2926 lines
76 KiB
Plaintext
2926 lines
76 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<166>;
|
|
.reg .b16 %rs<316>;
|
|
.reg .f32 %f<1423>;
|
|
.reg .b32 %r<380>;
|
|
.reg .f64 %fd<5>;
|
|
.reg .b64 %rd<159>;
|
|
|
|
|
|
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.u32 %r3, [params+540];
|
|
shl.b32 %r4, %r3, 1;
|
|
ld.const.u64 %rd37, [params+400];
|
|
cvta.to.global.u64 %rd38, %rd37;
|
|
ld.const.u32 %r59, [params+392];
|
|
mad.lo.s32 %r60, %r59, %r54, %r53;
|
|
mul.wide.u32 %rd39, %r60, 4;
|
|
add.s64 %rd2, %rd38, %rd39;
|
|
ld.global.v2.u8 {%rs7, %rs315}, [%rd2];
|
|
or.b16 %rs9, %rs7, %rs315;
|
|
and.b16 %rs10, %rs9, 255;
|
|
setp.eq.s16 %p7, %rs10, 0;
|
|
@%p7 bra $L__BB0_2;
|
|
|
|
ld.global.u8 %rs314, [%rd2+2];
|
|
bra.uni $L__BB0_3;
|
|
|
|
$L__BB0_2:
|
|
ld.global.u8 %rs314, [%rd2+2];
|
|
setp.eq.s16 %p8, %rs314, 0;
|
|
mov.f32 %f1303, 0f00000000;
|
|
mov.u16 %rs315, 0;
|
|
mov.f32 %f1304, %f1303;
|
|
mov.f32 %f1305, %f1303;
|
|
@%p8 bra $L__BB0_4;
|
|
|
|
$L__BB0_3:
|
|
cvt.rn.f32.u16 %f318, %rs7;
|
|
div.rn.f32 %f319, %f318, 0f437F0000;
|
|
fma.rn.f32 %f320, %f319, 0f40000000, 0fBF800000;
|
|
and.b16 %rs13, %rs315, 255;
|
|
cvt.rn.f32.u16 %f321, %rs13;
|
|
div.rn.f32 %f322, %f321, 0f437F0000;
|
|
fma.rn.f32 %f323, %f322, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f324, %rs314;
|
|
div.rn.f32 %f325, %f324, 0f437F0000;
|
|
fma.rn.f32 %f326, %f325, 0f40000000, 0fBF800000;
|
|
mul.f32 %f327, %f323, %f323;
|
|
fma.rn.f32 %f328, %f320, %f320, %f327;
|
|
fma.rn.f32 %f329, %f326, %f326, %f328;
|
|
sqrt.rn.f32 %f330, %f329;
|
|
rcp.rn.f32 %f331, %f330;
|
|
mul.f32 %f1305, %f331, %f326;
|
|
mul.f32 %f1304, %f331, %f323;
|
|
mul.f32 %f1303, %f320, %f331;
|
|
|
|
$L__BB0_4:
|
|
ld.const.v2.u32 {%r61, %r62}, [params];
|
|
add.s32 %r5, %r61, %r53;
|
|
add.s32 %r6, %r62, %r54;
|
|
setp.eq.f32 %p9, %f1303, 0f00000000;
|
|
setp.eq.f32 %p10, %f1304, 0f00000000;
|
|
and.pred %p11, %p9, %p10;
|
|
setp.eq.f32 %p12, %f1305, 0f00000000;
|
|
and.pred %p13, %p12, %p11;
|
|
@%p13 bra $L__BB0_143;
|
|
bra.uni $L__BB0_5;
|
|
|
|
$L__BB0_143:
|
|
ld.const.u32 %r51, [params+104];
|
|
and.b32 %r326, %r51, 1;
|
|
setp.eq.b32 %p150, %r326, 1;
|
|
mov.pred %p151, 0;
|
|
xor.pred %p152, %p150, %p151;
|
|
not.pred %p153, %p152;
|
|
@%p153 bra $L__BB0_145;
|
|
|
|
ld.const.u64 %rd115, [params+144];
|
|
cvta.to.global.u64 %rd116, %rd115;
|
|
ld.const.u32 %r327, [params+136];
|
|
mad.lo.s32 %r328, %r327, %r6, %r5;
|
|
mul.wide.u32 %rd117, %r328, 4;
|
|
add.s64 %rd118, %rd116, %rd117;
|
|
mov.u16 %rs170, 0;
|
|
st.global.v4.u8 [%rd118], {%rs170, %rs170, %rs170, %rs170};
|
|
|
|
$L__BB0_145:
|
|
and.b32 %r329, %r51, 4;
|
|
setp.eq.s32 %p154, %r329, 0;
|
|
@%p154 bra $L__BB0_147;
|
|
|
|
ld.const.u64 %rd119, [params+224];
|
|
cvta.to.global.u64 %rd120, %rd119;
|
|
ld.const.u32 %r330, [params+216];
|
|
mad.lo.s32 %r331, %r330, %r6, %r5;
|
|
mov.f32 %f1210, 0f00000000;
|
|
mul.wide.u32 %rd121, %r331, 8;
|
|
add.s64 %rd122, %rd120, %rd121;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f1210;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f1210;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs171, %f1210;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs174, 0;
|
|
st.global.v4.u16 [%rd122], {%rs171, %rs172, %rs173, %rs174};
|
|
|
|
$L__BB0_147:
|
|
and.b32 %r332, %r51, 16;
|
|
setp.eq.s32 %p155, %r332, 0;
|
|
@%p155 bra $L__BB0_149;
|
|
|
|
ld.const.u64 %rd123, [params+240];
|
|
cvta.to.global.u64 %rd124, %rd123;
|
|
ld.const.u32 %r333, [params+232];
|
|
mad.lo.s32 %r334, %r333, %r6, %r5;
|
|
mov.f32 %f1213, 0f00000000;
|
|
mul.wide.u32 %rd125, %r334, 8;
|
|
add.s64 %rd126, %rd124, %rd125;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs177, %f1213;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs176, %f1213;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs175, %f1213;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs178, 0;
|
|
st.global.v4.u16 [%rd126], {%rs175, %rs176, %rs177, %rs178};
|
|
|
|
$L__BB0_149:
|
|
ld.const.u32 %r52, [params+108];
|
|
setp.eq.s32 %p156, %r52, 0;
|
|
ld.const.u64 %rd127, [params+256];
|
|
cvta.to.global.u64 %rd128, %rd127;
|
|
ld.const.u32 %r335, [params+248];
|
|
mad.lo.s32 %r336, %r335, %r6, %r5;
|
|
mul.wide.u32 %rd129, %r336, 8;
|
|
add.s64 %rd27, %rd128, %rd129;
|
|
@%p156 bra $L__BB0_151;
|
|
|
|
ld.global.v4.u16 {%rs185, %rs186, %rs187, %rs188}, [%rd27];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1214, %rs185;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1215, %rs186;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1216, %rs187;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1217, %f1214, 0f00000000;
|
|
add.f32 %f1218, %f1215, 0f00000000;
|
|
add.f32 %f1219, %f1216, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs184, %f1219;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs183, %f1218;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs182, %f1217;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs189, 0;
|
|
st.global.v4.u16 [%rd27], {%rs182, %rs183, %rs184, %rs189};
|
|
bra.uni $L__BB0_152;
|
|
|
|
$L__BB0_5:
|
|
ld.const.u64 %rd40, [params+496];
|
|
cvta.to.global.u64 %rd41, %rd40;
|
|
ld.const.u32 %r65, [params+488];
|
|
mad.lo.s32 %r66, %r65, %r54, %r53;
|
|
mul.wide.u32 %rd42, %r66, 8;
|
|
add.s64 %rd43, %rd41, %rd42;
|
|
ld.global.v4.u16 {%rs19, %rs20, %rs21, %rs22}, [%rd43];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f332, %rs19;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f333, %rs20;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f334, %rs21;}
|
|
|
|
// end inline asm
|
|
ld.const.u64 %rd44, [params+432];
|
|
cvta.to.global.u64 %rd45, %rd44;
|
|
ld.const.u32 %r67, [params+424];
|
|
mad.lo.s32 %r68, %r67, %r54, %r53;
|
|
mul.wide.u32 %rd46, %r68, 12;
|
|
add.s64 %rd47, %rd45, %rd46;
|
|
ld.global.f32 %f366, [%rd47];
|
|
mul.f32 %f367, %f366, 0f3456BF95;
|
|
ld.global.f32 %f368, [%rd47+4];
|
|
mul.f32 %f369, %f368, 0f3456BF95;
|
|
ld.global.f32 %f370, [%rd47+8];
|
|
mul.f32 %f371, %f370, 0f3456BF95;
|
|
abs.f32 %f372, %f1303;
|
|
div.rn.f32 %f373, %f367, %f372;
|
|
abs.f32 %f374, %f1304;
|
|
div.rn.f32 %f375, %f369, %f374;
|
|
abs.f32 %f376, %f1305;
|
|
div.rn.f32 %f377, %f371, %f376;
|
|
abs.f32 %f378, %f373;
|
|
abs.f32 %f379, %f375;
|
|
abs.f32 %f380, %f377;
|
|
mov.f32 %f381, 0f38D1B717;
|
|
max.f32 %f382, %f378, %f381;
|
|
max.f32 %f383, %f379, %f381;
|
|
max.f32 %f384, %f380, %f381;
|
|
fma.rn.f32 %f13, %f1303, %f382, %f366;
|
|
fma.rn.f32 %f14, %f1304, %f383, %f368;
|
|
fma.rn.f32 %f15, %f1305, %f384, %f370;
|
|
setp.gt.f32 %p14, %f372, %f376;
|
|
neg.f32 %f385, %f1304;
|
|
selp.f32 %f386, %f385, 0f00000000, %p14;
|
|
mov.f32 %f1365, 0f00000000;
|
|
neg.f32 %f387, %f1305;
|
|
selp.f32 %f388, %f1303, %f387, %p14;
|
|
selp.f32 %f389, 0f00000000, %f1304, %p14;
|
|
mul.f32 %f390, %f388, %f388;
|
|
fma.rn.f32 %f391, %f386, %f386, %f390;
|
|
fma.rn.f32 %f392, %f389, %f389, %f391;
|
|
sqrt.rn.f32 %f393, %f392;
|
|
rcp.rn.f32 %f394, %f393;
|
|
mul.f32 %f16, %f386, %f394;
|
|
mul.f32 %f17, %f388, %f394;
|
|
mul.f32 %f18, %f389, %f394;
|
|
ld.const.u64 %rd48, [params+128];
|
|
cvta.to.global.u64 %rd49, %rd48;
|
|
ld.const.u32 %r69, [params+120];
|
|
mad.lo.s32 %r70, %r69, %r54, %r53;
|
|
mul.wide.u32 %rd50, %r70, 4;
|
|
add.s64 %rd3, %rd49, %rd50;
|
|
setp.lt.s32 %p15, %r3, 1;
|
|
mov.f32 %f1364, %f1365;
|
|
mov.f32 %f1363, %f1365;
|
|
mov.f32 %f1362, %f1365;
|
|
mov.f32 %f1361, %f1365;
|
|
mov.f32 %f1360, %f1365;
|
|
mov.f32 %f1359, %f1365;
|
|
mov.f32 %f1358, %f1365;
|
|
mov.f32 %f1357, %f1365;
|
|
mov.f32 %f1356, %f1365;
|
|
mov.f32 %f1355, %f1365;
|
|
mov.f32 %f1354, %f1365;
|
|
mov.f32 %f1353, %f1365;
|
|
mov.f32 %f1352, %f1365;
|
|
mov.f32 %f1351, %f1365;
|
|
mov.f32 %f1350, %f1365;
|
|
mov.f32 %f1349, %f1365;
|
|
mov.f32 %f1348, %f1365;
|
|
mov.f32 %f1347, %f1365;
|
|
mov.f32 %f1346, %f1365;
|
|
mov.f32 %f1345, %f1365;
|
|
mov.f32 %f1344, %f1365;
|
|
mov.f32 %f1343, %f1365;
|
|
mov.f32 %f1342, %f1365;
|
|
mov.f32 %f1341, %f1365;
|
|
mov.f32 %f1340, %f1365;
|
|
mov.f32 %f1339, %f1365;
|
|
mov.f32 %f1338, %f1365;
|
|
mov.f32 %f1337, %f1365;
|
|
mov.f32 %f1336, %f1365;
|
|
@%p15 bra $L__BB0_34;
|
|
|
|
cvt.rn.f32.s32 %f425, %r4;
|
|
rcp.rn.f32 %f19, %f425;
|
|
ld.global.u32 %r372, [%rd3];
|
|
mul.f32 %f20, %f13, 0f3456BF95;
|
|
mul.f32 %f21, %f14, 0f3456BF95;
|
|
mul.f32 %f22, %f15, 0f3456BF95;
|
|
ld.const.u64 %rd4, [params+96];
|
|
mul.f32 %f426, %f1303, %f17;
|
|
mul.f32 %f427, %f1304, %f16;
|
|
sub.f32 %f23, %f427, %f426;
|
|
mul.f32 %f428, %f1305, %f16;
|
|
mul.f32 %f429, %f1303, %f18;
|
|
sub.f32 %f24, %f429, %f428;
|
|
mul.f32 %f430, %f1304, %f18;
|
|
mul.f32 %f431, %f1305, %f17;
|
|
sub.f32 %f25, %f431, %f430;
|
|
mov.u32 %r71, 0;
|
|
add.s64 %rd5, %rd1, 24;
|
|
mov.u64 %rd51, __cudart_i2opi_f;
|
|
abs.f32 %f496, %f21;
|
|
abs.f32 %f497, %f20;
|
|
max.f32 %f498, %f497, %f496;
|
|
abs.f32 %f499, %f22;
|
|
max.f32 %f500, %f498, %f499;
|
|
mov.u32 %r369, %r71;
|
|
|
|
$L__BB0_7:
|
|
mov.u32 %r371, %r71;
|
|
|
|
$L__BB0_8:
|
|
cvt.rn.f32.s32 %f1301, %r369;
|
|
mad.lo.s32 %r73, %r372, 1664525, 1013904223;
|
|
and.b32 %r74, %r73, 16777215;
|
|
cvt.rn.f32.u32 %f432, %r74;
|
|
fma.rn.f32 %f433, %f432, 0f33800000, %f1301;
|
|
mul.f32 %f434, %f19, %f433;
|
|
mad.lo.s32 %r372, %r73, 1664525, 1013904223;
|
|
and.b32 %r75, %r372, 16777215;
|
|
cvt.rn.f32.u32 %f435, %r75;
|
|
cvt.rn.f32.s32 %f436, %r371;
|
|
fma.rn.f32 %f437, %f435, 0f33800000, %f436;
|
|
mul.f32 %f438, %f19, %f437;
|
|
fma.rn.f32 %f439, %f19, %f433, %f434;
|
|
mov.f32 %f440, 0f3F800000;
|
|
sub.f32 %f87, %f440, %f439;
|
|
mul.f32 %f441, %f87, %f87;
|
|
sub.f32 %f442, %f440, %f441;
|
|
mov.f32 %f443, 0f00000000;
|
|
max.f32 %f444, %f443, %f442;
|
|
sqrt.rn.f32 %f88, %f444;
|
|
mul.f32 %f89, %f438, 0f40C90FDB;
|
|
mul.f32 %f445, %f89, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r376, %f445;
|
|
cvt.rn.f32.s32 %f446, %r376;
|
|
mov.f32 %f447, 0fBFC90FDA;
|
|
fma.rn.f32 %f448, %f446, %f447, %f89;
|
|
mov.f32 %f449, 0fB3A22168;
|
|
fma.rn.f32 %f450, %f446, %f449, %f448;
|
|
mov.f32 %f451, 0fA7C234C5;
|
|
fma.rn.f32 %f1369, %f446, %f451, %f450;
|
|
abs.f32 %f91, %f89;
|
|
setp.ltu.f32 %p16, %f91, 0f47CE4780;
|
|
mov.f32 %f1366, %f1369;
|
|
@%p16 bra $L__BB0_16;
|
|
|
|
setp.eq.f32 %p17, %f91, 0f7F800000;
|
|
@%p17 bra $L__BB0_15;
|
|
bra.uni $L__BB0_10;
|
|
|
|
$L__BB0_15:
|
|
mov.f32 %f454, 0f00000000;
|
|
mul.rn.f32 %f1366, %f89, %f454;
|
|
mov.u32 %r376, 0;
|
|
bra.uni $L__BB0_16;
|
|
|
|
$L__BB0_10:
|
|
mov.b32 %r14, %f89;
|
|
bfe.u32 %r77, %r14, 23, 8;
|
|
mov.u64 %rd156, 0;
|
|
mov.u32 %r373, 0;
|
|
mov.u64 %rd154, %rd1;
|
|
mov.u64 %rd155, %rd51;
|
|
|
|
$L__BB0_11:
|
|
.pragma "nounroll";
|
|
mov.b32 %r356, %f89;
|
|
shl.b32 %r355, %r356, 8;
|
|
or.b32 %r354, %r355, -2147483648;
|
|
ld.global.nc.u32 %r79, [%rd155];
|
|
mad.wide.u32 %rd53, %r79, %r354, %rd156;
|
|
shr.u64 %rd156, %rd53, 32;
|
|
st.local.u32 [%rd154], %rd53;
|
|
add.s64 %rd155, %rd155, 4;
|
|
add.s64 %rd154, %rd154, 4;
|
|
add.s32 %r373, %r373, 1;
|
|
setp.ne.s32 %p18, %r373, 6;
|
|
@%p18 bra $L__BB0_11;
|
|
|
|
add.s32 %r361, %r77, -128;
|
|
mov.b32 %r360, %f89;
|
|
bfe.u32 %r359, %r360, 23, 8;
|
|
add.s32 %r358, %r359, -128;
|
|
shr.u32 %r357, %r358, 5;
|
|
st.local.u32 [%rd5], %rd156;
|
|
mov.u32 %r81, 6;
|
|
sub.s32 %r82, %r81, %r357;
|
|
mul.wide.s32 %rd54, %r82, 4;
|
|
add.s64 %rd55, %rd1, %rd54;
|
|
ld.local.u32 %r374, [%rd55];
|
|
ld.local.u32 %r375, [%rd55+-4];
|
|
and.b32 %r23, %r358, 31;
|
|
setp.eq.s32 %p19, %r23, 0;
|
|
@%p19 bra $L__BB0_14;
|
|
|
|
mov.b32 %r368, %f89;
|
|
bfe.u32 %r367, %r368, 23, 8;
|
|
add.s32 %r366, %r367, -128;
|
|
shr.u32 %r365, %r366, 5;
|
|
mov.u32 %r364, 4;
|
|
sub.s32 %r363, %r364, %r365;
|
|
mov.u32 %r83, 32;
|
|
sub.s32 %r84, %r83, %r23;
|
|
shr.u32 %r85, %r375, %r84;
|
|
shl.b32 %r86, %r374, %r23;
|
|
add.s32 %r374, %r85, %r86;
|
|
mul.wide.s32 %rd56, %r363, 4;
|
|
add.s64 %rd57, %rd1, %rd56;
|
|
ld.local.u32 %r87, [%rd57];
|
|
shr.u32 %r88, %r87, %r84;
|
|
shl.b32 %r89, %r375, %r23;
|
|
add.s32 %r375, %r88, %r89;
|
|
|
|
$L__BB0_14:
|
|
mov.b32 %r362, %f89;
|
|
and.b32 %r90, %r362, -2147483648;
|
|
shr.u32 %r91, %r375, 30;
|
|
shl.b32 %r92, %r374, 2;
|
|
or.b32 %r93, %r91, %r92;
|
|
shr.u32 %r94, %r93, 31;
|
|
shr.u32 %r95, %r374, 30;
|
|
add.s32 %r96, %r94, %r95;
|
|
neg.s32 %r97, %r96;
|
|
setp.eq.s32 %p20, %r90, 0;
|
|
selp.b32 %r376, %r96, %r97, %p20;
|
|
setp.ne.s32 %p21, %r94, 0;
|
|
xor.b32 %r98, %r90, -2147483648;
|
|
selp.b32 %r99, %r98, %r90, %p21;
|
|
selp.b32 %r100, -1, 0, %p21;
|
|
xor.b32 %r101, %r93, %r100;
|
|
shl.b32 %r102, %r375, 2;
|
|
xor.b32 %r103, %r102, %r100;
|
|
cvt.u64.u32 %rd58, %r101;
|
|
cvt.u64.u32 %rd59, %r103;
|
|
bfi.b64 %rd60, %rd58, %rd59, 32, 32;
|
|
cvt.rn.f64.s64 %fd1, %rd60;
|
|
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f452, %fd2;
|
|
setp.eq.s32 %p22, %r99, 0;
|
|
neg.f32 %f453, %f452;
|
|
selp.f32 %f1366, %f452, %f453, %p22;
|
|
|
|
$L__BB0_16:
|
|
add.s32 %r30, %r376, 1;
|
|
and.b32 %r31, %r30, 1;
|
|
setp.eq.s32 %p23, %r31, 0;
|
|
selp.f32 %f95, %f1366, 0f3F800000, %p23;
|
|
mul.rn.f32 %f96, %f1366, %f1366;
|
|
mov.f32 %f1367, 0fB94D4153;
|
|
@%p23 bra $L__BB0_18;
|
|
|
|
mov.f32 %f456, 0fBAB607ED;
|
|
mov.f32 %f457, 0f37CBAC00;
|
|
fma.rn.f32 %f1367, %f457, %f96, %f456;
|
|
|
|
$L__BB0_18:
|
|
selp.f32 %f458, 0f3C0885E4, 0f3D2AAABB, %p23;
|
|
fma.rn.f32 %f459, %f1367, %f96, %f458;
|
|
selp.f32 %f460, 0fBE2AAAA8, 0fBEFFFFFF, %p23;
|
|
fma.rn.f32 %f461, %f459, %f96, %f460;
|
|
mov.f32 %f462, 0f00000000;
|
|
fma.rn.f32 %f463, %f96, %f95, %f462;
|
|
fma.rn.f32 %f1368, %f461, %f463, %f95;
|
|
and.b32 %r105, %r30, 2;
|
|
setp.eq.s32 %p25, %r105, 0;
|
|
@%p25 bra $L__BB0_20;
|
|
|
|
mov.f32 %f465, 0fBF800000;
|
|
fma.rn.f32 %f1368, %f1368, %f465, %f462;
|
|
|
|
$L__BB0_20:
|
|
mul.f32 %f1302, %f89, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r379, %f1302;
|
|
setp.ltu.f32 %p165, %f91, 0f47CE4780;
|
|
@%p165 bra $L__BB0_28;
|
|
|
|
setp.eq.f32 %p27, %f91, 0f7F800000;
|
|
@%p27 bra $L__BB0_27;
|
|
bra.uni $L__BB0_22;
|
|
|
|
$L__BB0_27:
|
|
mov.f32 %f468, 0f00000000;
|
|
mul.rn.f32 %f1369, %f89, %f468;
|
|
mov.u32 %r379, 0;
|
|
bra.uni $L__BB0_28;
|
|
|
|
$L__BB0_22:
|
|
mov.b32 %r32, %f89;
|
|
bfe.u32 %r106, %r32, 23, 8;
|
|
add.s32 %r33, %r106, -128;
|
|
shl.b32 %r107, %r32, 8;
|
|
or.b32 %r34, %r107, -2147483648;
|
|
shr.u32 %r35, %r33, 5;
|
|
mov.u64 %rd157, 0;
|
|
mov.u64 %rd158, %rd157;
|
|
|
|
$L__BB0_23:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd63, %rd157, 2;
|
|
mov.u64 %rd64, __cudart_i2opi_f;
|
|
add.s64 %rd65, %rd64, %rd63;
|
|
ld.global.nc.u32 %r108, [%rd65];
|
|
mad.wide.u32 %rd66, %r108, %r34, %rd158;
|
|
shr.u64 %rd158, %rd66, 32;
|
|
add.s64 %rd67, %rd1, %rd63;
|
|
st.local.u32 [%rd67], %rd66;
|
|
cvt.u32.u64 %r109, %rd157;
|
|
add.s32 %r110, %r109, 1;
|
|
cvt.s64.s32 %rd157, %r110;
|
|
setp.ne.s32 %p28, %r110, 6;
|
|
@%p28 bra $L__BB0_23;
|
|
|
|
st.local.u32 [%rd5], %rd158;
|
|
mov.u32 %r111, 4;
|
|
sub.s32 %r36, %r111, %r35;
|
|
mov.u32 %r112, 6;
|
|
sub.s32 %r113, %r112, %r35;
|
|
mul.wide.s32 %rd68, %r113, 4;
|
|
add.s64 %rd69, %rd1, %rd68;
|
|
ld.local.u32 %r377, [%rd69];
|
|
ld.local.u32 %r378, [%rd69+-4];
|
|
and.b32 %r39, %r33, 31;
|
|
setp.eq.s32 %p29, %r39, 0;
|
|
@%p29 bra $L__BB0_26;
|
|
|
|
mov.u32 %r114, 32;
|
|
sub.s32 %r115, %r114, %r39;
|
|
shr.u32 %r116, %r378, %r115;
|
|
shl.b32 %r117, %r377, %r39;
|
|
add.s32 %r377, %r116, %r117;
|
|
mul.wide.s32 %rd70, %r36, 4;
|
|
add.s64 %rd71, %rd1, %rd70;
|
|
ld.local.u32 %r118, [%rd71];
|
|
shr.u32 %r119, %r118, %r115;
|
|
shl.b32 %r120, %r378, %r39;
|
|
add.s32 %r378, %r119, %r120;
|
|
|
|
$L__BB0_26:
|
|
and.b32 %r121, %r32, -2147483648;
|
|
shr.u32 %r122, %r378, 30;
|
|
shl.b32 %r123, %r377, 2;
|
|
or.b32 %r124, %r122, %r123;
|
|
shr.u32 %r125, %r124, 31;
|
|
shr.u32 %r126, %r377, 30;
|
|
add.s32 %r127, %r125, %r126;
|
|
neg.s32 %r128, %r127;
|
|
setp.eq.s32 %p30, %r121, 0;
|
|
selp.b32 %r379, %r127, %r128, %p30;
|
|
setp.ne.s32 %p31, %r125, 0;
|
|
xor.b32 %r129, %r121, -2147483648;
|
|
selp.b32 %r130, %r129, %r121, %p31;
|
|
selp.b32 %r131, -1, 0, %p31;
|
|
xor.b32 %r132, %r124, %r131;
|
|
shl.b32 %r133, %r378, 2;
|
|
xor.b32 %r134, %r133, %r131;
|
|
cvt.u64.u32 %rd72, %r132;
|
|
cvt.u64.u32 %rd73, %r134;
|
|
bfi.b64 %rd74, %rd72, %rd73, 32, 32;
|
|
cvt.rn.f64.s64 %fd3, %rd74;
|
|
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f466, %fd4;
|
|
setp.eq.s32 %p32, %r130, 0;
|
|
neg.f32 %f467, %f466;
|
|
selp.f32 %f1369, %f466, %f467, %p32;
|
|
|
|
$L__BB0_28:
|
|
mul.f32 %f105, %f88, %f1368;
|
|
and.b32 %r46, %r379, 1;
|
|
setp.eq.s32 %p33, %r46, 0;
|
|
selp.f32 %f106, %f1369, 0f3F800000, %p33;
|
|
mul.rn.f32 %f107, %f1369, %f1369;
|
|
mov.f32 %f1370, 0fB94D4153;
|
|
@%p33 bra $L__BB0_30;
|
|
|
|
mov.f32 %f470, 0fBAB607ED;
|
|
mov.f32 %f471, 0f37CBAC00;
|
|
fma.rn.f32 %f1370, %f471, %f107, %f470;
|
|
|
|
$L__BB0_30:
|
|
selp.f32 %f472, 0f3C0885E4, 0f3D2AAABB, %p33;
|
|
fma.rn.f32 %f473, %f1370, %f107, %f472;
|
|
selp.f32 %f474, 0fBE2AAAA8, 0fBEFFFFFF, %p33;
|
|
fma.rn.f32 %f475, %f473, %f107, %f474;
|
|
mov.f32 %f476, 0f00000000;
|
|
fma.rn.f32 %f477, %f107, %f106, %f476;
|
|
fma.rn.f32 %f1371, %f475, %f477, %f106;
|
|
and.b32 %r136, %r379, 2;
|
|
setp.eq.s32 %p35, %r136, 0;
|
|
@%p35 bra $L__BB0_32;
|
|
|
|
mov.f32 %f479, 0fBF800000;
|
|
fma.rn.f32 %f1371, %f1371, %f479, %f476;
|
|
|
|
$L__BB0_32:
|
|
mul.f32 %f489, %f88, %f1371;
|
|
mul.f32 %f490, %f16, %f489;
|
|
mul.f32 %f491, %f17, %f489;
|
|
mul.f32 %f492, %f18, %f489;
|
|
fma.rn.f32 %f493, %f25, %f105, %f490;
|
|
fma.rn.f32 %f494, %f24, %f105, %f491;
|
|
fma.rn.f32 %f495, %f23, %f105, %f492;
|
|
fma.rn.f32 %f483, %f1303, %f87, %f493;
|
|
fma.rn.f32 %f484, %f1304, %f87, %f494;
|
|
fma.rn.f32 %f485, %f1305, %f87, %f495;
|
|
mov.f32 %f501, 0f38D1B717;
|
|
max.f32 %f486, %f500, %f501;
|
|
mov.f32 %f487, 0f6C4ECB8F;
|
|
mov.u32 %r170, 1;
|
|
mov.u32 %r173, 2;
|
|
mov.u32 %r175, 4;
|
|
mov.u32 %r178, 1065353216;
|
|
mov.u32 %r179, 2139095039;
|
|
mov.u32 %r207, 0;
|
|
// begin inline asm
|
|
call(%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167,%r168),_optix_trace_typed_32,(%r207,%rd4,%f13,%f14,%f15,%f483,%f484,%f485,%f486,%f487,%f476,%r170,%r207,%r207,%r173,%r207,%r175,%r178,%r178,%r178,%r179,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207);
|
|
// end inline asm
|
|
mov.b32 %f502, %r137;
|
|
mov.b32 %f503, %r138;
|
|
mov.b32 %f504, %r139;
|
|
mov.b32 %f505, %r140;
|
|
setp.ltu.f32 %p36, %f505, 0f00000000;
|
|
selp.f32 %f506, %f504, 0f00000000, %p36;
|
|
selp.f32 %f507, %f503, 0f00000000, %p36;
|
|
selp.f32 %f508, %f502, 0f00000000, %p36;
|
|
mul.f32 %f509, %f1304, %f484;
|
|
fma.rn.f32 %f510, %f1303, %f483, %f509;
|
|
fma.rn.f32 %f511, %f1305, %f485, %f510;
|
|
add.f32 %f1341, %f1341, %f508;
|
|
add.f32 %f1340, %f1340, %f507;
|
|
add.f32 %f1339, %f1339, %f506;
|
|
fma.rn.f32 %f1344, %f484, %f508, %f1344;
|
|
fma.rn.f32 %f1343, %f484, %f507, %f1343;
|
|
fma.rn.f32 %f1342, %f484, %f506, %f1342;
|
|
fma.rn.f32 %f1347, %f485, %f508, %f1347;
|
|
fma.rn.f32 %f1346, %f485, %f507, %f1346;
|
|
fma.rn.f32 %f1345, %f485, %f506, %f1345;
|
|
fma.rn.f32 %f1350, %f483, %f508, %f1350;
|
|
fma.rn.f32 %f1349, %f483, %f507, %f1349;
|
|
fma.rn.f32 %f1348, %f483, %f506, %f1348;
|
|
mul.f32 %f512, %f483, %f484;
|
|
fma.rn.f32 %f1353, %f512, %f508, %f1353;
|
|
fma.rn.f32 %f1352, %f512, %f507, %f1352;
|
|
fma.rn.f32 %f1351, %f512, %f506, %f1351;
|
|
mul.f32 %f513, %f484, %f485;
|
|
fma.rn.f32 %f1356, %f513, %f508, %f1356;
|
|
fma.rn.f32 %f1355, %f513, %f507, %f1355;
|
|
fma.rn.f32 %f1354, %f513, %f506, %f1354;
|
|
fma.rn.f32 %f514, %f485, %f485, 0fBEAAAAAB;
|
|
fma.rn.f32 %f1359, %f514, %f508, %f1359;
|
|
fma.rn.f32 %f1358, %f514, %f507, %f1358;
|
|
fma.rn.f32 %f1357, %f514, %f506, %f1357;
|
|
mul.f32 %f515, %f483, %f485;
|
|
fma.rn.f32 %f1362, %f515, %f508, %f1362;
|
|
fma.rn.f32 %f1361, %f515, %f507, %f1361;
|
|
fma.rn.f32 %f1360, %f515, %f506, %f1360;
|
|
mul.f32 %f516, %f483, %f483;
|
|
mul.f32 %f517, %f484, %f484;
|
|
sub.f32 %f518, %f516, %f517;
|
|
fma.rn.f32 %f1365, %f518, %f508, %f1365;
|
|
fma.rn.f32 %f1364, %f518, %f507, %f1364;
|
|
fma.rn.f32 %f1363, %f518, %f506, %f1363;
|
|
cvt.sat.f32.f32 %f519, %f511;
|
|
fma.rn.f32 %f1338, %f519, %f508, %f1338;
|
|
fma.rn.f32 %f1337, %f519, %f507, %f1337;
|
|
fma.rn.f32 %f1336, %f519, %f506, %f1336;
|
|
add.s32 %r371, %r371, 1;
|
|
setp.lt.s32 %p37, %r371, %r4;
|
|
@%p37 bra $L__BB0_8;
|
|
|
|
add.s32 %r369, %r369, 1;
|
|
setp.lt.s32 %p38, %r369, %r4;
|
|
@%p38 bra $L__BB0_7;
|
|
|
|
$L__BB0_34:
|
|
mov.f32 %f1402, 0f00000000;
|
|
mul.lo.s32 %r208, %r4, %r4;
|
|
cvt.rn.f32.s32 %f523, %r208;
|
|
rcp.rn.f32 %f524, %f523;
|
|
mul.f32 %f525, %f524, %f1338;
|
|
mul.f32 %f526, %f524, %f1337;
|
|
mul.f32 %f527, %f524, %f1336;
|
|
mul.f32 %f173, %f524, %f1341;
|
|
mul.f32 %f174, %f524, %f1340;
|
|
mul.f32 %f175, %f524, %f1339;
|
|
mul.f32 %f176, %f524, %f1344;
|
|
mul.f32 %f177, %f524, %f1343;
|
|
mul.f32 %f178, %f524, %f1342;
|
|
mul.f32 %f179, %f524, %f1347;
|
|
mul.f32 %f180, %f524, %f1346;
|
|
mul.f32 %f181, %f524, %f1345;
|
|
mul.f32 %f182, %f524, %f1350;
|
|
mul.f32 %f183, %f524, %f1349;
|
|
mul.f32 %f184, %f524, %f1348;
|
|
mul.f32 %f185, %f524, %f1353;
|
|
mul.f32 %f186, %f524, %f1352;
|
|
mul.f32 %f187, %f524, %f1351;
|
|
mul.f32 %f188, %f524, %f1356;
|
|
mul.f32 %f189, %f524, %f1355;
|
|
mul.f32 %f190, %f524, %f1354;
|
|
mul.f32 %f191, %f524, %f1359;
|
|
mul.f32 %f192, %f524, %f1358;
|
|
mul.f32 %f193, %f524, %f1357;
|
|
mul.f32 %f194, %f524, %f1362;
|
|
mul.f32 %f195, %f524, %f1361;
|
|
mul.f32 %f196, %f524, %f1360;
|
|
mul.f32 %f197, %f524, %f1365;
|
|
mul.f32 %f198, %f524, %f1364;
|
|
mul.f32 %f199, %f524, %f1363;
|
|
fma.rn.f32 %f528, %f524, %f1338, %f525;
|
|
fma.rn.f32 %f529, %f524, %f1337, %f526;
|
|
fma.rn.f32 %f530, %f524, %f1336, %f527;
|
|
ld.const.u64 %rd76, [params+504];
|
|
cvt.rn.f32.u32 %f531, %r6;
|
|
cvt.rn.f32.u32 %f532, %r5;
|
|
tex.2d.v4.f32.f32 {%f533, %f534, %f535, %f536}, [%rd76, {%f532, %f531}];
|
|
mul.f32 %f200, %f528, %f533;
|
|
mul.f32 %f201, %f529, %f534;
|
|
mul.f32 %f202, %f530, %f535;
|
|
add.f32 %f203, %f332, %f200;
|
|
add.f32 %f204, %f333, %f201;
|
|
add.f32 %f205, %f334, %f202;
|
|
abs.f32 %f537, %f203;
|
|
setp.gtu.f32 %p39, %f537, 0f7F800000;
|
|
mov.f32 %f1403, %f1402;
|
|
mov.f32 %f1404, %f1402;
|
|
@%p39 bra $L__BB0_38;
|
|
|
|
mov.f32 %f1402, 0f00000000;
|
|
abs.f32 %f541, %f204;
|
|
setp.gtu.f32 %p40, %f541, 0f7F800000;
|
|
@%p40 bra $L__BB0_38;
|
|
|
|
mov.f32 %f1402, 0f00000000;
|
|
abs.f32 %f545, %f205;
|
|
setp.gtu.f32 %p41, %f545, 0f7F800000;
|
|
mov.f32 %f1403, %f1402;
|
|
mov.f32 %f1404, %f1402;
|
|
@%p41 bra $L__BB0_38;
|
|
|
|
mov.f32 %f1402, %f203;
|
|
mov.f32 %f1403, %f204;
|
|
mov.f32 %f1404, %f205;
|
|
|
|
$L__BB0_38:
|
|
ld.const.u32 %r49, [params+104];
|
|
and.b32 %r209, %r49, 1;
|
|
setp.eq.b32 %p42, %r209, 1;
|
|
mov.pred %p43, 0;
|
|
xor.pred %p44, %p42, %p43;
|
|
not.pred %p45, %p44;
|
|
@%p45 bra $L__BB0_112;
|
|
|
|
mov.f32 %f547, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f548, %f547;
|
|
add.f32 %f549, %f548, %f548;
|
|
mov.f32 %f550, 0f3EE66666;
|
|
sub.f32 %f551, %f550, %f549;
|
|
abs.f32 %f215, %f551;
|
|
abs.f32 %f216, %f1402;
|
|
setp.lt.f32 %p46, %f216, 0f00800000;
|
|
mul.f32 %f552, %f216, 0f4B800000;
|
|
selp.f32 %f553, %f552, %f216, %p46;
|
|
selp.f32 %f554, 0fC3170000, 0fC2FE0000, %p46;
|
|
mov.b32 %r210, %f553;
|
|
and.b32 %r211, %r210, 8388607;
|
|
or.b32 %r212, %r211, 1065353216;
|
|
mov.b32 %f555, %r212;
|
|
shr.u32 %r213, %r210, 23;
|
|
cvt.rn.f32.u32 %f556, %r213;
|
|
add.f32 %f557, %f554, %f556;
|
|
setp.gt.f32 %p47, %f555, 0f3FB504F3;
|
|
mul.f32 %f558, %f555, 0f3F000000;
|
|
add.f32 %f559, %f557, 0f3F800000;
|
|
selp.f32 %f560, %f559, %f557, %p47;
|
|
selp.f32 %f561, %f558, %f555, %p47;
|
|
add.f32 %f562, %f561, 0fBF800000;
|
|
add.f32 %f563, %f561, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f564, %f563;
|
|
add.f32 %f565, %f562, %f562;
|
|
mul.f32 %f566, %f565, %f564;
|
|
mul.f32 %f567, %f566, %f566;
|
|
mov.f32 %f568, 0f3C4CAF63;
|
|
mov.f32 %f569, 0f3B18F0FE;
|
|
fma.rn.f32 %f570, %f569, %f567, %f568;
|
|
mov.f32 %f571, 0f3DAAAABD;
|
|
fma.rn.f32 %f572, %f570, %f567, %f571;
|
|
mul.rn.f32 %f573, %f572, %f567;
|
|
mul.rn.f32 %f574, %f573, %f566;
|
|
sub.f32 %f575, %f562, %f566;
|
|
add.f32 %f576, %f575, %f575;
|
|
neg.f32 %f577, %f566;
|
|
fma.rn.f32 %f578, %f577, %f562, %f576;
|
|
mul.rn.f32 %f579, %f564, %f578;
|
|
add.f32 %f580, %f574, %f566;
|
|
sub.f32 %f581, %f566, %f580;
|
|
add.f32 %f582, %f574, %f581;
|
|
add.f32 %f583, %f579, %f582;
|
|
add.f32 %f584, %f580, %f583;
|
|
sub.f32 %f585, %f580, %f584;
|
|
add.f32 %f586, %f583, %f585;
|
|
mov.f32 %f587, 0f3F317200;
|
|
mul.rn.f32 %f588, %f560, %f587;
|
|
mov.f32 %f589, 0f35BFBE8E;
|
|
mul.rn.f32 %f590, %f560, %f589;
|
|
add.f32 %f591, %f588, %f584;
|
|
sub.f32 %f592, %f588, %f591;
|
|
add.f32 %f593, %f584, %f592;
|
|
add.f32 %f594, %f586, %f593;
|
|
add.f32 %f595, %f590, %f594;
|
|
add.f32 %f596, %f591, %f595;
|
|
sub.f32 %f597, %f591, %f596;
|
|
add.f32 %f598, %f595, %f597;
|
|
mul.rn.f32 %f599, %f550, %f596;
|
|
neg.f32 %f600, %f599;
|
|
fma.rn.f32 %f601, %f550, %f596, %f600;
|
|
fma.rn.f32 %f602, %f550, %f598, %f601;
|
|
mov.f32 %f603, 0f00000000;
|
|
fma.rn.f32 %f604, %f603, %f596, %f602;
|
|
add.rn.f32 %f605, %f599, %f604;
|
|
neg.f32 %f606, %f605;
|
|
add.rn.f32 %f607, %f599, %f606;
|
|
add.rn.f32 %f608, %f607, %f604;
|
|
mov.b32 %r214, %f605;
|
|
setp.eq.s32 %p48, %r214, 1118925336;
|
|
add.s32 %r215, %r214, -1;
|
|
mov.b32 %f609, %r215;
|
|
add.f32 %f610, %f608, 0f37000000;
|
|
selp.f32 %f217, %f610, %f608, %p48;
|
|
selp.f32 %f611, %f609, %f605, %p48;
|
|
mov.f32 %f612, 0f3FB8AA3B;
|
|
mul.rn.f32 %f613, %f611, %f612;
|
|
cvt.rzi.f32.f32 %f614, %f613;
|
|
abs.f32 %f615, %f614;
|
|
setp.gt.f32 %p49, %f615, 0f42FC0000;
|
|
mov.b32 %r216, %f614;
|
|
and.b32 %r217, %r216, -2147483648;
|
|
or.b32 %r218, %r217, 1123811328;
|
|
mov.b32 %f616, %r218;
|
|
selp.f32 %f617, %f616, %f614, %p49;
|
|
mov.f32 %f618, 0fBF317218;
|
|
fma.rn.f32 %f619, %f617, %f618, %f611;
|
|
mov.f32 %f620, 0f3102E308;
|
|
fma.rn.f32 %f621, %f617, %f620, %f619;
|
|
mul.f32 %f622, %f621, 0f3FB8AA3B;
|
|
add.f32 %f623, %f617, 0f4B40007F;
|
|
mov.b32 %r219, %f623;
|
|
shl.b32 %r220, %r219, 23;
|
|
mov.b32 %f624, %r220;
|
|
ex2.approx.ftz.f32 %f625, %f622;
|
|
mul.f32 %f218, %f625, %f624;
|
|
setp.eq.f32 %p50, %f218, 0f7F800000;
|
|
mov.f32 %f1405, 0f7F800000;
|
|
@%p50 bra $L__BB0_41;
|
|
|
|
fma.rn.f32 %f1405, %f218, %f217, %f218;
|
|
|
|
$L__BB0_41:
|
|
setp.lt.f32 %p51, %f1402, 0f00000000;
|
|
setp.eq.f32 %p52, %f215, 0f3F800000;
|
|
and.pred %p1, %p51, %p52;
|
|
setp.eq.f32 %p53, %f1402, 0f00000000;
|
|
@%p53 bra $L__BB0_45;
|
|
bra.uni $L__BB0_42;
|
|
|
|
$L__BB0_45:
|
|
add.f32 %f630, %f1402, %f1402;
|
|
selp.f32 %f1407, %f630, 0f00000000, %p52;
|
|
bra.uni $L__BB0_46;
|
|
|
|
$L__BB0_151:
|
|
mov.f32 %f1222, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs192, %f1222;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs191, %f1222;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs190, %f1222;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs193, 0;
|
|
st.global.v4.u16 [%rd27], {%rs190, %rs191, %rs192, %rs193};
|
|
|
|
$L__BB0_152:
|
|
ld.const.u64 %rd130, [params+272];
|
|
cvta.to.global.u64 %rd131, %rd130;
|
|
ld.const.u32 %r337, [params+264];
|
|
mad.lo.s32 %r338, %r337, %r6, %r5;
|
|
mul.wide.u32 %rd132, %r338, 8;
|
|
add.s64 %rd28, %rd131, %rd132;
|
|
@%p156 bra $L__BB0_154;
|
|
|
|
ld.global.v4.u16 {%rs200, %rs201, %rs202, %rs203}, [%rd28];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1223, %rs200;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1224, %rs201;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1225, %rs202;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1226, %f1223, 0f00000000;
|
|
add.f32 %f1227, %f1224, 0f00000000;
|
|
add.f32 %f1228, %f1225, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f1228;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f1227;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs197, %f1226;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs204, 0;
|
|
st.global.v4.u16 [%rd28], {%rs197, %rs198, %rs199, %rs204};
|
|
bra.uni $L__BB0_155;
|
|
|
|
$L__BB0_154:
|
|
mov.f32 %f1231, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs207, %f1231;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs206, %f1231;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs205, %f1231;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs208, 0;
|
|
st.global.v4.u16 [%rd28], {%rs205, %rs206, %rs207, %rs208};
|
|
|
|
$L__BB0_155:
|
|
ld.const.u64 %rd133, [params+288];
|
|
cvta.to.global.u64 %rd134, %rd133;
|
|
ld.const.u32 %r339, [params+280];
|
|
mad.lo.s32 %r340, %r339, %r6, %r5;
|
|
mul.wide.u32 %rd135, %r340, 8;
|
|
add.s64 %rd29, %rd134, %rd135;
|
|
@%p156 bra $L__BB0_157;
|
|
|
|
ld.global.v4.u16 {%rs215, %rs216, %rs217, %rs218}, [%rd29];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1232, %rs215;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1233, %rs216;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1234, %rs217;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1235, %f1232, 0f00000000;
|
|
add.f32 %f1236, %f1233, 0f00000000;
|
|
add.f32 %f1237, %f1234, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs214, %f1237;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs213, %f1236;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f1235;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs219, 0;
|
|
st.global.v4.u16 [%rd29], {%rs212, %rs213, %rs214, %rs219};
|
|
bra.uni $L__BB0_158;
|
|
|
|
$L__BB0_157:
|
|
mov.f32 %f1240, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs222, %f1240;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs221, %f1240;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs220, %f1240;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs223, 0;
|
|
st.global.v4.u16 [%rd29], {%rs220, %rs221, %rs222, %rs223};
|
|
|
|
$L__BB0_158:
|
|
ld.const.u64 %rd136, [params+304];
|
|
cvta.to.global.u64 %rd137, %rd136;
|
|
ld.const.u32 %r341, [params+296];
|
|
mad.lo.s32 %r342, %r341, %r6, %r5;
|
|
mul.wide.u32 %rd138, %r342, 8;
|
|
add.s64 %rd30, %rd137, %rd138;
|
|
@%p156 bra $L__BB0_160;
|
|
|
|
ld.global.v4.u16 {%rs230, %rs231, %rs232, %rs233}, [%rd30];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1241, %rs230;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1242, %rs231;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1243, %rs232;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1244, %f1241, 0f00000000;
|
|
add.f32 %f1245, %f1242, 0f00000000;
|
|
add.f32 %f1246, %f1243, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs229, %f1246;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs228, %f1245;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs227, %f1244;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs234, 0;
|
|
st.global.v4.u16 [%rd30], {%rs227, %rs228, %rs229, %rs234};
|
|
bra.uni $L__BB0_161;
|
|
|
|
$L__BB0_160:
|
|
mov.f32 %f1249, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs237, %f1249;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs236, %f1249;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs235, %f1249;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs238, 0;
|
|
st.global.v4.u16 [%rd30], {%rs235, %rs236, %rs237, %rs238};
|
|
|
|
$L__BB0_161:
|
|
ld.const.u64 %rd139, [params+320];
|
|
cvta.to.global.u64 %rd140, %rd139;
|
|
ld.const.u32 %r343, [params+312];
|
|
mad.lo.s32 %r344, %r343, %r6, %r5;
|
|
mul.wide.u32 %rd141, %r344, 8;
|
|
add.s64 %rd31, %rd140, %rd141;
|
|
@%p156 bra $L__BB0_163;
|
|
|
|
ld.global.v4.u16 {%rs245, %rs246, %rs247, %rs248}, [%rd31];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1250, %rs245;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1251, %rs246;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1252, %rs247;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1253, %f1250, 0f00000000;
|
|
add.f32 %f1254, %f1251, 0f00000000;
|
|
add.f32 %f1255, %f1252, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs244, %f1255;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs243, %f1254;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs242, %f1253;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs249, 0;
|
|
st.global.v4.u16 [%rd31], {%rs242, %rs243, %rs244, %rs249};
|
|
bra.uni $L__BB0_164;
|
|
|
|
$L__BB0_163:
|
|
mov.f32 %f1258, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs252, %f1258;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs251, %f1258;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs250, %f1258;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs253, 0;
|
|
st.global.v4.u16 [%rd31], {%rs250, %rs251, %rs252, %rs253};
|
|
|
|
$L__BB0_164:
|
|
ld.const.u64 %rd142, [params+336];
|
|
cvta.to.global.u64 %rd143, %rd142;
|
|
ld.const.u32 %r345, [params+328];
|
|
mad.lo.s32 %r346, %r345, %r6, %r5;
|
|
mul.wide.u32 %rd144, %r346, 8;
|
|
add.s64 %rd32, %rd143, %rd144;
|
|
@%p156 bra $L__BB0_166;
|
|
|
|
ld.global.v4.u16 {%rs260, %rs261, %rs262, %rs263}, [%rd32];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1259, %rs260;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1260, %rs261;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1261, %rs262;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1262, %f1259, 0f00000000;
|
|
add.f32 %f1263, %f1260, 0f00000000;
|
|
add.f32 %f1264, %f1261, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs259, %f1264;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs258, %f1263;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs257, %f1262;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs264, 0;
|
|
st.global.v4.u16 [%rd32], {%rs257, %rs258, %rs259, %rs264};
|
|
bra.uni $L__BB0_167;
|
|
|
|
$L__BB0_166:
|
|
mov.f32 %f1267, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs267, %f1267;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs266, %f1267;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs265, %f1267;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs268, 0;
|
|
st.global.v4.u16 [%rd32], {%rs265, %rs266, %rs267, %rs268};
|
|
|
|
$L__BB0_167:
|
|
ld.const.u64 %rd145, [params+352];
|
|
cvta.to.global.u64 %rd146, %rd145;
|
|
ld.const.u32 %r347, [params+344];
|
|
mad.lo.s32 %r348, %r347, %r6, %r5;
|
|
mul.wide.u32 %rd147, %r348, 8;
|
|
add.s64 %rd33, %rd146, %rd147;
|
|
@%p156 bra $L__BB0_169;
|
|
|
|
ld.global.v4.u16 {%rs275, %rs276, %rs277, %rs278}, [%rd33];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1268, %rs275;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1269, %rs276;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1270, %rs277;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1271, %f1268, 0f00000000;
|
|
add.f32 %f1272, %f1269, 0f00000000;
|
|
add.f32 %f1273, %f1270, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs274, %f1273;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs273, %f1272;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs272, %f1271;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs279, 0;
|
|
st.global.v4.u16 [%rd33], {%rs272, %rs273, %rs274, %rs279};
|
|
bra.uni $L__BB0_170;
|
|
|
|
$L__BB0_169:
|
|
mov.f32 %f1276, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs282, %f1276;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs281, %f1276;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs280, %f1276;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs283, 0;
|
|
st.global.v4.u16 [%rd33], {%rs280, %rs281, %rs282, %rs283};
|
|
|
|
$L__BB0_170:
|
|
ld.const.u64 %rd148, [params+368];
|
|
cvta.to.global.u64 %rd149, %rd148;
|
|
ld.const.u32 %r349, [params+360];
|
|
mad.lo.s32 %r350, %r349, %r6, %r5;
|
|
mul.wide.u32 %rd150, %r350, 8;
|
|
add.s64 %rd34, %rd149, %rd150;
|
|
@%p156 bra $L__BB0_172;
|
|
|
|
ld.global.v4.u16 {%rs290, %rs291, %rs292, %rs293}, [%rd34];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1277, %rs290;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1278, %rs291;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1279, %rs292;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1280, %f1277, 0f00000000;
|
|
add.f32 %f1281, %f1278, 0f00000000;
|
|
add.f32 %f1282, %f1279, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs289, %f1282;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs288, %f1281;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs287, %f1280;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs294, 0;
|
|
st.global.v4.u16 [%rd34], {%rs287, %rs288, %rs289, %rs294};
|
|
bra.uni $L__BB0_173;
|
|
|
|
$L__BB0_172:
|
|
mov.f32 %f1285, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs297, %f1285;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs296, %f1285;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs295, %f1285;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs298, 0;
|
|
st.global.v4.u16 [%rd34], {%rs295, %rs296, %rs297, %rs298};
|
|
|
|
$L__BB0_173:
|
|
ld.const.u64 %rd151, [params+384];
|
|
cvta.to.global.u64 %rd152, %rd151;
|
|
ld.const.u32 %r351, [params+376];
|
|
mad.lo.s32 %r352, %r351, %r6, %r5;
|
|
mul.wide.u32 %rd153, %r352, 8;
|
|
add.s64 %rd35, %rd152, %rd153;
|
|
@%p156 bra $L__BB0_175;
|
|
|
|
ld.global.v4.u16 {%rs305, %rs306, %rs307, %rs308}, [%rd35];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1286, %rs305;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1287, %rs306;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1288, %rs307;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1289, %f1286, 0f00000000;
|
|
add.f32 %f1290, %f1287, 0f00000000;
|
|
add.f32 %f1291, %f1288, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs304, %f1291;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs303, %f1290;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs302, %f1289;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs309, 0;
|
|
st.global.v4.u16 [%rd35], {%rs302, %rs303, %rs304, %rs309};
|
|
bra.uni $L__BB0_176;
|
|
|
|
$L__BB0_175:
|
|
mov.f32 %f1294, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs312, %f1294;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs311, %f1294;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs310, %f1294;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs313, 0;
|
|
st.global.v4.u16 [%rd35], {%rs310, %rs311, %rs312, %rs313};
|
|
bra.uni $L__BB0_176;
|
|
|
|
$L__BB0_42:
|
|
mov.b32 %r221, %f1405;
|
|
xor.b32 %r222, %r221, -2147483648;
|
|
mov.b32 %f626, %r222;
|
|
selp.f32 %f1407, %f626, %f1405, %p1;
|
|
setp.geu.f32 %p54, %f1402, 0f00000000;
|
|
@%p54 bra $L__BB0_46;
|
|
|
|
mov.f32 %f627, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f628, %f627;
|
|
setp.eq.f32 %p55, %f628, 0f3EE66666;
|
|
@%p55 bra $L__BB0_46;
|
|
|
|
mov.f32 %f1407, 0f7FFFFFFF;
|
|
|
|
$L__BB0_46:
|
|
add.f32 %f631, %f216, 0f3EE66666;
|
|
mov.b32 %r223, %f631;
|
|
setp.lt.s32 %p57, %r223, 2139095040;
|
|
@%p57 bra $L__BB0_51;
|
|
|
|
setp.gtu.f32 %p58, %f216, 0f7F800000;
|
|
@%p58 bra $L__BB0_50;
|
|
bra.uni $L__BB0_48;
|
|
|
|
$L__BB0_50:
|
|
add.f32 %f1407, %f1402, 0f3EE66666;
|
|
bra.uni $L__BB0_51;
|
|
|
|
$L__BB0_48:
|
|
setp.neu.f32 %p59, %f216, 0f7F800000;
|
|
@%p59 bra $L__BB0_51;
|
|
|
|
selp.f32 %f1407, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
$L__BB0_51:
|
|
setp.eq.f32 %p60, %f1402, 0f3F800000;
|
|
selp.f32 %f227, 0f3F800000, %f1407, %p60;
|
|
abs.f32 %f228, %f1403;
|
|
setp.lt.f32 %p61, %f228, 0f00800000;
|
|
mul.f32 %f633, %f228, 0f4B800000;
|
|
selp.f32 %f634, %f633, %f228, %p61;
|
|
selp.f32 %f635, 0fC3170000, 0fC2FE0000, %p61;
|
|
mov.b32 %r224, %f634;
|
|
and.b32 %r225, %r224, 8388607;
|
|
or.b32 %r226, %r225, 1065353216;
|
|
mov.b32 %f636, %r226;
|
|
shr.u32 %r227, %r224, 23;
|
|
cvt.rn.f32.u32 %f637, %r227;
|
|
add.f32 %f638, %f635, %f637;
|
|
setp.gt.f32 %p62, %f636, 0f3FB504F3;
|
|
mul.f32 %f639, %f636, 0f3F000000;
|
|
add.f32 %f640, %f638, 0f3F800000;
|
|
selp.f32 %f641, %f640, %f638, %p62;
|
|
selp.f32 %f642, %f639, %f636, %p62;
|
|
add.f32 %f643, %f642, 0fBF800000;
|
|
add.f32 %f644, %f642, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f645, %f644;
|
|
add.f32 %f646, %f643, %f643;
|
|
mul.f32 %f647, %f646, %f645;
|
|
mul.f32 %f648, %f647, %f647;
|
|
mov.f32 %f649, 0f3C4CAF63;
|
|
mov.f32 %f650, 0f3B18F0FE;
|
|
fma.rn.f32 %f651, %f650, %f648, %f649;
|
|
mov.f32 %f652, 0f3DAAAABD;
|
|
fma.rn.f32 %f653, %f651, %f648, %f652;
|
|
mul.rn.f32 %f654, %f653, %f648;
|
|
mul.rn.f32 %f655, %f654, %f647;
|
|
sub.f32 %f656, %f643, %f647;
|
|
add.f32 %f657, %f656, %f656;
|
|
neg.f32 %f658, %f647;
|
|
fma.rn.f32 %f659, %f658, %f643, %f657;
|
|
mul.rn.f32 %f660, %f645, %f659;
|
|
add.f32 %f661, %f655, %f647;
|
|
sub.f32 %f662, %f647, %f661;
|
|
add.f32 %f663, %f655, %f662;
|
|
add.f32 %f664, %f660, %f663;
|
|
add.f32 %f665, %f661, %f664;
|
|
sub.f32 %f666, %f661, %f665;
|
|
add.f32 %f667, %f664, %f666;
|
|
mov.f32 %f668, 0f3F317200;
|
|
mul.rn.f32 %f669, %f641, %f668;
|
|
mov.f32 %f670, 0f35BFBE8E;
|
|
mul.rn.f32 %f671, %f641, %f670;
|
|
add.f32 %f672, %f669, %f665;
|
|
sub.f32 %f673, %f669, %f672;
|
|
add.f32 %f674, %f665, %f673;
|
|
add.f32 %f675, %f667, %f674;
|
|
add.f32 %f676, %f671, %f675;
|
|
add.f32 %f677, %f672, %f676;
|
|
sub.f32 %f678, %f672, %f677;
|
|
add.f32 %f679, %f676, %f678;
|
|
mov.f32 %f680, 0f3EE66666;
|
|
mul.rn.f32 %f681, %f680, %f677;
|
|
neg.f32 %f682, %f681;
|
|
fma.rn.f32 %f683, %f680, %f677, %f682;
|
|
fma.rn.f32 %f684, %f680, %f679, %f683;
|
|
mov.f32 %f685, 0f00000000;
|
|
fma.rn.f32 %f686, %f685, %f677, %f684;
|
|
add.rn.f32 %f687, %f681, %f686;
|
|
neg.f32 %f688, %f687;
|
|
add.rn.f32 %f689, %f681, %f688;
|
|
add.rn.f32 %f690, %f689, %f686;
|
|
mov.b32 %r228, %f687;
|
|
setp.eq.s32 %p63, %r228, 1118925336;
|
|
add.s32 %r229, %r228, -1;
|
|
mov.b32 %f691, %r229;
|
|
add.f32 %f692, %f690, 0f37000000;
|
|
selp.f32 %f229, %f692, %f690, %p63;
|
|
selp.f32 %f693, %f691, %f687, %p63;
|
|
mov.f32 %f694, 0f3FB8AA3B;
|
|
mul.rn.f32 %f695, %f693, %f694;
|
|
cvt.rzi.f32.f32 %f696, %f695;
|
|
abs.f32 %f697, %f696;
|
|
setp.gt.f32 %p64, %f697, 0f42FC0000;
|
|
mov.b32 %r230, %f696;
|
|
and.b32 %r231, %r230, -2147483648;
|
|
or.b32 %r232, %r231, 1123811328;
|
|
mov.b32 %f698, %r232;
|
|
selp.f32 %f699, %f698, %f696, %p64;
|
|
mov.f32 %f700, 0fBF317218;
|
|
fma.rn.f32 %f701, %f699, %f700, %f693;
|
|
mov.f32 %f702, 0f3102E308;
|
|
fma.rn.f32 %f703, %f699, %f702, %f701;
|
|
mul.f32 %f704, %f703, 0f3FB8AA3B;
|
|
add.f32 %f705, %f699, 0f4B40007F;
|
|
mov.b32 %r233, %f705;
|
|
shl.b32 %r234, %r233, 23;
|
|
mov.b32 %f706, %r234;
|
|
ex2.approx.ftz.f32 %f707, %f704;
|
|
mul.f32 %f230, %f707, %f706;
|
|
setp.eq.f32 %p65, %f230, 0f7F800000;
|
|
mov.f32 %f1408, 0f7F800000;
|
|
@%p65 bra $L__BB0_53;
|
|
|
|
fma.rn.f32 %f1408, %f230, %f229, %f230;
|
|
|
|
$L__BB0_53:
|
|
setp.lt.f32 %p66, %f1403, 0f00000000;
|
|
and.pred %p2, %p66, %p52;
|
|
setp.eq.f32 %p68, %f1403, 0f00000000;
|
|
@%p68 bra $L__BB0_57;
|
|
bra.uni $L__BB0_54;
|
|
|
|
$L__BB0_57:
|
|
add.f32 %f712, %f1403, %f1403;
|
|
selp.f32 %f1410, %f712, 0f00000000, %p52;
|
|
bra.uni $L__BB0_58;
|
|
|
|
$L__BB0_54:
|
|
mov.b32 %r235, %f1408;
|
|
xor.b32 %r236, %r235, -2147483648;
|
|
mov.b32 %f708, %r236;
|
|
selp.f32 %f1410, %f708, %f1408, %p2;
|
|
setp.geu.f32 %p69, %f1403, 0f00000000;
|
|
@%p69 bra $L__BB0_58;
|
|
|
|
mov.f32 %f709, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f710, %f709;
|
|
setp.eq.f32 %p70, %f710, 0f3EE66666;
|
|
@%p70 bra $L__BB0_58;
|
|
|
|
mov.f32 %f1410, 0f7FFFFFFF;
|
|
|
|
$L__BB0_58:
|
|
add.f32 %f713, %f228, 0f3EE66666;
|
|
mov.b32 %r237, %f713;
|
|
setp.lt.s32 %p72, %r237, 2139095040;
|
|
@%p72 bra $L__BB0_63;
|
|
|
|
setp.gtu.f32 %p73, %f228, 0f7F800000;
|
|
@%p73 bra $L__BB0_62;
|
|
bra.uni $L__BB0_60;
|
|
|
|
$L__BB0_62:
|
|
add.f32 %f1410, %f1403, 0f3EE66666;
|
|
bra.uni $L__BB0_63;
|
|
|
|
$L__BB0_60:
|
|
setp.neu.f32 %p74, %f228, 0f7F800000;
|
|
@%p74 bra $L__BB0_63;
|
|
|
|
selp.f32 %f1410, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
$L__BB0_63:
|
|
setp.eq.f32 %p75, %f1403, 0f3F800000;
|
|
selp.f32 %f239, 0f3F800000, %f1410, %p75;
|
|
abs.f32 %f240, %f1404;
|
|
setp.lt.f32 %p76, %f240, 0f00800000;
|
|
mul.f32 %f715, %f240, 0f4B800000;
|
|
selp.f32 %f716, %f715, %f240, %p76;
|
|
selp.f32 %f717, 0fC3170000, 0fC2FE0000, %p76;
|
|
mov.b32 %r238, %f716;
|
|
and.b32 %r239, %r238, 8388607;
|
|
or.b32 %r240, %r239, 1065353216;
|
|
mov.b32 %f718, %r240;
|
|
shr.u32 %r241, %r238, 23;
|
|
cvt.rn.f32.u32 %f719, %r241;
|
|
add.f32 %f720, %f717, %f719;
|
|
setp.gt.f32 %p77, %f718, 0f3FB504F3;
|
|
mul.f32 %f721, %f718, 0f3F000000;
|
|
add.f32 %f722, %f720, 0f3F800000;
|
|
selp.f32 %f723, %f722, %f720, %p77;
|
|
selp.f32 %f724, %f721, %f718, %p77;
|
|
add.f32 %f725, %f724, 0fBF800000;
|
|
add.f32 %f726, %f724, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f727, %f726;
|
|
add.f32 %f728, %f725, %f725;
|
|
mul.f32 %f729, %f728, %f727;
|
|
mul.f32 %f730, %f729, %f729;
|
|
mov.f32 %f731, 0f3C4CAF63;
|
|
mov.f32 %f732, 0f3B18F0FE;
|
|
fma.rn.f32 %f733, %f732, %f730, %f731;
|
|
mov.f32 %f734, 0f3DAAAABD;
|
|
fma.rn.f32 %f735, %f733, %f730, %f734;
|
|
mul.rn.f32 %f736, %f735, %f730;
|
|
mul.rn.f32 %f737, %f736, %f729;
|
|
sub.f32 %f738, %f725, %f729;
|
|
add.f32 %f739, %f738, %f738;
|
|
neg.f32 %f740, %f729;
|
|
fma.rn.f32 %f741, %f740, %f725, %f739;
|
|
mul.rn.f32 %f742, %f727, %f741;
|
|
add.f32 %f743, %f737, %f729;
|
|
sub.f32 %f744, %f729, %f743;
|
|
add.f32 %f745, %f737, %f744;
|
|
add.f32 %f746, %f742, %f745;
|
|
add.f32 %f747, %f743, %f746;
|
|
sub.f32 %f748, %f743, %f747;
|
|
add.f32 %f749, %f746, %f748;
|
|
mov.f32 %f750, 0f3F317200;
|
|
mul.rn.f32 %f751, %f723, %f750;
|
|
mov.f32 %f752, 0f35BFBE8E;
|
|
mul.rn.f32 %f753, %f723, %f752;
|
|
add.f32 %f754, %f751, %f747;
|
|
sub.f32 %f755, %f751, %f754;
|
|
add.f32 %f756, %f747, %f755;
|
|
add.f32 %f757, %f749, %f756;
|
|
add.f32 %f758, %f753, %f757;
|
|
add.f32 %f759, %f754, %f758;
|
|
sub.f32 %f760, %f754, %f759;
|
|
add.f32 %f761, %f758, %f760;
|
|
mov.f32 %f762, 0f3EE66666;
|
|
mul.rn.f32 %f763, %f762, %f759;
|
|
neg.f32 %f764, %f763;
|
|
fma.rn.f32 %f765, %f762, %f759, %f764;
|
|
fma.rn.f32 %f766, %f762, %f761, %f765;
|
|
mov.f32 %f767, 0f00000000;
|
|
fma.rn.f32 %f768, %f767, %f759, %f766;
|
|
add.rn.f32 %f769, %f763, %f768;
|
|
neg.f32 %f770, %f769;
|
|
add.rn.f32 %f771, %f763, %f770;
|
|
add.rn.f32 %f772, %f771, %f768;
|
|
mov.b32 %r242, %f769;
|
|
setp.eq.s32 %p78, %r242, 1118925336;
|
|
add.s32 %r243, %r242, -1;
|
|
mov.b32 %f773, %r243;
|
|
add.f32 %f774, %f772, 0f37000000;
|
|
selp.f32 %f241, %f774, %f772, %p78;
|
|
selp.f32 %f775, %f773, %f769, %p78;
|
|
mov.f32 %f776, 0f3FB8AA3B;
|
|
mul.rn.f32 %f777, %f775, %f776;
|
|
cvt.rzi.f32.f32 %f778, %f777;
|
|
abs.f32 %f779, %f778;
|
|
setp.gt.f32 %p79, %f779, 0f42FC0000;
|
|
mov.b32 %r244, %f778;
|
|
and.b32 %r245, %r244, -2147483648;
|
|
or.b32 %r246, %r245, 1123811328;
|
|
mov.b32 %f780, %r246;
|
|
selp.f32 %f781, %f780, %f778, %p79;
|
|
mov.f32 %f782, 0fBF317218;
|
|
fma.rn.f32 %f783, %f781, %f782, %f775;
|
|
mov.f32 %f784, 0f3102E308;
|
|
fma.rn.f32 %f785, %f781, %f784, %f783;
|
|
mul.f32 %f786, %f785, 0f3FB8AA3B;
|
|
add.f32 %f787, %f781, 0f4B40007F;
|
|
mov.b32 %r247, %f787;
|
|
shl.b32 %r248, %r247, 23;
|
|
mov.b32 %f788, %r248;
|
|
ex2.approx.ftz.f32 %f789, %f786;
|
|
mul.f32 %f242, %f789, %f788;
|
|
setp.eq.f32 %p80, %f242, 0f7F800000;
|
|
mov.f32 %f1411, 0f7F800000;
|
|
@%p80 bra $L__BB0_65;
|
|
|
|
fma.rn.f32 %f1411, %f242, %f241, %f242;
|
|
|
|
$L__BB0_65:
|
|
setp.lt.f32 %p81, %f1404, 0f00000000;
|
|
and.pred %p3, %p81, %p52;
|
|
setp.eq.f32 %p83, %f1404, 0f00000000;
|
|
@%p83 bra $L__BB0_69;
|
|
bra.uni $L__BB0_66;
|
|
|
|
$L__BB0_69:
|
|
add.f32 %f794, %f1404, %f1404;
|
|
selp.f32 %f1413, %f794, 0f00000000, %p52;
|
|
bra.uni $L__BB0_70;
|
|
|
|
$L__BB0_66:
|
|
mov.b32 %r249, %f1411;
|
|
xor.b32 %r250, %r249, -2147483648;
|
|
mov.b32 %f790, %r250;
|
|
selp.f32 %f1413, %f790, %f1411, %p3;
|
|
setp.geu.f32 %p84, %f1404, 0f00000000;
|
|
@%p84 bra $L__BB0_70;
|
|
|
|
mov.f32 %f791, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f792, %f791;
|
|
setp.eq.f32 %p85, %f792, 0f3EE66666;
|
|
@%p85 bra $L__BB0_70;
|
|
|
|
mov.f32 %f1413, 0f7FFFFFFF;
|
|
|
|
$L__BB0_70:
|
|
add.f32 %f795, %f240, 0f3EE66666;
|
|
mov.b32 %r251, %f795;
|
|
setp.lt.s32 %p87, %r251, 2139095040;
|
|
@%p87 bra $L__BB0_75;
|
|
|
|
setp.gtu.f32 %p88, %f240, 0f7F800000;
|
|
@%p88 bra $L__BB0_74;
|
|
bra.uni $L__BB0_72;
|
|
|
|
$L__BB0_74:
|
|
add.f32 %f1413, %f1404, 0f3EE66666;
|
|
bra.uni $L__BB0_75;
|
|
|
|
$L__BB0_72:
|
|
setp.neu.f32 %p89, %f240, 0f7F800000;
|
|
@%p89 bra $L__BB0_75;
|
|
|
|
selp.f32 %f1413, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
$L__BB0_75:
|
|
setp.eq.f32 %p90, %f1404, 0f3F800000;
|
|
mov.f32 %f797, 0f3F800000;
|
|
selp.f32 %f798, 0f3F800000, %f1413, %p90;
|
|
ld.const.u64 %rd77, [params+144];
|
|
cvta.to.global.u64 %rd16, %rd77;
|
|
ld.const.u32 %r252, [params+136];
|
|
mad.lo.s32 %r253, %r252, %r6, %r5;
|
|
cvt.u64.u32 %rd17, %r253;
|
|
min.f32 %f799, %f227, %f797;
|
|
mov.f32 %f800, 0f00000000;
|
|
max.f32 %f251, %f800, %f799;
|
|
min.f32 %f801, %f239, %f797;
|
|
max.f32 %f252, %f800, %f801;
|
|
min.f32 %f802, %f798, %f797;
|
|
max.f32 %f253, %f800, %f802;
|
|
mov.f32 %f803, 0f3E555555;
|
|
cvt.rzi.f32.f32 %f804, %f803;
|
|
add.f32 %f805, %f804, %f804;
|
|
mov.f32 %f806, 0f3ED55555;
|
|
sub.f32 %f807, %f806, %f805;
|
|
abs.f32 %f254, %f807;
|
|
abs.f32 %f255, %f251;
|
|
setp.lt.f32 %p91, %f255, 0f00800000;
|
|
mul.f32 %f808, %f255, 0f4B800000;
|
|
selp.f32 %f809, %f808, %f255, %p91;
|
|
selp.f32 %f810, 0fC3170000, 0fC2FE0000, %p91;
|
|
mov.b32 %r254, %f809;
|
|
and.b32 %r255, %r254, 8388607;
|
|
or.b32 %r256, %r255, 1065353216;
|
|
mov.b32 %f811, %r256;
|
|
shr.u32 %r257, %r254, 23;
|
|
cvt.rn.f32.u32 %f812, %r257;
|
|
add.f32 %f813, %f810, %f812;
|
|
setp.gt.f32 %p92, %f811, 0f3FB504F3;
|
|
mul.f32 %f814, %f811, 0f3F000000;
|
|
add.f32 %f815, %f813, 0f3F800000;
|
|
selp.f32 %f816, %f815, %f813, %p92;
|
|
selp.f32 %f817, %f814, %f811, %p92;
|
|
add.f32 %f818, %f817, 0fBF800000;
|
|
add.f32 %f819, %f817, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f820, %f819;
|
|
add.f32 %f821, %f818, %f818;
|
|
mul.f32 %f822, %f821, %f820;
|
|
mul.f32 %f823, %f822, %f822;
|
|
mov.f32 %f824, 0f3C4CAF63;
|
|
mov.f32 %f825, 0f3B18F0FE;
|
|
fma.rn.f32 %f826, %f825, %f823, %f824;
|
|
mov.f32 %f827, 0f3DAAAABD;
|
|
fma.rn.f32 %f828, %f826, %f823, %f827;
|
|
mul.rn.f32 %f829, %f828, %f823;
|
|
mul.rn.f32 %f830, %f829, %f822;
|
|
sub.f32 %f831, %f818, %f822;
|
|
add.f32 %f832, %f831, %f831;
|
|
neg.f32 %f833, %f822;
|
|
fma.rn.f32 %f834, %f833, %f818, %f832;
|
|
mul.rn.f32 %f835, %f820, %f834;
|
|
add.f32 %f836, %f830, %f822;
|
|
sub.f32 %f837, %f822, %f836;
|
|
add.f32 %f838, %f830, %f837;
|
|
add.f32 %f839, %f835, %f838;
|
|
add.f32 %f840, %f836, %f839;
|
|
sub.f32 %f841, %f836, %f840;
|
|
add.f32 %f842, %f839, %f841;
|
|
mov.f32 %f843, 0f3F317200;
|
|
mul.rn.f32 %f844, %f816, %f843;
|
|
mov.f32 %f845, 0f35BFBE8E;
|
|
mul.rn.f32 %f846, %f816, %f845;
|
|
add.f32 %f847, %f844, %f840;
|
|
sub.f32 %f848, %f844, %f847;
|
|
add.f32 %f849, %f840, %f848;
|
|
add.f32 %f850, %f842, %f849;
|
|
add.f32 %f851, %f846, %f850;
|
|
add.f32 %f852, %f847, %f851;
|
|
sub.f32 %f853, %f847, %f852;
|
|
add.f32 %f854, %f851, %f853;
|
|
mul.rn.f32 %f855, %f806, %f852;
|
|
neg.f32 %f856, %f855;
|
|
fma.rn.f32 %f857, %f806, %f852, %f856;
|
|
fma.rn.f32 %f858, %f806, %f854, %f857;
|
|
fma.rn.f32 %f859, %f800, %f852, %f858;
|
|
add.rn.f32 %f860, %f855, %f859;
|
|
neg.f32 %f861, %f860;
|
|
add.rn.f32 %f862, %f855, %f861;
|
|
add.rn.f32 %f863, %f862, %f859;
|
|
mov.b32 %r258, %f860;
|
|
setp.eq.s32 %p93, %r258, 1118925336;
|
|
add.s32 %r259, %r258, -1;
|
|
mov.b32 %f864, %r259;
|
|
add.f32 %f865, %f863, 0f37000000;
|
|
selp.f32 %f256, %f865, %f863, %p93;
|
|
selp.f32 %f866, %f864, %f860, %p93;
|
|
mov.f32 %f867, 0f3FB8AA3B;
|
|
mul.rn.f32 %f868, %f866, %f867;
|
|
cvt.rzi.f32.f32 %f869, %f868;
|
|
abs.f32 %f870, %f869;
|
|
setp.gt.f32 %p94, %f870, 0f42FC0000;
|
|
mov.b32 %r260, %f869;
|
|
and.b32 %r261, %r260, -2147483648;
|
|
or.b32 %r262, %r261, 1123811328;
|
|
mov.b32 %f871, %r262;
|
|
selp.f32 %f872, %f871, %f869, %p94;
|
|
mov.f32 %f873, 0fBF317218;
|
|
fma.rn.f32 %f874, %f872, %f873, %f866;
|
|
mov.f32 %f875, 0f3102E308;
|
|
fma.rn.f32 %f876, %f872, %f875, %f874;
|
|
mul.f32 %f877, %f876, 0f3FB8AA3B;
|
|
add.f32 %f878, %f872, 0f4B40007F;
|
|
mov.b32 %r263, %f878;
|
|
shl.b32 %r264, %r263, 23;
|
|
mov.b32 %f879, %r264;
|
|
ex2.approx.ftz.f32 %f880, %f877;
|
|
mul.f32 %f257, %f880, %f879;
|
|
setp.eq.f32 %p95, %f257, 0f7F800000;
|
|
mov.f32 %f1414, 0f7F800000;
|
|
@%p95 bra $L__BB0_77;
|
|
|
|
fma.rn.f32 %f1414, %f257, %f256, %f257;
|
|
|
|
$L__BB0_77:
|
|
setp.lt.f32 %p96, %f251, 0f00000000;
|
|
setp.eq.f32 %p97, %f254, 0f3F800000;
|
|
and.pred %p4, %p96, %p97;
|
|
setp.eq.f32 %p98, %f251, 0f00000000;
|
|
@%p98 bra $L__BB0_81;
|
|
bra.uni $L__BB0_78;
|
|
|
|
$L__BB0_81:
|
|
add.f32 %f885, %f251, %f251;
|
|
selp.f32 %f1416, %f885, 0f00000000, %p97;
|
|
bra.uni $L__BB0_82;
|
|
|
|
$L__BB0_78:
|
|
mov.b32 %r265, %f1414;
|
|
xor.b32 %r266, %r265, -2147483648;
|
|
mov.b32 %f881, %r266;
|
|
selp.f32 %f1416, %f881, %f1414, %p4;
|
|
setp.geu.f32 %p99, %f251, 0f00000000;
|
|
@%p99 bra $L__BB0_82;
|
|
|
|
mov.f32 %f882, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f883, %f882;
|
|
setp.eq.f32 %p100, %f883, 0f3ED55555;
|
|
@%p100 bra $L__BB0_82;
|
|
|
|
mov.f32 %f1416, 0f7FFFFFFF;
|
|
|
|
$L__BB0_82:
|
|
add.f32 %f886, %f255, 0f3ED55555;
|
|
mov.b32 %r267, %f886;
|
|
setp.lt.s32 %p102, %r267, 2139095040;
|
|
@%p102 bra $L__BB0_87;
|
|
|
|
setp.gtu.f32 %p103, %f255, 0f7F800000;
|
|
@%p103 bra $L__BB0_86;
|
|
bra.uni $L__BB0_84;
|
|
|
|
$L__BB0_86:
|
|
add.f32 %f1416, %f251, 0f3ED55555;
|
|
bra.uni $L__BB0_87;
|
|
|
|
$L__BB0_84:
|
|
setp.neu.f32 %p104, %f255, 0f7F800000;
|
|
@%p104 bra $L__BB0_87;
|
|
|
|
selp.f32 %f1416, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
$L__BB0_87:
|
|
abs.f32 %f266, %f252;
|
|
setp.lt.f32 %p105, %f266, 0f00800000;
|
|
mul.f32 %f888, %f266, 0f4B800000;
|
|
selp.f32 %f889, %f888, %f266, %p105;
|
|
selp.f32 %f890, 0fC3170000, 0fC2FE0000, %p105;
|
|
mov.b32 %r268, %f889;
|
|
and.b32 %r269, %r268, 8388607;
|
|
or.b32 %r270, %r269, 1065353216;
|
|
mov.b32 %f891, %r270;
|
|
shr.u32 %r271, %r268, 23;
|
|
cvt.rn.f32.u32 %f892, %r271;
|
|
add.f32 %f893, %f890, %f892;
|
|
setp.gt.f32 %p106, %f891, 0f3FB504F3;
|
|
mul.f32 %f894, %f891, 0f3F000000;
|
|
add.f32 %f895, %f893, 0f3F800000;
|
|
selp.f32 %f896, %f895, %f893, %p106;
|
|
selp.f32 %f897, %f894, %f891, %p106;
|
|
add.f32 %f898, %f897, 0fBF800000;
|
|
add.f32 %f899, %f897, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f900, %f899;
|
|
add.f32 %f901, %f898, %f898;
|
|
mul.f32 %f902, %f901, %f900;
|
|
mul.f32 %f903, %f902, %f902;
|
|
mov.f32 %f904, 0f3C4CAF63;
|
|
mov.f32 %f905, 0f3B18F0FE;
|
|
fma.rn.f32 %f906, %f905, %f903, %f904;
|
|
mov.f32 %f907, 0f3DAAAABD;
|
|
fma.rn.f32 %f908, %f906, %f903, %f907;
|
|
mul.rn.f32 %f909, %f908, %f903;
|
|
mul.rn.f32 %f910, %f909, %f902;
|
|
sub.f32 %f911, %f898, %f902;
|
|
add.f32 %f912, %f911, %f911;
|
|
neg.f32 %f913, %f902;
|
|
fma.rn.f32 %f914, %f913, %f898, %f912;
|
|
mul.rn.f32 %f915, %f900, %f914;
|
|
add.f32 %f916, %f910, %f902;
|
|
sub.f32 %f917, %f902, %f916;
|
|
add.f32 %f918, %f910, %f917;
|
|
add.f32 %f919, %f915, %f918;
|
|
add.f32 %f920, %f916, %f919;
|
|
sub.f32 %f921, %f916, %f920;
|
|
add.f32 %f922, %f919, %f921;
|
|
mov.f32 %f923, 0f3F317200;
|
|
mul.rn.f32 %f924, %f896, %f923;
|
|
mov.f32 %f925, 0f35BFBE8E;
|
|
mul.rn.f32 %f926, %f896, %f925;
|
|
add.f32 %f927, %f924, %f920;
|
|
sub.f32 %f928, %f924, %f927;
|
|
add.f32 %f929, %f920, %f928;
|
|
add.f32 %f930, %f922, %f929;
|
|
add.f32 %f931, %f926, %f930;
|
|
add.f32 %f932, %f927, %f931;
|
|
sub.f32 %f933, %f927, %f932;
|
|
add.f32 %f934, %f931, %f933;
|
|
mov.f32 %f935, 0f3ED55555;
|
|
mul.rn.f32 %f936, %f935, %f932;
|
|
neg.f32 %f937, %f936;
|
|
fma.rn.f32 %f938, %f935, %f932, %f937;
|
|
fma.rn.f32 %f939, %f935, %f934, %f938;
|
|
mov.f32 %f940, 0f00000000;
|
|
fma.rn.f32 %f941, %f940, %f932, %f939;
|
|
add.rn.f32 %f942, %f936, %f941;
|
|
neg.f32 %f943, %f942;
|
|
add.rn.f32 %f944, %f936, %f943;
|
|
add.rn.f32 %f945, %f944, %f941;
|
|
mov.b32 %r272, %f942;
|
|
setp.eq.s32 %p107, %r272, 1118925336;
|
|
add.s32 %r273, %r272, -1;
|
|
mov.b32 %f946, %r273;
|
|
add.f32 %f947, %f945, 0f37000000;
|
|
selp.f32 %f267, %f947, %f945, %p107;
|
|
selp.f32 %f948, %f946, %f942, %p107;
|
|
mov.f32 %f949, 0f3FB8AA3B;
|
|
mul.rn.f32 %f950, %f948, %f949;
|
|
cvt.rzi.f32.f32 %f951, %f950;
|
|
abs.f32 %f952, %f951;
|
|
setp.gt.f32 %p108, %f952, 0f42FC0000;
|
|
mov.b32 %r274, %f951;
|
|
and.b32 %r275, %r274, -2147483648;
|
|
or.b32 %r276, %r275, 1123811328;
|
|
mov.b32 %f953, %r276;
|
|
selp.f32 %f954, %f953, %f951, %p108;
|
|
mov.f32 %f955, 0fBF317218;
|
|
fma.rn.f32 %f956, %f954, %f955, %f948;
|
|
mov.f32 %f957, 0f3102E308;
|
|
fma.rn.f32 %f958, %f954, %f957, %f956;
|
|
mul.f32 %f959, %f958, 0f3FB8AA3B;
|
|
add.f32 %f960, %f954, 0f4B40007F;
|
|
mov.b32 %r277, %f960;
|
|
shl.b32 %r278, %r277, 23;
|
|
mov.b32 %f961, %r278;
|
|
ex2.approx.ftz.f32 %f962, %f959;
|
|
mul.f32 %f268, %f962, %f961;
|
|
setp.eq.f32 %p109, %f268, 0f7F800000;
|
|
mov.f32 %f1417, 0f7F800000;
|
|
@%p109 bra $L__BB0_89;
|
|
|
|
fma.rn.f32 %f1417, %f268, %f267, %f268;
|
|
|
|
$L__BB0_89:
|
|
setp.lt.f32 %p110, %f252, 0f00000000;
|
|
and.pred %p5, %p110, %p97;
|
|
setp.eq.f32 %p112, %f252, 0f00000000;
|
|
@%p112 bra $L__BB0_93;
|
|
bra.uni $L__BB0_90;
|
|
|
|
$L__BB0_93:
|
|
add.f32 %f967, %f252, %f252;
|
|
selp.f32 %f1419, %f967, 0f00000000, %p97;
|
|
bra.uni $L__BB0_94;
|
|
|
|
$L__BB0_90:
|
|
mov.b32 %r279, %f1417;
|
|
xor.b32 %r280, %r279, -2147483648;
|
|
mov.b32 %f963, %r280;
|
|
selp.f32 %f1419, %f963, %f1417, %p5;
|
|
setp.geu.f32 %p113, %f252, 0f00000000;
|
|
@%p113 bra $L__BB0_94;
|
|
|
|
mov.f32 %f964, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f965, %f964;
|
|
setp.eq.f32 %p114, %f965, 0f3ED55555;
|
|
@%p114 bra $L__BB0_94;
|
|
|
|
mov.f32 %f1419, 0f7FFFFFFF;
|
|
|
|
$L__BB0_94:
|
|
add.f32 %f968, %f266, 0f3ED55555;
|
|
mov.b32 %r281, %f968;
|
|
setp.lt.s32 %p116, %r281, 2139095040;
|
|
@%p116 bra $L__BB0_99;
|
|
|
|
setp.gtu.f32 %p117, %f266, 0f7F800000;
|
|
@%p117 bra $L__BB0_98;
|
|
bra.uni $L__BB0_96;
|
|
|
|
$L__BB0_98:
|
|
add.f32 %f1419, %f252, 0f3ED55555;
|
|
bra.uni $L__BB0_99;
|
|
|
|
$L__BB0_96:
|
|
setp.neu.f32 %p118, %f266, 0f7F800000;
|
|
@%p118 bra $L__BB0_99;
|
|
|
|
selp.f32 %f1419, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
$L__BB0_99:
|
|
abs.f32 %f277, %f253;
|
|
setp.lt.f32 %p119, %f277, 0f00800000;
|
|
mul.f32 %f970, %f277, 0f4B800000;
|
|
selp.f32 %f971, %f970, %f277, %p119;
|
|
selp.f32 %f972, 0fC3170000, 0fC2FE0000, %p119;
|
|
mov.b32 %r282, %f971;
|
|
and.b32 %r283, %r282, 8388607;
|
|
or.b32 %r284, %r283, 1065353216;
|
|
mov.b32 %f973, %r284;
|
|
shr.u32 %r285, %r282, 23;
|
|
cvt.rn.f32.u32 %f974, %r285;
|
|
add.f32 %f975, %f972, %f974;
|
|
setp.gt.f32 %p120, %f973, 0f3FB504F3;
|
|
mul.f32 %f976, %f973, 0f3F000000;
|
|
add.f32 %f977, %f975, 0f3F800000;
|
|
selp.f32 %f978, %f977, %f975, %p120;
|
|
selp.f32 %f979, %f976, %f973, %p120;
|
|
add.f32 %f980, %f979, 0fBF800000;
|
|
add.f32 %f981, %f979, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f982, %f981;
|
|
add.f32 %f983, %f980, %f980;
|
|
mul.f32 %f984, %f983, %f982;
|
|
mul.f32 %f985, %f984, %f984;
|
|
mov.f32 %f986, 0f3C4CAF63;
|
|
mov.f32 %f987, 0f3B18F0FE;
|
|
fma.rn.f32 %f988, %f987, %f985, %f986;
|
|
mov.f32 %f989, 0f3DAAAABD;
|
|
fma.rn.f32 %f990, %f988, %f985, %f989;
|
|
mul.rn.f32 %f991, %f990, %f985;
|
|
mul.rn.f32 %f992, %f991, %f984;
|
|
sub.f32 %f993, %f980, %f984;
|
|
add.f32 %f994, %f993, %f993;
|
|
neg.f32 %f995, %f984;
|
|
fma.rn.f32 %f996, %f995, %f980, %f994;
|
|
mul.rn.f32 %f997, %f982, %f996;
|
|
add.f32 %f998, %f992, %f984;
|
|
sub.f32 %f999, %f984, %f998;
|
|
add.f32 %f1000, %f992, %f999;
|
|
add.f32 %f1001, %f997, %f1000;
|
|
add.f32 %f1002, %f998, %f1001;
|
|
sub.f32 %f1003, %f998, %f1002;
|
|
add.f32 %f1004, %f1001, %f1003;
|
|
mov.f32 %f1005, 0f3F317200;
|
|
mul.rn.f32 %f1006, %f978, %f1005;
|
|
mov.f32 %f1007, 0f35BFBE8E;
|
|
mul.rn.f32 %f1008, %f978, %f1007;
|
|
add.f32 %f1009, %f1006, %f1002;
|
|
sub.f32 %f1010, %f1006, %f1009;
|
|
add.f32 %f1011, %f1002, %f1010;
|
|
add.f32 %f1012, %f1004, %f1011;
|
|
add.f32 %f1013, %f1008, %f1012;
|
|
add.f32 %f1014, %f1009, %f1013;
|
|
sub.f32 %f1015, %f1009, %f1014;
|
|
add.f32 %f1016, %f1013, %f1015;
|
|
mov.f32 %f1017, 0f3ED55555;
|
|
mul.rn.f32 %f1018, %f1017, %f1014;
|
|
neg.f32 %f1019, %f1018;
|
|
fma.rn.f32 %f1020, %f1017, %f1014, %f1019;
|
|
fma.rn.f32 %f1021, %f1017, %f1016, %f1020;
|
|
mov.f32 %f1022, 0f00000000;
|
|
fma.rn.f32 %f1023, %f1022, %f1014, %f1021;
|
|
add.rn.f32 %f1024, %f1018, %f1023;
|
|
neg.f32 %f1025, %f1024;
|
|
add.rn.f32 %f1026, %f1018, %f1025;
|
|
add.rn.f32 %f1027, %f1026, %f1023;
|
|
mov.b32 %r286, %f1024;
|
|
setp.eq.s32 %p121, %r286, 1118925336;
|
|
add.s32 %r287, %r286, -1;
|
|
mov.b32 %f1028, %r287;
|
|
add.f32 %f1029, %f1027, 0f37000000;
|
|
selp.f32 %f278, %f1029, %f1027, %p121;
|
|
selp.f32 %f1030, %f1028, %f1024, %p121;
|
|
mov.f32 %f1031, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1032, %f1030, %f1031;
|
|
cvt.rzi.f32.f32 %f1033, %f1032;
|
|
abs.f32 %f1034, %f1033;
|
|
setp.gt.f32 %p122, %f1034, 0f42FC0000;
|
|
mov.b32 %r288, %f1033;
|
|
and.b32 %r289, %r288, -2147483648;
|
|
or.b32 %r290, %r289, 1123811328;
|
|
mov.b32 %f1035, %r290;
|
|
selp.f32 %f1036, %f1035, %f1033, %p122;
|
|
mov.f32 %f1037, 0fBF317218;
|
|
fma.rn.f32 %f1038, %f1036, %f1037, %f1030;
|
|
mov.f32 %f1039, 0f3102E308;
|
|
fma.rn.f32 %f1040, %f1036, %f1039, %f1038;
|
|
mul.f32 %f1041, %f1040, 0f3FB8AA3B;
|
|
add.f32 %f1042, %f1036, 0f4B40007F;
|
|
mov.b32 %r291, %f1042;
|
|
shl.b32 %r292, %r291, 23;
|
|
mov.b32 %f1043, %r292;
|
|
ex2.approx.ftz.f32 %f1044, %f1041;
|
|
mul.f32 %f279, %f1044, %f1043;
|
|
setp.eq.f32 %p123, %f279, 0f7F800000;
|
|
mov.f32 %f1420, 0f7F800000;
|
|
@%p123 bra $L__BB0_101;
|
|
|
|
fma.rn.f32 %f1420, %f279, %f278, %f279;
|
|
|
|
$L__BB0_101:
|
|
setp.lt.f32 %p124, %f253, 0f00000000;
|
|
and.pred %p6, %p124, %p97;
|
|
setp.eq.f32 %p126, %f253, 0f00000000;
|
|
@%p126 bra $L__BB0_105;
|
|
bra.uni $L__BB0_102;
|
|
|
|
$L__BB0_105:
|
|
add.f32 %f1049, %f253, %f253;
|
|
selp.f32 %f1422, %f1049, 0f00000000, %p97;
|
|
bra.uni $L__BB0_106;
|
|
|
|
$L__BB0_102:
|
|
mov.b32 %r293, %f1420;
|
|
xor.b32 %r294, %r293, -2147483648;
|
|
mov.b32 %f1045, %r294;
|
|
selp.f32 %f1422, %f1045, %f1420, %p6;
|
|
setp.geu.f32 %p127, %f253, 0f00000000;
|
|
@%p127 bra $L__BB0_106;
|
|
|
|
mov.f32 %f1046, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1047, %f1046;
|
|
setp.eq.f32 %p128, %f1047, 0f3ED55555;
|
|
@%p128 bra $L__BB0_106;
|
|
|
|
mov.f32 %f1422, 0f7FFFFFFF;
|
|
|
|
$L__BB0_106:
|
|
add.f32 %f1050, %f277, 0f3ED55555;
|
|
mov.b32 %r295, %f1050;
|
|
setp.lt.s32 %p130, %r295, 2139095040;
|
|
@%p130 bra $L__BB0_111;
|
|
|
|
setp.gtu.f32 %p131, %f277, 0f7F800000;
|
|
@%p131 bra $L__BB0_110;
|
|
bra.uni $L__BB0_108;
|
|
|
|
$L__BB0_110:
|
|
add.f32 %f1422, %f253, 0f3ED55555;
|
|
bra.uni $L__BB0_111;
|
|
|
|
$L__BB0_108:
|
|
setp.neu.f32 %p132, %f277, 0f7F800000;
|
|
@%p132 bra $L__BB0_111;
|
|
|
|
selp.f32 %f1422, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
$L__BB0_111:
|
|
fma.rn.f32 %f1051, %f1416, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p133, %f251, 0f3F800000;
|
|
mov.f32 %f1052, 0f3F800000;
|
|
selp.f32 %f1053, 0f3F7FFFFF, %f1051, %p133;
|
|
mul.f32 %f1054, %f251, 0f414EB852;
|
|
setp.lt.f32 %p134, %f251, 0f3B4D2E1C;
|
|
selp.f32 %f1055, %f1054, %f1053, %p134;
|
|
fma.rn.f32 %f1056, %f1419, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p135, %f252, 0f3F800000;
|
|
selp.f32 %f1057, 0f3F7FFFFF, %f1056, %p135;
|
|
mul.f32 %f1058, %f252, 0f414EB852;
|
|
setp.lt.f32 %p136, %f252, 0f3B4D2E1C;
|
|
selp.f32 %f1059, %f1058, %f1057, %p136;
|
|
fma.rn.f32 %f1060, %f1422, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p137, %f253, 0f3F800000;
|
|
selp.f32 %f1061, 0f3F7FFFFF, %f1060, %p137;
|
|
mul.f32 %f1062, %f253, 0f414EB852;
|
|
setp.lt.f32 %p138, %f253, 0f3B4D2E1C;
|
|
selp.f32 %f1063, %f1062, %f1061, %p138;
|
|
min.f32 %f1064, %f1055, %f1052;
|
|
mov.f32 %f1065, 0f00000000;
|
|
max.f32 %f1066, %f1065, %f1064;
|
|
mul.f32 %f1067, %f1066, 0f43800000;
|
|
cvt.rzi.u32.f32 %r296, %f1067;
|
|
min.u32 %r297, %r296, 255;
|
|
min.f32 %f1068, %f1059, %f1052;
|
|
max.f32 %f1069, %f1065, %f1068;
|
|
mul.f32 %f1070, %f1069, 0f43800000;
|
|
cvt.rzi.u32.f32 %r298, %f1070;
|
|
min.u32 %r299, %r298, 255;
|
|
min.f32 %f1071, %f1063, %f1052;
|
|
max.f32 %f1072, %f1065, %f1071;
|
|
mul.f32 %f1073, %f1072, 0f43800000;
|
|
cvt.rzi.u32.f32 %r300, %f1073;
|
|
min.u32 %r301, %r300, 255;
|
|
shl.b64 %rd78, %rd17, 2;
|
|
add.s64 %rd79, %rd16, %rd78;
|
|
cvt.u16.u32 %rs23, %r301;
|
|
cvt.u16.u32 %rs24, %r299;
|
|
cvt.u16.u32 %rs25, %r297;
|
|
mov.u16 %rs26, 255;
|
|
st.global.v4.u8 [%rd79], {%rs25, %rs24, %rs23, %rs26};
|
|
|
|
$L__BB0_112:
|
|
and.b32 %r302, %r49, 4;
|
|
setp.eq.s32 %p139, %r302, 0;
|
|
@%p139 bra $L__BB0_114;
|
|
|
|
ld.const.u64 %rd80, [params+224];
|
|
cvta.to.global.u64 %rd81, %rd80;
|
|
ld.const.u32 %r303, [params+216];
|
|
mad.lo.s32 %r304, %r303, %r6, %r5;
|
|
mov.f32 %f1077, 0f3F800000;
|
|
mul.wide.u32 %rd82, %r304, 8;
|
|
add.s64 %rd83, %rd81, %rd82;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f1077;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f1404;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f1403;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1402;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd83], {%rs27, %rs28, %rs29, %rs30};
|
|
|
|
$L__BB0_114:
|
|
and.b32 %r305, %r49, 16;
|
|
setp.eq.s32 %p140, %r305, 0;
|
|
@%p140 bra $L__BB0_116;
|
|
|
|
ld.const.u64 %rd84, [params+240];
|
|
cvta.to.global.u64 %rd85, %rd84;
|
|
ld.const.u32 %r306, [params+232];
|
|
mad.lo.s32 %r307, %r306, %r6, %r5;
|
|
mov.f32 %f1081, 0f3F800000;
|
|
mul.wide.u32 %rd86, %r307, 8;
|
|
add.s64 %rd87, %rd85, %rd86;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f1081;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f202;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f201;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f200;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd87], {%rs31, %rs32, %rs33, %rs34};
|
|
|
|
$L__BB0_116:
|
|
mul.f32 %f1082, %f173, 0f40800000;
|
|
mul.f32 %f288, %f1082, 0f3F558750;
|
|
mul.f32 %f1083, %f174, 0f40800000;
|
|
mul.f32 %f289, %f1083, 0f3F558750;
|
|
mul.f32 %f1084, %f175, 0f40800000;
|
|
mul.f32 %f290, %f1084, 0f3F558750;
|
|
ld.const.u32 %r50, [params+108];
|
|
setp.eq.s32 %p141, %r50, 0;
|
|
ld.const.u64 %rd88, [params+256];
|
|
cvta.to.global.u64 %rd89, %rd88;
|
|
ld.const.u32 %r308, [params+248];
|
|
mad.lo.s32 %r309, %r308, %r6, %r5;
|
|
mul.wide.u32 %rd90, %r309, 8;
|
|
add.s64 %rd18, %rd89, %rd90;
|
|
@%p141 bra $L__BB0_118;
|
|
|
|
ld.global.v4.u16 {%rs42, %rs43, %rs44, %rs45}, [%rd18];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1085, %rs42;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1086, %rs43;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1087, %rs44;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1088, %f288, %f1085;
|
|
add.f32 %f1089, %f289, %f1086;
|
|
add.f32 %f1090, %f290, %f1087;
|
|
mov.f32 %f1091, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f1090;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f1089;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f1088;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1091;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd18], {%rs38, %rs39, %rs40, %rs41};
|
|
bra.uni $L__BB0_119;
|
|
|
|
$L__BB0_118:
|
|
mov.f32 %f1095, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f1095;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f290;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f289;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f288;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd18], {%rs46, %rs47, %rs48, %rs49};
|
|
|
|
$L__BB0_119:
|
|
mul.f32 %f1096, %f176, 0f40800000;
|
|
mul.f32 %f291, %f1096, 0f3FB8EBD1;
|
|
mul.f32 %f1097, %f177, 0f40800000;
|
|
mul.f32 %f292, %f1097, 0f3FB8EBD1;
|
|
mul.f32 %f1098, %f178, 0f40800000;
|
|
mul.f32 %f293, %f1098, 0f3FB8EBD1;
|
|
ld.const.u64 %rd91, [params+272];
|
|
cvta.to.global.u64 %rd92, %rd91;
|
|
ld.const.u32 %r310, [params+264];
|
|
mad.lo.s32 %r311, %r310, %r6, %r5;
|
|
mul.wide.u32 %rd93, %r311, 8;
|
|
add.s64 %rd19, %rd92, %rd93;
|
|
@%p141 bra $L__BB0_121;
|
|
|
|
ld.global.v4.u16 {%rs57, %rs58, %rs59, %rs60}, [%rd19];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1099, %rs57;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1100, %rs58;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1101, %rs59;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1102, %f291, %f1099;
|
|
add.f32 %f1103, %f292, %f1100;
|
|
add.f32 %f1104, %f293, %f1101;
|
|
mov.f32 %f1105, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1104;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f1103;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f1102;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs56, %f1105;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd19], {%rs53, %rs54, %rs55, %rs56};
|
|
bra.uni $L__BB0_122;
|
|
|
|
$L__BB0_121:
|
|
mov.f32 %f1109, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f1109;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f293;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f292;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f291;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd19], {%rs61, %rs62, %rs63, %rs64};
|
|
|
|
$L__BB0_122:
|
|
mul.f32 %f1110, %f179, 0f40800000;
|
|
mul.f32 %f294, %f1110, 0f3FB8EBD1;
|
|
mul.f32 %f1111, %f180, 0f40800000;
|
|
mul.f32 %f295, %f1111, 0f3FB8EBD1;
|
|
mul.f32 %f1112, %f181, 0f40800000;
|
|
mul.f32 %f296, %f1112, 0f3FB8EBD1;
|
|
ld.const.u64 %rd94, [params+288];
|
|
cvta.to.global.u64 %rd95, %rd94;
|
|
ld.const.u32 %r312, [params+280];
|
|
mad.lo.s32 %r313, %r312, %r6, %r5;
|
|
mul.wide.u32 %rd96, %r313, 8;
|
|
add.s64 %rd20, %rd95, %rd96;
|
|
@%p141 bra $L__BB0_124;
|
|
|
|
ld.global.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd20];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1113, %rs72;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1114, %rs73;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1115, %rs74;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1116, %f294, %f1113;
|
|
add.f32 %f1117, %f295, %f1114;
|
|
add.f32 %f1118, %f296, %f1115;
|
|
mov.f32 %f1119, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f1118;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1117;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f1116;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f1119;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd20], {%rs68, %rs69, %rs70, %rs71};
|
|
bra.uni $L__BB0_125;
|
|
|
|
$L__BB0_124:
|
|
mov.f32 %f1123, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f1123;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f296;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f295;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f294;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd20], {%rs76, %rs77, %rs78, %rs79};
|
|
|
|
$L__BB0_125:
|
|
mul.f32 %f1124, %f182, 0f40800000;
|
|
mul.f32 %f297, %f1124, 0f3FB8EBD1;
|
|
mul.f32 %f1125, %f183, 0f40800000;
|
|
mul.f32 %f298, %f1125, 0f3FB8EBD1;
|
|
mul.f32 %f1126, %f184, 0f40800000;
|
|
mul.f32 %f299, %f1126, 0f3FB8EBD1;
|
|
ld.const.u64 %rd97, [params+304];
|
|
cvta.to.global.u64 %rd98, %rd97;
|
|
ld.const.u32 %r314, [params+296];
|
|
mad.lo.s32 %r315, %r314, %r6, %r5;
|
|
mul.wide.u32 %rd99, %r315, 8;
|
|
add.s64 %rd21, %rd98, %rd99;
|
|
@%p141 bra $L__BB0_127;
|
|
|
|
ld.global.v4.u16 {%rs87, %rs88, %rs89, %rs90}, [%rd21];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1127, %rs87;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1128, %rs88;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1129, %rs89;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1130, %f297, %f1127;
|
|
add.f32 %f1131, %f298, %f1128;
|
|
add.f32 %f1132, %f299, %f1129;
|
|
mov.f32 %f1133, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f1132;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs84, %f1131;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1130;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f1133;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd21], {%rs83, %rs84, %rs85, %rs86};
|
|
bra.uni $L__BB0_128;
|
|
|
|
$L__BB0_127:
|
|
mov.f32 %f1137, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f1137;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f299;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f298;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f297;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd21], {%rs91, %rs92, %rs93, %rs94};
|
|
|
|
$L__BB0_128:
|
|
mul.f32 %f1138, %f185, 0f40800000;
|
|
mul.f32 %f300, %f1138, 0f404EBF87;
|
|
mul.f32 %f1139, %f186, 0f40800000;
|
|
mul.f32 %f301, %f1139, 0f404EBF87;
|
|
mul.f32 %f1140, %f187, 0f40800000;
|
|
mul.f32 %f302, %f1140, 0f404EBF87;
|
|
ld.const.u64 %rd100, [params+320];
|
|
cvta.to.global.u64 %rd101, %rd100;
|
|
ld.const.u32 %r316, [params+312];
|
|
mad.lo.s32 %r317, %r316, %r6, %r5;
|
|
mul.wide.u32 %rd102, %r317, 8;
|
|
add.s64 %rd22, %rd101, %rd102;
|
|
@%p141 bra $L__BB0_130;
|
|
|
|
ld.global.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd22];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1141, %rs102;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1142, %rs103;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1143, %rs104;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1144, %f300, %f1141;
|
|
add.f32 %f1145, %f301, %f1142;
|
|
add.f32 %f1146, %f302, %f1143;
|
|
mov.f32 %f1147, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f1146;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f1145;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f1144;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f1147;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd22], {%rs98, %rs99, %rs100, %rs101};
|
|
bra.uni $L__BB0_131;
|
|
|
|
$L__BB0_130:
|
|
mov.f32 %f1151, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f1151;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f302;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f301;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs106, %f300;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd22], {%rs106, %rs107, %rs108, %rs109};
|
|
|
|
$L__BB0_131:
|
|
mul.f32 %f1152, %f188, 0f40800000;
|
|
mul.f32 %f303, %f1152, 0f404EBF87;
|
|
mul.f32 %f1153, %f189, 0f40800000;
|
|
mul.f32 %f304, %f1153, 0f404EBF87;
|
|
mul.f32 %f1154, %f190, 0f40800000;
|
|
mul.f32 %f305, %f1154, 0f404EBF87;
|
|
ld.const.u64 %rd103, [params+336];
|
|
cvta.to.global.u64 %rd104, %rd103;
|
|
ld.const.u32 %r318, [params+328];
|
|
mad.lo.s32 %r319, %r318, %r6, %r5;
|
|
mul.wide.u32 %rd105, %r319, 8;
|
|
add.s64 %rd23, %rd104, %rd105;
|
|
@%p141 bra $L__BB0_133;
|
|
|
|
ld.global.v4.u16 {%rs117, %rs118, %rs119, %rs120}, [%rd23];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1155, %rs117;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1156, %rs118;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1157, %rs119;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1158, %f303, %f1155;
|
|
add.f32 %f1159, %f304, %f1156;
|
|
add.f32 %f1160, %f305, %f1157;
|
|
mov.f32 %f1161, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f1160;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f1159;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f1158;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f1161;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd23], {%rs113, %rs114, %rs115, %rs116};
|
|
bra.uni $L__BB0_134;
|
|
|
|
$L__BB0_133:
|
|
mov.f32 %f1165, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f1165;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f305;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f304;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f303;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd23], {%rs121, %rs122, %rs123, %rs124};
|
|
|
|
$L__BB0_134:
|
|
mul.f32 %f1166, %f191, 0f40800000;
|
|
mul.f32 %f306, %f1166, 0f40330C94;
|
|
mul.f32 %f1167, %f192, 0f40800000;
|
|
mul.f32 %f307, %f1167, 0f40330C94;
|
|
mul.f32 %f1168, %f193, 0f40800000;
|
|
mul.f32 %f308, %f1168, 0f40330C94;
|
|
ld.const.u64 %rd106, [params+352];
|
|
cvta.to.global.u64 %rd107, %rd106;
|
|
ld.const.u32 %r320, [params+344];
|
|
mad.lo.s32 %r321, %r320, %r6, %r5;
|
|
mul.wide.u32 %rd108, %r321, 8;
|
|
add.s64 %rd24, %rd107, %rd108;
|
|
@%p141 bra $L__BB0_136;
|
|
|
|
ld.global.v4.u16 {%rs132, %rs133, %rs134, %rs135}, [%rd24];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1169, %rs132;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1170, %rs133;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1171, %rs134;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1172, %f306, %f1169;
|
|
add.f32 %f1173, %f307, %f1170;
|
|
add.f32 %f1174, %f308, %f1171;
|
|
mov.f32 %f1175, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs130, %f1174;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f1173;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f1172;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs131, %f1175;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd24], {%rs128, %rs129, %rs130, %rs131};
|
|
bra.uni $L__BB0_137;
|
|
|
|
$L__BB0_136:
|
|
mov.f32 %f1179, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs139, %f1179;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f308;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f307;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f306;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd24], {%rs136, %rs137, %rs138, %rs139};
|
|
|
|
$L__BB0_137:
|
|
mul.f32 %f1180, %f194, 0f40800000;
|
|
mul.f32 %f309, %f1180, 0f404EBF87;
|
|
mul.f32 %f1181, %f195, 0f40800000;
|
|
mul.f32 %f310, %f1181, 0f404EBF87;
|
|
mul.f32 %f1182, %f196, 0f40800000;
|
|
mul.f32 %f311, %f1182, 0f404EBF87;
|
|
ld.const.u64 %rd109, [params+368];
|
|
cvta.to.global.u64 %rd110, %rd109;
|
|
ld.const.u32 %r322, [params+360];
|
|
mad.lo.s32 %r323, %r322, %r6, %r5;
|
|
mul.wide.u32 %rd111, %r323, 8;
|
|
add.s64 %rd25, %rd110, %rd111;
|
|
@%p141 bra $L__BB0_139;
|
|
|
|
ld.global.v4.u16 {%rs147, %rs148, %rs149, %rs150}, [%rd25];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1183, %rs147;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1184, %rs148;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1185, %rs149;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1186, %f309, %f1183;
|
|
add.f32 %f1187, %f310, %f1184;
|
|
add.f32 %f1188, %f311, %f1185;
|
|
mov.f32 %f1189, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs145, %f1188;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f1187;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f1186;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f1189;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd25], {%rs143, %rs144, %rs145, %rs146};
|
|
bra.uni $L__BB0_140;
|
|
|
|
$L__BB0_139:
|
|
mov.f32 %f1193, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs154, %f1193;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs153, %f311;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f310;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f309;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd25], {%rs151, %rs152, %rs153, %rs154};
|
|
|
|
$L__BB0_140:
|
|
mul.f32 %f1194, %f197, 0f40800000;
|
|
mul.f32 %f312, %f1194, 0f3FCEBF87;
|
|
mul.f32 %f1195, %f198, 0f40800000;
|
|
mul.f32 %f313, %f1195, 0f3FCEBF87;
|
|
mul.f32 %f1196, %f199, 0f40800000;
|
|
mul.f32 %f314, %f1196, 0f3FCEBF87;
|
|
ld.const.u64 %rd112, [params+384];
|
|
cvta.to.global.u64 %rd113, %rd112;
|
|
ld.const.u32 %r324, [params+376];
|
|
mad.lo.s32 %r325, %r324, %r6, %r5;
|
|
mul.wide.u32 %rd114, %r325, 8;
|
|
add.s64 %rd26, %rd113, %rd114;
|
|
@%p141 bra $L__BB0_142;
|
|
|
|
ld.global.v4.u16 {%rs162, %rs163, %rs164, %rs165}, [%rd26];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1197, %rs162;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1198, %rs163;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1199, %rs164;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1200, %f312, %f1197;
|
|
add.f32 %f1201, %f313, %f1198;
|
|
add.f32 %f1202, %f314, %f1199;
|
|
mov.f32 %f1203, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs160, %f1202;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs159, %f1201;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs158, %f1200;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs161, %f1203;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd26], {%rs158, %rs159, %rs160, %rs161};
|
|
bra.uni $L__BB0_176;
|
|
|
|
$L__BB0_142:
|
|
mov.f32 %f1207, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs169, %f1207;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs168, %f314;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs167, %f313;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs166, %f312;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd26], {%rs166, %rs167, %rs168, %rs169};
|
|
|
|
$L__BB0_176:
|
|
ret;
|
|
|
|
}
|
|
|