %f1001, %f1000, %f997;
|
|
mov.b32 %r330, %f998;
|
|
setp.eq.s32 %p106, %r330, 1118925336;
|
|
add.s32 %r331, %r330, -1;
|
|
mov.b32 %f1002, %r331;
|
|
add.f32 %f1003, %f1001, 0f37000000;
|
|
selp.f32 %f194, %f1003, %f1001, %p106;
|
|
selp.f32 %f1004, %f1002, %f998, %p106;
|
|
mov.f32 %f1005, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1006, %f1004, %f1005;
|
|
cvt.rzi.f32.f32 %f1007, %f1006;
|
|
abs.f32 %f1008, %f1007;
|
|
setp.gt.f32 %p107, %f1008, 0f42FC0000;
|
|
mov.b32 %r332, %f1007;
|
|
and.b32 %r333, %r332, -2147483648;
|
|
or.b32 %r334, %r333, 1123811328;
|
|
mov.b32 %f1009, %r334;
|
|
selp.f32 %f1010, %f1009, %f1007, %p107;
|
|
mov.f32 %f1011, 0fBF317218;
|
|
fma.rn.f32 %f1012, %f1010, %f1011, %f1004;
|
|
mov.f32 %f1013, 0f3102E308;
|
|
fma.rn.f32 %f1014, %f1010, %f1013, %f1012;
|
|
mul.f32 %f1015, %f1014, 0f3FB8AA3B;
|
|
add.f32 %f1016, %f1010, 0f4B40007F;
|
|
mov.b32 %r335, %f1016;
|
|
shl.b32 %r336, %r335, 23;
|
|
mov.b32 %f1017, %r336;
|
|
ex2.approx.ftz.f32 %f1018, %f1015;
|
|
mul.f32 %f195, %f1018, %f1017;
|
|
setp.eq.f32 %p108, %f195, 0f7F800000;
|
|
mov.f32 %f1410, 0f7F800000;
|
|
@%p108 bra $L__BB0_88;
|
|
|
|
fma.rn.f32 %f1410, %f195, %f194, %f195;
|
|
|
|
$L__BB0_88:
|
|
setp.lt.f32 %p109, %f1389, 0f00000000;
|
|
and.pred %p6, %p109, %p80;
|
|
setp.eq.f32 %p111, %f1389, 0f00000000;
|
|
@%p111 bra $L__BB0_92;
|
|
bra.uni $L__BB0_89;
|
|
|
|
$L__BB0_92:
|
|
add.f32 %f1023, %f1389, %f1389;
|
|
selp.f32 %f1412, %f1023, 0f00000000, %p80;
|
|
bra.uni $L__BB0_93;
|
|
|
|
$L__BB0_89:
|
|
mov.b32 %r337, %f1410;
|
|
xor.b32 %r338, %r337, -2147483648;
|
|
mov.b32 %f1019, %r338;
|
|
selp.f32 %f1412, %f1019, %f1410, %p6;
|
|
setp.geu.f32 %p112, %f1389, 0f00000000;
|
|
@%p112 bra $L__BB0_93;
|
|
|
|
mov.f32 %f1020, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f1021, %f1020;
|
|
setp.eq.f32 %p113, %f1021, 0f3EE66666;
|
|
@%p113 bra $L__BB0_93;
|
|
|
|
mov.f32 %f1412, 0f7FFFFFFF;
|
|
|
|
$L__BB0_93:
|
|
add.f32 %f1024, %f193, 0f3EE66666;
|
|
mov.b32 %r339, %f1024;
|
|
setp.lt.s32 %p115, %r339, 2139095040;
|
|
@%p115 bra $L__BB0_98;
|
|
|
|
setp.gtu.f32 %p116, %f193, 0f7F800000;
|
|
@%p116 bra $L__BB0_97;
|
|
bra.uni $L__BB0_95;
|
|
|
|
$L__BB0_97:
|
|
add.f32 %f1412, %f1389, 0f3EE66666;
|
|
bra.uni $L__BB0_98;
|
|
|
|
$L__BB0_95:
|
|
setp.neu.f32 %p117, %f193, 0f7F800000;
|
|
@%p117 bra $L__BB0_98;
|
|
|
|
selp.f32 %f1412, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
$L__BB0_98:
|
|
setp.eq.f32 %p118, %f1389, 0f3F800000;
|
|
mov.f32 %f1026, 0f3F800000;
|
|
selp.f32 %f1027, 0f3F800000, %f1412, %p118;
|
|
ld.const.u64 %rd28, [params+144];
|
|
cvta.to.global.u64 %rd9, %rd28;
|
|
ld.const.u32 %r340, [params+136];
|
|
mad.lo.s32 %r341, %r340, %r4, %r3;
|
|
cvt.u64.u32 %rd10, %r341;
|
|
min.f32 %f1028, %f180, %f1026;
|
|
mov.f32 %f1029, 0f00000000;
|
|
max.f32 %f204, %f1029, %f1028;
|
|
min.f32 %f1030, %f192, %f1026;
|
|
max.f32 %f205, %f1029, %f1030;
|
|
min.f32 %f1031, %f1027, %f1026;
|
|
max.f32 %f206, %f1029, %f1031;
|
|
mov.f32 %f1032, 0f3E555555;
|
|
cvt.rzi.f32.f32 %f1033, %f1032;
|
|
add.f32 %f1034, %f1033, %f1033;
|
|
mov.f32 %f1035, 0f3ED55555;
|
|
sub.f32 %f1036, %f1035, %f1034;
|
|
abs.f32 %f207, %f1036;
|
|
abs.f32 %f208, %f204;
|
|
setp.lt.f32 %p119, %f208, 0f00800000;
|
|
mul.f32 %f1037, %f208, 0f4B800000;
|
|
selp.f32 %f1038, %f1037, %f208, %p119;
|
|
selp.f32 %f1039, 0fC3170000, 0fC2FE0000, %p119;
|
|
mov.b32 %r342, %f1038;
|
|
and.b32 %r343, %r342, 8388607;
|
|
or.b32 %r344, %r343, 1065353216;
|
|
mov.b32 %f1040, %r344;
|
|
shr.u32 %r345, %r342, 23;
|
|
cvt.rn.f32.u32 %f1041, %r345;
|
|
add.f32 %f1042, %f1039, %f1041;
|
|
setp.gt.f32 %p120, %f1040, 0f3FB504F3;
|
|
mul.f32 %f1043, %f1040, 0f3F000000;
|
|
add.f32 %f1044, %f1042, 0f3F800000;
|
|
selp.f32 %f1045, %f1044, %f1042, %p120;
|
|
selp.f32 %f1046, %f1043, %f1040, %p120;
|
|
add.f32 %f1047, %f1046, 0fBF800000;
|
|
add.f32 %f1048, %f1046, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1049, %f1048;
|
|
add.f32 %f1050, %f1047, %f1047;
|
|
mul.f32 %f1051, %f1050, %f1049;
|
|
mul.f32 %f1052, %f1051, %f1051;
|
|
mov.f32 %f1053, 0f3C4CAF63;
|
|
mov.f32 %f1054, 0f3B18F0FE;
|
|
fma.rn.f32 %f1055, %f1054, %f1052, %f1053;
|
|
mov.f32 %f1056, 0f3DAAAABD;
|
|
fma.rn.f32 %f1057, %f1055, %f1052, %f1056;
|
|
mul.rn.f32 %f1058, %f1057, %f1052;
|
|
mul.rn.f32 %f1059, %f1058, %f1051;
|
|
sub.f32 %f1060, %f1047, %f1051;
|
|
add.f32 %f1061, %f1060, %f1060;
|
|
neg.f32 %f1062, %f1051;
|
|
fma.rn.f32 %f1063, %f1062, %f1047, %f1061;
|
|
mul.rn.f32 %f1064, %f1049, %f1063;
|
|
add.f32 %f1065, %f1059, %f1051;
|
|
sub.f32 %f1066, %f1051, %f1065;
|
|
add.f32 %f1067, %f1059, %f1066;
|
|
add.f32 %f1068, %f1064, %f1067;
|
|
add.f32 %f1069, %f1065, %f1068;
|
|
sub.f32 %f1070, %f1065, %f1069;
|
|
add.f32 %f1071, %f1068, %f1070;
|
|
mov.f32 %f1072, 0f3F317200;
|
|
mul.rn.f32 %f1073, %f1045, %f1072;
|
|
mov.f32 %f1074, 0f35BFBE8E;
|
|
mul.rn.f32 %f1075, %f1045, %f1074;
|
|
add.f32 %f1076, %f1073, %f1069;
|
|
sub.f32 %f1077, %f1073, %f1076;
|
|
add.f32 %f1078, %f1069, %f1077;
|
|
add.f32 %f1079, %f1071, %f1078;
|
|
add.f32 %f1080, %f1075, %f1079;
|
|
add.f32 %f1081, %f1076, %f1080;
|
|
sub.f32 %f1082, %f1076, %f1081;
|
|
add.f32 %f1083, %f1080, %f1082;
|
|
mul.rn.f32 %f1084, %f1035, %f1081;
|
|
neg.f32 %f1085, %f1084;
|
|
fma.rn.f32 %f1086, %f1035, %f1081, %f1085;
|
|
fma.rn.f32 %f1087, %f1035, %f1083, %f1086;
|
|
fma.rn.f32 %f1088, %f1029, %f1081, %f1087;
|
|
add.rn.f32 %f1089, %f1084, %f1088;
|
|
neg.f32 %f1090, %f1089;
|
|
add.rn.f32 %f1091, %f1084, %f1090;
|
|
add.rn.f32 %f1092, %f1091, %f1088;
|
|
mov.b32 %r346, %f1089;
|
|
setp.eq.s32 %p121, %r346, 1118925336;
|
|
add.s32 %r347, %r346, -1;
|
|
mov.b32 %f1093, %r347;
|
|
add.f32 %f1094, %f1092, 0f37000000;
|
|
selp.f32 %f209, %f1094, %f1092, %p121;
|
|
selp.f32 %f1095, %f1093, %f1089, %p121;
|
|
mov.f32 %f1096, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1097, %f1095, %f1096;
|
|
cvt.rzi.f32.f32 %f1098, %f1097;
|
|
abs.f32 %f1099, %f1098;
|
|
setp.gt.f32 %p122, %f1099, 0f42FC0000;
|
|
mov.b32 %r348, %f1098;
|
|
and.b32 %r349, %r348, -2147483648;
|
|
or.b32 %r350, %r349, 1123811328;
|
|
mov.b32 %f1100, %r350;
|
|
selp.f32 %f1101, %f1100, %f1098, %p122;
|
|
mov.f32 %f1102, 0fBF317218;
|
|
fma.rn.f32 %f1103, %f1101, %f1102, %f1095;
|
|
mov.f32 %f1104, 0f3102E308;
|
|
fma.rn.f32 %f1105, %f1101, %f1104, %f1103;
|
|
mul.f32 %f1106, %f1105, 0f3FB8AA3B;
|
|
add.f32 %f1107, %f1101, 0f4B40007F;
|
|
mov.b32 %r351, %f1107;
|
|
shl.b32 %r352, %r351, 23;
|
|
mov.b32 %f1108, %r352;
|
|
ex2.approx.ftz.f32 %f1109, %f1106;
|
|
mul.f32 %f210, %f1109, %f1108;
|
|
setp.eq.f32 %p123, %f210, 0f7F800000;
|
|
mov.f32 %f1413, 0f7F800000;
|
|
@%p123 bra $L__BB0_100;
|
|
|
|
fma.rn.f32 %f1413, %f210, %f209, %f210;
|
|
|
|
$L__BB0_100:
|
|
setp.lt.f32 %p124, %f204, 0f00000000;
|
|
setp.eq.f32 %p125, %f207, 0f3F800000;
|
|
and.pred %p7, %p124, %p125;
|
|
setp.eq.f32 %p126, %f204, 0f00000000;
|
|
@%p126 bra $L__BB0_104;
|
|
bra.uni $L__BB0_101;
|
|
|
|
$L__BB0_104:
|
|
add.f32 %f1114, %f204, %f204;
|
|
selp.f32 %f1415, %f1114, 0f00000000, %p125;
|
|
bra.uni $L__BB0_105;
|
|
|
|
$L__BB0_101:
|
|
mov.b32 %r353, %f1413;
|
|
xor.b32 %r354, %r353, -2147483648;
|
|
mov.b32 %f1110, %r354;
|
|
selp.f32 %f1415, %f1110, %f1413, %p7;
|
|
setp.geu.f32 %p127, %f204, 0f00000000;
|
|
@%p127 bra $L__BB0_105;
|
|
|
|
mov.f32 %f1111, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1112, %f1111;
|
|
setp.eq.f32 %p128, %f1112, 0f3ED55555;
|
|
@%p128 bra $L__BB0_105;
|
|
|
|
mov.f32 %f1415, 0f7FFFFFFF;
|
|
|
|
$L__BB0_105:
|
|
add.f32 %f1115, %f208, 0f3ED55555;
|
|
mov.b32 %r355, %f1115;
|
|
setp.lt.s32 %p130, %r355, 2139095040;
|
|
@%p130 bra $L__BB0_110;
|
|
|
|
setp.gtu.f32 %p131, %f208, 0f7F800000;
|
|
@%p131 bra $L__BB0_109;
|
|
bra.uni $L__BB0_107;
|
|
|
|
$L__BB0_109:
|
|
add.f32 %f1415, %f204, 0f3ED55555;
|
|
bra.uni $L__BB0_110;
|
|
|
|
$L__BB0_107:
|
|
setp.neu.f32 %p132, %f208, 0f7F800000;
|
|
@%p132 bra $L__BB0_110;
|
|
|
|
selp.f32 %f1415, 0fFF800000, 0f7F800000, %p7;
|
|
|
|
$L__BB0_110:
|
|
abs.f32 %f219, %f205;
|
|
setp.lt.f32 %p133, %f219, 0f00800000;
|
|
mul.f32 %f1117, %f219, 0f4B800000;
|
|
selp.f32 %f1118, %f1117, %f219, %p133;
|
|
selp.f32 %f1119, 0fC3170000, 0fC2FE0000, %p133;
|
|
mov.b32 %r356, %f1118;
|
|
and.b32 %r357, %r356, 8388607;
|
|
or.b32 %r358, %r357, 1065353216;
|
|
mov.b32 %f1120, %r358;
|
|
shr.u32 %r359, %r356, 23;
|
|
cvt.rn.f32.u32 %f1121, %r359;
|
|
add.f32 %f1122, %f1119, %f1121;
|
|
setp.gt.f32 %p134, %f1120, 0f3FB504F3;
|
|
mul.f32 %f1123, %f1120, 0f3F000000;
|
|
add.f32 %f1124, %f1122, 0f3F800000;
|
|
selp.f32 %f1125, %f1124, %f1122, %p134;
|
|
selp.f32 %f1126, %f1123, %f1120, %p134;
|
|
add.f32 %f1127, %f1126, 0fBF800000;
|
|
add.f32 %f1128, %f1126, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1129, %f1128;
|
|
add.f32 %f1130, %f1127, %f1127;
|
|
mul.f32 %f1131, %f1130, %f1129;
|
|
mul.f32 %f1132, %f1131, %f1131;
|
|
mov.f32 %f1133, 0f3C4CAF63;
|
|
mov.f32 %f1134, 0f3B18F0FE;
|
|
fma.rn.f32 %f1135, %f1134, %f1132, %f1133;
|
|
mov.f32 %f1136, 0f3DAAAABD;
|
|
fma.rn.f32 %f1137, %f1135, %f1132, %f1136;
|
|
mul.rn.f32 %f1138, %f1137, %f1132;
|
|
mul.rn.f32 %f1139, %f1138, %f1131;
|
|
sub.f32 %f1140, %f1127, %f1131;
|
|
add.f32 %f1141, %f1140, %f1140;
|
|
neg.f32 %f1142, %f1131;
|
|
fma.rn.f32 %f1143, %f1142, %f1127, %f1141;
|
|
mul.rn.f32 %f1144, %f1129, %f1143;
|
|
add.f32 %f1145, %f1139, %f1131;
|
|
sub.f32 %f1146, %f1131, %f1145;
|
|
add.f32 %f1147, %f1139, %f1146;
|
|
add.f32 %f1148, %f1144, %f1147;
|
|
add.f32 %f1149, %f1145, %f1148;
|
|
sub.f32 %f1150, %f1145, %f1149;
|
|
add.f32 %f1151, %f1148, %f1150;
|
|
mov.f32 %f1152, 0f3F317200;
|
|
mul.rn.f32 %f1153, %f1125, %f1152;
|
|
mov.f32 %f1154, 0f35BFBE8E;
|
|
mul.rn.f32 %f1155, %f1125, %f1154;
|
|
add.f32 %f1156, %f1153, %f1149;
|
|
sub.f32 %f1157, %f1153, %f1156;
|
|
add.f32 %f1158, %f1149, %f1157;
|
|
add.f32 %f1159, %f1151, %f1158;
|
|
add.f32 %f1160, %f1155, %f1159;
|
|
add.f32 %f1161, %f1156, %f1160;
|
|
sub.f32 %f1162, %f1156, %f1161;
|
|
add.f32 %f1163, %f1160, %f1162;
|
|
mov.f32 %f1164, 0f3ED55555;
|
|
mul.rn.f32 %f1165, %f1164, %f1161;
|
|
neg.f32 %f1166, %f1165;
|
|
fma.rn.f32 %f1167, %f1164, %f1161, %f1166;
|
|
fma.rn.f32 %f1168, %f1164, %f1163, %f1167;
|
|
mov.f32 %f1169, 0f00000000;
|
|
fma.rn.f32 %f1170, %f1169, %f1161, %f1168;
|
|
add.rn.f32 %f1171, %f1165, %f1170;
|
|
neg.f32 %f1172, %f1171;
|
|
add.rn.f32 %f1173, %f1165, %f1172;
|
|
add.rn.f32 %f1174, %f1173, %f1170;
|
|
mov.b32 %r360, %f1171;
|
|
setp.eq.s32 %p135, %r360, 1118925336;
|
|
add.s32 %r361, %r360, -1;
|
|
mov.b32 %f1175, %r361;
|
|
add.f32 %f1176, %f1174, 0f37000000;
|
|
selp.f32 %f220, %f1176, %f1174, %p135;
|
|
selp.f32 %f1177, %f1175, %f1171, %p135;
|
|
mov.f32 %f1178, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1179, %f1177, %f1178;
|
|
cvt.rzi.f32.f32 %f1180, %f1179;
|
|
abs.f32 %f1181, %f1180;
|
|
setp.gt.f32 %p136, %f1181, 0f42FC0000;
|
|
mov.b32 %r362, %f1180;
|
|
and.b32 %r363, %r362, -2147483648;
|
|
or.b32 %r364, %r363, 1123811328;
|
|
mov.b32 %f1182, %r364;
|
|
selp.f32 %f1183, %f1182, %f1180, %p136;
|
|
mov.f32 %f1184, 0fBF317218;
|
|
fma.rn.f32 %f1185, %f1183, %f1184, %f1177;
|
|
mov.f32 %f1186, 0f3102E308;
|
|
fma.rn.f32 %f1187, %f1183, %f1186, %f1185;
|
|
mul.f32 %f1188, %f1187, 0f3FB8AA3B;
|
|
add.f32 %f1189, %f1183, 0f4B40007F;
|
|
mov.b32 %r365, %f1189;
|
|
shl.b32 %r366, %r365, 23;
|
|
mov.b32 %f1190, %r366;
|
|
ex2.approx.ftz.f32 %f1191, %f1188;
|
|
mul.f32 %f221, %f1191, %f1190;
|
|
setp.eq.f32 %p137, %f221, 0f7F800000;
|
|
mov.f32 %f1416, 0f7F800000;
|
|
@%p137 bra $L__BB0_112;
|
|
|
|
fma.rn.f32 %f1416, %f221, %f220, %f221;
|
|
|
|
$L__BB0_112:
|
|
setp.lt.f32 %p138, %f205, 0f00000000;
|
|
and.pred %p8, %p138, %p125;
|
|
setp.eq.f32 %p140, %f205, 0f00000000;
|
|
@%p140 bra $L__BB0_116;
|
|
bra.uni $L__BB0_113;
|
|
|
|
$L__BB0_116:
|
|
add.f32 %f1196, %f205, %f205;
|
|
selp.f32 %f1418, %f1196, 0f00000000, %p125;
|
|
bra.uni $L__BB0_117;
|
|
|
|
$L__BB0_113:
|
|
mov.b32 %r367, %f1416;
|
|
xor.b32 %r368, %r367, -2147483648;
|
|
mov.b32 %f1192, %r368;
|
|
selp.f32 %f1418, %f1192, %f1416, %p8;
|
|
setp.geu.f32 %p141, %f205, 0f00000000;
|
|
@%p141 bra $L__BB0_117;
|
|
|
|
mov.f32 %f1193, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1194, %f1193;
|
|
setp.eq.f32 %p142, %f1194, 0f3ED55555;
|
|
@%p142 bra $L__BB0_117;
|
|
|
|
mov.f32 %f1418, 0f7FFFFFFF;
|
|
|
|
$L__BB0_117:
|
|
add.f32 %f1197, %f219, 0f3ED55555;
|
|
mov.b32 %r369, %f1197;
|
|
setp.lt.s32 %p144, %r369, 2139095040;
|
|
@%p144 bra $L__BB0_122;
|
|
|
|
setp.gtu.f32 %p145, %f219, 0f7F800000;
|
|
@%p145 bra $L__BB0_121;
|
|
bra.uni $L__BB0_119;
|
|
|
|
$L__BB0_121:
|
|
add.f32 %f1418, %f205, 0f3ED55555;
|
|
bra.uni $L__BB0_122;
|
|
|
|
$L__BB0_119:
|
|
setp.neu.f32 %p146, %f219, 0f7F800000;
|
|
@%p146 bra $L__BB0_122;
|
|
|
|
selp.f32 %f1418, 0fFF800000, 0f7F800000, %p8;
|
|
|
|
$L__BB0_122:
|
|
abs.f32 %f230, %f206;
|
|
setp.lt.f32 %p147, %f230, 0f00800000;
|
|
mul.f32 %f1199, %f230, 0f4B800000;
|
|
selp.f32 %f1200, %f1199, %f230, %p147;
|
|
selp.f32 %f1201, 0fC3170000, 0fC2FE0000, %p147;
|
|
mov.b32 %r370, %f1200;
|
|
and.b32 %r371, %r370, 8388607;
|
|
or.b32 %r372, %r371, 1065353216;
|
|
mov.b32 %f1202, %r372;
|
|
shr.u32 %r373, %r370, 23;
|
|
cvt.rn.f32.u32 %f1203, %r373;
|
|
add.f32 %f1204, %f1201, %f1203;
|
|
setp.gt.f32 %p148, %f1202, 0f3FB504F3;
|
|
mul.f32 %f1205, %f1202, 0f3F000000;
|
|
add.f32 %f1206, %f1204, 0f3F800000;
|
|
selp.f32 %f1207, %f1206, %f1204, %p148;
|
|
selp.f32 %f1208, %f1205, %f1202, %p148;
|
|
add.f32 %f1209, %f1208, 0fBF800000;
|
|
add.f32 %f1210, %f1208, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1211, %f1210;
|
|
add.f32 %f1212, %f1209, %f1209;
|
|
mul.f32 %f1213, %f1212, %f1211;
|
|
mul.f32 %f1214, %f1213, %f1213;
|
|
mov.f32 %f1215, 0f3C4CAF63;
|
|
mov.f32 %f1216, 0f3B18F0FE;
|
|
fma.rn.f32 %f1217, %f1216, %f1214, %f1215;
|
|
mov.f32 %f1218, 0f3DAAAABD;
|
|
fma.rn.f32 %f1219, %f1217, %f1214, %f1218;
|
|
mul.rn.f32 %f1220, %f1219, %f1214;
|
|
mul.rn.f32 %f1221, %f1220, %f1213;
|
|
sub.f32 %f1222, %f1209, %f1213;
|
|
add.f32 %f1223, %f1222, %f1222;
|
|
neg.f32 %f1224, %f1213;
|
|
fma.rn.f32 %f1225, %f1224, %f1209, %f1223;
|
|
mul.rn.f32 %f1226, %f1211, %f1225;
|
|
add.f32 %f1227, %f1221, %f1213;
|
|
sub.f32 %f1228, %f1213, %f1227;
|
|
add.f32 %f1229, %f1221, %f1228;
|
|
add.f32 %f1230, %f1226, %f1229;
|
|
add.f32 %f1231, %f1227, %f1230;
|
|
sub.f32 %f1232, %f1227, %f1231;
|
|
add.f32 %f1233, %f1230, %f1232;
|
|
mov.f32 %f1234, 0f3F317200;
|
|
mul.rn.f32 %f1235, %f1207, %f1234;
|
|
mov.f32 %f1236, 0f35BFBE8E;
|
|
mul.rn.f32 %f1237, %f1207, %f1236;
|
|
add.f32 %f1238, %f1235, %f1231;
|
|
sub.f32 %f1239, %f1235, %f1238;
|
|
add.f32 %f1240, %f1231, %f1239;
|
|
add.f32 %f1241, %f1233, %f1240;
|
|
add.f32 %f1242, %f1237, %f1241;
|
|
add.f32 %f1243, %f1238, %f1242;
|
|
sub.f32 %f1244, %f1238, %f1243;
|
|
add.f32 %f1245, %f1242, %f1244;
|
|
mov.f32 %f1246, 0f3ED55555;
|
|
mul.rn.f32 %f1247, %f1246, %f1243;
|
|
neg.f32 %f1248, %f1247;
|
|
fma.rn.f32 %f1249, %f1246, %f1243, %f1248;
|
|
fma.rn.f32 %f1250, %f1246, %f1245, %f1249;
|
|
mov.f32 %f1251, 0f00000000;
|
|
fma.rn.f32 %f1252, %f1251, %f1243, %f1250;
|
|
add.rn.f32 %f1253, %f1247, %f1252;
|
|
neg.f32 %f1254, %f1253;
|
|
add.rn.f32 %f1255, %f1247, %f1254;
|
|
add.rn.f32 %f1256, %f1255, %f1252;
|
|
mov.b32 %r374, %f1253;
|
|
setp.eq.s32 %p149, %r374, 1118925336;
|
|
add.s32 %r375, %r374, -1;
|
|
mov.b32 %f1257, %r375;
|
|
add.f32 %f1258, %f1256, 0f37000000;
|
|
selp.f32 %f231, %f1258, %f1256, %p149;
|
|
selp.f32 %f1259, %f1257, %f1253, %p149;
|
|
mov.f32 %f1260, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1261, %f1259, %f1260;
|
|
cvt.rzi.f32.f32 %f1262, %f1261;
|
|
abs.f32 %f1263, %f1262;
|
|
setp.gt.f32 %p150, %f1263, 0f42FC0000;
|
|
mov.b32 %r376, %f1262;
|
|
and.b32 %r377, %r376, -2147483648;
|
|
or.b32 %r378, %r377, 1123811328;
|
|
mov.b32 %f1264, %r378;
|
|
selp.f32 %f1265, %f1264, %f1262, %p150;
|
|
mov.f32 %f1266, 0fBF317218;
|
|
fma.rn.f32 %f1267, %f1265, %f1266, %f1259;
|
|
mov.f32 %f1268, 0f3102E308;
|
|
fma.rn.f32 %f1269, %f1265, %f1268, %f1267;
|
|
mul.f32 %f1270, %f1269, 0f3FB8AA3B;
|
|
add.f32 %f1271, %f1265, 0f4B40007F;
|
|
mov.b32 %r379, %f1271;
|
|
shl.b32 %r380, %r379, 23;
|
|
mov.b32 %f1272, %r380;
|
|
ex2.approx.ftz.f32 %f1273, %f1270;
|
|
mul.f32 %f232, %f1273, %f1272;
|
|
setp.eq.f32 %p151, %f232, 0f7F800000;
|
|
mov.f32 %f1419, 0f7F800000;
|
|
@%p151 bra $L__BB0_124;
|
|
|
|
fma.rn.f32 %f1419, %f232, %f231, %f232;
|
|
|
|
$L__BB0_124:
|
|
setp.lt.f32 %p152, %f206, 0f00000000;
|
|
and.pred %p9, %p152, %p125;
|
|
setp.eq.f32 %p154, %f206, 0f00000000;
|
|
@%p154 bra $L__BB0_128;
|
|
bra.uni $L__BB0_125;
|
|
|
|
$L__BB0_128:
|
|
add.f32 %f1278, %f206, %f206;
|
|
selp.f32 %f1421, %f1278, 0f00000000, %p125;
|
|
bra.uni $L__BB0_129;
|
|
|
|
$L__BB0_125:
|
|
mov.b32 %r381, %f1419;
|
|
xor.b32 %r382, %r381, -2147483648;
|
|
mov.b32 %f1274, %r382;
|
|
selp.f32 %f1421, %f1274, %f1419, %p9;
|
|
setp.geu.f32 %p155, %f206, 0f00000000;
|
|
@%p155 bra $L__BB0_129;
|
|
|
|
mov.f32 %f1275, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1276, %f1275;
|
|
setp.eq.f32 %p156, %f1276, 0f3ED55555;
|
|
@%p156 bra $L__BB0_129;
|
|
|
|
mov.f32 %f1421, 0f7FFFFFFF;
|
|
|
|
$L__BB0_129:
|
|
add.f32 %f1279, %f230, 0f3ED55555;
|
|
mov.b32 %r383, %f1279;
|
|
setp.lt.s32 %p158, %r383, 2139095040;
|
|
@%p158 bra $L__BB0_134;
|
|
|
|
setp.gtu.f32 %p159, %f230, 0f7F800000;
|
|
@%p159 bra $L__BB0_133;
|
|
bra.uni $L__BB0_131;
|
|
|
|
$L__BB0_133:
|
|
add.f32 %f1421, %f206, 0f3ED55555;
|
|
bra.uni $L__BB0_134;
|
|
|
|
$L__BB0_131:
|
|
setp.neu.f32 %p160, %f230, 0f7F800000;
|
|
@%p160 bra $L__BB0_134;
|
|
|
|
selp.f32 %f1421, 0fFF800000, 0f7F800000, %p9;
|
|
|
|
$L__BB0_134:
|
|
fma.rn.f32 %f1280, %f1415, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p161, %f204, 0f3F800000;
|
|
mov.f32 %f1281, 0f3F800000;
|
|
selp.f32 %f1282, 0f3F7FFFFF, %f1280, %p161;
|
|
mul.f32 %f1283, %f204, 0f414EB852;
|
|
setp.lt.f32 %p162, %f204, 0f3B4D2E1C;
|
|
selp.f32 %f1284, %f1283, %f1282, %p162;
|
|
fma.rn.f32 %f1285, %f1418, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p163, %f205, 0f3F800000;
|
|
selp.f32 %f1286, 0f3F7FFFFF, %f1285, %p163;
|
|
mul.f32 %f1287, %f205, 0f414EB852;
|
|
setp.lt.f32 %p164, %f205, 0f3B4D2E1C;
|
|
selp.f32 %f1288, %f1287, %f1286, %p164;
|
|
fma.rn.f32 %f1289, %f1421, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p165, %f206, 0f3F800000;
|
|
selp.f32 %f1290, 0f3F7FFFFF, %f1289, %p165;
|
|
mul.f32 %f1291, %f206, 0f414EB852;
|
|
setp.lt.f32 %p166, %f206, 0f3B4D2E1C;
|
|
selp.f32 %f1292, %f1291, %f1290, %p166;
|
|
min.f32 %f1293, %f1284, %f1281;
|
|
mov.f32 %f1294, 0f00000000;
|
|
max.f32 %f1295, %f1294, %f1293;
|
|
mul.f32 %f1296, %f1295, 0f43800000;
|
|
cvt.rzi.u32.f32 %r384, %f1296;
|
|
min.u32 %r385, %r384, 255;
|
|
min.f32 %f1297, %f1288, %f1281;
|
|
max.f32 %f1298, %f1294, %f1297;
|
|
mul.f32 %f1299, %f1298, 0f43800000;
|
|
cvt.rzi.u32.f32 %r386, %f1299;
|
|
min.u32 %r387, %r386, 255;
|
|
min.f32 %f1300, %f1292, %f1281;
|
|
max.f32 %f1301, %f1294, %f1300;
|
|
mul.f32 %f1302, %f1301, 0f43800000;
|
|
cvt.rzi.u32.f32 %r388, %f1302;
|
|
min.u32 %r389, %r388, 255;
|
|
shl.b64 %rd29, %rd10, 2;
|
|
add.s64 %rd30, %rd9, %rd29;
|
|
cvt.u16.u32 %rs20, %r389;
|
|
cvt.u16.u32 %rs21, %r387;
|
|
cvt.u16.u32 %rs22, %r385;
|
|
mov.u16 %rs23, 255;
|
|
st.global.v4.u8 [%rd30], {%rs22, %rs21, %rs20, %rs23};
|
|
|
|
$L__BB0_135:
|
|
and.b32 %r390, %r418, 4;
|
|
setp.eq.s32 %p167, %r390, 0;
|
|
@%p167 bra $L__BB0_139;
|
|
|
|
ld.const.u32 %r391, [params+108];
|
|
setp.eq.s32 %p168, %r391, 0;
|
|
ld.const.u64 %rd31, [params+224];
|
|
cvta.to.global.u64 %rd32, %rd31;
|
|
ld.const.u32 %r392, [params+216];
|
|
mad.lo.s32 %r393, %r392, %r4, %r3;
|
|
mul.wide.u32 %rd33, %r393, 8;
|
|
add.s64 %rd11, %rd32, %rd33;
|
|
@%p168 bra $L__BB0_138;
|
|
|
|
ld.global.v4.u16 {%rs31, %rs32, %rs33, %rs34}, [%rd11];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1303, %rs31;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1304, %rs32;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1305, %rs33;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1306, %f1387, %f1303;
|
|
add.f32 %f1307, %f1388, %f1304;
|
|
add.f32 %f1308, %f1389, %f1305;
|
|
mov.f32 %f1309, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f1308;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f1307;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1306;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f1309;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd11], {%rs27, %rs28, %rs29, %rs30};
|
|
bra.uni $L__BB0_139;
|
|
|
|
$L__BB0_138:
|
|
mov.f32 %f1313, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f1313;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f1389;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f1388;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f1387;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd11], {%rs35, %rs36, %rs37, %rs38};
|
|
|
|
$L__BB0_139:
|
|
and.b32 %r394, %r418, 64;
|
|
setp.eq.s32 %p169, %r394, 0;
|
|
@%p169 bra $L__BB0_151;
|
|
|
|
mul.f32 %f1314, %f1391, %f1391;
|
|
fma.rn.f32 %f1315, %f1390, %f1390, %f1314;
|
|
fma.rn.f32 %f1316, %f1392, %f1392, %f1315;
|
|
sqrt.rn.f32 %f1317, %f1316;
|
|
rcp.rn.f32 %f1318, %f1317;
|
|
mul.f32 %f1319, %f1390, %f1318;
|
|
mul.f32 %f1320, %f1391, %f1318;
|
|
mul.f32 %f1321, %f1392, %f1318;
|
|
ld.const.u64 %rd34, [params+208];
|
|
cvta.to.global.u64 %rd35, %rd34;
|
|
ld.const.u32 %r395, [params+200];
|
|
mad.lo.s32 %r396, %r395, %r4, %r3;
|
|
fma.rn.f32 %f1322, %f1319, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f1323, %f1322, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r397, %f1323;
|
|
fma.rn.f32 %f1324, %f1320, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f1325, %f1324, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r398, %f1325;
|
|
fma.rn.f32 %f1326, %f1321, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f1327, %f1326, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r399, %f1327;
|
|
mul.wide.u32 %rd36, %r396, 4;
|
|
add.s64 %rd37, %rd35, %rd36;
|
|
cvt.u16.u32 %rs39, %r399;
|
|
cvt.u16.u32 %rs40, %r398;
|
|
cvt.u16.u32 %rs41, %r397;
|
|
mov.u16 %rs42, 255;
|
|
st.global.v4.u8 [%rd37], {%rs41, %rs40, %rs39, %rs42};
|
|
|
|
$L__BB0_151:
|
|
ret;
|
|
|
|
}
|
|
|