1667 lines
46 KiB
Plaintext
1667 lines
46 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<147>;
|
|
.reg .b16 %rs<46>;
|
|
.reg .f32 %f<921>;
|
|
.reg .b32 %r<329>;
|
|
.reg .f64 %fd<5>;
|
|
.reg .b64 %rd<90>;
|
|
|
|
|
|
mov.u64 %SPL, __local_depot0;
|
|
add.u64 %rd1, %SPL, 0;
|
|
// begin inline asm
|
|
call (%r50), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r51), _optix_get_launch_index_y, ();
|
|
// end inline asm
|
|
ld.const.u32 %r3, [params+540];
|
|
ld.const.u64 %rd19, [params+400];
|
|
cvta.to.global.u64 %rd20, %rd19;
|
|
ld.const.u32 %r56, [params+392];
|
|
mad.lo.s32 %r57, %r56, %r51, %r50;
|
|
mul.wide.u32 %rd21, %r57, 4;
|
|
add.s64 %rd2, %rd20, %rd21;
|
|
ld.global.v2.u8 {%rs7, %rs45}, [%rd2];
|
|
or.b16 %rs9, %rs7, %rs45;
|
|
and.b16 %rs10, %rs9, 255;
|
|
setp.eq.s16 %p7, %rs10, 0;
|
|
@%p7 bra $L__BB0_2;
|
|
|
|
ld.global.u8 %rs44, [%rd2+2];
|
|
bra.uni $L__BB0_3;
|
|
|
|
$L__BB0_2:
|
|
ld.global.u8 %rs44, [%rd2+2];
|
|
setp.eq.s16 %p8, %rs44, 0;
|
|
mov.f32 %f882, 0f00000000;
|
|
mov.u16 %rs45, 0;
|
|
mov.f32 %f883, %f882;
|
|
mov.f32 %f884, %f882;
|
|
@%p8 bra $L__BB0_4;
|
|
|
|
$L__BB0_3:
|
|
cvt.rn.f32.u16 %f155, %rs7;
|
|
div.rn.f32 %f156, %f155, 0f437F0000;
|
|
fma.rn.f32 %f157, %f156, 0f40000000, 0fBF800000;
|
|
and.b16 %rs13, %rs45, 255;
|
|
cvt.rn.f32.u16 %f158, %rs13;
|
|
div.rn.f32 %f159, %f158, 0f437F0000;
|
|
fma.rn.f32 %f160, %f159, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f161, %rs44;
|
|
div.rn.f32 %f162, %f161, 0f437F0000;
|
|
fma.rn.f32 %f163, %f162, 0f40000000, 0fBF800000;
|
|
mul.f32 %f164, %f160, %f160;
|
|
fma.rn.f32 %f165, %f157, %f157, %f164;
|
|
fma.rn.f32 %f166, %f163, %f163, %f165;
|
|
sqrt.rn.f32 %f167, %f166;
|
|
rcp.rn.f32 %f168, %f167;
|
|
mul.f32 %f884, %f168, %f163;
|
|
mul.f32 %f883, %f168, %f160;
|
|
mul.f32 %f882, %f157, %f168;
|
|
|
|
$L__BB0_4:
|
|
ld.const.v2.u32 {%r58, %r59}, [params];
|
|
add.s32 %r4, %r58, %r50;
|
|
add.s32 %r5, %r59, %r51;
|
|
setp.eq.f32 %p9, %f882, 0f00000000;
|
|
setp.eq.f32 %p10, %f883, 0f00000000;
|
|
and.pred %p11, %p9, %p10;
|
|
setp.eq.f32 %p12, %f884, 0f00000000;
|
|
and.pred %p13, %p12, %p11;
|
|
@%p13 bra $L__BB0_116;
|
|
bra.uni $L__BB0_5;
|
|
|
|
$L__BB0_116:
|
|
ld.const.u32 %r49, [params+104];
|
|
and.b32 %r305, %r49, 1;
|
|
setp.eq.b32 %p141, %r305, 1;
|
|
mov.pred %p142, 0;
|
|
xor.pred %p143, %p141, %p142;
|
|
not.pred %p144, %p143;
|
|
@%p144 bra $L__BB0_118;
|
|
|
|
ld.const.u64 %rd70, [params+144];
|
|
cvta.to.global.u64 %rd71, %rd70;
|
|
ld.const.u32 %r306, [params+136];
|
|
mad.lo.s32 %r307, %r306, %r5, %r4;
|
|
mul.wide.u32 %rd72, %r307, 4;
|
|
add.s64 %rd73, %rd71, %rd72;
|
|
mov.u16 %rs35, 0;
|
|
st.global.v4.u8 [%rd73], {%rs35, %rs35, %rs35, %rs35};
|
|
|
|
$L__BB0_118:
|
|
and.b32 %r308, %r49, 4;
|
|
setp.eq.s32 %p145, %r308, 0;
|
|
@%p145 bra $L__BB0_120;
|
|
|
|
ld.const.u64 %rd74, [params+224];
|
|
cvta.to.global.u64 %rd75, %rd74;
|
|
ld.const.u32 %r309, [params+216];
|
|
mad.lo.s32 %r310, %r309, %r5, %r4;
|
|
mov.f32 %f854, 0f00000000;
|
|
mul.wide.u32 %rd76, %r310, 8;
|
|
add.s64 %rd77, %rd75, %rd76;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f854;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f854;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f854;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs39, 0;
|
|
st.global.v4.u16 [%rd77], {%rs36, %rs37, %rs38, %rs39};
|
|
|
|
$L__BB0_120:
|
|
and.b32 %r311, %r49, 16;
|
|
setp.eq.s32 %p146, %r311, 0;
|
|
@%p146 bra $L__BB0_122;
|
|
|
|
ld.const.u64 %rd78, [params+240];
|
|
cvta.to.global.u64 %rd79, %rd78;
|
|
ld.const.u32 %r312, [params+232];
|
|
mad.lo.s32 %r313, %r312, %r5, %r4;
|
|
mov.f32 %f857, 0f00000000;
|
|
mul.wide.u32 %rd80, %r313, 8;
|
|
add.s64 %rd81, %rd79, %rd80;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs42, %f857;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f857;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f857;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs43, 0;
|
|
st.global.v4.u16 [%rd81], {%rs40, %rs41, %rs42, %rs43};
|
|
bra.uni $L__BB0_122;
|
|
|
|
$L__BB0_5:
|
|
ld.const.u64 %rd22, [params+496];
|
|
cvta.to.global.u64 %rd23, %rd22;
|
|
ld.const.u32 %r62, [params+488];
|
|
mad.lo.s32 %r63, %r62, %r51, %r50;
|
|
mul.wide.u32 %rd24, %r63, 8;
|
|
add.s64 %rd25, %rd23, %rd24;
|
|
ld.global.v4.u16 {%rs19, %rs20, %rs21, %rs22}, [%rd25];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f169, %rs19;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f170, %rs20;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f171, %rs21;}
|
|
|
|
// end inline asm
|
|
ld.const.u64 %rd26, [params+432];
|
|
cvta.to.global.u64 %rd27, %rd26;
|
|
ld.const.u32 %r64, [params+424];
|
|
mad.lo.s32 %r65, %r64, %r51, %r50;
|
|
mul.wide.u32 %rd28, %r65, 12;
|
|
add.s64 %rd29, %rd27, %rd28;
|
|
ld.global.f32 %f176, [%rd29];
|
|
mul.f32 %f177, %f176, 0f3456BF95;
|
|
ld.global.f32 %f178, [%rd29+4];
|
|
mul.f32 %f179, %f178, 0f3456BF95;
|
|
ld.global.f32 %f180, [%rd29+8];
|
|
mul.f32 %f181, %f180, 0f3456BF95;
|
|
abs.f32 %f182, %f882;
|
|
div.rn.f32 %f183, %f177, %f182;
|
|
abs.f32 %f184, %f883;
|
|
div.rn.f32 %f185, %f179, %f184;
|
|
abs.f32 %f186, %f884;
|
|
div.rn.f32 %f187, %f181, %f186;
|
|
abs.f32 %f188, %f183;
|
|
abs.f32 %f189, %f185;
|
|
abs.f32 %f190, %f187;
|
|
mov.f32 %f191, 0f38D1B717;
|
|
max.f32 %f192, %f188, %f191;
|
|
max.f32 %f193, %f189, %f191;
|
|
max.f32 %f194, %f190, %f191;
|
|
fma.rn.f32 %f13, %f882, %f192, %f176;
|
|
fma.rn.f32 %f14, %f883, %f193, %f178;
|
|
fma.rn.f32 %f15, %f884, %f194, %f180;
|
|
setp.gt.f32 %p14, %f182, %f186;
|
|
neg.f32 %f195, %f883;
|
|
selp.f32 %f196, %f195, 0f00000000, %p14;
|
|
mov.f32 %f900, 0f00000000;
|
|
neg.f32 %f197, %f884;
|
|
selp.f32 %f198, %f882, %f197, %p14;
|
|
selp.f32 %f199, 0f00000000, %f883, %p14;
|
|
mul.f32 %f200, %f198, %f198;
|
|
fma.rn.f32 %f201, %f196, %f196, %f200;
|
|
fma.rn.f32 %f202, %f199, %f199, %f201;
|
|
sqrt.rn.f32 %f203, %f202;
|
|
rcp.rn.f32 %f204, %f203;
|
|
mul.f32 %f16, %f196, %f204;
|
|
mul.f32 %f17, %f198, %f204;
|
|
mul.f32 %f18, %f199, %f204;
|
|
ld.const.u64 %rd30, [params+128];
|
|
cvta.to.global.u64 %rd31, %rd30;
|
|
ld.const.u32 %r66, [params+120];
|
|
mad.lo.s32 %r67, %r66, %r51, %r50;
|
|
mul.wide.u32 %rd32, %r67, 4;
|
|
add.s64 %rd3, %rd31, %rd32;
|
|
setp.lt.s32 %p15, %r3, 1;
|
|
mov.f32 %f890, %f900;
|
|
mov.f32 %f889, %f900;
|
|
mov.f32 %f888, %f900;
|
|
@%p15 bra $L__BB0_34;
|
|
|
|
cvt.rn.f32.s32 %f208, %r3;
|
|
rcp.rn.f32 %f19, %f208;
|
|
ld.global.u32 %r321, [%rd3];
|
|
mul.f32 %f20, %f13, 0f3456BF95;
|
|
mul.f32 %f21, %f14, 0f3456BF95;
|
|
mul.f32 %f22, %f15, 0f3456BF95;
|
|
ld.const.u64 %rd4, [params+96];
|
|
mul.f32 %f209, %f882, %f17;
|
|
mul.f32 %f210, %f883, %f16;
|
|
sub.f32 %f23, %f210, %f209;
|
|
mul.f32 %f211, %f884, %f16;
|
|
mul.f32 %f212, %f882, %f18;
|
|
sub.f32 %f24, %f212, %f211;
|
|
mul.f32 %f213, %f883, %f18;
|
|
mul.f32 %f214, %f884, %f17;
|
|
sub.f32 %f25, %f214, %f213;
|
|
mov.u32 %r68, 0;
|
|
add.s64 %rd5, %rd1, 24;
|
|
mov.u64 %rd33, __cudart_i2opi_f;
|
|
abs.f32 %f280, %f21;
|
|
abs.f32 %f281, %f20;
|
|
max.f32 %f282, %f281, %f280;
|
|
abs.f32 %f283, %f22;
|
|
max.f32 %f284, %f282, %f283;
|
|
mov.u32 %r318, %r68;
|
|
|
|
$L__BB0_7:
|
|
cvt.rn.f32.s32 %f29, %r318;
|
|
mov.u32 %r320, %r68;
|
|
|
|
$L__BB0_8:
|
|
mad.lo.s32 %r70, %r321, 1664525, 1013904223;
|
|
and.b32 %r71, %r70, 16777215;
|
|
cvt.rn.f32.u32 %f215, %r71;
|
|
fma.rn.f32 %f216, %f215, 0f33800000, %f29;
|
|
mul.f32 %f217, %f19, %f216;
|
|
mad.lo.s32 %r321, %r70, 1664525, 1013904223;
|
|
and.b32 %r72, %r321, 16777215;
|
|
cvt.rn.f32.u32 %f218, %r72;
|
|
cvt.rn.f32.s32 %f219, %r320;
|
|
fma.rn.f32 %f220, %f218, 0f33800000, %f219;
|
|
mul.f32 %f221, %f19, %f220;
|
|
sqrt.rn.f32 %f33, %f217;
|
|
mul.f32 %f34, %f221, 0f40C90FDB;
|
|
mul.f32 %f222, %f34, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r328, %f222;
|
|
cvt.rn.f32.s32 %f223, %r328;
|
|
mov.f32 %f224, 0fBFC90FDA;
|
|
fma.rn.f32 %f225, %f223, %f224, %f34;
|
|
mov.f32 %f226, 0fB3A22168;
|
|
fma.rn.f32 %f227, %f223, %f226, %f225;
|
|
mov.f32 %f228, 0fA7C234C5;
|
|
fma.rn.f32 %f894, %f223, %f228, %f227;
|
|
abs.f32 %f36, %f34;
|
|
setp.ltu.f32 %p16, %f36, 0f47CE4780;
|
|
mov.u32 %r325, %r328;
|
|
mov.f32 %f891, %f894;
|
|
@%p16 bra $L__BB0_16;
|
|
|
|
setp.eq.f32 %p17, %f36, 0f7F800000;
|
|
@%p17 bra $L__BB0_15;
|
|
bra.uni $L__BB0_10;
|
|
|
|
$L__BB0_15:
|
|
mov.f32 %f231, 0f00000000;
|
|
mul.rn.f32 %f891, %f34, %f231;
|
|
mov.u32 %r325, 0;
|
|
bra.uni $L__BB0_16;
|
|
|
|
$L__BB0_10:
|
|
mov.b32 %r13, %f34;
|
|
bfe.u32 %r74, %r13, 23, 8;
|
|
add.s32 %r14, %r74, -128;
|
|
shl.b32 %r75, %r13, 8;
|
|
or.b32 %r15, %r75, -2147483648;
|
|
shr.u32 %r16, %r14, 5;
|
|
mov.u64 %rd87, 0;
|
|
mov.u32 %r322, 0;
|
|
mov.u64 %rd85, %rd1;
|
|
mov.u64 %rd86, %rd33;
|
|
|
|
$L__BB0_11:
|
|
.pragma "nounroll";
|
|
ld.global.nc.u32 %r76, [%rd86];
|
|
mad.wide.u32 %rd35, %r76, %r15, %rd87;
|
|
shr.u64 %rd87, %rd35, 32;
|
|
st.local.u32 [%rd85], %rd35;
|
|
add.s64 %rd86, %rd86, 4;
|
|
add.s64 %rd85, %rd85, 4;
|
|
add.s32 %r322, %r322, 1;
|
|
setp.ne.s32 %p18, %r322, 6;
|
|
@%p18 bra $L__BB0_11;
|
|
|
|
st.local.u32 [%rd5], %rd87;
|
|
mov.u32 %r77, 4;
|
|
sub.s32 %r19, %r77, %r16;
|
|
mov.u32 %r78, 6;
|
|
sub.s32 %r79, %r78, %r16;
|
|
mul.wide.s32 %rd36, %r79, 4;
|
|
add.s64 %rd37, %rd1, %rd36;
|
|
ld.local.u32 %r323, [%rd37];
|
|
ld.local.u32 %r324, [%rd37+-4];
|
|
and.b32 %r22, %r14, 31;
|
|
setp.eq.s32 %p19, %r22, 0;
|
|
@%p19 bra $L__BB0_14;
|
|
|
|
mov.u32 %r80, 32;
|
|
sub.s32 %r81, %r80, %r22;
|
|
shr.u32 %r82, %r324, %r81;
|
|
shl.b32 %r83, %r323, %r22;
|
|
add.s32 %r323, %r82, %r83;
|
|
mul.wide.s32 %rd38, %r19, 4;
|
|
add.s64 %rd39, %rd1, %rd38;
|
|
ld.local.u32 %r84, [%rd39];
|
|
shr.u32 %r85, %r84, %r81;
|
|
shl.b32 %r86, %r324, %r22;
|
|
add.s32 %r324, %r85, %r86;
|
|
|
|
$L__BB0_14:
|
|
and.b32 %r87, %r13, -2147483648;
|
|
shr.u32 %r88, %r324, 30;
|
|
shl.b32 %r89, %r323, 2;
|
|
or.b32 %r90, %r88, %r89;
|
|
shr.u32 %r91, %r90, 31;
|
|
shr.u32 %r92, %r323, 30;
|
|
add.s32 %r93, %r91, %r92;
|
|
neg.s32 %r94, %r93;
|
|
setp.eq.s32 %p20, %r87, 0;
|
|
selp.b32 %r325, %r93, %r94, %p20;
|
|
setp.ne.s32 %p21, %r91, 0;
|
|
xor.b32 %r95, %r87, -2147483648;
|
|
selp.b32 %r96, %r95, %r87, %p21;
|
|
selp.b32 %r97, -1, 0, %p21;
|
|
xor.b32 %r98, %r90, %r97;
|
|
shl.b32 %r99, %r324, 2;
|
|
xor.b32 %r100, %r99, %r97;
|
|
cvt.u64.u32 %rd40, %r98;
|
|
cvt.u64.u32 %rd41, %r100;
|
|
bfi.b64 %rd42, %rd40, %rd41, 32, 32;
|
|
cvt.rn.f64.s64 %fd1, %rd42;
|
|
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f229, %fd2;
|
|
setp.eq.s32 %p22, %r96, 0;
|
|
neg.f32 %f230, %f229;
|
|
selp.f32 %f891, %f229, %f230, %p22;
|
|
|
|
$L__BB0_16:
|
|
add.s32 %r29, %r325, 1;
|
|
and.b32 %r30, %r29, 1;
|
|
setp.eq.s32 %p23, %r30, 0;
|
|
selp.f32 %f40, %f891, 0f3F800000, %p23;
|
|
mul.rn.f32 %f41, %f891, %f891;
|
|
mov.f32 %f892, 0fB94D4153;
|
|
@%p23 bra $L__BB0_18;
|
|
|
|
mov.f32 %f233, 0fBAB607ED;
|
|
mov.f32 %f234, 0f37CBAC00;
|
|
fma.rn.f32 %f892, %f234, %f41, %f233;
|
|
|
|
$L__BB0_18:
|
|
selp.f32 %f235, 0f3C0885E4, 0f3D2AAABB, %p23;
|
|
fma.rn.f32 %f236, %f892, %f41, %f235;
|
|
selp.f32 %f237, 0fBE2AAAA8, 0fBEFFFFFF, %p23;
|
|
fma.rn.f32 %f238, %f236, %f41, %f237;
|
|
mov.f32 %f239, 0f00000000;
|
|
fma.rn.f32 %f240, %f41, %f40, %f239;
|
|
fma.rn.f32 %f893, %f238, %f240, %f40;
|
|
and.b32 %r102, %r29, 2;
|
|
setp.eq.s32 %p25, %r102, 0;
|
|
@%p25 bra $L__BB0_20;
|
|
|
|
mov.f32 %f242, 0fBF800000;
|
|
fma.rn.f32 %f893, %f893, %f242, %f239;
|
|
|
|
$L__BB0_20:
|
|
@%p16 bra $L__BB0_28;
|
|
|
|
setp.eq.f32 %p27, %f36, 0f7F800000;
|
|
@%p27 bra $L__BB0_27;
|
|
bra.uni $L__BB0_22;
|
|
|
|
$L__BB0_27:
|
|
mov.f32 %f245, 0f00000000;
|
|
mul.rn.f32 %f894, %f34, %f245;
|
|
mov.u32 %r328, 0;
|
|
bra.uni $L__BB0_28;
|
|
|
|
$L__BB0_22:
|
|
mov.b32 %r31, %f34;
|
|
bfe.u32 %r103, %r31, 23, 8;
|
|
add.s32 %r32, %r103, -128;
|
|
shl.b32 %r104, %r31, 8;
|
|
or.b32 %r33, %r104, -2147483648;
|
|
shr.u32 %r34, %r32, 5;
|
|
mov.u64 %rd88, 0;
|
|
mov.u64 %rd89, %rd88;
|
|
|
|
$L__BB0_23:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd45, %rd88, 2;
|
|
mov.u64 %rd46, __cudart_i2opi_f;
|
|
add.s64 %rd47, %rd46, %rd45;
|
|
ld.global.nc.u32 %r105, [%rd47];
|
|
mad.wide.u32 %rd48, %r105, %r33, %rd89;
|
|
shr.u64 %rd89, %rd48, 32;
|
|
add.s64 %rd49, %rd1, %rd45;
|
|
st.local.u32 [%rd49], %rd48;
|
|
cvt.u32.u64 %r106, %rd88;
|
|
add.s32 %r107, %r106, 1;
|
|
cvt.s64.s32 %rd88, %r107;
|
|
setp.ne.s32 %p28, %r107, 6;
|
|
@%p28 bra $L__BB0_23;
|
|
|
|
st.local.u32 [%rd5], %rd89;
|
|
mov.u32 %r108, 4;
|
|
sub.s32 %r35, %r108, %r34;
|
|
mov.u32 %r109, 6;
|
|
sub.s32 %r110, %r109, %r34;
|
|
mul.wide.s32 %rd50, %r110, 4;
|
|
add.s64 %rd51, %rd1, %rd50;
|
|
ld.local.u32 %r326, [%rd51];
|
|
ld.local.u32 %r327, [%rd51+-4];
|
|
and.b32 %r38, %r32, 31;
|
|
setp.eq.s32 %p29, %r38, 0;
|
|
@%p29 bra $L__BB0_26;
|
|
|
|
mov.u32 %r111, 32;
|
|
sub.s32 %r112, %r111, %r38;
|
|
shr.u32 %r113, %r327, %r112;
|
|
shl.b32 %r114, %r326, %r38;
|
|
add.s32 %r326, %r113, %r114;
|
|
mul.wide.s32 %rd52, %r35, 4;
|
|
add.s64 %rd53, %rd1, %rd52;
|
|
ld.local.u32 %r115, [%rd53];
|
|
shr.u32 %r116, %r115, %r112;
|
|
shl.b32 %r117, %r327, %r38;
|
|
add.s32 %r327, %r116, %r117;
|
|
|
|
$L__BB0_26:
|
|
and.b32 %r118, %r31, -2147483648;
|
|
shr.u32 %r119, %r327, 30;
|
|
shl.b32 %r120, %r326, 2;
|
|
or.b32 %r121, %r119, %r120;
|
|
shr.u32 %r122, %r121, 31;
|
|
shr.u32 %r123, %r326, 30;
|
|
add.s32 %r124, %r122, %r123;
|
|
neg.s32 %r125, %r124;
|
|
setp.eq.s32 %p30, %r118, 0;
|
|
selp.b32 %r328, %r124, %r125, %p30;
|
|
setp.ne.s32 %p31, %r122, 0;
|
|
xor.b32 %r126, %r118, -2147483648;
|
|
selp.b32 %r127, %r126, %r118, %p31;
|
|
selp.b32 %r128, -1, 0, %p31;
|
|
xor.b32 %r129, %r121, %r128;
|
|
shl.b32 %r130, %r327, 2;
|
|
xor.b32 %r131, %r130, %r128;
|
|
cvt.u64.u32 %rd54, %r129;
|
|
cvt.u64.u32 %rd55, %r131;
|
|
bfi.b64 %rd56, %rd54, %rd55, 32, 32;
|
|
cvt.rn.f64.s64 %fd3, %rd56;
|
|
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f243, %fd4;
|
|
setp.eq.s32 %p32, %r127, 0;
|
|
neg.f32 %f244, %f243;
|
|
selp.f32 %f894, %f243, %f244, %p32;
|
|
|
|
$L__BB0_28:
|
|
mul.f32 %f50, %f33, %f893;
|
|
and.b32 %r45, %r328, 1;
|
|
setp.eq.s32 %p33, %r45, 0;
|
|
selp.f32 %f51, %f894, 0f3F800000, %p33;
|
|
mul.rn.f32 %f52, %f894, %f894;
|
|
mov.f32 %f895, 0fB94D4153;
|
|
@%p33 bra $L__BB0_30;
|
|
|
|
mov.f32 %f247, 0fBAB607ED;
|
|
mov.f32 %f248, 0f37CBAC00;
|
|
fma.rn.f32 %f895, %f248, %f52, %f247;
|
|
|
|
$L__BB0_30:
|
|
selp.f32 %f249, 0f3C0885E4, 0f3D2AAABB, %p33;
|
|
fma.rn.f32 %f250, %f895, %f52, %f249;
|
|
selp.f32 %f251, 0fBE2AAAA8, 0fBEFFFFFF, %p33;
|
|
fma.rn.f32 %f252, %f250, %f52, %f251;
|
|
mov.f32 %f253, 0f00000000;
|
|
fma.rn.f32 %f254, %f52, %f51, %f253;
|
|
fma.rn.f32 %f896, %f252, %f254, %f51;
|
|
and.b32 %r133, %r328, 2;
|
|
setp.eq.s32 %p35, %r133, 0;
|
|
@%p35 bra $L__BB0_32;
|
|
|
|
mov.f32 %f256, 0fBF800000;
|
|
fma.rn.f32 %f896, %f896, %f256, %f253;
|
|
|
|
$L__BB0_32:
|
|
mul.f32 %f266, %f50, %f50;
|
|
mov.f32 %f267, 0f3F800000;
|
|
sub.f32 %f268, %f267, %f266;
|
|
mul.f32 %f269, %f33, %f896;
|
|
mul.f32 %f270, %f269, %f269;
|
|
sub.f32 %f271, %f268, %f270;
|
|
max.f32 %f272, %f253, %f271;
|
|
sqrt.rn.f32 %f273, %f272;
|
|
mul.f32 %f274, %f16, %f269;
|
|
mul.f32 %f275, %f17, %f269;
|
|
mul.f32 %f276, %f18, %f269;
|
|
fma.rn.f32 %f277, %f25, %f50, %f274;
|
|
fma.rn.f32 %f278, %f24, %f50, %f275;
|
|
fma.rn.f32 %f279, %f23, %f50, %f276;
|
|
fma.rn.f32 %f260, %f882, %f273, %f277;
|
|
fma.rn.f32 %f261, %f883, %f273, %f278;
|
|
fma.rn.f32 %f262, %f884, %f273, %f279;
|
|
mov.f32 %f285, 0f38D1B717;
|
|
max.f32 %f263, %f284, %f285;
|
|
mov.f32 %f264, 0f6C4ECB8F;
|
|
mov.u32 %r167, 1;
|
|
mov.u32 %r170, 2;
|
|
mov.u32 %r172, 4;
|
|
mov.u32 %r175, 1065353216;
|
|
mov.u32 %r176, 2139095039;
|
|
mov.u32 %r204, 0;
|
|
// begin inline asm
|
|
call(%r134,%r135,%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165),_optix_trace_typed_32,(%r204,%rd4,%f13,%f14,%f15,%f260,%f261,%f262,%f263,%f264,%f253,%r167,%r204,%r204,%r170,%r204,%r172,%r175,%r175,%r175,%r176,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204);
|
|
// end inline asm
|
|
mov.b32 %f286, %r134;
|
|
mov.b32 %f287, %r135;
|
|
mov.b32 %f288, %r136;
|
|
mov.b32 %f289, %r137;
|
|
setp.ltu.f32 %p36, %f289, 0f00000000;
|
|
selp.f32 %f290, %f288, 0f00000000, %p36;
|
|
selp.f32 %f291, %f287, 0f00000000, %p36;
|
|
selp.f32 %f292, %f286, 0f00000000, %p36;
|
|
add.f32 %f890, %f890, %f292;
|
|
add.f32 %f889, %f889, %f291;
|
|
add.f32 %f888, %f888, %f290;
|
|
add.s32 %r320, %r320, 1;
|
|
setp.lt.s32 %p37, %r320, %r3;
|
|
@%p37 bra $L__BB0_8;
|
|
|
|
add.s32 %r318, %r318, 1;
|
|
setp.lt.s32 %p38, %r318, %r3;
|
|
@%p38 bra $L__BB0_7;
|
|
|
|
$L__BB0_34:
|
|
mul.lo.s32 %r205, %r3, %r3;
|
|
cvt.rn.f32.s32 %f296, %r205;
|
|
rcp.rn.f32 %f297, %f296;
|
|
mul.f32 %f298, %f297, %f890;
|
|
mul.f32 %f299, %f297, %f889;
|
|
mul.f32 %f300, %f297, %f888;
|
|
ld.const.u64 %rd58, [params+504];
|
|
cvt.rn.f32.u32 %f301, %r5;
|
|
cvt.rn.f32.u32 %f302, %r4;
|
|
tex.2d.v4.f32.f32 {%f303, %f304, %f305, %f306}, [%rd58, {%f302, %f301}];
|
|
mul.f32 %f64, %f298, %f303;
|
|
mul.f32 %f65, %f299, %f304;
|
|
mul.f32 %f66, %f300, %f305;
|
|
add.f32 %f67, %f169, %f64;
|
|
add.f32 %f68, %f170, %f65;
|
|
add.f32 %f69, %f171, %f66;
|
|
abs.f32 %f307, %f67;
|
|
setp.gtu.f32 %p39, %f307, 0f7F800000;
|
|
mov.f32 %f901, %f900;
|
|
mov.f32 %f902, %f900;
|
|
@%p39 bra $L__BB0_38;
|
|
|
|
abs.f32 %f311, %f68;
|
|
setp.gtu.f32 %p40, %f311, 0f7F800000;
|
|
mov.f32 %f901, %f900;
|
|
mov.f32 %f902, %f900;
|
|
@%p40 bra $L__BB0_38;
|
|
|
|
abs.f32 %f315, %f69;
|
|
setp.gtu.f32 %p41, %f315, 0f7F800000;
|
|
mov.f32 %f901, %f900;
|
|
mov.f32 %f902, %f900;
|
|
@%p41 bra $L__BB0_38;
|
|
|
|
mov.f32 %f900, %f67;
|
|
mov.f32 %f901, %f68;
|
|
mov.f32 %f902, %f69;
|
|
|
|
$L__BB0_38:
|
|
ld.const.u32 %r48, [params+104];
|
|
and.b32 %r206, %r48, 1;
|
|
setp.eq.b32 %p42, %r206, 1;
|
|
mov.pred %p43, 0;
|
|
xor.pred %p44, %p42, %p43;
|
|
not.pred %p45, %p44;
|
|
@%p45 bra $L__BB0_112;
|
|
|
|
mov.f32 %f320, 0f3EE66666;
|
|
abs.f32 %f80, %f900;
|
|
setp.lt.f32 %p46, %f80, 0f00800000;
|
|
mul.f32 %f322, %f80, 0f4B800000;
|
|
selp.f32 %f323, %f322, %f80, %p46;
|
|
selp.f32 %f324, 0fC3170000, 0fC2FE0000, %p46;
|
|
mov.b32 %r207, %f323;
|
|
and.b32 %r208, %r207, 8388607;
|
|
or.b32 %r209, %r208, 1065353216;
|
|
mov.b32 %f325, %r209;
|
|
shr.u32 %r210, %r207, 23;
|
|
cvt.rn.f32.u32 %f326, %r210;
|
|
add.f32 %f327, %f324, %f326;
|
|
setp.gt.f32 %p47, %f325, 0f3FB504F3;
|
|
mul.f32 %f328, %f325, 0f3F000000;
|
|
add.f32 %f329, %f327, 0f3F800000;
|
|
selp.f32 %f330, %f329, %f327, %p47;
|
|
selp.f32 %f331, %f328, %f325, %p47;
|
|
add.f32 %f332, %f331, 0fBF800000;
|
|
add.f32 %f333, %f331, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f334, %f333;
|
|
add.f32 %f335, %f332, %f332;
|
|
mul.f32 %f336, %f335, %f334;
|
|
mul.f32 %f337, %f336, %f336;
|
|
mov.f32 %f338, 0f3C4CAF63;
|
|
mov.f32 %f339, 0f3B18F0FE;
|
|
fma.rn.f32 %f340, %f339, %f337, %f338;
|
|
mov.f32 %f341, 0f3DAAAABD;
|
|
fma.rn.f32 %f342, %f340, %f337, %f341;
|
|
mul.rn.f32 %f343, %f342, %f337;
|
|
mul.rn.f32 %f344, %f343, %f336;
|
|
sub.f32 %f345, %f332, %f336;
|
|
add.f32 %f346, %f345, %f345;
|
|
neg.f32 %f347, %f336;
|
|
fma.rn.f32 %f348, %f347, %f332, %f346;
|
|
mul.rn.f32 %f349, %f334, %f348;
|
|
add.f32 %f350, %f344, %f336;
|
|
sub.f32 %f351, %f336, %f350;
|
|
add.f32 %f352, %f344, %f351;
|
|
add.f32 %f353, %f349, %f352;
|
|
add.f32 %f354, %f350, %f353;
|
|
sub.f32 %f355, %f350, %f354;
|
|
add.f32 %f356, %f353, %f355;
|
|
mov.f32 %f357, 0f3F317200;
|
|
mul.rn.f32 %f358, %f330, %f357;
|
|
mov.f32 %f359, 0f35BFBE8E;
|
|
mul.rn.f32 %f360, %f330, %f359;
|
|
add.f32 %f361, %f358, %f354;
|
|
sub.f32 %f362, %f358, %f361;
|
|
add.f32 %f363, %f354, %f362;
|
|
add.f32 %f364, %f356, %f363;
|
|
add.f32 %f365, %f360, %f364;
|
|
add.f32 %f366, %f361, %f365;
|
|
sub.f32 %f367, %f361, %f366;
|
|
add.f32 %f368, %f365, %f367;
|
|
mul.rn.f32 %f369, %f320, %f366;
|
|
neg.f32 %f370, %f369;
|
|
fma.rn.f32 %f371, %f320, %f366, %f370;
|
|
fma.rn.f32 %f372, %f320, %f368, %f371;
|
|
mov.f32 %f373, 0f00000000;
|
|
fma.rn.f32 %f374, %f373, %f366, %f372;
|
|
add.rn.f32 %f375, %f369, %f374;
|
|
neg.f32 %f376, %f375;
|
|
add.rn.f32 %f377, %f369, %f376;
|
|
add.rn.f32 %f378, %f377, %f374;
|
|
mov.b32 %r211, %f375;
|
|
setp.eq.s32 %p48, %r211, 1118925336;
|
|
add.s32 %r212, %r211, -1;
|
|
mov.b32 %f379, %r212;
|
|
add.f32 %f380, %f378, 0f37000000;
|
|
selp.f32 %f81, %f380, %f378, %p48;
|
|
selp.f32 %f381, %f379, %f375, %p48;
|
|
mov.f32 %f382, 0f3FB8AA3B;
|
|
mul.rn.f32 %f383, %f381, %f382;
|
|
cvt.rzi.f32.f32 %f384, %f383;
|
|
abs.f32 %f385, %f384;
|
|
setp.gt.f32 %p49, %f385, 0f42FC0000;
|
|
mov.b32 %r213, %f384;
|
|
and.b32 %r214, %r213, -2147483648;
|
|
or.b32 %r215, %r214, 1123811328;
|
|
mov.b32 %f386, %r215;
|
|
selp.f32 %f387, %f386, %f384, %p49;
|
|
mov.f32 %f388, 0fBF317218;
|
|
fma.rn.f32 %f389, %f387, %f388, %f381;
|
|
mov.f32 %f390, 0f3102E308;
|
|
fma.rn.f32 %f391, %f387, %f390, %f389;
|
|
mul.f32 %f392, %f391, 0f3FB8AA3B;
|
|
add.f32 %f393, %f387, 0f4B40007F;
|
|
mov.b32 %r216, %f393;
|
|
shl.b32 %r217, %r216, 23;
|
|
mov.b32 %f394, %r217;
|
|
ex2.approx.ftz.f32 %f395, %f392;
|
|
mul.f32 %f82, %f395, %f394;
|
|
setp.eq.f32 %p50, %f82, 0f7F800000;
|
|
mov.f32 %f903, 0f7F800000;
|
|
@%p50 bra $L__BB0_41;
|
|
|
|
fma.rn.f32 %f903, %f82, %f81, %f82;
|
|
|
|
$L__BB0_41:
|
|
mov.f32 %f872, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f871, %f872;
|
|
add.f32 %f870, %f871, %f871;
|
|
mov.f32 %f869, 0f3EE66666;
|
|
sub.f32 %f868, %f869, %f870;
|
|
abs.f32 %f867, %f868;
|
|
setp.lt.f32 %p51, %f900, 0f00000000;
|
|
setp.eq.f32 %p52, %f867, 0f3F800000;
|
|
and.pred %p1, %p51, %p52;
|
|
setp.eq.f32 %p53, %f900, 0f00000000;
|
|
@%p53 bra $L__BB0_45;
|
|
bra.uni $L__BB0_42;
|
|
|
|
$L__BB0_45:
|
|
add.f32 %f400, %f900, %f900;
|
|
selp.f32 %f905, %f400, 0f00000000, %p52;
|
|
bra.uni $L__BB0_46;
|
|
|
|
$L__BB0_42:
|
|
mov.b32 %r218, %f903;
|
|
xor.b32 %r219, %r218, -2147483648;
|
|
mov.b32 %f396, %r219;
|
|
selp.f32 %f905, %f396, %f903, %p1;
|
|
setp.geu.f32 %p54, %f900, 0f00000000;
|
|
@%p54 bra $L__BB0_46;
|
|
|
|
mov.f32 %f397, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f398, %f397;
|
|
setp.eq.f32 %p55, %f398, 0f3EE66666;
|
|
@%p55 bra $L__BB0_46;
|
|
|
|
mov.f32 %f905, 0f7FFFFFFF;
|
|
|
|
$L__BB0_46:
|
|
abs.f32 %f873, %f900;
|
|
add.f32 %f401, %f873, 0f3EE66666;
|
|
mov.b32 %r220, %f401;
|
|
setp.lt.s32 %p57, %r220, 2139095040;
|
|
@%p57 bra $L__BB0_51;
|
|
|
|
abs.f32 %f874, %f900;
|
|
setp.gtu.f32 %p58, %f874, 0f7F800000;
|
|
@%p58 bra $L__BB0_50;
|
|
bra.uni $L__BB0_48;
|
|
|
|
$L__BB0_50:
|
|
add.f32 %f905, %f900, 0f3EE66666;
|
|
bra.uni $L__BB0_51;
|
|
|
|
$L__BB0_48:
|
|
abs.f32 %f875, %f900;
|
|
setp.neu.f32 %p59, %f875, 0f7F800000;
|
|
@%p59 bra $L__BB0_51;
|
|
|
|
selp.f32 %f905, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
$L__BB0_51:
|
|
setp.eq.f32 %p60, %f900, 0f3F800000;
|
|
selp.f32 %f91, 0f3F800000, %f905, %p60;
|
|
abs.f32 %f92, %f901;
|
|
setp.lt.f32 %p61, %f92, 0f00800000;
|
|
mul.f32 %f403, %f92, 0f4B800000;
|
|
selp.f32 %f404, %f403, %f92, %p61;
|
|
selp.f32 %f405, 0fC3170000, 0fC2FE0000, %p61;
|
|
mov.b32 %r221, %f404;
|
|
and.b32 %r222, %r221, 8388607;
|
|
or.b32 %r223, %r222, 1065353216;
|
|
mov.b32 %f406, %r223;
|
|
shr.u32 %r224, %r221, 23;
|
|
cvt.rn.f32.u32 %f407, %r224;
|
|
add.f32 %f408, %f405, %f407;
|
|
setp.gt.f32 %p62, %f406, 0f3FB504F3;
|
|
mul.f32 %f409, %f406, 0f3F000000;
|
|
add.f32 %f410, %f408, 0f3F800000;
|
|
selp.f32 %f411, %f410, %f408, %p62;
|
|
selp.f32 %f412, %f409, %f406, %p62;
|
|
add.f32 %f413, %f412, 0fBF800000;
|
|
add.f32 %f414, %f412, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f415, %f414;
|
|
add.f32 %f416, %f413, %f413;
|
|
mul.f32 %f417, %f416, %f415;
|
|
mul.f32 %f418, %f417, %f417;
|
|
mov.f32 %f419, 0f3C4CAF63;
|
|
mov.f32 %f420, 0f3B18F0FE;
|
|
fma.rn.f32 %f421, %f420, %f418, %f419;
|
|
mov.f32 %f422, 0f3DAAAABD;
|
|
fma.rn.f32 %f423, %f421, %f418, %f422;
|
|
mul.rn.f32 %f424, %f423, %f418;
|
|
mul.rn.f32 %f425, %f424, %f417;
|
|
sub.f32 %f426, %f413, %f417;
|
|
add.f32 %f427, %f426, %f426;
|
|
neg.f32 %f428, %f417;
|
|
fma.rn.f32 %f429, %f428, %f413, %f427;
|
|
mul.rn.f32 %f430, %f415, %f429;
|
|
add.f32 %f431, %f425, %f417;
|
|
sub.f32 %f432, %f417, %f431;
|
|
add.f32 %f433, %f425, %f432;
|
|
add.f32 %f434, %f430, %f433;
|
|
add.f32 %f435, %f431, %f434;
|
|
sub.f32 %f436, %f431, %f435;
|
|
add.f32 %f437, %f434, %f436;
|
|
mov.f32 %f438, 0f3F317200;
|
|
mul.rn.f32 %f439, %f411, %f438;
|
|
mov.f32 %f440, 0f35BFBE8E;
|
|
mul.rn.f32 %f441, %f411, %f440;
|
|
add.f32 %f442, %f439, %f435;
|
|
sub.f32 %f443, %f439, %f442;
|
|
add.f32 %f444, %f435, %f443;
|
|
add.f32 %f445, %f437, %f444;
|
|
add.f32 %f446, %f441, %f445;
|
|
add.f32 %f447, %f442, %f446;
|
|
sub.f32 %f448, %f442, %f447;
|
|
add.f32 %f449, %f446, %f448;
|
|
mov.f32 %f450, 0f3EE66666;
|
|
mul.rn.f32 %f451, %f450, %f447;
|
|
neg.f32 %f452, %f451;
|
|
fma.rn.f32 %f453, %f450, %f447, %f452;
|
|
fma.rn.f32 %f454, %f450, %f449, %f453;
|
|
mov.f32 %f455, 0f00000000;
|
|
fma.rn.f32 %f456, %f455, %f447, %f454;
|
|
add.rn.f32 %f457, %f451, %f456;
|
|
neg.f32 %f458, %f457;
|
|
add.rn.f32 %f459, %f451, %f458;
|
|
add.rn.f32 %f460, %f459, %f456;
|
|
mov.b32 %r225, %f457;
|
|
setp.eq.s32 %p63, %r225, 1118925336;
|
|
add.s32 %r226, %r225, -1;
|
|
mov.b32 %f461, %r226;
|
|
add.f32 %f462, %f460, 0f37000000;
|
|
selp.f32 %f93, %f462, %f460, %p63;
|
|
selp.f32 %f463, %f461, %f457, %p63;
|
|
mov.f32 %f464, 0f3FB8AA3B;
|
|
mul.rn.f32 %f465, %f463, %f464;
|
|
cvt.rzi.f32.f32 %f466, %f465;
|
|
abs.f32 %f467, %f466;
|
|
setp.gt.f32 %p64, %f467, 0f42FC0000;
|
|
mov.b32 %r227, %f466;
|
|
and.b32 %r228, %r227, -2147483648;
|
|
or.b32 %r229, %r228, 1123811328;
|
|
mov.b32 %f468, %r229;
|
|
selp.f32 %f469, %f468, %f466, %p64;
|
|
mov.f32 %f470, 0fBF317218;
|
|
fma.rn.f32 %f471, %f469, %f470, %f463;
|
|
mov.f32 %f472, 0f3102E308;
|
|
fma.rn.f32 %f473, %f469, %f472, %f471;
|
|
mul.f32 %f474, %f473, 0f3FB8AA3B;
|
|
add.f32 %f475, %f469, 0f4B40007F;
|
|
mov.b32 %r230, %f475;
|
|
shl.b32 %r231, %r230, 23;
|
|
mov.b32 %f476, %r231;
|
|
ex2.approx.ftz.f32 %f477, %f474;
|
|
mul.f32 %f94, %f477, %f476;
|
|
setp.eq.f32 %p65, %f94, 0f7F800000;
|
|
mov.f32 %f906, 0f7F800000;
|
|
@%p65 bra $L__BB0_53;
|
|
|
|
fma.rn.f32 %f906, %f94, %f93, %f94;
|
|
|
|
$L__BB0_53:
|
|
setp.lt.f32 %p66, %f901, 0f00000000;
|
|
and.pred %p2, %p66, %p52;
|
|
setp.eq.f32 %p68, %f901, 0f00000000;
|
|
@%p68 bra $L__BB0_57;
|
|
bra.uni $L__BB0_54;
|
|
|
|
$L__BB0_57:
|
|
add.f32 %f482, %f901, %f901;
|
|
selp.f32 %f908, %f482, 0f00000000, %p52;
|
|
bra.uni $L__BB0_58;
|
|
|
|
$L__BB0_54:
|
|
mov.b32 %r232, %f906;
|
|
xor.b32 %r233, %r232, -2147483648;
|
|
mov.b32 %f478, %r233;
|
|
selp.f32 %f908, %f478, %f906, %p2;
|
|
setp.geu.f32 %p69, %f901, 0f00000000;
|
|
@%p69 bra $L__BB0_58;
|
|
|
|
mov.f32 %f479, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f480, %f479;
|
|
setp.eq.f32 %p70, %f480, 0f3EE66666;
|
|
@%p70 bra $L__BB0_58;
|
|
|
|
mov.f32 %f908, 0f7FFFFFFF;
|
|
|
|
$L__BB0_58:
|
|
abs.f32 %f876, %f901;
|
|
add.f32 %f483, %f876, 0f3EE66666;
|
|
mov.b32 %r234, %f483;
|
|
setp.lt.s32 %p72, %r234, 2139095040;
|
|
@%p72 bra $L__BB0_63;
|
|
|
|
abs.f32 %f877, %f901;
|
|
setp.gtu.f32 %p73, %f877, 0f7F800000;
|
|
@%p73 bra $L__BB0_62;
|
|
bra.uni $L__BB0_60;
|
|
|
|
$L__BB0_62:
|
|
add.f32 %f908, %f901, 0f3EE66666;
|
|
bra.uni $L__BB0_63;
|
|
|
|
$L__BB0_60:
|
|
abs.f32 %f878, %f901;
|
|
setp.neu.f32 %p74, %f878, 0f7F800000;
|
|
@%p74 bra $L__BB0_63;
|
|
|
|
selp.f32 %f908, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
$L__BB0_63:
|
|
setp.eq.f32 %p75, %f901, 0f3F800000;
|
|
selp.f32 %f103, 0f3F800000, %f908, %p75;
|
|
abs.f32 %f104, %f902;
|
|
setp.lt.f32 %p76, %f104, 0f00800000;
|
|
mul.f32 %f485, %f104, 0f4B800000;
|
|
selp.f32 %f486, %f485, %f104, %p76;
|
|
selp.f32 %f487, 0fC3170000, 0fC2FE0000, %p76;
|
|
mov.b32 %r235, %f486;
|
|
and.b32 %r236, %r235, 8388607;
|
|
or.b32 %r237, %r236, 1065353216;
|
|
mov.b32 %f488, %r237;
|
|
shr.u32 %r238, %r235, 23;
|
|
cvt.rn.f32.u32 %f489, %r238;
|
|
add.f32 %f490, %f487, %f489;
|
|
setp.gt.f32 %p77, %f488, 0f3FB504F3;
|
|
mul.f32 %f491, %f488, 0f3F000000;
|
|
add.f32 %f492, %f490, 0f3F800000;
|
|
selp.f32 %f493, %f492, %f490, %p77;
|
|
selp.f32 %f494, %f491, %f488, %p77;
|
|
add.f32 %f495, %f494, 0fBF800000;
|
|
add.f32 %f496, %f494, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f497, %f496;
|
|
add.f32 %f498, %f495, %f495;
|
|
mul.f32 %f499, %f498, %f497;
|
|
mul.f32 %f500, %f499, %f499;
|
|
mov.f32 %f501, 0f3C4CAF63;
|
|
mov.f32 %f502, 0f3B18F0FE;
|
|
fma.rn.f32 %f503, %f502, %f500, %f501;
|
|
mov.f32 %f504, 0f3DAAAABD;
|
|
fma.rn.f32 %f505, %f503, %f500, %f504;
|
|
mul.rn.f32 %f506, %f505, %f500;
|
|
mul.rn.f32 %f507, %f506, %f499;
|
|
sub.f32 %f508, %f495, %f499;
|
|
add.f32 %f509, %f508, %f508;
|
|
neg.f32 %f510, %f499;
|
|
fma.rn.f32 %f511, %f510, %f495, %f509;
|
|
mul.rn.f32 %f512, %f497, %f511;
|
|
add.f32 %f513, %f507, %f499;
|
|
sub.f32 %f514, %f499, %f513;
|
|
add.f32 %f515, %f507, %f514;
|
|
add.f32 %f516, %f512, %f515;
|
|
add.f32 %f517, %f513, %f516;
|
|
sub.f32 %f518, %f513, %f517;
|
|
add.f32 %f519, %f516, %f518;
|
|
mov.f32 %f520, 0f3F317200;
|
|
mul.rn.f32 %f521, %f493, %f520;
|
|
mov.f32 %f522, 0f35BFBE8E;
|
|
mul.rn.f32 %f523, %f493, %f522;
|
|
add.f32 %f524, %f521, %f517;
|
|
sub.f32 %f525, %f521, %f524;
|
|
add.f32 %f526, %f517, %f525;
|
|
add.f32 %f527, %f519, %f526;
|
|
add.f32 %f528, %f523, %f527;
|
|
add.f32 %f529, %f524, %f528;
|
|
sub.f32 %f530, %f524, %f529;
|
|
add.f32 %f531, %f528, %f530;
|
|
mov.f32 %f532, 0f3EE66666;
|
|
mul.rn.f32 %f533, %f532, %f529;
|
|
neg.f32 %f534, %f533;
|
|
fma.rn.f32 %f535, %f532, %f529, %f534;
|
|
fma.rn.f32 %f536, %f532, %f531, %f535;
|
|
mov.f32 %f537, 0f00000000;
|
|
fma.rn.f32 %f538, %f537, %f529, %f536;
|
|
add.rn.f32 %f539, %f533, %f538;
|
|
neg.f32 %f540, %f539;
|
|
add.rn.f32 %f541, %f533, %f540;
|
|
add.rn.f32 %f542, %f541, %f538;
|
|
mov.b32 %r239, %f539;
|
|
setp.eq.s32 %p78, %r239, 1118925336;
|
|
add.s32 %r240, %r239, -1;
|
|
mov.b32 %f543, %r240;
|
|
add.f32 %f544, %f542, 0f37000000;
|
|
selp.f32 %f105, %f544, %f542, %p78;
|
|
selp.f32 %f545, %f543, %f539, %p78;
|
|
mov.f32 %f546, 0f3FB8AA3B;
|
|
mul.rn.f32 %f547, %f545, %f546;
|
|
cvt.rzi.f32.f32 %f548, %f547;
|
|
abs.f32 %f549, %f548;
|
|
setp.gt.f32 %p79, %f549, 0f42FC0000;
|
|
mov.b32 %r241, %f548;
|
|
and.b32 %r242, %r241, -2147483648;
|
|
or.b32 %r243, %r242, 1123811328;
|
|
mov.b32 %f550, %r243;
|
|
selp.f32 %f551, %f550, %f548, %p79;
|
|
mov.f32 %f552, 0fBF317218;
|
|
fma.rn.f32 %f553, %f551, %f552, %f545;
|
|
mov.f32 %f554, 0f3102E308;
|
|
fma.rn.f32 %f555, %f551, %f554, %f553;
|
|
mul.f32 %f556, %f555, 0f3FB8AA3B;
|
|
add.f32 %f557, %f551, 0f4B40007F;
|
|
mov.b32 %r244, %f557;
|
|
shl.b32 %r245, %r244, 23;
|
|
mov.b32 %f558, %r245;
|
|
ex2.approx.ftz.f32 %f559, %f556;
|
|
mul.f32 %f106, %f559, %f558;
|
|
setp.eq.f32 %p80, %f106, 0f7F800000;
|
|
mov.f32 %f909, 0f7F800000;
|
|
@%p80 bra $L__BB0_65;
|
|
|
|
fma.rn.f32 %f909, %f106, %f105, %f106;
|
|
|
|
$L__BB0_65:
|
|
setp.lt.f32 %p81, %f902, 0f00000000;
|
|
and.pred %p3, %p81, %p52;
|
|
setp.eq.f32 %p83, %f902, 0f00000000;
|
|
@%p83 bra $L__BB0_69;
|
|
bra.uni $L__BB0_66;
|
|
|
|
$L__BB0_69:
|
|
add.f32 %f564, %f902, %f902;
|
|
selp.f32 %f911, %f564, 0f00000000, %p52;
|
|
bra.uni $L__BB0_70;
|
|
|
|
$L__BB0_66:
|
|
mov.b32 %r246, %f909;
|
|
xor.b32 %r247, %r246, -2147483648;
|
|
mov.b32 %f560, %r247;
|
|
selp.f32 %f911, %f560, %f909, %p3;
|
|
setp.geu.f32 %p84, %f902, 0f00000000;
|
|
@%p84 bra $L__BB0_70;
|
|
|
|
mov.f32 %f561, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f562, %f561;
|
|
setp.eq.f32 %p85, %f562, 0f3EE66666;
|
|
@%p85 bra $L__BB0_70;
|
|
|
|
mov.f32 %f911, 0f7FFFFFFF;
|
|
|
|
$L__BB0_70:
|
|
abs.f32 %f879, %f902;
|
|
add.f32 %f565, %f879, 0f3EE66666;
|
|
mov.b32 %r248, %f565;
|
|
setp.lt.s32 %p87, %r248, 2139095040;
|
|
@%p87 bra $L__BB0_75;
|
|
|
|
abs.f32 %f880, %f902;
|
|
setp.gtu.f32 %p88, %f880, 0f7F800000;
|
|
@%p88 bra $L__BB0_74;
|
|
bra.uni $L__BB0_72;
|
|
|
|
$L__BB0_74:
|
|
add.f32 %f911, %f902, 0f3EE66666;
|
|
bra.uni $L__BB0_75;
|
|
|
|
$L__BB0_72:
|
|
abs.f32 %f881, %f902;
|
|
setp.neu.f32 %p89, %f881, 0f7F800000;
|
|
@%p89 bra $L__BB0_75;
|
|
|
|
selp.f32 %f911, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
$L__BB0_75:
|
|
setp.eq.f32 %p90, %f902, 0f3F800000;
|
|
mov.f32 %f567, 0f3F800000;
|
|
selp.f32 %f568, 0f3F800000, %f911, %p90;
|
|
min.f32 %f569, %f91, %f567;
|
|
mov.f32 %f570, 0f00000000;
|
|
max.f32 %f115, %f570, %f569;
|
|
min.f32 %f571, %f103, %f567;
|
|
max.f32 %f116, %f570, %f571;
|
|
min.f32 %f572, %f568, %f567;
|
|
max.f32 %f117, %f570, %f572;
|
|
mov.f32 %f576, 0f3ED55555;
|
|
abs.f32 %f119, %f115;
|
|
setp.lt.f32 %p91, %f119, 0f00800000;
|
|
mul.f32 %f578, %f119, 0f4B800000;
|
|
selp.f32 %f579, %f578, %f119, %p91;
|
|
selp.f32 %f580, 0fC3170000, 0fC2FE0000, %p91;
|
|
mov.b32 %r251, %f579;
|
|
and.b32 %r252, %r251, 8388607;
|
|
or.b32 %r253, %r252, 1065353216;
|
|
mov.b32 %f581, %r253;
|
|
shr.u32 %r254, %r251, 23;
|
|
cvt.rn.f32.u32 %f582, %r254;
|
|
add.f32 %f583, %f580, %f582;
|
|
setp.gt.f32 %p92, %f581, 0f3FB504F3;
|
|
mul.f32 %f584, %f581, 0f3F000000;
|
|
add.f32 %f585, %f583, 0f3F800000;
|
|
selp.f32 %f586, %f585, %f583, %p92;
|
|
selp.f32 %f587, %f584, %f581, %p92;
|
|
add.f32 %f588, %f587, 0fBF800000;
|
|
add.f32 %f589, %f587, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f590, %f589;
|
|
add.f32 %f591, %f588, %f588;
|
|
mul.f32 %f592, %f591, %f590;
|
|
mul.f32 %f593, %f592, %f592;
|
|
mov.f32 %f594, 0f3C4CAF63;
|
|
mov.f32 %f595, 0f3B18F0FE;
|
|
fma.rn.f32 %f596, %f595, %f593, %f594;
|
|
mov.f32 %f597, 0f3DAAAABD;
|
|
fma.rn.f32 %f598, %f596, %f593, %f597;
|
|
mul.rn.f32 %f599, %f598, %f593;
|
|
mul.rn.f32 %f600, %f599, %f592;
|
|
sub.f32 %f601, %f588, %f592;
|
|
add.f32 %f602, %f601, %f601;
|
|
neg.f32 %f603, %f592;
|
|
fma.rn.f32 %f604, %f603, %f588, %f602;
|
|
mul.rn.f32 %f605, %f590, %f604;
|
|
add.f32 %f606, %f600, %f592;
|
|
sub.f32 %f607, %f592, %f606;
|
|
add.f32 %f608, %f600, %f607;
|
|
add.f32 %f609, %f605, %f608;
|
|
add.f32 %f610, %f606, %f609;
|
|
sub.f32 %f611, %f606, %f610;
|
|
add.f32 %f612, %f609, %f611;
|
|
mov.f32 %f613, 0f3F317200;
|
|
mul.rn.f32 %f614, %f586, %f613;
|
|
mov.f32 %f615, 0f35BFBE8E;
|
|
mul.rn.f32 %f616, %f586, %f615;
|
|
add.f32 %f617, %f614, %f610;
|
|
sub.f32 %f618, %f614, %f617;
|
|
add.f32 %f619, %f610, %f618;
|
|
add.f32 %f620, %f612, %f619;
|
|
add.f32 %f621, %f616, %f620;
|
|
add.f32 %f622, %f617, %f621;
|
|
sub.f32 %f623, %f617, %f622;
|
|
add.f32 %f624, %f621, %f623;
|
|
mul.rn.f32 %f625, %f576, %f622;
|
|
neg.f32 %f626, %f625;
|
|
fma.rn.f32 %f627, %f576, %f622, %f626;
|
|
fma.rn.f32 %f628, %f576, %f624, %f627;
|
|
fma.rn.f32 %f629, %f570, %f622, %f628;
|
|
add.rn.f32 %f630, %f625, %f629;
|
|
neg.f32 %f631, %f630;
|
|
add.rn.f32 %f632, %f625, %f631;
|
|
add.rn.f32 %f633, %f632, %f629;
|
|
mov.b32 %r255, %f630;
|
|
setp.eq.s32 %p93, %r255, 1118925336;
|
|
add.s32 %r256, %r255, -1;
|
|
mov.b32 %f634, %r256;
|
|
add.f32 %f635, %f633, 0f37000000;
|
|
selp.f32 %f120, %f635, %f633, %p93;
|
|
selp.f32 %f636, %f634, %f630, %p93;
|
|
mov.f32 %f637, 0f3FB8AA3B;
|
|
mul.rn.f32 %f638, %f636, %f637;
|
|
cvt.rzi.f32.f32 %f639, %f638;
|
|
abs.f32 %f640, %f639;
|
|
setp.gt.f32 %p94, %f640, 0f42FC0000;
|
|
mov.b32 %r257, %f639;
|
|
and.b32 %r258, %r257, -2147483648;
|
|
or.b32 %r259, %r258, 1123811328;
|
|
mov.b32 %f641, %r259;
|
|
selp.f32 %f642, %f641, %f639, %p94;
|
|
mov.f32 %f643, 0fBF317218;
|
|
fma.rn.f32 %f644, %f642, %f643, %f636;
|
|
mov.f32 %f645, 0f3102E308;
|
|
fma.rn.f32 %f646, %f642, %f645, %f644;
|
|
mul.f32 %f647, %f646, 0f3FB8AA3B;
|
|
add.f32 %f648, %f642, 0f4B40007F;
|
|
mov.b32 %r260, %f648;
|
|
shl.b32 %r261, %r260, 23;
|
|
mov.b32 %f649, %r261;
|
|
ex2.approx.ftz.f32 %f650, %f647;
|
|
mul.f32 %f121, %f650, %f649;
|
|
setp.eq.f32 %p95, %f121, 0f7F800000;
|
|
mov.f32 %f912, 0f7F800000;
|
|
@%p95 bra $L__BB0_77;
|
|
|
|
fma.rn.f32 %f912, %f121, %f120, %f121;
|
|
|
|
$L__BB0_77:
|
|
mov.f32 %f866, 0f3E555555;
|
|
cvt.rzi.f32.f32 %f865, %f866;
|
|
add.f32 %f864, %f865, %f865;
|
|
mov.f32 %f863, 0f3ED55555;
|
|
sub.f32 %f862, %f863, %f864;
|
|
abs.f32 %f861, %f862;
|
|
setp.lt.f32 %p96, %f115, 0f00000000;
|
|
setp.eq.f32 %p97, %f861, 0f3F800000;
|
|
and.pred %p4, %p96, %p97;
|
|
setp.eq.f32 %p98, %f115, 0f00000000;
|
|
@%p98 bra $L__BB0_81;
|
|
bra.uni $L__BB0_78;
|
|
|
|
$L__BB0_81:
|
|
add.f32 %f655, %f115, %f115;
|
|
selp.f32 %f914, %f655, 0f00000000, %p97;
|
|
bra.uni $L__BB0_82;
|
|
|
|
$L__BB0_78:
|
|
mov.b32 %r262, %f912;
|
|
xor.b32 %r263, %r262, -2147483648;
|
|
mov.b32 %f651, %r263;
|
|
selp.f32 %f914, %f651, %f912, %p4;
|
|
setp.geu.f32 %p99, %f115, 0f00000000;
|
|
@%p99 bra $L__BB0_82;
|
|
|
|
mov.f32 %f652, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f653, %f652;
|
|
setp.eq.f32 %p100, %f653, 0f3ED55555;
|
|
@%p100 bra $L__BB0_82;
|
|
|
|
mov.f32 %f914, 0f7FFFFFFF;
|
|
|
|
$L__BB0_82:
|
|
add.f32 %f656, %f119, 0f3ED55555;
|
|
mov.b32 %r264, %f656;
|
|
setp.lt.s32 %p102, %r264, 2139095040;
|
|
@%p102 bra $L__BB0_87;
|
|
|
|
setp.gtu.f32 %p103, %f119, 0f7F800000;
|
|
@%p103 bra $L__BB0_86;
|
|
bra.uni $L__BB0_84;
|
|
|
|
$L__BB0_86:
|
|
add.f32 %f914, %f115, 0f3ED55555;
|
|
bra.uni $L__BB0_87;
|
|
|
|
$L__BB0_84:
|
|
setp.neu.f32 %p104, %f119, 0f7F800000;
|
|
@%p104 bra $L__BB0_87;
|
|
|
|
selp.f32 %f914, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
$L__BB0_87:
|
|
abs.f32 %f130, %f116;
|
|
setp.lt.f32 %p105, %f130, 0f00800000;
|
|
mul.f32 %f658, %f130, 0f4B800000;
|
|
selp.f32 %f659, %f658, %f130, %p105;
|
|
selp.f32 %f660, 0fC3170000, 0fC2FE0000, %p105;
|
|
mov.b32 %r265, %f659;
|
|
and.b32 %r266, %r265, 8388607;
|
|
or.b32 %r267, %r266, 1065353216;
|
|
mov.b32 %f661, %r267;
|
|
shr.u32 %r268, %r265, 23;
|
|
cvt.rn.f32.u32 %f662, %r268;
|
|
add.f32 %f663, %f660, %f662;
|
|
setp.gt.f32 %p106, %f661, 0f3FB504F3;
|
|
mul.f32 %f664, %f661, 0f3F000000;
|
|
add.f32 %f665, %f663, 0f3F800000;
|
|
selp.f32 %f666, %f665, %f663, %p106;
|
|
selp.f32 %f667, %f664, %f661, %p106;
|
|
add.f32 %f668, %f667, 0fBF800000;
|
|
add.f32 %f669, %f667, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f670, %f669;
|
|
add.f32 %f671, %f668, %f668;
|
|
mul.f32 %f672, %f671, %f670;
|
|
mul.f32 %f673, %f672, %f672;
|
|
mov.f32 %f674, 0f3C4CAF63;
|
|
mov.f32 %f675, 0f3B18F0FE;
|
|
fma.rn.f32 %f676, %f675, %f673, %f674;
|
|
mov.f32 %f677, 0f3DAAAABD;
|
|
fma.rn.f32 %f678, %f676, %f673, %f677;
|
|
mul.rn.f32 %f679, %f678, %f673;
|
|
mul.rn.f32 %f680, %f679, %f672;
|
|
sub.f32 %f681, %f668, %f672;
|
|
add.f32 %f682, %f681, %f681;
|
|
neg.f32 %f683, %f672;
|
|
fma.rn.f32 %f684, %f683, %f668, %f682;
|
|
mul.rn.f32 %f685, %f670, %f684;
|
|
add.f32 %f686, %f680, %f672;
|
|
sub.f32 %f687, %f672, %f686;
|
|
add.f32 %f688, %f680, %f687;
|
|
add.f32 %f689, %f685, %f688;
|
|
add.f32 %f690, %f686, %f689;
|
|
sub.f32 %f691, %f686, %f690;
|
|
add.f32 %f692, %f689, %f691;
|
|
mov.f32 %f693, 0f3F317200;
|
|
mul.rn.f32 %f694, %f666, %f693;
|
|
mov.f32 %f695, 0f35BFBE8E;
|
|
mul.rn.f32 %f696, %f666, %f695;
|
|
add.f32 %f697, %f694, %f690;
|
|
sub.f32 %f698, %f694, %f697;
|
|
add.f32 %f699, %f690, %f698;
|
|
add.f32 %f700, %f692, %f699;
|
|
add.f32 %f701, %f696, %f700;
|
|
add.f32 %f702, %f697, %f701;
|
|
sub.f32 %f703, %f697, %f702;
|
|
add.f32 %f704, %f701, %f703;
|
|
mov.f32 %f705, 0f3ED55555;
|
|
mul.rn.f32 %f706, %f705, %f702;
|
|
neg.f32 %f707, %f706;
|
|
fma.rn.f32 %f708, %f705, %f702, %f707;
|
|
fma.rn.f32 %f709, %f705, %f704, %f708;
|
|
mov.f32 %f710, 0f00000000;
|
|
fma.rn.f32 %f711, %f710, %f702, %f709;
|
|
add.rn.f32 %f712, %f706, %f711;
|
|
neg.f32 %f713, %f712;
|
|
add.rn.f32 %f714, %f706, %f713;
|
|
add.rn.f32 %f715, %f714, %f711;
|
|
mov.b32 %r269, %f712;
|
|
setp.eq.s32 %p107, %r269, 1118925336;
|
|
add.s32 %r270, %r269, -1;
|
|
mov.b32 %f716, %r270;
|
|
add.f32 %f717, %f715, 0f37000000;
|
|
selp.f32 %f131, %f717, %f715, %p107;
|
|
selp.f32 %f718, %f716, %f712, %p107;
|
|
mov.f32 %f719, 0f3FB8AA3B;
|
|
mul.rn.f32 %f720, %f718, %f719;
|
|
cvt.rzi.f32.f32 %f721, %f720;
|
|
abs.f32 %f722, %f721;
|
|
setp.gt.f32 %p108, %f722, 0f42FC0000;
|
|
mov.b32 %r271, %f721;
|
|
and.b32 %r272, %r271, -2147483648;
|
|
or.b32 %r273, %r272, 1123811328;
|
|
mov.b32 %f723, %r273;
|
|
selp.f32 %f724, %f723, %f721, %p108;
|
|
mov.f32 %f725, 0fBF317218;
|
|
fma.rn.f32 %f726, %f724, %f725, %f718;
|
|
mov.f32 %f727, 0f3102E308;
|
|
fma.rn.f32 %f728, %f724, %f727, %f726;
|
|
mul.f32 %f729, %f728, 0f3FB8AA3B;
|
|
add.f32 %f730, %f724, 0f4B40007F;
|
|
mov.b32 %r274, %f730;
|
|
shl.b32 %r275, %r274, 23;
|
|
mov.b32 %f731, %r275;
|
|
ex2.approx.ftz.f32 %f732, %f729;
|
|
mul.f32 %f132, %f732, %f731;
|
|
setp.eq.f32 %p109, %f132, 0f7F800000;
|
|
mov.f32 %f915, 0f7F800000;
|
|
@%p109 bra $L__BB0_89;
|
|
|
|
fma.rn.f32 %f915, %f132, %f131, %f132;
|
|
|
|
$L__BB0_89:
|
|
setp.lt.f32 %p110, %f116, 0f00000000;
|
|
and.pred %p5, %p110, %p97;
|
|
setp.eq.f32 %p112, %f116, 0f00000000;
|
|
@%p112 bra $L__BB0_93;
|
|
bra.uni $L__BB0_90;
|
|
|
|
$L__BB0_93:
|
|
add.f32 %f737, %f116, %f116;
|
|
selp.f32 %f917, %f737, 0f00000000, %p97;
|
|
bra.uni $L__BB0_94;
|
|
|
|
$L__BB0_90:
|
|
mov.b32 %r276, %f915;
|
|
xor.b32 %r277, %r276, -2147483648;
|
|
mov.b32 %f733, %r277;
|
|
selp.f32 %f917, %f733, %f915, %p5;
|
|
setp.geu.f32 %p113, %f116, 0f00000000;
|
|
@%p113 bra $L__BB0_94;
|
|
|
|
mov.f32 %f734, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f735, %f734;
|
|
setp.eq.f32 %p114, %f735, 0f3ED55555;
|
|
@%p114 bra $L__BB0_94;
|
|
|
|
mov.f32 %f917, 0f7FFFFFFF;
|
|
|
|
$L__BB0_94:
|
|
add.f32 %f738, %f130, 0f3ED55555;
|
|
mov.b32 %r278, %f738;
|
|
setp.lt.s32 %p116, %r278, 2139095040;
|
|
@%p116 bra $L__BB0_99;
|
|
|
|
setp.gtu.f32 %p117, %f130, 0f7F800000;
|
|
@%p117 bra $L__BB0_98;
|
|
bra.uni $L__BB0_96;
|
|
|
|
$L__BB0_98:
|
|
add.f32 %f917, %f116, 0f3ED55555;
|
|
bra.uni $L__BB0_99;
|
|
|
|
$L__BB0_96:
|
|
setp.neu.f32 %p118, %f130, 0f7F800000;
|
|
@%p118 bra $L__BB0_99;
|
|
|
|
selp.f32 %f917, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
$L__BB0_99:
|
|
abs.f32 %f141, %f117;
|
|
setp.lt.f32 %p119, %f141, 0f00800000;
|
|
mul.f32 %f740, %f141, 0f4B800000;
|
|
selp.f32 %f741, %f740, %f141, %p119;
|
|
selp.f32 %f742, 0fC3170000, 0fC2FE0000, %p119;
|
|
mov.b32 %r279, %f741;
|
|
and.b32 %r280, %r279, 8388607;
|
|
or.b32 %r281, %r280, 1065353216;
|
|
mov.b32 %f743, %r281;
|
|
shr.u32 %r282, %r279, 23;
|
|
cvt.rn.f32.u32 %f744, %r282;
|
|
add.f32 %f745, %f742, %f744;
|
|
setp.gt.f32 %p120, %f743, 0f3FB504F3;
|
|
mul.f32 %f746, %f743, 0f3F000000;
|
|
add.f32 %f747, %f745, 0f3F800000;
|
|
selp.f32 %f748, %f747, %f745, %p120;
|
|
selp.f32 %f749, %f746, %f743, %p120;
|
|
add.f32 %f750, %f749, 0fBF800000;
|
|
add.f32 %f751, %f749, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f752, %f751;
|
|
add.f32 %f753, %f750, %f750;
|
|
mul.f32 %f754, %f753, %f752;
|
|
mul.f32 %f755, %f754, %f754;
|
|
mov.f32 %f756, 0f3C4CAF63;
|
|
mov.f32 %f757, 0f3B18F0FE;
|
|
fma.rn.f32 %f758, %f757, %f755, %f756;
|
|
mov.f32 %f759, 0f3DAAAABD;
|
|
fma.rn.f32 %f760, %f758, %f755, %f759;
|
|
mul.rn.f32 %f761, %f760, %f755;
|
|
mul.rn.f32 %f762, %f761, %f754;
|
|
sub.f32 %f763, %f750, %f754;
|
|
add.f32 %f764, %f763, %f763;
|
|
neg.f32 %f765, %f754;
|
|
fma.rn.f32 %f766, %f765, %f750, %f764;
|
|
mul.rn.f32 %f767, %f752, %f766;
|
|
add.f32 %f768, %f762, %f754;
|
|
sub.f32 %f769, %f754, %f768;
|
|
add.f32 %f770, %f762, %f769;
|
|
add.f32 %f771, %f767, %f770;
|
|
add.f32 %f772, %f768, %f771;
|
|
sub.f32 %f773, %f768, %f772;
|
|
add.f32 %f774, %f771, %f773;
|
|
mov.f32 %f775, 0f3F317200;
|
|
mul.rn.f32 %f776, %f748, %f775;
|
|
mov.f32 %f777, 0f35BFBE8E;
|
|
mul.rn.f32 %f778, %f748, %f777;
|
|
add.f32 %f779, %f776, %f772;
|
|
sub.f32 %f780, %f776, %f779;
|
|
add.f32 %f781, %f772, %f780;
|
|
add.f32 %f782, %f774, %f781;
|
|
add.f32 %f783, %f778, %f782;
|
|
add.f32 %f784, %f779, %f783;
|
|
sub.f32 %f785, %f779, %f784;
|
|
add.f32 %f786, %f783, %f785;
|
|
mov.f32 %f787, 0f3ED55555;
|
|
mul.rn.f32 %f788, %f787, %f784;
|
|
neg.f32 %f789, %f788;
|
|
fma.rn.f32 %f790, %f787, %f784, %f789;
|
|
fma.rn.f32 %f791, %f787, %f786, %f790;
|
|
mov.f32 %f792, 0f00000000;
|
|
fma.rn.f32 %f793, %f792, %f784, %f791;
|
|
add.rn.f32 %f794, %f788, %f793;
|
|
neg.f32 %f795, %f794;
|
|
add.rn.f32 %f796, %f788, %f795;
|
|
add.rn.f32 %f797, %f796, %f793;
|
|
mov.b32 %r283, %f794;
|
|
setp.eq.s32 %p121, %r283, 1118925336;
|
|
add.s32 %r284, %r283, -1;
|
|
mov.b32 %f798, %r284;
|
|
add.f32 %f799, %f797, 0f37000000;
|
|
selp.f32 %f142, %f799, %f797, %p121;
|
|
selp.f32 %f800, %f798, %f794, %p121;
|
|
mov.f32 %f801, 0f3FB8AA3B;
|
|
mul.rn.f32 %f802, %f800, %f801;
|
|
cvt.rzi.f32.f32 %f803, %f802;
|
|
abs.f32 %f804, %f803;
|
|
setp.gt.f32 %p122, %f804, 0f42FC0000;
|
|
mov.b32 %r285, %f803;
|
|
and.b32 %r286, %r285, -2147483648;
|
|
or.b32 %r287, %r286, 1123811328;
|
|
mov.b32 %f805, %r287;
|
|
selp.f32 %f806, %f805, %f803, %p122;
|
|
mov.f32 %f807, 0fBF317218;
|
|
fma.rn.f32 %f808, %f806, %f807, %f800;
|
|
mov.f32 %f809, 0f3102E308;
|
|
fma.rn.f32 %f810, %f806, %f809, %f808;
|
|
mul.f32 %f811, %f810, 0f3FB8AA3B;
|
|
add.f32 %f812, %f806, 0f4B40007F;
|
|
mov.b32 %r288, %f812;
|
|
shl.b32 %r289, %r288, 23;
|
|
mov.b32 %f813, %r289;
|
|
ex2.approx.ftz.f32 %f814, %f811;
|
|
mul.f32 %f143, %f814, %f813;
|
|
setp.eq.f32 %p123, %f143, 0f7F800000;
|
|
mov.f32 %f918, 0f7F800000;
|
|
@%p123 bra $L__BB0_101;
|
|
|
|
fma.rn.f32 %f918, %f143, %f142, %f143;
|
|
|
|
$L__BB0_101:
|
|
setp.lt.f32 %p124, %f117, 0f00000000;
|
|
and.pred %p6, %p124, %p97;
|
|
setp.eq.f32 %p126, %f117, 0f00000000;
|
|
@%p126 bra $L__BB0_105;
|
|
bra.uni $L__BB0_102;
|
|
|
|
$L__BB0_105:
|
|
add.f32 %f819, %f117, %f117;
|
|
selp.f32 %f920, %f819, 0f00000000, %p97;
|
|
bra.uni $L__BB0_106;
|
|
|
|
$L__BB0_102:
|
|
mov.b32 %r290, %f918;
|
|
xor.b32 %r291, %r290, -2147483648;
|
|
mov.b32 %f815, %r291;
|
|
selp.f32 %f920, %f815, %f918, %p6;
|
|
setp.geu.f32 %p127, %f117, 0f00000000;
|
|
@%p127 bra $L__BB0_106;
|
|
|
|
mov.f32 %f816, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f817, %f816;
|
|
setp.eq.f32 %p128, %f817, 0f3ED55555;
|
|
@%p128 bra $L__BB0_106;
|
|
|
|
mov.f32 %f920, 0f7FFFFFFF;
|
|
|
|
$L__BB0_106:
|
|
add.f32 %f820, %f141, 0f3ED55555;
|
|
mov.b32 %r292, %f820;
|
|
setp.lt.s32 %p130, %r292, 2139095040;
|
|
@%p130 bra $L__BB0_111;
|
|
|
|
setp.gtu.f32 %p131, %f141, 0f7F800000;
|
|
@%p131 bra $L__BB0_110;
|
|
bra.uni $L__BB0_108;
|
|
|
|
$L__BB0_110:
|
|
add.f32 %f920, %f117, 0f3ED55555;
|
|
bra.uni $L__BB0_111;
|
|
|
|
$L__BB0_108:
|
|
setp.neu.f32 %p132, %f141, 0f7F800000;
|
|
@%p132 bra $L__BB0_111;
|
|
|
|
selp.f32 %f920, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
$L__BB0_111:
|
|
ld.const.u64 %rd84, [params+144];
|
|
cvta.to.global.u64 %rd83, %rd84;
|
|
ld.const.u32 %r315, [params+136];
|
|
mad.lo.s32 %r314, %r315, %r5, %r4;
|
|
cvt.u64.u32 %rd82, %r314;
|
|
fma.rn.f32 %f821, %f914, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p133, %f115, 0f3F800000;
|
|
mov.f32 %f822, 0f3F800000;
|
|
selp.f32 %f823, 0f3F7FFFFF, %f821, %p133;
|
|
mul.f32 %f824, %f115, 0f414EB852;
|
|
setp.lt.f32 %p134, %f115, 0f3B4D2E1C;
|
|
selp.f32 %f825, %f824, %f823, %p134;
|
|
fma.rn.f32 %f826, %f917, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p135, %f116, 0f3F800000;
|
|
selp.f32 %f827, 0f3F7FFFFF, %f826, %p135;
|
|
mul.f32 %f828, %f116, 0f414EB852;
|
|
setp.lt.f32 %p136, %f116, 0f3B4D2E1C;
|
|
selp.f32 %f829, %f828, %f827, %p136;
|
|
fma.rn.f32 %f830, %f920, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p137, %f117, 0f3F800000;
|
|
selp.f32 %f831, 0f3F7FFFFF, %f830, %p137;
|
|
mul.f32 %f832, %f117, 0f414EB852;
|
|
setp.lt.f32 %p138, %f117, 0f3B4D2E1C;
|
|
selp.f32 %f833, %f832, %f831, %p138;
|
|
min.f32 %f834, %f825, %f822;
|
|
mov.f32 %f835, 0f00000000;
|
|
max.f32 %f836, %f835, %f834;
|
|
mul.f32 %f837, %f836, 0f43800000;
|
|
cvt.rzi.u32.f32 %r293, %f837;
|
|
min.u32 %r294, %r293, 255;
|
|
min.f32 %f838, %f829, %f822;
|
|
max.f32 %f839, %f835, %f838;
|
|
mul.f32 %f840, %f839, 0f43800000;
|
|
cvt.rzi.u32.f32 %r295, %f840;
|
|
min.u32 %r296, %r295, 255;
|
|
min.f32 %f841, %f833, %f822;
|
|
max.f32 %f842, %f835, %f841;
|
|
mul.f32 %f843, %f842, 0f43800000;
|
|
cvt.rzi.u32.f32 %r297, %f843;
|
|
min.u32 %r298, %r297, 255;
|
|
shl.b64 %rd60, %rd82, 2;
|
|
add.s64 %rd61, %rd83, %rd60;
|
|
cvt.u16.u32 %rs23, %r298;
|
|
cvt.u16.u32 %rs24, %r296;
|
|
cvt.u16.u32 %rs25, %r294;
|
|
mov.u16 %rs26, 255;
|
|
st.global.v4.u8 [%rd61], {%rs25, %rs24, %rs23, %rs26};
|
|
|
|
$L__BB0_112:
|
|
ld.const.u32 %r316, [params+104];
|
|
and.b32 %r299, %r316, 4;
|
|
setp.eq.s32 %p139, %r299, 0;
|
|
@%p139 bra $L__BB0_114;
|
|
|
|
ld.const.u64 %rd62, [params+224];
|
|
cvta.to.global.u64 %rd63, %rd62;
|
|
ld.const.u32 %r300, [params+216];
|
|
mad.lo.s32 %r301, %r300, %r5, %r4;
|
|
mov.f32 %f847, 0f3F800000;
|
|
mul.wide.u32 %rd64, %r301, 8;
|
|
add.s64 %rd65, %rd63, %rd64;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f847;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f902;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f901;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f900;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd65], {%rs27, %rs28, %rs29, %rs30};
|
|
|
|
$L__BB0_114:
|
|
ld.const.u32 %r317, [params+104];
|
|
and.b32 %r302, %r317, 16;
|
|
setp.eq.s32 %p140, %r302, 0;
|
|
@%p140 bra $L__BB0_122;
|
|
|
|
ld.const.u64 %rd66, [params+240];
|
|
cvta.to.global.u64 %rd67, %rd66;
|
|
ld.const.u32 %r303, [params+232];
|
|
mad.lo.s32 %r304, %r303, %r5, %r4;
|
|
mov.f32 %f851, 0f3F800000;
|
|
mul.wide.u32 %rd68, %r304, 8;
|
|
add.s64 %rd69, %rd67, %rd68;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f851;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f66;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f65;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f64;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd69], {%rs31, %rs32, %rs33, %rs34};
|
|
|
|
$L__BB0_122:
|
|
ret;
|
|
|
|
}
|
|
|