3438 lines
108 KiB
Plaintext
3438 lines
108 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<256>;
|
|
.reg .b16 %rs<206>;
|
|
.reg .f32 %f<2136>;
|
|
.reg .b32 %r<1551>;
|
|
.reg .f64 %fd<5>;
|
|
.reg .b64 %rd<131>;
|
|
|
|
|
|
mov.u64 %SPL, __local_depot0;
|
|
add.u64 %rd1, %SPL, 0;
|
|
// begin inline asm
|
|
call (%r63), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r64), _optix_get_launch_index_y, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r66), _optix_get_launch_dimension_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r67), _optix_get_launch_dimension_y, ();
|
|
// end inline asm
|
|
ld.const.u64 %rd29, [params+400];
|
|
cvta.to.global.u64 %rd30, %rd29;
|
|
ld.const.u32 %r69, [params+392];
|
|
mad.lo.s32 %r70, %r69, %r64, %r63;
|
|
mul.wide.u32 %rd31, %r70, 4;
|
|
add.s64 %rd2, %rd30, %rd31;
|
|
ld.global.v2.u8 {%rs24, %rs193}, [%rd2];
|
|
or.b16 %rs26, %rs24, %rs193;
|
|
and.b16 %rs27, %rs26, 255;
|
|
setp.eq.s16 %p9, %rs27, 0;
|
|
@%p9 bra $L__BB0_2;
|
|
|
|
ld.global.u8 %rs192, [%rd2+2];
|
|
bra.uni $L__BB0_3;
|
|
|
|
$L__BB0_2:
|
|
ld.global.u8 %rs192, [%rd2+2];
|
|
setp.eq.s16 %p10, %rs192, 0;
|
|
mov.f32 %f2043, 0f00000000;
|
|
mov.u16 %rs193, 0;
|
|
mov.f32 %f2044, %f2043;
|
|
mov.f32 %f2045, %f2043;
|
|
@%p10 bra $L__BB0_4;
|
|
|
|
$L__BB0_3:
|
|
cvt.rn.f32.u16 %f327, %rs24;
|
|
div.rn.f32 %f328, %f327, 0f437F0000;
|
|
fma.rn.f32 %f329, %f328, 0f40000000, 0fBF800000;
|
|
and.b16 %rs30, %rs193, 255;
|
|
cvt.rn.f32.u16 %f330, %rs30;
|
|
div.rn.f32 %f331, %f330, 0f437F0000;
|
|
fma.rn.f32 %f332, %f331, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f333, %rs192;
|
|
div.rn.f32 %f334, %f333, 0f437F0000;
|
|
fma.rn.f32 %f335, %f334, 0f40000000, 0fBF800000;
|
|
mul.f32 %f336, %f332, %f332;
|
|
fma.rn.f32 %f337, %f329, %f329, %f336;
|
|
fma.rn.f32 %f338, %f335, %f335, %f337;
|
|
sqrt.rn.f32 %f339, %f338;
|
|
rcp.rn.f32 %f340, %f339;
|
|
mul.f32 %f2045, %f340, %f335;
|
|
mul.f32 %f2044, %f340, %f332;
|
|
mul.f32 %f2043, %f329, %f340;
|
|
|
|
$L__BB0_4:
|
|
ld.const.v2.u32 {%r71, %r72}, [params];
|
|
add.s32 %r5, %r71, %r63;
|
|
add.s32 %r6, %r72, %r64;
|
|
setp.eq.f32 %p11, %f2043, 0f00000000;
|
|
setp.eq.f32 %p12, %f2044, 0f00000000;
|
|
and.pred %p13, %p11, %p12;
|
|
setp.eq.f32 %p14, %f2045, 0f00000000;
|
|
and.pred %p15, %p14, %p13;
|
|
@%p15 bra $L__BB0_175;
|
|
bra.uni $L__BB0_5;
|
|
|
|
$L__BB0_175:
|
|
ld.const.u32 %r60, [params+104];
|
|
and.b32 %r1522, %r60, 1;
|
|
setp.eq.b32 %p246, %r1522, 1;
|
|
mov.pred %p247, 0;
|
|
xor.pred %p248, %p246, %p247;
|
|
not.pred %p249, %p248;
|
|
@%p249 bra $L__BB0_177;
|
|
|
|
ld.const.u64 %rd106, [params+144];
|
|
cvta.to.global.u64 %rd107, %rd106;
|
|
ld.const.u32 %r1523, [params+136];
|
|
mad.lo.s32 %r1524, %r1523, %r6, %r5;
|
|
mul.wide.u32 %rd108, %r1524, 4;
|
|
add.s64 %rd109, %rd107, %rd108;
|
|
mov.u16 %rs129, 0;
|
|
st.global.v4.u8 [%rd109], {%rs129, %rs129, %rs129, %rs129};
|
|
|
|
$L__BB0_177:
|
|
and.b32 %r1525, %r60, 8;
|
|
setp.eq.s32 %p250, %r1525, 0;
|
|
@%p250 bra $L__BB0_179;
|
|
|
|
ld.const.u64 %rd110, [params+192];
|
|
cvta.to.global.u64 %rd111, %rd110;
|
|
ld.const.u32 %r1526, [params+184];
|
|
mad.lo.s32 %r1527, %r1526, %r6, %r5;
|
|
mov.f32 %f2005, 0f00000000;
|
|
cvt.rzi.u32.f32 %r1528, %f2005;
|
|
mul.wide.u32 %rd112, %r1527, 2;
|
|
add.s64 %rd113, %rd111, %rd112;
|
|
mov.u16 %rs130, 0;
|
|
cvt.u16.u32 %rs131, %r1528;
|
|
st.global.v2.u8 [%rd113], {%rs131, %rs130};
|
|
|
|
$L__BB0_179:
|
|
and.b32 %r1529, %r60, 4;
|
|
setp.eq.s32 %p251, %r1529, 0;
|
|
ld.const.u32 %r1550, [params+108];
|
|
@%p251 bra $L__BB0_183;
|
|
|
|
setp.eq.s32 %p252, %r1550, 0;
|
|
ld.const.u64 %rd114, [params+224];
|
|
cvta.to.global.u64 %rd115, %rd114;
|
|
ld.const.u32 %r1530, [params+216];
|
|
mad.lo.s32 %r1531, %r1530, %r6, %r5;
|
|
mul.wide.u32 %rd116, %r1531, 8;
|
|
add.s64 %rd24, %rd115, %rd116;
|
|
@%p252 bra $L__BB0_182;
|
|
|
|
ld.global.v4.u16 {%rs138, %rs139, %rs140, %rs141}, [%rd24];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2006, %rs138;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2007, %rs139;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2008, %rs140;}
|
|
|
|
// end inline asm
|
|
add.f32 %f2009, %f2006, 0f00000000;
|
|
add.f32 %f2010, %f2007, 0f00000000;
|
|
add.f32 %f2011, %f2008, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f2011;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f2010;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f2009;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs142, 0;
|
|
st.global.v4.u16 [%rd24], {%rs135, %rs136, %rs137, %rs142};
|
|
bra.uni $L__BB0_183;
|
|
|
|
$L__BB0_5:
|
|
ld.const.v4.f32 {%f345, %f346, %f347, %f348}, [params+512];
|
|
neg.f32 %f349, %f346;
|
|
neg.f32 %f350, %f347;
|
|
mul.f32 %f351, %f2043, %f345;
|
|
mul.f32 %f352, %f2044, %f349;
|
|
sub.f32 %f353, %f352, %f351;
|
|
mul.f32 %f354, %f2045, %f347;
|
|
sub.f32 %f13, %f353, %f354;
|
|
ld.const.u64 %rd32, [params+432];
|
|
cvta.to.global.u64 %rd33, %rd32;
|
|
ld.const.u32 %r75, [params+424];
|
|
mul.lo.s32 %r76, %r75, %r64;
|
|
add.s32 %r77, %r76, %r63;
|
|
mul.wide.u32 %rd34, %r77, 12;
|
|
add.s64 %rd35, %rd33, %rd34;
|
|
ld.global.f32 %f355, [%rd35];
|
|
mul.f32 %f356, %f355, 0f3456BF95;
|
|
ld.global.f32 %f357, [%rd35+4];
|
|
mul.f32 %f358, %f357, 0f3456BF95;
|
|
ld.global.f32 %f359, [%rd35+8];
|
|
mul.f32 %f360, %f359, 0f3456BF95;
|
|
abs.f32 %f361, %f2043;
|
|
div.rn.f32 %f362, %f356, %f361;
|
|
abs.f32 %f363, %f2044;
|
|
div.rn.f32 %f364, %f358, %f363;
|
|
abs.f32 %f365, %f2045;
|
|
div.rn.f32 %f366, %f360, %f365;
|
|
abs.f32 %f367, %f362;
|
|
abs.f32 %f368, %f364;
|
|
abs.f32 %f369, %f366;
|
|
mov.f32 %f370, 0f38D1B717;
|
|
max.f32 %f371, %f367, %f370;
|
|
max.f32 %f372, %f368, %f370;
|
|
max.f32 %f373, %f369, %f370;
|
|
fma.rn.f32 %f14, %f2043, %f371, %f355;
|
|
fma.rn.f32 %f15, %f2044, %f372, %f357;
|
|
fma.rn.f32 %f16, %f2045, %f373, %f359;
|
|
setp.gt.f32 %p16, %f361, %f365;
|
|
neg.f32 %f374, %f2044;
|
|
selp.f32 %f375, %f374, 0f00000000, %p16;
|
|
mov.f32 %f2060, 0f00000000;
|
|
neg.f32 %f376, %f2045;
|
|
selp.f32 %f377, %f2043, %f376, %p16;
|
|
selp.f32 %f378, 0f00000000, %f2044, %p16;
|
|
mul.f32 %f379, %f377, %f377;
|
|
fma.rn.f32 %f380, %f375, %f375, %f379;
|
|
fma.rn.f32 %f381, %f378, %f378, %f380;
|
|
sqrt.rn.f32 %f382, %f381;
|
|
rcp.rn.f32 %f383, %f382;
|
|
mul.f32 %f17, %f375, %f383;
|
|
mul.f32 %f18, %f377, %f383;
|
|
mul.f32 %f19, %f378, %f383;
|
|
mul.f32 %f384, %f2045, %f18;
|
|
mul.f32 %f385, %f2044, %f19;
|
|
sub.f32 %f20, %f384, %f385;
|
|
mul.f32 %f386, %f2043, %f19;
|
|
mul.f32 %f387, %f2045, %f17;
|
|
sub.f32 %f21, %f386, %f387;
|
|
mul.f32 %f388, %f2044, %f17;
|
|
mul.f32 %f389, %f2043, %f18;
|
|
sub.f32 %f22, %f388, %f389;
|
|
add.s32 %r78, %r66, -1;
|
|
add.s32 %r79, %r63, 1;
|
|
min.u32 %r80, %r79, %r78;
|
|
add.s32 %r81, %r76, %r80;
|
|
mul.wide.u32 %rd36, %r81, 12;
|
|
add.s64 %rd37, %rd33, %rd36;
|
|
ld.global.f32 %f390, [%rd37];
|
|
setp.eq.f32 %p17, %f390, 0f00000000;
|
|
ld.global.f32 %f391, [%rd37+4];
|
|
setp.eq.f32 %p18, %f391, 0f00000000;
|
|
and.pred %p19, %p17, %p18;
|
|
ld.global.f32 %f392, [%rd37+8];
|
|
setp.eq.f32 %p20, %f392, 0f00000000;
|
|
and.pred %p21, %p19, %p20;
|
|
selp.f32 %f393, %f355, %f390, %p21;
|
|
selp.f32 %f394, %f357, %f391, %p21;
|
|
selp.f32 %f395, %f359, %f392, %p21;
|
|
add.s32 %r82, %r67, -1;
|
|
add.s32 %r83, %r64, 1;
|
|
min.u32 %r84, %r83, %r82;
|
|
mad.lo.s32 %r85, %r75, %r84, %r63;
|
|
mul.wide.u32 %rd38, %r85, 12;
|
|
add.s64 %rd39, %rd33, %rd38;
|
|
ld.global.f32 %f396, [%rd39];
|
|
setp.eq.f32 %p22, %f396, 0f00000000;
|
|
ld.global.f32 %f397, [%rd39+4];
|
|
setp.eq.f32 %p23, %f397, 0f00000000;
|
|
and.pred %p24, %p22, %p23;
|
|
ld.global.f32 %f398, [%rd39+8];
|
|
setp.eq.f32 %p25, %f398, 0f00000000;
|
|
and.pred %p26, %p24, %p25;
|
|
selp.f32 %f399, %f355, %f396, %p26;
|
|
selp.f32 %f400, %f357, %f397, %p26;
|
|
selp.f32 %f401, %f359, %f398, %p26;
|
|
sub.f32 %f402, %f393, %f355;
|
|
sub.f32 %f403, %f394, %f357;
|
|
sub.f32 %f404, %f395, %f359;
|
|
sub.f32 %f405, %f399, %f355;
|
|
sub.f32 %f406, %f400, %f357;
|
|
sub.f32 %f407, %f401, %f359;
|
|
abs.f32 %f408, %f402;
|
|
abs.f32 %f409, %f405;
|
|
max.f32 %f410, %f408, %f409;
|
|
abs.f32 %f411, %f403;
|
|
abs.f32 %f412, %f406;
|
|
max.f32 %f413, %f411, %f412;
|
|
abs.f32 %f414, %f404;
|
|
abs.f32 %f415, %f407;
|
|
max.f32 %f416, %f414, %f415;
|
|
max.f32 %f417, %f410, %f413;
|
|
max.f32 %f418, %f417, %f416;
|
|
setp.eq.f32 %p27, %f402, 0f00000000;
|
|
setp.eq.f32 %p28, %f403, 0f00000000;
|
|
and.pred %p29, %p27, %p28;
|
|
setp.eq.f32 %p30, %f404, 0f00000000;
|
|
and.pred %p31, %p30, %p29;
|
|
setp.eq.f32 %p32, %f405, 0f00000000;
|
|
setp.eq.f32 %p33, %f406, 0f00000000;
|
|
and.pred %p34, %p32, %p33;
|
|
setp.eq.f32 %p35, %f407, 0f00000000;
|
|
and.pred %p36, %p35, %p34;
|
|
or.pred %p37, %p31, %p36;
|
|
selp.f32 %f23, 0f00000000, %f418, %p37;
|
|
abs.f32 %f419, %f345;
|
|
abs.f32 %f420, %f347;
|
|
setp.gt.f32 %p38, %f419, %f420;
|
|
selp.f32 %f421, %f349, 0f00000000, %p38;
|
|
selp.f32 %f422, %f345, %f350, %p38;
|
|
selp.f32 %f423, 0f00000000, %f346, %p38;
|
|
mul.f32 %f424, %f422, %f422;
|
|
fma.rn.f32 %f425, %f421, %f421, %f424;
|
|
fma.rn.f32 %f426, %f423, %f423, %f425;
|
|
sqrt.rn.f32 %f427, %f426;
|
|
rcp.rn.f32 %f428, %f427;
|
|
mul.f32 %f24, %f421, %f428;
|
|
mul.f32 %f25, %f422, %f428;
|
|
mul.f32 %f26, %f423, %f428;
|
|
ld.const.u32 %r7, [params+540];
|
|
setp.lt.s32 %p39, %r7, 1;
|
|
mov.f32 %f2061, %f2060;
|
|
mov.f32 %f2062, %f2060;
|
|
mov.f32 %f2063, %f2060;
|
|
@%p39 bra $L__BB0_52;
|
|
|
|
ld.const.u64 %rd40, [params+128];
|
|
ld.const.u32 %r87, [params+120];
|
|
cvt.rn.f32.s32 %f433, %r7;
|
|
rcp.rn.f32 %f27, %f433;
|
|
mad.lo.s32 %r88, %r87, %r64, %r63;
|
|
cvta.to.global.u64 %rd41, %rd40;
|
|
mul.wide.u32 %rd42, %r88, 4;
|
|
add.s64 %rd43, %rd41, %rd42;
|
|
ld.global.u32 %r1542, [%rd43];
|
|
ld.const.f32 %f28, [params+536];
|
|
ld.const.u8 %rs32, [params+104];
|
|
and.b16 %rs33, %rs32, 32;
|
|
setp.eq.s16 %p40, %rs33, 0;
|
|
mov.u32 %r86, 0;
|
|
selp.f32 %f29, 0f3F800000, 0f41200000, %p40;
|
|
mul.f32 %f30, %f14, 0f3456BF95;
|
|
mul.f32 %f31, %f15, 0f3456BF95;
|
|
mul.f32 %f32, %f16, 0f3456BF95;
|
|
ld.const.u64 %rd3, [params+96];
|
|
mul.f32 %f434, %f345, %f25;
|
|
mul.f32 %f435, %f346, %f24;
|
|
sub.f32 %f33, %f435, %f434;
|
|
mul.f32 %f436, %f347, %f24;
|
|
mul.f32 %f437, %f345, %f26;
|
|
sub.f32 %f34, %f437, %f436;
|
|
mul.f32 %f438, %f346, %f26;
|
|
mul.f32 %f439, %f347, %f25;
|
|
sub.f32 %f35, %f439, %f438;
|
|
abs.f32 %f522, %f31;
|
|
abs.f32 %f523, %f30;
|
|
max.f32 %f524, %f523, %f522;
|
|
abs.f32 %f525, %f32;
|
|
max.f32 %f526, %f524, %f525;
|
|
mov.u32 %r1539, %r86;
|
|
|
|
$L__BB0_7:
|
|
mov.u32 %r1541, %r86;
|
|
|
|
$L__BB0_8:
|
|
cvt.rn.f32.s32 %f2042, %r1539;
|
|
mad.lo.s32 %r90, %r1542, 1664525, 1013904223;
|
|
and.b32 %r91, %r90, 16777215;
|
|
cvt.rn.f32.u32 %f440, %r91;
|
|
fma.rn.f32 %f441, %f440, 0f33800000, %f2042;
|
|
mul.f32 %f442, %f27, %f441;
|
|
mad.lo.s32 %r13, %r90, 1664525, 1013904223;
|
|
and.b32 %r92, %r13, 16777215;
|
|
cvt.rn.f32.u32 %f443, %r92;
|
|
cvt.rn.f32.s32 %f444, %r1541;
|
|
fma.rn.f32 %f445, %f443, 0f33800000, %f444;
|
|
mul.f32 %f446, %f27, %f445;
|
|
sqrt.rn.f32 %f45, %f442;
|
|
mul.f32 %f46, %f446, 0f40C90FDB;
|
|
mul.f32 %f447, %f46, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r1549, %f447;
|
|
cvt.rn.f32.s32 %f448, %r1549;
|
|
mov.f32 %f449, 0fBFC90FDA;
|
|
fma.rn.f32 %f450, %f448, %f449, %f46;
|
|
mov.f32 %f451, 0fB3A22168;
|
|
fma.rn.f32 %f452, %f448, %f451, %f450;
|
|
mov.f32 %f453, 0fA7C234C5;
|
|
fma.rn.f32 %f2057, %f448, %f453, %f452;
|
|
abs.f32 %f48, %f46;
|
|
setp.ltu.f32 %p41, %f48, 0f47CE4780;
|
|
add.s64 %rd4, %rd1, 24;
|
|
mov.u32 %r1546, %r1549;
|
|
mov.f32 %f2054, %f2057;
|
|
@%p41 bra $L__BB0_16;
|
|
|
|
setp.eq.f32 %p42, %f48, 0f7F800000;
|
|
@%p42 bra $L__BB0_15;
|
|
bra.uni $L__BB0_10;
|
|
|
|
$L__BB0_15:
|
|
mov.f32 %f456, 0f00000000;
|
|
mul.rn.f32 %f2054, %f46, %f456;
|
|
mov.u32 %r1546, 0;
|
|
bra.uni $L__BB0_16;
|
|
|
|
$L__BB0_10:
|
|
mov.b32 %r15, %f46;
|
|
bfe.u32 %r94, %r15, 23, 8;
|
|
add.s32 %r16, %r94, -128;
|
|
shl.b32 %r95, %r15, 8;
|
|
or.b32 %r17, %r95, -2147483648;
|
|
shr.u32 %r18, %r16, 5;
|
|
mov.u64 %rd128, 0;
|
|
mov.u32 %r1543, 0;
|
|
mov.u64 %rd127, __cudart_i2opi_f;
|
|
mov.u64 %rd126, %rd1;
|
|
|
|
$L__BB0_11:
|
|
.pragma "nounroll";
|
|
ld.global.nc.u32 %r96, [%rd127];
|
|
mad.wide.u32 %rd46, %r96, %r17, %rd128;
|
|
shr.u64 %rd128, %rd46, 32;
|
|
st.local.u32 [%rd126], %rd46;
|
|
add.s64 %rd127, %rd127, 4;
|
|
add.s64 %rd126, %rd126, 4;
|
|
add.s32 %r1543, %r1543, 1;
|
|
setp.ne.s32 %p43, %r1543, 6;
|
|
@%p43 bra $L__BB0_11;
|
|
|
|
st.local.u32 [%rd4], %rd128;
|
|
mov.u32 %r97, 4;
|
|
sub.s32 %r21, %r97, %r18;
|
|
mov.u32 %r98, 6;
|
|
sub.s32 %r99, %r98, %r18;
|
|
mul.wide.s32 %rd47, %r99, 4;
|
|
add.s64 %rd48, %rd1, %rd47;
|
|
ld.local.u32 %r1544, [%rd48];
|
|
ld.local.u32 %r1545, [%rd48+-4];
|
|
and.b32 %r24, %r16, 31;
|
|
setp.eq.s32 %p44, %r24, 0;
|
|
@%p44 bra $L__BB0_14;
|
|
|
|
mov.u32 %r100, 32;
|
|
sub.s32 %r101, %r100, %r24;
|
|
shr.u32 %r102, %r1545, %r101;
|
|
shl.b32 %r103, %r1544, %r24;
|
|
add.s32 %r1544, %r102, %r103;
|
|
mul.wide.s32 %rd49, %r21, 4;
|
|
add.s64 %rd50, %rd1, %rd49;
|
|
ld.local.u32 %r104, [%rd50];
|
|
shr.u32 %r105, %r104, %r101;
|
|
shl.b32 %r106, %r1545, %r24;
|
|
add.s32 %r1545, %r105, %r106;
|
|
|
|
$L__BB0_14:
|
|
and.b32 %r107, %r15, -2147483648;
|
|
shr.u32 %r108, %r1545, 30;
|
|
shl.b32 %r109, %r1544, 2;
|
|
or.b32 %r110, %r108, %r109;
|
|
shr.u32 %r111, %r110, 31;
|
|
shr.u32 %r112, %r1544, 30;
|
|
add.s32 %r113, %r111, %r112;
|
|
neg.s32 %r114, %r113;
|
|
setp.eq.s32 %p45, %r107, 0;
|
|
selp.b32 %r1546, %r113, %r114, %p45;
|
|
setp.ne.s32 %p46, %r111, 0;
|
|
xor.b32 %r115, %r107, -2147483648;
|
|
selp.b32 %r116, %r115, %r107, %p46;
|
|
selp.b32 %r117, -1, 0, %p46;
|
|
xor.b32 %r118, %r110, %r117;
|
|
shl.b32 %r119, %r1545, 2;
|
|
xor.b32 %r120, %r119, %r117;
|
|
cvt.u64.u32 %rd51, %r118;
|
|
cvt.u64.u32 %rd52, %r120;
|
|
bfi.b64 %rd53, %rd51, %rd52, 32, 32;
|
|
cvt.rn.f64.s64 %fd1, %rd53;
|
|
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f454, %fd2;
|
|
setp.eq.s32 %p47, %r116, 0;
|
|
neg.f32 %f455, %f454;
|
|
selp.f32 %f2054, %f454, %f455, %p47;
|
|
|
|
$L__BB0_16:
|
|
add.s32 %r31, %r1546, 1;
|
|
and.b32 %r32, %r31, 1;
|
|
setp.eq.s32 %p48, %r32, 0;
|
|
selp.f32 %f52, %f2054, 0f3F800000, %p48;
|
|
mul.rn.f32 %f53, %f2054, %f2054;
|
|
mov.f32 %f2055, 0fB94D4153;
|
|
@%p48 bra $L__BB0_18;
|
|
|
|
mov.f32 %f458, 0fBAB607ED;
|
|
mov.f32 %f459, 0f37CBAC00;
|
|
fma.rn.f32 %f2055, %f459, %f53, %f458;
|
|
|
|
$L__BB0_18:
|
|
selp.f32 %f460, 0f3C0885E4, 0f3D2AAABB, %p48;
|
|
fma.rn.f32 %f461, %f2055, %f53, %f460;
|
|
selp.f32 %f462, 0fBE2AAAA8, 0fBEFFFFFF, %p48;
|
|
fma.rn.f32 %f463, %f461, %f53, %f462;
|
|
mov.f32 %f464, 0f00000000;
|
|
fma.rn.f32 %f465, %f53, %f52, %f464;
|
|
fma.rn.f32 %f2056, %f463, %f465, %f52;
|
|
and.b32 %r122, %r31, 2;
|
|
setp.eq.s32 %p50, %r122, 0;
|
|
@%p50 bra $L__BB0_20;
|
|
|
|
mov.f32 %f467, 0fBF800000;
|
|
fma.rn.f32 %f2056, %f2056, %f467, %f464;
|
|
|
|
$L__BB0_20:
|
|
@%p41 bra $L__BB0_28;
|
|
|
|
setp.eq.f32 %p52, %f48, 0f7F800000;
|
|
@%p52 bra $L__BB0_27;
|
|
bra.uni $L__BB0_22;
|
|
|
|
$L__BB0_27:
|
|
mov.f32 %f470, 0f00000000;
|
|
mul.rn.f32 %f2057, %f46, %f470;
|
|
mov.u32 %r1549, 0;
|
|
bra.uni $L__BB0_28;
|
|
|
|
$L__BB0_22:
|
|
mov.b32 %r33, %f46;
|
|
bfe.u32 %r123, %r33, 23, 8;
|
|
add.s32 %r34, %r123, -128;
|
|
shl.b32 %r124, %r33, 8;
|
|
or.b32 %r35, %r124, -2147483648;
|
|
shr.u32 %r36, %r34, 5;
|
|
mov.u64 %rd129, 0;
|
|
mov.u64 %rd130, %rd129;
|
|
|
|
$L__BB0_23:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd56, %rd129, 2;
|
|
mov.u64 %rd57, __cudart_i2opi_f;
|
|
add.s64 %rd58, %rd57, %rd56;
|
|
ld.global.nc.u32 %r125, [%rd58];
|
|
mad.wide.u32 %rd59, %r125, %r35, %rd130;
|
|
shr.u64 %rd130, %rd59, 32;
|
|
add.s64 %rd60, %rd1, %rd56;
|
|
st.local.u32 [%rd60], %rd59;
|
|
cvt.u32.u64 %r126, %rd129;
|
|
add.s32 %r127, %r126, 1;
|
|
cvt.s64.s32 %rd129, %r127;
|
|
setp.ne.s32 %p53, %r127, 6;
|
|
@%p53 bra $L__BB0_23;
|
|
|
|
st.local.u32 [%rd4], %rd130;
|
|
mov.u32 %r128, 4;
|
|
sub.s32 %r37, %r128, %r36;
|
|
mov.u32 %r129, 6;
|
|
sub.s32 %r130, %r129, %r36;
|
|
mul.wide.s32 %rd61, %r130, 4;
|
|
add.s64 %rd62, %rd1, %rd61;
|
|
ld.local.u32 %r1547, [%rd62];
|
|
ld.local.u32 %r1548, [%rd62+-4];
|
|
and.b32 %r40, %r34, 31;
|
|
setp.eq.s32 %p54, %r40, 0;
|
|
@%p54 bra $L__BB0_26;
|
|
|
|
mov.u32 %r131, 32;
|
|
sub.s32 %r132, %r131, %r40;
|
|
shr.u32 %r133, %r1548, %r132;
|
|
shl.b32 %r134, %r1547, %r40;
|
|
add.s32 %r1547, %r133, %r134;
|
|
mul.wide.s32 %rd63, %r37, 4;
|
|
add.s64 %rd64, %rd1, %rd63;
|
|
ld.local.u32 %r135, [%rd64];
|
|
shr.u32 %r136, %r135, %r132;
|
|
shl.b32 %r137, %r1548, %r40;
|
|
add.s32 %r1548, %r136, %r137;
|
|
|
|
$L__BB0_26:
|
|
and.b32 %r138, %r33, -2147483648;
|
|
shr.u32 %r139, %r1548, 30;
|
|
shl.b32 %r140, %r1547, 2;
|
|
or.b32 %r141, %r139, %r140;
|
|
shr.u32 %r142, %r141, 31;
|
|
shr.u32 %r143, %r1547, 30;
|
|
add.s32 %r144, %r142, %r143;
|
|
neg.s32 %r145, %r144;
|
|
setp.eq.s32 %p55, %r138, 0;
|
|
selp.b32 %r1549, %r144, %r145, %p55;
|
|
setp.ne.s32 %p56, %r142, 0;
|
|
xor.b32 %r146, %r138, -2147483648;
|
|
selp.b32 %r147, %r146, %r138, %p56;
|
|
selp.b32 %r148, -1, 0, %p56;
|
|
xor.b32 %r149, %r141, %r148;
|
|
shl.b32 %r150, %r1548, 2;
|
|
xor.b32 %r151, %r150, %r148;
|
|
cvt.u64.u32 %rd65, %r149;
|
|
cvt.u64.u32 %rd66, %r151;
|
|
bfi.b64 %rd67, %rd65, %rd66, 32, 32;
|
|
cvt.rn.f64.s64 %fd3, %rd67;
|
|
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f468, %fd4;
|
|
setp.eq.s32 %p57, %r147, 0;
|
|
neg.f32 %f469, %f468;
|
|
selp.f32 %f2057, %f468, %f469, %p57;
|
|
|
|
$L__BB0_28:
|
|
mul.f32 %f62, %f45, %f2056;
|
|
and.b32 %r47, %r1549, 1;
|
|
setp.eq.s32 %p58, %r47, 0;
|
|
selp.f32 %f63, %f2057, 0f3F800000, %p58;
|
|
mul.rn.f32 %f64, %f2057, %f2057;
|
|
mov.f32 %f2058, 0fB94D4153;
|
|
@%p58 bra $L__BB0_30;
|
|
|
|
mov.f32 %f472, 0fBAB607ED;
|
|
mov.f32 %f473, 0f37CBAC00;
|
|
fma.rn.f32 %f2058, %f473, %f64, %f472;
|
|
|
|
$L__BB0_30:
|
|
selp.f32 %f474, 0f3C0885E4, 0f3D2AAABB, %p58;
|
|
fma.rn.f32 %f475, %f2058, %f64, %f474;
|
|
selp.f32 %f476, 0fBE2AAAA8, 0fBEFFFFFF, %p58;
|
|
fma.rn.f32 %f477, %f475, %f64, %f476;
|
|
mov.f32 %f478, 0f00000000;
|
|
fma.rn.f32 %f479, %f64, %f63, %f478;
|
|
fma.rn.f32 %f2059, %f477, %f479, %f63;
|
|
and.b32 %r153, %r1549, 2;
|
|
setp.eq.s32 %p60, %r153, 0;
|
|
@%p60 bra $L__BB0_32;
|
|
|
|
mov.f32 %f481, 0fBF800000;
|
|
fma.rn.f32 %f2059, %f2059, %f481, %f478;
|
|
|
|
$L__BB0_32:
|
|
mul.f32 %f491, %f62, %f62;
|
|
mov.f32 %f492, 0f3F800000;
|
|
sub.f32 %f493, %f492, %f491;
|
|
mul.f32 %f494, %f45, %f2059;
|
|
mul.f32 %f495, %f494, %f494;
|
|
sub.f32 %f496, %f493, %f495;
|
|
max.f32 %f497, %f478, %f496;
|
|
sqrt.rn.f32 %f498, %f497;
|
|
mul.f32 %f499, %f24, %f494;
|
|
mul.f32 %f500, %f25, %f494;
|
|
mul.f32 %f501, %f26, %f494;
|
|
fma.rn.f32 %f502, %f35, %f62, %f499;
|
|
fma.rn.f32 %f503, %f34, %f62, %f500;
|
|
fma.rn.f32 %f504, %f33, %f62, %f501;
|
|
fma.rn.f32 %f505, %f345, %f498, %f502;
|
|
fma.rn.f32 %f506, %f346, %f498, %f503;
|
|
fma.rn.f32 %f507, %f347, %f498, %f504;
|
|
add.f32 %f508, %f345, %f505;
|
|
add.f32 %f509, %f346, %f506;
|
|
add.f32 %f510, %f347, %f507;
|
|
mul.f32 %f511, %f28, %f508;
|
|
mul.f32 %f512, %f28, %f509;
|
|
mul.f32 %f513, %f28, %f510;
|
|
sub.f32 %f514, %f511, %f345;
|
|
sub.f32 %f515, %f512, %f346;
|
|
sub.f32 %f516, %f513, %f347;
|
|
mul.f32 %f517, %f515, %f515;
|
|
fma.rn.f32 %f518, %f514, %f514, %f517;
|
|
fma.rn.f32 %f519, %f516, %f516, %f518;
|
|
sqrt.rn.f32 %f520, %f519;
|
|
rcp.rn.f32 %f521, %f520;
|
|
mul.f32 %f70, %f521, %f514;
|
|
mul.f32 %f71, %f521, %f515;
|
|
mul.f32 %f72, %f521, %f516;
|
|
mov.f32 %f527, 0f38D1B717;
|
|
max.f32 %f528, %f526, %f527;
|
|
mul.f32 %f488, %f29, %f528;
|
|
mad.lo.s32 %r225, %r13, 1664525, 1013904223;
|
|
and.b32 %r226, %r225, 16777215;
|
|
cvt.rn.f32.u32 %f529, %r226;
|
|
mul.f32 %f530, %f529, 0f33800000;
|
|
fma.rn.f32 %f531, %f530, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r48, %r225, 1664525, 1013904223;
|
|
and.b32 %r227, %r48, 16777215;
|
|
cvt.rn.f32.u32 %f532, %r227;
|
|
mul.f32 %f533, %f532, 0f33800000;
|
|
fma.rn.f32 %f534, %f533, 0f40000000, 0fBF800000;
|
|
mul.f32 %f535, %f17, %f534;
|
|
mul.f32 %f536, %f18, %f534;
|
|
mul.f32 %f537, %f19, %f534;
|
|
fma.rn.f32 %f538, %f20, %f531, %f535;
|
|
fma.rn.f32 %f539, %f21, %f531, %f536;
|
|
fma.rn.f32 %f540, %f22, %f531, %f537;
|
|
mul.f32 %f541, %f539, %f539;
|
|
fma.rn.f32 %f542, %f538, %f538, %f541;
|
|
fma.rn.f32 %f543, %f540, %f540, %f542;
|
|
sqrt.rn.f32 %f544, %f543;
|
|
rcp.rn.f32 %f545, %f544;
|
|
mul.f32 %f546, %f538, %f545;
|
|
mul.f32 %f547, %f539, %f545;
|
|
mul.f32 %f548, %f540, %f545;
|
|
mul.f32 %f549, %f2044, %f548;
|
|
mul.f32 %f550, %f2045, %f547;
|
|
sub.f32 %f551, %f549, %f550;
|
|
mul.f32 %f552, %f2045, %f546;
|
|
mul.f32 %f553, %f2043, %f548;
|
|
sub.f32 %f554, %f552, %f553;
|
|
mul.f32 %f555, %f2043, %f547;
|
|
mul.f32 %f556, %f2044, %f546;
|
|
sub.f32 %f557, %f555, %f556;
|
|
mul.f32 %f558, %f546, 0f3E24A490;
|
|
mul.f32 %f559, %f547, 0f3E24A490;
|
|
mul.f32 %f560, %f548, 0f3E24A490;
|
|
fma.rn.f32 %f561, %f23, %f558, %f14;
|
|
fma.rn.f32 %f562, %f23, %f559, %f15;
|
|
fma.rn.f32 %f563, %f23, %f560, %f16;
|
|
mul.f32 %f564, %f551, 0fBECECEDE;
|
|
mul.f32 %f565, %f554, 0fBECECEDE;
|
|
mul.f32 %f566, %f557, 0fBECECEDE;
|
|
fma.rn.f32 %f74, %f23, %f564, %f561;
|
|
fma.rn.f32 %f75, %f23, %f565, %f562;
|
|
fma.rn.f32 %f76, %f23, %f566, %f563;
|
|
sub.f32 %f567, %f74, %f14;
|
|
sub.f32 %f568, %f75, %f15;
|
|
sub.f32 %f569, %f76, %f16;
|
|
mul.f32 %f570, %f568, %f568;
|
|
fma.rn.f32 %f571, %f567, %f567, %f570;
|
|
fma.rn.f32 %f572, %f569, %f569, %f571;
|
|
sqrt.rn.f32 %f489, %f572;
|
|
rcp.rn.f32 %f573, %f489;
|
|
mul.f32 %f485, %f573, %f567;
|
|
mul.f32 %f486, %f573, %f568;
|
|
mul.f32 %f487, %f573, %f569;
|
|
mov.u32 %r190, 2;
|
|
mov.u32 %r191, 1;
|
|
mov.u32 %r192, 3;
|
|
mov.u32 %r195, 1065353216;
|
|
mov.u32 %r224, 0;
|
|
// begin inline asm
|
|
call(%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167,%r168,%r169,%r170,%r171,%r172,%r173,%r174,%r175,%r176,%r177,%r178,%r179,%r180,%r181,%r182,%r183,%r184,%r185),_optix_trace_typed_32,(%r224,%rd3,%f14,%f15,%f16,%f485,%f486,%f487,%f488,%f489,%f478,%r191,%r224,%r191,%r190,%r191,%r192,%r195,%r195,%r195,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224);
|
|
// end inline asm
|
|
mov.b32 %f574, %r154;
|
|
mov.b32 %f575, %r155;
|
|
mov.b32 %f576, %r156;
|
|
setp.neu.f32 %p61, %f574, 0f3F800000;
|
|
setp.neu.f32 %p62, %f575, 0f3F800000;
|
|
or.pred %p63, %p61, %p62;
|
|
setp.neu.f32 %p64, %f576, 0f3F800000;
|
|
mov.u16 %rs195, 0;
|
|
or.pred %p65, %p64, %p63;
|
|
@%p65 bra $L__BB0_34;
|
|
|
|
mov.f32 %f584, 0f6C4ECB8F;
|
|
mov.f32 %f585, 0f00000000;
|
|
// begin inline asm
|
|
call(%r228,%r229,%r230,%r231,%r232,%r233,%r234,%r235,%r236,%r237,%r238,%r239,%r240,%r241,%r242,%r243,%r244,%r245,%r246,%r247,%r248,%r249,%r250,%r251,%r252,%r253,%r254,%r255,%r256,%r257,%r258,%r259),_optix_trace_typed_32,(%r224,%rd3,%f74,%f75,%f76,%f70,%f71,%f72,%f488,%f584,%f585,%r191,%r224,%r191,%r190,%r191,%r192,%r195,%r195,%r195,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224);
|
|
// end inline asm
|
|
mov.b32 %f586, %r228;
|
|
mov.b32 %f587, %r229;
|
|
mov.b32 %f588, %r230;
|
|
add.f32 %f2060, %f2060, %f586;
|
|
add.f32 %f2061, %f2061, %f587;
|
|
add.f32 %f2062, %f2062, %f588;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_34:
|
|
mad.lo.s32 %r370, %r48, 1664525, 1013904223;
|
|
and.b32 %r371, %r370, 16777215;
|
|
cvt.rn.f32.u32 %f598, %r371;
|
|
mul.f32 %f599, %f598, 0f33800000;
|
|
fma.rn.f32 %f600, %f599, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r49, %r370, 1664525, 1013904223;
|
|
and.b32 %r372, %r49, 16777215;
|
|
cvt.rn.f32.u32 %f601, %r372;
|
|
mul.f32 %f602, %f601, 0f33800000;
|
|
fma.rn.f32 %f603, %f602, 0f40000000, 0fBF800000;
|
|
mul.f32 %f604, %f17, %f603;
|
|
mul.f32 %f605, %f18, %f603;
|
|
mul.f32 %f606, %f19, %f603;
|
|
fma.rn.f32 %f607, %f20, %f600, %f604;
|
|
fma.rn.f32 %f608, %f21, %f600, %f605;
|
|
fma.rn.f32 %f609, %f22, %f600, %f606;
|
|
mul.f32 %f610, %f608, %f608;
|
|
fma.rn.f32 %f611, %f607, %f607, %f610;
|
|
fma.rn.f32 %f612, %f609, %f609, %f611;
|
|
sqrt.rn.f32 %f613, %f612;
|
|
rcp.rn.f32 %f614, %f613;
|
|
mul.f32 %f615, %f607, %f614;
|
|
mul.f32 %f616, %f608, %f614;
|
|
mul.f32 %f617, %f609, %f614;
|
|
mul.f32 %f618, %f2044, %f617;
|
|
mul.f32 %f619, %f2045, %f616;
|
|
sub.f32 %f620, %f618, %f619;
|
|
mul.f32 %f621, %f2045, %f615;
|
|
mul.f32 %f622, %f2043, %f617;
|
|
sub.f32 %f623, %f621, %f622;
|
|
mul.f32 %f624, %f2043, %f616;
|
|
mul.f32 %f625, %f2044, %f615;
|
|
sub.f32 %f626, %f624, %f625;
|
|
mul.f32 %f627, %f615, 0fBE24A490;
|
|
mul.f32 %f628, %f616, 0fBE24A490;
|
|
mul.f32 %f629, %f617, 0fBE24A490;
|
|
fma.rn.f32 %f630, %f23, %f627, %f14;
|
|
fma.rn.f32 %f631, %f23, %f628, %f15;
|
|
fma.rn.f32 %f632, %f23, %f629, %f16;
|
|
mul.f32 %f633, %f620, 0f3ECACABC;
|
|
mul.f32 %f634, %f623, 0f3ECACABC;
|
|
mul.f32 %f635, %f626, 0f3ECACABC;
|
|
fma.rn.f32 %f85, %f23, %f633, %f630;
|
|
fma.rn.f32 %f86, %f23, %f634, %f631;
|
|
fma.rn.f32 %f87, %f23, %f635, %f632;
|
|
sub.f32 %f636, %f85, %f14;
|
|
sub.f32 %f637, %f86, %f15;
|
|
sub.f32 %f638, %f87, %f16;
|
|
mul.f32 %f639, %f637, %f637;
|
|
fma.rn.f32 %f640, %f636, %f636, %f639;
|
|
fma.rn.f32 %f641, %f638, %f638, %f640;
|
|
sqrt.rn.f32 %f596, %f641;
|
|
rcp.rn.f32 %f642, %f596;
|
|
mul.f32 %f592, %f642, %f636;
|
|
mul.f32 %f593, %f642, %f637;
|
|
mul.f32 %f594, %f642, %f638;
|
|
mov.f32 %f597, 0f00000000;
|
|
// begin inline asm
|
|
call(%r299,%r300,%r301,%r302,%r303,%r304,%r305,%r306,%r307,%r308,%r309,%r310,%r311,%r312,%r313,%r314,%r315,%r316,%r317,%r318,%r319,%r320,%r321,%r322,%r323,%r324,%r325,%r326,%r327,%r328,%r329,%r330),_optix_trace_typed_32,(%r224,%rd3,%f14,%f15,%f16,%f592,%f593,%f594,%f488,%f596,%f597,%r191,%r224,%r191,%r190,%r191,%r192,%r195,%r195,%r195,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224);
|
|
// end inline asm
|
|
mov.b32 %f643, %r299;
|
|
mov.b32 %f644, %r300;
|
|
mov.b32 %f645, %r301;
|
|
setp.neu.f32 %p66, %f643, 0f3F800000;
|
|
setp.neu.f32 %p67, %f644, 0f3F800000;
|
|
or.pred %p68, %p66, %p67;
|
|
setp.neu.f32 %p69, %f645, 0f3F800000;
|
|
or.pred %p70, %p69, %p68;
|
|
@%p70 bra $L__BB0_36;
|
|
|
|
mov.f32 %f653, 0f6C4ECB8F;
|
|
mov.u32 %r409, 2;
|
|
mov.u32 %r410, 1;
|
|
mov.u32 %r411, 3;
|
|
mov.u32 %r414, 1065353216;
|
|
mov.u32 %r443, 0;
|
|
// begin inline asm
|
|
call(%r373,%r374,%r375,%r376,%r377,%r378,%r379,%r380,%r381,%r382,%r383,%r384,%r385,%r386,%r387,%r388,%r389,%r390,%r391,%r392,%r393,%r394,%r395,%r396,%r397,%r398,%r399,%r400,%r401,%r402,%r403,%r404),_optix_trace_typed_32,(%r443,%rd3,%f85,%f86,%f87,%f70,%f71,%f72,%f488,%f653,%f597,%r410,%r443,%r410,%r409,%r410,%r411,%r414,%r414,%r414,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443);
|
|
// end inline asm
|
|
mov.b32 %f655, %r373;
|
|
mov.b32 %f656, %r374;
|
|
mov.b32 %f657, %r375;
|
|
add.f32 %f2060, %f2060, %f655;
|
|
add.f32 %f2061, %f2061, %f656;
|
|
add.f32 %f2062, %f2062, %f657;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_36:
|
|
mad.lo.s32 %r515, %r49, 1664525, 1013904223;
|
|
and.b32 %r516, %r515, 16777215;
|
|
cvt.rn.f32.u32 %f667, %r516;
|
|
mul.f32 %f668, %f667, 0f33800000;
|
|
fma.rn.f32 %f669, %f668, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r50, %r515, 1664525, 1013904223;
|
|
and.b32 %r517, %r50, 16777215;
|
|
cvt.rn.f32.u32 %f670, %r517;
|
|
mul.f32 %f671, %f670, 0f33800000;
|
|
fma.rn.f32 %f672, %f671, 0f40000000, 0fBF800000;
|
|
mul.f32 %f673, %f17, %f672;
|
|
mul.f32 %f674, %f18, %f672;
|
|
mul.f32 %f675, %f19, %f672;
|
|
fma.rn.f32 %f676, %f20, %f669, %f673;
|
|
fma.rn.f32 %f677, %f21, %f669, %f674;
|
|
fma.rn.f32 %f678, %f22, %f669, %f675;
|
|
mul.f32 %f679, %f677, %f677;
|
|
fma.rn.f32 %f680, %f676, %f676, %f679;
|
|
fma.rn.f32 %f681, %f678, %f678, %f680;
|
|
sqrt.rn.f32 %f682, %f681;
|
|
rcp.rn.f32 %f683, %f682;
|
|
mul.f32 %f684, %f676, %f683;
|
|
mul.f32 %f685, %f677, %f683;
|
|
mul.f32 %f686, %f678, %f683;
|
|
mul.f32 %f687, %f2044, %f686;
|
|
mul.f32 %f688, %f2045, %f685;
|
|
sub.f32 %f689, %f687, %f688;
|
|
mul.f32 %f690, %f2045, %f684;
|
|
mul.f32 %f691, %f2043, %f686;
|
|
sub.f32 %f692, %f690, %f691;
|
|
mul.f32 %f693, %f2043, %f685;
|
|
mul.f32 %f694, %f2044, %f684;
|
|
sub.f32 %f695, %f693, %f694;
|
|
mul.f32 %f696, %f684, 0f3F23A3A8;
|
|
mul.f32 %f697, %f685, 0f3F23A3A8;
|
|
mul.f32 %f698, %f686, 0f3F23A3A8;
|
|
fma.rn.f32 %f699, %f23, %f696, %f14;
|
|
fma.rn.f32 %f700, %f23, %f697, %f15;
|
|
fma.rn.f32 %f701, %f23, %f698, %f16;
|
|
mul.f32 %f702, %f689, 0f3E24A490;
|
|
mul.f32 %f703, %f692, 0f3E24A490;
|
|
mul.f32 %f704, %f695, 0f3E24A490;
|
|
fma.rn.f32 %f96, %f23, %f702, %f699;
|
|
fma.rn.f32 %f97, %f23, %f703, %f700;
|
|
fma.rn.f32 %f98, %f23, %f704, %f701;
|
|
sub.f32 %f705, %f96, %f14;
|
|
sub.f32 %f706, %f97, %f15;
|
|
sub.f32 %f707, %f98, %f16;
|
|
mul.f32 %f708, %f706, %f706;
|
|
fma.rn.f32 %f709, %f705, %f705, %f708;
|
|
fma.rn.f32 %f710, %f707, %f707, %f709;
|
|
sqrt.rn.f32 %f665, %f710;
|
|
rcp.rn.f32 %f711, %f665;
|
|
mul.f32 %f661, %f711, %f705;
|
|
mul.f32 %f662, %f711, %f706;
|
|
mul.f32 %f663, %f711, %f707;
|
|
mov.u32 %r480, 2;
|
|
mov.u32 %r481, 1;
|
|
mov.u32 %r482, 3;
|
|
mov.u32 %r485, 1065353216;
|
|
mov.u32 %r514, 0;
|
|
// begin inline asm
|
|
call(%r444,%r445,%r446,%r447,%r448,%r449,%r450,%r451,%r452,%r453,%r454,%r455,%r456,%r457,%r458,%r459,%r460,%r461,%r462,%r463,%r464,%r465,%r466,%r467,%r468,%r469,%r470,%r471,%r472,%r473,%r474,%r475),_optix_trace_typed_32,(%r514,%rd3,%f14,%f15,%f16,%f661,%f662,%f663,%f488,%f665,%f597,%r481,%r514,%r481,%r480,%r481,%r482,%r485,%r485,%r485,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514);
|
|
// end inline asm
|
|
mov.b32 %f712, %r444;
|
|
mov.b32 %f713, %r445;
|
|
mov.b32 %f714, %r446;
|
|
setp.neu.f32 %p71, %f712, 0f3F800000;
|
|
setp.neu.f32 %p72, %f713, 0f3F800000;
|
|
or.pred %p73, %p71, %p72;
|
|
setp.neu.f32 %p74, %f714, 0f3F800000;
|
|
or.pred %p75, %p74, %p73;
|
|
@%p75 bra $L__BB0_38;
|
|
|
|
mov.f32 %f722, 0f6C4ECB8F;
|
|
mov.f32 %f723, 0f00000000;
|
|
// begin inline asm
|
|
call(%r518,%r519,%r520,%r521,%r522,%r523,%r524,%r525,%r526,%r527,%r528,%r529,%r530,%r531,%r532,%r533,%r534,%r535,%r536,%r537,%r538,%r539,%r540,%r541,%r542,%r543,%r544,%r545,%r546,%r547,%r548,%r549),_optix_trace_typed_32,(%r514,%rd3,%f96,%f97,%f98,%f70,%f71,%f72,%f488,%f722,%f723,%r481,%r514,%r481,%r480,%r481,%r482,%r485,%r485,%r485,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514);
|
|
// end inline asm
|
|
mov.b32 %f724, %r518;
|
|
mov.b32 %f725, %r519;
|
|
mov.b32 %f726, %r520;
|
|
add.f32 %f2060, %f2060, %f724;
|
|
add.f32 %f2061, %f2061, %f725;
|
|
add.f32 %f2062, %f2062, %f726;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_38:
|
|
mad.lo.s32 %r660, %r50, 1664525, 1013904223;
|
|
and.b32 %r661, %r660, 16777215;
|
|
cvt.rn.f32.u32 %f736, %r661;
|
|
mul.f32 %f737, %f736, 0f33800000;
|
|
fma.rn.f32 %f738, %f737, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r51, %r660, 1664525, 1013904223;
|
|
and.b32 %r662, %r51, 16777215;
|
|
cvt.rn.f32.u32 %f739, %r662;
|
|
mul.f32 %f740, %f739, 0f33800000;
|
|
fma.rn.f32 %f741, %f740, 0f40000000, 0fBF800000;
|
|
mul.f32 %f742, %f17, %f741;
|
|
mul.f32 %f743, %f18, %f741;
|
|
mul.f32 %f744, %f19, %f741;
|
|
fma.rn.f32 %f745, %f20, %f738, %f742;
|
|
fma.rn.f32 %f746, %f21, %f738, %f743;
|
|
fma.rn.f32 %f747, %f22, %f738, %f744;
|
|
mul.f32 %f748, %f746, %f746;
|
|
fma.rn.f32 %f749, %f745, %f745, %f748;
|
|
fma.rn.f32 %f750, %f747, %f747, %f749;
|
|
sqrt.rn.f32 %f751, %f750;
|
|
rcp.rn.f32 %f752, %f751;
|
|
mul.f32 %f753, %f745, %f752;
|
|
mul.f32 %f754, %f746, %f752;
|
|
mul.f32 %f755, %f747, %f752;
|
|
mul.f32 %f756, %f2044, %f755;
|
|
mul.f32 %f757, %f2045, %f754;
|
|
sub.f32 %f758, %f756, %f757;
|
|
mul.f32 %f759, %f2045, %f753;
|
|
mul.f32 %f760, %f2043, %f755;
|
|
sub.f32 %f761, %f759, %f760;
|
|
mul.f32 %f762, %f2043, %f754;
|
|
mul.f32 %f763, %f2044, %f753;
|
|
sub.f32 %f764, %f762, %f763;
|
|
mul.f32 %f765, %f753, 0fBECECEDE;
|
|
mul.f32 %f766, %f754, 0fBECECEDE;
|
|
mul.f32 %f767, %f755, 0fBECECEDE;
|
|
fma.rn.f32 %f768, %f23, %f765, %f14;
|
|
fma.rn.f32 %f769, %f23, %f766, %f15;
|
|
fma.rn.f32 %f770, %f23, %f767, %f16;
|
|
mul.f32 %f771, %f758, 0fBF23A3A9;
|
|
mul.f32 %f772, %f761, 0fBF23A3A9;
|
|
mul.f32 %f773, %f764, 0fBF23A3A9;
|
|
fma.rn.f32 %f107, %f23, %f771, %f768;
|
|
fma.rn.f32 %f108, %f23, %f772, %f769;
|
|
fma.rn.f32 %f109, %f23, %f773, %f770;
|
|
sub.f32 %f774, %f107, %f14;
|
|
sub.f32 %f775, %f108, %f15;
|
|
sub.f32 %f776, %f109, %f16;
|
|
mul.f32 %f777, %f775, %f775;
|
|
fma.rn.f32 %f778, %f774, %f774, %f777;
|
|
fma.rn.f32 %f779, %f776, %f776, %f778;
|
|
sqrt.rn.f32 %f734, %f779;
|
|
rcp.rn.f32 %f780, %f734;
|
|
mul.f32 %f730, %f780, %f774;
|
|
mul.f32 %f731, %f780, %f775;
|
|
mul.f32 %f732, %f780, %f776;
|
|
mov.f32 %f735, 0f00000000;
|
|
// begin inline asm
|
|
call(%r589,%r590,%r591,%r592,%r593,%r594,%r595,%r596,%r597,%r598,%r599,%r600,%r601,%r602,%r603,%r604,%r605,%r606,%r607,%r608,%r609,%r610,%r611,%r612,%r613,%r614,%r615,%r616,%r617,%r618,%r619,%r620),_optix_trace_typed_32,(%r514,%rd3,%f14,%f15,%f16,%f730,%f731,%f732,%f488,%f734,%f735,%r481,%r514,%r481,%r480,%r481,%r482,%r485,%r485,%r485,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514);
|
|
// end inline asm
|
|
mov.b32 %f781, %r589;
|
|
mov.b32 %f782, %r590;
|
|
mov.b32 %f783, %r591;
|
|
setp.neu.f32 %p76, %f781, 0f3F800000;
|
|
setp.neu.f32 %p77, %f782, 0f3F800000;
|
|
or.pred %p78, %p76, %p77;
|
|
setp.neu.f32 %p79, %f783, 0f3F800000;
|
|
or.pred %p80, %p79, %p78;
|
|
@%p80 bra $L__BB0_40;
|
|
|
|
mov.f32 %f791, 0f6C4ECB8F;
|
|
mov.u32 %r699, 2;
|
|
mov.u32 %r700, 1;
|
|
mov.u32 %r701, 3;
|
|
mov.u32 %r704, 1065353216;
|
|
mov.u32 %r733, 0;
|
|
// begin inline asm
|
|
call(%r663,%r664,%r665,%r666,%r667,%r668,%r669,%r670,%r671,%r672,%r673,%r674,%r675,%r676,%r677,%r678,%r679,%r680,%r681,%r682,%r683,%r684,%r685,%r686,%r687,%r688,%r689,%r690,%r691,%r692,%r693,%r694),_optix_trace_typed_32,(%r733,%rd3,%f107,%f108,%f109,%f70,%f71,%f72,%f488,%f791,%f735,%r700,%r733,%r700,%r699,%r700,%r701,%r704,%r704,%r704,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733);
|
|
// end inline asm
|
|
mov.b32 %f793, %r663;
|
|
mov.b32 %f794, %r664;
|
|
mov.b32 %f795, %r665;
|
|
add.f32 %f2060, %f2060, %f793;
|
|
add.f32 %f2061, %f2061, %f794;
|
|
add.f32 %f2062, %f2062, %f795;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_40:
|
|
mad.lo.s32 %r805, %r51, 1664525, 1013904223;
|
|
and.b32 %r806, %r805, 16777215;
|
|
cvt.rn.f32.u32 %f805, %r806;
|
|
mul.f32 %f806, %f805, 0f33800000;
|
|
fma.rn.f32 %f807, %f806, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r52, %r805, 1664525, 1013904223;
|
|
and.b32 %r807, %r52, 16777215;
|
|
cvt.rn.f32.u32 %f808, %r807;
|
|
mul.f32 %f809, %f808, 0f33800000;
|
|
fma.rn.f32 %f810, %f809, 0f40000000, 0fBF800000;
|
|
mul.f32 %f811, %f17, %f810;
|
|
mul.f32 %f812, %f18, %f810;
|
|
mul.f32 %f813, %f19, %f810;
|
|
fma.rn.f32 %f814, %f20, %f807, %f811;
|
|
fma.rn.f32 %f815, %f21, %f807, %f812;
|
|
fma.rn.f32 %f816, %f22, %f807, %f813;
|
|
mul.f32 %f817, %f815, %f815;
|
|
fma.rn.f32 %f818, %f814, %f814, %f817;
|
|
fma.rn.f32 %f819, %f816, %f816, %f818;
|
|
sqrt.rn.f32 %f820, %f819;
|
|
rcp.rn.f32 %f821, %f820;
|
|
mul.f32 %f822, %f814, %f821;
|
|
mul.f32 %f823, %f815, %f821;
|
|
mul.f32 %f824, %f816, %f821;
|
|
mul.f32 %f825, %f2044, %f824;
|
|
mul.f32 %f826, %f2045, %f823;
|
|
sub.f32 %f827, %f825, %f826;
|
|
mul.f32 %f828, %f2045, %f822;
|
|
mul.f32 %f829, %f2043, %f824;
|
|
sub.f32 %f830, %f828, %f829;
|
|
mul.f32 %f831, %f2043, %f823;
|
|
mul.f32 %f832, %f2044, %f822;
|
|
sub.f32 %f833, %f831, %f832;
|
|
mul.f32 %f834, %f822, 0fBF23A3A9;
|
|
mul.f32 %f835, %f823, 0fBF23A3A9;
|
|
mul.f32 %f836, %f824, 0fBF23A3A9;
|
|
fma.rn.f32 %f837, %f23, %f834, %f14;
|
|
fma.rn.f32 %f838, %f23, %f835, %f15;
|
|
fma.rn.f32 %f839, %f23, %f836, %f16;
|
|
mul.f32 %f840, %f827, 0f3F23A3A8;
|
|
mul.f32 %f841, %f830, 0f3F23A3A8;
|
|
mul.f32 %f842, %f833, 0f3F23A3A8;
|
|
fma.rn.f32 %f118, %f23, %f840, %f837;
|
|
fma.rn.f32 %f119, %f23, %f841, %f838;
|
|
fma.rn.f32 %f120, %f23, %f842, %f839;
|
|
sub.f32 %f843, %f118, %f14;
|
|
sub.f32 %f844, %f119, %f15;
|
|
sub.f32 %f845, %f120, %f16;
|
|
mul.f32 %f846, %f844, %f844;
|
|
fma.rn.f32 %f847, %f843, %f843, %f846;
|
|
fma.rn.f32 %f848, %f845, %f845, %f847;
|
|
sqrt.rn.f32 %f803, %f848;
|
|
rcp.rn.f32 %f849, %f803;
|
|
mul.f32 %f799, %f849, %f843;
|
|
mul.f32 %f800, %f849, %f844;
|
|
mul.f32 %f801, %f849, %f845;
|
|
mov.u32 %r770, 2;
|
|
mov.u32 %r771, 1;
|
|
mov.u32 %r772, 3;
|
|
mov.u32 %r775, 1065353216;
|
|
mov.u32 %r804, 0;
|
|
// begin inline asm
|
|
call(%r734,%r735,%r736,%r737,%r738,%r739,%r740,%r741,%r742,%r743,%r744,%r745,%r746,%r747,%r748,%r749,%r750,%r751,%r752,%r753,%r754,%r755,%r756,%r757,%r758,%r759,%r760,%r761,%r762,%r763,%r764,%r765),_optix_trace_typed_32,(%r804,%rd3,%f14,%f15,%f16,%f799,%f800,%f801,%f488,%f803,%f735,%r771,%r804,%r771,%r770,%r771,%r772,%r775,%r775,%r775,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804);
|
|
// end inline asm
|
|
mov.b32 %f850, %r734;
|
|
mov.b32 %f851, %r735;
|
|
mov.b32 %f852, %r736;
|
|
setp.neu.f32 %p81, %f850, 0f3F800000;
|
|
setp.neu.f32 %p82, %f851, 0f3F800000;
|
|
or.pred %p83, %p81, %p82;
|
|
setp.neu.f32 %p84, %f852, 0f3F800000;
|
|
or.pred %p85, %p84, %p83;
|
|
@%p85 bra $L__BB0_42;
|
|
|
|
mov.f32 %f860, 0f6C4ECB8F;
|
|
mov.f32 %f861, 0f00000000;
|
|
// begin inline asm
|
|
call(%r808,%r809,%r810,%r811,%r812,%r813,%r814,%r815,%r816,%r817,%r818,%r819,%r820,%r821,%r822,%r823,%r824,%r825,%r826,%r827,%r828,%r829,%r830,%r831,%r832,%r833,%r834,%r835,%r836,%r837,%r838,%r839),_optix_trace_typed_32,(%r804,%rd3,%f118,%f119,%f120,%f70,%f71,%f72,%f488,%f860,%f861,%r771,%r804,%r771,%r770,%r771,%r772,%r775,%r775,%r775,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804);
|
|
// end inline asm
|
|
mov.b32 %f862, %r808;
|
|
mov.b32 %f863, %r809;
|
|
mov.b32 %f864, %r810;
|
|
add.f32 %f2060, %f2060, %f862;
|
|
add.f32 %f2061, %f2061, %f863;
|
|
add.f32 %f2062, %f2062, %f864;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_42:
|
|
mad.lo.s32 %r950, %r52, 1664525, 1013904223;
|
|
and.b32 %r951, %r950, 16777215;
|
|
cvt.rn.f32.u32 %f874, %r951;
|
|
mul.f32 %f875, %f874, 0f33800000;
|
|
fma.rn.f32 %f876, %f875, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r53, %r950, 1664525, 1013904223;
|
|
and.b32 %r952, %r53, 16777215;
|
|
cvt.rn.f32.u32 %f877, %r952;
|
|
mul.f32 %f878, %f877, 0f33800000;
|
|
fma.rn.f32 %f879, %f878, 0f40000000, 0fBF800000;
|
|
mul.f32 %f880, %f17, %f879;
|
|
mul.f32 %f881, %f18, %f879;
|
|
mul.f32 %f882, %f19, %f879;
|
|
fma.rn.f32 %f883, %f20, %f876, %f880;
|
|
fma.rn.f32 %f884, %f21, %f876, %f881;
|
|
fma.rn.f32 %f885, %f22, %f876, %f882;
|
|
mul.f32 %f886, %f884, %f884;
|
|
fma.rn.f32 %f887, %f883, %f883, %f886;
|
|
fma.rn.f32 %f888, %f885, %f885, %f887;
|
|
sqrt.rn.f32 %f889, %f888;
|
|
rcp.rn.f32 %f890, %f889;
|
|
mul.f32 %f891, %f883, %f890;
|
|
mul.f32 %f892, %f884, %f890;
|
|
mul.f32 %f893, %f885, %f890;
|
|
mul.f32 %f894, %f2044, %f893;
|
|
mul.f32 %f895, %f2045, %f892;
|
|
sub.f32 %f896, %f894, %f895;
|
|
mul.f32 %f897, %f2045, %f891;
|
|
mul.f32 %f898, %f2043, %f893;
|
|
sub.f32 %f899, %f897, %f898;
|
|
mul.f32 %f900, %f2043, %f892;
|
|
mul.f32 %f901, %f2044, %f891;
|
|
sub.f32 %f902, %f900, %f901;
|
|
mul.f32 %f903, %f891, 0fBF61E1E3;
|
|
mul.f32 %f904, %f892, 0fBF61E1E3;
|
|
mul.f32 %f905, %f893, 0fBF61E1E3;
|
|
fma.rn.f32 %f906, %f23, %f903, %f14;
|
|
fma.rn.f32 %f907, %f23, %f904, %f15;
|
|
fma.rn.f32 %f908, %f23, %f905, %f16;
|
|
mul.f32 %f909, %f896, 0fBE24A490;
|
|
mul.f32 %f910, %f899, 0fBE24A490;
|
|
mul.f32 %f911, %f902, 0fBE24A490;
|
|
fma.rn.f32 %f129, %f23, %f909, %f906;
|
|
fma.rn.f32 %f130, %f23, %f910, %f907;
|
|
fma.rn.f32 %f131, %f23, %f911, %f908;
|
|
sub.f32 %f912, %f129, %f14;
|
|
sub.f32 %f913, %f130, %f15;
|
|
sub.f32 %f914, %f131, %f16;
|
|
mul.f32 %f915, %f913, %f913;
|
|
fma.rn.f32 %f916, %f912, %f912, %f915;
|
|
fma.rn.f32 %f917, %f914, %f914, %f916;
|
|
sqrt.rn.f32 %f872, %f917;
|
|
rcp.rn.f32 %f918, %f872;
|
|
mul.f32 %f868, %f918, %f912;
|
|
mul.f32 %f869, %f918, %f913;
|
|
mul.f32 %f870, %f918, %f914;
|
|
mov.f32 %f873, 0f00000000;
|
|
// begin inline asm
|
|
call(%r879,%r880,%r881,%r882,%r883,%r884,%r885,%r886,%r887,%r888,%r889,%r890,%r891,%r892,%r893,%r894,%r895,%r896,%r897,%r898,%r899,%r900,%r901,%r902,%r903,%r904,%r905,%r906,%r907,%r908,%r909,%r910),_optix_trace_typed_32,(%r804,%rd3,%f14,%f15,%f16,%f868,%f869,%f870,%f488,%f872,%f873,%r771,%r804,%r771,%r770,%r771,%r772,%r775,%r775,%r775,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804);
|
|
// end inline asm
|
|
mov.b32 %f919, %r879;
|
|
mov.b32 %f920, %r880;
|
|
mov.b32 %f921, %r881;
|
|
setp.neu.f32 %p86, %f919, 0f3F800000;
|
|
setp.neu.f32 %p87, %f920, 0f3F800000;
|
|
or.pred %p88, %p86, %p87;
|
|
setp.neu.f32 %p89, %f921, 0f3F800000;
|
|
or.pred %p90, %p89, %p88;
|
|
@%p90 bra $L__BB0_44;
|
|
|
|
mov.f32 %f929, 0f6C4ECB8F;
|
|
mov.u32 %r989, 2;
|
|
mov.u32 %r990, 1;
|
|
mov.u32 %r991, 3;
|
|
mov.u32 %r994, 1065353216;
|
|
mov.u32 %r1023, 0;
|
|
// begin inline asm
|
|
call(%r953,%r954,%r955,%r956,%r957,%r958,%r959,%r960,%r961,%r962,%r963,%r964,%r965,%r966,%r967,%r968,%r969,%r970,%r971,%r972,%r973,%r974,%r975,%r976,%r977,%r978,%r979,%r980,%r981,%r982,%r983,%r984),_optix_trace_typed_32,(%r1023,%rd3,%f129,%f130,%f131,%f70,%f71,%f72,%f488,%f929,%f873,%r990,%r1023,%r990,%r989,%r990,%r991,%r994,%r994,%r994,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023);
|
|
// end inline asm
|
|
mov.b32 %f931, %r953;
|
|
mov.b32 %f932, %r954;
|
|
mov.b32 %f933, %r955;
|
|
add.f32 %f2060, %f2060, %f931;
|
|
add.f32 %f2061, %f2061, %f932;
|
|
add.f32 %f2062, %f2062, %f933;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_44:
|
|
mad.lo.s32 %r1095, %r53, 1664525, 1013904223;
|
|
and.b32 %r1096, %r1095, 16777215;
|
|
cvt.rn.f32.u32 %f943, %r1096;
|
|
mul.f32 %f944, %f943, 0f33800000;
|
|
fma.rn.f32 %f945, %f944, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r54, %r1095, 1664525, 1013904223;
|
|
and.b32 %r1097, %r54, 16777215;
|
|
cvt.rn.f32.u32 %f946, %r1097;
|
|
mul.f32 %f947, %f946, 0f33800000;
|
|
fma.rn.f32 %f948, %f947, 0f40000000, 0fBF800000;
|
|
mul.f32 %f949, %f17, %f948;
|
|
mul.f32 %f950, %f18, %f948;
|
|
mul.f32 %f951, %f19, %f948;
|
|
fma.rn.f32 %f952, %f20, %f945, %f949;
|
|
fma.rn.f32 %f953, %f21, %f945, %f950;
|
|
fma.rn.f32 %f954, %f22, %f945, %f951;
|
|
mul.f32 %f955, %f953, %f953;
|
|
fma.rn.f32 %f956, %f952, %f952, %f955;
|
|
fma.rn.f32 %f957, %f954, %f954, %f956;
|
|
sqrt.rn.f32 %f958, %f957;
|
|
rcp.rn.f32 %f959, %f958;
|
|
mul.f32 %f960, %f952, %f959;
|
|
mul.f32 %f961, %f953, %f959;
|
|
mul.f32 %f962, %f954, %f959;
|
|
mul.f32 %f963, %f2044, %f962;
|
|
mul.f32 %f964, %f2045, %f961;
|
|
sub.f32 %f965, %f963, %f964;
|
|
mul.f32 %f966, %f2045, %f960;
|
|
mul.f32 %f967, %f2043, %f962;
|
|
sub.f32 %f968, %f966, %f967;
|
|
mul.f32 %f969, %f2043, %f961;
|
|
mul.f32 %f970, %f2044, %f960;
|
|
sub.f32 %f971, %f969, %f970;
|
|
mul.f32 %f972, %f960, 0f3ECACABC;
|
|
mul.f32 %f973, %f961, 0f3ECACABC;
|
|
mul.f32 %f974, %f962, 0f3ECACABC;
|
|
fma.rn.f32 %f975, %f23, %f972, %f14;
|
|
fma.rn.f32 %f976, %f23, %f973, %f15;
|
|
fma.rn.f32 %f977, %f23, %f974, %f16;
|
|
mul.f32 %f978, %f965, 0f3F61E1D2;
|
|
mul.f32 %f979, %f968, 0f3F61E1D2;
|
|
mul.f32 %f980, %f971, 0f3F61E1D2;
|
|
fma.rn.f32 %f140, %f23, %f978, %f975;
|
|
fma.rn.f32 %f141, %f23, %f979, %f976;
|
|
fma.rn.f32 %f142, %f23, %f980, %f977;
|
|
sub.f32 %f981, %f140, %f14;
|
|
sub.f32 %f982, %f141, %f15;
|
|
sub.f32 %f983, %f142, %f16;
|
|
mul.f32 %f984, %f982, %f982;
|
|
fma.rn.f32 %f985, %f981, %f981, %f984;
|
|
fma.rn.f32 %f986, %f983, %f983, %f985;
|
|
sqrt.rn.f32 %f941, %f986;
|
|
rcp.rn.f32 %f987, %f941;
|
|
mul.f32 %f937, %f987, %f981;
|
|
mul.f32 %f938, %f987, %f982;
|
|
mul.f32 %f939, %f987, %f983;
|
|
mov.u32 %r1060, 2;
|
|
mov.u32 %r1061, 1;
|
|
mov.u32 %r1062, 3;
|
|
mov.u32 %r1065, 1065353216;
|
|
mov.u32 %r1094, 0;
|
|
// begin inline asm
|
|
call(%r1024,%r1025,%r1026,%r1027,%r1028,%r1029,%r1030,%r1031,%r1032,%r1033,%r1034,%r1035,%r1036,%r1037,%r1038,%r1039,%r1040,%r1041,%r1042,%r1043,%r1044,%r1045,%r1046,%r1047,%r1048,%r1049,%r1050,%r1051,%r1052,%r1053,%r1054,%r1055),_optix_trace_typed_32,(%r1094,%rd3,%f14,%f15,%f16,%f937,%f938,%f939,%f488,%f941,%f873,%r1061,%r1094,%r1061,%r1060,%r1061,%r1062,%r1065,%r1065,%r1065,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094);
|
|
// end inline asm
|
|
mov.b32 %f988, %r1024;
|
|
mov.b32 %f989, %r1025;
|
|
mov.b32 %f990, %r1026;
|
|
setp.neu.f32 %p91, %f988, 0f3F800000;
|
|
setp.neu.f32 %p92, %f989, 0f3F800000;
|
|
or.pred %p93, %p91, %p92;
|
|
setp.neu.f32 %p94, %f990, 0f3F800000;
|
|
or.pred %p95, %p94, %p93;
|
|
@%p95 bra $L__BB0_46;
|
|
|
|
mov.f32 %f998, 0f6C4ECB8F;
|
|
mov.f32 %f999, 0f00000000;
|
|
// begin inline asm
|
|
call(%r1098,%r1099,%r1100,%r1101,%r1102,%r1103,%r1104,%r1105,%r1106,%r1107,%r1108,%r1109,%r1110,%r1111,%r1112,%r1113,%r1114,%r1115,%r1116,%r1117,%r1118,%r1119,%r1120,%r1121,%r1122,%r1123,%r1124,%r1125,%r1126,%r1127,%r1128,%r1129),_optix_trace_typed_32,(%r1094,%rd3,%f140,%f141,%f142,%f70,%f71,%f72,%f488,%f998,%f999,%r1061,%r1094,%r1061,%r1060,%r1061,%r1062,%r1065,%r1065,%r1065,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094);
|
|
// end inline asm
|
|
mov.b32 %f1000, %r1098;
|
|
mov.b32 %f1001, %r1099;
|
|
mov.b32 %f1002, %r1100;
|
|
add.f32 %f2060, %f2060, %f1000;
|
|
add.f32 %f2061, %f2061, %f1001;
|
|
add.f32 %f2062, %f2062, %f1002;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_46:
|
|
mad.lo.s32 %r1240, %r54, 1664525, 1013904223;
|
|
and.b32 %r1241, %r1240, 16777215;
|
|
cvt.rn.f32.u32 %f1012, %r1241;
|
|
mul.f32 %f1013, %f1012, 0f33800000;
|
|
fma.rn.f32 %f1014, %f1013, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r1542, %r1240, 1664525, 1013904223;
|
|
and.b32 %r1242, %r1542, 16777215;
|
|
cvt.rn.f32.u32 %f1015, %r1242;
|
|
mul.f32 %f1016, %f1015, 0f33800000;
|
|
fma.rn.f32 %f1017, %f1016, 0f40000000, 0fBF800000;
|
|
mul.f32 %f1018, %f17, %f1017;
|
|
mul.f32 %f1019, %f18, %f1017;
|
|
mul.f32 %f1020, %f19, %f1017;
|
|
fma.rn.f32 %f1021, %f20, %f1014, %f1018;
|
|
fma.rn.f32 %f1022, %f21, %f1014, %f1019;
|
|
fma.rn.f32 %f1023, %f22, %f1014, %f1020;
|
|
mul.f32 %f1024, %f1022, %f1022;
|
|
fma.rn.f32 %f1025, %f1021, %f1021, %f1024;
|
|
fma.rn.f32 %f1026, %f1023, %f1023, %f1025;
|
|
sqrt.rn.f32 %f1027, %f1026;
|
|
rcp.rn.f32 %f1028, %f1027;
|
|
mul.f32 %f1029, %f1021, %f1028;
|
|
mul.f32 %f1030, %f1022, %f1028;
|
|
mul.f32 %f1031, %f1023, %f1028;
|
|
mul.f32 %f1032, %f2044, %f1031;
|
|
mul.f32 %f1033, %f2045, %f1030;
|
|
sub.f32 %f1034, %f1032, %f1033;
|
|
mul.f32 %f1035, %f2045, %f1029;
|
|
mul.f32 %f1036, %f2043, %f1031;
|
|
sub.f32 %f1037, %f1035, %f1036;
|
|
mul.f32 %f1038, %f2043, %f1030;
|
|
mul.f32 %f1039, %f2044, %f1029;
|
|
sub.f32 %f1040, %f1038, %f1039;
|
|
mul.f32 %f1041, %f1029, 0f3F61E1D2;
|
|
mul.f32 %f1042, %f1030, 0f3F61E1D2;
|
|
mul.f32 %f1043, %f1031, 0f3F61E1D2;
|
|
fma.rn.f32 %f1044, %f23, %f1041, %f14;
|
|
fma.rn.f32 %f1045, %f23, %f1042, %f15;
|
|
fma.rn.f32 %f1046, %f23, %f1043, %f16;
|
|
mul.f32 %f1047, %f1034, 0fBF61E1E3;
|
|
mul.f32 %f1048, %f1037, 0fBF61E1E3;
|
|
mul.f32 %f1049, %f1040, 0fBF61E1E3;
|
|
fma.rn.f32 %f151, %f23, %f1047, %f1044;
|
|
fma.rn.f32 %f152, %f23, %f1048, %f1045;
|
|
fma.rn.f32 %f153, %f23, %f1049, %f1046;
|
|
sub.f32 %f1050, %f151, %f14;
|
|
sub.f32 %f1051, %f152, %f15;
|
|
sub.f32 %f1052, %f153, %f16;
|
|
mul.f32 %f1053, %f1051, %f1051;
|
|
fma.rn.f32 %f1054, %f1050, %f1050, %f1053;
|
|
fma.rn.f32 %f1055, %f1052, %f1052, %f1054;
|
|
sqrt.rn.f32 %f1010, %f1055;
|
|
rcp.rn.f32 %f1056, %f1010;
|
|
mul.f32 %f1006, %f1056, %f1050;
|
|
mul.f32 %f1007, %f1056, %f1051;
|
|
mul.f32 %f1008, %f1056, %f1052;
|
|
mov.f32 %f1011, 0f00000000;
|
|
// begin inline asm
|
|
call(%r1169,%r1170,%r1171,%r1172,%r1173,%r1174,%r1175,%r1176,%r1177,%r1178,%r1179,%r1180,%r1181,%r1182,%r1183,%r1184,%r1185,%r1186,%r1187,%r1188,%r1189,%r1190,%r1191,%r1192,%r1193,%r1194,%r1195,%r1196,%r1197,%r1198,%r1199,%r1200),_optix_trace_typed_32,(%r1094,%rd3,%f14,%f15,%f16,%f1006,%f1007,%f1008,%f488,%f1010,%f1011,%r1061,%r1094,%r1061,%r1060,%r1061,%r1062,%r1065,%r1065,%r1065,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094);
|
|
// end inline asm
|
|
mov.b32 %f1057, %r1169;
|
|
mov.b32 %f1058, %r1170;
|
|
mov.b32 %f1059, %r1171;
|
|
setp.neu.f32 %p96, %f1057, 0f3F800000;
|
|
setp.neu.f32 %p97, %f1058, 0f3F800000;
|
|
or.pred %p98, %p96, %p97;
|
|
setp.neu.f32 %p99, %f1059, 0f3F800000;
|
|
or.pred %p100, %p99, %p98;
|
|
@%p100 bra $L__BB0_48;
|
|
|
|
mov.f32 %f1067, 0f6C4ECB8F;
|
|
mov.u32 %r1279, 2;
|
|
mov.u32 %r1280, 1;
|
|
mov.u32 %r1281, 3;
|
|
mov.u32 %r1284, 1065353216;
|
|
mov.u32 %r1313, 0;
|
|
// begin inline asm
|
|
call(%r1243,%r1244,%r1245,%r1246,%r1247,%r1248,%r1249,%r1250,%r1251,%r1252,%r1253,%r1254,%r1255,%r1256,%r1257,%r1258,%r1259,%r1260,%r1261,%r1262,%r1263,%r1264,%r1265,%r1266,%r1267,%r1268,%r1269,%r1270,%r1271,%r1272,%r1273,%r1274),_optix_trace_typed_32,(%r1313,%rd3,%f151,%f152,%f153,%f70,%f71,%f72,%f488,%f1067,%f1011,%r1280,%r1313,%r1280,%r1279,%r1280,%r1281,%r1284,%r1284,%r1284,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313);
|
|
// end inline asm
|
|
mov.b32 %f1069, %r1243;
|
|
mov.b32 %f1070, %r1244;
|
|
mov.b32 %f1071, %r1245;
|
|
add.f32 %f2060, %f2060, %f1069;
|
|
add.f32 %f2061, %f2061, %f1070;
|
|
add.f32 %f2062, %f2062, %f1071;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_48:
|
|
setp.ne.s16 %p101, %rs195, 0;
|
|
@%p101 bra $L__BB0_50;
|
|
|
|
mov.f32 %f1079, 0f6C4ECB8F;
|
|
mov.f32 %f1080, 0f00000000;
|
|
mov.u32 %r1350, 2;
|
|
mov.u32 %r1351, 1;
|
|
mov.u32 %r1352, 3;
|
|
mov.u32 %r1355, 1065353216;
|
|
mov.u32 %r1384, 0;
|
|
// begin inline asm
|
|
call(%r1314,%r1315,%r1316,%r1317,%r1318,%r1319,%r1320,%r1321,%r1322,%r1323,%r1324,%r1325,%r1326,%r1327,%r1328,%r1329,%r1330,%r1331,%r1332,%r1333,%r1334,%r1335,%r1336,%r1337,%r1338,%r1339,%r1340,%r1341,%r1342,%r1343,%r1344,%r1345),_optix_trace_typed_32,(%r1384,%rd3,%f14,%f15,%f16,%f70,%f71,%f72,%f488,%f1079,%f1080,%r1351,%r1384,%r1351,%r1350,%r1351,%r1352,%r1355,%r1355,%r1355,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384);
|
|
// end inline asm
|
|
mov.b32 %f1081, %r1314;
|
|
mov.b32 %f1082, %r1315;
|
|
mov.b32 %f1083, %r1316;
|
|
add.f32 %f2060, %f2060, %f1081;
|
|
add.f32 %f2061, %f2061, %f1082;
|
|
add.f32 %f2062, %f2062, %f1083;
|
|
add.f32 %f2063, %f2063, 0f3F800000;
|
|
|
|
$L__BB0_50:
|
|
add.s32 %r1541, %r1541, 1;
|
|
setp.lt.s32 %p102, %r1541, %r7;
|
|
@%p102 bra $L__BB0_8;
|
|
|
|
add.s32 %r1539, %r1539, 1;
|
|
setp.lt.s32 %p103, %r1539, %r7;
|
|
@%p103 bra $L__BB0_7;
|
|
|
|
$L__BB0_52:
|
|
setp.eq.s32 %p104, %r7, 0;
|
|
mov.f32 %f2100, 0f3F800000;
|
|
mov.f32 %f2101, %f2100;
|
|
mov.f32 %f2102, %f2100;
|
|
@%p104 bra $L__BB0_54;
|
|
|
|
mov.f32 %f1087, 0f3F800000;
|
|
max.f32 %f1088, %f2063, %f1087;
|
|
rcp.rn.f32 %f1089, %f1088;
|
|
mul.f32 %f2100, %f2060, %f1089;
|
|
mul.f32 %f2101, %f2061, %f1089;
|
|
mul.f32 %f2102, %f2062, %f1089;
|
|
|
|
$L__BB0_54:
|
|
ld.const.f32 %f1091, [params+524];
|
|
mul.f32 %f180, %f2100, %f1091;
|
|
ld.const.v2.f32 {%f1092, %f1093}, [params+528];
|
|
mul.f32 %f181, %f2101, %f1092;
|
|
mul.f32 %f182, %f2102, %f1093;
|
|
cvt.sat.f32.f32 %f1096, %f13;
|
|
mul.f32 %f183, %f180, %f1096;
|
|
mul.f32 %f184, %f181, %f1096;
|
|
mul.f32 %f185, %f182, %f1096;
|
|
fma.rn.f32 %f1097, %f13, 0f3F000000, 0f3F000000;
|
|
cvt.sat.f32.f32 %f1098, %f1097;
|
|
add.f32 %f186, %f1098, %f1098;
|
|
mov.f32 %f1099, 0f41200000;
|
|
cvt.rzi.f32.f32 %f1100, %f1099;
|
|
add.f32 %f1101, %f1100, %f1100;
|
|
mov.f32 %f1102, 0f41A00000;
|
|
sub.f32 %f1103, %f1102, %f1101;
|
|
abs.f32 %f187, %f1103;
|
|
abs.f32 %f188, %f186;
|
|
setp.lt.f32 %p105, %f188, 0f00800000;
|
|
mul.f32 %f1104, %f188, 0f4B800000;
|
|
selp.f32 %f1105, %f1104, %f188, %p105;
|
|
selp.f32 %f1106, 0fC3170000, 0fC2FE0000, %p105;
|
|
mov.b32 %r1385, %f1105;
|
|
and.b32 %r1386, %r1385, 8388607;
|
|
or.b32 %r1387, %r1386, 1065353216;
|
|
mov.b32 %f1107, %r1387;
|
|
shr.u32 %r1388, %r1385, 23;
|
|
cvt.rn.f32.u32 %f1108, %r1388;
|
|
add.f32 %f1109, %f1106, %f1108;
|
|
setp.gt.f32 %p106, %f1107, 0f3FB504F3;
|
|
mul.f32 %f1110, %f1107, 0f3F000000;
|
|
add.f32 %f1111, %f1109, 0f3F800000;
|
|
selp.f32 %f1112, %f1111, %f1109, %p106;
|
|
selp.f32 %f1113, %f1110, %f1107, %p106;
|
|
add.f32 %f1114, %f1113, 0fBF800000;
|
|
add.f32 %f1115, %f1113, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1116, %f1115;
|
|
add.f32 %f1117, %f1114, %f1114;
|
|
mul.f32 %f1118, %f1117, %f1116;
|
|
mul.f32 %f1119, %f1118, %f1118;
|
|
mov.f32 %f1120, 0f3C4CAF63;
|
|
mov.f32 %f1121, 0f3B18F0FE;
|
|
fma.rn.f32 %f1122, %f1121, %f1119, %f1120;
|
|
mov.f32 %f1123, 0f3DAAAABD;
|
|
fma.rn.f32 %f1124, %f1122, %f1119, %f1123;
|
|
mul.rn.f32 %f1125, %f1124, %f1119;
|
|
mul.rn.f32 %f1126, %f1125, %f1118;
|
|
sub.f32 %f1127, %f1114, %f1118;
|
|
add.f32 %f1128, %f1127, %f1127;
|
|
neg.f32 %f1129, %f1118;
|
|
fma.rn.f32 %f1130, %f1129, %f1114, %f1128;
|
|
mul.rn.f32 %f1131, %f1116, %f1130;
|
|
add.f32 %f1132, %f1126, %f1118;
|
|
sub.f32 %f1133, %f1118, %f1132;
|
|
add.f32 %f1134, %f1126, %f1133;
|
|
add.f32 %f1135, %f1131, %f1134;
|
|
add.f32 %f1136, %f1132, %f1135;
|
|
sub.f32 %f1137, %f1132, %f1136;
|
|
add.f32 %f1138, %f1135, %f1137;
|
|
mov.f32 %f1139, 0f3F317200;
|
|
mul.rn.f32 %f1140, %f1112, %f1139;
|
|
mov.f32 %f1141, 0f35BFBE8E;
|
|
mul.rn.f32 %f1142, %f1112, %f1141;
|
|
add.f32 %f1143, %f1140, %f1136;
|
|
sub.f32 %f1144, %f1140, %f1143;
|
|
add.f32 %f1145, %f1136, %f1144;
|
|
add.f32 %f1146, %f1138, %f1145;
|
|
add.f32 %f1147, %f1142, %f1146;
|
|
add.f32 %f1148, %f1143, %f1147;
|
|
sub.f32 %f1149, %f1143, %f1148;
|
|
add.f32 %f1150, %f1147, %f1149;
|
|
mul.rn.f32 %f1151, %f1102, %f1148;
|
|
neg.f32 %f1152, %f1151;
|
|
fma.rn.f32 %f1153, %f1102, %f1148, %f1152;
|
|
fma.rn.f32 %f1154, %f1102, %f1150, %f1153;
|
|
mov.f32 %f1155, 0f00000000;
|
|
fma.rn.f32 %f1156, %f1155, %f1148, %f1154;
|
|
add.rn.f32 %f1157, %f1151, %f1156;
|
|
neg.f32 %f1158, %f1157;
|
|
add.rn.f32 %f1159, %f1151, %f1158;
|
|
add.rn.f32 %f1160, %f1159, %f1156;
|
|
mov.b32 %r1389, %f1157;
|
|
setp.eq.s32 %p107, %r1389, 1118925336;
|
|
add.s32 %r1390, %r1389, -1;
|
|
mov.b32 %f1161, %r1390;
|
|
add.f32 %f1162, %f1160, 0f37000000;
|
|
selp.f32 %f189, %f1162, %f1160, %p107;
|
|
selp.f32 %f1163, %f1161, %f1157, %p107;
|
|
mov.f32 %f1164, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1165, %f1163, %f1164;
|
|
cvt.rzi.f32.f32 %f1166, %f1165;
|
|
abs.f32 %f1167, %f1166;
|
|
setp.gt.f32 %p108, %f1167, 0f42FC0000;
|
|
mov.b32 %r1391, %f1166;
|
|
and.b32 %r1392, %r1391, -2147483648;
|
|
or.b32 %r1393, %r1392, 1123811328;
|
|
mov.b32 %f1168, %r1393;
|
|
selp.f32 %f1169, %f1168, %f1166, %p108;
|
|
mov.f32 %f1170, 0fBF317218;
|
|
fma.rn.f32 %f1171, %f1169, %f1170, %f1163;
|
|
mov.f32 %f1172, 0f3102E308;
|
|
fma.rn.f32 %f1173, %f1169, %f1172, %f1171;
|
|
mul.f32 %f1174, %f1173, 0f3FB8AA3B;
|
|
add.f32 %f1175, %f1169, 0f4B40007F;
|
|
mov.b32 %r1394, %f1175;
|
|
shl.b32 %r1395, %r1394, 23;
|
|
mov.b32 %f1176, %r1395;
|
|
ex2.approx.ftz.f32 %f1177, %f1174;
|
|
mul.f32 %f190, %f1177, %f1176;
|
|
setp.eq.f32 %p109, %f190, 0f7F800000;
|
|
mov.f32 %f2103, 0f7F800000;
|
|
@%p109 bra $L__BB0_56;
|
|
|
|
fma.rn.f32 %f2103, %f190, %f189, %f190;
|
|
|
|
$L__BB0_56:
|
|
setp.lt.f32 %p110, %f186, 0f00000000;
|
|
setp.eq.f32 %p111, %f187, 0f3F800000;
|
|
and.pred %p1, %p110, %p111;
|
|
setp.eq.f32 %p112, %f186, 0f00000000;
|
|
@%p112 bra $L__BB0_60;
|
|
bra.uni $L__BB0_57;
|
|
|
|
$L__BB0_60:
|
|
add.f32 %f1182, %f186, %f186;
|
|
selp.f32 %f2105, %f1182, 0f00000000, %p111;
|
|
bra.uni $L__BB0_61;
|
|
|
|
$L__BB0_57:
|
|
mov.b32 %r1396, %f2103;
|
|
xor.b32 %r1397, %r1396, -2147483648;
|
|
mov.b32 %f1178, %r1397;
|
|
selp.f32 %f2105, %f1178, %f2103, %p1;
|
|
setp.geu.f32 %p113, %f186, 0f00000000;
|
|
@%p113 bra $L__BB0_61;
|
|
|
|
mov.f32 %f1179, 0f41A00000;
|
|
cvt.rzi.f32.f32 %f1180, %f1179;
|
|
setp.eq.f32 %p114, %f1180, 0f41A00000;
|
|
@%p114 bra $L__BB0_61;
|
|
|
|
mov.f32 %f2105, 0f7FFFFFFF;
|
|
|
|
$L__BB0_61:
|
|
add.f32 %f1183, %f188, 0f41A00000;
|
|
mov.b32 %r1398, %f1183;
|
|
setp.lt.s32 %p116, %r1398, 2139095040;
|
|
@%p116 bra $L__BB0_66;
|
|
|
|
setp.gtu.f32 %p117, %f188, 0f7F800000;
|
|
@%p117 bra $L__BB0_65;
|
|
bra.uni $L__BB0_63;
|
|
|
|
$L__BB0_65:
|
|
add.f32 %f2105, %f186, 0f41A00000;
|
|
bra.uni $L__BB0_66;
|
|
|
|
$L__BB0_63:
|
|
setp.neu.f32 %p118, %f188, 0f7F800000;
|
|
@%p118 bra $L__BB0_66;
|
|
|
|
selp.f32 %f2105, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
$L__BB0_66:
|
|
setp.eq.f32 %p119, %f186, 0f3F800000;
|
|
selp.f32 %f1184, 0f3F800000, %f2105, %p119;
|
|
cvt.sat.f32.f32 %f1185, %f1184;
|
|
mul.f32 %f199, %f180, %f1185;
|
|
mul.f32 %f200, %f181, %f1185;
|
|
mul.f32 %f201, %f182, %f1185;
|
|
ld.const.u32 %r58, [params+104];
|
|
and.b32 %r1399, %r58, 8;
|
|
setp.eq.s32 %p120, %r1399, 0;
|
|
@%p120 bra $L__BB0_80;
|
|
|
|
ld.const.u64 %rd85, [params+192];
|
|
cvta.to.global.u64 %rd15, %rd85;
|
|
ld.const.u32 %r1400, [params+184];
|
|
mad.lo.s32 %r1401, %r1400, %r6, %r5;
|
|
cvt.u64.u32 %rd16, %r1401;
|
|
mov.f32 %f1187, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f1188, %f1187;
|
|
add.f32 %f1189, %f1188, %f1188;
|
|
mov.f32 %f1190, 0f3EE8BA2E;
|
|
sub.f32 %f1191, %f1190, %f1189;
|
|
abs.f32 %f202, %f1191;
|
|
abs.f32 %f203, %f2100;
|
|
setp.lt.f32 %p121, %f203, 0f00800000;
|
|
mul.f32 %f1192, %f203, 0f4B800000;
|
|
selp.f32 %f1193, %f1192, %f203, %p121;
|
|
selp.f32 %f1194, 0fC3170000, 0fC2FE0000, %p121;
|
|
mov.b32 %r1402, %f1193;
|
|
and.b32 %r1403, %r1402, 8388607;
|
|
or.b32 %r1404, %r1403, 1065353216;
|
|
mov.b32 %f1195, %r1404;
|
|
shr.u32 %r1405, %r1402, 23;
|
|
cvt.rn.f32.u32 %f1196, %r1405;
|
|
add.f32 %f1197, %f1194, %f1196;
|
|
setp.gt.f32 %p122, %f1195, 0f3FB504F3;
|
|
mul.f32 %f1198, %f1195, 0f3F000000;
|
|
add.f32 %f1199, %f1197, 0f3F800000;
|
|
selp.f32 %f1200, %f1199, %f1197, %p122;
|
|
selp.f32 %f1201, %f1198, %f1195, %p122;
|
|
add.f32 %f1202, %f1201, 0fBF800000;
|
|
add.f32 %f1203, %f1201, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1204, %f1203;
|
|
add.f32 %f1205, %f1202, %f1202;
|
|
mul.f32 %f1206, %f1205, %f1204;
|
|
mul.f32 %f1207, %f1206, %f1206;
|
|
mov.f32 %f1208, 0f3C4CAF63;
|
|
mov.f32 %f1209, 0f3B18F0FE;
|
|
fma.rn.f32 %f1210, %f1209, %f1207, %f1208;
|
|
mov.f32 %f1211, 0f3DAAAABD;
|
|
fma.rn.f32 %f1212, %f1210, %f1207, %f1211;
|
|
mul.rn.f32 %f1213, %f1212, %f1207;
|
|
mul.rn.f32 %f1214, %f1213, %f1206;
|
|
sub.f32 %f1215, %f1202, %f1206;
|
|
add.f32 %f1216, %f1215, %f1215;
|
|
neg.f32 %f1217, %f1206;
|
|
fma.rn.f32 %f1218, %f1217, %f1202, %f1216;
|
|
mul.rn.f32 %f1219, %f1204, %f1218;
|
|
add.f32 %f1220, %f1214, %f1206;
|
|
sub.f32 %f1221, %f1206, %f1220;
|
|
add.f32 %f1222, %f1214, %f1221;
|
|
add.f32 %f1223, %f1219, %f1222;
|
|
add.f32 %f1224, %f1220, %f1223;
|
|
sub.f32 %f1225, %f1220, %f1224;
|
|
add.f32 %f1226, %f1223, %f1225;
|
|
mov.f32 %f1227, 0f3F317200;
|
|
mul.rn.f32 %f1228, %f1200, %f1227;
|
|
mov.f32 %f1229, 0f35BFBE8E;
|
|
mul.rn.f32 %f1230, %f1200, %f1229;
|
|
add.f32 %f1231, %f1228, %f1224;
|
|
sub.f32 %f1232, %f1228, %f1231;
|
|
add.f32 %f1233, %f1224, %f1232;
|
|
add.f32 %f1234, %f1226, %f1233;
|
|
add.f32 %f1235, %f1230, %f1234;
|
|
add.f32 %f1236, %f1231, %f1235;
|
|
sub.f32 %f1237, %f1231, %f1236;
|
|
add.f32 %f1238, %f1235, %f1237;
|
|
mul.rn.f32 %f1239, %f1190, %f1236;
|
|
neg.f32 %f1240, %f1239;
|
|
fma.rn.f32 %f1241, %f1190, %f1236, %f1240;
|
|
fma.rn.f32 %f1242, %f1190, %f1238, %f1241;
|
|
mov.f32 %f1243, 0f00000000;
|
|
fma.rn.f32 %f1244, %f1243, %f1236, %f1242;
|
|
add.rn.f32 %f1245, %f1239, %f1244;
|
|
neg.f32 %f1246, %f1245;
|
|
add.rn.f32 %f1247, %f1239, %f1246;
|
|
add.rn.f32 %f1248, %f1247, %f1244;
|
|
mov.b32 %r1406, %f1245;
|
|
setp.eq.s32 %p123, %r1406, 1118925336;
|
|
add.s32 %r1407, %r1406, -1;
|
|
mov.b32 %f1249, %r1407;
|
|
add.f32 %f1250, %f1248, 0f37000000;
|
|
selp.f32 %f204, %f1250, %f1248, %p123;
|
|
selp.f32 %f1251, %f1249, %f1245, %p123;
|
|
mov.f32 %f1252, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1253, %f1251, %f1252;
|
|
cvt.rzi.f32.f32 %f1254, %f1253;
|
|
abs.f32 %f1255, %f1254;
|
|
setp.gt.f32 %p124, %f1255, 0f42FC0000;
|
|
mov.b32 %r1408, %f1254;
|
|
and.b32 %r1409, %r1408, -2147483648;
|
|
or.b32 %r1410, %r1409, 1123811328;
|
|
mov.b32 %f1256, %r1410;
|
|
selp.f32 %f1257, %f1256, %f1254, %p124;
|
|
mov.f32 %f1258, 0fBF317218;
|
|
fma.rn.f32 %f1259, %f1257, %f1258, %f1251;
|
|
mov.f32 %f1260, 0f3102E308;
|
|
fma.rn.f32 %f1261, %f1257, %f1260, %f1259;
|
|
mul.f32 %f1262, %f1261, 0f3FB8AA3B;
|
|
add.f32 %f1263, %f1257, 0f4B40007F;
|
|
mov.b32 %r1411, %f1263;
|
|
shl.b32 %r1412, %r1411, 23;
|
|
mov.b32 %f1264, %r1412;
|
|
ex2.approx.ftz.f32 %f1265, %f1262;
|
|
mul.f32 %f205, %f1265, %f1264;
|
|
setp.eq.f32 %p125, %f205, 0f7F800000;
|
|
mov.f32 %f2106, 0f7F800000;
|
|
@%p125 bra $L__BB0_69;
|
|
|
|
fma.rn.f32 %f2106, %f205, %f204, %f205;
|
|
|
|
$L__BB0_69:
|
|
setp.lt.f32 %p126, %f2100, 0f00000000;
|
|
setp.eq.f32 %p127, %f202, 0f3F800000;
|
|
and.pred %p2, %p126, %p127;
|
|
setp.eq.f32 %p128, %f2100, 0f00000000;
|
|
@%p128 bra $L__BB0_73;
|
|
bra.uni $L__BB0_70;
|
|
|
|
$L__BB0_73:
|
|
add.f32 %f1270, %f2100, %f2100;
|
|
selp.f32 %f2108, %f1270, 0f00000000, %p127;
|
|
bra.uni $L__BB0_74;
|
|
|
|
$L__BB0_182:
|
|
mov.f32 %f2014, 0f00000000;
|
|
mov.u32 %r1550, 0;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs145, %f2014;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f2014;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f2014;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs146, 0;
|
|
st.global.v4.u16 [%rd24], {%rs143, %rs144, %rs145, %rs146};
|
|
|
|
$L__BB0_183:
|
|
ld.const.u64 %rd117, [params+256];
|
|
cvta.to.global.u64 %rd118, %rd117;
|
|
ld.const.u32 %r1533, [params+248];
|
|
mad.lo.s32 %r1534, %r1533, %r6, %r5;
|
|
mul.wide.u32 %rd119, %r1534, 8;
|
|
add.s64 %rd25, %rd118, %rd119;
|
|
setp.eq.s32 %p253, %r1550, 0;
|
|
@%p253 bra $L__BB0_185;
|
|
|
|
ld.global.v4.u16 {%rs153, %rs154, %rs155, %rs156}, [%rd25];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2015, %rs153;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2016, %rs154;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2017, %rs155;}
|
|
|
|
// end inline asm
|
|
add.f32 %f2018, %f2015, 0f00000000;
|
|
add.f32 %f2019, %f2016, 0f00000000;
|
|
add.f32 %f2020, %f2017, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f2020;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f2019;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f2018;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs157, 0;
|
|
st.global.v4.u16 [%rd25], {%rs150, %rs151, %rs152, %rs157};
|
|
bra.uni $L__BB0_186;
|
|
|
|
$L__BB0_185:
|
|
mov.f32 %f2023, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs160, %f2023;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs159, %f2023;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs158, %f2023;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs161, 0;
|
|
st.global.v4.u16 [%rd25], {%rs158, %rs159, %rs160, %rs161};
|
|
|
|
$L__BB0_186:
|
|
ld.const.u64 %rd120, [params+272];
|
|
cvta.to.global.u64 %rd121, %rd120;
|
|
ld.const.u32 %r1535, [params+264];
|
|
mad.lo.s32 %r1536, %r1535, %r6, %r5;
|
|
mul.wide.u32 %rd122, %r1536, 8;
|
|
add.s64 %rd26, %rd121, %rd122;
|
|
@%p253 bra $L__BB0_188;
|
|
|
|
ld.global.v4.u16 {%rs168, %rs169, %rs170, %rs171}, [%rd26];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2024, %rs168;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2025, %rs169;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2026, %rs170;}
|
|
|
|
// end inline asm
|
|
add.f32 %f2027, %f2024, 0f00000000;
|
|
add.f32 %f2028, %f2025, 0f00000000;
|
|
add.f32 %f2029, %f2026, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs167, %f2029;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs166, %f2028;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs165, %f2027;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs172, 0;
|
|
st.global.v4.u16 [%rd26], {%rs165, %rs166, %rs167, %rs172};
|
|
bra.uni $L__BB0_189;
|
|
|
|
$L__BB0_188:
|
|
mov.f32 %f2032, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs175, %f2032;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs174, %f2032;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f2032;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs176, 0;
|
|
st.global.v4.u16 [%rd26], {%rs173, %rs174, %rs175, %rs176};
|
|
|
|
$L__BB0_189:
|
|
ld.const.u64 %rd123, [params+288];
|
|
cvta.to.global.u64 %rd124, %rd123;
|
|
ld.const.u32 %r1537, [params+280];
|
|
mad.lo.s32 %r1538, %r1537, %r6, %r5;
|
|
mul.wide.u32 %rd125, %r1538, 8;
|
|
add.s64 %rd27, %rd124, %rd125;
|
|
@%p253 bra $L__BB0_191;
|
|
|
|
ld.global.v4.u16 {%rs183, %rs184, %rs185, %rs186}, [%rd27];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2033, %rs183;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2034, %rs184;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f2035, %rs185;}
|
|
|
|
// end inline asm
|
|
add.f32 %f2036, %f2033, 0f00000000;
|
|
add.f32 %f2037, %f2034, 0f00000000;
|
|
add.f32 %f2038, %f2035, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs182, %f2038;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs181, %f2037;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs180, %f2036;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs187, 0;
|
|
st.global.v4.u16 [%rd27], {%rs180, %rs181, %rs182, %rs187};
|
|
bra.uni $L__BB0_192;
|
|
|
|
$L__BB0_191:
|
|
mov.f32 %f2041, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs190, %f2041;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs189, %f2041;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs188, %f2041;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs191, 0;
|
|
st.global.v4.u16 [%rd27], {%rs188, %rs189, %rs190, %rs191};
|
|
bra.uni $L__BB0_192;
|
|
|
|
$L__BB0_70:
|
|
mov.b32 %r1413, %f2106;
|
|
xor.b32 %r1414, %r1413, -2147483648;
|
|
mov.b32 %f1266, %r1414;
|
|
selp.f32 %f2108, %f1266, %f2106, %p2;
|
|
setp.geu.f32 %p129, %f2100, 0f00000000;
|
|
@%p129 bra $L__BB0_74;
|
|
|
|
mov.f32 %f1267, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f1268, %f1267;
|
|
setp.eq.f32 %p130, %f1268, 0f3EE8BA2E;
|
|
@%p130 bra $L__BB0_74;
|
|
|
|
mov.f32 %f2108, 0f7FFFFFFF;
|
|
|
|
$L__BB0_74:
|
|
add.f32 %f1271, %f203, 0f3EE8BA2E;
|
|
mov.b32 %r1415, %f1271;
|
|
setp.lt.s32 %p132, %r1415, 2139095040;
|
|
@%p132 bra $L__BB0_79;
|
|
|
|
setp.gtu.f32 %p133, %f203, 0f7F800000;
|
|
@%p133 bra $L__BB0_78;
|
|
bra.uni $L__BB0_76;
|
|
|
|
$L__BB0_78:
|
|
add.f32 %f2108, %f2100, 0f3EE8BA2E;
|
|
bra.uni $L__BB0_79;
|
|
|
|
$L__BB0_76:
|
|
setp.neu.f32 %p134, %f203, 0f7F800000;
|
|
@%p134 bra $L__BB0_79;
|
|
|
|
selp.f32 %f2108, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
$L__BB0_79:
|
|
mul.f32 %f1272, %f2108, 0f437F0000;
|
|
setp.eq.f32 %p135, %f2100, 0f3F800000;
|
|
selp.f32 %f1273, 0f437F0000, %f1272, %p135;
|
|
cvt.rzi.u32.f32 %r1416, %f1273;
|
|
shl.b64 %rd86, %rd16, 1;
|
|
add.s64 %rd87, %rd15, %rd86;
|
|
cvt.u16.u32 %rs43, %r1416;
|
|
mov.u16 %rs44, 255;
|
|
st.global.v2.u8 [%rd87], {%rs43, %rs44};
|
|
|
|
$L__BB0_80:
|
|
and.b32 %r1417, %r58, 1;
|
|
setp.eq.b32 %p136, %r1417, 1;
|
|
mov.pred %p137, 0;
|
|
xor.pred %p138, %p136, %p137;
|
|
not.pred %p139, %p138;
|
|
@%p139 bra $L__BB0_154;
|
|
|
|
mov.f32 %f1275, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f1276, %f1275;
|
|
add.f32 %f1277, %f1276, %f1276;
|
|
mov.f32 %f1278, 0f3EE66666;
|
|
sub.f32 %f1279, %f1278, %f1277;
|
|
abs.f32 %f214, %f1279;
|
|
abs.f32 %f215, %f183;
|
|
setp.lt.f32 %p140, %f215, 0f00800000;
|
|
mul.f32 %f1280, %f215, 0f4B800000;
|
|
selp.f32 %f1281, %f1280, %f215, %p140;
|
|
selp.f32 %f1282, 0fC3170000, 0fC2FE0000, %p140;
|
|
mov.b32 %r1418, %f1281;
|
|
and.b32 %r1419, %r1418, 8388607;
|
|
or.b32 %r1420, %r1419, 1065353216;
|
|
mov.b32 %f1283, %r1420;
|
|
shr.u32 %r1421, %r1418, 23;
|
|
cvt.rn.f32.u32 %f1284, %r1421;
|
|
add.f32 %f1285, %f1282, %f1284;
|
|
setp.gt.f32 %p141, %f1283, 0f3FB504F3;
|
|
mul.f32 %f1286, %f1283, 0f3F000000;
|
|
add.f32 %f1287, %f1285, 0f3F800000;
|
|
selp.f32 %f1288, %f1287, %f1285, %p141;
|
|
selp.f32 %f1289, %f1286, %f1283, %p141;
|
|
add.f32 %f1290, %f1289, 0fBF800000;
|
|
add.f32 %f1291, %f1289, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1292, %f1291;
|
|
add.f32 %f1293, %f1290, %f1290;
|
|
mul.f32 %f1294, %f1293, %f1292;
|
|
mul.f32 %f1295, %f1294, %f1294;
|
|
mov.f32 %f1296, 0f3C4CAF63;
|
|
mov.f32 %f1297, 0f3B18F0FE;
|
|
fma.rn.f32 %f1298, %f1297, %f1295, %f1296;
|
|
mov.f32 %f1299, 0f3DAAAABD;
|
|
fma.rn.f32 %f1300, %f1298, %f1295, %f1299;
|
|
mul.rn.f32 %f1301, %f1300, %f1295;
|
|
mul.rn.f32 %f1302, %f1301, %f1294;
|
|
sub.f32 %f1303, %f1290, %f1294;
|
|
add.f32 %f1304, %f1303, %f1303;
|
|
neg.f32 %f1305, %f1294;
|
|
fma.rn.f32 %f1306, %f1305, %f1290, %f1304;
|
|
mul.rn.f32 %f1307, %f1292, %f1306;
|
|
add.f32 %f1308, %f1302, %f1294;
|
|
sub.f32 %f1309, %f1294, %f1308;
|
|
add.f32 %f1310, %f1302, %f1309;
|
|
add.f32 %f1311, %f1307, %f1310;
|
|
add.f32 %f1312, %f1308, %f1311;
|
|
sub.f32 %f1313, %f1308, %f1312;
|
|
add.f32 %f1314, %f1311, %f1313;
|
|
mov.f32 %f1315, 0f3F317200;
|
|
mul.rn.f32 %f1316, %f1288, %f1315;
|
|
mov.f32 %f1317, 0f35BFBE8E;
|
|
mul.rn.f32 %f1318, %f1288, %f1317;
|
|
add.f32 %f1319, %f1316, %f1312;
|
|
sub.f32 %f1320, %f1316, %f1319;
|
|
add.f32 %f1321, %f1312, %f1320;
|
|
add.f32 %f1322, %f1314, %f1321;
|
|
add.f32 %f1323, %f1318, %f1322;
|
|
add.f32 %f1324, %f1319, %f1323;
|
|
sub.f32 %f1325, %f1319, %f1324;
|
|
add.f32 %f1326, %f1323, %f1325;
|
|
mul.rn.f32 %f1327, %f1278, %f1324;
|
|
neg.f32 %f1328, %f1327;
|
|
fma.rn.f32 %f1329, %f1278, %f1324, %f1328;
|
|
fma.rn.f32 %f1330, %f1278, %f1326, %f1329;
|
|
mov.f32 %f1331, 0f00000000;
|
|
fma.rn.f32 %f1332, %f1331, %f1324, %f1330;
|
|
add.rn.f32 %f1333, %f1327, %f1332;
|
|
neg.f32 %f1334, %f1333;
|
|
add.rn.f32 %f1335, %f1327, %f1334;
|
|
add.rn.f32 %f1336, %f1335, %f1332;
|
|
mov.b32 %r1422, %f1333;
|
|
setp.eq.s32 %p142, %r1422, 1118925336;
|
|
add.s32 %r1423, %r1422, -1;
|
|
mov.b32 %f1337, %r1423;
|
|
add.f32 %f1338, %f1336, 0f37000000;
|
|
selp.f32 %f216, %f1338, %f1336, %p142;
|
|
selp.f32 %f1339, %f1337, %f1333, %p142;
|
|
mov.f32 %f1340, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1341, %f1339, %f1340;
|
|
cvt.rzi.f32.f32 %f1342, %f1341;
|
|
abs.f32 %f1343, %f1342;
|
|
setp.gt.f32 %p143, %f1343, 0f42FC0000;
|
|
mov.b32 %r1424, %f1342;
|
|
and.b32 %r1425, %r1424, -2147483648;
|
|
or.b32 %r1426, %r1425, 1123811328;
|
|
mov.b32 %f1344, %r1426;
|
|
selp.f32 %f1345, %f1344, %f1342, %p143;
|
|
mov.f32 %f1346, 0fBF317218;
|
|
fma.rn.f32 %f1347, %f1345, %f1346, %f1339;
|
|
mov.f32 %f1348, 0f3102E308;
|
|
fma.rn.f32 %f1349, %f1345, %f1348, %f1347;
|
|
mul.f32 %f1350, %f1349, 0f3FB8AA3B;
|
|
add.f32 %f1351, %f1345, 0f4B40007F;
|
|
mov.b32 %r1427, %f1351;
|
|
shl.b32 %r1428, %r1427, 23;
|
|
mov.b32 %f1352, %r1428;
|
|
ex2.approx.ftz.f32 %f1353, %f1350;
|
|
mul.f32 %f217, %f1353, %f1352;
|
|
setp.eq.f32 %p144, %f217, 0f7F800000;
|
|
mov.f32 %f2109, 0f7F800000;
|
|
@%p144 bra $L__BB0_83;
|
|
|
|
fma.rn.f32 %f2109, %f217, %f216, %f217;
|
|
|
|
$L__BB0_83:
|
|
setp.lt.f32 %p145, %f183, 0f00000000;
|
|
setp.eq.f32 %p146, %f214, 0f3F800000;
|
|
and.pred %p3, %p145, %p146;
|
|
setp.eq.f32 %p147, %f183, 0f00000000;
|
|
@%p147 bra $L__BB0_87;
|
|
bra.uni $L__BB0_84;
|
|
|
|
$L__BB0_87:
|
|
add.f32 %f1358, %f183, %f183;
|
|
selp.f32 %f2111, %f1358, 0f00000000, %p146;
|
|
bra.uni $L__BB0_88;
|
|
|
|
$L__BB0_84:
|
|
mov.b32 %r1429, %f2109;
|
|
xor.b32 %r1430, %r1429, -2147483648;
|
|
mov.b32 %f1354, %r1430;
|
|
selp.f32 %f2111, %f1354, %f2109, %p3;
|
|
setp.geu.f32 %p148, %f183, 0f00000000;
|
|
@%p148 bra $L__BB0_88;
|
|
|
|
mov.f32 %f1355, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f1356, %f1355;
|
|
setp.eq.f32 %p149, %f1356, 0f3EE66666;
|
|
@%p149 bra $L__BB0_88;
|
|
|
|
mov.f32 %f2111, 0f7FFFFFFF;
|
|
|
|
$L__BB0_88:
|
|
add.f32 %f1359, %f215, 0f3EE66666;
|
|
mov.b32 %r1431, %f1359;
|
|
setp.lt.s32 %p151, %r1431, 2139095040;
|
|
@%p151 bra $L__BB0_93;
|
|
|
|
setp.gtu.f32 %p152, %f215, 0f7F800000;
|
|
@%p152 bra $L__BB0_92;
|
|
bra.uni $L__BB0_90;
|
|
|
|
$L__BB0_92:
|
|
add.f32 %f2111, %f183, 0f3EE66666;
|
|
bra.uni $L__BB0_93;
|
|
|
|
$L__BB0_90:
|
|
setp.neu.f32 %p153, %f215, 0f7F800000;
|
|
@%p153 bra $L__BB0_93;
|
|
|
|
selp.f32 %f2111, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
$L__BB0_93:
|
|
setp.eq.f32 %p154, %f183, 0f3F800000;
|
|
selp.f32 %f226, 0f3F800000, %f2111, %p154;
|
|
abs.f32 %f227, %f184;
|
|
setp.lt.f32 %p155, %f227, 0f00800000;
|
|
mul.f32 %f1361, %f227, 0f4B800000;
|
|
selp.f32 %f1362, %f1361, %f227, %p155;
|
|
selp.f32 %f1363, 0fC3170000, 0fC2FE0000, %p155;
|
|
mov.b32 %r1432, %f1362;
|
|
and.b32 %r1433, %r1432, 8388607;
|
|
or.b32 %r1434, %r1433, 1065353216;
|
|
mov.b32 %f1364, %r1434;
|
|
shr.u32 %r1435, %r1432, 23;
|
|
cvt.rn.f32.u32 %f1365, %r1435;
|
|
add.f32 %f1366, %f1363, %f1365;
|
|
setp.gt.f32 %p156, %f1364, 0f3FB504F3;
|
|
mul.f32 %f1367, %f1364, 0f3F000000;
|
|
add.f32 %f1368, %f1366, 0f3F800000;
|
|
selp.f32 %f1369, %f1368, %f1366, %p156;
|
|
selp.f32 %f1370, %f1367, %f1364, %p156;
|
|
add.f32 %f1371, %f1370, 0fBF800000;
|
|
add.f32 %f1372, %f1370, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1373, %f1372;
|
|
add.f32 %f1374, %f1371, %f1371;
|
|
mul.f32 %f1375, %f1374, %f1373;
|
|
mul.f32 %f1376, %f1375, %f1375;
|
|
mov.f32 %f1377, 0f3C4CAF63;
|
|
mov.f32 %f1378, 0f3B18F0FE;
|
|
fma.rn.f32 %f1379, %f1378, %f1376, %f1377;
|
|
mov.f32 %f1380, 0f3DAAAABD;
|
|
fma.rn.f32 %f1381, %f1379, %f1376, %f1380;
|
|
mul.rn.f32 %f1382, %f1381, %f1376;
|
|
mul.rn.f32 %f1383, %f1382, %f1375;
|
|
sub.f32 %f1384, %f1371, %f1375;
|
|
add.f32 %f1385, %f1384, %f1384;
|
|
neg.f32 %f1386, %f1375;
|
|
fma.rn.f32 %f1387, %f1386, %f1371, %f1385;
|
|
mul.rn.f32 %f1388, %f1373, %f1387;
|
|
add.f32 %f1389, %f1383, %f1375;
|
|
sub.f32 %f1390, %f1375, %f1389;
|
|
add.f32 %f1391, %f1383, %f1390;
|
|
add.f32 %f1392, %f1388, %f1391;
|
|
add.f32 %f1393, %f1389, %f1392;
|
|
sub.f32 %f1394, %f1389, %f1393;
|
|
add.f32 %f1395, %f1392, %f1394;
|
|
mov.f32 %f1396, 0f3F317200;
|
|
mul.rn.f32 %f1397, %f1369, %f1396;
|
|
mov.f32 %f1398, 0f35BFBE8E;
|
|
mul.rn.f32 %f1399, %f1369, %f1398;
|
|
add.f32 %f1400, %f1397, %f1393;
|
|
sub.f32 %f1401, %f1397, %f1400;
|
|
add.f32 %f1402, %f1393, %f1401;
|
|
add.f32 %f1403, %f1395, %f1402;
|
|
add.f32 %f1404, %f1399, %f1403;
|
|
add.f32 %f1405, %f1400, %f1404;
|
|
sub.f32 %f1406, %f1400, %f1405;
|
|
add.f32 %f1407, %f1404, %f1406;
|
|
mov.f32 %f1408, 0f3EE66666;
|
|
mul.rn.f32 %f1409, %f1408, %f1405;
|
|
neg.f32 %f1410, %f1409;
|
|
fma.rn.f32 %f1411, %f1408, %f1405, %f1410;
|
|
fma.rn.f32 %f1412, %f1408, %f1407, %f1411;
|
|
mov.f32 %f1413, 0f00000000;
|
|
fma.rn.f32 %f1414, %f1413, %f1405, %f1412;
|
|
add.rn.f32 %f1415, %f1409, %f1414;
|
|
neg.f32 %f1416, %f1415;
|
|
add.rn.f32 %f1417, %f1409, %f1416;
|
|
add.rn.f32 %f1418, %f1417, %f1414;
|
|
mov.b32 %r1436, %f1415;
|
|
setp.eq.s32 %p157, %r1436, 1118925336;
|
|
add.s32 %r1437, %r1436, -1;
|
|
mov.b32 %f1419, %r1437;
|
|
add.f32 %f1420, %f1418, 0f37000000;
|
|
selp.f32 %f228, %f1420, %f1418, %p157;
|
|
selp.f32 %f1421, %f1419, %f1415, %p157;
|
|
mov.f32 %f1422, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1423, %f1421, %f1422;
|
|
cvt.rzi.f32.f32 %f1424, %f1423;
|
|
abs.f32 %f1425, %f1424;
|
|
setp.gt.f32 %p158, %f1425, 0f42FC0000;
|
|
mov.b32 %r1438, %f1424;
|
|
and.b32 %r1439, %r1438, -2147483648;
|
|
or.b32 %r1440, %r1439, 1123811328;
|
|
mov.b32 %f1426, %r1440;
|
|
selp.f32 %f1427, %f1426, %f1424, %p158;
|
|
mov.f32 %f1428, 0fBF317218;
|
|
fma.rn.f32 %f1429, %f1427, %f1428, %f1421;
|
|
mov.f32 %f1430, 0f3102E308;
|
|
fma.rn.f32 %f1431, %f1427, %f1430, %f1429;
|
|
mul.f32 %f1432, %f1431, 0f3FB8AA3B;
|
|
add.f32 %f1433, %f1427, 0f4B40007F;
|
|
mov.b32 %r1441, %f1433;
|
|
shl.b32 %r1442, %r1441, 23;
|
|
mov.b32 %f1434, %r1442;
|
|
ex2.approx.ftz.f32 %f1435, %f1432;
|
|
mul.f32 %f229, %f1435, %f1434;
|
|
setp.eq.f32 %p159, %f229, 0f7F800000;
|
|
mov.f32 %f2112, 0f7F800000;
|
|
@%p159 bra $L__BB0_95;
|
|
|
|
fma.rn.f32 %f2112, %f229, %f228, %f229;
|
|
|
|
$L__BB0_95:
|
|
setp.lt.f32 %p160, %f184, 0f00000000;
|
|
and.pred %p4, %p160, %p146;
|
|
setp.eq.f32 %p162, %f184, 0f00000000;
|
|
@%p162 bra $L__BB0_99;
|
|
bra.uni $L__BB0_96;
|
|
|
|
$L__BB0_99:
|
|
add.f32 %f1440, %f184, %f184;
|
|
selp.f32 %f2114, %f1440, 0f00000000, %p146;
|
|
bra.uni $L__BB0_100;
|
|
|
|
$L__BB0_96:
|
|
mov.b32 %r1443, %f2112;
|
|
xor.b32 %r1444, %r1443, -2147483648;
|
|
mov.b32 %f1436, %r1444;
|
|
selp.f32 %f2114, %f1436, %f2112, %p4;
|
|
setp.geu.f32 %p163, %f184, 0f00000000;
|
|
@%p163 bra $L__BB0_100;
|
|
|
|
mov.f32 %f1437, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f1438, %f1437;
|
|
setp.eq.f32 %p164, %f1438, 0f3EE66666;
|
|
@%p164 bra $L__BB0_100;
|
|
|
|
mov.f32 %f2114, 0f7FFFFFFF;
|
|
|
|
$L__BB0_100:
|
|
add.f32 %f1441, %f227, 0f3EE66666;
|
|
mov.b32 %r1445, %f1441;
|
|
setp.lt.s32 %p166, %r1445, 2139095040;
|
|
@%p166 bra $L__BB0_105;
|
|
|
|
setp.gtu.f32 %p167, %f227, 0f7F800000;
|
|
@%p167 bra $L__BB0_104;
|
|
bra.uni $L__BB0_102;
|
|
|
|
$L__BB0_104:
|
|
add.f32 %f2114, %f184, 0f3EE66666;
|
|
bra.uni $L__BB0_105;
|
|
|
|
$L__BB0_102:
|
|
setp.neu.f32 %p168, %f227, 0f7F800000;
|
|
@%p168 bra $L__BB0_105;
|
|
|
|
selp.f32 %f2114, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
$L__BB0_105:
|
|
setp.eq.f32 %p169, %f184, 0f3F800000;
|
|
selp.f32 %f238, 0f3F800000, %f2114, %p169;
|
|
abs.f32 %f239, %f185;
|
|
setp.lt.f32 %p170, %f239, 0f00800000;
|
|
mul.f32 %f1443, %f239, 0f4B800000;
|
|
selp.f32 %f1444, %f1443, %f239, %p170;
|
|
selp.f32 %f1445, 0fC3170000, 0fC2FE0000, %p170;
|
|
mov.b32 %r1446, %f1444;
|
|
and.b32 %r1447, %r1446, 8388607;
|
|
or.b32 %r1448, %r1447, 1065353216;
|
|
mov.b32 %f1446, %r1448;
|
|
shr.u32 %r1449, %r1446, 23;
|
|
cvt.rn.f32.u32 %f1447, %r1449;
|
|
add.f32 %f1448, %f1445, %f1447;
|
|
setp.gt.f32 %p171, %f1446, 0f3FB504F3;
|
|
mul.f32 %f1449, %f1446, 0f3F000000;
|
|
add.f32 %f1450, %f1448, 0f3F800000;
|
|
selp.f32 %f1451, %f1450, %f1448, %p171;
|
|
selp.f32 %f1452, %f1449, %f1446, %p171;
|
|
add.f32 %f1453, %f1452, 0fBF800000;
|
|
add.f32 %f1454, %f1452, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1455, %f1454;
|
|
add.f32 %f1456, %f1453, %f1453;
|
|
mul.f32 %f1457, %f1456, %f1455;
|
|
mul.f32 %f1458, %f1457, %f1457;
|
|
mov.f32 %f1459, 0f3C4CAF63;
|
|
mov.f32 %f1460, 0f3B18F0FE;
|
|
fma.rn.f32 %f1461, %f1460, %f1458, %f1459;
|
|
mov.f32 %f1462, 0f3DAAAABD;
|
|
fma.rn.f32 %f1463, %f1461, %f1458, %f1462;
|
|
mul.rn.f32 %f1464, %f1463, %f1458;
|
|
mul.rn.f32 %f1465, %f1464, %f1457;
|
|
sub.f32 %f1466, %f1453, %f1457;
|
|
add.f32 %f1467, %f1466, %f1466;
|
|
neg.f32 %f1468, %f1457;
|
|
fma.rn.f32 %f1469, %f1468, %f1453, %f1467;
|
|
mul.rn.f32 %f1470, %f1455, %f1469;
|
|
add.f32 %f1471, %f1465, %f1457;
|
|
sub.f32 %f1472, %f1457, %f1471;
|
|
add.f32 %f1473, %f1465, %f1472;
|
|
add.f32 %f1474, %f1470, %f1473;
|
|
add.f32 %f1475, %f1471, %f1474;
|
|
sub.f32 %f1476, %f1471, %f1475;
|
|
add.f32 %f1477, %f1474, %f1476;
|
|
mov.f32 %f1478, 0f3F317200;
|
|
mul.rn.f32 %f1479, %f1451, %f1478;
|
|
mov.f32 %f1480, 0f35BFBE8E;
|
|
mul.rn.f32 %f1481, %f1451, %f1480;
|
|
add.f32 %f1482, %f1479, %f1475;
|
|
sub.f32 %f1483, %f1479, %f1482;
|
|
add.f32 %f1484, %f1475, %f1483;
|
|
add.f32 %f1485, %f1477, %f1484;
|
|
add.f32 %f1486, %f1481, %f1485;
|
|
add.f32 %f1487, %f1482, %f1486;
|
|
sub.f32 %f1488, %f1482, %f1487;
|
|
add.f32 %f1489, %f1486, %f1488;
|
|
mov.f32 %f1490, 0f3EE66666;
|
|
mul.rn.f32 %f1491, %f1490, %f1487;
|
|
neg.f32 %f1492, %f1491;
|
|
fma.rn.f32 %f1493, %f1490, %f1487, %f1492;
|
|
fma.rn.f32 %f1494, %f1490, %f1489, %f1493;
|
|
mov.f32 %f1495, 0f00000000;
|
|
fma.rn.f32 %f1496, %f1495, %f1487, %f1494;
|
|
add.rn.f32 %f1497, %f1491, %f1496;
|
|
neg.f32 %f1498, %f1497;
|
|
add.rn.f32 %f1499, %f1491, %f1498;
|
|
add.rn.f32 %f1500, %f1499, %f1496;
|
|
mov.b32 %r1450, %f1497;
|
|
setp.eq.s32 %p172, %r1450, 1118925336;
|
|
add.s32 %r1451, %r1450, -1;
|
|
mov.b32 %f1501, %r1451;
|
|
add.f32 %f1502, %f1500, 0f37000000;
|
|
selp.f32 %f240, %f1502, %f1500, %p172;
|
|
selp.f32 %f1503, %f1501, %f1497, %p172;
|
|
mov.f32 %f1504, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1505, %f1503, %f1504;
|
|
cvt.rzi.f32.f32 %f1506, %f1505;
|
|
abs.f32 %f1507, %f1506;
|
|
setp.gt.f32 %p173, %f1507, 0f42FC0000;
|
|
mov.b32 %r1452, %f1506;
|
|
and.b32 %r1453, %r1452, -2147483648;
|
|
or.b32 %r1454, %r1453, 1123811328;
|
|
mov.b32 %f1508, %r1454;
|
|
selp.f32 %f1509, %f1508, %f1506, %p173;
|
|
mov.f32 %f1510, 0fBF317218;
|
|
fma.rn.f32 %f1511, %f1509, %f1510, %f1503;
|
|
mov.f32 %f1512, 0f3102E308;
|
|
fma.rn.f32 %f1513, %f1509, %f1512, %f1511;
|
|
mul.f32 %f1514, %f1513, 0f3FB8AA3B;
|
|
add.f32 %f1515, %f1509, 0f4B40007F;
|
|
mov.b32 %r1455, %f1515;
|
|
shl.b32 %r1456, %r1455, 23;
|
|
mov.b32 %f1516, %r1456;
|
|
ex2.approx.ftz.f32 %f1517, %f1514;
|
|
mul.f32 %f241, %f1517, %f1516;
|
|
setp.eq.f32 %p174, %f241, 0f7F800000;
|
|
mov.f32 %f2115, 0f7F800000;
|
|
@%p174 bra $L__BB0_107;
|
|
|
|
fma.rn.f32 %f2115, %f241, %f240, %f241;
|
|
|
|
$L__BB0_107:
|
|
setp.lt.f32 %p175, %f185, 0f00000000;
|
|
and.pred %p5, %p175, %p146;
|
|
setp.eq.f32 %p177, %f185, 0f00000000;
|
|
@%p177 bra $L__BB0_111;
|
|
bra.uni $L__BB0_108;
|
|
|
|
$L__BB0_111:
|
|
add.f32 %f1522, %f185, %f185;
|
|
selp.f32 %f2117, %f1522, 0f00000000, %p146;
|
|
bra.uni $L__BB0_112;
|
|
|
|
$L__BB0_108:
|
|
mov.b32 %r1457, %f2115;
|
|
xor.b32 %r1458, %r1457, -2147483648;
|
|
mov.b32 %f1518, %r1458;
|
|
selp.f32 %f2117, %f1518, %f2115, %p5;
|
|
setp.geu.f32 %p178, %f185, 0f00000000;
|
|
@%p178 bra $L__BB0_112;
|
|
|
|
mov.f32 %f1519, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f1520, %f1519;
|
|
setp.eq.f32 %p179, %f1520, 0f3EE66666;
|
|
@%p179 bra $L__BB0_112;
|
|
|
|
mov.f32 %f2117, 0f7FFFFFFF;
|
|
|
|
$L__BB0_112:
|
|
add.f32 %f1523, %f239, 0f3EE66666;
|
|
mov.b32 %r1459, %f1523;
|
|
setp.lt.s32 %p181, %r1459, 2139095040;
|
|
@%p181 bra $L__BB0_117;
|
|
|
|
setp.gtu.f32 %p182, %f239, 0f7F800000;
|
|
@%p182 bra $L__BB0_116;
|
|
bra.uni $L__BB0_114;
|
|
|
|
$L__BB0_116:
|
|
add.f32 %f2117, %f185, 0f3EE66666;
|
|
bra.uni $L__BB0_117;
|
|
|
|
$L__BB0_114:
|
|
setp.neu.f32 %p183, %f239, 0f7F800000;
|
|
@%p183 bra $L__BB0_117;
|
|
|
|
selp.f32 %f2117, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
$L__BB0_117:
|
|
setp.eq.f32 %p184, %f185, 0f3F800000;
|
|
mov.f32 %f1525, 0f3F800000;
|
|
selp.f32 %f1526, 0f3F800000, %f2117, %p184;
|
|
ld.const.u64 %rd88, [params+144];
|
|
cvta.to.global.u64 %rd17, %rd88;
|
|
ld.const.u32 %r1460, [params+136];
|
|
mad.lo.s32 %r1461, %r1460, %r6, %r5;
|
|
cvt.u64.u32 %rd18, %r1461;
|
|
min.f32 %f1527, %f226, %f1525;
|
|
mov.f32 %f1528, 0f00000000;
|
|
max.f32 %f250, %f1528, %f1527;
|
|
min.f32 %f1529, %f238, %f1525;
|
|
max.f32 %f251, %f1528, %f1529;
|
|
min.f32 %f1530, %f1526, %f1525;
|
|
max.f32 %f252, %f1528, %f1530;
|
|
mov.f32 %f1531, 0f3E555555;
|
|
cvt.rzi.f32.f32 %f1532, %f1531;
|
|
add.f32 %f1533, %f1532, %f1532;
|
|
mov.f32 %f1534, 0f3ED55555;
|
|
sub.f32 %f1535, %f1534, %f1533;
|
|
abs.f32 %f253, %f1535;
|
|
abs.f32 %f254, %f250;
|
|
setp.lt.f32 %p185, %f254, 0f00800000;
|
|
mul.f32 %f1536, %f254, 0f4B800000;
|
|
selp.f32 %f1537, %f1536, %f254, %p185;
|
|
selp.f32 %f1538, 0fC3170000, 0fC2FE0000, %p185;
|
|
mov.b32 %r1462, %f1537;
|
|
and.b32 %r1463, %r1462, 8388607;
|
|
or.b32 %r1464, %r1463, 1065353216;
|
|
mov.b32 %f1539, %r1464;
|
|
shr.u32 %r1465, %r1462, 23;
|
|
cvt.rn.f32.u32 %f1540, %r1465;
|
|
add.f32 %f1541, %f1538, %f1540;
|
|
setp.gt.f32 %p186, %f1539, 0f3FB504F3;
|
|
mul.f32 %f1542, %f1539, 0f3F000000;
|
|
add.f32 %f1543, %f1541, 0f3F800000;
|
|
selp.f32 %f1544, %f1543, %f1541, %p186;
|
|
selp.f32 %f1545, %f1542, %f1539, %p186;
|
|
add.f32 %f1546, %f1545, 0fBF800000;
|
|
add.f32 %f1547, %f1545, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1548, %f1547;
|
|
add.f32 %f1549, %f1546, %f1546;
|
|
mul.f32 %f1550, %f1549, %f1548;
|
|
mul.f32 %f1551, %f1550, %f1550;
|
|
mov.f32 %f1552, 0f3C4CAF63;
|
|
mov.f32 %f1553, 0f3B18F0FE;
|
|
fma.rn.f32 %f1554, %f1553, %f1551, %f1552;
|
|
mov.f32 %f1555, 0f3DAAAABD;
|
|
fma.rn.f32 %f1556, %f1554, %f1551, %f1555;
|
|
mul.rn.f32 %f1557, %f1556, %f1551;
|
|
mul.rn.f32 %f1558, %f1557, %f1550;
|
|
sub.f32 %f1559, %f1546, %f1550;
|
|
add.f32 %f1560, %f1559, %f1559;
|
|
neg.f32 %f1561, %f1550;
|
|
fma.rn.f32 %f1562, %f1561, %f1546, %f1560;
|
|
mul.rn.f32 %f1563, %f1548, %f1562;
|
|
add.f32 %f1564, %f1558, %f1550;
|
|
sub.f32 %f1565, %f1550, %f1564;
|
|
add.f32 %f1566, %f1558, %f1565;
|
|
add.f32 %f1567, %f1563, %f1566;
|
|
add.f32 %f1568, %f1564, %f1567;
|
|
sub.f32 %f1569, %f1564, %f1568;
|
|
add.f32 %f1570, %f1567, %f1569;
|
|
mov.f32 %f1571, 0f3F317200;
|
|
mul.rn.f32 %f1572, %f1544, %f1571;
|
|
mov.f32 %f1573, 0f35BFBE8E;
|
|
mul.rn.f32 %f1574, %f1544, %f1573;
|
|
add.f32 %f1575, %f1572, %f1568;
|
|
sub.f32 %f1576, %f1572, %f1575;
|
|
add.f32 %f1577, %f1568, %f1576;
|
|
add.f32 %f1578, %f1570, %f1577;
|
|
add.f32 %f1579, %f1574, %f1578;
|
|
add.f32 %f1580, %f1575, %f1579;
|
|
sub.f32 %f1581, %f1575, %f1580;
|
|
add.f32 %f1582, %f1579, %f1581;
|
|
mul.rn.f32 %f1583, %f1534, %f1580;
|
|
neg.f32 %f1584, %f1583;
|
|
fma.rn.f32 %f1585, %f1534, %f1580, %f1584;
|
|
fma.rn.f32 %f1586, %f1534, %f1582, %f1585;
|
|
fma.rn.f32 %f1587, %f1528, %f1580, %f1586;
|
|
add.rn.f32 %f1588, %f1583, %f1587;
|
|
neg.f32 %f1589, %f1588;
|
|
add.rn.f32 %f1590, %f1583, %f1589;
|
|
add.rn.f32 %f1591, %f1590, %f1587;
|
|
mov.b32 %r1466, %f1588;
|
|
setp.eq.s32 %p187, %r1466, 1118925336;
|
|
add.s32 %r1467, %r1466, -1;
|
|
mov.b32 %f1592, %r1467;
|
|
add.f32 %f1593, %f1591, 0f37000000;
|
|
selp.f32 %f255, %f1593, %f1591, %p187;
|
|
selp.f32 %f1594, %f1592, %f1588, %p187;
|
|
mov.f32 %f1595, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1596, %f1594, %f1595;
|
|
cvt.rzi.f32.f32 %f1597, %f1596;
|
|
abs.f32 %f1598, %f1597;
|
|
setp.gt.f32 %p188, %f1598, 0f42FC0000;
|
|
mov.b32 %r1468, %f1597;
|
|
and.b32 %r1469, %r1468, -2147483648;
|
|
or.b32 %r1470, %r1469, 1123811328;
|
|
mov.b32 %f1599, %r1470;
|
|
selp.f32 %f1600, %f1599, %f1597, %p188;
|
|
mov.f32 %f1601, 0fBF317218;
|
|
fma.rn.f32 %f1602, %f1600, %f1601, %f1594;
|
|
mov.f32 %f1603, 0f3102E308;
|
|
fma.rn.f32 %f1604, %f1600, %f1603, %f1602;
|
|
mul.f32 %f1605, %f1604, 0f3FB8AA3B;
|
|
add.f32 %f1606, %f1600, 0f4B40007F;
|
|
mov.b32 %r1471, %f1606;
|
|
shl.b32 %r1472, %r1471, 23;
|
|
mov.b32 %f1607, %r1472;
|
|
ex2.approx.ftz.f32 %f1608, %f1605;
|
|
mul.f32 %f256, %f1608, %f1607;
|
|
setp.eq.f32 %p189, %f256, 0f7F800000;
|
|
mov.f32 %f2118, 0f7F800000;
|
|
@%p189 bra $L__BB0_119;
|
|
|
|
fma.rn.f32 %f2118, %f256, %f255, %f256;
|
|
|
|
$L__BB0_119:
|
|
setp.lt.f32 %p190, %f250, 0f00000000;
|
|
setp.eq.f32 %p191, %f253, 0f3F800000;
|
|
and.pred %p6, %p190, %p191;
|
|
setp.eq.f32 %p192, %f250, 0f00000000;
|
|
@%p192 bra $L__BB0_123;
|
|
bra.uni $L__BB0_120;
|
|
|
|
$L__BB0_123:
|
|
add.f32 %f1613, %f250, %f250;
|
|
selp.f32 %f2120, %f1613, 0f00000000, %p191;
|
|
bra.uni $L__BB0_124;
|
|
|
|
$L__BB0_120:
|
|
mov.b32 %r1473, %f2118;
|
|
xor.b32 %r1474, %r1473, -2147483648;
|
|
mov.b32 %f1609, %r1474;
|
|
selp.f32 %f2120, %f1609, %f2118, %p6;
|
|
setp.geu.f32 %p193, %f250, 0f00000000;
|
|
@%p193 bra $L__BB0_124;
|
|
|
|
mov.f32 %f1610, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1611, %f1610;
|
|
setp.eq.f32 %p194, %f1611, 0f3ED55555;
|
|
@%p194 bra $L__BB0_124;
|
|
|
|
mov.f32 %f2120, 0f7FFFFFFF;
|
|
|
|
$L__BB0_124:
|
|
add.f32 %f1614, %f254, 0f3ED55555;
|
|
mov.b32 %r1475, %f1614;
|
|
setp.lt.s32 %p196, %r1475, 2139095040;
|
|
@%p196 bra $L__BB0_129;
|
|
|
|
setp.gtu.f32 %p197, %f254, 0f7F800000;
|
|
@%p197 bra $L__BB0_128;
|
|
bra.uni $L__BB0_126;
|
|
|
|
$L__BB0_128:
|
|
add.f32 %f2120, %f250, 0f3ED55555;
|
|
bra.uni $L__BB0_129;
|
|
|
|
$L__BB0_126:
|
|
setp.neu.f32 %p198, %f254, 0f7F800000;
|
|
@%p198 bra $L__BB0_129;
|
|
|
|
selp.f32 %f2120, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
$L__BB0_129:
|
|
abs.f32 %f265, %f251;
|
|
setp.lt.f32 %p199, %f265, 0f00800000;
|
|
mul.f32 %f1616, %f265, 0f4B800000;
|
|
selp.f32 %f1617, %f1616, %f265, %p199;
|
|
selp.f32 %f1618, 0fC3170000, 0fC2FE0000, %p199;
|
|
mov.b32 %r1476, %f1617;
|
|
and.b32 %r1477, %r1476, 8388607;
|
|
or.b32 %r1478, %r1477, 1065353216;
|
|
mov.b32 %f1619, %r1478;
|
|
shr.u32 %r1479, %r1476, 23;
|
|
cvt.rn.f32.u32 %f1620, %r1479;
|
|
add.f32 %f1621, %f1618, %f1620;
|
|
setp.gt.f32 %p200, %f1619, 0f3FB504F3;
|
|
mul.f32 %f1622, %f1619, 0f3F000000;
|
|
add.f32 %f1623, %f1621, 0f3F800000;
|
|
selp.f32 %f1624, %f1623, %f1621, %p200;
|
|
selp.f32 %f1625, %f1622, %f1619, %p200;
|
|
add.f32 %f1626, %f1625, 0fBF800000;
|
|
add.f32 %f1627, %f1625, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1628, %f1627;
|
|
add.f32 %f1629, %f1626, %f1626;
|
|
mul.f32 %f1630, %f1629, %f1628;
|
|
mul.f32 %f1631, %f1630, %f1630;
|
|
mov.f32 %f1632, 0f3C4CAF63;
|
|
mov.f32 %f1633, 0f3B18F0FE;
|
|
fma.rn.f32 %f1634, %f1633, %f1631, %f1632;
|
|
mov.f32 %f1635, 0f3DAAAABD;
|
|
fma.rn.f32 %f1636, %f1634, %f1631, %f1635;
|
|
mul.rn.f32 %f1637, %f1636, %f1631;
|
|
mul.rn.f32 %f1638, %f1637, %f1630;
|
|
sub.f32 %f1639, %f1626, %f1630;
|
|
add.f32 %f1640, %f1639, %f1639;
|
|
neg.f32 %f1641, %f1630;
|
|
fma.rn.f32 %f1642, %f1641, %f1626, %f1640;
|
|
mul.rn.f32 %f1643, %f1628, %f1642;
|
|
add.f32 %f1644, %f1638, %f1630;
|
|
sub.f32 %f1645, %f1630, %f1644;
|
|
add.f32 %f1646, %f1638, %f1645;
|
|
add.f32 %f1647, %f1643, %f1646;
|
|
add.f32 %f1648, %f1644, %f1647;
|
|
sub.f32 %f1649, %f1644, %f1648;
|
|
add.f32 %f1650, %f1647, %f1649;
|
|
mov.f32 %f1651, 0f3F317200;
|
|
mul.rn.f32 %f1652, %f1624, %f1651;
|
|
mov.f32 %f1653, 0f35BFBE8E;
|
|
mul.rn.f32 %f1654, %f1624, %f1653;
|
|
add.f32 %f1655, %f1652, %f1648;
|
|
sub.f32 %f1656, %f1652, %f1655;
|
|
add.f32 %f1657, %f1648, %f1656;
|
|
add.f32 %f1658, %f1650, %f1657;
|
|
add.f32 %f1659, %f1654, %f1658;
|
|
add.f32 %f1660, %f1655, %f1659;
|
|
sub.f32 %f1661, %f1655, %f1660;
|
|
add.f32 %f1662, %f1659, %f1661;
|
|
mov.f32 %f1663, 0f3ED55555;
|
|
mul.rn.f32 %f1664, %f1663, %f1660;
|
|
neg.f32 %f1665, %f1664;
|
|
fma.rn.f32 %f1666, %f1663, %f1660, %f1665;
|
|
fma.rn.f32 %f1667, %f1663, %f1662, %f1666;
|
|
mov.f32 %f1668, 0f00000000;
|
|
fma.rn.f32 %f1669, %f1668, %f1660, %f1667;
|
|
add.rn.f32 %f1670, %f1664, %f1669;
|
|
neg.f32 %f1671, %f1670;
|
|
add.rn.f32 %f1672, %f1664, %f1671;
|
|
add.rn.f32 %f1673, %f1672, %f1669;
|
|
mov.b32 %r1480, %f1670;
|
|
setp.eq.s32 %p201, %r1480, 1118925336;
|
|
add.s32 %r1481, %r1480, -1;
|
|
mov.b32 %f1674, %r1481;
|
|
add.f32 %f1675, %f1673, 0f37000000;
|
|
selp.f32 %f266, %f1675, %f1673, %p201;
|
|
selp.f32 %f1676, %f1674, %f1670, %p201;
|
|
mov.f32 %f1677, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1678, %f1676, %f1677;
|
|
cvt.rzi.f32.f32 %f1679, %f1678;
|
|
abs.f32 %f1680, %f1679;
|
|
setp.gt.f32 %p202, %f1680, 0f42FC0000;
|
|
mov.b32 %r1482, %f1679;
|
|
and.b32 %r1483, %r1482, -2147483648;
|
|
or.b32 %r1484, %r1483, 1123811328;
|
|
mov.b32 %f1681, %r1484;
|
|
selp.f32 %f1682, %f1681, %f1679, %p202;
|
|
mov.f32 %f1683, 0fBF317218;
|
|
fma.rn.f32 %f1684, %f1682, %f1683, %f1676;
|
|
mov.f32 %f1685, 0f3102E308;
|
|
fma.rn.f32 %f1686, %f1682, %f1685, %f1684;
|
|
mul.f32 %f1687, %f1686, 0f3FB8AA3B;
|
|
add.f32 %f1688, %f1682, 0f4B40007F;
|
|
mov.b32 %r1485, %f1688;
|
|
shl.b32 %r1486, %r1485, 23;
|
|
mov.b32 %f1689, %r1486;
|
|
ex2.approx.ftz.f32 %f1690, %f1687;
|
|
mul.f32 %f267, %f1690, %f1689;
|
|
setp.eq.f32 %p203, %f267, 0f7F800000;
|
|
mov.f32 %f2121, 0f7F800000;
|
|
@%p203 bra $L__BB0_131;
|
|
|
|
fma.rn.f32 %f2121, %f267, %f266, %f267;
|
|
|
|
$L__BB0_131:
|
|
setp.lt.f32 %p204, %f251, 0f00000000;
|
|
and.pred %p7, %p204, %p191;
|
|
setp.eq.f32 %p206, %f251, 0f00000000;
|
|
@%p206 bra $L__BB0_135;
|
|
bra.uni $L__BB0_132;
|
|
|
|
$L__BB0_135:
|
|
add.f32 %f1695, %f251, %f251;
|
|
selp.f32 %f2123, %f1695, 0f00000000, %p191;
|
|
bra.uni $L__BB0_136;
|
|
|
|
$L__BB0_132:
|
|
mov.b32 %r1487, %f2121;
|
|
xor.b32 %r1488, %r1487, -2147483648;
|
|
mov.b32 %f1691, %r1488;
|
|
selp.f32 %f2123, %f1691, %f2121, %p7;
|
|
setp.geu.f32 %p207, %f251, 0f00000000;
|
|
@%p207 bra $L__BB0_136;
|
|
|
|
mov.f32 %f1692, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1693, %f1692;
|
|
setp.eq.f32 %p208, %f1693, 0f3ED55555;
|
|
@%p208 bra $L__BB0_136;
|
|
|
|
mov.f32 %f2123, 0f7FFFFFFF;
|
|
|
|
$L__BB0_136:
|
|
add.f32 %f1696, %f265, 0f3ED55555;
|
|
mov.b32 %r1489, %f1696;
|
|
setp.lt.s32 %p210, %r1489, 2139095040;
|
|
@%p210 bra $L__BB0_141;
|
|
|
|
setp.gtu.f32 %p211, %f265, 0f7F800000;
|
|
@%p211 bra $L__BB0_140;
|
|
bra.uni $L__BB0_138;
|
|
|
|
$L__BB0_140:
|
|
add.f32 %f2123, %f251, 0f3ED55555;
|
|
bra.uni $L__BB0_141;
|
|
|
|
$L__BB0_138:
|
|
setp.neu.f32 %p212, %f265, 0f7F800000;
|
|
@%p212 bra $L__BB0_141;
|
|
|
|
selp.f32 %f2123, 0fFF800000, 0f7F800000, %p7;
|
|
|
|
$L__BB0_141:
|
|
abs.f32 %f276, %f252;
|
|
setp.lt.f32 %p213, %f276, 0f00800000;
|
|
mul.f32 %f1698, %f276, 0f4B800000;
|
|
selp.f32 %f1699, %f1698, %f276, %p213;
|
|
selp.f32 %f1700, 0fC3170000, 0fC2FE0000, %p213;
|
|
mov.b32 %r1490, %f1699;
|
|
and.b32 %r1491, %r1490, 8388607;
|
|
or.b32 %r1492, %r1491, 1065353216;
|
|
mov.b32 %f1701, %r1492;
|
|
shr.u32 %r1493, %r1490, 23;
|
|
cvt.rn.f32.u32 %f1702, %r1493;
|
|
add.f32 %f1703, %f1700, %f1702;
|
|
setp.gt.f32 %p214, %f1701, 0f3FB504F3;
|
|
mul.f32 %f1704, %f1701, 0f3F000000;
|
|
add.f32 %f1705, %f1703, 0f3F800000;
|
|
selp.f32 %f1706, %f1705, %f1703, %p214;
|
|
selp.f32 %f1707, %f1704, %f1701, %p214;
|
|
add.f32 %f1708, %f1707, 0fBF800000;
|
|
add.f32 %f1709, %f1707, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1710, %f1709;
|
|
add.f32 %f1711, %f1708, %f1708;
|
|
mul.f32 %f1712, %f1711, %f1710;
|
|
mul.f32 %f1713, %f1712, %f1712;
|
|
mov.f32 %f1714, 0f3C4CAF63;
|
|
mov.f32 %f1715, 0f3B18F0FE;
|
|
fma.rn.f32 %f1716, %f1715, %f1713, %f1714;
|
|
mov.f32 %f1717, 0f3DAAAABD;
|
|
fma.rn.f32 %f1718, %f1716, %f1713, %f1717;
|
|
mul.rn.f32 %f1719, %f1718, %f1713;
|
|
mul.rn.f32 %f1720, %f1719, %f1712;
|
|
sub.f32 %f1721, %f1708, %f1712;
|
|
add.f32 %f1722, %f1721, %f1721;
|
|
neg.f32 %f1723, %f1712;
|
|
fma.rn.f32 %f1724, %f1723, %f1708, %f1722;
|
|
mul.rn.f32 %f1725, %f1710, %f1724;
|
|
add.f32 %f1726, %f1720, %f1712;
|
|
sub.f32 %f1727, %f1712, %f1726;
|
|
add.f32 %f1728, %f1720, %f1727;
|
|
add.f32 %f1729, %f1725, %f1728;
|
|
add.f32 %f1730, %f1726, %f1729;
|
|
sub.f32 %f1731, %f1726, %f1730;
|
|
add.f32 %f1732, %f1729, %f1731;
|
|
mov.f32 %f1733, 0f3F317200;
|
|
mul.rn.f32 %f1734, %f1706, %f1733;
|
|
mov.f32 %f1735, 0f35BFBE8E;
|
|
mul.rn.f32 %f1736, %f1706, %f1735;
|
|
add.f32 %f1737, %f1734, %f1730;
|
|
sub.f32 %f1738, %f1734, %f1737;
|
|
add.f32 %f1739, %f1730, %f1738;
|
|
add.f32 %f1740, %f1732, %f1739;
|
|
add.f32 %f1741, %f1736, %f1740;
|
|
add.f32 %f1742, %f1737, %f1741;
|
|
sub.f32 %f1743, %f1737, %f1742;
|
|
add.f32 %f1744, %f1741, %f1743;
|
|
mov.f32 %f1745, 0f3ED55555;
|
|
mul.rn.f32 %f1746, %f1745, %f1742;
|
|
neg.f32 %f1747, %f1746;
|
|
fma.rn.f32 %f1748, %f1745, %f1742, %f1747;
|
|
fma.rn.f32 %f1749, %f1745, %f1744, %f1748;
|
|
mov.f32 %f1750, 0f00000000;
|
|
fma.rn.f32 %f1751, %f1750, %f1742, %f1749;
|
|
add.rn.f32 %f1752, %f1746, %f1751;
|
|
neg.f32 %f1753, %f1752;
|
|
add.rn.f32 %f1754, %f1746, %f1753;
|
|
add.rn.f32 %f1755, %f1754, %f1751;
|
|
mov.b32 %r1494, %f1752;
|
|
setp.eq.s32 %p215, %r1494, 1118925336;
|
|
add.s32 %r1495, %r1494, -1;
|
|
mov.b32 %f1756, %r1495;
|
|
add.f32 %f1757, %f1755, 0f37000000;
|
|
selp.f32 %f277, %f1757, %f1755, %p215;
|
|
selp.f32 %f1758, %f1756, %f1752, %p215;
|
|
mov.f32 %f1759, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1760, %f1758, %f1759;
|
|
cvt.rzi.f32.f32 %f1761, %f1760;
|
|
abs.f32 %f1762, %f1761;
|
|
setp.gt.f32 %p216, %f1762, 0f42FC0000;
|
|
mov.b32 %r1496, %f1761;
|
|
and.b32 %r1497, %r1496, -2147483648;
|
|
or.b32 %r1498, %r1497, 1123811328;
|
|
mov.b32 %f1763, %r1498;
|
|
selp.f32 %f1764, %f1763, %f1761, %p216;
|
|
mov.f32 %f1765, 0fBF317218;
|
|
fma.rn.f32 %f1766, %f1764, %f1765, %f1758;
|
|
mov.f32 %f1767, 0f3102E308;
|
|
fma.rn.f32 %f1768, %f1764, %f1767, %f1766;
|
|
mul.f32 %f1769, %f1768, 0f3FB8AA3B;
|
|
add.f32 %f1770, %f1764, 0f4B40007F;
|
|
mov.b32 %r1499, %f1770;
|
|
shl.b32 %r1500, %r1499, 23;
|
|
mov.b32 %f1771, %r1500;
|
|
ex2.approx.ftz.f32 %f1772, %f1769;
|
|
mul.f32 %f278, %f1772, %f1771;
|
|
setp.eq.f32 %p217, %f278, 0f7F800000;
|
|
mov.f32 %f2124, 0f7F800000;
|
|
@%p217 bra $L__BB0_143;
|
|
|
|
fma.rn.f32 %f2124, %f278, %f277, %f278;
|
|
|
|
$L__BB0_143:
|
|
setp.lt.f32 %p218, %f252, 0f00000000;
|
|
and.pred %p8, %p218, %p191;
|
|
setp.eq.f32 %p220, %f252, 0f00000000;
|
|
@%p220 bra $L__BB0_147;
|
|
bra.uni $L__BB0_144;
|
|
|
|
$L__BB0_147:
|
|
add.f32 %f1777, %f252, %f252;
|
|
selp.f32 %f2126, %f1777, 0f00000000, %p191;
|
|
bra.uni $L__BB0_148;
|
|
|
|
$L__BB0_144:
|
|
mov.b32 %r1501, %f2124;
|
|
xor.b32 %r1502, %r1501, -2147483648;
|
|
mov.b32 %f1773, %r1502;
|
|
selp.f32 %f2126, %f1773, %f2124, %p8;
|
|
setp.geu.f32 %p221, %f252, 0f00000000;
|
|
@%p221 bra $L__BB0_148;
|
|
|
|
mov.f32 %f1774, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1775, %f1774;
|
|
setp.eq.f32 %p222, %f1775, 0f3ED55555;
|
|
@%p222 bra $L__BB0_148;
|
|
|
|
mov.f32 %f2126, 0f7FFFFFFF;
|
|
|
|
$L__BB0_148:
|
|
add.f32 %f1778, %f276, 0f3ED55555;
|
|
mov.b32 %r1503, %f1778;
|
|
setp.lt.s32 %p224, %r1503, 2139095040;
|
|
@%p224 bra $L__BB0_153;
|
|
|
|
setp.gtu.f32 %p225, %f276, 0f7F800000;
|
|
@%p225 bra $L__BB0_152;
|
|
bra.uni $L__BB0_150;
|
|
|
|
$L__BB0_152:
|
|
add.f32 %f2126, %f252, 0f3ED55555;
|
|
bra.uni $L__BB0_153;
|
|
|
|
$L__BB0_150:
|
|
setp.neu.f32 %p226, %f276, 0f7F800000;
|
|
@%p226 bra $L__BB0_153;
|
|
|
|
selp.f32 %f2126, 0fFF800000, 0f7F800000, %p8;
|
|
|
|
$L__BB0_153:
|
|
fma.rn.f32 %f1779, %f2120, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p227, %f250, 0f3F800000;
|
|
mov.f32 %f1780, 0f3F800000;
|
|
selp.f32 %f1781, 0f3F7FFFFF, %f1779, %p227;
|
|
mul.f32 %f1782, %f250, 0f414EB852;
|
|
setp.lt.f32 %p228, %f250, 0f3B4D2E1C;
|
|
selp.f32 %f1783, %f1782, %f1781, %p228;
|
|
fma.rn.f32 %f1784, %f2123, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p229, %f251, 0f3F800000;
|
|
selp.f32 %f1785, 0f3F7FFFFF, %f1784, %p229;
|
|
mul.f32 %f1786, %f251, 0f414EB852;
|
|
setp.lt.f32 %p230, %f251, 0f3B4D2E1C;
|
|
selp.f32 %f1787, %f1786, %f1785, %p230;
|
|
fma.rn.f32 %f1788, %f2126, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p231, %f252, 0f3F800000;
|
|
selp.f32 %f1789, 0f3F7FFFFF, %f1788, %p231;
|
|
mul.f32 %f1790, %f252, 0f414EB852;
|
|
setp.lt.f32 %p232, %f252, 0f3B4D2E1C;
|
|
selp.f32 %f1791, %f1790, %f1789, %p232;
|
|
min.f32 %f1792, %f1783, %f1780;
|
|
mov.f32 %f1793, 0f00000000;
|
|
max.f32 %f1794, %f1793, %f1792;
|
|
mul.f32 %f1795, %f1794, 0f43800000;
|
|
cvt.rzi.u32.f32 %r1504, %f1795;
|
|
min.u32 %r1505, %r1504, 255;
|
|
min.f32 %f1796, %f1787, %f1780;
|
|
max.f32 %f1797, %f1793, %f1796;
|
|
mul.f32 %f1798, %f1797, 0f43800000;
|
|
cvt.rzi.u32.f32 %r1506, %f1798;
|
|
min.u32 %r1507, %r1506, 255;
|
|
min.f32 %f1799, %f1791, %f1780;
|
|
max.f32 %f1800, %f1793, %f1799;
|
|
mul.f32 %f1801, %f1800, 0f43800000;
|
|
cvt.rzi.u32.f32 %r1508, %f1801;
|
|
min.u32 %r1509, %r1508, 255;
|
|
shl.b64 %rd89, %rd18, 2;
|
|
add.s64 %rd90, %rd17, %rd89;
|
|
cvt.u16.u32 %rs45, %r1509;
|
|
cvt.u16.u32 %rs46, %r1507;
|
|
cvt.u16.u32 %rs47, %r1505;
|
|
mov.u16 %rs48, 255;
|
|
st.global.v4.u8 [%rd90], {%rs47, %rs46, %rs45, %rs48};
|
|
|
|
$L__BB0_154:
|
|
and.b32 %r1510, %r58, 4;
|
|
setp.eq.s32 %p233, %r1510, 0;
|
|
@%p233 bra $L__BB0_158;
|
|
|
|
ld.const.u32 %r1511, [params+108];
|
|
setp.eq.s32 %p234, %r1511, 0;
|
|
ld.const.u64 %rd91, [params+224];
|
|
cvta.to.global.u64 %rd92, %rd91;
|
|
ld.const.u32 %r1512, [params+216];
|
|
mad.lo.s32 %r1513, %r1512, %r6, %r5;
|
|
mul.wide.u32 %rd93, %r1513, 8;
|
|
add.s64 %rd19, %rd92, %rd93;
|
|
@%p234 bra $L__BB0_157;
|
|
|
|
ld.global.v4.u16 {%rs56, %rs57, %rs58, %rs59}, [%rd19];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1802, %rs56;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1803, %rs57;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1804, %rs58;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1805, %f183, %f1802;
|
|
add.f32 %f1806, %f184, %f1803;
|
|
add.f32 %f1807, %f185, %f1804;
|
|
mov.f32 %f1808, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f1807;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f1806;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f1805;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1808;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd19], {%rs52, %rs53, %rs54, %rs55};
|
|
bra.uni $L__BB0_158;
|
|
|
|
$L__BB0_157:
|
|
mov.f32 %f1812, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f1812;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f185;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f184;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f183;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd19], {%rs60, %rs61, %rs62, %rs63};
|
|
|
|
$L__BB0_158:
|
|
ld.const.u64 %rd94, [params+464];
|
|
cvta.to.global.u64 %rd95, %rd94;
|
|
ld.const.u32 %r1514, [params+456];
|
|
mad.lo.s32 %r1515, %r1514, %r64, %r63;
|
|
mul.wide.u32 %rd96, %r1515, 4;
|
|
add.s64 %rd20, %rd95, %rd96;
|
|
ld.global.v2.u8 {%rs64, %rs204}, [%rd20];
|
|
or.b16 %rs66, %rs64, %rs204;
|
|
and.b16 %rs67, %rs66, 255;
|
|
setp.eq.s16 %p235, %rs67, 0;
|
|
@%p235 bra $L__BB0_160;
|
|
|
|
ld.global.u8 %rs202, [%rd20+2];
|
|
bra.uni $L__BB0_161;
|
|
|
|
$L__BB0_160:
|
|
ld.global.u8 %rs202, [%rd20+2];
|
|
setp.eq.s16 %p236, %rs202, 0;
|
|
mov.f32 %f2127, 0f00000000;
|
|
mov.u16 %rs204, 0;
|
|
mov.u16 %rs205, %rs204;
|
|
mov.f32 %f2128, %f2127;
|
|
mov.f32 %f2129, %f2127;
|
|
@%p236 bra $L__BB0_162;
|
|
|
|
$L__BB0_161:
|
|
mov.u16 %rs205, %rs204;
|
|
cvt.rn.f32.u16 %f1816, %rs64;
|
|
div.rn.f32 %f1817, %f1816, 0f437F0000;
|
|
fma.rn.f32 %f1818, %f1817, 0f40000000, 0fBF800000;
|
|
and.b16 %rs72, %rs205, 255;
|
|
cvt.rn.f32.u16 %f1819, %rs72;
|
|
div.rn.f32 %f1820, %f1819, 0f437F0000;
|
|
fma.rn.f32 %f1821, %f1820, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f1822, %rs202;
|
|
div.rn.f32 %f1823, %f1822, 0f437F0000;
|
|
fma.rn.f32 %f1824, %f1823, 0f40000000, 0fBF800000;
|
|
mul.f32 %f1825, %f1821, %f1821;
|
|
fma.rn.f32 %f1826, %f1818, %f1818, %f1825;
|
|
fma.rn.f32 %f1827, %f1824, %f1824, %f1826;
|
|
sqrt.rn.f32 %f1828, %f1827;
|
|
rcp.rn.f32 %f1829, %f1828;
|
|
mul.f32 %f2129, %f1829, %f1824;
|
|
mul.f32 %f2128, %f1829, %f1821;
|
|
mul.f32 %f2127, %f1818, %f1829;
|
|
mov.u16 %rs204, %rs202;
|
|
|
|
$L__BB0_162:
|
|
mul.f32 %f1833, %f2044, %f2129;
|
|
mul.f32 %f1834, %f2045, %f2128;
|
|
sub.f32 %f1835, %f1833, %f1834;
|
|
mul.f32 %f1836, %f2045, %f2127;
|
|
mul.f32 %f1837, %f2043, %f2129;
|
|
sub.f32 %f1838, %f1836, %f1837;
|
|
mul.f32 %f1839, %f2043, %f2128;
|
|
mul.f32 %f1840, %f2044, %f2127;
|
|
sub.f32 %f1841, %f1839, %f1840;
|
|
ld.global.u8 %rs74, [%rd20+3];
|
|
setp.eq.s16 %p237, %rs74, 0;
|
|
selp.f32 %f293, 0fBF800000, 0f3F800000, %p237;
|
|
mul.f32 %f1842, %f1835, %f293;
|
|
mul.f32 %f1843, %f1838, %f293;
|
|
mul.f32 %f1844, %f1841, %f293;
|
|
mul.f32 %f1845, %f1842, 0f00000000;
|
|
mul.f32 %f1846, %f1843, 0f00000000;
|
|
mul.f32 %f1847, %f1844, 0f00000000;
|
|
fma.rn.f32 %f1848, %f2127, 0f3F5105EC, %f1845;
|
|
fma.rn.f32 %f1849, %f2128, 0f3F5105EC, %f1846;
|
|
fma.rn.f32 %f1850, %f2129, 0f3F5105EC, %f1847;
|
|
mul.f32 %f294, %f2043, 0f3F13CD3A;
|
|
add.f32 %f295, %f294, %f1848;
|
|
mul.f32 %f296, %f2044, 0f3F13CD3A;
|
|
add.f32 %f297, %f296, %f1849;
|
|
mul.f32 %f298, %f2045, 0f3F13CD3A;
|
|
add.f32 %f299, %f298, %f1850;
|
|
or.b16 %rs75, %rs64, %rs205;
|
|
or.b16 %rs23, %rs75, %rs204;
|
|
and.b16 %rs76, %rs23, 255;
|
|
setp.eq.s16 %p238, %rs76, 0;
|
|
mov.f32 %f2133, 0f00000000;
|
|
mov.f32 %f2130, %f2133;
|
|
mov.f32 %f2131, %f2133;
|
|
mov.f32 %f2132, %f2133;
|
|
@%p238 bra $L__BB0_164;
|
|
|
|
cvt.rn.f32.u16 %f1851, %rs64;
|
|
div.rn.f32 %f1852, %f1851, 0f437F0000;
|
|
fma.rn.f32 %f1853, %f1852, 0f40000000, 0fBF800000;
|
|
and.b16 %rs78, %rs205, 255;
|
|
cvt.rn.f32.u16 %f1854, %rs78;
|
|
div.rn.f32 %f1855, %f1854, 0f437F0000;
|
|
fma.rn.f32 %f1856, %f1855, 0f40000000, 0fBF800000;
|
|
and.b16 %rs79, %rs204, 255;
|
|
cvt.rn.f32.u16 %f1857, %rs79;
|
|
div.rn.f32 %f1858, %f1857, 0f437F0000;
|
|
fma.rn.f32 %f1859, %f1858, 0f40000000, 0fBF800000;
|
|
mul.f32 %f1860, %f1856, %f1856;
|
|
fma.rn.f32 %f1861, %f1853, %f1853, %f1860;
|
|
fma.rn.f32 %f1862, %f1859, %f1859, %f1861;
|
|
sqrt.rn.f32 %f1863, %f1862;
|
|
rcp.rn.f32 %f1864, %f1863;
|
|
mul.f32 %f2132, %f1864, %f1859;
|
|
mul.f32 %f2131, %f1864, %f1856;
|
|
mul.f32 %f2130, %f1853, %f1864;
|
|
|
|
$L__BB0_164:
|
|
mul.f32 %f1868, %f2044, %f2132;
|
|
mul.f32 %f1869, %f2045, %f2131;
|
|
sub.f32 %f1870, %f1868, %f1869;
|
|
mul.f32 %f1871, %f2045, %f2130;
|
|
mul.f32 %f1872, %f2043, %f2132;
|
|
sub.f32 %f1873, %f1871, %f1872;
|
|
mul.f32 %f1874, %f2043, %f2131;
|
|
mul.f32 %f1875, %f2044, %f2130;
|
|
sub.f32 %f1876, %f1874, %f1875;
|
|
mul.f32 %f1877, %f1870, %f293;
|
|
mul.f32 %f1878, %f1873, %f293;
|
|
mul.f32 %f1879, %f1876, %f293;
|
|
mul.f32 %f1880, %f2130, 0f3ED105EC;
|
|
mul.f32 %f1881, %f2131, 0f3ED105EC;
|
|
mul.f32 %f1882, %f2132, 0f3ED105EC;
|
|
mul.f32 %f1883, %f1877, 0f3F3504F3;
|
|
mul.f32 %f1884, %f1878, 0f3F3504F3;
|
|
mul.f32 %f1885, %f1879, 0f3F3504F3;
|
|
sub.f32 %f1886, %f1883, %f1880;
|
|
sub.f32 %f1887, %f1884, %f1881;
|
|
sub.f32 %f1888, %f1885, %f1882;
|
|
add.f32 %f306, %f294, %f1886;
|
|
add.f32 %f307, %f296, %f1887;
|
|
add.f32 %f308, %f298, %f1888;
|
|
mov.f32 %f2134, %f2133;
|
|
mov.f32 %f2135, %f2133;
|
|
@%p238 bra $L__BB0_166;
|
|
|
|
cvt.rn.f32.u16 %f1889, %rs64;
|
|
div.rn.f32 %f1890, %f1889, 0f437F0000;
|
|
fma.rn.f32 %f1891, %f1890, 0f40000000, 0fBF800000;
|
|
and.b16 %rs82, %rs205, 255;
|
|
cvt.rn.f32.u16 %f1892, %rs82;
|
|
div.rn.f32 %f1893, %f1892, 0f437F0000;
|
|
fma.rn.f32 %f1894, %f1893, 0f40000000, 0fBF800000;
|
|
and.b16 %rs83, %rs204, 255;
|
|
cvt.rn.f32.u16 %f1895, %rs83;
|
|
div.rn.f32 %f1896, %f1895, 0f437F0000;
|
|
fma.rn.f32 %f1897, %f1896, 0f40000000, 0fBF800000;
|
|
mul.f32 %f1898, %f1894, %f1894;
|
|
fma.rn.f32 %f1899, %f1891, %f1891, %f1898;
|
|
fma.rn.f32 %f1900, %f1897, %f1897, %f1899;
|
|
sqrt.rn.f32 %f1901, %f1900;
|
|
rcp.rn.f32 %f1902, %f1901;
|
|
mul.f32 %f2135, %f1902, %f1897;
|
|
mul.f32 %f2134, %f1902, %f1894;
|
|
mul.f32 %f2133, %f1891, %f1902;
|
|
|
|
$L__BB0_166:
|
|
mul.f32 %f1903, %f2044, %f2135;
|
|
mul.f32 %f1904, %f2045, %f2134;
|
|
sub.f32 %f1905, %f1903, %f1904;
|
|
mul.f32 %f1906, %f2045, %f2133;
|
|
mul.f32 %f1907, %f2043, %f2135;
|
|
sub.f32 %f1908, %f1906, %f1907;
|
|
mul.f32 %f1909, %f2043, %f2134;
|
|
mul.f32 %f1910, %f2044, %f2133;
|
|
sub.f32 %f1911, %f1909, %f1910;
|
|
mul.f32 %f1912, %f1905, %f293;
|
|
mul.f32 %f1913, %f1908, %f293;
|
|
mul.f32 %f1914, %f1911, %f293;
|
|
mul.f32 %f1915, %f2133, 0f3ED105EC;
|
|
mul.f32 %f1916, %f2134, 0f3ED105EC;
|
|
mul.f32 %f1917, %f2135, 0f3ED105EC;
|
|
mul.f32 %f1918, %f1912, 0fBF3504F3;
|
|
mul.f32 %f1919, %f1913, 0fBF3504F3;
|
|
mul.f32 %f1920, %f1914, 0fBF3504F3;
|
|
sub.f32 %f1921, %f1918, %f1915;
|
|
sub.f32 %f1922, %f1919, %f1916;
|
|
sub.f32 %f1923, %f1920, %f1917;
|
|
add.f32 %f1924, %f294, %f1921;
|
|
add.f32 %f1925, %f296, %f1922;
|
|
add.f32 %f1926, %f298, %f1923;
|
|
mul.f32 %f1927, %f297, %f346;
|
|
neg.f32 %f1928, %f1927;
|
|
mul.f32 %f1929, %f295, %f345;
|
|
sub.f32 %f1930, %f1928, %f1929;
|
|
mul.f32 %f1931, %f299, %f347;
|
|
sub.f32 %f1932, %f1930, %f1931;
|
|
cvt.sat.f32.f32 %f1933, %f1932;
|
|
mul.f32 %f1934, %f199, %f1933;
|
|
mul.f32 %f1935, %f200, %f1933;
|
|
mul.f32 %f1936, %f201, %f1933;
|
|
mul.f32 %f1937, %f307, %f346;
|
|
neg.f32 %f1938, %f1937;
|
|
mul.f32 %f1939, %f306, %f345;
|
|
sub.f32 %f1940, %f1938, %f1939;
|
|
mul.f32 %f1941, %f308, %f347;
|
|
sub.f32 %f1942, %f1940, %f1941;
|
|
cvt.sat.f32.f32 %f1943, %f1942;
|
|
mul.f32 %f1944, %f199, %f1943;
|
|
mul.f32 %f1945, %f200, %f1943;
|
|
mul.f32 %f1946, %f201, %f1943;
|
|
mul.f32 %f1947, %f1924, %f345;
|
|
mul.f32 %f1948, %f1925, %f346;
|
|
neg.f32 %f1949, %f1948;
|
|
sub.f32 %f1950, %f1949, %f1947;
|
|
mul.f32 %f1951, %f1926, %f347;
|
|
sub.f32 %f1952, %f1950, %f1951;
|
|
cvt.sat.f32.f32 %f1953, %f1952;
|
|
mul.f32 %f1954, %f199, %f1953;
|
|
mul.f32 %f1955, %f200, %f1953;
|
|
mul.f32 %f1956, %f201, %f1953;
|
|
add.f32 %f1957, %f1934, %f1944;
|
|
add.f32 %f1958, %f1935, %f1945;
|
|
add.f32 %f1959, %f1936, %f1946;
|
|
add.f32 %f1960, %f1957, %f1954;
|
|
add.f32 %f1961, %f1958, %f1955;
|
|
add.f32 %f1962, %f1959, %f1956;
|
|
mul.f32 %f1963, %f1960, 0f3F13CD3A;
|
|
mul.f32 %f1964, %f1961, 0f3F13CD3A;
|
|
mul.f32 %f1965, %f1962, 0f3F13CD3A;
|
|
div.rn.f32 %f1966, %f183, %f1963;
|
|
div.rn.f32 %f1967, %f184, %f1964;
|
|
div.rn.f32 %f1968, %f185, %f1965;
|
|
setp.eq.f32 %p240, %f183, 0f00000000;
|
|
selp.f32 %f1969, 0f00000000, %f1966, %p240;
|
|
setp.eq.f32 %p241, %f184, 0f00000000;
|
|
selp.f32 %f1970, 0f00000000, %f1967, %p241;
|
|
setp.eq.f32 %p242, %f185, 0f00000000;
|
|
selp.f32 %f1971, 0f00000000, %f1968, %p242;
|
|
mul.f32 %f315, %f1934, %f1969;
|
|
mul.f32 %f316, %f1935, %f1970;
|
|
mul.f32 %f317, %f1936, %f1971;
|
|
mul.f32 %f318, %f1944, %f1969;
|
|
mul.f32 %f319, %f1945, %f1970;
|
|
mul.f32 %f320, %f1946, %f1971;
|
|
mul.f32 %f321, %f1954, %f1969;
|
|
mul.f32 %f322, %f1955, %f1970;
|
|
mul.f32 %f323, %f1956, %f1971;
|
|
ld.const.u32 %r59, [params+108];
|
|
setp.eq.s32 %p243, %r59, 0;
|
|
ld.const.u64 %rd97, [params+256];
|
|
cvta.to.global.u64 %rd98, %rd97;
|
|
ld.const.u32 %r1516, [params+248];
|
|
mad.lo.s32 %r1517, %r1516, %r6, %r5;
|
|
mul.wide.u32 %rd99, %r1517, 8;
|
|
add.s64 %rd21, %rd98, %rd99;
|
|
@%p243 bra $L__BB0_168;
|
|
|
|
ld.global.v4.u16 {%rs91, %rs92, %rs93, %rs94}, [%rd21];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1972, %rs91;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1973, %rs92;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1974, %rs93;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1975, %f315, %f1972;
|
|
add.f32 %f1976, %f316, %f1973;
|
|
add.f32 %f1977, %f317, %f1974;
|
|
mov.f32 %f1978, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs89, %f1977;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f1976;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f1975;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f1978;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd21], {%rs87, %rs88, %rs89, %rs90};
|
|
bra.uni $L__BB0_169;
|
|
|
|
$L__BB0_168:
|
|
mov.f32 %f1982, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f1982;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f317;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs96, %f316;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f315;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd21], {%rs95, %rs96, %rs97, %rs98};
|
|
|
|
$L__BB0_169:
|
|
ld.const.u64 %rd100, [params+272];
|
|
cvta.to.global.u64 %rd101, %rd100;
|
|
ld.const.u32 %r1518, [params+264];
|
|
mad.lo.s32 %r1519, %r1518, %r6, %r5;
|
|
mul.wide.u32 %rd102, %r1519, 8;
|
|
add.s64 %rd22, %rd101, %rd102;
|
|
@%p243 bra $L__BB0_171;
|
|
|
|
ld.global.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd22];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1983, %rs106;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1984, %rs107;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1985, %rs108;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1986, %f318, %f1983;
|
|
add.f32 %f1987, %f319, %f1984;
|
|
add.f32 %f1988, %f320, %f1985;
|
|
mov.f32 %f1989, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f1988;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs103, %f1987;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs102, %f1986;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f1989;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd22], {%rs102, %rs103, %rs104, %rs105};
|
|
bra.uni $L__BB0_172;
|
|
|
|
$L__BB0_171:
|
|
mov.f32 %f1993, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f1993;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f320;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f319;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f318;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd22], {%rs110, %rs111, %rs112, %rs113};
|
|
|
|
$L__BB0_172:
|
|
ld.const.u64 %rd103, [params+288];
|
|
cvta.to.global.u64 %rd104, %rd103;
|
|
ld.const.u32 %r1520, [params+280];
|
|
mad.lo.s32 %r1521, %r1520, %r6, %r5;
|
|
mul.wide.u32 %rd105, %r1521, 8;
|
|
add.s64 %rd23, %rd104, %rd105;
|
|
@%p243 bra $L__BB0_174;
|
|
|
|
ld.global.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd23];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1994, %rs121;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1995, %rs122;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1996, %rs123;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1997, %f321, %f1994;
|
|
add.f32 %f1998, %f322, %f1995;
|
|
add.f32 %f1999, %f323, %f1996;
|
|
mov.f32 %f2000, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs119, %f1999;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f1998;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs117, %f1997;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f2000;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd23], {%rs117, %rs118, %rs119, %rs120};
|
|
bra.uni $L__BB0_192;
|
|
|
|
$L__BB0_174:
|
|
mov.f32 %f2004, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f2004;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f323;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f322;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f321;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd23], {%rs125, %rs126, %rs127, %rs128};
|
|
|
|
$L__BB0_192:
|
|
ret;
|
|
|
|
}
|
|
|