2742 lines
76 KiB
Plaintext
2742 lines
76 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-31968024
|
|
// Cuda compilation tools, release 12.0, V12.0.76
|
|
// Based on NVVM 7.0.1
|
|
//
|
|
|
|
.version 8.0
|
|
.target sm_52
|
|
.address_size 64
|
|
|
|
// .globl __raygen__oxMain
|
|
.const .align 16 .b8 params[1184];
|
|
.global .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
|
|
|
.visible .entry __raygen__oxMain()
|
|
{
|
|
.local .align 4 .b8 __local_depot0[28];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<199>;
|
|
.reg .b16 %rs<177>;
|
|
.reg .f32 %f<1639>;
|
|
.reg .b32 %r<469>;
|
|
.reg .f64 %fd<5>;
|
|
.reg .b64 %rd<125>;
|
|
|
|
|
|
mov.u64 %SPL, __local_depot0;
|
|
add.u64 %rd1, %SPL, 0;
|
|
// begin inline asm
|
|
call (%r62), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r63), _optix_get_launch_index_y, ();
|
|
// end inline asm
|
|
ld.const.u32 %r3, [params+616];
|
|
abs.s32 %r4, %r3;
|
|
ld.const.u32 %r5, [params+540];
|
|
ld.const.u64 %rd34, [params+400];
|
|
cvta.to.global.u64 %rd35, %rd34;
|
|
ld.const.u32 %r68, [params+392];
|
|
mad.lo.s32 %r69, %r68, %r63, %r62;
|
|
mul.wide.u32 %rd36, %r69, 4;
|
|
add.s64 %rd2, %rd35, %rd36;
|
|
ld.global.v2.u8 {%rs7, %rs176}, [%rd2];
|
|
or.b16 %rs9, %rs7, %rs176;
|
|
and.b16 %rs10, %rs9, 255;
|
|
setp.eq.s16 %p10, %rs10, 0;
|
|
@%p10 bra $L__BB0_2;
|
|
|
|
ld.global.u8 %rs175, [%rd2+2];
|
|
bra.uni $L__BB0_3;
|
|
|
|
$L__BB0_2:
|
|
ld.global.u8 %rs175, [%rd2+2];
|
|
setp.eq.s16 %p11, %rs175, 0;
|
|
mov.f32 %f1481, 0f00000000;
|
|
mov.u16 %rs176, 0;
|
|
mov.f32 %f1482, %f1481;
|
|
mov.f32 %f1483, %f1481;
|
|
@%p11 bra $L__BB0_4;
|
|
|
|
$L__BB0_3:
|
|
cvt.rn.f32.u16 %f384, %rs7;
|
|
div.rn.f32 %f385, %f384, 0f437F0000;
|
|
fma.rn.f32 %f386, %f385, 0f40000000, 0fBF800000;
|
|
and.b16 %rs13, %rs176, 255;
|
|
cvt.rn.f32.u16 %f387, %rs13;
|
|
div.rn.f32 %f388, %f387, 0f437F0000;
|
|
fma.rn.f32 %f389, %f388, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f390, %rs175;
|
|
div.rn.f32 %f391, %f390, 0f437F0000;
|
|
fma.rn.f32 %f392, %f391, 0f40000000, 0fBF800000;
|
|
mul.f32 %f393, %f389, %f389;
|
|
fma.rn.f32 %f394, %f386, %f386, %f393;
|
|
fma.rn.f32 %f395, %f392, %f392, %f394;
|
|
sqrt.rn.f32 %f396, %f395;
|
|
rcp.rn.f32 %f397, %f396;
|
|
mul.f32 %f1483, %f397, %f392;
|
|
mul.f32 %f1482, %f397, %f389;
|
|
mul.f32 %f1481, %f386, %f397;
|
|
|
|
$L__BB0_4:
|
|
ld.const.v2.u32 {%r70, %r71}, [params];
|
|
add.s32 %r6, %r70, %r62;
|
|
add.s32 %r7, %r71, %r63;
|
|
setp.eq.f32 %p12, %f1481, 0f00000000;
|
|
setp.eq.f32 %p13, %f1482, 0f00000000;
|
|
and.pred %p14, %p12, %p13;
|
|
setp.eq.f32 %p15, %f1483, 0f00000000;
|
|
and.pred %p16, %p15, %p14;
|
|
@%p16 bra $L__BB0_160;
|
|
bra.uni $L__BB0_5;
|
|
|
|
$L__BB0_160:
|
|
ld.const.u32 %r59, [params+104];
|
|
and.b32 %r432, %r59, 1;
|
|
setp.eq.b32 %p187, %r432, 1;
|
|
mov.pred %p188, 0;
|
|
xor.pred %p189, %p187, %p188;
|
|
not.pred %p190, %p189;
|
|
@%p190 bra $L__BB0_162;
|
|
|
|
ld.const.u64 %rd97, [params+144];
|
|
cvta.to.global.u64 %rd98, %rd97;
|
|
ld.const.u32 %r433, [params+136];
|
|
mad.lo.s32 %r434, %r433, %r7, %r6;
|
|
mul.wide.u32 %rd99, %r434, 4;
|
|
add.s64 %rd100, %rd98, %rd99;
|
|
mov.u16 %rs97, 0;
|
|
st.global.v4.u8 [%rd100], {%rs97, %rs97, %rs97, %rs97};
|
|
|
|
$L__BB0_162:
|
|
and.b32 %r435, %r59, 8;
|
|
setp.eq.s32 %p191, %r435, 0;
|
|
@%p191 bra $L__BB0_164;
|
|
|
|
ld.const.u64 %rd101, [params+192];
|
|
cvta.to.global.u64 %rd102, %rd101;
|
|
ld.const.u32 %r436, [params+184];
|
|
mad.lo.s32 %r437, %r436, %r7, %r6;
|
|
mov.f32 %f1427, 0f00000000;
|
|
cvt.rzi.u32.f32 %r438, %f1427;
|
|
mul.wide.u32 %rd103, %r437, 2;
|
|
add.s64 %rd104, %rd102, %rd103;
|
|
mov.u16 %rs98, 0;
|
|
cvt.u16.u32 %rs99, %r438;
|
|
st.global.v2.u8 [%rd104], {%rs99, %rs98};
|
|
|
|
$L__BB0_164:
|
|
and.b32 %r439, %r59, 4;
|
|
setp.eq.s32 %p192, %r439, 0;
|
|
ld.const.u32 %r468, [params+108];
|
|
@%p192 bra $L__BB0_168;
|
|
|
|
setp.eq.s32 %p193, %r468, 0;
|
|
ld.const.u64 %rd105, [params+224];
|
|
cvta.to.global.u64 %rd106, %rd105;
|
|
ld.const.u32 %r440, [params+216];
|
|
mad.lo.s32 %r441, %r440, %r7, %r6;
|
|
mul.wide.u32 %rd107, %r441, 8;
|
|
add.s64 %rd28, %rd106, %rd107;
|
|
@%p193 bra $L__BB0_167;
|
|
|
|
ld.global.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd28];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1428, %rs106;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1429, %rs107;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1430, %rs108;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1431, %f1428, 0f00000000;
|
|
add.f32 %f1432, %f1429, 0f00000000;
|
|
add.f32 %f1433, %f1430, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f1433;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f1432;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs103, %f1431;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs110, 0;
|
|
st.global.v4.u16 [%rd28], {%rs103, %rs104, %rs105, %rs110};
|
|
bra.uni $L__BB0_168;
|
|
|
|
$L__BB0_5:
|
|
ld.const.u64 %rd37, [params+432];
|
|
cvta.to.global.u64 %rd38, %rd37;
|
|
ld.const.u32 %r75, [params+424];
|
|
mad.lo.s32 %r76, %r75, %r63, %r62;
|
|
mul.wide.u32 %rd39, %r76, 12;
|
|
add.s64 %rd40, %rd38, %rd39;
|
|
ld.global.f32 %f10, [%rd40];
|
|
mul.f32 %f414, %f10, 0f3456BF95;
|
|
ld.global.f32 %f11, [%rd40+4];
|
|
mul.f32 %f415, %f11, 0f3456BF95;
|
|
ld.global.f32 %f12, [%rd40+8];
|
|
mul.f32 %f416, %f12, 0f3456BF95;
|
|
abs.f32 %f13, %f1481;
|
|
div.rn.f32 %f417, %f414, %f13;
|
|
abs.f32 %f418, %f1482;
|
|
div.rn.f32 %f419, %f415, %f418;
|
|
abs.f32 %f14, %f1483;
|
|
div.rn.f32 %f420, %f416, %f14;
|
|
abs.f32 %f421, %f417;
|
|
abs.f32 %f422, %f419;
|
|
abs.f32 %f423, %f420;
|
|
mov.f32 %f424, 0f38D1B717;
|
|
max.f32 %f425, %f421, %f424;
|
|
max.f32 %f426, %f422, %f424;
|
|
max.f32 %f427, %f423, %f424;
|
|
fma.rn.f32 %f15, %f1481, %f425, %f10;
|
|
fma.rn.f32 %f16, %f1482, %f426, %f11;
|
|
fma.rn.f32 %f17, %f1483, %f427, %f12;
|
|
ld.const.u64 %rd41, [params+128];
|
|
cvta.to.global.u64 %rd42, %rd41;
|
|
ld.const.u32 %r77, [params+120];
|
|
mad.lo.s32 %r78, %r77, %r63, %r62;
|
|
mul.wide.u32 %rd43, %r78, 4;
|
|
add.s64 %rd44, %rd42, %rd43;
|
|
ld.global.u32 %r459, [%rd44];
|
|
setp.lt.s32 %p18, %r4, 1;
|
|
mov.pred %p17, 0;
|
|
mov.u32 %r455, 0;
|
|
mov.f32 %f1503, 0f00000000;
|
|
mov.f32 %f1504, %f1503;
|
|
mov.f32 %f1505, %f1503;
|
|
mov.f32 %f1506, %f1503;
|
|
mov.f32 %f1507, %f1503;
|
|
mov.f32 %f1508, %f1503;
|
|
mov.f32 %f1509, %f1503;
|
|
mov.f32 %f1510, %f1503;
|
|
mov.f32 %f1511, %f1503;
|
|
mov.f32 %f1512, %f1503;
|
|
mov.f32 %f1513, %f1503;
|
|
mov.f32 %f1514, %f1503;
|
|
mov.f32 %f1515, %f1503;
|
|
mov.f32 %f1516, %f1503;
|
|
mov.f32 %f1517, %f1503;
|
|
mov.f32 %f1518, %f1503;
|
|
mov.pred %p198, %p17;
|
|
@%p18 bra $L__BB0_25;
|
|
|
|
ld.const.f32 %f18, [params+620];
|
|
ld.const.u64 %rd45, [params+624];
|
|
cvta.to.global.u64 %rd3, %rd45;
|
|
ld.const.u64 %rd46, [params+640];
|
|
cvta.to.global.u64 %rd47, %rd46;
|
|
ld.const.u32 %r81, [params+632];
|
|
and.b32 %r82, %r63, 255;
|
|
and.b32 %r83, %r62, 255;
|
|
mad.lo.s32 %r84, %r81, %r82, %r83;
|
|
mul.wide.u32 %rd48, %r84, 3;
|
|
add.s64 %rd4, %rd47, %rd48;
|
|
ld.const.f32 %f19, [params+660];
|
|
mul.f32 %f20, %f15, 0f3456BF95;
|
|
mul.f32 %f21, %f16, 0f3456BF95;
|
|
mul.f32 %f22, %f17, 0f3456BF95;
|
|
ld.const.u64 %rd5, [params+96];
|
|
abs.f32 %f571, %f21;
|
|
abs.f32 %f572, %f20;
|
|
max.f32 %f573, %f572, %f571;
|
|
abs.f32 %f574, %f22;
|
|
max.f32 %f575, %f573, %f574;
|
|
mov.u32 %r452, %r455;
|
|
|
|
$L__BB0_7:
|
|
mul.lo.s32 %r85, %r452, 3;
|
|
mul.wide.s32 %rd49, %r85, 12;
|
|
add.s64 %rd6, %rd3, %rd49;
|
|
ld.global.f32 %f444, [%rd6];
|
|
sub.f32 %f445, %f444, %f10;
|
|
ld.global.f32 %f446, [%rd6+4];
|
|
sub.f32 %f447, %f446, %f11;
|
|
ld.global.f32 %f448, [%rd6+8];
|
|
sub.f32 %f449, %f448, %f12;
|
|
mul.f32 %f450, %f447, %f447;
|
|
fma.rn.f32 %f451, %f445, %f445, %f450;
|
|
fma.rn.f32 %f452, %f449, %f449, %f451;
|
|
sqrt.rn.f32 %f39, %f452;
|
|
rcp.rn.f32 %f453, %f39;
|
|
mul.f32 %f40, %f445, %f453;
|
|
mul.f32 %f41, %f447, %f453;
|
|
mul.f32 %f42, %f449, %f453;
|
|
mul.f32 %f454, %f1482, %f41;
|
|
fma.rn.f32 %f455, %f1481, %f40, %f454;
|
|
fma.rn.f32 %f43, %f1483, %f42, %f455;
|
|
setp.leu.f32 %p19, %f43, 0f00000000;
|
|
@%p19 bra $L__BB0_24;
|
|
|
|
setp.ne.s32 %p21, %r5, 0;
|
|
mul.f32 %f456, %f39, %f39;
|
|
div.rn.f32 %f44, %f18, %f456;
|
|
ld.global.u8 %rs15, [%rd4];
|
|
cvt.rn.f32.u16 %f457, %rs15;
|
|
div.rn.f32 %f458, %f457, 0fC37F0000;
|
|
fma.rn.f32 %f459, %f458, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p22, %f44, %f459;
|
|
and.pred %p23, %p21, %p22;
|
|
mov.pred %p198, -1;
|
|
@%p23 bra $L__BB0_25;
|
|
|
|
mul.f32 %f45, %f39, %f19;
|
|
mov.f32 %f464, 0f40800000;
|
|
abs.f32 %f47, %f45;
|
|
setp.lt.f32 %p24, %f47, 0f00800000;
|
|
mul.f32 %f466, %f47, 0f4B800000;
|
|
selp.f32 %f467, %f466, %f47, %p24;
|
|
selp.f32 %f468, 0fC3170000, 0fC2FE0000, %p24;
|
|
mov.b32 %r86, %f467;
|
|
and.b32 %r87, %r86, 8388607;
|
|
or.b32 %r88, %r87, 1065353216;
|
|
mov.b32 %f469, %r88;
|
|
shr.u32 %r89, %r86, 23;
|
|
cvt.rn.f32.u32 %f470, %r89;
|
|
add.f32 %f471, %f468, %f470;
|
|
setp.gt.f32 %p25, %f469, 0f3FB504F3;
|
|
mul.f32 %f472, %f469, 0f3F000000;
|
|
add.f32 %f473, %f471, 0f3F800000;
|
|
selp.f32 %f474, %f473, %f471, %p25;
|
|
selp.f32 %f475, %f472, %f469, %p25;
|
|
add.f32 %f476, %f475, 0fBF800000;
|
|
add.f32 %f477, %f475, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f478, %f477;
|
|
add.f32 %f479, %f476, %f476;
|
|
mul.f32 %f480, %f479, %f478;
|
|
mul.f32 %f481, %f480, %f480;
|
|
mov.f32 %f482, 0f3C4CAF63;
|
|
mov.f32 %f483, 0f3B18F0FE;
|
|
fma.rn.f32 %f484, %f483, %f481, %f482;
|
|
mov.f32 %f485, 0f3DAAAABD;
|
|
fma.rn.f32 %f486, %f484, %f481, %f485;
|
|
mul.rn.f32 %f487, %f486, %f481;
|
|
mul.rn.f32 %f488, %f487, %f480;
|
|
sub.f32 %f489, %f476, %f480;
|
|
add.f32 %f490, %f489, %f489;
|
|
neg.f32 %f491, %f480;
|
|
fma.rn.f32 %f492, %f491, %f476, %f490;
|
|
mul.rn.f32 %f493, %f478, %f492;
|
|
add.f32 %f494, %f488, %f480;
|
|
sub.f32 %f495, %f480, %f494;
|
|
add.f32 %f496, %f488, %f495;
|
|
add.f32 %f497, %f493, %f496;
|
|
add.f32 %f498, %f494, %f497;
|
|
sub.f32 %f499, %f494, %f498;
|
|
add.f32 %f500, %f497, %f499;
|
|
mov.f32 %f501, 0f3F317200;
|
|
mul.rn.f32 %f502, %f474, %f501;
|
|
mov.f32 %f503, 0f35BFBE8E;
|
|
mul.rn.f32 %f504, %f474, %f503;
|
|
add.f32 %f505, %f502, %f498;
|
|
sub.f32 %f506, %f502, %f505;
|
|
add.f32 %f507, %f498, %f506;
|
|
add.f32 %f508, %f500, %f507;
|
|
add.f32 %f509, %f504, %f508;
|
|
add.f32 %f510, %f505, %f509;
|
|
sub.f32 %f511, %f505, %f510;
|
|
add.f32 %f512, %f509, %f511;
|
|
mul.rn.f32 %f513, %f464, %f510;
|
|
neg.f32 %f514, %f513;
|
|
fma.rn.f32 %f515, %f464, %f510, %f514;
|
|
fma.rn.f32 %f516, %f464, %f512, %f515;
|
|
mov.f32 %f517, 0f00000000;
|
|
fma.rn.f32 %f518, %f517, %f510, %f516;
|
|
add.rn.f32 %f519, %f513, %f518;
|
|
neg.f32 %f520, %f519;
|
|
add.rn.f32 %f521, %f513, %f520;
|
|
add.rn.f32 %f522, %f521, %f518;
|
|
mov.b32 %r90, %f519;
|
|
setp.eq.s32 %p26, %r90, 1118925336;
|
|
add.s32 %r91, %r90, -1;
|
|
mov.b32 %f523, %r91;
|
|
add.f32 %f524, %f522, 0f37000000;
|
|
selp.f32 %f48, %f524, %f522, %p26;
|
|
selp.f32 %f525, %f523, %f519, %p26;
|
|
mov.f32 %f526, 0f3FB8AA3B;
|
|
mul.rn.f32 %f527, %f525, %f526;
|
|
cvt.rzi.f32.f32 %f528, %f527;
|
|
abs.f32 %f529, %f528;
|
|
setp.gt.f32 %p27, %f529, 0f42FC0000;
|
|
mov.b32 %r92, %f528;
|
|
and.b32 %r93, %r92, -2147483648;
|
|
or.b32 %r94, %r93, 1123811328;
|
|
mov.b32 %f530, %r94;
|
|
selp.f32 %f531, %f530, %f528, %p27;
|
|
mov.f32 %f532, 0fBF317218;
|
|
fma.rn.f32 %f533, %f531, %f532, %f525;
|
|
mov.f32 %f534, 0f3102E308;
|
|
fma.rn.f32 %f535, %f531, %f534, %f533;
|
|
mul.f32 %f536, %f535, 0f3FB8AA3B;
|
|
add.f32 %f537, %f531, 0f4B40007F;
|
|
mov.b32 %r95, %f537;
|
|
shl.b32 %r96, %r95, 23;
|
|
mov.b32 %f538, %r96;
|
|
ex2.approx.ftz.f32 %f539, %f536;
|
|
mul.f32 %f49, %f539, %f538;
|
|
setp.eq.f32 %p28, %f49, 0f7F800000;
|
|
mov.f32 %f1500, 0f7F800000;
|
|
@%p28 bra $L__BB0_11;
|
|
|
|
fma.rn.f32 %f1500, %f49, %f48, %f49;
|
|
|
|
$L__BB0_11:
|
|
mov.f32 %f1478, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1477, %f1478;
|
|
add.f32 %f1476, %f1477, %f1477;
|
|
mov.f32 %f1475, 0f40800000;
|
|
sub.f32 %f1474, %f1475, %f1476;
|
|
abs.f32 %f1473, %f1474;
|
|
setp.lt.f32 %p29, %f45, 0f00000000;
|
|
setp.eq.f32 %p30, %f1473, 0f3F800000;
|
|
and.pred %p1, %p29, %p30;
|
|
setp.eq.f32 %p31, %f45, 0f00000000;
|
|
@%p31 bra $L__BB0_15;
|
|
bra.uni $L__BB0_12;
|
|
|
|
$L__BB0_15:
|
|
add.f32 %f544, %f45, %f45;
|
|
selp.f32 %f1502, %f544, 0f00000000, %p30;
|
|
bra.uni $L__BB0_16;
|
|
|
|
$L__BB0_12:
|
|
mov.b32 %r97, %f1500;
|
|
xor.b32 %r98, %r97, -2147483648;
|
|
mov.b32 %f540, %r98;
|
|
selp.f32 %f1502, %f540, %f1500, %p1;
|
|
setp.geu.f32 %p32, %f45, 0f00000000;
|
|
@%p32 bra $L__BB0_16;
|
|
|
|
mov.f32 %f541, 0f40800000;
|
|
cvt.rzi.f32.f32 %f542, %f541;
|
|
setp.eq.f32 %p33, %f542, 0f40800000;
|
|
@%p33 bra $L__BB0_16;
|
|
|
|
mov.f32 %f1502, 0f7FFFFFFF;
|
|
|
|
$L__BB0_16:
|
|
add.f32 %f545, %f47, 0f40800000;
|
|
mov.b32 %r99, %f545;
|
|
setp.lt.s32 %p35, %r99, 2139095040;
|
|
@%p35 bra $L__BB0_21;
|
|
|
|
setp.gtu.f32 %p36, %f47, 0f7F800000;
|
|
@%p36 bra $L__BB0_20;
|
|
bra.uni $L__BB0_18;
|
|
|
|
$L__BB0_20:
|
|
add.f32 %f1502, %f45, 0f40800000;
|
|
bra.uni $L__BB0_21;
|
|
|
|
$L__BB0_18:
|
|
setp.neu.f32 %p37, %f47, 0f7F800000;
|
|
@%p37 bra $L__BB0_21;
|
|
|
|
selp.f32 %f1502, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
$L__BB0_21:
|
|
mov.f32 %f546, 0f3F800000;
|
|
sub.f32 %f547, %f546, %f1502;
|
|
setp.eq.f32 %p38, %f45, 0f3F800000;
|
|
selp.f32 %f548, 0f00000000, %f547, %p38;
|
|
cvt.sat.f32.f32 %f549, %f548;
|
|
mul.f32 %f550, %f44, %f549;
|
|
ld.global.f32 %f551, [%rd6+12];
|
|
mul.f32 %f552, %f40, %f551;
|
|
ld.global.f32 %f553, [%rd6+16];
|
|
mul.f32 %f554, %f41, %f553;
|
|
neg.f32 %f555, %f554;
|
|
sub.f32 %f556, %f555, %f552;
|
|
ld.global.f32 %f557, [%rd6+20];
|
|
mul.f32 %f558, %f42, %f557;
|
|
sub.f32 %f559, %f556, %f558;
|
|
cvt.sat.f32.f32 %f560, %f559;
|
|
mul.f32 %f58, %f550, %f560;
|
|
setp.leu.f32 %p39, %f58, 0f3727C5AC;
|
|
@%p39 bra $L__BB0_23;
|
|
|
|
cvt.sat.f32.f32 %f570, %f43;
|
|
mov.f32 %f576, 0f38D1B717;
|
|
max.f32 %f567, %f575, %f576;
|
|
sub.f32 %f568, %f39, %f567;
|
|
mov.f32 %f569, 0f00000000;
|
|
mov.u32 %r136, 2;
|
|
mov.u32 %r138, 1;
|
|
mov.u32 %r139, 1065353216;
|
|
mov.u32 %r170, 0;
|
|
// begin inline asm
|
|
call(%r100,%r101,%r102,%r103,%r104,%r105,%r106,%r107,%r108,%r109,%r110,%r111,%r112,%r113,%r114,%r115,%r116,%r117,%r118,%r119,%r120,%r121,%r122,%r123,%r124,%r125,%r126,%r127,%r128,%r129,%r130,%r131),_optix_trace_typed_32,(%r170,%rd5,%f15,%f16,%f17,%f40,%f41,%f42,%f567,%f568,%f569,%r138,%r170,%r138,%r136,%r138,%r138,%r139,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170,%r170);
|
|
// end inline asm
|
|
mov.b32 %f577, %r100;
|
|
ld.global.f32 %f578, [%rd6+24];
|
|
mul.f32 %f579, %f578, %f577;
|
|
ld.global.f32 %f580, [%rd6+28];
|
|
mul.f32 %f581, %f580, %f577;
|
|
ld.global.f32 %f582, [%rd6+32];
|
|
mul.f32 %f583, %f582, %f577;
|
|
mul.f32 %f584, %f58, 0f3EA2F983;
|
|
mul.f32 %f585, %f584, %f579;
|
|
mul.f32 %f586, %f584, %f581;
|
|
mul.f32 %f587, %f584, %f583;
|
|
fma.rn.f32 %f1515, %f570, %f585, %f1515;
|
|
fma.rn.f32 %f1516, %f570, %f586, %f1516;
|
|
fma.rn.f32 %f1517, %f570, %f587, %f1517;
|
|
mul.f32 %f588, %f43, 0f40800000;
|
|
cvt.sat.f32.f32 %f589, %f588;
|
|
mul.f32 %f590, %f585, %f589;
|
|
mul.f32 %f591, %f586, %f589;
|
|
mul.f32 %f592, %f589, %f587;
|
|
fma.rn.f32 %f1509, %f40, %f590, %f1509;
|
|
fma.rn.f32 %f1510, %f40, %f591, %f1510;
|
|
fma.rn.f32 %f1511, %f40, %f592, %f1511;
|
|
fma.rn.f32 %f1506, %f41, %f590, %f1506;
|
|
fma.rn.f32 %f1507, %f41, %f591, %f1507;
|
|
fma.rn.f32 %f1508, %f41, %f592, %f1508;
|
|
fma.rn.f32 %f1503, %f42, %f590, %f1503;
|
|
fma.rn.f32 %f1504, %f42, %f591, %f1504;
|
|
fma.rn.f32 %f1505, %f42, %f592, %f1505;
|
|
add.f32 %f1512, %f1512, %f590;
|
|
add.f32 %f1513, %f1513, %f591;
|
|
add.f32 %f1514, %f1514, %f592;
|
|
add.f32 %f1518, %f1518, %f577;
|
|
|
|
$L__BB0_23:
|
|
add.s32 %r455, %r455, 1;
|
|
|
|
$L__BB0_24:
|
|
add.s32 %r452, %r452, 1;
|
|
setp.lt.s32 %p41, %r452, %r4;
|
|
mov.pred %p198, %p17;
|
|
@%p41 bra $L__BB0_7;
|
|
|
|
$L__BB0_25:
|
|
cvt.rn.f32.s32 %f593, %r455;
|
|
mov.f32 %f594, 0f3F800000;
|
|
max.f32 %f595, %f593, %f594;
|
|
rcp.rn.f32 %f596, %f595;
|
|
mul.f32 %f1614, %f1515, %f596;
|
|
mul.f32 %f1615, %f1516, %f596;
|
|
mul.f32 %f1616, %f1517, %f596;
|
|
div.rn.f32 %f1617, %f1518, %f595;
|
|
mul.f32 %f1611, %f1512, %f596;
|
|
mul.f32 %f1612, %f1513, %f596;
|
|
mul.f32 %f1613, %f1514, %f596;
|
|
mul.f32 %f1608, %f1509, %f596;
|
|
mul.f32 %f1609, %f1510, %f596;
|
|
mul.f32 %f1610, %f1511, %f596;
|
|
mul.f32 %f1605, %f1506, %f596;
|
|
mul.f32 %f1606, %f1507, %f596;
|
|
mul.f32 %f1607, %f1508, %f596;
|
|
mul.f32 %f1602, %f1503, %f596;
|
|
mul.f32 %f1603, %f1504, %f596;
|
|
mul.f32 %f1604, %f1505, %f596;
|
|
not.pred %p42, %p198;
|
|
@%p42 bra $L__BB0_56;
|
|
|
|
abs.f32 %f1480, %f1483;
|
|
abs.f32 %f1479, %f1481;
|
|
setp.gt.f32 %p43, %f1479, %f1480;
|
|
selp.f32 %f612, 0f00000000, %f1482, %p43;
|
|
mov.f32 %f611, 0f00000000;
|
|
neg.f32 %f613, %f1483;
|
|
selp.f32 %f614, %f1481, %f613, %p43;
|
|
neg.f32 %f615, %f1482;
|
|
selp.f32 %f616, %f615, 0f00000000, %p43;
|
|
mul.f32 %f617, %f614, %f614;
|
|
fma.rn.f32 %f618, %f616, %f616, %f617;
|
|
fma.rn.f32 %f619, %f612, %f612, %f618;
|
|
sqrt.rn.f32 %f620, %f619;
|
|
rcp.rn.f32 %f621, %f620;
|
|
mul.f32 %f139, %f616, %f621;
|
|
mul.f32 %f140, %f614, %f621;
|
|
mul.f32 %f141, %f612, %f621;
|
|
setp.lt.s32 %p44, %r5, 1;
|
|
mov.f32 %f1580, %f611;
|
|
mov.f32 %f1579, %f611;
|
|
mov.f32 %f1578, %f611;
|
|
mov.f32 %f1577, %f611;
|
|
mov.f32 %f1576, %f611;
|
|
mov.f32 %f1575, %f611;
|
|
mov.f32 %f1574, %f611;
|
|
mov.f32 %f1573, %f611;
|
|
mov.f32 %f1572, %f611;
|
|
mov.f32 %f1571, %f611;
|
|
mov.f32 %f1570, %f611;
|
|
mov.f32 %f1569, %f611;
|
|
mov.f32 %f1568, %f611;
|
|
mov.f32 %f1567, %f611;
|
|
mov.f32 %f1566, %f611;
|
|
@%p44 bra $L__BB0_55;
|
|
|
|
cvt.rn.f32.s32 %f637, %r5;
|
|
rcp.rn.f32 %f142, %f637;
|
|
mul.f32 %f143, %f15, 0f3456BF95;
|
|
mul.f32 %f144, %f16, 0f3456BF95;
|
|
mul.f32 %f145, %f17, 0f3456BF95;
|
|
ld.const.u64 %rd7, [params+96];
|
|
mul.f32 %f638, %f1481, %f140;
|
|
mul.f32 %f639, %f1482, %f139;
|
|
sub.f32 %f146, %f639, %f638;
|
|
mul.f32 %f640, %f1483, %f139;
|
|
mul.f32 %f641, %f1481, %f141;
|
|
sub.f32 %f147, %f641, %f640;
|
|
mul.f32 %f642, %f1482, %f141;
|
|
mul.f32 %f643, %f1483, %f140;
|
|
sub.f32 %f148, %f643, %f642;
|
|
mov.u32 %r171, 0;
|
|
add.s64 %rd8, %rd1, 24;
|
|
mov.u64 %rd51, __cudart_i2opi_f;
|
|
abs.f32 %f706, %f144;
|
|
abs.f32 %f707, %f143;
|
|
max.f32 %f708, %f707, %f706;
|
|
abs.f32 %f709, %f145;
|
|
max.f32 %f710, %f708, %f709;
|
|
mov.u32 %r456, %r171;
|
|
|
|
$L__BB0_28:
|
|
cvt.rn.f32.s32 %f164, %r456;
|
|
mov.u32 %r458, %r171;
|
|
|
|
$L__BB0_29:
|
|
mad.lo.s32 %r173, %r459, 1664525, 1013904223;
|
|
and.b32 %r174, %r173, 16777215;
|
|
cvt.rn.f32.u32 %f644, %r174;
|
|
fma.rn.f32 %f645, %f644, 0f33800000, %f164;
|
|
mul.f32 %f180, %f142, %f645;
|
|
mad.lo.s32 %r459, %r173, 1664525, 1013904223;
|
|
and.b32 %r175, %r459, 16777215;
|
|
cvt.rn.f32.u32 %f646, %r175;
|
|
cvt.rn.f32.s32 %f647, %r458;
|
|
fma.rn.f32 %f648, %f646, 0f33800000, %f647;
|
|
mul.f32 %f649, %f142, %f648;
|
|
mul.f32 %f650, %f180, %f180;
|
|
mov.f32 %f651, 0f3F800000;
|
|
sub.f32 %f652, %f651, %f650;
|
|
mov.f32 %f653, 0f00000000;
|
|
max.f32 %f654, %f653, %f652;
|
|
sqrt.rn.f32 %f181, %f654;
|
|
mul.f32 %f182, %f649, 0f40C90FDB;
|
|
mul.f32 %f655, %f182, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r466, %f655;
|
|
cvt.rn.f32.s32 %f656, %r466;
|
|
mov.f32 %f657, 0fBFC90FDA;
|
|
fma.rn.f32 %f658, %f656, %f657, %f182;
|
|
mov.f32 %f659, 0fB3A22168;
|
|
fma.rn.f32 %f660, %f656, %f659, %f658;
|
|
mov.f32 %f661, 0fA7C234C5;
|
|
fma.rn.f32 %f1584, %f656, %f661, %f660;
|
|
abs.f32 %f184, %f182;
|
|
setp.ltu.f32 %p45, %f184, 0f47CE4780;
|
|
mov.u32 %r463, %r466;
|
|
mov.f32 %f1581, %f1584;
|
|
@%p45 bra $L__BB0_37;
|
|
|
|
setp.eq.f32 %p46, %f184, 0f7F800000;
|
|
@%p46 bra $L__BB0_36;
|
|
bra.uni $L__BB0_31;
|
|
|
|
$L__BB0_36:
|
|
mov.f32 %f664, 0f00000000;
|
|
mul.rn.f32 %f1581, %f182, %f664;
|
|
mov.u32 %r463, 0;
|
|
bra.uni $L__BB0_37;
|
|
|
|
$L__BB0_31:
|
|
mov.b32 %r21, %f182;
|
|
bfe.u32 %r177, %r21, 23, 8;
|
|
add.s32 %r22, %r177, -128;
|
|
shl.b32 %r178, %r21, 8;
|
|
or.b32 %r23, %r178, -2147483648;
|
|
shr.u32 %r24, %r22, 5;
|
|
mov.u64 %rd122, 0;
|
|
mov.u32 %r460, 0;
|
|
mov.u64 %rd120, %rd1;
|
|
mov.u64 %rd121, %rd51;
|
|
|
|
$L__BB0_32:
|
|
.pragma "nounroll";
|
|
ld.global.nc.u32 %r179, [%rd121];
|
|
mad.wide.u32 %rd53, %r179, %r23, %rd122;
|
|
shr.u64 %rd122, %rd53, 32;
|
|
st.local.u32 [%rd120], %rd53;
|
|
add.s64 %rd121, %rd121, 4;
|
|
add.s64 %rd120, %rd120, 4;
|
|
add.s32 %r460, %r460, 1;
|
|
setp.ne.s32 %p47, %r460, 6;
|
|
@%p47 bra $L__BB0_32;
|
|
|
|
st.local.u32 [%rd8], %rd122;
|
|
mov.u32 %r180, 4;
|
|
sub.s32 %r27, %r180, %r24;
|
|
mov.u32 %r181, 6;
|
|
sub.s32 %r182, %r181, %r24;
|
|
mul.wide.s32 %rd54, %r182, 4;
|
|
add.s64 %rd55, %rd1, %rd54;
|
|
ld.local.u32 %r461, [%rd55];
|
|
ld.local.u32 %r462, [%rd55+-4];
|
|
and.b32 %r30, %r22, 31;
|
|
setp.eq.s32 %p48, %r30, 0;
|
|
@%p48 bra $L__BB0_35;
|
|
|
|
mov.u32 %r183, 32;
|
|
sub.s32 %r184, %r183, %r30;
|
|
shr.u32 %r185, %r462, %r184;
|
|
shl.b32 %r186, %r461, %r30;
|
|
add.s32 %r461, %r185, %r186;
|
|
mul.wide.s32 %rd56, %r27, 4;
|
|
add.s64 %rd57, %rd1, %rd56;
|
|
ld.local.u32 %r187, [%rd57];
|
|
shr.u32 %r188, %r187, %r184;
|
|
shl.b32 %r189, %r462, %r30;
|
|
add.s32 %r462, %r188, %r189;
|
|
|
|
$L__BB0_35:
|
|
and.b32 %r190, %r21, -2147483648;
|
|
shr.u32 %r191, %r462, 30;
|
|
shl.b32 %r192, %r461, 2;
|
|
or.b32 %r193, %r191, %r192;
|
|
shr.u32 %r194, %r193, 31;
|
|
shr.u32 %r195, %r461, 30;
|
|
add.s32 %r196, %r194, %r195;
|
|
neg.s32 %r197, %r196;
|
|
setp.eq.s32 %p49, %r190, 0;
|
|
selp.b32 %r463, %r196, %r197, %p49;
|
|
setp.ne.s32 %p50, %r194, 0;
|
|
xor.b32 %r198, %r190, -2147483648;
|
|
selp.b32 %r199, %r198, %r190, %p50;
|
|
selp.b32 %r200, -1, 0, %p50;
|
|
xor.b32 %r201, %r193, %r200;
|
|
shl.b32 %r202, %r462, 2;
|
|
xor.b32 %r203, %r202, %r200;
|
|
cvt.u64.u32 %rd58, %r201;
|
|
cvt.u64.u32 %rd59, %r203;
|
|
bfi.b64 %rd60, %rd58, %rd59, 32, 32;
|
|
cvt.rn.f64.s64 %fd1, %rd60;
|
|
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f662, %fd2;
|
|
setp.eq.s32 %p51, %r199, 0;
|
|
neg.f32 %f663, %f662;
|
|
selp.f32 %f1581, %f662, %f663, %p51;
|
|
|
|
$L__BB0_37:
|
|
add.s32 %r37, %r463, 1;
|
|
and.b32 %r38, %r37, 1;
|
|
setp.eq.s32 %p52, %r38, 0;
|
|
selp.f32 %f188, %f1581, 0f3F800000, %p52;
|
|
mul.rn.f32 %f189, %f1581, %f1581;
|
|
mov.f32 %f1582, 0fB94D4153;
|
|
@%p52 bra $L__BB0_39;
|
|
|
|
mov.f32 %f666, 0fBAB607ED;
|
|
mov.f32 %f667, 0f37CBAC00;
|
|
fma.rn.f32 %f1582, %f667, %f189, %f666;
|
|
|
|
$L__BB0_39:
|
|
selp.f32 %f668, 0f3C0885E4, 0f3D2AAABB, %p52;
|
|
fma.rn.f32 %f669, %f1582, %f189, %f668;
|
|
selp.f32 %f670, 0fBE2AAAA8, 0fBEFFFFFF, %p52;
|
|
fma.rn.f32 %f671, %f669, %f189, %f670;
|
|
mov.f32 %f672, 0f00000000;
|
|
fma.rn.f32 %f673, %f189, %f188, %f672;
|
|
fma.rn.f32 %f1583, %f671, %f673, %f188;
|
|
and.b32 %r205, %r37, 2;
|
|
setp.eq.s32 %p54, %r205, 0;
|
|
@%p54 bra $L__BB0_41;
|
|
|
|
mov.f32 %f675, 0fBF800000;
|
|
fma.rn.f32 %f1583, %f1583, %f675, %f672;
|
|
|
|
$L__BB0_41:
|
|
@%p45 bra $L__BB0_49;
|
|
|
|
setp.eq.f32 %p56, %f184, 0f7F800000;
|
|
@%p56 bra $L__BB0_48;
|
|
bra.uni $L__BB0_43;
|
|
|
|
$L__BB0_48:
|
|
mov.f32 %f678, 0f00000000;
|
|
mul.rn.f32 %f1584, %f182, %f678;
|
|
mov.u32 %r466, 0;
|
|
bra.uni $L__BB0_49;
|
|
|
|
$L__BB0_43:
|
|
mov.b32 %r39, %f182;
|
|
bfe.u32 %r206, %r39, 23, 8;
|
|
add.s32 %r40, %r206, -128;
|
|
shl.b32 %r207, %r39, 8;
|
|
or.b32 %r41, %r207, -2147483648;
|
|
shr.u32 %r42, %r40, 5;
|
|
mov.u64 %rd123, 0;
|
|
mov.u64 %rd124, %rd123;
|
|
|
|
$L__BB0_44:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd63, %rd123, 2;
|
|
mov.u64 %rd64, __cudart_i2opi_f;
|
|
add.s64 %rd65, %rd64, %rd63;
|
|
ld.global.nc.u32 %r208, [%rd65];
|
|
mad.wide.u32 %rd66, %r208, %r41, %rd124;
|
|
shr.u64 %rd124, %rd66, 32;
|
|
add.s64 %rd67, %rd1, %rd63;
|
|
st.local.u32 [%rd67], %rd66;
|
|
cvt.u32.u64 %r209, %rd123;
|
|
add.s32 %r210, %r209, 1;
|
|
cvt.s64.s32 %rd123, %r210;
|
|
setp.ne.s32 %p57, %r210, 6;
|
|
@%p57 bra $L__BB0_44;
|
|
|
|
st.local.u32 [%rd8], %rd124;
|
|
mov.u32 %r211, 4;
|
|
sub.s32 %r43, %r211, %r42;
|
|
mov.u32 %r212, 6;
|
|
sub.s32 %r213, %r212, %r42;
|
|
mul.wide.s32 %rd68, %r213, 4;
|
|
add.s64 %rd69, %rd1, %rd68;
|
|
ld.local.u32 %r464, [%rd69];
|
|
ld.local.u32 %r465, [%rd69+-4];
|
|
and.b32 %r46, %r40, 31;
|
|
setp.eq.s32 %p58, %r46, 0;
|
|
@%p58 bra $L__BB0_47;
|
|
|
|
mov.u32 %r214, 32;
|
|
sub.s32 %r215, %r214, %r46;
|
|
shr.u32 %r216, %r465, %r215;
|
|
shl.b32 %r217, %r464, %r46;
|
|
add.s32 %r464, %r216, %r217;
|
|
mul.wide.s32 %rd70, %r43, 4;
|
|
add.s64 %rd71, %rd1, %rd70;
|
|
ld.local.u32 %r218, [%rd71];
|
|
shr.u32 %r219, %r218, %r215;
|
|
shl.b32 %r220, %r465, %r46;
|
|
add.s32 %r465, %r219, %r220;
|
|
|
|
$L__BB0_47:
|
|
and.b32 %r221, %r39, -2147483648;
|
|
shr.u32 %r222, %r465, 30;
|
|
shl.b32 %r223, %r464, 2;
|
|
or.b32 %r224, %r222, %r223;
|
|
shr.u32 %r225, %r224, 31;
|
|
shr.u32 %r226, %r464, 30;
|
|
add.s32 %r227, %r225, %r226;
|
|
neg.s32 %r228, %r227;
|
|
setp.eq.s32 %p59, %r221, 0;
|
|
selp.b32 %r466, %r227, %r228, %p59;
|
|
setp.ne.s32 %p60, %r225, 0;
|
|
xor.b32 %r229, %r221, -2147483648;
|
|
selp.b32 %r230, %r229, %r221, %p60;
|
|
selp.b32 %r231, -1, 0, %p60;
|
|
xor.b32 %r232, %r224, %r231;
|
|
shl.b32 %r233, %r465, 2;
|
|
xor.b32 %r234, %r233, %r231;
|
|
cvt.u64.u32 %rd72, %r232;
|
|
cvt.u64.u32 %rd73, %r234;
|
|
bfi.b64 %rd74, %rd72, %rd73, 32, 32;
|
|
cvt.rn.f64.s64 %fd3, %rd74;
|
|
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f676, %fd4;
|
|
setp.eq.s32 %p61, %r230, 0;
|
|
neg.f32 %f677, %f676;
|
|
selp.f32 %f1584, %f676, %f677, %p61;
|
|
|
|
$L__BB0_49:
|
|
mul.f32 %f198, %f181, %f1583;
|
|
and.b32 %r53, %r466, 1;
|
|
setp.eq.s32 %p62, %r53, 0;
|
|
selp.f32 %f199, %f1584, 0f3F800000, %p62;
|
|
mul.rn.f32 %f200, %f1584, %f1584;
|
|
mov.f32 %f1585, 0fB94D4153;
|
|
@%p62 bra $L__BB0_51;
|
|
|
|
mov.f32 %f680, 0fBAB607ED;
|
|
mov.f32 %f681, 0f37CBAC00;
|
|
fma.rn.f32 %f1585, %f681, %f200, %f680;
|
|
|
|
$L__BB0_51:
|
|
selp.f32 %f682, 0f3C0885E4, 0f3D2AAABB, %p62;
|
|
fma.rn.f32 %f683, %f1585, %f200, %f682;
|
|
selp.f32 %f684, 0fBE2AAAA8, 0fBEFFFFFF, %p62;
|
|
fma.rn.f32 %f685, %f683, %f200, %f684;
|
|
mov.f32 %f686, 0f00000000;
|
|
fma.rn.f32 %f687, %f200, %f199, %f686;
|
|
fma.rn.f32 %f1586, %f685, %f687, %f199;
|
|
and.b32 %r236, %r466, 2;
|
|
setp.eq.s32 %p64, %r236, 0;
|
|
@%p64 bra $L__BB0_53;
|
|
|
|
mov.f32 %f689, 0fBF800000;
|
|
fma.rn.f32 %f1586, %f1586, %f689, %f686;
|
|
|
|
$L__BB0_53:
|
|
mul.f32 %f699, %f181, %f1586;
|
|
mul.f32 %f700, %f139, %f699;
|
|
mul.f32 %f701, %f140, %f699;
|
|
mul.f32 %f702, %f141, %f699;
|
|
fma.rn.f32 %f703, %f148, %f198, %f700;
|
|
fma.rn.f32 %f704, %f147, %f198, %f701;
|
|
fma.rn.f32 %f705, %f146, %f198, %f702;
|
|
fma.rn.f32 %f693, %f1481, %f180, %f703;
|
|
fma.rn.f32 %f694, %f1482, %f180, %f704;
|
|
fma.rn.f32 %f695, %f1483, %f180, %f705;
|
|
mov.f32 %f711, 0f38D1B717;
|
|
max.f32 %f696, %f710, %f711;
|
|
mov.f32 %f697, 0f6C4ECB8F;
|
|
mov.u32 %r270, 1;
|
|
mov.u32 %r273, 2;
|
|
mov.u32 %r275, 3;
|
|
mov.u32 %r307, 0;
|
|
// begin inline asm
|
|
call(%r237,%r238,%r239,%r240,%r241,%r242,%r243,%r244,%r245,%r246,%r247,%r248,%r249,%r250,%r251,%r252,%r253,%r254,%r255,%r256,%r257,%r258,%r259,%r260,%r261,%r262,%r263,%r264,%r265,%r266,%r267,%r268),_optix_trace_typed_32,(%r307,%rd7,%f15,%f16,%f17,%f693,%f694,%f695,%f696,%f697,%f686,%r270,%r307,%r307,%r273,%r307,%r275,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307,%r307);
|
|
// end inline asm
|
|
mov.b32 %f712, %r237;
|
|
mov.b32 %f713, %r238;
|
|
mov.b32 %f714, %r239;
|
|
max.f32 %f715, %f712, %f686;
|
|
max.f32 %f716, %f713, %f686;
|
|
max.f32 %f717, %f714, %f686;
|
|
fma.rn.f32 %f1574, %f693, %f715, %f1574;
|
|
fma.rn.f32 %f1573, %f693, %f716, %f1573;
|
|
fma.rn.f32 %f1572, %f693, %f717, %f1572;
|
|
fma.rn.f32 %f1577, %f694, %f715, %f1577;
|
|
fma.rn.f32 %f1576, %f694, %f716, %f1576;
|
|
fma.rn.f32 %f1575, %f694, %f717, %f1575;
|
|
fma.rn.f32 %f1580, %f695, %f715, %f1580;
|
|
fma.rn.f32 %f1579, %f695, %f716, %f1579;
|
|
fma.rn.f32 %f1578, %f695, %f717, %f1578;
|
|
add.f32 %f1571, %f1571, %f715;
|
|
add.f32 %f1570, %f1570, %f716;
|
|
add.f32 %f1569, %f1569, %f717;
|
|
mul.f32 %f718, %f1482, %f694;
|
|
fma.rn.f32 %f719, %f1481, %f693, %f718;
|
|
fma.rn.f32 %f720, %f1483, %f695, %f719;
|
|
cvt.sat.f32.f32 %f721, %f720;
|
|
fma.rn.f32 %f1568, %f715, %f721, %f1568;
|
|
fma.rn.f32 %f1567, %f716, %f721, %f1567;
|
|
fma.rn.f32 %f1566, %f717, %f721, %f1566;
|
|
add.s32 %r458, %r458, 1;
|
|
setp.lt.s32 %p65, %r458, %r5;
|
|
@%p65 bra $L__BB0_29;
|
|
|
|
add.s32 %r456, %r456, 1;
|
|
setp.lt.s32 %p66, %r456, %r5;
|
|
@%p66 bra $L__BB0_28;
|
|
|
|
$L__BB0_55:
|
|
mul.lo.s32 %r308, %r5, %r5;
|
|
cvt.rn.f32.s32 %f722, %r308;
|
|
rcp.rn.f32 %f723, %f722;
|
|
mul.f32 %f724, %f723, %f1568;
|
|
mul.f32 %f725, %f723, %f1567;
|
|
mul.f32 %f726, %f723, %f1566;
|
|
div.rn.f32 %f1617, %f611, %f722;
|
|
mul.f32 %f1611, %f723, %f1571;
|
|
mul.f32 %f1612, %f723, %f1570;
|
|
mul.f32 %f1613, %f723, %f1569;
|
|
mul.f32 %f1608, %f723, %f1574;
|
|
mul.f32 %f1609, %f723, %f1573;
|
|
mul.f32 %f1610, %f723, %f1572;
|
|
mul.f32 %f1605, %f723, %f1577;
|
|
mul.f32 %f1606, %f723, %f1576;
|
|
mul.f32 %f1607, %f723, %f1575;
|
|
mul.f32 %f1602, %f723, %f1580;
|
|
mul.f32 %f1603, %f723, %f1579;
|
|
mul.f32 %f1604, %f723, %f1578;
|
|
fma.rn.f32 %f1614, %f723, %f1568, %f724;
|
|
fma.rn.f32 %f1615, %f723, %f1567, %f725;
|
|
fma.rn.f32 %f1616, %f723, %f1566, %f726;
|
|
|
|
$L__BB0_56:
|
|
ld.const.u32 %r451, [params+616];
|
|
setp.lt.s32 %p67, %r451, 0;
|
|
selp.f32 %f268, %f1614, %f1617, %p67;
|
|
ld.const.u32 %r56, [params+104];
|
|
and.b32 %r309, %r56, 8;
|
|
setp.eq.s32 %p68, %r309, 0;
|
|
@%p68 bra $L__BB0_70;
|
|
|
|
ld.const.u64 %rd76, [params+192];
|
|
cvta.to.global.u64 %rd19, %rd76;
|
|
ld.const.u32 %r310, [params+184];
|
|
mad.lo.s32 %r311, %r310, %r7, %r6;
|
|
cvt.u64.u32 %rd20, %r311;
|
|
mov.f32 %f729, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f730, %f729;
|
|
add.f32 %f731, %f730, %f730;
|
|
mov.f32 %f732, 0f3EE8BA2E;
|
|
sub.f32 %f733, %f732, %f731;
|
|
abs.f32 %f269, %f733;
|
|
abs.f32 %f270, %f268;
|
|
setp.lt.f32 %p69, %f270, 0f00800000;
|
|
mul.f32 %f734, %f270, 0f4B800000;
|
|
selp.f32 %f735, %f734, %f270, %p69;
|
|
selp.f32 %f736, 0fC3170000, 0fC2FE0000, %p69;
|
|
mov.b32 %r312, %f735;
|
|
and.b32 %r313, %r312, 8388607;
|
|
or.b32 %r314, %r313, 1065353216;
|
|
mov.b32 %f737, %r314;
|
|
shr.u32 %r315, %r312, 23;
|
|
cvt.rn.f32.u32 %f738, %r315;
|
|
add.f32 %f739, %f736, %f738;
|
|
setp.gt.f32 %p70, %f737, 0f3FB504F3;
|
|
mul.f32 %f740, %f737, 0f3F000000;
|
|
add.f32 %f741, %f739, 0f3F800000;
|
|
selp.f32 %f742, %f741, %f739, %p70;
|
|
selp.f32 %f743, %f740, %f737, %p70;
|
|
add.f32 %f744, %f743, 0fBF800000;
|
|
add.f32 %f745, %f743, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f746, %f745;
|
|
add.f32 %f747, %f744, %f744;
|
|
mul.f32 %f748, %f747, %f746;
|
|
mul.f32 %f749, %f748, %f748;
|
|
mov.f32 %f750, 0f3C4CAF63;
|
|
mov.f32 %f751, 0f3B18F0FE;
|
|
fma.rn.f32 %f752, %f751, %f749, %f750;
|
|
mov.f32 %f753, 0f3DAAAABD;
|
|
fma.rn.f32 %f754, %f752, %f749, %f753;
|
|
mul.rn.f32 %f755, %f754, %f749;
|
|
mul.rn.f32 %f756, %f755, %f748;
|
|
sub.f32 %f757, %f744, %f748;
|
|
add.f32 %f758, %f757, %f757;
|
|
neg.f32 %f759, %f748;
|
|
fma.rn.f32 %f760, %f759, %f744, %f758;
|
|
mul.rn.f32 %f761, %f746, %f760;
|
|
add.f32 %f762, %f756, %f748;
|
|
sub.f32 %f763, %f748, %f762;
|
|
add.f32 %f764, %f756, %f763;
|
|
add.f32 %f765, %f761, %f764;
|
|
add.f32 %f766, %f762, %f765;
|
|
sub.f32 %f767, %f762, %f766;
|
|
add.f32 %f768, %f765, %f767;
|
|
mov.f32 %f769, 0f3F317200;
|
|
mul.rn.f32 %f770, %f742, %f769;
|
|
mov.f32 %f771, 0f35BFBE8E;
|
|
mul.rn.f32 %f772, %f742, %f771;
|
|
add.f32 %f773, %f770, %f766;
|
|
sub.f32 %f774, %f770, %f773;
|
|
add.f32 %f775, %f766, %f774;
|
|
add.f32 %f776, %f768, %f775;
|
|
add.f32 %f777, %f772, %f776;
|
|
add.f32 %f778, %f773, %f777;
|
|
sub.f32 %f779, %f773, %f778;
|
|
add.f32 %f780, %f777, %f779;
|
|
mul.rn.f32 %f781, %f732, %f778;
|
|
neg.f32 %f782, %f781;
|
|
fma.rn.f32 %f783, %f732, %f778, %f782;
|
|
fma.rn.f32 %f784, %f732, %f780, %f783;
|
|
mov.f32 %f785, 0f00000000;
|
|
fma.rn.f32 %f786, %f785, %f778, %f784;
|
|
add.rn.f32 %f787, %f781, %f786;
|
|
neg.f32 %f788, %f787;
|
|
add.rn.f32 %f789, %f781, %f788;
|
|
add.rn.f32 %f790, %f789, %f786;
|
|
mov.b32 %r316, %f787;
|
|
setp.eq.s32 %p71, %r316, 1118925336;
|
|
add.s32 %r317, %r316, -1;
|
|
mov.b32 %f791, %r317;
|
|
add.f32 %f792, %f790, 0f37000000;
|
|
selp.f32 %f271, %f792, %f790, %p71;
|
|
selp.f32 %f793, %f791, %f787, %p71;
|
|
mov.f32 %f794, 0f3FB8AA3B;
|
|
mul.rn.f32 %f795, %f793, %f794;
|
|
cvt.rzi.f32.f32 %f796, %f795;
|
|
abs.f32 %f797, %f796;
|
|
setp.gt.f32 %p72, %f797, 0f42FC0000;
|
|
mov.b32 %r318, %f796;
|
|
and.b32 %r319, %r318, -2147483648;
|
|
or.b32 %r320, %r319, 1123811328;
|
|
mov.b32 %f798, %r320;
|
|
selp.f32 %f799, %f798, %f796, %p72;
|
|
mov.f32 %f800, 0fBF317218;
|
|
fma.rn.f32 %f801, %f799, %f800, %f793;
|
|
mov.f32 %f802, 0f3102E308;
|
|
fma.rn.f32 %f803, %f799, %f802, %f801;
|
|
mul.f32 %f804, %f803, 0f3FB8AA3B;
|
|
add.f32 %f805, %f799, 0f4B40007F;
|
|
mov.b32 %r321, %f805;
|
|
shl.b32 %r322, %r321, 23;
|
|
mov.b32 %f806, %r322;
|
|
ex2.approx.ftz.f32 %f807, %f804;
|
|
mul.f32 %f272, %f807, %f806;
|
|
setp.eq.f32 %p73, %f272, 0f7F800000;
|
|
mov.f32 %f1618, 0f7F800000;
|
|
@%p73 bra $L__BB0_59;
|
|
|
|
fma.rn.f32 %f1618, %f272, %f271, %f272;
|
|
|
|
$L__BB0_59:
|
|
setp.lt.f32 %p74, %f268, 0f00000000;
|
|
setp.eq.f32 %p75, %f269, 0f3F800000;
|
|
and.pred %p3, %p74, %p75;
|
|
setp.eq.f32 %p76, %f268, 0f00000000;
|
|
@%p76 bra $L__BB0_63;
|
|
bra.uni $L__BB0_60;
|
|
|
|
$L__BB0_63:
|
|
add.f32 %f812, %f268, %f268;
|
|
selp.f32 %f1620, %f812, 0f00000000, %p75;
|
|
bra.uni $L__BB0_64;
|
|
|
|
$L__BB0_167:
|
|
mov.f32 %f1436, 0f00000000;
|
|
mov.u32 %r468, 0;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f1436;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f1436;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f1436;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs114, 0;
|
|
st.global.v4.u16 [%rd28], {%rs111, %rs112, %rs113, %rs114};
|
|
|
|
$L__BB0_168:
|
|
ld.const.u64 %rd108, [params+256];
|
|
cvta.to.global.u64 %rd109, %rd108;
|
|
ld.const.u32 %r443, [params+248];
|
|
mad.lo.s32 %r444, %r443, %r7, %r6;
|
|
mul.wide.u32 %rd110, %r444, 8;
|
|
add.s64 %rd29, %rd109, %rd110;
|
|
setp.eq.s32 %p194, %r468, 0;
|
|
@%p194 bra $L__BB0_170;
|
|
|
|
ld.global.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd29];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1437, %rs121;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1438, %rs122;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1439, %rs123;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1440, %f1437, 0f00000000;
|
|
add.f32 %f1441, %f1438, 0f00000000;
|
|
add.f32 %f1442, %f1439, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f1442;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs119, %f1441;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f1440;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs125, 0;
|
|
st.global.v4.u16 [%rd29], {%rs118, %rs119, %rs120, %rs125};
|
|
bra.uni $L__BB0_171;
|
|
|
|
$L__BB0_170:
|
|
mov.f32 %f1445, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f1445;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f1445;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f1445;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs129, 0;
|
|
st.global.v4.u16 [%rd29], {%rs126, %rs127, %rs128, %rs129};
|
|
|
|
$L__BB0_171:
|
|
ld.const.u64 %rd111, [params+272];
|
|
cvta.to.global.u64 %rd112, %rd111;
|
|
ld.const.u32 %r445, [params+264];
|
|
mad.lo.s32 %r446, %r445, %r7, %r6;
|
|
mul.wide.u32 %rd113, %r446, 8;
|
|
add.s64 %rd30, %rd112, %rd113;
|
|
@%p194 bra $L__BB0_173;
|
|
|
|
ld.global.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd30];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1446, %rs136;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1447, %rs137;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1448, %rs138;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1449, %f1446, 0f00000000;
|
|
add.f32 %f1450, %f1447, 0f00000000;
|
|
add.f32 %f1451, %f1448, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f1451;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f1450;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f1449;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs140, 0;
|
|
st.global.v4.u16 [%rd30], {%rs133, %rs134, %rs135, %rs140};
|
|
bra.uni $L__BB0_174;
|
|
|
|
$L__BB0_173:
|
|
mov.f32 %f1454, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f1454;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f1454;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs141, %f1454;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs144, 0;
|
|
st.global.v4.u16 [%rd30], {%rs141, %rs142, %rs143, %rs144};
|
|
|
|
$L__BB0_174:
|
|
ld.const.u64 %rd114, [params+288];
|
|
cvta.to.global.u64 %rd115, %rd114;
|
|
ld.const.u32 %r447, [params+280];
|
|
mad.lo.s32 %r448, %r447, %r7, %r6;
|
|
mul.wide.u32 %rd116, %r448, 8;
|
|
add.s64 %rd31, %rd115, %rd116;
|
|
@%p194 bra $L__BB0_176;
|
|
|
|
ld.global.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd31];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1455, %rs151;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1456, %rs152;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1457, %rs153;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1458, %f1455, 0f00000000;
|
|
add.f32 %f1459, %f1456, 0f00000000;
|
|
add.f32 %f1460, %f1457, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f1460;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f1459;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs148, %f1458;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs155, 0;
|
|
st.global.v4.u16 [%rd31], {%rs148, %rs149, %rs150, %rs155};
|
|
bra.uni $L__BB0_177;
|
|
|
|
$L__BB0_176:
|
|
mov.f32 %f1463, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs158, %f1463;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f1463;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs156, %f1463;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs159, 0;
|
|
st.global.v4.u16 [%rd31], {%rs156, %rs157, %rs158, %rs159};
|
|
|
|
$L__BB0_177:
|
|
ld.const.u64 %rd117, [params+304];
|
|
cvta.to.global.u64 %rd118, %rd117;
|
|
ld.const.u32 %r449, [params+296];
|
|
mad.lo.s32 %r450, %r449, %r7, %r6;
|
|
mul.wide.u32 %rd119, %r450, 8;
|
|
add.s64 %rd32, %rd118, %rd119;
|
|
@%p194 bra $L__BB0_179;
|
|
|
|
ld.global.v4.u16 {%rs166, %rs167, %rs168, %rs169}, [%rd32];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1464, %rs166;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1465, %rs167;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1466, %rs168;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1467, %f1464, 0f00000000;
|
|
add.f32 %f1468, %f1465, 0f00000000;
|
|
add.f32 %f1469, %f1466, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs165, %f1469;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs164, %f1468;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs163, %f1467;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs170, 0;
|
|
st.global.v4.u16 [%rd32], {%rs163, %rs164, %rs165, %rs170};
|
|
bra.uni $L__BB0_180;
|
|
|
|
$L__BB0_179:
|
|
mov.f32 %f1472, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f1472;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f1472;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs171, %f1472;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs174, 0;
|
|
st.global.v4.u16 [%rd32], {%rs171, %rs172, %rs173, %rs174};
|
|
bra.uni $L__BB0_180;
|
|
|
|
$L__BB0_60:
|
|
mov.b32 %r323, %f1618;
|
|
xor.b32 %r324, %r323, -2147483648;
|
|
mov.b32 %f808, %r324;
|
|
selp.f32 %f1620, %f808, %f1618, %p3;
|
|
setp.geu.f32 %p77, %f268, 0f00000000;
|
|
@%p77 bra $L__BB0_64;
|
|
|
|
mov.f32 %f809, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f810, %f809;
|
|
setp.eq.f32 %p78, %f810, 0f3EE8BA2E;
|
|
@%p78 bra $L__BB0_64;
|
|
|
|
mov.f32 %f1620, 0f7FFFFFFF;
|
|
|
|
$L__BB0_64:
|
|
add.f32 %f813, %f270, 0f3EE8BA2E;
|
|
mov.b32 %r325, %f813;
|
|
setp.lt.s32 %p80, %r325, 2139095040;
|
|
@%p80 bra $L__BB0_69;
|
|
|
|
setp.gtu.f32 %p81, %f270, 0f7F800000;
|
|
@%p81 bra $L__BB0_68;
|
|
bra.uni $L__BB0_66;
|
|
|
|
$L__BB0_68:
|
|
add.f32 %f1620, %f268, 0f3EE8BA2E;
|
|
bra.uni $L__BB0_69;
|
|
|
|
$L__BB0_66:
|
|
setp.neu.f32 %p82, %f270, 0f7F800000;
|
|
@%p82 bra $L__BB0_69;
|
|
|
|
selp.f32 %f1620, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
$L__BB0_69:
|
|
mul.f32 %f814, %f1620, 0f437F0000;
|
|
setp.eq.f32 %p83, %f268, 0f3F800000;
|
|
selp.f32 %f815, 0f437F0000, %f814, %p83;
|
|
cvt.rzi.u32.f32 %r326, %f815;
|
|
shl.b64 %rd77, %rd20, 1;
|
|
add.s64 %rd78, %rd19, %rd77;
|
|
cvt.u16.u32 %rs16, %r326;
|
|
mov.u16 %rs17, 255;
|
|
st.global.v2.u8 [%rd78], {%rs16, %rs17};
|
|
|
|
$L__BB0_70:
|
|
ld.const.v2.f32 {%f816, %f817}, [params+648];
|
|
mul.f32 %f283, %f1614, %f816;
|
|
mul.f32 %f284, %f1615, %f817;
|
|
ld.const.f32 %f285, [params+656];
|
|
mul.f32 %f286, %f1616, %f285;
|
|
and.b32 %r327, %r56, 1;
|
|
setp.eq.b32 %p84, %r327, 1;
|
|
mov.pred %p85, 0;
|
|
xor.pred %p86, %p84, %p85;
|
|
not.pred %p87, %p86;
|
|
@%p87 bra $L__BB0_144;
|
|
|
|
mov.f32 %f819, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f820, %f819;
|
|
add.f32 %f821, %f820, %f820;
|
|
mov.f32 %f822, 0f3EE66666;
|
|
sub.f32 %f823, %f822, %f821;
|
|
abs.f32 %f287, %f823;
|
|
abs.f32 %f288, %f283;
|
|
setp.lt.f32 %p88, %f288, 0f00800000;
|
|
mul.f32 %f824, %f288, 0f4B800000;
|
|
selp.f32 %f825, %f824, %f288, %p88;
|
|
selp.f32 %f826, 0fC3170000, 0fC2FE0000, %p88;
|
|
mov.b32 %r328, %f825;
|
|
and.b32 %r329, %r328, 8388607;
|
|
or.b32 %r330, %r329, 1065353216;
|
|
mov.b32 %f827, %r330;
|
|
shr.u32 %r331, %r328, 23;
|
|
cvt.rn.f32.u32 %f828, %r331;
|
|
add.f32 %f829, %f826, %f828;
|
|
setp.gt.f32 %p89, %f827, 0f3FB504F3;
|
|
mul.f32 %f830, %f827, 0f3F000000;
|
|
add.f32 %f831, %f829, 0f3F800000;
|
|
selp.f32 %f832, %f831, %f829, %p89;
|
|
selp.f32 %f833, %f830, %f827, %p89;
|
|
add.f32 %f834, %f833, 0fBF800000;
|
|
add.f32 %f835, %f833, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f836, %f835;
|
|
add.f32 %f837, %f834, %f834;
|
|
mul.f32 %f838, %f837, %f836;
|
|
mul.f32 %f839, %f838, %f838;
|
|
mov.f32 %f840, 0f3C4CAF63;
|
|
mov.f32 %f841, 0f3B18F0FE;
|
|
fma.rn.f32 %f842, %f841, %f839, %f840;
|
|
mov.f32 %f843, 0f3DAAAABD;
|
|
fma.rn.f32 %f844, %f842, %f839, %f843;
|
|
mul.rn.f32 %f845, %f844, %f839;
|
|
mul.rn.f32 %f846, %f845, %f838;
|
|
sub.f32 %f847, %f834, %f838;
|
|
add.f32 %f848, %f847, %f847;
|
|
neg.f32 %f849, %f838;
|
|
fma.rn.f32 %f850, %f849, %f834, %f848;
|
|
mul.rn.f32 %f851, %f836, %f850;
|
|
add.f32 %f852, %f846, %f838;
|
|
sub.f32 %f853, %f838, %f852;
|
|
add.f32 %f854, %f846, %f853;
|
|
add.f32 %f855, %f851, %f854;
|
|
add.f32 %f856, %f852, %f855;
|
|
sub.f32 %f857, %f852, %f856;
|
|
add.f32 %f858, %f855, %f857;
|
|
mov.f32 %f859, 0f3F317200;
|
|
mul.rn.f32 %f860, %f832, %f859;
|
|
mov.f32 %f861, 0f35BFBE8E;
|
|
mul.rn.f32 %f862, %f832, %f861;
|
|
add.f32 %f863, %f860, %f856;
|
|
sub.f32 %f864, %f860, %f863;
|
|
add.f32 %f865, %f856, %f864;
|
|
add.f32 %f866, %f858, %f865;
|
|
add.f32 %f867, %f862, %f866;
|
|
add.f32 %f868, %f863, %f867;
|
|
sub.f32 %f869, %f863, %f868;
|
|
add.f32 %f870, %f867, %f869;
|
|
mul.rn.f32 %f871, %f822, %f868;
|
|
neg.f32 %f872, %f871;
|
|
fma.rn.f32 %f873, %f822, %f868, %f872;
|
|
fma.rn.f32 %f874, %f822, %f870, %f873;
|
|
mov.f32 %f875, 0f00000000;
|
|
fma.rn.f32 %f876, %f875, %f868, %f874;
|
|
add.rn.f32 %f877, %f871, %f876;
|
|
neg.f32 %f878, %f877;
|
|
add.rn.f32 %f879, %f871, %f878;
|
|
add.rn.f32 %f880, %f879, %f876;
|
|
mov.b32 %r332, %f877;
|
|
setp.eq.s32 %p90, %r332, 1118925336;
|
|
add.s32 %r333, %r332, -1;
|
|
mov.b32 %f881, %r333;
|
|
add.f32 %f882, %f880, 0f37000000;
|
|
selp.f32 %f289, %f882, %f880, %p90;
|
|
selp.f32 %f883, %f881, %f877, %p90;
|
|
mov.f32 %f884, 0f3FB8AA3B;
|
|
mul.rn.f32 %f885, %f883, %f884;
|
|
cvt.rzi.f32.f32 %f886, %f885;
|
|
abs.f32 %f887, %f886;
|
|
setp.gt.f32 %p91, %f887, 0f42FC0000;
|
|
mov.b32 %r334, %f886;
|
|
and.b32 %r335, %r334, -2147483648;
|
|
or.b32 %r336, %r335, 1123811328;
|
|
mov.b32 %f888, %r336;
|
|
selp.f32 %f889, %f888, %f886, %p91;
|
|
mov.f32 %f890, 0fBF317218;
|
|
fma.rn.f32 %f891, %f889, %f890, %f883;
|
|
mov.f32 %f892, 0f3102E308;
|
|
fma.rn.f32 %f893, %f889, %f892, %f891;
|
|
mul.f32 %f894, %f893, 0f3FB8AA3B;
|
|
add.f32 %f895, %f889, 0f4B40007F;
|
|
mov.b32 %r337, %f895;
|
|
shl.b32 %r338, %r337, 23;
|
|
mov.b32 %f896, %r338;
|
|
ex2.approx.ftz.f32 %f897, %f894;
|
|
mul.f32 %f290, %f897, %f896;
|
|
setp.eq.f32 %p92, %f290, 0f7F800000;
|
|
mov.f32 %f1621, 0f7F800000;
|
|
@%p92 bra $L__BB0_73;
|
|
|
|
fma.rn.f32 %f1621, %f290, %f289, %f290;
|
|
|
|
$L__BB0_73:
|
|
setp.lt.f32 %p93, %f283, 0f00000000;
|
|
setp.eq.f32 %p94, %f287, 0f3F800000;
|
|
and.pred %p4, %p93, %p94;
|
|
setp.eq.f32 %p95, %f283, 0f00000000;
|
|
@%p95 bra $L__BB0_77;
|
|
bra.uni $L__BB0_74;
|
|
|
|
$L__BB0_77:
|
|
add.f32 %f902, %f283, %f283;
|
|
selp.f32 %f1623, %f902, 0f00000000, %p94;
|
|
bra.uni $L__BB0_78;
|
|
|
|
$L__BB0_74:
|
|
mov.b32 %r339, %f1621;
|
|
xor.b32 %r340, %r339, -2147483648;
|
|
mov.b32 %f898, %r340;
|
|
selp.f32 %f1623, %f898, %f1621, %p4;
|
|
setp.geu.f32 %p96, %f283, 0f00000000;
|
|
@%p96 bra $L__BB0_78;
|
|
|
|
mov.f32 %f899, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f900, %f899;
|
|
setp.eq.f32 %p97, %f900, 0f3EE66666;
|
|
@%p97 bra $L__BB0_78;
|
|
|
|
mov.f32 %f1623, 0f7FFFFFFF;
|
|
|
|
$L__BB0_78:
|
|
add.f32 %f903, %f288, 0f3EE66666;
|
|
mov.b32 %r341, %f903;
|
|
setp.lt.s32 %p99, %r341, 2139095040;
|
|
@%p99 bra $L__BB0_83;
|
|
|
|
setp.gtu.f32 %p100, %f288, 0f7F800000;
|
|
@%p100 bra $L__BB0_82;
|
|
bra.uni $L__BB0_80;
|
|
|
|
$L__BB0_82:
|
|
add.f32 %f1623, %f283, 0f3EE66666;
|
|
bra.uni $L__BB0_83;
|
|
|
|
$L__BB0_80:
|
|
setp.neu.f32 %p101, %f288, 0f7F800000;
|
|
@%p101 bra $L__BB0_83;
|
|
|
|
selp.f32 %f1623, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
$L__BB0_83:
|
|
setp.eq.f32 %p102, %f283, 0f3F800000;
|
|
selp.f32 %f299, 0f3F800000, %f1623, %p102;
|
|
abs.f32 %f300, %f284;
|
|
setp.lt.f32 %p103, %f300, 0f00800000;
|
|
mul.f32 %f905, %f300, 0f4B800000;
|
|
selp.f32 %f906, %f905, %f300, %p103;
|
|
selp.f32 %f907, 0fC3170000, 0fC2FE0000, %p103;
|
|
mov.b32 %r342, %f906;
|
|
and.b32 %r343, %r342, 8388607;
|
|
or.b32 %r344, %r343, 1065353216;
|
|
mov.b32 %f908, %r344;
|
|
shr.u32 %r345, %r342, 23;
|
|
cvt.rn.f32.u32 %f909, %r345;
|
|
add.f32 %f910, %f907, %f909;
|
|
setp.gt.f32 %p104, %f908, 0f3FB504F3;
|
|
mul.f32 %f911, %f908, 0f3F000000;
|
|
add.f32 %f912, %f910, 0f3F800000;
|
|
selp.f32 %f913, %f912, %f910, %p104;
|
|
selp.f32 %f914, %f911, %f908, %p104;
|
|
add.f32 %f915, %f914, 0fBF800000;
|
|
add.f32 %f916, %f914, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f917, %f916;
|
|
add.f32 %f918, %f915, %f915;
|
|
mul.f32 %f919, %f918, %f917;
|
|
mul.f32 %f920, %f919, %f919;
|
|
mov.f32 %f921, 0f3C4CAF63;
|
|
mov.f32 %f922, 0f3B18F0FE;
|
|
fma.rn.f32 %f923, %f922, %f920, %f921;
|
|
mov.f32 %f924, 0f3DAAAABD;
|
|
fma.rn.f32 %f925, %f923, %f920, %f924;
|
|
mul.rn.f32 %f926, %f925, %f920;
|
|
mul.rn.f32 %f927, %f926, %f919;
|
|
sub.f32 %f928, %f915, %f919;
|
|
add.f32 %f929, %f928, %f928;
|
|
neg.f32 %f930, %f919;
|
|
fma.rn.f32 %f931, %f930, %f915, %f929;
|
|
mul.rn.f32 %f932, %f917, %f931;
|
|
add.f32 %f933, %f927, %f919;
|
|
sub.f32 %f934, %f919, %f933;
|
|
add.f32 %f935, %f927, %f934;
|
|
add.f32 %f936, %f932, %f935;
|
|
add.f32 %f937, %f933, %f936;
|
|
sub.f32 %f938, %f933, %f937;
|
|
add.f32 %f939, %f936, %f938;
|
|
mov.f32 %f940, 0f3F317200;
|
|
mul.rn.f32 %f941, %f913, %f940;
|
|
mov.f32 %f942, 0f35BFBE8E;
|
|
mul.rn.f32 %f943, %f913, %f942;
|
|
add.f32 %f944, %f941, %f937;
|
|
sub.f32 %f945, %f941, %f944;
|
|
add.f32 %f946, %f937, %f945;
|
|
add.f32 %f947, %f939, %f946;
|
|
add.f32 %f948, %f943, %f947;
|
|
add.f32 %f949, %f944, %f948;
|
|
sub.f32 %f950, %f944, %f949;
|
|
add.f32 %f951, %f948, %f950;
|
|
mov.f32 %f952, 0f3EE66666;
|
|
mul.rn.f32 %f953, %f952, %f949;
|
|
neg.f32 %f954, %f953;
|
|
fma.rn.f32 %f955, %f952, %f949, %f954;
|
|
fma.rn.f32 %f956, %f952, %f951, %f955;
|
|
mov.f32 %f957, 0f00000000;
|
|
fma.rn.f32 %f958, %f957, %f949, %f956;
|
|
add.rn.f32 %f959, %f953, %f958;
|
|
neg.f32 %f960, %f959;
|
|
add.rn.f32 %f961, %f953, %f960;
|
|
add.rn.f32 %f962, %f961, %f958;
|
|
mov.b32 %r346, %f959;
|
|
setp.eq.s32 %p105, %r346, 1118925336;
|
|
add.s32 %r347, %r346, -1;
|
|
mov.b32 %f963, %r347;
|
|
add.f32 %f964, %f962, 0f37000000;
|
|
selp.f32 %f301, %f964, %f962, %p105;
|
|
selp.f32 %f965, %f963, %f959, %p105;
|
|
mov.f32 %f966, 0f3FB8AA3B;
|
|
mul.rn.f32 %f967, %f965, %f966;
|
|
cvt.rzi.f32.f32 %f968, %f967;
|
|
abs.f32 %f969, %f968;
|
|
setp.gt.f32 %p106, %f969, 0f42FC0000;
|
|
mov.b32 %r348, %f968;
|
|
and.b32 %r349, %r348, -2147483648;
|
|
or.b32 %r350, %r349, 1123811328;
|
|
mov.b32 %f970, %r350;
|
|
selp.f32 %f971, %f970, %f968, %p106;
|
|
mov.f32 %f972, 0fBF317218;
|
|
fma.rn.f32 %f973, %f971, %f972, %f965;
|
|
mov.f32 %f974, 0f3102E308;
|
|
fma.rn.f32 %f975, %f971, %f974, %f973;
|
|
mul.f32 %f976, %f975, 0f3FB8AA3B;
|
|
add.f32 %f977, %f971, 0f4B40007F;
|
|
mov.b32 %r351, %f977;
|
|
shl.b32 %r352, %r351, 23;
|
|
mov.b32 %f978, %r352;
|
|
ex2.approx.ftz.f32 %f979, %f976;
|
|
mul.f32 %f302, %f979, %f978;
|
|
setp.eq.f32 %p107, %f302, 0f7F800000;
|
|
mov.f32 %f1624, 0f7F800000;
|
|
@%p107 bra $L__BB0_85;
|
|
|
|
fma.rn.f32 %f1624, %f302, %f301, %f302;
|
|
|
|
$L__BB0_85:
|
|
setp.lt.f32 %p108, %f284, 0f00000000;
|
|
and.pred %p5, %p108, %p94;
|
|
setp.eq.f32 %p110, %f284, 0f00000000;
|
|
@%p110 bra $L__BB0_89;
|
|
bra.uni $L__BB0_86;
|
|
|
|
$L__BB0_89:
|
|
add.f32 %f984, %f284, %f284;
|
|
selp.f32 %f1626, %f984, 0f00000000, %p94;
|
|
bra.uni $L__BB0_90;
|
|
|
|
$L__BB0_86:
|
|
mov.b32 %r353, %f1624;
|
|
xor.b32 %r354, %r353, -2147483648;
|
|
mov.b32 %f980, %r354;
|
|
selp.f32 %f1626, %f980, %f1624, %p5;
|
|
setp.geu.f32 %p111, %f284, 0f00000000;
|
|
@%p111 bra $L__BB0_90;
|
|
|
|
mov.f32 %f981, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f982, %f981;
|
|
setp.eq.f32 %p112, %f982, 0f3EE66666;
|
|
@%p112 bra $L__BB0_90;
|
|
|
|
mov.f32 %f1626, 0f7FFFFFFF;
|
|
|
|
$L__BB0_90:
|
|
add.f32 %f985, %f300, 0f3EE66666;
|
|
mov.b32 %r355, %f985;
|
|
setp.lt.s32 %p114, %r355, 2139095040;
|
|
@%p114 bra $L__BB0_95;
|
|
|
|
setp.gtu.f32 %p115, %f300, 0f7F800000;
|
|
@%p115 bra $L__BB0_94;
|
|
bra.uni $L__BB0_92;
|
|
|
|
$L__BB0_94:
|
|
add.f32 %f1626, %f284, 0f3EE66666;
|
|
bra.uni $L__BB0_95;
|
|
|
|
$L__BB0_92:
|
|
setp.neu.f32 %p116, %f300, 0f7F800000;
|
|
@%p116 bra $L__BB0_95;
|
|
|
|
selp.f32 %f1626, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
$L__BB0_95:
|
|
setp.eq.f32 %p117, %f284, 0f3F800000;
|
|
selp.f32 %f311, 0f3F800000, %f1626, %p117;
|
|
abs.f32 %f312, %f286;
|
|
setp.lt.f32 %p118, %f312, 0f00800000;
|
|
mul.f32 %f987, %f312, 0f4B800000;
|
|
selp.f32 %f988, %f987, %f312, %p118;
|
|
selp.f32 %f989, 0fC3170000, 0fC2FE0000, %p118;
|
|
mov.b32 %r356, %f988;
|
|
and.b32 %r357, %r356, 8388607;
|
|
or.b32 %r358, %r357, 1065353216;
|
|
mov.b32 %f990, %r358;
|
|
shr.u32 %r359, %r356, 23;
|
|
cvt.rn.f32.u32 %f991, %r359;
|
|
add.f32 %f992, %f989, %f991;
|
|
setp.gt.f32 %p119, %f990, 0f3FB504F3;
|
|
mul.f32 %f993, %f990, 0f3F000000;
|
|
add.f32 %f994, %f992, 0f3F800000;
|
|
selp.f32 %f995, %f994, %f992, %p119;
|
|
selp.f32 %f996, %f993, %f990, %p119;
|
|
add.f32 %f997, %f996, 0fBF800000;
|
|
add.f32 %f998, %f996, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f999, %f998;
|
|
add.f32 %f1000, %f997, %f997;
|
|
mul.f32 %f1001, %f1000, %f999;
|
|
mul.f32 %f1002, %f1001, %f1001;
|
|
mov.f32 %f1003, 0f3C4CAF63;
|
|
mov.f32 %f1004, 0f3B18F0FE;
|
|
fma.rn.f32 %f1005, %f1004, %f1002, %f1003;
|
|
mov.f32 %f1006, 0f3DAAAABD;
|
|
fma.rn.f32 %f1007, %f1005, %f1002, %f1006;
|
|
mul.rn.f32 %f1008, %f1007, %f1002;
|
|
mul.rn.f32 %f1009, %f1008, %f1001;
|
|
sub.f32 %f1010, %f997, %f1001;
|
|
add.f32 %f1011, %f1010, %f1010;
|
|
neg.f32 %f1012, %f1001;
|
|
fma.rn.f32 %f1013, %f1012, %f997, %f1011;
|
|
mul.rn.f32 %f1014, %f999, %f1013;
|
|
add.f32 %f1015, %f1009, %f1001;
|
|
sub.f32 %f1016, %f1001, %f1015;
|
|
add.f32 %f1017, %f1009, %f1016;
|
|
add.f32 %f1018, %f1014, %f1017;
|
|
add.f32 %f1019, %f1015, %f1018;
|
|
sub.f32 %f1020, %f1015, %f1019;
|
|
add.f32 %f1021, %f1018, %f1020;
|
|
mov.f32 %f1022, 0f3F317200;
|
|
mul.rn.f32 %f1023, %f995, %f1022;
|
|
mov.f32 %f1024, 0f35BFBE8E;
|
|
mul.rn.f32 %f1025, %f995, %f1024;
|
|
add.f32 %f1026, %f1023, %f1019;
|
|
sub.f32 %f1027, %f1023, %f1026;
|
|
add.f32 %f1028, %f1019, %f1027;
|
|
add.f32 %f1029, %f1021, %f1028;
|
|
add.f32 %f1030, %f1025, %f1029;
|
|
add.f32 %f1031, %f1026, %f1030;
|
|
sub.f32 %f1032, %f1026, %f1031;
|
|
add.f32 %f1033, %f1030, %f1032;
|
|
mov.f32 %f1034, 0f3EE66666;
|
|
mul.rn.f32 %f1035, %f1034, %f1031;
|
|
neg.f32 %f1036, %f1035;
|
|
fma.rn.f32 %f1037, %f1034, %f1031, %f1036;
|
|
fma.rn.f32 %f1038, %f1034, %f1033, %f1037;
|
|
mov.f32 %f1039, 0f00000000;
|
|
fma.rn.f32 %f1040, %f1039, %f1031, %f1038;
|
|
add.rn.f32 %f1041, %f1035, %f1040;
|
|
neg.f32 %f1042, %f1041;
|
|
add.rn.f32 %f1043, %f1035, %f1042;
|
|
add.rn.f32 %f1044, %f1043, %f1040;
|
|
mov.b32 %r360, %f1041;
|
|
setp.eq.s32 %p120, %r360, 1118925336;
|
|
add.s32 %r361, %r360, -1;
|
|
mov.b32 %f1045, %r361;
|
|
add.f32 %f1046, %f1044, 0f37000000;
|
|
selp.f32 %f313, %f1046, %f1044, %p120;
|
|
selp.f32 %f1047, %f1045, %f1041, %p120;
|
|
mov.f32 %f1048, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1049, %f1047, %f1048;
|
|
cvt.rzi.f32.f32 %f1050, %f1049;
|
|
abs.f32 %f1051, %f1050;
|
|
setp.gt.f32 %p121, %f1051, 0f42FC0000;
|
|
mov.b32 %r362, %f1050;
|
|
and.b32 %r363, %r362, -2147483648;
|
|
or.b32 %r364, %r363, 1123811328;
|
|
mov.b32 %f1052, %r364;
|
|
selp.f32 %f1053, %f1052, %f1050, %p121;
|
|
mov.f32 %f1054, 0fBF317218;
|
|
fma.rn.f32 %f1055, %f1053, %f1054, %f1047;
|
|
mov.f32 %f1056, 0f3102E308;
|
|
fma.rn.f32 %f1057, %f1053, %f1056, %f1055;
|
|
mul.f32 %f1058, %f1057, 0f3FB8AA3B;
|
|
add.f32 %f1059, %f1053, 0f4B40007F;
|
|
mov.b32 %r365, %f1059;
|
|
shl.b32 %r366, %r365, 23;
|
|
mov.b32 %f1060, %r366;
|
|
ex2.approx.ftz.f32 %f1061, %f1058;
|
|
mul.f32 %f314, %f1061, %f1060;
|
|
setp.eq.f32 %p122, %f314, 0f7F800000;
|
|
mov.f32 %f1627, 0f7F800000;
|
|
@%p122 bra $L__BB0_97;
|
|
|
|
fma.rn.f32 %f1627, %f314, %f313, %f314;
|
|
|
|
$L__BB0_97:
|
|
setp.lt.f32 %p123, %f286, 0f00000000;
|
|
and.pred %p6, %p123, %p94;
|
|
setp.eq.f32 %p125, %f286, 0f00000000;
|
|
@%p125 bra $L__BB0_101;
|
|
bra.uni $L__BB0_98;
|
|
|
|
$L__BB0_101:
|
|
add.f32 %f1066, %f286, %f286;
|
|
selp.f32 %f1629, %f1066, 0f00000000, %p94;
|
|
bra.uni $L__BB0_102;
|
|
|
|
$L__BB0_98:
|
|
mov.b32 %r367, %f1627;
|
|
xor.b32 %r368, %r367, -2147483648;
|
|
mov.b32 %f1062, %r368;
|
|
selp.f32 %f1629, %f1062, %f1627, %p6;
|
|
setp.geu.f32 %p126, %f286, 0f00000000;
|
|
@%p126 bra $L__BB0_102;
|
|
|
|
mov.f32 %f1063, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f1064, %f1063;
|
|
setp.eq.f32 %p127, %f1064, 0f3EE66666;
|
|
@%p127 bra $L__BB0_102;
|
|
|
|
mov.f32 %f1629, 0f7FFFFFFF;
|
|
|
|
$L__BB0_102:
|
|
add.f32 %f1067, %f312, 0f3EE66666;
|
|
mov.b32 %r369, %f1067;
|
|
setp.lt.s32 %p129, %r369, 2139095040;
|
|
@%p129 bra $L__BB0_107;
|
|
|
|
setp.gtu.f32 %p130, %f312, 0f7F800000;
|
|
@%p130 bra $L__BB0_106;
|
|
bra.uni $L__BB0_104;
|
|
|
|
$L__BB0_106:
|
|
add.f32 %f1629, %f286, 0f3EE66666;
|
|
bra.uni $L__BB0_107;
|
|
|
|
$L__BB0_104:
|
|
setp.neu.f32 %p131, %f312, 0f7F800000;
|
|
@%p131 bra $L__BB0_107;
|
|
|
|
selp.f32 %f1629, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
$L__BB0_107:
|
|
setp.eq.f32 %p132, %f286, 0f3F800000;
|
|
mov.f32 %f1069, 0f3F800000;
|
|
selp.f32 %f1070, 0f3F800000, %f1629, %p132;
|
|
ld.const.u64 %rd79, [params+144];
|
|
cvta.to.global.u64 %rd21, %rd79;
|
|
ld.const.u32 %r370, [params+136];
|
|
mad.lo.s32 %r371, %r370, %r7, %r6;
|
|
cvt.u64.u32 %rd22, %r371;
|
|
min.f32 %f1071, %f299, %f1069;
|
|
mov.f32 %f1072, 0f00000000;
|
|
max.f32 %f323, %f1072, %f1071;
|
|
min.f32 %f1073, %f311, %f1069;
|
|
max.f32 %f324, %f1072, %f1073;
|
|
min.f32 %f1074, %f1070, %f1069;
|
|
max.f32 %f325, %f1072, %f1074;
|
|
mov.f32 %f1075, 0f3E555555;
|
|
cvt.rzi.f32.f32 %f1076, %f1075;
|
|
add.f32 %f1077, %f1076, %f1076;
|
|
mov.f32 %f1078, 0f3ED55555;
|
|
sub.f32 %f1079, %f1078, %f1077;
|
|
abs.f32 %f326, %f1079;
|
|
abs.f32 %f327, %f323;
|
|
setp.lt.f32 %p133, %f327, 0f00800000;
|
|
mul.f32 %f1080, %f327, 0f4B800000;
|
|
selp.f32 %f1081, %f1080, %f327, %p133;
|
|
selp.f32 %f1082, 0fC3170000, 0fC2FE0000, %p133;
|
|
mov.b32 %r372, %f1081;
|
|
and.b32 %r373, %r372, 8388607;
|
|
or.b32 %r374, %r373, 1065353216;
|
|
mov.b32 %f1083, %r374;
|
|
shr.u32 %r375, %r372, 23;
|
|
cvt.rn.f32.u32 %f1084, %r375;
|
|
add.f32 %f1085, %f1082, %f1084;
|
|
setp.gt.f32 %p134, %f1083, 0f3FB504F3;
|
|
mul.f32 %f1086, %f1083, 0f3F000000;
|
|
add.f32 %f1087, %f1085, 0f3F800000;
|
|
selp.f32 %f1088, %f1087, %f1085, %p134;
|
|
selp.f32 %f1089, %f1086, %f1083, %p134;
|
|
add.f32 %f1090, %f1089, 0fBF800000;
|
|
add.f32 %f1091, %f1089, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1092, %f1091;
|
|
add.f32 %f1093, %f1090, %f1090;
|
|
mul.f32 %f1094, %f1093, %f1092;
|
|
mul.f32 %f1095, %f1094, %f1094;
|
|
mov.f32 %f1096, 0f3C4CAF63;
|
|
mov.f32 %f1097, 0f3B18F0FE;
|
|
fma.rn.f32 %f1098, %f1097, %f1095, %f1096;
|
|
mov.f32 %f1099, 0f3DAAAABD;
|
|
fma.rn.f32 %f1100, %f1098, %f1095, %f1099;
|
|
mul.rn.f32 %f1101, %f1100, %f1095;
|
|
mul.rn.f32 %f1102, %f1101, %f1094;
|
|
sub.f32 %f1103, %f1090, %f1094;
|
|
add.f32 %f1104, %f1103, %f1103;
|
|
neg.f32 %f1105, %f1094;
|
|
fma.rn.f32 %f1106, %f1105, %f1090, %f1104;
|
|
mul.rn.f32 %f1107, %f1092, %f1106;
|
|
add.f32 %f1108, %f1102, %f1094;
|
|
sub.f32 %f1109, %f1094, %f1108;
|
|
add.f32 %f1110, %f1102, %f1109;
|
|
add.f32 %f1111, %f1107, %f1110;
|
|
add.f32 %f1112, %f1108, %f1111;
|
|
sub.f32 %f1113, %f1108, %f1112;
|
|
add.f32 %f1114, %f1111, %f1113;
|
|
mov.f32 %f1115, 0f3F317200;
|
|
mul.rn.f32 %f1116, %f1088, %f1115;
|
|
mov.f32 %f1117, 0f35BFBE8E;
|
|
mul.rn.f32 %f1118, %f1088, %f1117;
|
|
add.f32 %f1119, %f1116, %f1112;
|
|
sub.f32 %f1120, %f1116, %f1119;
|
|
add.f32 %f1121, %f1112, %f1120;
|
|
add.f32 %f1122, %f1114, %f1121;
|
|
add.f32 %f1123, %f1118, %f1122;
|
|
add.f32 %f1124, %f1119, %f1123;
|
|
sub.f32 %f1125, %f1119, %f1124;
|
|
add.f32 %f1126, %f1123, %f1125;
|
|
mul.rn.f32 %f1127, %f1078, %f1124;
|
|
neg.f32 %f1128, %f1127;
|
|
fma.rn.f32 %f1129, %f1078, %f1124, %f1128;
|
|
fma.rn.f32 %f1130, %f1078, %f1126, %f1129;
|
|
fma.rn.f32 %f1131, %f1072, %f1124, %f1130;
|
|
add.rn.f32 %f1132, %f1127, %f1131;
|
|
neg.f32 %f1133, %f1132;
|
|
add.rn.f32 %f1134, %f1127, %f1133;
|
|
add.rn.f32 %f1135, %f1134, %f1131;
|
|
mov.b32 %r376, %f1132;
|
|
setp.eq.s32 %p135, %r376, 1118925336;
|
|
add.s32 %r377, %r376, -1;
|
|
mov.b32 %f1136, %r377;
|
|
add.f32 %f1137, %f1135, 0f37000000;
|
|
selp.f32 %f328, %f1137, %f1135, %p135;
|
|
selp.f32 %f1138, %f1136, %f1132, %p135;
|
|
mov.f32 %f1139, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1140, %f1138, %f1139;
|
|
cvt.rzi.f32.f32 %f1141, %f1140;
|
|
abs.f32 %f1142, %f1141;
|
|
setp.gt.f32 %p136, %f1142, 0f42FC0000;
|
|
mov.b32 %r378, %f1141;
|
|
and.b32 %r379, %r378, -2147483648;
|
|
or.b32 %r380, %r379, 1123811328;
|
|
mov.b32 %f1143, %r380;
|
|
selp.f32 %f1144, %f1143, %f1141, %p136;
|
|
mov.f32 %f1145, 0fBF317218;
|
|
fma.rn.f32 %f1146, %f1144, %f1145, %f1138;
|
|
mov.f32 %f1147, 0f3102E308;
|
|
fma.rn.f32 %f1148, %f1144, %f1147, %f1146;
|
|
mul.f32 %f1149, %f1148, 0f3FB8AA3B;
|
|
add.f32 %f1150, %f1144, 0f4B40007F;
|
|
mov.b32 %r381, %f1150;
|
|
shl.b32 %r382, %r381, 23;
|
|
mov.b32 %f1151, %r382;
|
|
ex2.approx.ftz.f32 %f1152, %f1149;
|
|
mul.f32 %f329, %f1152, %f1151;
|
|
setp.eq.f32 %p137, %f329, 0f7F800000;
|
|
mov.f32 %f1630, 0f7F800000;
|
|
@%p137 bra $L__BB0_109;
|
|
|
|
fma.rn.f32 %f1630, %f329, %f328, %f329;
|
|
|
|
$L__BB0_109:
|
|
setp.lt.f32 %p138, %f323, 0f00000000;
|
|
setp.eq.f32 %p139, %f326, 0f3F800000;
|
|
and.pred %p7, %p138, %p139;
|
|
setp.eq.f32 %p140, %f323, 0f00000000;
|
|
@%p140 bra $L__BB0_113;
|
|
bra.uni $L__BB0_110;
|
|
|
|
$L__BB0_113:
|
|
add.f32 %f1157, %f323, %f323;
|
|
selp.f32 %f1632, %f1157, 0f00000000, %p139;
|
|
bra.uni $L__BB0_114;
|
|
|
|
$L__BB0_110:
|
|
mov.b32 %r383, %f1630;
|
|
xor.b32 %r384, %r383, -2147483648;
|
|
mov.b32 %f1153, %r384;
|
|
selp.f32 %f1632, %f1153, %f1630, %p7;
|
|
setp.geu.f32 %p141, %f323, 0f00000000;
|
|
@%p141 bra $L__BB0_114;
|
|
|
|
mov.f32 %f1154, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1155, %f1154;
|
|
setp.eq.f32 %p142, %f1155, 0f3ED55555;
|
|
@%p142 bra $L__BB0_114;
|
|
|
|
mov.f32 %f1632, 0f7FFFFFFF;
|
|
|
|
$L__BB0_114:
|
|
add.f32 %f1158, %f327, 0f3ED55555;
|
|
mov.b32 %r385, %f1158;
|
|
setp.lt.s32 %p144, %r385, 2139095040;
|
|
@%p144 bra $L__BB0_119;
|
|
|
|
setp.gtu.f32 %p145, %f327, 0f7F800000;
|
|
@%p145 bra $L__BB0_118;
|
|
bra.uni $L__BB0_116;
|
|
|
|
$L__BB0_118:
|
|
add.f32 %f1632, %f323, 0f3ED55555;
|
|
bra.uni $L__BB0_119;
|
|
|
|
$L__BB0_116:
|
|
setp.neu.f32 %p146, %f327, 0f7F800000;
|
|
@%p146 bra $L__BB0_119;
|
|
|
|
selp.f32 %f1632, 0fFF800000, 0f7F800000, %p7;
|
|
|
|
$L__BB0_119:
|
|
abs.f32 %f338, %f324;
|
|
setp.lt.f32 %p147, %f338, 0f00800000;
|
|
mul.f32 %f1160, %f338, 0f4B800000;
|
|
selp.f32 %f1161, %f1160, %f338, %p147;
|
|
selp.f32 %f1162, 0fC3170000, 0fC2FE0000, %p147;
|
|
mov.b32 %r386, %f1161;
|
|
and.b32 %r387, %r386, 8388607;
|
|
or.b32 %r388, %r387, 1065353216;
|
|
mov.b32 %f1163, %r388;
|
|
shr.u32 %r389, %r386, 23;
|
|
cvt.rn.f32.u32 %f1164, %r389;
|
|
add.f32 %f1165, %f1162, %f1164;
|
|
setp.gt.f32 %p148, %f1163, 0f3FB504F3;
|
|
mul.f32 %f1166, %f1163, 0f3F000000;
|
|
add.f32 %f1167, %f1165, 0f3F800000;
|
|
selp.f32 %f1168, %f1167, %f1165, %p148;
|
|
selp.f32 %f1169, %f1166, %f1163, %p148;
|
|
add.f32 %f1170, %f1169, 0fBF800000;
|
|
add.f32 %f1171, %f1169, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1172, %f1171;
|
|
add.f32 %f1173, %f1170, %f1170;
|
|
mul.f32 %f1174, %f1173, %f1172;
|
|
mul.f32 %f1175, %f1174, %f1174;
|
|
mov.f32 %f1176, 0f3C4CAF63;
|
|
mov.f32 %f1177, 0f3B18F0FE;
|
|
fma.rn.f32 %f1178, %f1177, %f1175, %f1176;
|
|
mov.f32 %f1179, 0f3DAAAABD;
|
|
fma.rn.f32 %f1180, %f1178, %f1175, %f1179;
|
|
mul.rn.f32 %f1181, %f1180, %f1175;
|
|
mul.rn.f32 %f1182, %f1181, %f1174;
|
|
sub.f32 %f1183, %f1170, %f1174;
|
|
add.f32 %f1184, %f1183, %f1183;
|
|
neg.f32 %f1185, %f1174;
|
|
fma.rn.f32 %f1186, %f1185, %f1170, %f1184;
|
|
mul.rn.f32 %f1187, %f1172, %f1186;
|
|
add.f32 %f1188, %f1182, %f1174;
|
|
sub.f32 %f1189, %f1174, %f1188;
|
|
add.f32 %f1190, %f1182, %f1189;
|
|
add.f32 %f1191, %f1187, %f1190;
|
|
add.f32 %f1192, %f1188, %f1191;
|
|
sub.f32 %f1193, %f1188, %f1192;
|
|
add.f32 %f1194, %f1191, %f1193;
|
|
mov.f32 %f1195, 0f3F317200;
|
|
mul.rn.f32 %f1196, %f1168, %f1195;
|
|
mov.f32 %f1197, 0f35BFBE8E;
|
|
mul.rn.f32 %f1198, %f1168, %f1197;
|
|
add.f32 %f1199, %f1196, %f1192;
|
|
sub.f32 %f1200, %f1196, %f1199;
|
|
add.f32 %f1201, %f1192, %f1200;
|
|
add.f32 %f1202, %f1194, %f1201;
|
|
add.f32 %f1203, %f1198, %f1202;
|
|
add.f32 %f1204, %f1199, %f1203;
|
|
sub.f32 %f1205, %f1199, %f1204;
|
|
add.f32 %f1206, %f1203, %f1205;
|
|
mov.f32 %f1207, 0f3ED55555;
|
|
mul.rn.f32 %f1208, %f1207, %f1204;
|
|
neg.f32 %f1209, %f1208;
|
|
fma.rn.f32 %f1210, %f1207, %f1204, %f1209;
|
|
fma.rn.f32 %f1211, %f1207, %f1206, %f1210;
|
|
mov.f32 %f1212, 0f00000000;
|
|
fma.rn.f32 %f1213, %f1212, %f1204, %f1211;
|
|
add.rn.f32 %f1214, %f1208, %f1213;
|
|
neg.f32 %f1215, %f1214;
|
|
add.rn.f32 %f1216, %f1208, %f1215;
|
|
add.rn.f32 %f1217, %f1216, %f1213;
|
|
mov.b32 %r390, %f1214;
|
|
setp.eq.s32 %p149, %r390, 1118925336;
|
|
add.s32 %r391, %r390, -1;
|
|
mov.b32 %f1218, %r391;
|
|
add.f32 %f1219, %f1217, 0f37000000;
|
|
selp.f32 %f339, %f1219, %f1217, %p149;
|
|
selp.f32 %f1220, %f1218, %f1214, %p149;
|
|
mov.f32 %f1221, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1222, %f1220, %f1221;
|
|
cvt.rzi.f32.f32 %f1223, %f1222;
|
|
abs.f32 %f1224, %f1223;
|
|
setp.gt.f32 %p150, %f1224, 0f42FC0000;
|
|
mov.b32 %r392, %f1223;
|
|
and.b32 %r393, %r392, -2147483648;
|
|
or.b32 %r394, %r393, 1123811328;
|
|
mov.b32 %f1225, %r394;
|
|
selp.f32 %f1226, %f1225, %f1223, %p150;
|
|
mov.f32 %f1227, 0fBF317218;
|
|
fma.rn.f32 %f1228, %f1226, %f1227, %f1220;
|
|
mov.f32 %f1229, 0f3102E308;
|
|
fma.rn.f32 %f1230, %f1226, %f1229, %f1228;
|
|
mul.f32 %f1231, %f1230, 0f3FB8AA3B;
|
|
add.f32 %f1232, %f1226, 0f4B40007F;
|
|
mov.b32 %r395, %f1232;
|
|
shl.b32 %r396, %r395, 23;
|
|
mov.b32 %f1233, %r396;
|
|
ex2.approx.ftz.f32 %f1234, %f1231;
|
|
mul.f32 %f340, %f1234, %f1233;
|
|
setp.eq.f32 %p151, %f340, 0f7F800000;
|
|
mov.f32 %f1633, 0f7F800000;
|
|
@%p151 bra $L__BB0_121;
|
|
|
|
fma.rn.f32 %f1633, %f340, %f339, %f340;
|
|
|
|
$L__BB0_121:
|
|
setp.lt.f32 %p152, %f324, 0f00000000;
|
|
and.pred %p8, %p152, %p139;
|
|
setp.eq.f32 %p154, %f324, 0f00000000;
|
|
@%p154 bra $L__BB0_125;
|
|
bra.uni $L__BB0_122;
|
|
|
|
$L__BB0_125:
|
|
add.f32 %f1239, %f324, %f324;
|
|
selp.f32 %f1635, %f1239, 0f00000000, %p139;
|
|
bra.uni $L__BB0_126;
|
|
|
|
$L__BB0_122:
|
|
mov.b32 %r397, %f1633;
|
|
xor.b32 %r398, %r397, -2147483648;
|
|
mov.b32 %f1235, %r398;
|
|
selp.f32 %f1635, %f1235, %f1633, %p8;
|
|
setp.geu.f32 %p155, %f324, 0f00000000;
|
|
@%p155 bra $L__BB0_126;
|
|
|
|
mov.f32 %f1236, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1237, %f1236;
|
|
setp.eq.f32 %p156, %f1237, 0f3ED55555;
|
|
@%p156 bra $L__BB0_126;
|
|
|
|
mov.f32 %f1635, 0f7FFFFFFF;
|
|
|
|
$L__BB0_126:
|
|
add.f32 %f1240, %f338, 0f3ED55555;
|
|
mov.b32 %r399, %f1240;
|
|
setp.lt.s32 %p158, %r399, 2139095040;
|
|
@%p158 bra $L__BB0_131;
|
|
|
|
setp.gtu.f32 %p159, %f338, 0f7F800000;
|
|
@%p159 bra $L__BB0_130;
|
|
bra.uni $L__BB0_128;
|
|
|
|
$L__BB0_130:
|
|
add.f32 %f1635, %f324, 0f3ED55555;
|
|
bra.uni $L__BB0_131;
|
|
|
|
$L__BB0_128:
|
|
setp.neu.f32 %p160, %f338, 0f7F800000;
|
|
@%p160 bra $L__BB0_131;
|
|
|
|
selp.f32 %f1635, 0fFF800000, 0f7F800000, %p8;
|
|
|
|
$L__BB0_131:
|
|
abs.f32 %f349, %f325;
|
|
setp.lt.f32 %p161, %f349, 0f00800000;
|
|
mul.f32 %f1242, %f349, 0f4B800000;
|
|
selp.f32 %f1243, %f1242, %f349, %p161;
|
|
selp.f32 %f1244, 0fC3170000, 0fC2FE0000, %p161;
|
|
mov.b32 %r400, %f1243;
|
|
and.b32 %r401, %r400, 8388607;
|
|
or.b32 %r402, %r401, 1065353216;
|
|
mov.b32 %f1245, %r402;
|
|
shr.u32 %r403, %r400, 23;
|
|
cvt.rn.f32.u32 %f1246, %r403;
|
|
add.f32 %f1247, %f1244, %f1246;
|
|
setp.gt.f32 %p162, %f1245, 0f3FB504F3;
|
|
mul.f32 %f1248, %f1245, 0f3F000000;
|
|
add.f32 %f1249, %f1247, 0f3F800000;
|
|
selp.f32 %f1250, %f1249, %f1247, %p162;
|
|
selp.f32 %f1251, %f1248, %f1245, %p162;
|
|
add.f32 %f1252, %f1251, 0fBF800000;
|
|
add.f32 %f1253, %f1251, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1254, %f1253;
|
|
add.f32 %f1255, %f1252, %f1252;
|
|
mul.f32 %f1256, %f1255, %f1254;
|
|
mul.f32 %f1257, %f1256, %f1256;
|
|
mov.f32 %f1258, 0f3C4CAF63;
|
|
mov.f32 %f1259, 0f3B18F0FE;
|
|
fma.rn.f32 %f1260, %f1259, %f1257, %f1258;
|
|
mov.f32 %f1261, 0f3DAAAABD;
|
|
fma.rn.f32 %f1262, %f1260, %f1257, %f1261;
|
|
mul.rn.f32 %f1263, %f1262, %f1257;
|
|
mul.rn.f32 %f1264, %f1263, %f1256;
|
|
sub.f32 %f1265, %f1252, %f1256;
|
|
add.f32 %f1266, %f1265, %f1265;
|
|
neg.f32 %f1267, %f1256;
|
|
fma.rn.f32 %f1268, %f1267, %f1252, %f1266;
|
|
mul.rn.f32 %f1269, %f1254, %f1268;
|
|
add.f32 %f1270, %f1264, %f1256;
|
|
sub.f32 %f1271, %f1256, %f1270;
|
|
add.f32 %f1272, %f1264, %f1271;
|
|
add.f32 %f1273, %f1269, %f1272;
|
|
add.f32 %f1274, %f1270, %f1273;
|
|
sub.f32 %f1275, %f1270, %f1274;
|
|
add.f32 %f1276, %f1273, %f1275;
|
|
mov.f32 %f1277, 0f3F317200;
|
|
mul.rn.f32 %f1278, %f1250, %f1277;
|
|
mov.f32 %f1279, 0f35BFBE8E;
|
|
mul.rn.f32 %f1280, %f1250, %f1279;
|
|
add.f32 %f1281, %f1278, %f1274;
|
|
sub.f32 %f1282, %f1278, %f1281;
|
|
add.f32 %f1283, %f1274, %f1282;
|
|
add.f32 %f1284, %f1276, %f1283;
|
|
add.f32 %f1285, %f1280, %f1284;
|
|
add.f32 %f1286, %f1281, %f1285;
|
|
sub.f32 %f1287, %f1281, %f1286;
|
|
add.f32 %f1288, %f1285, %f1287;
|
|
mov.f32 %f1289, 0f3ED55555;
|
|
mul.rn.f32 %f1290, %f1289, %f1286;
|
|
neg.f32 %f1291, %f1290;
|
|
fma.rn.f32 %f1292, %f1289, %f1286, %f1291;
|
|
fma.rn.f32 %f1293, %f1289, %f1288, %f1292;
|
|
mov.f32 %f1294, 0f00000000;
|
|
fma.rn.f32 %f1295, %f1294, %f1286, %f1293;
|
|
add.rn.f32 %f1296, %f1290, %f1295;
|
|
neg.f32 %f1297, %f1296;
|
|
add.rn.f32 %f1298, %f1290, %f1297;
|
|
add.rn.f32 %f1299, %f1298, %f1295;
|
|
mov.b32 %r404, %f1296;
|
|
setp.eq.s32 %p163, %r404, 1118925336;
|
|
add.s32 %r405, %r404, -1;
|
|
mov.b32 %f1300, %r405;
|
|
add.f32 %f1301, %f1299, 0f37000000;
|
|
selp.f32 %f350, %f1301, %f1299, %p163;
|
|
selp.f32 %f1302, %f1300, %f1296, %p163;
|
|
mov.f32 %f1303, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1304, %f1302, %f1303;
|
|
cvt.rzi.f32.f32 %f1305, %f1304;
|
|
abs.f32 %f1306, %f1305;
|
|
setp.gt.f32 %p164, %f1306, 0f42FC0000;
|
|
mov.b32 %r406, %f1305;
|
|
and.b32 %r407, %r406, -2147483648;
|
|
or.b32 %r408, %r407, 1123811328;
|
|
mov.b32 %f1307, %r408;
|
|
selp.f32 %f1308, %f1307, %f1305, %p164;
|
|
mov.f32 %f1309, 0fBF317218;
|
|
fma.rn.f32 %f1310, %f1308, %f1309, %f1302;
|
|
mov.f32 %f1311, 0f3102E308;
|
|
fma.rn.f32 %f1312, %f1308, %f1311, %f1310;
|
|
mul.f32 %f1313, %f1312, 0f3FB8AA3B;
|
|
add.f32 %f1314, %f1308, 0f4B40007F;
|
|
mov.b32 %r409, %f1314;
|
|
shl.b32 %r410, %r409, 23;
|
|
mov.b32 %f1315, %r410;
|
|
ex2.approx.ftz.f32 %f1316, %f1313;
|
|
mul.f32 %f351, %f1316, %f1315;
|
|
setp.eq.f32 %p165, %f351, 0f7F800000;
|
|
mov.f32 %f1636, 0f7F800000;
|
|
@%p165 bra $L__BB0_133;
|
|
|
|
fma.rn.f32 %f1636, %f351, %f350, %f351;
|
|
|
|
$L__BB0_133:
|
|
setp.lt.f32 %p166, %f325, 0f00000000;
|
|
and.pred %p9, %p166, %p139;
|
|
setp.eq.f32 %p168, %f325, 0f00000000;
|
|
@%p168 bra $L__BB0_137;
|
|
bra.uni $L__BB0_134;
|
|
|
|
$L__BB0_137:
|
|
add.f32 %f1321, %f325, %f325;
|
|
selp.f32 %f1638, %f1321, 0f00000000, %p139;
|
|
bra.uni $L__BB0_138;
|
|
|
|
$L__BB0_134:
|
|
mov.b32 %r411, %f1636;
|
|
xor.b32 %r412, %r411, -2147483648;
|
|
mov.b32 %f1317, %r412;
|
|
selp.f32 %f1638, %f1317, %f1636, %p9;
|
|
setp.geu.f32 %p169, %f325, 0f00000000;
|
|
@%p169 bra $L__BB0_138;
|
|
|
|
mov.f32 %f1318, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1319, %f1318;
|
|
setp.eq.f32 %p170, %f1319, 0f3ED55555;
|
|
@%p170 bra $L__BB0_138;
|
|
|
|
mov.f32 %f1638, 0f7FFFFFFF;
|
|
|
|
$L__BB0_138:
|
|
add.f32 %f1322, %f349, 0f3ED55555;
|
|
mov.b32 %r413, %f1322;
|
|
setp.lt.s32 %p172, %r413, 2139095040;
|
|
@%p172 bra $L__BB0_143;
|
|
|
|
setp.gtu.f32 %p173, %f349, 0f7F800000;
|
|
@%p173 bra $L__BB0_142;
|
|
bra.uni $L__BB0_140;
|
|
|
|
$L__BB0_142:
|
|
add.f32 %f1638, %f325, 0f3ED55555;
|
|
bra.uni $L__BB0_143;
|
|
|
|
$L__BB0_140:
|
|
setp.neu.f32 %p174, %f349, 0f7F800000;
|
|
@%p174 bra $L__BB0_143;
|
|
|
|
selp.f32 %f1638, 0fFF800000, 0f7F800000, %p9;
|
|
|
|
$L__BB0_143:
|
|
fma.rn.f32 %f1323, %f1632, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p175, %f323, 0f3F800000;
|
|
mov.f32 %f1324, 0f3F800000;
|
|
selp.f32 %f1325, 0f3F7FFFFF, %f1323, %p175;
|
|
mul.f32 %f1326, %f323, 0f414EB852;
|
|
setp.lt.f32 %p176, %f323, 0f3B4D2E1C;
|
|
selp.f32 %f1327, %f1326, %f1325, %p176;
|
|
fma.rn.f32 %f1328, %f1635, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p177, %f324, 0f3F800000;
|
|
selp.f32 %f1329, 0f3F7FFFFF, %f1328, %p177;
|
|
mul.f32 %f1330, %f324, 0f414EB852;
|
|
setp.lt.f32 %p178, %f324, 0f3B4D2E1C;
|
|
selp.f32 %f1331, %f1330, %f1329, %p178;
|
|
fma.rn.f32 %f1332, %f1638, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p179, %f325, 0f3F800000;
|
|
selp.f32 %f1333, 0f3F7FFFFF, %f1332, %p179;
|
|
mul.f32 %f1334, %f325, 0f414EB852;
|
|
setp.lt.f32 %p180, %f325, 0f3B4D2E1C;
|
|
selp.f32 %f1335, %f1334, %f1333, %p180;
|
|
min.f32 %f1336, %f1327, %f1324;
|
|
mov.f32 %f1337, 0f00000000;
|
|
max.f32 %f1338, %f1337, %f1336;
|
|
mul.f32 %f1339, %f1338, 0f43800000;
|
|
cvt.rzi.u32.f32 %r414, %f1339;
|
|
min.u32 %r415, %r414, 255;
|
|
min.f32 %f1340, %f1331, %f1324;
|
|
max.f32 %f1341, %f1337, %f1340;
|
|
mul.f32 %f1342, %f1341, 0f43800000;
|
|
cvt.rzi.u32.f32 %r416, %f1342;
|
|
min.u32 %r417, %r416, 255;
|
|
min.f32 %f1343, %f1335, %f1324;
|
|
max.f32 %f1344, %f1337, %f1343;
|
|
mul.f32 %f1345, %f1344, 0f43800000;
|
|
cvt.rzi.u32.f32 %r418, %f1345;
|
|
min.u32 %r419, %r418, 255;
|
|
shl.b64 %rd80, %rd22, 2;
|
|
add.s64 %rd81, %rd21, %rd80;
|
|
cvt.u16.u32 %rs18, %r419;
|
|
cvt.u16.u32 %rs19, %r417;
|
|
cvt.u16.u32 %rs20, %r415;
|
|
mov.u16 %rs21, 255;
|
|
st.global.v4.u8 [%rd81], {%rs20, %rs19, %rs18, %rs21};
|
|
|
|
$L__BB0_144:
|
|
and.b32 %r420, %r56, 4;
|
|
setp.eq.s32 %p181, %r420, 0;
|
|
ld.const.u32 %r467, [params+108];
|
|
@%p181 bra $L__BB0_148;
|
|
|
|
setp.eq.s32 %p182, %r467, 0;
|
|
ld.const.u64 %rd82, [params+224];
|
|
cvta.to.global.u64 %rd83, %rd82;
|
|
ld.const.u32 %r421, [params+216];
|
|
mad.lo.s32 %r422, %r421, %r7, %r6;
|
|
mul.wide.u32 %rd84, %r422, 8;
|
|
add.s64 %rd23, %rd83, %rd84;
|
|
@%p182 bra $L__BB0_147;
|
|
|
|
ld.global.v4.u16 {%rs29, %rs30, %rs31, %rs32}, [%rd23];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1346, %rs29;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1347, %rs30;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1348, %rs31;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1349, %f283, %f1346;
|
|
add.f32 %f1350, %f284, %f1347;
|
|
add.f32 %f1351, %f286, %f1348;
|
|
mov.f32 %f1352, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1351;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f1350;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f1349;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f1352;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd23], {%rs25, %rs26, %rs27, %rs28};
|
|
bra.uni $L__BB0_148;
|
|
|
|
$L__BB0_147:
|
|
mov.f32 %f1356, 0f3F800000;
|
|
mov.u32 %r467, 0;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f1356;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f286;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f284;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f283;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd23], {%rs33, %rs34, %rs35, %rs36};
|
|
|
|
$L__BB0_148:
|
|
selp.f32 %f1357, 0f3F000000, 0f3E800000, %p198;
|
|
mul.f32 %f1358, %f1357, %f1611;
|
|
mul.f32 %f1359, %f1357, %f1612;
|
|
mul.f32 %f1360, %f1357, %f1613;
|
|
mul.f32 %f1361, %f1357, %f1608;
|
|
mul.f32 %f1362, %f1357, %f1609;
|
|
mul.f32 %f1363, %f1357, %f1610;
|
|
mul.f32 %f1364, %f1357, %f1605;
|
|
mul.f32 %f1365, %f1357, %f1606;
|
|
mul.f32 %f1366, %f1357, %f1607;
|
|
mul.f32 %f1367, %f1357, %f1602;
|
|
mul.f32 %f1368, %f1357, %f1603;
|
|
mul.f32 %f1369, %f1357, %f1604;
|
|
mul.f32 %f360, %f1358, %f816;
|
|
mul.f32 %f361, %f1359, %f817;
|
|
mul.f32 %f362, %f1360, %f285;
|
|
mul.f32 %f363, %f1361, %f816;
|
|
mul.f32 %f364, %f1362, %f817;
|
|
mul.f32 %f365, %f1363, %f285;
|
|
mul.f32 %f366, %f1364, %f816;
|
|
mul.f32 %f367, %f1365, %f817;
|
|
mul.f32 %f368, %f1366, %f285;
|
|
mul.f32 %f369, %f1367, %f816;
|
|
mul.f32 %f370, %f1368, %f817;
|
|
mul.f32 %f371, %f1369, %f285;
|
|
ld.const.u64 %rd85, [params+256];
|
|
cvta.to.global.u64 %rd86, %rd85;
|
|
ld.const.u32 %r424, [params+248];
|
|
mad.lo.s32 %r425, %r424, %r7, %r6;
|
|
mul.wide.u32 %rd87, %r425, 8;
|
|
add.s64 %rd24, %rd86, %rd87;
|
|
setp.eq.s32 %p183, %r467, 0;
|
|
@%p183 bra $L__BB0_150;
|
|
|
|
ld.global.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd24];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1370, %rs44;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1371, %rs45;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1372, %rs46;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1373, %f360, %f1370;
|
|
add.f32 %f1374, %f361, %f1371;
|
|
add.f32 %f1375, %f362, %f1372;
|
|
mov.f32 %f1376, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs42, %f1375;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1374;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f1373;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f1376;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd24], {%rs40, %rs41, %rs42, %rs43};
|
|
bra.uni $L__BB0_151;
|
|
|
|
$L__BB0_150:
|
|
mov.f32 %f1380, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f1380;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f362;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f361;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f360;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd24], {%rs48, %rs49, %rs50, %rs51};
|
|
|
|
$L__BB0_151:
|
|
mov.f32 %f1381, 0f34000000;
|
|
max.f32 %f1382, %f360, %f1381;
|
|
div.rn.f32 %f1383, %f363, %f1382;
|
|
max.f32 %f1384, %f361, %f1381;
|
|
div.rn.f32 %f1385, %f364, %f1384;
|
|
max.f32 %f1386, %f362, %f1381;
|
|
div.rn.f32 %f1387, %f365, %f1386;
|
|
fma.rn.f32 %f372, %f1383, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f373, %f1385, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f374, %f1387, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1388, %f366, %f1382;
|
|
div.rn.f32 %f1389, %f367, %f1384;
|
|
div.rn.f32 %f1390, %f368, %f1386;
|
|
fma.rn.f32 %f375, %f1388, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f376, %f1389, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f377, %f1390, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1391, %f369, %f1382;
|
|
div.rn.f32 %f1392, %f370, %f1384;
|
|
div.rn.f32 %f1393, %f371, %f1386;
|
|
fma.rn.f32 %f378, %f1391, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f379, %f1392, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f380, %f1393, 0f3F000000, 0f3F000000;
|
|
ld.const.u64 %rd88, [params+272];
|
|
cvta.to.global.u64 %rd89, %rd88;
|
|
ld.const.u32 %r426, [params+264];
|
|
mad.lo.s32 %r427, %r426, %r7, %r6;
|
|
mul.wide.u32 %rd90, %r427, 8;
|
|
add.s64 %rd25, %rd89, %rd90;
|
|
@%p183 bra $L__BB0_153;
|
|
|
|
ld.global.v4.u16 {%rs59, %rs60, %rs61, %rs62}, [%rd25];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1394, %rs59;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1395, %rs60;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1396, %rs61;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1397, %f372, %f1394;
|
|
add.f32 %f1398, %f373, %f1395;
|
|
add.f32 %f1399, %f374, %f1396;
|
|
mov.f32 %f1400, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f1399;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs56, %f1398;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1397;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f1400;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd25], {%rs55, %rs56, %rs57, %rs58};
|
|
bra.uni $L__BB0_154;
|
|
|
|
$L__BB0_153:
|
|
mov.f32 %f1404, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f1404;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f374;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f373;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f372;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd25], {%rs63, %rs64, %rs65, %rs66};
|
|
|
|
$L__BB0_154:
|
|
ld.const.u64 %rd91, [params+288];
|
|
cvta.to.global.u64 %rd92, %rd91;
|
|
ld.const.u32 %r428, [params+280];
|
|
mad.lo.s32 %r429, %r428, %r7, %r6;
|
|
mul.wide.u32 %rd93, %r429, 8;
|
|
add.s64 %rd26, %rd92, %rd93;
|
|
@%p183 bra $L__BB0_156;
|
|
|
|
ld.global.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd26];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1405, %rs74;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1406, %rs75;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1407, %rs76;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1408, %f375, %f1405;
|
|
add.f32 %f1409, %f376, %f1406;
|
|
add.f32 %f1410, %f377, %f1407;
|
|
mov.f32 %f1411, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f1410;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f1409;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f1408;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f1411;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd26], {%rs70, %rs71, %rs72, %rs73};
|
|
bra.uni $L__BB0_157;
|
|
|
|
$L__BB0_156:
|
|
mov.f32 %f1415, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f1415;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f377;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f376;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f375;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd26], {%rs78, %rs79, %rs80, %rs81};
|
|
|
|
$L__BB0_157:
|
|
ld.const.u64 %rd94, [params+304];
|
|
cvta.to.global.u64 %rd95, %rd94;
|
|
ld.const.u32 %r430, [params+296];
|
|
mad.lo.s32 %r431, %r430, %r7, %r6;
|
|
mul.wide.u32 %rd96, %r431, 8;
|
|
add.s64 %rd27, %rd95, %rd96;
|
|
@%p183 bra $L__BB0_159;
|
|
|
|
ld.global.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd27];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1416, %rs89;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1417, %rs90;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1418, %rs91;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1419, %f378, %f1416;
|
|
add.f32 %f1420, %f379, %f1417;
|
|
add.f32 %f1421, %f380, %f1418;
|
|
mov.f32 %f1422, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f1421;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f1420;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f1419;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f1422;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd27], {%rs85, %rs86, %rs87, %rs88};
|
|
bra.uni $L__BB0_180;
|
|
|
|
$L__BB0_159:
|
|
mov.f32 %f1426, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs96, %f1426;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f380;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f379;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f378;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd27], {%rs93, %rs94, %rs95, %rs96};
|
|
|
|
$L__BB0_180:
|
|
ret;
|
|
|
|
}
|
|
|