3338 lines
104 KiB
Plaintext
3338 lines
104 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<224>;
|
|
.reg .b16 %rs<206>;
|
|
.reg .f32 %f<2023>;
|
|
.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 %f1956, 0f00000000;
|
|
mov.u16 %rs193, 0;
|
|
mov.f32 %f1957, %f1956;
|
|
mov.f32 %f1958, %f1956;
|
|
@%p10 bra $L__BB0_4;
|
|
|
|
$L__BB0_3:
|
|
cvt.rn.f32.u16 %f281, %rs24;
|
|
div.rn.f32 %f282, %f281, 0f437F0000;
|
|
fma.rn.f32 %f283, %f282, 0f40000000, 0fBF800000;
|
|
and.b16 %rs30, %rs193, 255;
|
|
cvt.rn.f32.u16 %f284, %rs30;
|
|
div.rn.f32 %f285, %f284, 0f437F0000;
|
|
fma.rn.f32 %f286, %f285, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f287, %rs192;
|
|
div.rn.f32 %f288, %f287, 0f437F0000;
|
|
fma.rn.f32 %f289, %f288, 0f40000000, 0fBF800000;
|
|
mul.f32 %f290, %f286, %f286;
|
|
fma.rn.f32 %f291, %f283, %f283, %f290;
|
|
fma.rn.f32 %f292, %f289, %f289, %f291;
|
|
sqrt.rn.f32 %f293, %f292;
|
|
rcp.rn.f32 %f294, %f293;
|
|
mul.f32 %f1958, %f294, %f289;
|
|
mul.f32 %f1957, %f294, %f286;
|
|
mul.f32 %f1956, %f283, %f294;
|
|
|
|
$L__BB0_4:
|
|
ld.const.v2.u32 {%r71, %r72}, [params];
|
|
add.s32 %r5, %r71, %r63;
|
|
add.s32 %r6, %r72, %r64;
|
|
setp.eq.f32 %p11, %f1956, 0f00000000;
|
|
setp.eq.f32 %p12, %f1957, 0f00000000;
|
|
and.pred %p13, %p11, %p12;
|
|
setp.eq.f32 %p14, %f1958, 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 %p214, %r1522, 1;
|
|
mov.pred %p215, 0;
|
|
xor.pred %p216, %p214, %p215;
|
|
not.pred %p217, %p216;
|
|
@%p217 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 %p218, %r1525, 0;
|
|
@%p218 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 %f1918, 0f00000000;
|
|
cvt.rzi.u32.f32 %r1528, %f1918;
|
|
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 %p219, %r1529, 0;
|
|
ld.const.u32 %r1550, [params+108];
|
|
@%p219 bra $L__BB0_183;
|
|
|
|
setp.eq.s32 %p220, %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;
|
|
@%p220 bra $L__BB0_182;
|
|
|
|
ld.global.v4.u16 {%rs138, %rs139, %rs140, %rs141}, [%rd24];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1919, %rs138;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1920, %rs139;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1921, %rs140;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1922, %f1919, 0f00000000;
|
|
add.f32 %f1923, %f1920, 0f00000000;
|
|
add.f32 %f1924, %f1921, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f1924;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f1923;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f1922;}
|
|
|
|
// 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 {%f297, %f298, %f299, %f300}, [params+512];
|
|
neg.f32 %f301, %f298;
|
|
neg.f32 %f302, %f299;
|
|
mul.f32 %f303, %f1956, %f297;
|
|
mul.f32 %f304, %f1957, %f301;
|
|
sub.f32 %f305, %f304, %f303;
|
|
mul.f32 %f306, %f1958, %f299;
|
|
sub.f32 %f13, %f305, %f306;
|
|
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 %f307, [%rd35];
|
|
mul.f32 %f308, %f307, 0f3456BF95;
|
|
ld.global.f32 %f309, [%rd35+4];
|
|
mul.f32 %f310, %f309, 0f3456BF95;
|
|
ld.global.f32 %f311, [%rd35+8];
|
|
mul.f32 %f312, %f311, 0f3456BF95;
|
|
abs.f32 %f313, %f1956;
|
|
div.rn.f32 %f314, %f308, %f313;
|
|
abs.f32 %f315, %f1957;
|
|
div.rn.f32 %f316, %f310, %f315;
|
|
abs.f32 %f317, %f1958;
|
|
div.rn.f32 %f318, %f312, %f317;
|
|
abs.f32 %f319, %f314;
|
|
abs.f32 %f320, %f316;
|
|
abs.f32 %f321, %f318;
|
|
mov.f32 %f322, 0f38D1B717;
|
|
max.f32 %f323, %f319, %f322;
|
|
max.f32 %f324, %f320, %f322;
|
|
max.f32 %f325, %f321, %f322;
|
|
fma.rn.f32 %f14, %f1956, %f323, %f307;
|
|
fma.rn.f32 %f15, %f1957, %f324, %f309;
|
|
fma.rn.f32 %f16, %f1958, %f325, %f311;
|
|
setp.gt.f32 %p16, %f313, %f317;
|
|
neg.f32 %f326, %f1957;
|
|
selp.f32 %f327, %f326, 0f00000000, %p16;
|
|
mov.f32 %f1969, 0f00000000;
|
|
neg.f32 %f328, %f1958;
|
|
selp.f32 %f329, %f1956, %f328, %p16;
|
|
selp.f32 %f330, 0f00000000, %f1957, %p16;
|
|
mul.f32 %f331, %f329, %f329;
|
|
fma.rn.f32 %f332, %f327, %f327, %f331;
|
|
fma.rn.f32 %f333, %f330, %f330, %f332;
|
|
sqrt.rn.f32 %f334, %f333;
|
|
rcp.rn.f32 %f335, %f334;
|
|
mul.f32 %f17, %f327, %f335;
|
|
mul.f32 %f18, %f329, %f335;
|
|
mul.f32 %f19, %f330, %f335;
|
|
mul.f32 %f336, %f1958, %f18;
|
|
mul.f32 %f337, %f1957, %f19;
|
|
sub.f32 %f20, %f336, %f337;
|
|
mul.f32 %f338, %f1956, %f19;
|
|
mul.f32 %f339, %f1958, %f17;
|
|
sub.f32 %f21, %f338, %f339;
|
|
mul.f32 %f340, %f1957, %f17;
|
|
mul.f32 %f341, %f1956, %f18;
|
|
sub.f32 %f22, %f340, %f341;
|
|
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 %f342, [%rd37];
|
|
setp.eq.f32 %p17, %f342, 0f00000000;
|
|
ld.global.f32 %f343, [%rd37+4];
|
|
setp.eq.f32 %p18, %f343, 0f00000000;
|
|
and.pred %p19, %p17, %p18;
|
|
ld.global.f32 %f344, [%rd37+8];
|
|
setp.eq.f32 %p20, %f344, 0f00000000;
|
|
and.pred %p21, %p19, %p20;
|
|
selp.f32 %f345, %f307, %f342, %p21;
|
|
selp.f32 %f346, %f309, %f343, %p21;
|
|
selp.f32 %f347, %f311, %f344, %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 %f348, [%rd39];
|
|
setp.eq.f32 %p22, %f348, 0f00000000;
|
|
ld.global.f32 %f349, [%rd39+4];
|
|
setp.eq.f32 %p23, %f349, 0f00000000;
|
|
and.pred %p24, %p22, %p23;
|
|
ld.global.f32 %f350, [%rd39+8];
|
|
setp.eq.f32 %p25, %f350, 0f00000000;
|
|
and.pred %p26, %p24, %p25;
|
|
selp.f32 %f351, %f307, %f348, %p26;
|
|
selp.f32 %f352, %f309, %f349, %p26;
|
|
selp.f32 %f353, %f311, %f350, %p26;
|
|
sub.f32 %f354, %f345, %f307;
|
|
sub.f32 %f355, %f346, %f309;
|
|
sub.f32 %f356, %f347, %f311;
|
|
sub.f32 %f357, %f351, %f307;
|
|
sub.f32 %f358, %f352, %f309;
|
|
sub.f32 %f359, %f353, %f311;
|
|
abs.f32 %f360, %f354;
|
|
abs.f32 %f361, %f357;
|
|
max.f32 %f362, %f360, %f361;
|
|
abs.f32 %f363, %f355;
|
|
abs.f32 %f364, %f358;
|
|
max.f32 %f365, %f363, %f364;
|
|
abs.f32 %f366, %f356;
|
|
abs.f32 %f367, %f359;
|
|
max.f32 %f368, %f366, %f367;
|
|
max.f32 %f369, %f362, %f365;
|
|
max.f32 %f370, %f369, %f368;
|
|
setp.eq.f32 %p27, %f354, 0f00000000;
|
|
setp.eq.f32 %p28, %f355, 0f00000000;
|
|
and.pred %p29, %p27, %p28;
|
|
setp.eq.f32 %p30, %f356, 0f00000000;
|
|
and.pred %p31, %p30, %p29;
|
|
setp.eq.f32 %p32, %f357, 0f00000000;
|
|
setp.eq.f32 %p33, %f358, 0f00000000;
|
|
and.pred %p34, %p32, %p33;
|
|
setp.eq.f32 %p35, %f359, 0f00000000;
|
|
and.pred %p36, %p35, %p34;
|
|
or.pred %p37, %p31, %p36;
|
|
selp.f32 %f23, 0f00000000, %f370, %p37;
|
|
abs.f32 %f371, %f297;
|
|
abs.f32 %f372, %f299;
|
|
setp.gt.f32 %p38, %f371, %f372;
|
|
selp.f32 %f373, %f301, 0f00000000, %p38;
|
|
selp.f32 %f374, %f297, %f302, %p38;
|
|
selp.f32 %f375, 0f00000000, %f298, %p38;
|
|
mul.f32 %f376, %f374, %f374;
|
|
fma.rn.f32 %f377, %f373, %f373, %f376;
|
|
fma.rn.f32 %f378, %f375, %f375, %f377;
|
|
sqrt.rn.f32 %f379, %f378;
|
|
rcp.rn.f32 %f380, %f379;
|
|
mul.f32 %f24, %f373, %f380;
|
|
mul.f32 %f25, %f374, %f380;
|
|
mul.f32 %f26, %f375, %f380;
|
|
ld.const.u32 %r7, [params+540];
|
|
setp.lt.s32 %p39, %r7, 1;
|
|
mov.f32 %f1970, %f1969;
|
|
@%p39 bra $L__BB0_52;
|
|
|
|
ld.const.u64 %rd40, [params+128];
|
|
ld.const.u32 %r87, [params+120];
|
|
cvt.rn.f32.s32 %f383, %r7;
|
|
rcp.rn.f32 %f27, %f383;
|
|
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 %f384, %f297, %f25;
|
|
mul.f32 %f385, %f298, %f24;
|
|
sub.f32 %f33, %f385, %f384;
|
|
mul.f32 %f386, %f299, %f24;
|
|
mul.f32 %f387, %f297, %f26;
|
|
sub.f32 %f34, %f387, %f386;
|
|
mul.f32 %f388, %f298, %f26;
|
|
mul.f32 %f389, %f299, %f25;
|
|
sub.f32 %f35, %f389, %f388;
|
|
abs.f32 %f472, %f31;
|
|
abs.f32 %f473, %f30;
|
|
max.f32 %f474, %f473, %f472;
|
|
abs.f32 %f475, %f32;
|
|
max.f32 %f476, %f474, %f475;
|
|
mov.u32 %r1539, %r86;
|
|
|
|
$L__BB0_7:
|
|
mov.u32 %r1541, %r86;
|
|
|
|
$L__BB0_8:
|
|
cvt.rn.f32.s32 %f1955, %r1539;
|
|
mad.lo.s32 %r90, %r1542, 1664525, 1013904223;
|
|
and.b32 %r91, %r90, 16777215;
|
|
cvt.rn.f32.u32 %f390, %r91;
|
|
fma.rn.f32 %f391, %f390, 0f33800000, %f1955;
|
|
mul.f32 %f392, %f27, %f391;
|
|
mad.lo.s32 %r13, %r90, 1664525, 1013904223;
|
|
and.b32 %r92, %r13, 16777215;
|
|
cvt.rn.f32.u32 %f393, %r92;
|
|
cvt.rn.f32.s32 %f394, %r1541;
|
|
fma.rn.f32 %f395, %f393, 0f33800000, %f394;
|
|
mul.f32 %f396, %f27, %f395;
|
|
sqrt.rn.f32 %f41, %f392;
|
|
mul.f32 %f42, %f396, 0f40C90FDB;
|
|
mul.f32 %f397, %f42, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r1549, %f397;
|
|
cvt.rn.f32.s32 %f398, %r1549;
|
|
mov.f32 %f399, 0fBFC90FDA;
|
|
fma.rn.f32 %f400, %f398, %f399, %f42;
|
|
mov.f32 %f401, 0fB3A22168;
|
|
fma.rn.f32 %f402, %f398, %f401, %f400;
|
|
mov.f32 %f403, 0fA7C234C5;
|
|
fma.rn.f32 %f1966, %f398, %f403, %f402;
|
|
abs.f32 %f44, %f42;
|
|
setp.ltu.f32 %p41, %f44, 0f47CE4780;
|
|
add.s64 %rd4, %rd1, 24;
|
|
mov.u32 %r1546, %r1549;
|
|
mov.f32 %f1963, %f1966;
|
|
@%p41 bra $L__BB0_16;
|
|
|
|
setp.eq.f32 %p42, %f44, 0f7F800000;
|
|
@%p42 bra $L__BB0_15;
|
|
bra.uni $L__BB0_10;
|
|
|
|
$L__BB0_15:
|
|
mov.f32 %f406, 0f00000000;
|
|
mul.rn.f32 %f1963, %f42, %f406;
|
|
mov.u32 %r1546, 0;
|
|
bra.uni $L__BB0_16;
|
|
|
|
$L__BB0_10:
|
|
mov.b32 %r15, %f42;
|
|
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 %f404, %fd2;
|
|
setp.eq.s32 %p47, %r116, 0;
|
|
neg.f32 %f405, %f404;
|
|
selp.f32 %f1963, %f404, %f405, %p47;
|
|
|
|
$L__BB0_16:
|
|
add.s32 %r31, %r1546, 1;
|
|
and.b32 %r32, %r31, 1;
|
|
setp.eq.s32 %p48, %r32, 0;
|
|
selp.f32 %f48, %f1963, 0f3F800000, %p48;
|
|
mul.rn.f32 %f49, %f1963, %f1963;
|
|
mov.f32 %f1964, 0fB94D4153;
|
|
@%p48 bra $L__BB0_18;
|
|
|
|
mov.f32 %f408, 0fBAB607ED;
|
|
mov.f32 %f409, 0f37CBAC00;
|
|
fma.rn.f32 %f1964, %f409, %f49, %f408;
|
|
|
|
$L__BB0_18:
|
|
selp.f32 %f410, 0f3C0885E4, 0f3D2AAABB, %p48;
|
|
fma.rn.f32 %f411, %f1964, %f49, %f410;
|
|
selp.f32 %f412, 0fBE2AAAA8, 0fBEFFFFFF, %p48;
|
|
fma.rn.f32 %f413, %f411, %f49, %f412;
|
|
mov.f32 %f414, 0f00000000;
|
|
fma.rn.f32 %f415, %f49, %f48, %f414;
|
|
fma.rn.f32 %f1965, %f413, %f415, %f48;
|
|
and.b32 %r122, %r31, 2;
|
|
setp.eq.s32 %p50, %r122, 0;
|
|
@%p50 bra $L__BB0_20;
|
|
|
|
mov.f32 %f417, 0fBF800000;
|
|
fma.rn.f32 %f1965, %f1965, %f417, %f414;
|
|
|
|
$L__BB0_20:
|
|
@%p41 bra $L__BB0_28;
|
|
|
|
setp.eq.f32 %p52, %f44, 0f7F800000;
|
|
@%p52 bra $L__BB0_27;
|
|
bra.uni $L__BB0_22;
|
|
|
|
$L__BB0_27:
|
|
mov.f32 %f420, 0f00000000;
|
|
mul.rn.f32 %f1966, %f42, %f420;
|
|
mov.u32 %r1549, 0;
|
|
bra.uni $L__BB0_28;
|
|
|
|
$L__BB0_22:
|
|
mov.b32 %r33, %f42;
|
|
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 %f418, %fd4;
|
|
setp.eq.s32 %p57, %r147, 0;
|
|
neg.f32 %f419, %f418;
|
|
selp.f32 %f1966, %f418, %f419, %p57;
|
|
|
|
$L__BB0_28:
|
|
mul.f32 %f58, %f41, %f1965;
|
|
and.b32 %r47, %r1549, 1;
|
|
setp.eq.s32 %p58, %r47, 0;
|
|
selp.f32 %f59, %f1966, 0f3F800000, %p58;
|
|
mul.rn.f32 %f60, %f1966, %f1966;
|
|
mov.f32 %f1967, 0fB94D4153;
|
|
@%p58 bra $L__BB0_30;
|
|
|
|
mov.f32 %f422, 0fBAB607ED;
|
|
mov.f32 %f423, 0f37CBAC00;
|
|
fma.rn.f32 %f1967, %f423, %f60, %f422;
|
|
|
|
$L__BB0_30:
|
|
selp.f32 %f424, 0f3C0885E4, 0f3D2AAABB, %p58;
|
|
fma.rn.f32 %f425, %f1967, %f60, %f424;
|
|
selp.f32 %f426, 0fBE2AAAA8, 0fBEFFFFFF, %p58;
|
|
fma.rn.f32 %f427, %f425, %f60, %f426;
|
|
mov.f32 %f428, 0f00000000;
|
|
fma.rn.f32 %f429, %f60, %f59, %f428;
|
|
fma.rn.f32 %f1968, %f427, %f429, %f59;
|
|
and.b32 %r153, %r1549, 2;
|
|
setp.eq.s32 %p60, %r153, 0;
|
|
@%p60 bra $L__BB0_32;
|
|
|
|
mov.f32 %f431, 0fBF800000;
|
|
fma.rn.f32 %f1968, %f1968, %f431, %f428;
|
|
|
|
$L__BB0_32:
|
|
mul.f32 %f441, %f58, %f58;
|
|
mov.f32 %f442, 0f3F800000;
|
|
sub.f32 %f443, %f442, %f441;
|
|
mul.f32 %f444, %f41, %f1968;
|
|
mul.f32 %f445, %f444, %f444;
|
|
sub.f32 %f446, %f443, %f445;
|
|
max.f32 %f447, %f428, %f446;
|
|
sqrt.rn.f32 %f448, %f447;
|
|
mul.f32 %f449, %f24, %f444;
|
|
mul.f32 %f450, %f25, %f444;
|
|
mul.f32 %f451, %f26, %f444;
|
|
fma.rn.f32 %f452, %f35, %f58, %f449;
|
|
fma.rn.f32 %f453, %f34, %f58, %f450;
|
|
fma.rn.f32 %f454, %f33, %f58, %f451;
|
|
fma.rn.f32 %f455, %f297, %f448, %f452;
|
|
fma.rn.f32 %f456, %f298, %f448, %f453;
|
|
fma.rn.f32 %f457, %f299, %f448, %f454;
|
|
add.f32 %f458, %f297, %f455;
|
|
add.f32 %f459, %f298, %f456;
|
|
add.f32 %f460, %f299, %f457;
|
|
mul.f32 %f461, %f28, %f458;
|
|
mul.f32 %f462, %f28, %f459;
|
|
mul.f32 %f463, %f28, %f460;
|
|
sub.f32 %f464, %f461, %f297;
|
|
sub.f32 %f465, %f462, %f298;
|
|
sub.f32 %f466, %f463, %f299;
|
|
mul.f32 %f467, %f465, %f465;
|
|
fma.rn.f32 %f468, %f464, %f464, %f467;
|
|
fma.rn.f32 %f469, %f466, %f466, %f468;
|
|
sqrt.rn.f32 %f470, %f469;
|
|
rcp.rn.f32 %f471, %f470;
|
|
mul.f32 %f66, %f471, %f464;
|
|
mul.f32 %f67, %f471, %f465;
|
|
mul.f32 %f68, %f471, %f466;
|
|
mov.f32 %f477, 0f38D1B717;
|
|
max.f32 %f478, %f476, %f477;
|
|
mul.f32 %f438, %f29, %f478;
|
|
mad.lo.s32 %r225, %r13, 1664525, 1013904223;
|
|
and.b32 %r226, %r225, 16777215;
|
|
cvt.rn.f32.u32 %f479, %r226;
|
|
mul.f32 %f480, %f479, 0f33800000;
|
|
fma.rn.f32 %f481, %f480, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r48, %r225, 1664525, 1013904223;
|
|
and.b32 %r227, %r48, 16777215;
|
|
cvt.rn.f32.u32 %f482, %r227;
|
|
mul.f32 %f483, %f482, 0f33800000;
|
|
fma.rn.f32 %f484, %f483, 0f40000000, 0fBF800000;
|
|
mul.f32 %f485, %f17, %f484;
|
|
mul.f32 %f486, %f18, %f484;
|
|
mul.f32 %f487, %f19, %f484;
|
|
fma.rn.f32 %f488, %f20, %f481, %f485;
|
|
fma.rn.f32 %f489, %f21, %f481, %f486;
|
|
fma.rn.f32 %f490, %f22, %f481, %f487;
|
|
mul.f32 %f491, %f489, %f489;
|
|
fma.rn.f32 %f492, %f488, %f488, %f491;
|
|
fma.rn.f32 %f493, %f490, %f490, %f492;
|
|
sqrt.rn.f32 %f494, %f493;
|
|
rcp.rn.f32 %f495, %f494;
|
|
mul.f32 %f496, %f488, %f495;
|
|
mul.f32 %f497, %f489, %f495;
|
|
mul.f32 %f498, %f490, %f495;
|
|
mul.f32 %f499, %f1957, %f498;
|
|
mul.f32 %f500, %f1958, %f497;
|
|
sub.f32 %f501, %f499, %f500;
|
|
mul.f32 %f502, %f1958, %f496;
|
|
mul.f32 %f503, %f1956, %f498;
|
|
sub.f32 %f504, %f502, %f503;
|
|
mul.f32 %f505, %f1956, %f497;
|
|
mul.f32 %f506, %f1957, %f496;
|
|
sub.f32 %f507, %f505, %f506;
|
|
mul.f32 %f508, %f496, 0f3E24A490;
|
|
mul.f32 %f509, %f497, 0f3E24A490;
|
|
mul.f32 %f510, %f498, 0f3E24A490;
|
|
fma.rn.f32 %f511, %f23, %f508, %f14;
|
|
fma.rn.f32 %f512, %f23, %f509, %f15;
|
|
fma.rn.f32 %f513, %f23, %f510, %f16;
|
|
mul.f32 %f514, %f501, 0fBECECEDE;
|
|
mul.f32 %f515, %f504, 0fBECECEDE;
|
|
mul.f32 %f516, %f507, 0fBECECEDE;
|
|
fma.rn.f32 %f70, %f23, %f514, %f511;
|
|
fma.rn.f32 %f71, %f23, %f515, %f512;
|
|
fma.rn.f32 %f72, %f23, %f516, %f513;
|
|
sub.f32 %f517, %f70, %f14;
|
|
sub.f32 %f518, %f71, %f15;
|
|
sub.f32 %f519, %f72, %f16;
|
|
mul.f32 %f520, %f518, %f518;
|
|
fma.rn.f32 %f521, %f517, %f517, %f520;
|
|
fma.rn.f32 %f522, %f519, %f519, %f521;
|
|
sqrt.rn.f32 %f439, %f522;
|
|
rcp.rn.f32 %f523, %f439;
|
|
mul.f32 %f435, %f523, %f517;
|
|
mul.f32 %f436, %f523, %f518;
|
|
mul.f32 %f437, %f523, %f519;
|
|
mov.u32 %r190, 2;
|
|
mov.u32 %r192, 1;
|
|
mov.u32 %r193, 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,%f435,%f436,%f437,%f438,%f439,%f428,%r192,%r224,%r192,%r190,%r192,%r192,%r193,%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,%r224,%r224);
|
|
// end inline asm
|
|
mov.b32 %f524, %r154;
|
|
setp.neu.f32 %p61, %f524, 0f3F800000;
|
|
mov.u16 %rs195, 0;
|
|
@%p61 bra $L__BB0_34;
|
|
|
|
mov.f32 %f532, 0f6C4ECB8F;
|
|
mov.f32 %f533, 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,%f70,%f71,%f72,%f66,%f67,%f68,%f438,%f532,%f533,%r192,%r224,%r192,%r190,%r192,%r192,%r193,%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,%r224,%r224);
|
|
// end inline asm
|
|
mov.b32 %f534, %r228;
|
|
add.f32 %f1969, %f1969, %f534;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_34:
|
|
mad.lo.s32 %r370, %r48, 1664525, 1013904223;
|
|
and.b32 %r371, %r370, 16777215;
|
|
cvt.rn.f32.u32 %f544, %r371;
|
|
mul.f32 %f545, %f544, 0f33800000;
|
|
fma.rn.f32 %f546, %f545, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r49, %r370, 1664525, 1013904223;
|
|
and.b32 %r372, %r49, 16777215;
|
|
cvt.rn.f32.u32 %f547, %r372;
|
|
mul.f32 %f548, %f547, 0f33800000;
|
|
fma.rn.f32 %f549, %f548, 0f40000000, 0fBF800000;
|
|
mul.f32 %f550, %f17, %f549;
|
|
mul.f32 %f551, %f18, %f549;
|
|
mul.f32 %f552, %f19, %f549;
|
|
fma.rn.f32 %f553, %f20, %f546, %f550;
|
|
fma.rn.f32 %f554, %f21, %f546, %f551;
|
|
fma.rn.f32 %f555, %f22, %f546, %f552;
|
|
mul.f32 %f556, %f554, %f554;
|
|
fma.rn.f32 %f557, %f553, %f553, %f556;
|
|
fma.rn.f32 %f558, %f555, %f555, %f557;
|
|
sqrt.rn.f32 %f559, %f558;
|
|
rcp.rn.f32 %f560, %f559;
|
|
mul.f32 %f561, %f553, %f560;
|
|
mul.f32 %f562, %f554, %f560;
|
|
mul.f32 %f563, %f555, %f560;
|
|
mul.f32 %f564, %f1957, %f563;
|
|
mul.f32 %f565, %f1958, %f562;
|
|
sub.f32 %f566, %f564, %f565;
|
|
mul.f32 %f567, %f1958, %f561;
|
|
mul.f32 %f568, %f1956, %f563;
|
|
sub.f32 %f569, %f567, %f568;
|
|
mul.f32 %f570, %f1956, %f562;
|
|
mul.f32 %f571, %f1957, %f561;
|
|
sub.f32 %f572, %f570, %f571;
|
|
mul.f32 %f573, %f561, 0fBE24A490;
|
|
mul.f32 %f574, %f562, 0fBE24A490;
|
|
mul.f32 %f575, %f563, 0fBE24A490;
|
|
fma.rn.f32 %f576, %f23, %f573, %f14;
|
|
fma.rn.f32 %f577, %f23, %f574, %f15;
|
|
fma.rn.f32 %f578, %f23, %f575, %f16;
|
|
mul.f32 %f579, %f566, 0f3ECACABC;
|
|
mul.f32 %f580, %f569, 0f3ECACABC;
|
|
mul.f32 %f581, %f572, 0f3ECACABC;
|
|
fma.rn.f32 %f77, %f23, %f579, %f576;
|
|
fma.rn.f32 %f78, %f23, %f580, %f577;
|
|
fma.rn.f32 %f79, %f23, %f581, %f578;
|
|
sub.f32 %f582, %f77, %f14;
|
|
sub.f32 %f583, %f78, %f15;
|
|
sub.f32 %f584, %f79, %f16;
|
|
mul.f32 %f585, %f583, %f583;
|
|
fma.rn.f32 %f586, %f582, %f582, %f585;
|
|
fma.rn.f32 %f587, %f584, %f584, %f586;
|
|
sqrt.rn.f32 %f542, %f587;
|
|
rcp.rn.f32 %f588, %f542;
|
|
mul.f32 %f538, %f588, %f582;
|
|
mul.f32 %f539, %f588, %f583;
|
|
mul.f32 %f540, %f588, %f584;
|
|
mov.f32 %f543, 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,%f538,%f539,%f540,%f438,%f542,%f543,%r192,%r224,%r192,%r190,%r192,%r192,%r193,%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,%r224,%r224);
|
|
// end inline asm
|
|
mov.b32 %f589, %r299;
|
|
setp.neu.f32 %p62, %f589, 0f3F800000;
|
|
@%p62 bra $L__BB0_36;
|
|
|
|
mov.f32 %f597, 0f6C4ECB8F;
|
|
mov.u32 %r409, 2;
|
|
mov.u32 %r411, 1;
|
|
mov.u32 %r412, 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,%f77,%f78,%f79,%f66,%f67,%f68,%f438,%f597,%f543,%r411,%r443,%r411,%r409,%r411,%r411,%r412,%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,%r443,%r443);
|
|
// end inline asm
|
|
mov.b32 %f599, %r373;
|
|
add.f32 %f1969, %f1969, %f599;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_36:
|
|
mad.lo.s32 %r515, %r49, 1664525, 1013904223;
|
|
and.b32 %r516, %r515, 16777215;
|
|
cvt.rn.f32.u32 %f609, %r516;
|
|
mul.f32 %f610, %f609, 0f33800000;
|
|
fma.rn.f32 %f611, %f610, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r50, %r515, 1664525, 1013904223;
|
|
and.b32 %r517, %r50, 16777215;
|
|
cvt.rn.f32.u32 %f612, %r517;
|
|
mul.f32 %f613, %f612, 0f33800000;
|
|
fma.rn.f32 %f614, %f613, 0f40000000, 0fBF800000;
|
|
mul.f32 %f615, %f17, %f614;
|
|
mul.f32 %f616, %f18, %f614;
|
|
mul.f32 %f617, %f19, %f614;
|
|
fma.rn.f32 %f618, %f20, %f611, %f615;
|
|
fma.rn.f32 %f619, %f21, %f611, %f616;
|
|
fma.rn.f32 %f620, %f22, %f611, %f617;
|
|
mul.f32 %f621, %f619, %f619;
|
|
fma.rn.f32 %f622, %f618, %f618, %f621;
|
|
fma.rn.f32 %f623, %f620, %f620, %f622;
|
|
sqrt.rn.f32 %f624, %f623;
|
|
rcp.rn.f32 %f625, %f624;
|
|
mul.f32 %f626, %f618, %f625;
|
|
mul.f32 %f627, %f619, %f625;
|
|
mul.f32 %f628, %f620, %f625;
|
|
mul.f32 %f629, %f1957, %f628;
|
|
mul.f32 %f630, %f1958, %f627;
|
|
sub.f32 %f631, %f629, %f630;
|
|
mul.f32 %f632, %f1958, %f626;
|
|
mul.f32 %f633, %f1956, %f628;
|
|
sub.f32 %f634, %f632, %f633;
|
|
mul.f32 %f635, %f1956, %f627;
|
|
mul.f32 %f636, %f1957, %f626;
|
|
sub.f32 %f637, %f635, %f636;
|
|
mul.f32 %f638, %f626, 0f3F23A3A8;
|
|
mul.f32 %f639, %f627, 0f3F23A3A8;
|
|
mul.f32 %f640, %f628, 0f3F23A3A8;
|
|
fma.rn.f32 %f641, %f23, %f638, %f14;
|
|
fma.rn.f32 %f642, %f23, %f639, %f15;
|
|
fma.rn.f32 %f643, %f23, %f640, %f16;
|
|
mul.f32 %f644, %f631, 0f3E24A490;
|
|
mul.f32 %f645, %f634, 0f3E24A490;
|
|
mul.f32 %f646, %f637, 0f3E24A490;
|
|
fma.rn.f32 %f84, %f23, %f644, %f641;
|
|
fma.rn.f32 %f85, %f23, %f645, %f642;
|
|
fma.rn.f32 %f86, %f23, %f646, %f643;
|
|
sub.f32 %f647, %f84, %f14;
|
|
sub.f32 %f648, %f85, %f15;
|
|
sub.f32 %f649, %f86, %f16;
|
|
mul.f32 %f650, %f648, %f648;
|
|
fma.rn.f32 %f651, %f647, %f647, %f650;
|
|
fma.rn.f32 %f652, %f649, %f649, %f651;
|
|
sqrt.rn.f32 %f607, %f652;
|
|
rcp.rn.f32 %f653, %f607;
|
|
mul.f32 %f603, %f653, %f647;
|
|
mul.f32 %f604, %f653, %f648;
|
|
mul.f32 %f605, %f653, %f649;
|
|
mov.u32 %r480, 2;
|
|
mov.u32 %r482, 1;
|
|
mov.u32 %r483, 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,%f603,%f604,%f605,%f438,%f607,%f543,%r482,%r514,%r482,%r480,%r482,%r482,%r483,%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,%r514,%r514);
|
|
// end inline asm
|
|
mov.b32 %f654, %r444;
|
|
setp.neu.f32 %p63, %f654, 0f3F800000;
|
|
@%p63 bra $L__BB0_38;
|
|
|
|
mov.f32 %f662, 0f6C4ECB8F;
|
|
mov.f32 %f663, 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,%f84,%f85,%f86,%f66,%f67,%f68,%f438,%f662,%f663,%r482,%r514,%r482,%r480,%r482,%r482,%r483,%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,%r514,%r514);
|
|
// end inline asm
|
|
mov.b32 %f664, %r518;
|
|
add.f32 %f1969, %f1969, %f664;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_38:
|
|
mad.lo.s32 %r660, %r50, 1664525, 1013904223;
|
|
and.b32 %r661, %r660, 16777215;
|
|
cvt.rn.f32.u32 %f674, %r661;
|
|
mul.f32 %f675, %f674, 0f33800000;
|
|
fma.rn.f32 %f676, %f675, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r51, %r660, 1664525, 1013904223;
|
|
and.b32 %r662, %r51, 16777215;
|
|
cvt.rn.f32.u32 %f677, %r662;
|
|
mul.f32 %f678, %f677, 0f33800000;
|
|
fma.rn.f32 %f679, %f678, 0f40000000, 0fBF800000;
|
|
mul.f32 %f680, %f17, %f679;
|
|
mul.f32 %f681, %f18, %f679;
|
|
mul.f32 %f682, %f19, %f679;
|
|
fma.rn.f32 %f683, %f20, %f676, %f680;
|
|
fma.rn.f32 %f684, %f21, %f676, %f681;
|
|
fma.rn.f32 %f685, %f22, %f676, %f682;
|
|
mul.f32 %f686, %f684, %f684;
|
|
fma.rn.f32 %f687, %f683, %f683, %f686;
|
|
fma.rn.f32 %f688, %f685, %f685, %f687;
|
|
sqrt.rn.f32 %f689, %f688;
|
|
rcp.rn.f32 %f690, %f689;
|
|
mul.f32 %f691, %f683, %f690;
|
|
mul.f32 %f692, %f684, %f690;
|
|
mul.f32 %f693, %f685, %f690;
|
|
mul.f32 %f694, %f1957, %f693;
|
|
mul.f32 %f695, %f1958, %f692;
|
|
sub.f32 %f696, %f694, %f695;
|
|
mul.f32 %f697, %f1958, %f691;
|
|
mul.f32 %f698, %f1956, %f693;
|
|
sub.f32 %f699, %f697, %f698;
|
|
mul.f32 %f700, %f1956, %f692;
|
|
mul.f32 %f701, %f1957, %f691;
|
|
sub.f32 %f702, %f700, %f701;
|
|
mul.f32 %f703, %f691, 0fBECECEDE;
|
|
mul.f32 %f704, %f692, 0fBECECEDE;
|
|
mul.f32 %f705, %f693, 0fBECECEDE;
|
|
fma.rn.f32 %f706, %f23, %f703, %f14;
|
|
fma.rn.f32 %f707, %f23, %f704, %f15;
|
|
fma.rn.f32 %f708, %f23, %f705, %f16;
|
|
mul.f32 %f709, %f696, 0fBF23A3A9;
|
|
mul.f32 %f710, %f699, 0fBF23A3A9;
|
|
mul.f32 %f711, %f702, 0fBF23A3A9;
|
|
fma.rn.f32 %f91, %f23, %f709, %f706;
|
|
fma.rn.f32 %f92, %f23, %f710, %f707;
|
|
fma.rn.f32 %f93, %f23, %f711, %f708;
|
|
sub.f32 %f712, %f91, %f14;
|
|
sub.f32 %f713, %f92, %f15;
|
|
sub.f32 %f714, %f93, %f16;
|
|
mul.f32 %f715, %f713, %f713;
|
|
fma.rn.f32 %f716, %f712, %f712, %f715;
|
|
fma.rn.f32 %f717, %f714, %f714, %f716;
|
|
sqrt.rn.f32 %f672, %f717;
|
|
rcp.rn.f32 %f718, %f672;
|
|
mul.f32 %f668, %f718, %f712;
|
|
mul.f32 %f669, %f718, %f713;
|
|
mul.f32 %f670, %f718, %f714;
|
|
mov.f32 %f673, 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,%f668,%f669,%f670,%f438,%f672,%f673,%r482,%r514,%r482,%r480,%r482,%r482,%r483,%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,%r514,%r514);
|
|
// end inline asm
|
|
mov.b32 %f719, %r589;
|
|
setp.neu.f32 %p64, %f719, 0f3F800000;
|
|
@%p64 bra $L__BB0_40;
|
|
|
|
mov.f32 %f727, 0f6C4ECB8F;
|
|
mov.u32 %r699, 2;
|
|
mov.u32 %r701, 1;
|
|
mov.u32 %r702, 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,%f91,%f92,%f93,%f66,%f67,%f68,%f438,%f727,%f673,%r701,%r733,%r701,%r699,%r701,%r701,%r702,%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,%r733,%r733);
|
|
// end inline asm
|
|
mov.b32 %f729, %r663;
|
|
add.f32 %f1969, %f1969, %f729;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_40:
|
|
mad.lo.s32 %r805, %r51, 1664525, 1013904223;
|
|
and.b32 %r806, %r805, 16777215;
|
|
cvt.rn.f32.u32 %f739, %r806;
|
|
mul.f32 %f740, %f739, 0f33800000;
|
|
fma.rn.f32 %f741, %f740, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r52, %r805, 1664525, 1013904223;
|
|
and.b32 %r807, %r52, 16777215;
|
|
cvt.rn.f32.u32 %f742, %r807;
|
|
mul.f32 %f743, %f742, 0f33800000;
|
|
fma.rn.f32 %f744, %f743, 0f40000000, 0fBF800000;
|
|
mul.f32 %f745, %f17, %f744;
|
|
mul.f32 %f746, %f18, %f744;
|
|
mul.f32 %f747, %f19, %f744;
|
|
fma.rn.f32 %f748, %f20, %f741, %f745;
|
|
fma.rn.f32 %f749, %f21, %f741, %f746;
|
|
fma.rn.f32 %f750, %f22, %f741, %f747;
|
|
mul.f32 %f751, %f749, %f749;
|
|
fma.rn.f32 %f752, %f748, %f748, %f751;
|
|
fma.rn.f32 %f753, %f750, %f750, %f752;
|
|
sqrt.rn.f32 %f754, %f753;
|
|
rcp.rn.f32 %f755, %f754;
|
|
mul.f32 %f756, %f748, %f755;
|
|
mul.f32 %f757, %f749, %f755;
|
|
mul.f32 %f758, %f750, %f755;
|
|
mul.f32 %f759, %f1957, %f758;
|
|
mul.f32 %f760, %f1958, %f757;
|
|
sub.f32 %f761, %f759, %f760;
|
|
mul.f32 %f762, %f1958, %f756;
|
|
mul.f32 %f763, %f1956, %f758;
|
|
sub.f32 %f764, %f762, %f763;
|
|
mul.f32 %f765, %f1956, %f757;
|
|
mul.f32 %f766, %f1957, %f756;
|
|
sub.f32 %f767, %f765, %f766;
|
|
mul.f32 %f768, %f756, 0fBF23A3A9;
|
|
mul.f32 %f769, %f757, 0fBF23A3A9;
|
|
mul.f32 %f770, %f758, 0fBF23A3A9;
|
|
fma.rn.f32 %f771, %f23, %f768, %f14;
|
|
fma.rn.f32 %f772, %f23, %f769, %f15;
|
|
fma.rn.f32 %f773, %f23, %f770, %f16;
|
|
mul.f32 %f774, %f761, 0f3F23A3A8;
|
|
mul.f32 %f775, %f764, 0f3F23A3A8;
|
|
mul.f32 %f776, %f767, 0f3F23A3A8;
|
|
fma.rn.f32 %f98, %f23, %f774, %f771;
|
|
fma.rn.f32 %f99, %f23, %f775, %f772;
|
|
fma.rn.f32 %f100, %f23, %f776, %f773;
|
|
sub.f32 %f777, %f98, %f14;
|
|
sub.f32 %f778, %f99, %f15;
|
|
sub.f32 %f779, %f100, %f16;
|
|
mul.f32 %f780, %f778, %f778;
|
|
fma.rn.f32 %f781, %f777, %f777, %f780;
|
|
fma.rn.f32 %f782, %f779, %f779, %f781;
|
|
sqrt.rn.f32 %f737, %f782;
|
|
rcp.rn.f32 %f783, %f737;
|
|
mul.f32 %f733, %f783, %f777;
|
|
mul.f32 %f734, %f783, %f778;
|
|
mul.f32 %f735, %f783, %f779;
|
|
mov.u32 %r770, 2;
|
|
mov.u32 %r772, 1;
|
|
mov.u32 %r773, 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,%f733,%f734,%f735,%f438,%f737,%f673,%r772,%r804,%r772,%r770,%r772,%r772,%r773,%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,%r804,%r804);
|
|
// end inline asm
|
|
mov.b32 %f784, %r734;
|
|
setp.neu.f32 %p65, %f784, 0f3F800000;
|
|
@%p65 bra $L__BB0_42;
|
|
|
|
mov.f32 %f792, 0f6C4ECB8F;
|
|
mov.f32 %f793, 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,%f98,%f99,%f100,%f66,%f67,%f68,%f438,%f792,%f793,%r772,%r804,%r772,%r770,%r772,%r772,%r773,%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,%r804,%r804);
|
|
// end inline asm
|
|
mov.b32 %f794, %r808;
|
|
add.f32 %f1969, %f1969, %f794;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_42:
|
|
mad.lo.s32 %r950, %r52, 1664525, 1013904223;
|
|
and.b32 %r951, %r950, 16777215;
|
|
cvt.rn.f32.u32 %f804, %r951;
|
|
mul.f32 %f805, %f804, 0f33800000;
|
|
fma.rn.f32 %f806, %f805, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r53, %r950, 1664525, 1013904223;
|
|
and.b32 %r952, %r53, 16777215;
|
|
cvt.rn.f32.u32 %f807, %r952;
|
|
mul.f32 %f808, %f807, 0f33800000;
|
|
fma.rn.f32 %f809, %f808, 0f40000000, 0fBF800000;
|
|
mul.f32 %f810, %f17, %f809;
|
|
mul.f32 %f811, %f18, %f809;
|
|
mul.f32 %f812, %f19, %f809;
|
|
fma.rn.f32 %f813, %f20, %f806, %f810;
|
|
fma.rn.f32 %f814, %f21, %f806, %f811;
|
|
fma.rn.f32 %f815, %f22, %f806, %f812;
|
|
mul.f32 %f816, %f814, %f814;
|
|
fma.rn.f32 %f817, %f813, %f813, %f816;
|
|
fma.rn.f32 %f818, %f815, %f815, %f817;
|
|
sqrt.rn.f32 %f819, %f818;
|
|
rcp.rn.f32 %f820, %f819;
|
|
mul.f32 %f821, %f813, %f820;
|
|
mul.f32 %f822, %f814, %f820;
|
|
mul.f32 %f823, %f815, %f820;
|
|
mul.f32 %f824, %f1957, %f823;
|
|
mul.f32 %f825, %f1958, %f822;
|
|
sub.f32 %f826, %f824, %f825;
|
|
mul.f32 %f827, %f1958, %f821;
|
|
mul.f32 %f828, %f1956, %f823;
|
|
sub.f32 %f829, %f827, %f828;
|
|
mul.f32 %f830, %f1956, %f822;
|
|
mul.f32 %f831, %f1957, %f821;
|
|
sub.f32 %f832, %f830, %f831;
|
|
mul.f32 %f833, %f821, 0fBF61E1E3;
|
|
mul.f32 %f834, %f822, 0fBF61E1E3;
|
|
mul.f32 %f835, %f823, 0fBF61E1E3;
|
|
fma.rn.f32 %f836, %f23, %f833, %f14;
|
|
fma.rn.f32 %f837, %f23, %f834, %f15;
|
|
fma.rn.f32 %f838, %f23, %f835, %f16;
|
|
mul.f32 %f839, %f826, 0fBE24A490;
|
|
mul.f32 %f840, %f829, 0fBE24A490;
|
|
mul.f32 %f841, %f832, 0fBE24A490;
|
|
fma.rn.f32 %f105, %f23, %f839, %f836;
|
|
fma.rn.f32 %f106, %f23, %f840, %f837;
|
|
fma.rn.f32 %f107, %f23, %f841, %f838;
|
|
sub.f32 %f842, %f105, %f14;
|
|
sub.f32 %f843, %f106, %f15;
|
|
sub.f32 %f844, %f107, %f16;
|
|
mul.f32 %f845, %f843, %f843;
|
|
fma.rn.f32 %f846, %f842, %f842, %f845;
|
|
fma.rn.f32 %f847, %f844, %f844, %f846;
|
|
sqrt.rn.f32 %f802, %f847;
|
|
rcp.rn.f32 %f848, %f802;
|
|
mul.f32 %f798, %f848, %f842;
|
|
mul.f32 %f799, %f848, %f843;
|
|
mul.f32 %f800, %f848, %f844;
|
|
mov.f32 %f803, 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,%f798,%f799,%f800,%f438,%f802,%f803,%r772,%r804,%r772,%r770,%r772,%r772,%r773,%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,%r804,%r804);
|
|
// end inline asm
|
|
mov.b32 %f849, %r879;
|
|
setp.neu.f32 %p66, %f849, 0f3F800000;
|
|
@%p66 bra $L__BB0_44;
|
|
|
|
mov.f32 %f857, 0f6C4ECB8F;
|
|
mov.u32 %r989, 2;
|
|
mov.u32 %r991, 1;
|
|
mov.u32 %r992, 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,%f105,%f106,%f107,%f66,%f67,%f68,%f438,%f857,%f803,%r991,%r1023,%r991,%r989,%r991,%r991,%r992,%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,%r1023,%r1023);
|
|
// end inline asm
|
|
mov.b32 %f859, %r953;
|
|
add.f32 %f1969, %f1969, %f859;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_44:
|
|
mad.lo.s32 %r1095, %r53, 1664525, 1013904223;
|
|
and.b32 %r1096, %r1095, 16777215;
|
|
cvt.rn.f32.u32 %f869, %r1096;
|
|
mul.f32 %f870, %f869, 0f33800000;
|
|
fma.rn.f32 %f871, %f870, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r54, %r1095, 1664525, 1013904223;
|
|
and.b32 %r1097, %r54, 16777215;
|
|
cvt.rn.f32.u32 %f872, %r1097;
|
|
mul.f32 %f873, %f872, 0f33800000;
|
|
fma.rn.f32 %f874, %f873, 0f40000000, 0fBF800000;
|
|
mul.f32 %f875, %f17, %f874;
|
|
mul.f32 %f876, %f18, %f874;
|
|
mul.f32 %f877, %f19, %f874;
|
|
fma.rn.f32 %f878, %f20, %f871, %f875;
|
|
fma.rn.f32 %f879, %f21, %f871, %f876;
|
|
fma.rn.f32 %f880, %f22, %f871, %f877;
|
|
mul.f32 %f881, %f879, %f879;
|
|
fma.rn.f32 %f882, %f878, %f878, %f881;
|
|
fma.rn.f32 %f883, %f880, %f880, %f882;
|
|
sqrt.rn.f32 %f884, %f883;
|
|
rcp.rn.f32 %f885, %f884;
|
|
mul.f32 %f886, %f878, %f885;
|
|
mul.f32 %f887, %f879, %f885;
|
|
mul.f32 %f888, %f880, %f885;
|
|
mul.f32 %f889, %f1957, %f888;
|
|
mul.f32 %f890, %f1958, %f887;
|
|
sub.f32 %f891, %f889, %f890;
|
|
mul.f32 %f892, %f1958, %f886;
|
|
mul.f32 %f893, %f1956, %f888;
|
|
sub.f32 %f894, %f892, %f893;
|
|
mul.f32 %f895, %f1956, %f887;
|
|
mul.f32 %f896, %f1957, %f886;
|
|
sub.f32 %f897, %f895, %f896;
|
|
mul.f32 %f898, %f886, 0f3ECACABC;
|
|
mul.f32 %f899, %f887, 0f3ECACABC;
|
|
mul.f32 %f900, %f888, 0f3ECACABC;
|
|
fma.rn.f32 %f901, %f23, %f898, %f14;
|
|
fma.rn.f32 %f902, %f23, %f899, %f15;
|
|
fma.rn.f32 %f903, %f23, %f900, %f16;
|
|
mul.f32 %f904, %f891, 0f3F61E1D2;
|
|
mul.f32 %f905, %f894, 0f3F61E1D2;
|
|
mul.f32 %f906, %f897, 0f3F61E1D2;
|
|
fma.rn.f32 %f112, %f23, %f904, %f901;
|
|
fma.rn.f32 %f113, %f23, %f905, %f902;
|
|
fma.rn.f32 %f114, %f23, %f906, %f903;
|
|
sub.f32 %f907, %f112, %f14;
|
|
sub.f32 %f908, %f113, %f15;
|
|
sub.f32 %f909, %f114, %f16;
|
|
mul.f32 %f910, %f908, %f908;
|
|
fma.rn.f32 %f911, %f907, %f907, %f910;
|
|
fma.rn.f32 %f912, %f909, %f909, %f911;
|
|
sqrt.rn.f32 %f867, %f912;
|
|
rcp.rn.f32 %f913, %f867;
|
|
mul.f32 %f863, %f913, %f907;
|
|
mul.f32 %f864, %f913, %f908;
|
|
mul.f32 %f865, %f913, %f909;
|
|
mov.u32 %r1060, 2;
|
|
mov.u32 %r1062, 1;
|
|
mov.u32 %r1063, 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,%f863,%f864,%f865,%f438,%f867,%f803,%r1062,%r1094,%r1062,%r1060,%r1062,%r1062,%r1063,%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,%r1094,%r1094);
|
|
// end inline asm
|
|
mov.b32 %f914, %r1024;
|
|
setp.neu.f32 %p67, %f914, 0f3F800000;
|
|
@%p67 bra $L__BB0_46;
|
|
|
|
mov.f32 %f922, 0f6C4ECB8F;
|
|
mov.f32 %f923, 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,%f112,%f113,%f114,%f66,%f67,%f68,%f438,%f922,%f923,%r1062,%r1094,%r1062,%r1060,%r1062,%r1062,%r1063,%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,%r1094,%r1094);
|
|
// end inline asm
|
|
mov.b32 %f924, %r1098;
|
|
add.f32 %f1969, %f1969, %f924;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_46:
|
|
mad.lo.s32 %r1240, %r54, 1664525, 1013904223;
|
|
and.b32 %r1241, %r1240, 16777215;
|
|
cvt.rn.f32.u32 %f934, %r1241;
|
|
mul.f32 %f935, %f934, 0f33800000;
|
|
fma.rn.f32 %f936, %f935, 0f40000000, 0fBF800000;
|
|
mad.lo.s32 %r1542, %r1240, 1664525, 1013904223;
|
|
and.b32 %r1242, %r1542, 16777215;
|
|
cvt.rn.f32.u32 %f937, %r1242;
|
|
mul.f32 %f938, %f937, 0f33800000;
|
|
fma.rn.f32 %f939, %f938, 0f40000000, 0fBF800000;
|
|
mul.f32 %f940, %f17, %f939;
|
|
mul.f32 %f941, %f18, %f939;
|
|
mul.f32 %f942, %f19, %f939;
|
|
fma.rn.f32 %f943, %f20, %f936, %f940;
|
|
fma.rn.f32 %f944, %f21, %f936, %f941;
|
|
fma.rn.f32 %f945, %f22, %f936, %f942;
|
|
mul.f32 %f946, %f944, %f944;
|
|
fma.rn.f32 %f947, %f943, %f943, %f946;
|
|
fma.rn.f32 %f948, %f945, %f945, %f947;
|
|
sqrt.rn.f32 %f949, %f948;
|
|
rcp.rn.f32 %f950, %f949;
|
|
mul.f32 %f951, %f943, %f950;
|
|
mul.f32 %f952, %f944, %f950;
|
|
mul.f32 %f953, %f945, %f950;
|
|
mul.f32 %f954, %f1957, %f953;
|
|
mul.f32 %f955, %f1958, %f952;
|
|
sub.f32 %f956, %f954, %f955;
|
|
mul.f32 %f957, %f1958, %f951;
|
|
mul.f32 %f958, %f1956, %f953;
|
|
sub.f32 %f959, %f957, %f958;
|
|
mul.f32 %f960, %f1956, %f952;
|
|
mul.f32 %f961, %f1957, %f951;
|
|
sub.f32 %f962, %f960, %f961;
|
|
mul.f32 %f963, %f951, 0f3F61E1D2;
|
|
mul.f32 %f964, %f952, 0f3F61E1D2;
|
|
mul.f32 %f965, %f953, 0f3F61E1D2;
|
|
fma.rn.f32 %f966, %f23, %f963, %f14;
|
|
fma.rn.f32 %f967, %f23, %f964, %f15;
|
|
fma.rn.f32 %f968, %f23, %f965, %f16;
|
|
mul.f32 %f969, %f956, 0fBF61E1E3;
|
|
mul.f32 %f970, %f959, 0fBF61E1E3;
|
|
mul.f32 %f971, %f962, 0fBF61E1E3;
|
|
fma.rn.f32 %f119, %f23, %f969, %f966;
|
|
fma.rn.f32 %f120, %f23, %f970, %f967;
|
|
fma.rn.f32 %f121, %f23, %f971, %f968;
|
|
sub.f32 %f972, %f119, %f14;
|
|
sub.f32 %f973, %f120, %f15;
|
|
sub.f32 %f974, %f121, %f16;
|
|
mul.f32 %f975, %f973, %f973;
|
|
fma.rn.f32 %f976, %f972, %f972, %f975;
|
|
fma.rn.f32 %f977, %f974, %f974, %f976;
|
|
sqrt.rn.f32 %f932, %f977;
|
|
rcp.rn.f32 %f978, %f932;
|
|
mul.f32 %f928, %f978, %f972;
|
|
mul.f32 %f929, %f978, %f973;
|
|
mul.f32 %f930, %f978, %f974;
|
|
mov.f32 %f933, 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,%f928,%f929,%f930,%f438,%f932,%f933,%r1062,%r1094,%r1062,%r1060,%r1062,%r1062,%r1063,%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,%r1094,%r1094);
|
|
// end inline asm
|
|
mov.b32 %f979, %r1169;
|
|
setp.neu.f32 %p68, %f979, 0f3F800000;
|
|
@%p68 bra $L__BB0_48;
|
|
|
|
mov.f32 %f987, 0f6C4ECB8F;
|
|
mov.u32 %r1279, 2;
|
|
mov.u32 %r1281, 1;
|
|
mov.u32 %r1282, 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,%f119,%f120,%f121,%f66,%f67,%f68,%f438,%f987,%f933,%r1281,%r1313,%r1281,%r1279,%r1281,%r1281,%r1282,%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,%r1313,%r1313);
|
|
// end inline asm
|
|
mov.b32 %f989, %r1243;
|
|
add.f32 %f1969, %f1969, %f989;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
mov.u16 %rs195, 1;
|
|
|
|
$L__BB0_48:
|
|
setp.ne.s16 %p69, %rs195, 0;
|
|
@%p69 bra $L__BB0_50;
|
|
|
|
mov.f32 %f997, 0f6C4ECB8F;
|
|
mov.f32 %f998, 0f00000000;
|
|
mov.u32 %r1350, 2;
|
|
mov.u32 %r1352, 1;
|
|
mov.u32 %r1353, 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,%f66,%f67,%f68,%f438,%f997,%f998,%r1352,%r1384,%r1352,%r1350,%r1352,%r1352,%r1353,%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,%r1384,%r1384);
|
|
// end inline asm
|
|
mov.b32 %f999, %r1314;
|
|
add.f32 %f1969, %f1969, %f999;
|
|
add.f32 %f1970, %f1970, 0f3F800000;
|
|
|
|
$L__BB0_50:
|
|
add.s32 %r1541, %r1541, 1;
|
|
setp.lt.s32 %p70, %r1541, %r7;
|
|
@%p70 bra $L__BB0_8;
|
|
|
|
add.s32 %r1539, %r1539, 1;
|
|
setp.lt.s32 %p71, %r1539, %r7;
|
|
@%p71 bra $L__BB0_7;
|
|
|
|
$L__BB0_52:
|
|
setp.eq.s32 %p72, %r7, 0;
|
|
mov.f32 %f1989, 0f3F800000;
|
|
@%p72 bra $L__BB0_54;
|
|
|
|
mov.f32 %f1001, 0f3F800000;
|
|
max.f32 %f1002, %f1970, %f1001;
|
|
div.rn.f32 %f1989, %f1969, %f1002;
|
|
|
|
$L__BB0_54:
|
|
ld.const.f32 %f1004, [params+524];
|
|
mul.f32 %f134, %f1989, %f1004;
|
|
ld.const.v2.f32 {%f1005, %f1006}, [params+528];
|
|
mul.f32 %f135, %f1989, %f1005;
|
|
mul.f32 %f136, %f1989, %f1006;
|
|
cvt.sat.f32.f32 %f1009, %f13;
|
|
mul.f32 %f137, %f134, %f1009;
|
|
mul.f32 %f138, %f135, %f1009;
|
|
mul.f32 %f139, %f136, %f1009;
|
|
fma.rn.f32 %f1010, %f13, 0f3F000000, 0f3F000000;
|
|
cvt.sat.f32.f32 %f1011, %f1010;
|
|
add.f32 %f140, %f1011, %f1011;
|
|
mov.f32 %f1012, 0f41200000;
|
|
cvt.rzi.f32.f32 %f1013, %f1012;
|
|
add.f32 %f1014, %f1013, %f1013;
|
|
mov.f32 %f1015, 0f41A00000;
|
|
sub.f32 %f1016, %f1015, %f1014;
|
|
abs.f32 %f141, %f1016;
|
|
abs.f32 %f142, %f140;
|
|
setp.lt.f32 %p73, %f142, 0f00800000;
|
|
mul.f32 %f1017, %f142, 0f4B800000;
|
|
selp.f32 %f1018, %f1017, %f142, %p73;
|
|
selp.f32 %f1019, 0fC3170000, 0fC2FE0000, %p73;
|
|
mov.b32 %r1385, %f1018;
|
|
and.b32 %r1386, %r1385, 8388607;
|
|
or.b32 %r1387, %r1386, 1065353216;
|
|
mov.b32 %f1020, %r1387;
|
|
shr.u32 %r1388, %r1385, 23;
|
|
cvt.rn.f32.u32 %f1021, %r1388;
|
|
add.f32 %f1022, %f1019, %f1021;
|
|
setp.gt.f32 %p74, %f1020, 0f3FB504F3;
|
|
mul.f32 %f1023, %f1020, 0f3F000000;
|
|
add.f32 %f1024, %f1022, 0f3F800000;
|
|
selp.f32 %f1025, %f1024, %f1022, %p74;
|
|
selp.f32 %f1026, %f1023, %f1020, %p74;
|
|
add.f32 %f1027, %f1026, 0fBF800000;
|
|
add.f32 %f1028, %f1026, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1029, %f1028;
|
|
add.f32 %f1030, %f1027, %f1027;
|
|
mul.f32 %f1031, %f1030, %f1029;
|
|
mul.f32 %f1032, %f1031, %f1031;
|
|
mov.f32 %f1033, 0f3C4CAF63;
|
|
mov.f32 %f1034, 0f3B18F0FE;
|
|
fma.rn.f32 %f1035, %f1034, %f1032, %f1033;
|
|
mov.f32 %f1036, 0f3DAAAABD;
|
|
fma.rn.f32 %f1037, %f1035, %f1032, %f1036;
|
|
mul.rn.f32 %f1038, %f1037, %f1032;
|
|
mul.rn.f32 %f1039, %f1038, %f1031;
|
|
sub.f32 %f1040, %f1027, %f1031;
|
|
add.f32 %f1041, %f1040, %f1040;
|
|
neg.f32 %f1042, %f1031;
|
|
fma.rn.f32 %f1043, %f1042, %f1027, %f1041;
|
|
mul.rn.f32 %f1044, %f1029, %f1043;
|
|
add.f32 %f1045, %f1039, %f1031;
|
|
sub.f32 %f1046, %f1031, %f1045;
|
|
add.f32 %f1047, %f1039, %f1046;
|
|
add.f32 %f1048, %f1044, %f1047;
|
|
add.f32 %f1049, %f1045, %f1048;
|
|
sub.f32 %f1050, %f1045, %f1049;
|
|
add.f32 %f1051, %f1048, %f1050;
|
|
mov.f32 %f1052, 0f3F317200;
|
|
mul.rn.f32 %f1053, %f1025, %f1052;
|
|
mov.f32 %f1054, 0f35BFBE8E;
|
|
mul.rn.f32 %f1055, %f1025, %f1054;
|
|
add.f32 %f1056, %f1053, %f1049;
|
|
sub.f32 %f1057, %f1053, %f1056;
|
|
add.f32 %f1058, %f1049, %f1057;
|
|
add.f32 %f1059, %f1051, %f1058;
|
|
add.f32 %f1060, %f1055, %f1059;
|
|
add.f32 %f1061, %f1056, %f1060;
|
|
sub.f32 %f1062, %f1056, %f1061;
|
|
add.f32 %f1063, %f1060, %f1062;
|
|
mul.rn.f32 %f1064, %f1015, %f1061;
|
|
neg.f32 %f1065, %f1064;
|
|
fma.rn.f32 %f1066, %f1015, %f1061, %f1065;
|
|
fma.rn.f32 %f1067, %f1015, %f1063, %f1066;
|
|
mov.f32 %f1068, 0f00000000;
|
|
fma.rn.f32 %f1069, %f1068, %f1061, %f1067;
|
|
add.rn.f32 %f1070, %f1064, %f1069;
|
|
neg.f32 %f1071, %f1070;
|
|
add.rn.f32 %f1072, %f1064, %f1071;
|
|
add.rn.f32 %f1073, %f1072, %f1069;
|
|
mov.b32 %r1389, %f1070;
|
|
setp.eq.s32 %p75, %r1389, 1118925336;
|
|
add.s32 %r1390, %r1389, -1;
|
|
mov.b32 %f1074, %r1390;
|
|
add.f32 %f1075, %f1073, 0f37000000;
|
|
selp.f32 %f143, %f1075, %f1073, %p75;
|
|
selp.f32 %f1076, %f1074, %f1070, %p75;
|
|
mov.f32 %f1077, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1078, %f1076, %f1077;
|
|
cvt.rzi.f32.f32 %f1079, %f1078;
|
|
abs.f32 %f1080, %f1079;
|
|
setp.gt.f32 %p76, %f1080, 0f42FC0000;
|
|
mov.b32 %r1391, %f1079;
|
|
and.b32 %r1392, %r1391, -2147483648;
|
|
or.b32 %r1393, %r1392, 1123811328;
|
|
mov.b32 %f1081, %r1393;
|
|
selp.f32 %f1082, %f1081, %f1079, %p76;
|
|
mov.f32 %f1083, 0fBF317218;
|
|
fma.rn.f32 %f1084, %f1082, %f1083, %f1076;
|
|
mov.f32 %f1085, 0f3102E308;
|
|
fma.rn.f32 %f1086, %f1082, %f1085, %f1084;
|
|
mul.f32 %f1087, %f1086, 0f3FB8AA3B;
|
|
add.f32 %f1088, %f1082, 0f4B40007F;
|
|
mov.b32 %r1394, %f1088;
|
|
shl.b32 %r1395, %r1394, 23;
|
|
mov.b32 %f1089, %r1395;
|
|
ex2.approx.ftz.f32 %f1090, %f1087;
|
|
mul.f32 %f144, %f1090, %f1089;
|
|
setp.eq.f32 %p77, %f144, 0f7F800000;
|
|
mov.f32 %f1990, 0f7F800000;
|
|
@%p77 bra $L__BB0_56;
|
|
|
|
fma.rn.f32 %f1990, %f144, %f143, %f144;
|
|
|
|
$L__BB0_56:
|
|
setp.lt.f32 %p78, %f140, 0f00000000;
|
|
setp.eq.f32 %p79, %f141, 0f3F800000;
|
|
and.pred %p1, %p78, %p79;
|
|
setp.eq.f32 %p80, %f140, 0f00000000;
|
|
@%p80 bra $L__BB0_60;
|
|
bra.uni $L__BB0_57;
|
|
|
|
$L__BB0_60:
|
|
add.f32 %f1095, %f140, %f140;
|
|
selp.f32 %f1992, %f1095, 0f00000000, %p79;
|
|
bra.uni $L__BB0_61;
|
|
|
|
$L__BB0_57:
|
|
mov.b32 %r1396, %f1990;
|
|
xor.b32 %r1397, %r1396, -2147483648;
|
|
mov.b32 %f1091, %r1397;
|
|
selp.f32 %f1992, %f1091, %f1990, %p1;
|
|
setp.geu.f32 %p81, %f140, 0f00000000;
|
|
@%p81 bra $L__BB0_61;
|
|
|
|
mov.f32 %f1092, 0f41A00000;
|
|
cvt.rzi.f32.f32 %f1093, %f1092;
|
|
setp.eq.f32 %p82, %f1093, 0f41A00000;
|
|
@%p82 bra $L__BB0_61;
|
|
|
|
mov.f32 %f1992, 0f7FFFFFFF;
|
|
|
|
$L__BB0_61:
|
|
add.f32 %f1096, %f142, 0f41A00000;
|
|
mov.b32 %r1398, %f1096;
|
|
setp.lt.s32 %p84, %r1398, 2139095040;
|
|
@%p84 bra $L__BB0_66;
|
|
|
|
setp.gtu.f32 %p85, %f142, 0f7F800000;
|
|
@%p85 bra $L__BB0_65;
|
|
bra.uni $L__BB0_63;
|
|
|
|
$L__BB0_65:
|
|
add.f32 %f1992, %f140, 0f41A00000;
|
|
bra.uni $L__BB0_66;
|
|
|
|
$L__BB0_63:
|
|
setp.neu.f32 %p86, %f142, 0f7F800000;
|
|
@%p86 bra $L__BB0_66;
|
|
|
|
selp.f32 %f1992, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
$L__BB0_66:
|
|
setp.eq.f32 %p87, %f140, 0f3F800000;
|
|
selp.f32 %f1097, 0f3F800000, %f1992, %p87;
|
|
cvt.sat.f32.f32 %f1098, %f1097;
|
|
mul.f32 %f153, %f134, %f1098;
|
|
mul.f32 %f154, %f135, %f1098;
|
|
mul.f32 %f155, %f136, %f1098;
|
|
ld.const.u32 %r58, [params+104];
|
|
and.b32 %r1399, %r58, 8;
|
|
setp.eq.s32 %p88, %r1399, 0;
|
|
@%p88 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 %f1100, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f1101, %f1100;
|
|
add.f32 %f1102, %f1101, %f1101;
|
|
mov.f32 %f1103, 0f3EE8BA2E;
|
|
sub.f32 %f1104, %f1103, %f1102;
|
|
abs.f32 %f156, %f1104;
|
|
abs.f32 %f157, %f1989;
|
|
setp.lt.f32 %p89, %f157, 0f00800000;
|
|
mul.f32 %f1105, %f157, 0f4B800000;
|
|
selp.f32 %f1106, %f1105, %f157, %p89;
|
|
selp.f32 %f1107, 0fC3170000, 0fC2FE0000, %p89;
|
|
mov.b32 %r1402, %f1106;
|
|
and.b32 %r1403, %r1402, 8388607;
|
|
or.b32 %r1404, %r1403, 1065353216;
|
|
mov.b32 %f1108, %r1404;
|
|
shr.u32 %r1405, %r1402, 23;
|
|
cvt.rn.f32.u32 %f1109, %r1405;
|
|
add.f32 %f1110, %f1107, %f1109;
|
|
setp.gt.f32 %p90, %f1108, 0f3FB504F3;
|
|
mul.f32 %f1111, %f1108, 0f3F000000;
|
|
add.f32 %f1112, %f1110, 0f3F800000;
|
|
selp.f32 %f1113, %f1112, %f1110, %p90;
|
|
selp.f32 %f1114, %f1111, %f1108, %p90;
|
|
add.f32 %f1115, %f1114, 0fBF800000;
|
|
add.f32 %f1116, %f1114, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1117, %f1116;
|
|
add.f32 %f1118, %f1115, %f1115;
|
|
mul.f32 %f1119, %f1118, %f1117;
|
|
mul.f32 %f1120, %f1119, %f1119;
|
|
mov.f32 %f1121, 0f3C4CAF63;
|
|
mov.f32 %f1122, 0f3B18F0FE;
|
|
fma.rn.f32 %f1123, %f1122, %f1120, %f1121;
|
|
mov.f32 %f1124, 0f3DAAAABD;
|
|
fma.rn.f32 %f1125, %f1123, %f1120, %f1124;
|
|
mul.rn.f32 %f1126, %f1125, %f1120;
|
|
mul.rn.f32 %f1127, %f1126, %f1119;
|
|
sub.f32 %f1128, %f1115, %f1119;
|
|
add.f32 %f1129, %f1128, %f1128;
|
|
neg.f32 %f1130, %f1119;
|
|
fma.rn.f32 %f1131, %f1130, %f1115, %f1129;
|
|
mul.rn.f32 %f1132, %f1117, %f1131;
|
|
add.f32 %f1133, %f1127, %f1119;
|
|
sub.f32 %f1134, %f1119, %f1133;
|
|
add.f32 %f1135, %f1127, %f1134;
|
|
add.f32 %f1136, %f1132, %f1135;
|
|
add.f32 %f1137, %f1133, %f1136;
|
|
sub.f32 %f1138, %f1133, %f1137;
|
|
add.f32 %f1139, %f1136, %f1138;
|
|
mov.f32 %f1140, 0f3F317200;
|
|
mul.rn.f32 %f1141, %f1113, %f1140;
|
|
mov.f32 %f1142, 0f35BFBE8E;
|
|
mul.rn.f32 %f1143, %f1113, %f1142;
|
|
add.f32 %f1144, %f1141, %f1137;
|
|
sub.f32 %f1145, %f1141, %f1144;
|
|
add.f32 %f1146, %f1137, %f1145;
|
|
add.f32 %f1147, %f1139, %f1146;
|
|
add.f32 %f1148, %f1143, %f1147;
|
|
add.f32 %f1149, %f1144, %f1148;
|
|
sub.f32 %f1150, %f1144, %f1149;
|
|
add.f32 %f1151, %f1148, %f1150;
|
|
mul.rn.f32 %f1152, %f1103, %f1149;
|
|
neg.f32 %f1153, %f1152;
|
|
fma.rn.f32 %f1154, %f1103, %f1149, %f1153;
|
|
fma.rn.f32 %f1155, %f1103, %f1151, %f1154;
|
|
mov.f32 %f1156, 0f00000000;
|
|
fma.rn.f32 %f1157, %f1156, %f1149, %f1155;
|
|
add.rn.f32 %f1158, %f1152, %f1157;
|
|
neg.f32 %f1159, %f1158;
|
|
add.rn.f32 %f1160, %f1152, %f1159;
|
|
add.rn.f32 %f1161, %f1160, %f1157;
|
|
mov.b32 %r1406, %f1158;
|
|
setp.eq.s32 %p91, %r1406, 1118925336;
|
|
add.s32 %r1407, %r1406, -1;
|
|
mov.b32 %f1162, %r1407;
|
|
add.f32 %f1163, %f1161, 0f37000000;
|
|
selp.f32 %f158, %f1163, %f1161, %p91;
|
|
selp.f32 %f1164, %f1162, %f1158, %p91;
|
|
mov.f32 %f1165, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1166, %f1164, %f1165;
|
|
cvt.rzi.f32.f32 %f1167, %f1166;
|
|
abs.f32 %f1168, %f1167;
|
|
setp.gt.f32 %p92, %f1168, 0f42FC0000;
|
|
mov.b32 %r1408, %f1167;
|
|
and.b32 %r1409, %r1408, -2147483648;
|
|
or.b32 %r1410, %r1409, 1123811328;
|
|
mov.b32 %f1169, %r1410;
|
|
selp.f32 %f1170, %f1169, %f1167, %p92;
|
|
mov.f32 %f1171, 0fBF317218;
|
|
fma.rn.f32 %f1172, %f1170, %f1171, %f1164;
|
|
mov.f32 %f1173, 0f3102E308;
|
|
fma.rn.f32 %f1174, %f1170, %f1173, %f1172;
|
|
mul.f32 %f1175, %f1174, 0f3FB8AA3B;
|
|
add.f32 %f1176, %f1170, 0f4B40007F;
|
|
mov.b32 %r1411, %f1176;
|
|
shl.b32 %r1412, %r1411, 23;
|
|
mov.b32 %f1177, %r1412;
|
|
ex2.approx.ftz.f32 %f1178, %f1175;
|
|
mul.f32 %f159, %f1178, %f1177;
|
|
setp.eq.f32 %p93, %f159, 0f7F800000;
|
|
mov.f32 %f1993, 0f7F800000;
|
|
@%p93 bra $L__BB0_69;
|
|
|
|
fma.rn.f32 %f1993, %f159, %f158, %f159;
|
|
|
|
$L__BB0_69:
|
|
setp.lt.f32 %p94, %f1989, 0f00000000;
|
|
setp.eq.f32 %p95, %f156, 0f3F800000;
|
|
and.pred %p2, %p94, %p95;
|
|
setp.eq.f32 %p96, %f1989, 0f00000000;
|
|
@%p96 bra $L__BB0_73;
|
|
bra.uni $L__BB0_70;
|
|
|
|
$L__BB0_73:
|
|
add.f32 %f1183, %f1989, %f1989;
|
|
selp.f32 %f1995, %f1183, 0f00000000, %p95;
|
|
bra.uni $L__BB0_74;
|
|
|
|
$L__BB0_182:
|
|
mov.f32 %f1927, 0f00000000;
|
|
mov.u32 %r1550, 0;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs145, %f1927;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f1927;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f1927;}
|
|
|
|
// 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 %p221, %r1550, 0;
|
|
@%p221 bra $L__BB0_185;
|
|
|
|
ld.global.v4.u16 {%rs153, %rs154, %rs155, %rs156}, [%rd25];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1928, %rs153;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1929, %rs154;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1930, %rs155;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1931, %f1928, 0f00000000;
|
|
add.f32 %f1932, %f1929, 0f00000000;
|
|
add.f32 %f1933, %f1930, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f1933;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f1932;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f1931;}
|
|
|
|
// 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 %f1936, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs160, %f1936;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs159, %f1936;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs158, %f1936;}
|
|
|
|
// 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;
|
|
@%p221 bra $L__BB0_188;
|
|
|
|
ld.global.v4.u16 {%rs168, %rs169, %rs170, %rs171}, [%rd26];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1937, %rs168;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1938, %rs169;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1939, %rs170;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1940, %f1937, 0f00000000;
|
|
add.f32 %f1941, %f1938, 0f00000000;
|
|
add.f32 %f1942, %f1939, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs167, %f1942;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs166, %f1941;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs165, %f1940;}
|
|
|
|
// 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 %f1945, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs175, %f1945;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs174, %f1945;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f1945;}
|
|
|
|
// 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;
|
|
@%p221 bra $L__BB0_191;
|
|
|
|
ld.global.v4.u16 {%rs183, %rs184, %rs185, %rs186}, [%rd27];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1946, %rs183;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1947, %rs184;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1948, %rs185;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1949, %f1946, 0f00000000;
|
|
add.f32 %f1950, %f1947, 0f00000000;
|
|
add.f32 %f1951, %f1948, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs182, %f1951;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs181, %f1950;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs180, %f1949;}
|
|
|
|
// 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 %f1954, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs190, %f1954;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs189, %f1954;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs188, %f1954;}
|
|
|
|
// 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, %f1993;
|
|
xor.b32 %r1414, %r1413, -2147483648;
|
|
mov.b32 %f1179, %r1414;
|
|
selp.f32 %f1995, %f1179, %f1993, %p2;
|
|
setp.geu.f32 %p97, %f1989, 0f00000000;
|
|
@%p97 bra $L__BB0_74;
|
|
|
|
mov.f32 %f1180, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f1181, %f1180;
|
|
setp.eq.f32 %p98, %f1181, 0f3EE8BA2E;
|
|
@%p98 bra $L__BB0_74;
|
|
|
|
mov.f32 %f1995, 0f7FFFFFFF;
|
|
|
|
$L__BB0_74:
|
|
add.f32 %f1184, %f157, 0f3EE8BA2E;
|
|
mov.b32 %r1415, %f1184;
|
|
setp.lt.s32 %p100, %r1415, 2139095040;
|
|
@%p100 bra $L__BB0_79;
|
|
|
|
setp.gtu.f32 %p101, %f157, 0f7F800000;
|
|
@%p101 bra $L__BB0_78;
|
|
bra.uni $L__BB0_76;
|
|
|
|
$L__BB0_78:
|
|
add.f32 %f1995, %f1989, 0f3EE8BA2E;
|
|
bra.uni $L__BB0_79;
|
|
|
|
$L__BB0_76:
|
|
setp.neu.f32 %p102, %f157, 0f7F800000;
|
|
@%p102 bra $L__BB0_79;
|
|
|
|
selp.f32 %f1995, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
$L__BB0_79:
|
|
mul.f32 %f1185, %f1995, 0f437F0000;
|
|
setp.eq.f32 %p103, %f1989, 0f3F800000;
|
|
selp.f32 %f1186, 0f437F0000, %f1185, %p103;
|
|
cvt.rzi.u32.f32 %r1416, %f1186;
|
|
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 %p104, %r1417, 1;
|
|
mov.pred %p105, 0;
|
|
xor.pred %p106, %p104, %p105;
|
|
not.pred %p107, %p106;
|
|
@%p107 bra $L__BB0_154;
|
|
|
|
mov.f32 %f1188, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f1189, %f1188;
|
|
add.f32 %f1190, %f1189, %f1189;
|
|
mov.f32 %f1191, 0f3EE66666;
|
|
sub.f32 %f1192, %f1191, %f1190;
|
|
abs.f32 %f168, %f1192;
|
|
abs.f32 %f169, %f137;
|
|
setp.lt.f32 %p108, %f169, 0f00800000;
|
|
mul.f32 %f1193, %f169, 0f4B800000;
|
|
selp.f32 %f1194, %f1193, %f169, %p108;
|
|
selp.f32 %f1195, 0fC3170000, 0fC2FE0000, %p108;
|
|
mov.b32 %r1418, %f1194;
|
|
and.b32 %r1419, %r1418, 8388607;
|
|
or.b32 %r1420, %r1419, 1065353216;
|
|
mov.b32 %f1196, %r1420;
|
|
shr.u32 %r1421, %r1418, 23;
|
|
cvt.rn.f32.u32 %f1197, %r1421;
|
|
add.f32 %f1198, %f1195, %f1197;
|
|
setp.gt.f32 %p109, %f1196, 0f3FB504F3;
|
|
mul.f32 %f1199, %f1196, 0f3F000000;
|
|
add.f32 %f1200, %f1198, 0f3F800000;
|
|
selp.f32 %f1201, %f1200, %f1198, %p109;
|
|
selp.f32 %f1202, %f1199, %f1196, %p109;
|
|
add.f32 %f1203, %f1202, 0fBF800000;
|
|
add.f32 %f1204, %f1202, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1205, %f1204;
|
|
add.f32 %f1206, %f1203, %f1203;
|
|
mul.f32 %f1207, %f1206, %f1205;
|
|
mul.f32 %f1208, %f1207, %f1207;
|
|
mov.f32 %f1209, 0f3C4CAF63;
|
|
mov.f32 %f1210, 0f3B18F0FE;
|
|
fma.rn.f32 %f1211, %f1210, %f1208, %f1209;
|
|
mov.f32 %f1212, 0f3DAAAABD;
|
|
fma.rn.f32 %f1213, %f1211, %f1208, %f1212;
|
|
mul.rn.f32 %f1214, %f1213, %f1208;
|
|
mul.rn.f32 %f1215, %f1214, %f1207;
|
|
sub.f32 %f1216, %f1203, %f1207;
|
|
add.f32 %f1217, %f1216, %f1216;
|
|
neg.f32 %f1218, %f1207;
|
|
fma.rn.f32 %f1219, %f1218, %f1203, %f1217;
|
|
mul.rn.f32 %f1220, %f1205, %f1219;
|
|
add.f32 %f1221, %f1215, %f1207;
|
|
sub.f32 %f1222, %f1207, %f1221;
|
|
add.f32 %f1223, %f1215, %f1222;
|
|
add.f32 %f1224, %f1220, %f1223;
|
|
add.f32 %f1225, %f1221, %f1224;
|
|
sub.f32 %f1226, %f1221, %f1225;
|
|
add.f32 %f1227, %f1224, %f1226;
|
|
mov.f32 %f1228, 0f3F317200;
|
|
mul.rn.f32 %f1229, %f1201, %f1228;
|
|
mov.f32 %f1230, 0f35BFBE8E;
|
|
mul.rn.f32 %f1231, %f1201, %f1230;
|
|
add.f32 %f1232, %f1229, %f1225;
|
|
sub.f32 %f1233, %f1229, %f1232;
|
|
add.f32 %f1234, %f1225, %f1233;
|
|
add.f32 %f1235, %f1227, %f1234;
|
|
add.f32 %f1236, %f1231, %f1235;
|
|
add.f32 %f1237, %f1232, %f1236;
|
|
sub.f32 %f1238, %f1232, %f1237;
|
|
add.f32 %f1239, %f1236, %f1238;
|
|
mul.rn.f32 %f1240, %f1191, %f1237;
|
|
neg.f32 %f1241, %f1240;
|
|
fma.rn.f32 %f1242, %f1191, %f1237, %f1241;
|
|
fma.rn.f32 %f1243, %f1191, %f1239, %f1242;
|
|
mov.f32 %f1244, 0f00000000;
|
|
fma.rn.f32 %f1245, %f1244, %f1237, %f1243;
|
|
add.rn.f32 %f1246, %f1240, %f1245;
|
|
neg.f32 %f1247, %f1246;
|
|
add.rn.f32 %f1248, %f1240, %f1247;
|
|
add.rn.f32 %f1249, %f1248, %f1245;
|
|
mov.b32 %r1422, %f1246;
|
|
setp.eq.s32 %p110, %r1422, 1118925336;
|
|
add.s32 %r1423, %r1422, -1;
|
|
mov.b32 %f1250, %r1423;
|
|
add.f32 %f1251, %f1249, 0f37000000;
|
|
selp.f32 %f170, %f1251, %f1249, %p110;
|
|
selp.f32 %f1252, %f1250, %f1246, %p110;
|
|
mov.f32 %f1253, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1254, %f1252, %f1253;
|
|
cvt.rzi.f32.f32 %f1255, %f1254;
|
|
abs.f32 %f1256, %f1255;
|
|
setp.gt.f32 %p111, %f1256, 0f42FC0000;
|
|
mov.b32 %r1424, %f1255;
|
|
and.b32 %r1425, %r1424, -2147483648;
|
|
or.b32 %r1426, %r1425, 1123811328;
|
|
mov.b32 %f1257, %r1426;
|
|
selp.f32 %f1258, %f1257, %f1255, %p111;
|
|
mov.f32 %f1259, 0fBF317218;
|
|
fma.rn.f32 %f1260, %f1258, %f1259, %f1252;
|
|
mov.f32 %f1261, 0f3102E308;
|
|
fma.rn.f32 %f1262, %f1258, %f1261, %f1260;
|
|
mul.f32 %f1263, %f1262, 0f3FB8AA3B;
|
|
add.f32 %f1264, %f1258, 0f4B40007F;
|
|
mov.b32 %r1427, %f1264;
|
|
shl.b32 %r1428, %r1427, 23;
|
|
mov.b32 %f1265, %r1428;
|
|
ex2.approx.ftz.f32 %f1266, %f1263;
|
|
mul.f32 %f171, %f1266, %f1265;
|
|
setp.eq.f32 %p112, %f171, 0f7F800000;
|
|
mov.f32 %f1996, 0f7F800000;
|
|
@%p112 bra $L__BB0_83;
|
|
|
|
fma.rn.f32 %f1996, %f171, %f170, %f171;
|
|
|
|
$L__BB0_83:
|
|
setp.lt.f32 %p113, %f137, 0f00000000;
|
|
setp.eq.f32 %p114, %f168, 0f3F800000;
|
|
and.pred %p3, %p113, %p114;
|
|
setp.eq.f32 %p115, %f137, 0f00000000;
|
|
@%p115 bra $L__BB0_87;
|
|
bra.uni $L__BB0_84;
|
|
|
|
$L__BB0_87:
|
|
add.f32 %f1271, %f137, %f137;
|
|
selp.f32 %f1998, %f1271, 0f00000000, %p114;
|
|
bra.uni $L__BB0_88;
|
|
|
|
$L__BB0_84:
|
|
mov.b32 %r1429, %f1996;
|
|
xor.b32 %r1430, %r1429, -2147483648;
|
|
mov.b32 %f1267, %r1430;
|
|
selp.f32 %f1998, %f1267, %f1996, %p3;
|
|
setp.geu.f32 %p116, %f137, 0f00000000;
|
|
@%p116 bra $L__BB0_88;
|
|
|
|
mov.f32 %f1268, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f1269, %f1268;
|
|
setp.eq.f32 %p117, %f1269, 0f3EE66666;
|
|
@%p117 bra $L__BB0_88;
|
|
|
|
mov.f32 %f1998, 0f7FFFFFFF;
|
|
|
|
$L__BB0_88:
|
|
add.f32 %f1272, %f169, 0f3EE66666;
|
|
mov.b32 %r1431, %f1272;
|
|
setp.lt.s32 %p119, %r1431, 2139095040;
|
|
@%p119 bra $L__BB0_93;
|
|
|
|
setp.gtu.f32 %p120, %f169, 0f7F800000;
|
|
@%p120 bra $L__BB0_92;
|
|
bra.uni $L__BB0_90;
|
|
|
|
$L__BB0_92:
|
|
add.f32 %f1998, %f137, 0f3EE66666;
|
|
bra.uni $L__BB0_93;
|
|
|
|
$L__BB0_90:
|
|
setp.neu.f32 %p121, %f169, 0f7F800000;
|
|
@%p121 bra $L__BB0_93;
|
|
|
|
selp.f32 %f1998, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
$L__BB0_93:
|
|
setp.eq.f32 %p122, %f137, 0f3F800000;
|
|
selp.f32 %f180, 0f3F800000, %f1998, %p122;
|
|
abs.f32 %f181, %f138;
|
|
setp.lt.f32 %p123, %f181, 0f00800000;
|
|
mul.f32 %f1274, %f181, 0f4B800000;
|
|
selp.f32 %f1275, %f1274, %f181, %p123;
|
|
selp.f32 %f1276, 0fC3170000, 0fC2FE0000, %p123;
|
|
mov.b32 %r1432, %f1275;
|
|
and.b32 %r1433, %r1432, 8388607;
|
|
or.b32 %r1434, %r1433, 1065353216;
|
|
mov.b32 %f1277, %r1434;
|
|
shr.u32 %r1435, %r1432, 23;
|
|
cvt.rn.f32.u32 %f1278, %r1435;
|
|
add.f32 %f1279, %f1276, %f1278;
|
|
setp.gt.f32 %p124, %f1277, 0f3FB504F3;
|
|
mul.f32 %f1280, %f1277, 0f3F000000;
|
|
add.f32 %f1281, %f1279, 0f3F800000;
|
|
selp.f32 %f1282, %f1281, %f1279, %p124;
|
|
selp.f32 %f1283, %f1280, %f1277, %p124;
|
|
add.f32 %f1284, %f1283, 0fBF800000;
|
|
add.f32 %f1285, %f1283, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1286, %f1285;
|
|
add.f32 %f1287, %f1284, %f1284;
|
|
mul.f32 %f1288, %f1287, %f1286;
|
|
mul.f32 %f1289, %f1288, %f1288;
|
|
mov.f32 %f1290, 0f3C4CAF63;
|
|
mov.f32 %f1291, 0f3B18F0FE;
|
|
fma.rn.f32 %f1292, %f1291, %f1289, %f1290;
|
|
mov.f32 %f1293, 0f3DAAAABD;
|
|
fma.rn.f32 %f1294, %f1292, %f1289, %f1293;
|
|
mul.rn.f32 %f1295, %f1294, %f1289;
|
|
mul.rn.f32 %f1296, %f1295, %f1288;
|
|
sub.f32 %f1297, %f1284, %f1288;
|
|
add.f32 %f1298, %f1297, %f1297;
|
|
neg.f32 %f1299, %f1288;
|
|
fma.rn.f32 %f1300, %f1299, %f1284, %f1298;
|
|
mul.rn.f32 %f1301, %f1286, %f1300;
|
|
add.f32 %f1302, %f1296, %f1288;
|
|
sub.f32 %f1303, %f1288, %f1302;
|
|
add.f32 %f1304, %f1296, %f1303;
|
|
add.f32 %f1305, %f1301, %f1304;
|
|
add.f32 %f1306, %f1302, %f1305;
|
|
sub.f32 %f1307, %f1302, %f1306;
|
|
add.f32 %f1308, %f1305, %f1307;
|
|
mov.f32 %f1309, 0f3F317200;
|
|
mul.rn.f32 %f1310, %f1282, %f1309;
|
|
mov.f32 %f1311, 0f35BFBE8E;
|
|
mul.rn.f32 %f1312, %f1282, %f1311;
|
|
add.f32 %f1313, %f1310, %f1306;
|
|
sub.f32 %f1314, %f1310, %f1313;
|
|
add.f32 %f1315, %f1306, %f1314;
|
|
add.f32 %f1316, %f1308, %f1315;
|
|
add.f32 %f1317, %f1312, %f1316;
|
|
add.f32 %f1318, %f1313, %f1317;
|
|
sub.f32 %f1319, %f1313, %f1318;
|
|
add.f32 %f1320, %f1317, %f1319;
|
|
mov.f32 %f1321, 0f3EE66666;
|
|
mul.rn.f32 %f1322, %f1321, %f1318;
|
|
neg.f32 %f1323, %f1322;
|
|
fma.rn.f32 %f1324, %f1321, %f1318, %f1323;
|
|
fma.rn.f32 %f1325, %f1321, %f1320, %f1324;
|
|
mov.f32 %f1326, 0f00000000;
|
|
fma.rn.f32 %f1327, %f1326, %f1318, %f1325;
|
|
add.rn.f32 %f1328, %f1322, %f1327;
|
|
neg.f32 %f1329, %f1328;
|
|
add.rn.f32 %f1330, %f1322, %f1329;
|
|
add.rn.f32 %f1331, %f1330, %f1327;
|
|
mov.b32 %r1436, %f1328;
|
|
setp.eq.s32 %p125, %r1436, 1118925336;
|
|
add.s32 %r1437, %r1436, -1;
|
|
mov.b32 %f1332, %r1437;
|
|
add.f32 %f1333, %f1331, 0f37000000;
|
|
selp.f32 %f182, %f1333, %f1331, %p125;
|
|
selp.f32 %f1334, %f1332, %f1328, %p125;
|
|
mov.f32 %f1335, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1336, %f1334, %f1335;
|
|
cvt.rzi.f32.f32 %f1337, %f1336;
|
|
abs.f32 %f1338, %f1337;
|
|
setp.gt.f32 %p126, %f1338, 0f42FC0000;
|
|
mov.b32 %r1438, %f1337;
|
|
and.b32 %r1439, %r1438, -2147483648;
|
|
or.b32 %r1440, %r1439, 1123811328;
|
|
mov.b32 %f1339, %r1440;
|
|
selp.f32 %f1340, %f1339, %f1337, %p126;
|
|
mov.f32 %f1341, 0fBF317218;
|
|
fma.rn.f32 %f1342, %f1340, %f1341, %f1334;
|
|
mov.f32 %f1343, 0f3102E308;
|
|
fma.rn.f32 %f1344, %f1340, %f1343, %f1342;
|
|
mul.f32 %f1345, %f1344, 0f3FB8AA3B;
|
|
add.f32 %f1346, %f1340, 0f4B40007F;
|
|
mov.b32 %r1441, %f1346;
|
|
shl.b32 %r1442, %r1441, 23;
|
|
mov.b32 %f1347, %r1442;
|
|
ex2.approx.ftz.f32 %f1348, %f1345;
|
|
mul.f32 %f183, %f1348, %f1347;
|
|
setp.eq.f32 %p127, %f183, 0f7F800000;
|
|
mov.f32 %f1999, 0f7F800000;
|
|
@%p127 bra $L__BB0_95;
|
|
|
|
fma.rn.f32 %f1999, %f183, %f182, %f183;
|
|
|
|
$L__BB0_95:
|
|
setp.lt.f32 %p128, %f138, 0f00000000;
|
|
and.pred %p4, %p128, %p114;
|
|
setp.eq.f32 %p130, %f138, 0f00000000;
|
|
@%p130 bra $L__BB0_99;
|
|
bra.uni $L__BB0_96;
|
|
|
|
$L__BB0_99:
|
|
add.f32 %f1353, %f138, %f138;
|
|
selp.f32 %f2001, %f1353, 0f00000000, %p114;
|
|
bra.uni $L__BB0_100;
|
|
|
|
$L__BB0_96:
|
|
mov.b32 %r1443, %f1999;
|
|
xor.b32 %r1444, %r1443, -2147483648;
|
|
mov.b32 %f1349, %r1444;
|
|
selp.f32 %f2001, %f1349, %f1999, %p4;
|
|
setp.geu.f32 %p131, %f138, 0f00000000;
|
|
@%p131 bra $L__BB0_100;
|
|
|
|
mov.f32 %f1350, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f1351, %f1350;
|
|
setp.eq.f32 %p132, %f1351, 0f3EE66666;
|
|
@%p132 bra $L__BB0_100;
|
|
|
|
mov.f32 %f2001, 0f7FFFFFFF;
|
|
|
|
$L__BB0_100:
|
|
add.f32 %f1354, %f181, 0f3EE66666;
|
|
mov.b32 %r1445, %f1354;
|
|
setp.lt.s32 %p134, %r1445, 2139095040;
|
|
@%p134 bra $L__BB0_105;
|
|
|
|
setp.gtu.f32 %p135, %f181, 0f7F800000;
|
|
@%p135 bra $L__BB0_104;
|
|
bra.uni $L__BB0_102;
|
|
|
|
$L__BB0_104:
|
|
add.f32 %f2001, %f138, 0f3EE66666;
|
|
bra.uni $L__BB0_105;
|
|
|
|
$L__BB0_102:
|
|
setp.neu.f32 %p136, %f181, 0f7F800000;
|
|
@%p136 bra $L__BB0_105;
|
|
|
|
selp.f32 %f2001, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
$L__BB0_105:
|
|
setp.eq.f32 %p137, %f138, 0f3F800000;
|
|
selp.f32 %f192, 0f3F800000, %f2001, %p137;
|
|
abs.f32 %f193, %f139;
|
|
setp.lt.f32 %p138, %f193, 0f00800000;
|
|
mul.f32 %f1356, %f193, 0f4B800000;
|
|
selp.f32 %f1357, %f1356, %f193, %p138;
|
|
selp.f32 %f1358, 0fC3170000, 0fC2FE0000, %p138;
|
|
mov.b32 %r1446, %f1357;
|
|
and.b32 %r1447, %r1446, 8388607;
|
|
or.b32 %r1448, %r1447, 1065353216;
|
|
mov.b32 %f1359, %r1448;
|
|
shr.u32 %r1449, %r1446, 23;
|
|
cvt.rn.f32.u32 %f1360, %r1449;
|
|
add.f32 %f1361, %f1358, %f1360;
|
|
setp.gt.f32 %p139, %f1359, 0f3FB504F3;
|
|
mul.f32 %f1362, %f1359, 0f3F000000;
|
|
add.f32 %f1363, %f1361, 0f3F800000;
|
|
selp.f32 %f1364, %f1363, %f1361, %p139;
|
|
selp.f32 %f1365, %f1362, %f1359, %p139;
|
|
add.f32 %f1366, %f1365, 0fBF800000;
|
|
add.f32 %f1367, %f1365, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1368, %f1367;
|
|
add.f32 %f1369, %f1366, %f1366;
|
|
mul.f32 %f1370, %f1369, %f1368;
|
|
mul.f32 %f1371, %f1370, %f1370;
|
|
mov.f32 %f1372, 0f3C4CAF63;
|
|
mov.f32 %f1373, 0f3B18F0FE;
|
|
fma.rn.f32 %f1374, %f1373, %f1371, %f1372;
|
|
mov.f32 %f1375, 0f3DAAAABD;
|
|
fma.rn.f32 %f1376, %f1374, %f1371, %f1375;
|
|
mul.rn.f32 %f1377, %f1376, %f1371;
|
|
mul.rn.f32 %f1378, %f1377, %f1370;
|
|
sub.f32 %f1379, %f1366, %f1370;
|
|
add.f32 %f1380, %f1379, %f1379;
|
|
neg.f32 %f1381, %f1370;
|
|
fma.rn.f32 %f1382, %f1381, %f1366, %f1380;
|
|
mul.rn.f32 %f1383, %f1368, %f1382;
|
|
add.f32 %f1384, %f1378, %f1370;
|
|
sub.f32 %f1385, %f1370, %f1384;
|
|
add.f32 %f1386, %f1378, %f1385;
|
|
add.f32 %f1387, %f1383, %f1386;
|
|
add.f32 %f1388, %f1384, %f1387;
|
|
sub.f32 %f1389, %f1384, %f1388;
|
|
add.f32 %f1390, %f1387, %f1389;
|
|
mov.f32 %f1391, 0f3F317200;
|
|
mul.rn.f32 %f1392, %f1364, %f1391;
|
|
mov.f32 %f1393, 0f35BFBE8E;
|
|
mul.rn.f32 %f1394, %f1364, %f1393;
|
|
add.f32 %f1395, %f1392, %f1388;
|
|
sub.f32 %f1396, %f1392, %f1395;
|
|
add.f32 %f1397, %f1388, %f1396;
|
|
add.f32 %f1398, %f1390, %f1397;
|
|
add.f32 %f1399, %f1394, %f1398;
|
|
add.f32 %f1400, %f1395, %f1399;
|
|
sub.f32 %f1401, %f1395, %f1400;
|
|
add.f32 %f1402, %f1399, %f1401;
|
|
mov.f32 %f1403, 0f3EE66666;
|
|
mul.rn.f32 %f1404, %f1403, %f1400;
|
|
neg.f32 %f1405, %f1404;
|
|
fma.rn.f32 %f1406, %f1403, %f1400, %f1405;
|
|
fma.rn.f32 %f1407, %f1403, %f1402, %f1406;
|
|
mov.f32 %f1408, 0f00000000;
|
|
fma.rn.f32 %f1409, %f1408, %f1400, %f1407;
|
|
add.rn.f32 %f1410, %f1404, %f1409;
|
|
neg.f32 %f1411, %f1410;
|
|
add.rn.f32 %f1412, %f1404, %f1411;
|
|
add.rn.f32 %f1413, %f1412, %f1409;
|
|
mov.b32 %r1450, %f1410;
|
|
setp.eq.s32 %p140, %r1450, 1118925336;
|
|
add.s32 %r1451, %r1450, -1;
|
|
mov.b32 %f1414, %r1451;
|
|
add.f32 %f1415, %f1413, 0f37000000;
|
|
selp.f32 %f194, %f1415, %f1413, %p140;
|
|
selp.f32 %f1416, %f1414, %f1410, %p140;
|
|
mov.f32 %f1417, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1418, %f1416, %f1417;
|
|
cvt.rzi.f32.f32 %f1419, %f1418;
|
|
abs.f32 %f1420, %f1419;
|
|
setp.gt.f32 %p141, %f1420, 0f42FC0000;
|
|
mov.b32 %r1452, %f1419;
|
|
and.b32 %r1453, %r1452, -2147483648;
|
|
or.b32 %r1454, %r1453, 1123811328;
|
|
mov.b32 %f1421, %r1454;
|
|
selp.f32 %f1422, %f1421, %f1419, %p141;
|
|
mov.f32 %f1423, 0fBF317218;
|
|
fma.rn.f32 %f1424, %f1422, %f1423, %f1416;
|
|
mov.f32 %f1425, 0f3102E308;
|
|
fma.rn.f32 %f1426, %f1422, %f1425, %f1424;
|
|
mul.f32 %f1427, %f1426, 0f3FB8AA3B;
|
|
add.f32 %f1428, %f1422, 0f4B40007F;
|
|
mov.b32 %r1455, %f1428;
|
|
shl.b32 %r1456, %r1455, 23;
|
|
mov.b32 %f1429, %r1456;
|
|
ex2.approx.ftz.f32 %f1430, %f1427;
|
|
mul.f32 %f195, %f1430, %f1429;
|
|
setp.eq.f32 %p142, %f195, 0f7F800000;
|
|
mov.f32 %f2002, 0f7F800000;
|
|
@%p142 bra $L__BB0_107;
|
|
|
|
fma.rn.f32 %f2002, %f195, %f194, %f195;
|
|
|
|
$L__BB0_107:
|
|
setp.lt.f32 %p143, %f139, 0f00000000;
|
|
and.pred %p5, %p143, %p114;
|
|
setp.eq.f32 %p145, %f139, 0f00000000;
|
|
@%p145 bra $L__BB0_111;
|
|
bra.uni $L__BB0_108;
|
|
|
|
$L__BB0_111:
|
|
add.f32 %f1435, %f139, %f139;
|
|
selp.f32 %f2004, %f1435, 0f00000000, %p114;
|
|
bra.uni $L__BB0_112;
|
|
|
|
$L__BB0_108:
|
|
mov.b32 %r1457, %f2002;
|
|
xor.b32 %r1458, %r1457, -2147483648;
|
|
mov.b32 %f1431, %r1458;
|
|
selp.f32 %f2004, %f1431, %f2002, %p5;
|
|
setp.geu.f32 %p146, %f139, 0f00000000;
|
|
@%p146 bra $L__BB0_112;
|
|
|
|
mov.f32 %f1432, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f1433, %f1432;
|
|
setp.eq.f32 %p147, %f1433, 0f3EE66666;
|
|
@%p147 bra $L__BB0_112;
|
|
|
|
mov.f32 %f2004, 0f7FFFFFFF;
|
|
|
|
$L__BB0_112:
|
|
add.f32 %f1436, %f193, 0f3EE66666;
|
|
mov.b32 %r1459, %f1436;
|
|
setp.lt.s32 %p149, %r1459, 2139095040;
|
|
@%p149 bra $L__BB0_117;
|
|
|
|
setp.gtu.f32 %p150, %f193, 0f7F800000;
|
|
@%p150 bra $L__BB0_116;
|
|
bra.uni $L__BB0_114;
|
|
|
|
$L__BB0_116:
|
|
add.f32 %f2004, %f139, 0f3EE66666;
|
|
bra.uni $L__BB0_117;
|
|
|
|
$L__BB0_114:
|
|
setp.neu.f32 %p151, %f193, 0f7F800000;
|
|
@%p151 bra $L__BB0_117;
|
|
|
|
selp.f32 %f2004, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
$L__BB0_117:
|
|
setp.eq.f32 %p152, %f139, 0f3F800000;
|
|
mov.f32 %f1438, 0f3F800000;
|
|
selp.f32 %f1439, 0f3F800000, %f2004, %p152;
|
|
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 %f1440, %f180, %f1438;
|
|
mov.f32 %f1441, 0f00000000;
|
|
max.f32 %f204, %f1441, %f1440;
|
|
min.f32 %f1442, %f192, %f1438;
|
|
max.f32 %f205, %f1441, %f1442;
|
|
min.f32 %f1443, %f1439, %f1438;
|
|
max.f32 %f206, %f1441, %f1443;
|
|
mov.f32 %f1444, 0f3E555555;
|
|
cvt.rzi.f32.f32 %f1445, %f1444;
|
|
add.f32 %f1446, %f1445, %f1445;
|
|
mov.f32 %f1447, 0f3ED55555;
|
|
sub.f32 %f1448, %f1447, %f1446;
|
|
abs.f32 %f207, %f1448;
|
|
abs.f32 %f208, %f204;
|
|
setp.lt.f32 %p153, %f208, 0f00800000;
|
|
mul.f32 %f1449, %f208, 0f4B800000;
|
|
selp.f32 %f1450, %f1449, %f208, %p153;
|
|
selp.f32 %f1451, 0fC3170000, 0fC2FE0000, %p153;
|
|
mov.b32 %r1462, %f1450;
|
|
and.b32 %r1463, %r1462, 8388607;
|
|
or.b32 %r1464, %r1463, 1065353216;
|
|
mov.b32 %f1452, %r1464;
|
|
shr.u32 %r1465, %r1462, 23;
|
|
cvt.rn.f32.u32 %f1453, %r1465;
|
|
add.f32 %f1454, %f1451, %f1453;
|
|
setp.gt.f32 %p154, %f1452, 0f3FB504F3;
|
|
mul.f32 %f1455, %f1452, 0f3F000000;
|
|
add.f32 %f1456, %f1454, 0f3F800000;
|
|
selp.f32 %f1457, %f1456, %f1454, %p154;
|
|
selp.f32 %f1458, %f1455, %f1452, %p154;
|
|
add.f32 %f1459, %f1458, 0fBF800000;
|
|
add.f32 %f1460, %f1458, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1461, %f1460;
|
|
add.f32 %f1462, %f1459, %f1459;
|
|
mul.f32 %f1463, %f1462, %f1461;
|
|
mul.f32 %f1464, %f1463, %f1463;
|
|
mov.f32 %f1465, 0f3C4CAF63;
|
|
mov.f32 %f1466, 0f3B18F0FE;
|
|
fma.rn.f32 %f1467, %f1466, %f1464, %f1465;
|
|
mov.f32 %f1468, 0f3DAAAABD;
|
|
fma.rn.f32 %f1469, %f1467, %f1464, %f1468;
|
|
mul.rn.f32 %f1470, %f1469, %f1464;
|
|
mul.rn.f32 %f1471, %f1470, %f1463;
|
|
sub.f32 %f1472, %f1459, %f1463;
|
|
add.f32 %f1473, %f1472, %f1472;
|
|
neg.f32 %f1474, %f1463;
|
|
fma.rn.f32 %f1475, %f1474, %f1459, %f1473;
|
|
mul.rn.f32 %f1476, %f1461, %f1475;
|
|
add.f32 %f1477, %f1471, %f1463;
|
|
sub.f32 %f1478, %f1463, %f1477;
|
|
add.f32 %f1479, %f1471, %f1478;
|
|
add.f32 %f1480, %f1476, %f1479;
|
|
add.f32 %f1481, %f1477, %f1480;
|
|
sub.f32 %f1482, %f1477, %f1481;
|
|
add.f32 %f1483, %f1480, %f1482;
|
|
mov.f32 %f1484, 0f3F317200;
|
|
mul.rn.f32 %f1485, %f1457, %f1484;
|
|
mov.f32 %f1486, 0f35BFBE8E;
|
|
mul.rn.f32 %f1487, %f1457, %f1486;
|
|
add.f32 %f1488, %f1485, %f1481;
|
|
sub.f32 %f1489, %f1485, %f1488;
|
|
add.f32 %f1490, %f1481, %f1489;
|
|
add.f32 %f1491, %f1483, %f1490;
|
|
add.f32 %f1492, %f1487, %f1491;
|
|
add.f32 %f1493, %f1488, %f1492;
|
|
sub.f32 %f1494, %f1488, %f1493;
|
|
add.f32 %f1495, %f1492, %f1494;
|
|
mul.rn.f32 %f1496, %f1447, %f1493;
|
|
neg.f32 %f1497, %f1496;
|
|
fma.rn.f32 %f1498, %f1447, %f1493, %f1497;
|
|
fma.rn.f32 %f1499, %f1447, %f1495, %f1498;
|
|
fma.rn.f32 %f1500, %f1441, %f1493, %f1499;
|
|
add.rn.f32 %f1501, %f1496, %f1500;
|
|
neg.f32 %f1502, %f1501;
|
|
add.rn.f32 %f1503, %f1496, %f1502;
|
|
add.rn.f32 %f1504, %f1503, %f1500;
|
|
mov.b32 %r1466, %f1501;
|
|
setp.eq.s32 %p155, %r1466, 1118925336;
|
|
add.s32 %r1467, %r1466, -1;
|
|
mov.b32 %f1505, %r1467;
|
|
add.f32 %f1506, %f1504, 0f37000000;
|
|
selp.f32 %f209, %f1506, %f1504, %p155;
|
|
selp.f32 %f1507, %f1505, %f1501, %p155;
|
|
mov.f32 %f1508, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1509, %f1507, %f1508;
|
|
cvt.rzi.f32.f32 %f1510, %f1509;
|
|
abs.f32 %f1511, %f1510;
|
|
setp.gt.f32 %p156, %f1511, 0f42FC0000;
|
|
mov.b32 %r1468, %f1510;
|
|
and.b32 %r1469, %r1468, -2147483648;
|
|
or.b32 %r1470, %r1469, 1123811328;
|
|
mov.b32 %f1512, %r1470;
|
|
selp.f32 %f1513, %f1512, %f1510, %p156;
|
|
mov.f32 %f1514, 0fBF317218;
|
|
fma.rn.f32 %f1515, %f1513, %f1514, %f1507;
|
|
mov.f32 %f1516, 0f3102E308;
|
|
fma.rn.f32 %f1517, %f1513, %f1516, %f1515;
|
|
mul.f32 %f1518, %f1517, 0f3FB8AA3B;
|
|
add.f32 %f1519, %f1513, 0f4B40007F;
|
|
mov.b32 %r1471, %f1519;
|
|
shl.b32 %r1472, %r1471, 23;
|
|
mov.b32 %f1520, %r1472;
|
|
ex2.approx.ftz.f32 %f1521, %f1518;
|
|
mul.f32 %f210, %f1521, %f1520;
|
|
setp.eq.f32 %p157, %f210, 0f7F800000;
|
|
mov.f32 %f2005, 0f7F800000;
|
|
@%p157 bra $L__BB0_119;
|
|
|
|
fma.rn.f32 %f2005, %f210, %f209, %f210;
|
|
|
|
$L__BB0_119:
|
|
setp.lt.f32 %p158, %f204, 0f00000000;
|
|
setp.eq.f32 %p159, %f207, 0f3F800000;
|
|
and.pred %p6, %p158, %p159;
|
|
setp.eq.f32 %p160, %f204, 0f00000000;
|
|
@%p160 bra $L__BB0_123;
|
|
bra.uni $L__BB0_120;
|
|
|
|
$L__BB0_123:
|
|
add.f32 %f1526, %f204, %f204;
|
|
selp.f32 %f2007, %f1526, 0f00000000, %p159;
|
|
bra.uni $L__BB0_124;
|
|
|
|
$L__BB0_120:
|
|
mov.b32 %r1473, %f2005;
|
|
xor.b32 %r1474, %r1473, -2147483648;
|
|
mov.b32 %f1522, %r1474;
|
|
selp.f32 %f2007, %f1522, %f2005, %p6;
|
|
setp.geu.f32 %p161, %f204, 0f00000000;
|
|
@%p161 bra $L__BB0_124;
|
|
|
|
mov.f32 %f1523, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1524, %f1523;
|
|
setp.eq.f32 %p162, %f1524, 0f3ED55555;
|
|
@%p162 bra $L__BB0_124;
|
|
|
|
mov.f32 %f2007, 0f7FFFFFFF;
|
|
|
|
$L__BB0_124:
|
|
add.f32 %f1527, %f208, 0f3ED55555;
|
|
mov.b32 %r1475, %f1527;
|
|
setp.lt.s32 %p164, %r1475, 2139095040;
|
|
@%p164 bra $L__BB0_129;
|
|
|
|
setp.gtu.f32 %p165, %f208, 0f7F800000;
|
|
@%p165 bra $L__BB0_128;
|
|
bra.uni $L__BB0_126;
|
|
|
|
$L__BB0_128:
|
|
add.f32 %f2007, %f204, 0f3ED55555;
|
|
bra.uni $L__BB0_129;
|
|
|
|
$L__BB0_126:
|
|
setp.neu.f32 %p166, %f208, 0f7F800000;
|
|
@%p166 bra $L__BB0_129;
|
|
|
|
selp.f32 %f2007, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
$L__BB0_129:
|
|
abs.f32 %f219, %f205;
|
|
setp.lt.f32 %p167, %f219, 0f00800000;
|
|
mul.f32 %f1529, %f219, 0f4B800000;
|
|
selp.f32 %f1530, %f1529, %f219, %p167;
|
|
selp.f32 %f1531, 0fC3170000, 0fC2FE0000, %p167;
|
|
mov.b32 %r1476, %f1530;
|
|
and.b32 %r1477, %r1476, 8388607;
|
|
or.b32 %r1478, %r1477, 1065353216;
|
|
mov.b32 %f1532, %r1478;
|
|
shr.u32 %r1479, %r1476, 23;
|
|
cvt.rn.f32.u32 %f1533, %r1479;
|
|
add.f32 %f1534, %f1531, %f1533;
|
|
setp.gt.f32 %p168, %f1532, 0f3FB504F3;
|
|
mul.f32 %f1535, %f1532, 0f3F000000;
|
|
add.f32 %f1536, %f1534, 0f3F800000;
|
|
selp.f32 %f1537, %f1536, %f1534, %p168;
|
|
selp.f32 %f1538, %f1535, %f1532, %p168;
|
|
add.f32 %f1539, %f1538, 0fBF800000;
|
|
add.f32 %f1540, %f1538, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1541, %f1540;
|
|
add.f32 %f1542, %f1539, %f1539;
|
|
mul.f32 %f1543, %f1542, %f1541;
|
|
mul.f32 %f1544, %f1543, %f1543;
|
|
mov.f32 %f1545, 0f3C4CAF63;
|
|
mov.f32 %f1546, 0f3B18F0FE;
|
|
fma.rn.f32 %f1547, %f1546, %f1544, %f1545;
|
|
mov.f32 %f1548, 0f3DAAAABD;
|
|
fma.rn.f32 %f1549, %f1547, %f1544, %f1548;
|
|
mul.rn.f32 %f1550, %f1549, %f1544;
|
|
mul.rn.f32 %f1551, %f1550, %f1543;
|
|
sub.f32 %f1552, %f1539, %f1543;
|
|
add.f32 %f1553, %f1552, %f1552;
|
|
neg.f32 %f1554, %f1543;
|
|
fma.rn.f32 %f1555, %f1554, %f1539, %f1553;
|
|
mul.rn.f32 %f1556, %f1541, %f1555;
|
|
add.f32 %f1557, %f1551, %f1543;
|
|
sub.f32 %f1558, %f1543, %f1557;
|
|
add.f32 %f1559, %f1551, %f1558;
|
|
add.f32 %f1560, %f1556, %f1559;
|
|
add.f32 %f1561, %f1557, %f1560;
|
|
sub.f32 %f1562, %f1557, %f1561;
|
|
add.f32 %f1563, %f1560, %f1562;
|
|
mov.f32 %f1564, 0f3F317200;
|
|
mul.rn.f32 %f1565, %f1537, %f1564;
|
|
mov.f32 %f1566, 0f35BFBE8E;
|
|
mul.rn.f32 %f1567, %f1537, %f1566;
|
|
add.f32 %f1568, %f1565, %f1561;
|
|
sub.f32 %f1569, %f1565, %f1568;
|
|
add.f32 %f1570, %f1561, %f1569;
|
|
add.f32 %f1571, %f1563, %f1570;
|
|
add.f32 %f1572, %f1567, %f1571;
|
|
add.f32 %f1573, %f1568, %f1572;
|
|
sub.f32 %f1574, %f1568, %f1573;
|
|
add.f32 %f1575, %f1572, %f1574;
|
|
mov.f32 %f1576, 0f3ED55555;
|
|
mul.rn.f32 %f1577, %f1576, %f1573;
|
|
neg.f32 %f1578, %f1577;
|
|
fma.rn.f32 %f1579, %f1576, %f1573, %f1578;
|
|
fma.rn.f32 %f1580, %f1576, %f1575, %f1579;
|
|
mov.f32 %f1581, 0f00000000;
|
|
fma.rn.f32 %f1582, %f1581, %f1573, %f1580;
|
|
add.rn.f32 %f1583, %f1577, %f1582;
|
|
neg.f32 %f1584, %f1583;
|
|
add.rn.f32 %f1585, %f1577, %f1584;
|
|
add.rn.f32 %f1586, %f1585, %f1582;
|
|
mov.b32 %r1480, %f1583;
|
|
setp.eq.s32 %p169, %r1480, 1118925336;
|
|
add.s32 %r1481, %r1480, -1;
|
|
mov.b32 %f1587, %r1481;
|
|
add.f32 %f1588, %f1586, 0f37000000;
|
|
selp.f32 %f220, %f1588, %f1586, %p169;
|
|
selp.f32 %f1589, %f1587, %f1583, %p169;
|
|
mov.f32 %f1590, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1591, %f1589, %f1590;
|
|
cvt.rzi.f32.f32 %f1592, %f1591;
|
|
abs.f32 %f1593, %f1592;
|
|
setp.gt.f32 %p170, %f1593, 0f42FC0000;
|
|
mov.b32 %r1482, %f1592;
|
|
and.b32 %r1483, %r1482, -2147483648;
|
|
or.b32 %r1484, %r1483, 1123811328;
|
|
mov.b32 %f1594, %r1484;
|
|
selp.f32 %f1595, %f1594, %f1592, %p170;
|
|
mov.f32 %f1596, 0fBF317218;
|
|
fma.rn.f32 %f1597, %f1595, %f1596, %f1589;
|
|
mov.f32 %f1598, 0f3102E308;
|
|
fma.rn.f32 %f1599, %f1595, %f1598, %f1597;
|
|
mul.f32 %f1600, %f1599, 0f3FB8AA3B;
|
|
add.f32 %f1601, %f1595, 0f4B40007F;
|
|
mov.b32 %r1485, %f1601;
|
|
shl.b32 %r1486, %r1485, 23;
|
|
mov.b32 %f1602, %r1486;
|
|
ex2.approx.ftz.f32 %f1603, %f1600;
|
|
mul.f32 %f221, %f1603, %f1602;
|
|
setp.eq.f32 %p171, %f221, 0f7F800000;
|
|
mov.f32 %f2008, 0f7F800000;
|
|
@%p171 bra $L__BB0_131;
|
|
|
|
fma.rn.f32 %f2008, %f221, %f220, %f221;
|
|
|
|
$L__BB0_131:
|
|
setp.lt.f32 %p172, %f205, 0f00000000;
|
|
and.pred %p7, %p172, %p159;
|
|
setp.eq.f32 %p174, %f205, 0f00000000;
|
|
@%p174 bra $L__BB0_135;
|
|
bra.uni $L__BB0_132;
|
|
|
|
$L__BB0_135:
|
|
add.f32 %f1608, %f205, %f205;
|
|
selp.f32 %f2010, %f1608, 0f00000000, %p159;
|
|
bra.uni $L__BB0_136;
|
|
|
|
$L__BB0_132:
|
|
mov.b32 %r1487, %f2008;
|
|
xor.b32 %r1488, %r1487, -2147483648;
|
|
mov.b32 %f1604, %r1488;
|
|
selp.f32 %f2010, %f1604, %f2008, %p7;
|
|
setp.geu.f32 %p175, %f205, 0f00000000;
|
|
@%p175 bra $L__BB0_136;
|
|
|
|
mov.f32 %f1605, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1606, %f1605;
|
|
setp.eq.f32 %p176, %f1606, 0f3ED55555;
|
|
@%p176 bra $L__BB0_136;
|
|
|
|
mov.f32 %f2010, 0f7FFFFFFF;
|
|
|
|
$L__BB0_136:
|
|
add.f32 %f1609, %f219, 0f3ED55555;
|
|
mov.b32 %r1489, %f1609;
|
|
setp.lt.s32 %p178, %r1489, 2139095040;
|
|
@%p178 bra $L__BB0_141;
|
|
|
|
setp.gtu.f32 %p179, %f219, 0f7F800000;
|
|
@%p179 bra $L__BB0_140;
|
|
bra.uni $L__BB0_138;
|
|
|
|
$L__BB0_140:
|
|
add.f32 %f2010, %f205, 0f3ED55555;
|
|
bra.uni $L__BB0_141;
|
|
|
|
$L__BB0_138:
|
|
setp.neu.f32 %p180, %f219, 0f7F800000;
|
|
@%p180 bra $L__BB0_141;
|
|
|
|
selp.f32 %f2010, 0fFF800000, 0f7F800000, %p7;
|
|
|
|
$L__BB0_141:
|
|
abs.f32 %f230, %f206;
|
|
setp.lt.f32 %p181, %f230, 0f00800000;
|
|
mul.f32 %f1611, %f230, 0f4B800000;
|
|
selp.f32 %f1612, %f1611, %f230, %p181;
|
|
selp.f32 %f1613, 0fC3170000, 0fC2FE0000, %p181;
|
|
mov.b32 %r1490, %f1612;
|
|
and.b32 %r1491, %r1490, 8388607;
|
|
or.b32 %r1492, %r1491, 1065353216;
|
|
mov.b32 %f1614, %r1492;
|
|
shr.u32 %r1493, %r1490, 23;
|
|
cvt.rn.f32.u32 %f1615, %r1493;
|
|
add.f32 %f1616, %f1613, %f1615;
|
|
setp.gt.f32 %p182, %f1614, 0f3FB504F3;
|
|
mul.f32 %f1617, %f1614, 0f3F000000;
|
|
add.f32 %f1618, %f1616, 0f3F800000;
|
|
selp.f32 %f1619, %f1618, %f1616, %p182;
|
|
selp.f32 %f1620, %f1617, %f1614, %p182;
|
|
add.f32 %f1621, %f1620, 0fBF800000;
|
|
add.f32 %f1622, %f1620, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f1623, %f1622;
|
|
add.f32 %f1624, %f1621, %f1621;
|
|
mul.f32 %f1625, %f1624, %f1623;
|
|
mul.f32 %f1626, %f1625, %f1625;
|
|
mov.f32 %f1627, 0f3C4CAF63;
|
|
mov.f32 %f1628, 0f3B18F0FE;
|
|
fma.rn.f32 %f1629, %f1628, %f1626, %f1627;
|
|
mov.f32 %f1630, 0f3DAAAABD;
|
|
fma.rn.f32 %f1631, %f1629, %f1626, %f1630;
|
|
mul.rn.f32 %f1632, %f1631, %f1626;
|
|
mul.rn.f32 %f1633, %f1632, %f1625;
|
|
sub.f32 %f1634, %f1621, %f1625;
|
|
add.f32 %f1635, %f1634, %f1634;
|
|
neg.f32 %f1636, %f1625;
|
|
fma.rn.f32 %f1637, %f1636, %f1621, %f1635;
|
|
mul.rn.f32 %f1638, %f1623, %f1637;
|
|
add.f32 %f1639, %f1633, %f1625;
|
|
sub.f32 %f1640, %f1625, %f1639;
|
|
add.f32 %f1641, %f1633, %f1640;
|
|
add.f32 %f1642, %f1638, %f1641;
|
|
add.f32 %f1643, %f1639, %f1642;
|
|
sub.f32 %f1644, %f1639, %f1643;
|
|
add.f32 %f1645, %f1642, %f1644;
|
|
mov.f32 %f1646, 0f3F317200;
|
|
mul.rn.f32 %f1647, %f1619, %f1646;
|
|
mov.f32 %f1648, 0f35BFBE8E;
|
|
mul.rn.f32 %f1649, %f1619, %f1648;
|
|
add.f32 %f1650, %f1647, %f1643;
|
|
sub.f32 %f1651, %f1647, %f1650;
|
|
add.f32 %f1652, %f1643, %f1651;
|
|
add.f32 %f1653, %f1645, %f1652;
|
|
add.f32 %f1654, %f1649, %f1653;
|
|
add.f32 %f1655, %f1650, %f1654;
|
|
sub.f32 %f1656, %f1650, %f1655;
|
|
add.f32 %f1657, %f1654, %f1656;
|
|
mov.f32 %f1658, 0f3ED55555;
|
|
mul.rn.f32 %f1659, %f1658, %f1655;
|
|
neg.f32 %f1660, %f1659;
|
|
fma.rn.f32 %f1661, %f1658, %f1655, %f1660;
|
|
fma.rn.f32 %f1662, %f1658, %f1657, %f1661;
|
|
mov.f32 %f1663, 0f00000000;
|
|
fma.rn.f32 %f1664, %f1663, %f1655, %f1662;
|
|
add.rn.f32 %f1665, %f1659, %f1664;
|
|
neg.f32 %f1666, %f1665;
|
|
add.rn.f32 %f1667, %f1659, %f1666;
|
|
add.rn.f32 %f1668, %f1667, %f1664;
|
|
mov.b32 %r1494, %f1665;
|
|
setp.eq.s32 %p183, %r1494, 1118925336;
|
|
add.s32 %r1495, %r1494, -1;
|
|
mov.b32 %f1669, %r1495;
|
|
add.f32 %f1670, %f1668, 0f37000000;
|
|
selp.f32 %f231, %f1670, %f1668, %p183;
|
|
selp.f32 %f1671, %f1669, %f1665, %p183;
|
|
mov.f32 %f1672, 0f3FB8AA3B;
|
|
mul.rn.f32 %f1673, %f1671, %f1672;
|
|
cvt.rzi.f32.f32 %f1674, %f1673;
|
|
abs.f32 %f1675, %f1674;
|
|
setp.gt.f32 %p184, %f1675, 0f42FC0000;
|
|
mov.b32 %r1496, %f1674;
|
|
and.b32 %r1497, %r1496, -2147483648;
|
|
or.b32 %r1498, %r1497, 1123811328;
|
|
mov.b32 %f1676, %r1498;
|
|
selp.f32 %f1677, %f1676, %f1674, %p184;
|
|
mov.f32 %f1678, 0fBF317218;
|
|
fma.rn.f32 %f1679, %f1677, %f1678, %f1671;
|
|
mov.f32 %f1680, 0f3102E308;
|
|
fma.rn.f32 %f1681, %f1677, %f1680, %f1679;
|
|
mul.f32 %f1682, %f1681, 0f3FB8AA3B;
|
|
add.f32 %f1683, %f1677, 0f4B40007F;
|
|
mov.b32 %r1499, %f1683;
|
|
shl.b32 %r1500, %r1499, 23;
|
|
mov.b32 %f1684, %r1500;
|
|
ex2.approx.ftz.f32 %f1685, %f1682;
|
|
mul.f32 %f232, %f1685, %f1684;
|
|
setp.eq.f32 %p185, %f232, 0f7F800000;
|
|
mov.f32 %f2011, 0f7F800000;
|
|
@%p185 bra $L__BB0_143;
|
|
|
|
fma.rn.f32 %f2011, %f232, %f231, %f232;
|
|
|
|
$L__BB0_143:
|
|
setp.lt.f32 %p186, %f206, 0f00000000;
|
|
and.pred %p8, %p186, %p159;
|
|
setp.eq.f32 %p188, %f206, 0f00000000;
|
|
@%p188 bra $L__BB0_147;
|
|
bra.uni $L__BB0_144;
|
|
|
|
$L__BB0_147:
|
|
add.f32 %f1690, %f206, %f206;
|
|
selp.f32 %f2013, %f1690, 0f00000000, %p159;
|
|
bra.uni $L__BB0_148;
|
|
|
|
$L__BB0_144:
|
|
mov.b32 %r1501, %f2011;
|
|
xor.b32 %r1502, %r1501, -2147483648;
|
|
mov.b32 %f1686, %r1502;
|
|
selp.f32 %f2013, %f1686, %f2011, %p8;
|
|
setp.geu.f32 %p189, %f206, 0f00000000;
|
|
@%p189 bra $L__BB0_148;
|
|
|
|
mov.f32 %f1687, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f1688, %f1687;
|
|
setp.eq.f32 %p190, %f1688, 0f3ED55555;
|
|
@%p190 bra $L__BB0_148;
|
|
|
|
mov.f32 %f2013, 0f7FFFFFFF;
|
|
|
|
$L__BB0_148:
|
|
add.f32 %f1691, %f230, 0f3ED55555;
|
|
mov.b32 %r1503, %f1691;
|
|
setp.lt.s32 %p192, %r1503, 2139095040;
|
|
@%p192 bra $L__BB0_153;
|
|
|
|
setp.gtu.f32 %p193, %f230, 0f7F800000;
|
|
@%p193 bra $L__BB0_152;
|
|
bra.uni $L__BB0_150;
|
|
|
|
$L__BB0_152:
|
|
add.f32 %f2013, %f206, 0f3ED55555;
|
|
bra.uni $L__BB0_153;
|
|
|
|
$L__BB0_150:
|
|
setp.neu.f32 %p194, %f230, 0f7F800000;
|
|
@%p194 bra $L__BB0_153;
|
|
|
|
selp.f32 %f2013, 0fFF800000, 0f7F800000, %p8;
|
|
|
|
$L__BB0_153:
|
|
fma.rn.f32 %f1692, %f2007, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p195, %f204, 0f3F800000;
|
|
mov.f32 %f1693, 0f3F800000;
|
|
selp.f32 %f1694, 0f3F7FFFFF, %f1692, %p195;
|
|
mul.f32 %f1695, %f204, 0f414EB852;
|
|
setp.lt.f32 %p196, %f204, 0f3B4D2E1C;
|
|
selp.f32 %f1696, %f1695, %f1694, %p196;
|
|
fma.rn.f32 %f1697, %f2010, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p197, %f205, 0f3F800000;
|
|
selp.f32 %f1698, 0f3F7FFFFF, %f1697, %p197;
|
|
mul.f32 %f1699, %f205, 0f414EB852;
|
|
setp.lt.f32 %p198, %f205, 0f3B4D2E1C;
|
|
selp.f32 %f1700, %f1699, %f1698, %p198;
|
|
fma.rn.f32 %f1701, %f2013, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p199, %f206, 0f3F800000;
|
|
selp.f32 %f1702, 0f3F7FFFFF, %f1701, %p199;
|
|
mul.f32 %f1703, %f206, 0f414EB852;
|
|
setp.lt.f32 %p200, %f206, 0f3B4D2E1C;
|
|
selp.f32 %f1704, %f1703, %f1702, %p200;
|
|
min.f32 %f1705, %f1696, %f1693;
|
|
mov.f32 %f1706, 0f00000000;
|
|
max.f32 %f1707, %f1706, %f1705;
|
|
mul.f32 %f1708, %f1707, 0f43800000;
|
|
cvt.rzi.u32.f32 %r1504, %f1708;
|
|
min.u32 %r1505, %r1504, 255;
|
|
min.f32 %f1709, %f1700, %f1693;
|
|
max.f32 %f1710, %f1706, %f1709;
|
|
mul.f32 %f1711, %f1710, 0f43800000;
|
|
cvt.rzi.u32.f32 %r1506, %f1711;
|
|
min.u32 %r1507, %r1506, 255;
|
|
min.f32 %f1712, %f1704, %f1693;
|
|
max.f32 %f1713, %f1706, %f1712;
|
|
mul.f32 %f1714, %f1713, 0f43800000;
|
|
cvt.rzi.u32.f32 %r1508, %f1714;
|
|
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 %p201, %r1510, 0;
|
|
@%p201 bra $L__BB0_158;
|
|
|
|
ld.const.u32 %r1511, [params+108];
|
|
setp.eq.s32 %p202, %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;
|
|
@%p202 bra $L__BB0_157;
|
|
|
|
ld.global.v4.u16 {%rs56, %rs57, %rs58, %rs59}, [%rd19];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1715, %rs56;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1716, %rs57;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1717, %rs58;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1718, %f137, %f1715;
|
|
add.f32 %f1719, %f138, %f1716;
|
|
add.f32 %f1720, %f139, %f1717;
|
|
mov.f32 %f1721, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f1720;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f1719;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f1718;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1721;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd19], {%rs52, %rs53, %rs54, %rs55};
|
|
bra.uni $L__BB0_158;
|
|
|
|
$L__BB0_157:
|
|
mov.f32 %f1725, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f1725;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f139;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f138;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f137;}
|
|
|
|
// 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 %p203, %rs67, 0;
|
|
@%p203 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 %p204, %rs202, 0;
|
|
mov.f32 %f2014, 0f00000000;
|
|
mov.u16 %rs204, 0;
|
|
mov.u16 %rs205, %rs204;
|
|
mov.f32 %f2015, %f2014;
|
|
mov.f32 %f2016, %f2014;
|
|
@%p204 bra $L__BB0_162;
|
|
|
|
$L__BB0_161:
|
|
mov.u16 %rs205, %rs204;
|
|
cvt.rn.f32.u16 %f1729, %rs64;
|
|
div.rn.f32 %f1730, %f1729, 0f437F0000;
|
|
fma.rn.f32 %f1731, %f1730, 0f40000000, 0fBF800000;
|
|
and.b16 %rs72, %rs205, 255;
|
|
cvt.rn.f32.u16 %f1732, %rs72;
|
|
div.rn.f32 %f1733, %f1732, 0f437F0000;
|
|
fma.rn.f32 %f1734, %f1733, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f1735, %rs202;
|
|
div.rn.f32 %f1736, %f1735, 0f437F0000;
|
|
fma.rn.f32 %f1737, %f1736, 0f40000000, 0fBF800000;
|
|
mul.f32 %f1738, %f1734, %f1734;
|
|
fma.rn.f32 %f1739, %f1731, %f1731, %f1738;
|
|
fma.rn.f32 %f1740, %f1737, %f1737, %f1739;
|
|
sqrt.rn.f32 %f1741, %f1740;
|
|
rcp.rn.f32 %f1742, %f1741;
|
|
mul.f32 %f2016, %f1742, %f1737;
|
|
mul.f32 %f2015, %f1742, %f1734;
|
|
mul.f32 %f2014, %f1731, %f1742;
|
|
mov.u16 %rs204, %rs202;
|
|
|
|
$L__BB0_162:
|
|
mul.f32 %f1746, %f1957, %f2016;
|
|
mul.f32 %f1747, %f1958, %f2015;
|
|
sub.f32 %f1748, %f1746, %f1747;
|
|
mul.f32 %f1749, %f1958, %f2014;
|
|
mul.f32 %f1750, %f1956, %f2016;
|
|
sub.f32 %f1751, %f1749, %f1750;
|
|
mul.f32 %f1752, %f1956, %f2015;
|
|
mul.f32 %f1753, %f1957, %f2014;
|
|
sub.f32 %f1754, %f1752, %f1753;
|
|
ld.global.u8 %rs74, [%rd20+3];
|
|
setp.eq.s16 %p205, %rs74, 0;
|
|
selp.f32 %f247, 0fBF800000, 0f3F800000, %p205;
|
|
mul.f32 %f1755, %f1748, %f247;
|
|
mul.f32 %f1756, %f1751, %f247;
|
|
mul.f32 %f1757, %f1754, %f247;
|
|
mul.f32 %f1758, %f1755, 0f00000000;
|
|
mul.f32 %f1759, %f1756, 0f00000000;
|
|
mul.f32 %f1760, %f1757, 0f00000000;
|
|
fma.rn.f32 %f1761, %f2014, 0f3F5105EC, %f1758;
|
|
fma.rn.f32 %f1762, %f2015, 0f3F5105EC, %f1759;
|
|
fma.rn.f32 %f1763, %f2016, 0f3F5105EC, %f1760;
|
|
mul.f32 %f248, %f1956, 0f3F13CD3A;
|
|
add.f32 %f249, %f248, %f1761;
|
|
mul.f32 %f250, %f1957, 0f3F13CD3A;
|
|
add.f32 %f251, %f250, %f1762;
|
|
mul.f32 %f252, %f1958, 0f3F13CD3A;
|
|
add.f32 %f253, %f252, %f1763;
|
|
or.b16 %rs75, %rs64, %rs205;
|
|
or.b16 %rs23, %rs75, %rs204;
|
|
and.b16 %rs76, %rs23, 255;
|
|
setp.eq.s16 %p206, %rs76, 0;
|
|
mov.f32 %f2020, 0f00000000;
|
|
mov.f32 %f2017, %f2020;
|
|
mov.f32 %f2018, %f2020;
|
|
mov.f32 %f2019, %f2020;
|
|
@%p206 bra $L__BB0_164;
|
|
|
|
cvt.rn.f32.u16 %f1764, %rs64;
|
|
div.rn.f32 %f1765, %f1764, 0f437F0000;
|
|
fma.rn.f32 %f1766, %f1765, 0f40000000, 0fBF800000;
|
|
and.b16 %rs78, %rs205, 255;
|
|
cvt.rn.f32.u16 %f1767, %rs78;
|
|
div.rn.f32 %f1768, %f1767, 0f437F0000;
|
|
fma.rn.f32 %f1769, %f1768, 0f40000000, 0fBF800000;
|
|
and.b16 %rs79, %rs204, 255;
|
|
cvt.rn.f32.u16 %f1770, %rs79;
|
|
div.rn.f32 %f1771, %f1770, 0f437F0000;
|
|
fma.rn.f32 %f1772, %f1771, 0f40000000, 0fBF800000;
|
|
mul.f32 %f1773, %f1769, %f1769;
|
|
fma.rn.f32 %f1774, %f1766, %f1766, %f1773;
|
|
fma.rn.f32 %f1775, %f1772, %f1772, %f1774;
|
|
sqrt.rn.f32 %f1776, %f1775;
|
|
rcp.rn.f32 %f1777, %f1776;
|
|
mul.f32 %f2019, %f1777, %f1772;
|
|
mul.f32 %f2018, %f1777, %f1769;
|
|
mul.f32 %f2017, %f1766, %f1777;
|
|
|
|
$L__BB0_164:
|
|
mul.f32 %f1781, %f1957, %f2019;
|
|
mul.f32 %f1782, %f1958, %f2018;
|
|
sub.f32 %f1783, %f1781, %f1782;
|
|
mul.f32 %f1784, %f1958, %f2017;
|
|
mul.f32 %f1785, %f1956, %f2019;
|
|
sub.f32 %f1786, %f1784, %f1785;
|
|
mul.f32 %f1787, %f1956, %f2018;
|
|
mul.f32 %f1788, %f1957, %f2017;
|
|
sub.f32 %f1789, %f1787, %f1788;
|
|
mul.f32 %f1790, %f1783, %f247;
|
|
mul.f32 %f1791, %f1786, %f247;
|
|
mul.f32 %f1792, %f1789, %f247;
|
|
mul.f32 %f1793, %f2017, 0f3ED105EC;
|
|
mul.f32 %f1794, %f2018, 0f3ED105EC;
|
|
mul.f32 %f1795, %f2019, 0f3ED105EC;
|
|
mul.f32 %f1796, %f1790, 0f3F3504F3;
|
|
mul.f32 %f1797, %f1791, 0f3F3504F3;
|
|
mul.f32 %f1798, %f1792, 0f3F3504F3;
|
|
sub.f32 %f1799, %f1796, %f1793;
|
|
sub.f32 %f1800, %f1797, %f1794;
|
|
sub.f32 %f1801, %f1798, %f1795;
|
|
add.f32 %f260, %f248, %f1799;
|
|
add.f32 %f261, %f250, %f1800;
|
|
add.f32 %f262, %f252, %f1801;
|
|
mov.f32 %f2021, %f2020;
|
|
mov.f32 %f2022, %f2020;
|
|
@%p206 bra $L__BB0_166;
|
|
|
|
cvt.rn.f32.u16 %f1802, %rs64;
|
|
div.rn.f32 %f1803, %f1802, 0f437F0000;
|
|
fma.rn.f32 %f1804, %f1803, 0f40000000, 0fBF800000;
|
|
and.b16 %rs82, %rs205, 255;
|
|
cvt.rn.f32.u16 %f1805, %rs82;
|
|
div.rn.f32 %f1806, %f1805, 0f437F0000;
|
|
fma.rn.f32 %f1807, %f1806, 0f40000000, 0fBF800000;
|
|
and.b16 %rs83, %rs204, 255;
|
|
cvt.rn.f32.u16 %f1808, %rs83;
|
|
div.rn.f32 %f1809, %f1808, 0f437F0000;
|
|
fma.rn.f32 %f1810, %f1809, 0f40000000, 0fBF800000;
|
|
mul.f32 %f1811, %f1807, %f1807;
|
|
fma.rn.f32 %f1812, %f1804, %f1804, %f1811;
|
|
fma.rn.f32 %f1813, %f1810, %f1810, %f1812;
|
|
sqrt.rn.f32 %f1814, %f1813;
|
|
rcp.rn.f32 %f1815, %f1814;
|
|
mul.f32 %f2022, %f1815, %f1810;
|
|
mul.f32 %f2021, %f1815, %f1807;
|
|
mul.f32 %f2020, %f1804, %f1815;
|
|
|
|
$L__BB0_166:
|
|
mul.f32 %f1816, %f1957, %f2022;
|
|
mul.f32 %f1817, %f1958, %f2021;
|
|
sub.f32 %f1818, %f1816, %f1817;
|
|
mul.f32 %f1819, %f1958, %f2020;
|
|
mul.f32 %f1820, %f1956, %f2022;
|
|
sub.f32 %f1821, %f1819, %f1820;
|
|
mul.f32 %f1822, %f1956, %f2021;
|
|
mul.f32 %f1823, %f1957, %f2020;
|
|
sub.f32 %f1824, %f1822, %f1823;
|
|
mul.f32 %f1825, %f1818, %f247;
|
|
mul.f32 %f1826, %f1821, %f247;
|
|
mul.f32 %f1827, %f1824, %f247;
|
|
mul.f32 %f1828, %f2020, 0f3ED105EC;
|
|
mul.f32 %f1829, %f2021, 0f3ED105EC;
|
|
mul.f32 %f1830, %f2022, 0f3ED105EC;
|
|
mul.f32 %f1831, %f1825, 0fBF3504F3;
|
|
mul.f32 %f1832, %f1826, 0fBF3504F3;
|
|
mul.f32 %f1833, %f1827, 0fBF3504F3;
|
|
sub.f32 %f1834, %f1831, %f1828;
|
|
sub.f32 %f1835, %f1832, %f1829;
|
|
sub.f32 %f1836, %f1833, %f1830;
|
|
add.f32 %f1837, %f248, %f1834;
|
|
add.f32 %f1838, %f250, %f1835;
|
|
add.f32 %f1839, %f252, %f1836;
|
|
mul.f32 %f1840, %f251, %f298;
|
|
neg.f32 %f1841, %f1840;
|
|
mul.f32 %f1842, %f249, %f297;
|
|
sub.f32 %f1843, %f1841, %f1842;
|
|
mul.f32 %f1844, %f253, %f299;
|
|
sub.f32 %f1845, %f1843, %f1844;
|
|
cvt.sat.f32.f32 %f1846, %f1845;
|
|
mul.f32 %f1847, %f153, %f1846;
|
|
mul.f32 %f1848, %f154, %f1846;
|
|
mul.f32 %f1849, %f155, %f1846;
|
|
mul.f32 %f1850, %f261, %f298;
|
|
neg.f32 %f1851, %f1850;
|
|
mul.f32 %f1852, %f260, %f297;
|
|
sub.f32 %f1853, %f1851, %f1852;
|
|
mul.f32 %f1854, %f262, %f299;
|
|
sub.f32 %f1855, %f1853, %f1854;
|
|
cvt.sat.f32.f32 %f1856, %f1855;
|
|
mul.f32 %f1857, %f153, %f1856;
|
|
mul.f32 %f1858, %f154, %f1856;
|
|
mul.f32 %f1859, %f155, %f1856;
|
|
mul.f32 %f1860, %f1837, %f297;
|
|
mul.f32 %f1861, %f1838, %f298;
|
|
neg.f32 %f1862, %f1861;
|
|
sub.f32 %f1863, %f1862, %f1860;
|
|
mul.f32 %f1864, %f1839, %f299;
|
|
sub.f32 %f1865, %f1863, %f1864;
|
|
cvt.sat.f32.f32 %f1866, %f1865;
|
|
mul.f32 %f1867, %f153, %f1866;
|
|
mul.f32 %f1868, %f154, %f1866;
|
|
mul.f32 %f1869, %f155, %f1866;
|
|
add.f32 %f1870, %f1847, %f1857;
|
|
add.f32 %f1871, %f1848, %f1858;
|
|
add.f32 %f1872, %f1849, %f1859;
|
|
add.f32 %f1873, %f1870, %f1867;
|
|
add.f32 %f1874, %f1871, %f1868;
|
|
add.f32 %f1875, %f1872, %f1869;
|
|
mul.f32 %f1876, %f1873, 0f3F13CD3A;
|
|
mul.f32 %f1877, %f1874, 0f3F13CD3A;
|
|
mul.f32 %f1878, %f1875, 0f3F13CD3A;
|
|
div.rn.f32 %f1879, %f137, %f1876;
|
|
div.rn.f32 %f1880, %f138, %f1877;
|
|
div.rn.f32 %f1881, %f139, %f1878;
|
|
setp.eq.f32 %p208, %f137, 0f00000000;
|
|
selp.f32 %f1882, 0f00000000, %f1879, %p208;
|
|
setp.eq.f32 %p209, %f138, 0f00000000;
|
|
selp.f32 %f1883, 0f00000000, %f1880, %p209;
|
|
setp.eq.f32 %p210, %f139, 0f00000000;
|
|
selp.f32 %f1884, 0f00000000, %f1881, %p210;
|
|
mul.f32 %f269, %f1847, %f1882;
|
|
mul.f32 %f270, %f1848, %f1883;
|
|
mul.f32 %f271, %f1849, %f1884;
|
|
mul.f32 %f272, %f1857, %f1882;
|
|
mul.f32 %f273, %f1858, %f1883;
|
|
mul.f32 %f274, %f1859, %f1884;
|
|
mul.f32 %f275, %f1867, %f1882;
|
|
mul.f32 %f276, %f1868, %f1883;
|
|
mul.f32 %f277, %f1869, %f1884;
|
|
ld.const.u32 %r59, [params+108];
|
|
setp.eq.s32 %p211, %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;
|
|
@%p211 bra $L__BB0_168;
|
|
|
|
ld.global.v4.u16 {%rs91, %rs92, %rs93, %rs94}, [%rd21];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1885, %rs91;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1886, %rs92;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1887, %rs93;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1888, %f269, %f1885;
|
|
add.f32 %f1889, %f270, %f1886;
|
|
add.f32 %f1890, %f271, %f1887;
|
|
mov.f32 %f1891, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs89, %f1890;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f1889;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f1888;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f1891;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd21], {%rs87, %rs88, %rs89, %rs90};
|
|
bra.uni $L__BB0_169;
|
|
|
|
$L__BB0_168:
|
|
mov.f32 %f1895, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f1895;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f271;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs96, %f270;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f269;}
|
|
|
|
// 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;
|
|
@%p211 bra $L__BB0_171;
|
|
|
|
ld.global.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd22];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1896, %rs106;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1897, %rs107;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1898, %rs108;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1899, %f272, %f1896;
|
|
add.f32 %f1900, %f273, %f1897;
|
|
add.f32 %f1901, %f274, %f1898;
|
|
mov.f32 %f1902, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f1901;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs103, %f1900;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs102, %f1899;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f1902;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd22], {%rs102, %rs103, %rs104, %rs105};
|
|
bra.uni $L__BB0_172;
|
|
|
|
$L__BB0_171:
|
|
mov.f32 %f1906, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f1906;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f274;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f273;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f272;}
|
|
|
|
// 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;
|
|
@%p211 bra $L__BB0_174;
|
|
|
|
ld.global.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd23];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1907, %rs121;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1908, %rs122;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1909, %rs123;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1910, %f275, %f1907;
|
|
add.f32 %f1911, %f276, %f1908;
|
|
add.f32 %f1912, %f277, %f1909;
|
|
mov.f32 %f1913, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs119, %f1912;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f1911;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs117, %f1910;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f1913;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd23], {%rs117, %rs118, %rs119, %rs120};
|
|
bra.uni $L__BB0_192;
|
|
|
|
$L__BB0_174:
|
|
mov.f32 %f1917, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f1917;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f277;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f276;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f275;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd23], {%rs125, %rs126, %rs127, %rs128};
|
|
|
|
$L__BB0_192:
|
|
ret;
|
|
|
|
}
|
|
|