1084 lines
29 KiB
Plaintext
1084 lines
29 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<85>;
|
|
.reg .b16 %rs<56>;
|
|
.reg .f32 %f<476>;
|
|
.reg .b32 %r<279>;
|
|
.reg .f64 %fd<5>;
|
|
.reg .b64 %rd<82>;
|
|
|
|
|
|
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.f32 %f1, [params+700];
|
|
ld.const.u64 %rd21, [params+400];
|
|
cvta.to.global.u64 %rd22, %rd21;
|
|
ld.const.u32 %r56, [params+392];
|
|
mad.lo.s32 %r57, %r56, %r51, %r50;
|
|
mul.wide.u32 %rd23, %r57, 4;
|
|
add.s64 %rd2, %rd22, %rd23;
|
|
ld.global.v2.u8 {%rs7, %rs55}, [%rd2];
|
|
or.b16 %rs9, %rs7, %rs55;
|
|
and.b16 %rs10, %rs9, 255;
|
|
setp.eq.s16 %p3, %rs10, 0;
|
|
@%p3 bra $L__BB0_2;
|
|
|
|
ld.global.u8 %rs54, [%rd2+2];
|
|
bra.uni $L__BB0_3;
|
|
|
|
$L__BB0_2:
|
|
ld.global.u8 %rs54, [%rd2+2];
|
|
setp.eq.s16 %p4, %rs54, 0;
|
|
mov.f32 %f448, 0f00000000;
|
|
mov.u16 %rs55, 0;
|
|
mov.f32 %f449, %f448;
|
|
mov.f32 %f450, %f448;
|
|
@%p4 bra $L__BB0_4;
|
|
|
|
$L__BB0_3:
|
|
cvt.rn.f32.u16 %f102, %rs7;
|
|
div.rn.f32 %f103, %f102, 0f437F0000;
|
|
fma.rn.f32 %f104, %f103, 0f40000000, 0fBF800000;
|
|
and.b16 %rs13, %rs55, 255;
|
|
cvt.rn.f32.u16 %f105, %rs13;
|
|
div.rn.f32 %f106, %f105, 0f437F0000;
|
|
fma.rn.f32 %f107, %f106, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f108, %rs54;
|
|
div.rn.f32 %f109, %f108, 0f437F0000;
|
|
fma.rn.f32 %f110, %f109, 0f40000000, 0fBF800000;
|
|
mul.f32 %f111, %f107, %f107;
|
|
fma.rn.f32 %f112, %f104, %f104, %f111;
|
|
fma.rn.f32 %f113, %f110, %f110, %f112;
|
|
sqrt.rn.f32 %f114, %f113;
|
|
rcp.rn.f32 %f115, %f114;
|
|
mul.f32 %f450, %f115, %f110;
|
|
mul.f32 %f449, %f115, %f107;
|
|
mul.f32 %f448, %f104, %f115;
|
|
|
|
$L__BB0_4:
|
|
ld.const.v2.u32 {%r58, %r59}, [params];
|
|
add.s32 %r3, %r58, %r50;
|
|
add.s32 %r4, %r59, %r51;
|
|
setp.eq.f32 %p5, %f448, 0f00000000;
|
|
setp.eq.f32 %p6, %f449, 0f00000000;
|
|
and.pred %p7, %p5, %p6;
|
|
setp.eq.f32 %p8, %f450, 0f00000000;
|
|
and.pred %p9, %p8, %p7;
|
|
@%p9 bra $L__BB0_73;
|
|
bra.uni $L__BB0_5;
|
|
|
|
$L__BB0_73:
|
|
ld.const.u32 %r49, [params+104];
|
|
and.b32 %r256, %r49, 4;
|
|
setp.eq.s32 %p81, %r256, 0;
|
|
@%p81 bra $L__BB0_77;
|
|
|
|
ld.const.u32 %r257, [params+108];
|
|
setp.eq.s32 %p82, %r257, 0;
|
|
ld.const.u64 %rd66, [params+224];
|
|
cvta.to.global.u64 %rd67, %rd66;
|
|
ld.const.u32 %r258, [params+216];
|
|
mad.lo.s32 %r259, %r258, %r4, %r3;
|
|
mul.wide.u32 %rd68, %r259, 8;
|
|
add.s64 %rd19, %rd67, %rd68;
|
|
@%p82 bra $L__BB0_76;
|
|
|
|
ld.global.v4.u16 {%rs42, %rs43, %rs44, %rs45}, [%rd19];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f438, %rs42;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f439, %rs43;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f440, %rs44;}
|
|
|
|
// end inline asm
|
|
add.f32 %f441, %f438, 0f00000000;
|
|
add.f32 %f442, %f439, 0f00000000;
|
|
add.f32 %f443, %f440, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f443;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f442;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f441;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs46, 0;
|
|
st.global.v4.u16 [%rd19], {%rs39, %rs40, %rs41, %rs46};
|
|
bra.uni $L__BB0_77;
|
|
|
|
$L__BB0_5:
|
|
ld.const.u64 %rd24, [params+432];
|
|
cvta.to.global.u64 %rd25, %rd24;
|
|
ld.const.u32 %r62, [params+424];
|
|
mad.lo.s32 %r63, %r62, %r51, %r50;
|
|
mul.wide.u32 %rd26, %r63, 12;
|
|
add.s64 %rd27, %rd25, %rd26;
|
|
ld.global.f32 %f120, [%rd27];
|
|
mul.f32 %f121, %f120, 0f3456BF95;
|
|
ld.global.f32 %f122, [%rd27+4];
|
|
mul.f32 %f123, %f122, 0f3456BF95;
|
|
ld.global.f32 %f124, [%rd27+8];
|
|
mul.f32 %f125, %f124, 0f3456BF95;
|
|
abs.f32 %f126, %f448;
|
|
div.rn.f32 %f127, %f121, %f126;
|
|
abs.f32 %f128, %f449;
|
|
div.rn.f32 %f129, %f123, %f128;
|
|
abs.f32 %f130, %f450;
|
|
div.rn.f32 %f131, %f125, %f130;
|
|
abs.f32 %f132, %f127;
|
|
abs.f32 %f133, %f129;
|
|
abs.f32 %f134, %f131;
|
|
mov.f32 %f135, 0f38D1B717;
|
|
max.f32 %f136, %f132, %f135;
|
|
max.f32 %f137, %f133, %f135;
|
|
max.f32 %f138, %f134, %f135;
|
|
fma.rn.f32 %f11, %f448, %f136, %f120;
|
|
fma.rn.f32 %f12, %f449, %f137, %f122;
|
|
fma.rn.f32 %f13, %f450, %f138, %f124;
|
|
setp.gt.f32 %p10, %f126, %f130;
|
|
neg.f32 %f139, %f449;
|
|
selp.f32 %f140, %f139, 0f00000000, %p10;
|
|
mov.f32 %f458, 0f00000000;
|
|
neg.f32 %f141, %f450;
|
|
selp.f32 %f142, %f448, %f141, %p10;
|
|
selp.f32 %f143, 0f00000000, %f449, %p10;
|
|
mul.f32 %f144, %f142, %f142;
|
|
fma.rn.f32 %f145, %f140, %f140, %f144;
|
|
fma.rn.f32 %f146, %f143, %f143, %f145;
|
|
sqrt.rn.f32 %f147, %f146;
|
|
rcp.rn.f32 %f148, %f147;
|
|
mul.f32 %f14, %f140, %f148;
|
|
mul.f32 %f15, %f142, %f148;
|
|
mul.f32 %f16, %f143, %f148;
|
|
ld.const.u64 %rd28, [params+128];
|
|
cvta.to.global.u64 %rd29, %rd28;
|
|
ld.const.u32 %r64, [params+120];
|
|
mad.lo.s32 %r65, %r64, %r51, %r50;
|
|
mul.wide.u32 %rd30, %r65, 4;
|
|
add.s64 %rd3, %rd29, %rd30;
|
|
ld.const.u32 %r5, [params+540];
|
|
setp.lt.s32 %p11, %r5, 1;
|
|
mov.f32 %f457, %f458;
|
|
mov.f32 %f456, %f458;
|
|
mov.f32 %f455, %f458;
|
|
@%p11 bra $L__BB0_34;
|
|
|
|
cvt.rn.f32.s32 %f153, %r5;
|
|
rcp.rn.f32 %f17, %f153;
|
|
ld.global.u32 %r271, [%rd3];
|
|
mul.f32 %f18, %f11, 0f3456BF95;
|
|
mul.f32 %f19, %f12, 0f3456BF95;
|
|
mul.f32 %f20, %f13, 0f3456BF95;
|
|
ld.const.u64 %rd4, [params+96];
|
|
mul.f32 %f154, %f448, %f15;
|
|
mul.f32 %f155, %f449, %f14;
|
|
sub.f32 %f21, %f155, %f154;
|
|
mul.f32 %f156, %f450, %f14;
|
|
mul.f32 %f157, %f448, %f16;
|
|
sub.f32 %f22, %f157, %f156;
|
|
mul.f32 %f158, %f449, %f16;
|
|
mul.f32 %f159, %f450, %f15;
|
|
sub.f32 %f23, %f159, %f158;
|
|
mov.u32 %r66, 0;
|
|
add.s64 %rd5, %rd1, 24;
|
|
mov.u64 %rd31, __cudart_i2opi_f;
|
|
abs.f32 %f225, %f19;
|
|
abs.f32 %f226, %f18;
|
|
max.f32 %f227, %f226, %f225;
|
|
abs.f32 %f228, %f20;
|
|
max.f32 %f229, %f227, %f228;
|
|
mov.u32 %r268, %r66;
|
|
|
|
$L__BB0_7:
|
|
cvt.rn.f32.s32 %f28, %r268;
|
|
mov.u32 %r270, %r66;
|
|
|
|
$L__BB0_8:
|
|
mad.lo.s32 %r68, %r271, 1664525, 1013904223;
|
|
and.b32 %r69, %r68, 16777215;
|
|
cvt.rn.f32.u32 %f160, %r69;
|
|
fma.rn.f32 %f161, %f160, 0f33800000, %f28;
|
|
mul.f32 %f162, %f17, %f161;
|
|
mad.lo.s32 %r271, %r68, 1664525, 1013904223;
|
|
and.b32 %r70, %r271, 16777215;
|
|
cvt.rn.f32.u32 %f163, %r70;
|
|
cvt.rn.f32.s32 %f164, %r270;
|
|
fma.rn.f32 %f165, %f163, 0f33800000, %f164;
|
|
mul.f32 %f166, %f17, %f165;
|
|
sqrt.rn.f32 %f33, %f162;
|
|
mul.f32 %f34, %f166, 0f40C90FDB;
|
|
mul.f32 %f167, %f34, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r278, %f167;
|
|
cvt.rn.f32.s32 %f168, %r278;
|
|
mov.f32 %f169, 0fBFC90FDA;
|
|
fma.rn.f32 %f170, %f168, %f169, %f34;
|
|
mov.f32 %f171, 0fB3A22168;
|
|
fma.rn.f32 %f172, %f168, %f171, %f170;
|
|
mov.f32 %f173, 0fA7C234C5;
|
|
fma.rn.f32 %f462, %f168, %f173, %f172;
|
|
abs.f32 %f36, %f34;
|
|
setp.ltu.f32 %p12, %f36, 0f47CE4780;
|
|
mov.u32 %r275, %r278;
|
|
mov.f32 %f459, %f462;
|
|
@%p12 bra $L__BB0_16;
|
|
|
|
setp.eq.f32 %p13, %f36, 0f7F800000;
|
|
@%p13 bra $L__BB0_15;
|
|
bra.uni $L__BB0_10;
|
|
|
|
$L__BB0_15:
|
|
mov.f32 %f176, 0f00000000;
|
|
mul.rn.f32 %f459, %f34, %f176;
|
|
mov.u32 %r275, 0;
|
|
bra.uni $L__BB0_16;
|
|
|
|
$L__BB0_10:
|
|
mov.b32 %r13, %f34;
|
|
bfe.u32 %r72, %r13, 23, 8;
|
|
add.s32 %r14, %r72, -128;
|
|
shl.b32 %r73, %r13, 8;
|
|
or.b32 %r15, %r73, -2147483648;
|
|
shr.u32 %r16, %r14, 5;
|
|
mov.u64 %rd79, 0;
|
|
mov.u32 %r272, 0;
|
|
mov.u64 %rd77, %rd1;
|
|
mov.u64 %rd78, %rd31;
|
|
|
|
$L__BB0_11:
|
|
.pragma "nounroll";
|
|
ld.global.nc.u32 %r74, [%rd78];
|
|
mad.wide.u32 %rd33, %r74, %r15, %rd79;
|
|
shr.u64 %rd79, %rd33, 32;
|
|
st.local.u32 [%rd77], %rd33;
|
|
add.s64 %rd78, %rd78, 4;
|
|
add.s64 %rd77, %rd77, 4;
|
|
add.s32 %r272, %r272, 1;
|
|
setp.ne.s32 %p14, %r272, 6;
|
|
@%p14 bra $L__BB0_11;
|
|
|
|
st.local.u32 [%rd5], %rd79;
|
|
mov.u32 %r75, 4;
|
|
sub.s32 %r19, %r75, %r16;
|
|
mov.u32 %r76, 6;
|
|
sub.s32 %r77, %r76, %r16;
|
|
mul.wide.s32 %rd34, %r77, 4;
|
|
add.s64 %rd35, %rd1, %rd34;
|
|
ld.local.u32 %r273, [%rd35];
|
|
ld.local.u32 %r274, [%rd35+-4];
|
|
and.b32 %r22, %r14, 31;
|
|
setp.eq.s32 %p15, %r22, 0;
|
|
@%p15 bra $L__BB0_14;
|
|
|
|
mov.u32 %r78, 32;
|
|
sub.s32 %r79, %r78, %r22;
|
|
shr.u32 %r80, %r274, %r79;
|
|
shl.b32 %r81, %r273, %r22;
|
|
add.s32 %r273, %r80, %r81;
|
|
mul.wide.s32 %rd36, %r19, 4;
|
|
add.s64 %rd37, %rd1, %rd36;
|
|
ld.local.u32 %r82, [%rd37];
|
|
shr.u32 %r83, %r82, %r79;
|
|
shl.b32 %r84, %r274, %r22;
|
|
add.s32 %r274, %r83, %r84;
|
|
|
|
$L__BB0_14:
|
|
and.b32 %r85, %r13, -2147483648;
|
|
shr.u32 %r86, %r274, 30;
|
|
shl.b32 %r87, %r273, 2;
|
|
or.b32 %r88, %r86, %r87;
|
|
shr.u32 %r89, %r88, 31;
|
|
shr.u32 %r90, %r273, 30;
|
|
add.s32 %r91, %r89, %r90;
|
|
neg.s32 %r92, %r91;
|
|
setp.eq.s32 %p16, %r85, 0;
|
|
selp.b32 %r275, %r91, %r92, %p16;
|
|
setp.ne.s32 %p17, %r89, 0;
|
|
xor.b32 %r93, %r85, -2147483648;
|
|
selp.b32 %r94, %r93, %r85, %p17;
|
|
selp.b32 %r95, -1, 0, %p17;
|
|
xor.b32 %r96, %r88, %r95;
|
|
shl.b32 %r97, %r274, 2;
|
|
xor.b32 %r98, %r97, %r95;
|
|
cvt.u64.u32 %rd38, %r96;
|
|
cvt.u64.u32 %rd39, %r98;
|
|
bfi.b64 %rd40, %rd38, %rd39, 32, 32;
|
|
cvt.rn.f64.s64 %fd1, %rd40;
|
|
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f174, %fd2;
|
|
setp.eq.s32 %p18, %r94, 0;
|
|
neg.f32 %f175, %f174;
|
|
selp.f32 %f459, %f174, %f175, %p18;
|
|
|
|
$L__BB0_16:
|
|
add.s32 %r29, %r275, 1;
|
|
and.b32 %r30, %r29, 1;
|
|
setp.eq.s32 %p19, %r30, 0;
|
|
selp.f32 %f40, %f459, 0f3F800000, %p19;
|
|
mul.rn.f32 %f41, %f459, %f459;
|
|
mov.f32 %f460, 0fB94D4153;
|
|
@%p19 bra $L__BB0_18;
|
|
|
|
mov.f32 %f178, 0fBAB607ED;
|
|
mov.f32 %f179, 0f37CBAC00;
|
|
fma.rn.f32 %f460, %f179, %f41, %f178;
|
|
|
|
$L__BB0_18:
|
|
selp.f32 %f180, 0f3C0885E4, 0f3D2AAABB, %p19;
|
|
fma.rn.f32 %f181, %f460, %f41, %f180;
|
|
selp.f32 %f182, 0fBE2AAAA8, 0fBEFFFFFF, %p19;
|
|
fma.rn.f32 %f183, %f181, %f41, %f182;
|
|
mov.f32 %f184, 0f00000000;
|
|
fma.rn.f32 %f185, %f41, %f40, %f184;
|
|
fma.rn.f32 %f461, %f183, %f185, %f40;
|
|
and.b32 %r100, %r29, 2;
|
|
setp.eq.s32 %p21, %r100, 0;
|
|
@%p21 bra $L__BB0_20;
|
|
|
|
mov.f32 %f187, 0fBF800000;
|
|
fma.rn.f32 %f461, %f461, %f187, %f184;
|
|
|
|
$L__BB0_20:
|
|
@%p12 bra $L__BB0_28;
|
|
|
|
setp.eq.f32 %p23, %f36, 0f7F800000;
|
|
@%p23 bra $L__BB0_27;
|
|
bra.uni $L__BB0_22;
|
|
|
|
$L__BB0_27:
|
|
mov.f32 %f190, 0f00000000;
|
|
mul.rn.f32 %f462, %f34, %f190;
|
|
mov.u32 %r278, 0;
|
|
bra.uni $L__BB0_28;
|
|
|
|
$L__BB0_22:
|
|
mov.b32 %r31, %f34;
|
|
bfe.u32 %r101, %r31, 23, 8;
|
|
add.s32 %r32, %r101, -128;
|
|
shl.b32 %r102, %r31, 8;
|
|
or.b32 %r33, %r102, -2147483648;
|
|
shr.u32 %r34, %r32, 5;
|
|
mov.u64 %rd80, 0;
|
|
mov.u64 %rd81, %rd80;
|
|
|
|
$L__BB0_23:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd43, %rd80, 2;
|
|
mov.u64 %rd44, __cudart_i2opi_f;
|
|
add.s64 %rd45, %rd44, %rd43;
|
|
ld.global.nc.u32 %r103, [%rd45];
|
|
mad.wide.u32 %rd46, %r103, %r33, %rd81;
|
|
shr.u64 %rd81, %rd46, 32;
|
|
add.s64 %rd47, %rd1, %rd43;
|
|
st.local.u32 [%rd47], %rd46;
|
|
cvt.u32.u64 %r104, %rd80;
|
|
add.s32 %r105, %r104, 1;
|
|
cvt.s64.s32 %rd80, %r105;
|
|
setp.ne.s32 %p24, %r105, 6;
|
|
@%p24 bra $L__BB0_23;
|
|
|
|
st.local.u32 [%rd5], %rd81;
|
|
mov.u32 %r106, 4;
|
|
sub.s32 %r35, %r106, %r34;
|
|
mov.u32 %r107, 6;
|
|
sub.s32 %r108, %r107, %r34;
|
|
mul.wide.s32 %rd48, %r108, 4;
|
|
add.s64 %rd49, %rd1, %rd48;
|
|
ld.local.u32 %r276, [%rd49];
|
|
ld.local.u32 %r277, [%rd49+-4];
|
|
and.b32 %r38, %r32, 31;
|
|
setp.eq.s32 %p25, %r38, 0;
|
|
@%p25 bra $L__BB0_26;
|
|
|
|
mov.u32 %r109, 32;
|
|
sub.s32 %r110, %r109, %r38;
|
|
shr.u32 %r111, %r277, %r110;
|
|
shl.b32 %r112, %r276, %r38;
|
|
add.s32 %r276, %r111, %r112;
|
|
mul.wide.s32 %rd50, %r35, 4;
|
|
add.s64 %rd51, %rd1, %rd50;
|
|
ld.local.u32 %r113, [%rd51];
|
|
shr.u32 %r114, %r113, %r110;
|
|
shl.b32 %r115, %r277, %r38;
|
|
add.s32 %r277, %r114, %r115;
|
|
|
|
$L__BB0_26:
|
|
and.b32 %r116, %r31, -2147483648;
|
|
shr.u32 %r117, %r277, 30;
|
|
shl.b32 %r118, %r276, 2;
|
|
or.b32 %r119, %r117, %r118;
|
|
shr.u32 %r120, %r119, 31;
|
|
shr.u32 %r121, %r276, 30;
|
|
add.s32 %r122, %r120, %r121;
|
|
neg.s32 %r123, %r122;
|
|
setp.eq.s32 %p26, %r116, 0;
|
|
selp.b32 %r278, %r122, %r123, %p26;
|
|
setp.ne.s32 %p27, %r120, 0;
|
|
xor.b32 %r124, %r116, -2147483648;
|
|
selp.b32 %r125, %r124, %r116, %p27;
|
|
selp.b32 %r126, -1, 0, %p27;
|
|
xor.b32 %r127, %r119, %r126;
|
|
shl.b32 %r128, %r277, 2;
|
|
xor.b32 %r129, %r128, %r126;
|
|
cvt.u64.u32 %rd52, %r127;
|
|
cvt.u64.u32 %rd53, %r129;
|
|
bfi.b64 %rd54, %rd52, %rd53, 32, 32;
|
|
cvt.rn.f64.s64 %fd3, %rd54;
|
|
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f188, %fd4;
|
|
setp.eq.s32 %p28, %r125, 0;
|
|
neg.f32 %f189, %f188;
|
|
selp.f32 %f462, %f188, %f189, %p28;
|
|
|
|
$L__BB0_28:
|
|
mul.f32 %f50, %f33, %f461;
|
|
and.b32 %r45, %r278, 1;
|
|
setp.eq.s32 %p29, %r45, 0;
|
|
selp.f32 %f51, %f462, 0f3F800000, %p29;
|
|
mul.rn.f32 %f52, %f462, %f462;
|
|
mov.f32 %f463, 0fB94D4153;
|
|
@%p29 bra $L__BB0_30;
|
|
|
|
mov.f32 %f192, 0fBAB607ED;
|
|
mov.f32 %f193, 0f37CBAC00;
|
|
fma.rn.f32 %f463, %f193, %f52, %f192;
|
|
|
|
$L__BB0_30:
|
|
selp.f32 %f194, 0f3C0885E4, 0f3D2AAABB, %p29;
|
|
fma.rn.f32 %f195, %f463, %f52, %f194;
|
|
selp.f32 %f196, 0fBE2AAAA8, 0fBEFFFFFF, %p29;
|
|
fma.rn.f32 %f197, %f195, %f52, %f196;
|
|
mov.f32 %f198, 0f00000000;
|
|
fma.rn.f32 %f199, %f52, %f51, %f198;
|
|
fma.rn.f32 %f464, %f197, %f199, %f51;
|
|
and.b32 %r131, %r278, 2;
|
|
setp.eq.s32 %p31, %r131, 0;
|
|
@%p31 bra $L__BB0_32;
|
|
|
|
mov.f32 %f201, 0fBF800000;
|
|
fma.rn.f32 %f464, %f464, %f201, %f198;
|
|
|
|
$L__BB0_32:
|
|
mul.f32 %f211, %f50, %f50;
|
|
mov.f32 %f212, 0f3F800000;
|
|
sub.f32 %f213, %f212, %f211;
|
|
mul.f32 %f214, %f33, %f464;
|
|
mul.f32 %f215, %f214, %f214;
|
|
sub.f32 %f216, %f213, %f215;
|
|
max.f32 %f217, %f198, %f216;
|
|
sqrt.rn.f32 %f218, %f217;
|
|
mul.f32 %f219, %f14, %f214;
|
|
mul.f32 %f220, %f15, %f214;
|
|
mul.f32 %f221, %f16, %f214;
|
|
fma.rn.f32 %f222, %f23, %f50, %f219;
|
|
fma.rn.f32 %f223, %f22, %f50, %f220;
|
|
fma.rn.f32 %f224, %f21, %f50, %f221;
|
|
fma.rn.f32 %f205, %f448, %f218, %f222;
|
|
fma.rn.f32 %f206, %f449, %f218, %f223;
|
|
fma.rn.f32 %f207, %f450, %f218, %f224;
|
|
mov.f32 %f230, 0f38D1B717;
|
|
max.f32 %f208, %f229, %f230;
|
|
mov.u32 %r168, 2;
|
|
mov.u32 %r170, 1;
|
|
mov.u32 %r171, 1065353216;
|
|
mov.u32 %r202, 0;
|
|
// begin inline asm
|
|
call(%r132,%r133,%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),_optix_trace_typed_32,(%r202,%rd4,%f11,%f12,%f13,%f205,%f206,%f207,%f208,%f1,%f198,%r170,%r202,%r170,%r168,%r170,%r170,%r171,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202,%r202);
|
|
// end inline asm
|
|
mov.b32 %f231, %r132;
|
|
add.f32 %f455, %f455, %f231;
|
|
fma.rn.f32 %f458, %f205, %f231, %f458;
|
|
fma.rn.f32 %f457, %f206, %f231, %f457;
|
|
fma.rn.f32 %f456, %f207, %f231, %f456;
|
|
add.s32 %r270, %r270, 1;
|
|
setp.lt.s32 %p32, %r270, %r5;
|
|
@%p32 bra $L__BB0_8;
|
|
|
|
add.s32 %r268, %r268, 1;
|
|
setp.lt.s32 %p33, %r268, %r5;
|
|
@%p33 bra $L__BB0_7;
|
|
|
|
$L__BB0_34:
|
|
mul.lo.s32 %r203, %r5, %r5;
|
|
cvt.rn.f32.s32 %f232, %r203;
|
|
div.rn.f32 %f66, %f455, %f232;
|
|
ld.const.f32 %f67, [params+704];
|
|
setp.lt.f32 %p34, %f67, 0f00000000;
|
|
@%p34 bra $L__BB0_36;
|
|
bra.uni $L__BB0_35;
|
|
|
|
$L__BB0_36:
|
|
neg.f32 %f69, %f67;
|
|
setp.lt.f32 %p35, %f67, 0fBF800000;
|
|
@%p35 bra $L__BB0_38;
|
|
bra.uni $L__BB0_37;
|
|
|
|
$L__BB0_38:
|
|
mul.f32 %f238, %f67, 0fBF000000;
|
|
cvt.rzi.f32.f32 %f239, %f238;
|
|
add.f32 %f240, %f239, %f239;
|
|
sub.f32 %f241, %f69, %f240;
|
|
abs.f32 %f71, %f241;
|
|
abs.f32 %f72, %f66;
|
|
setp.lt.f32 %p36, %f72, 0f00800000;
|
|
mul.f32 %f242, %f72, 0f4B800000;
|
|
selp.f32 %f243, %f242, %f72, %p36;
|
|
selp.f32 %f244, 0fC3170000, 0fC2FE0000, %p36;
|
|
mov.b32 %r204, %f243;
|
|
and.b32 %r205, %r204, 8388607;
|
|
or.b32 %r206, %r205, 1065353216;
|
|
mov.b32 %f245, %r206;
|
|
shr.u32 %r207, %r204, 23;
|
|
cvt.rn.f32.u32 %f246, %r207;
|
|
add.f32 %f247, %f244, %f246;
|
|
setp.gt.f32 %p37, %f245, 0f3FB504F3;
|
|
mul.f32 %f248, %f245, 0f3F000000;
|
|
add.f32 %f249, %f247, 0f3F800000;
|
|
selp.f32 %f250, %f249, %f247, %p37;
|
|
selp.f32 %f251, %f248, %f245, %p37;
|
|
add.f32 %f252, %f251, 0fBF800000;
|
|
add.f32 %f253, %f251, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f254, %f253;
|
|
add.f32 %f255, %f252, %f252;
|
|
mul.f32 %f256, %f255, %f254;
|
|
mul.f32 %f257, %f256, %f256;
|
|
mov.f32 %f258, 0f3C4CAF63;
|
|
mov.f32 %f259, 0f3B18F0FE;
|
|
fma.rn.f32 %f260, %f259, %f257, %f258;
|
|
mov.f32 %f261, 0f3DAAAABD;
|
|
fma.rn.f32 %f262, %f260, %f257, %f261;
|
|
mul.rn.f32 %f263, %f262, %f257;
|
|
mul.rn.f32 %f264, %f263, %f256;
|
|
sub.f32 %f265, %f252, %f256;
|
|
add.f32 %f266, %f265, %f265;
|
|
neg.f32 %f267, %f256;
|
|
fma.rn.f32 %f268, %f267, %f252, %f266;
|
|
mul.rn.f32 %f269, %f254, %f268;
|
|
add.f32 %f270, %f264, %f256;
|
|
sub.f32 %f271, %f256, %f270;
|
|
add.f32 %f272, %f264, %f271;
|
|
add.f32 %f273, %f269, %f272;
|
|
add.f32 %f274, %f270, %f273;
|
|
sub.f32 %f275, %f270, %f274;
|
|
add.f32 %f276, %f273, %f275;
|
|
mov.f32 %f277, 0f3F317200;
|
|
mul.rn.f32 %f278, %f250, %f277;
|
|
mov.f32 %f279, 0f35BFBE8E;
|
|
mul.rn.f32 %f280, %f250, %f279;
|
|
add.f32 %f281, %f278, %f274;
|
|
sub.f32 %f282, %f278, %f281;
|
|
add.f32 %f283, %f274, %f282;
|
|
add.f32 %f284, %f276, %f283;
|
|
add.f32 %f285, %f280, %f284;
|
|
add.f32 %f286, %f281, %f285;
|
|
sub.f32 %f287, %f281, %f286;
|
|
add.f32 %f288, %f285, %f287;
|
|
abs.f32 %f73, %f69;
|
|
setp.gt.f32 %p38, %f73, 0f77F684DF;
|
|
mul.f32 %f289, %f67, 0fB9000000;
|
|
selp.f32 %f290, %f289, %f69, %p38;
|
|
mul.rn.f32 %f291, %f290, %f286;
|
|
neg.f32 %f292, %f291;
|
|
fma.rn.f32 %f293, %f290, %f286, %f292;
|
|
fma.rn.f32 %f294, %f290, %f288, %f293;
|
|
mov.f32 %f295, 0f00000000;
|
|
fma.rn.f32 %f296, %f295, %f286, %f294;
|
|
add.rn.f32 %f297, %f291, %f296;
|
|
neg.f32 %f298, %f297;
|
|
add.rn.f32 %f299, %f291, %f298;
|
|
add.rn.f32 %f300, %f299, %f296;
|
|
mov.b32 %r208, %f297;
|
|
setp.eq.s32 %p39, %r208, 1118925336;
|
|
add.s32 %r209, %r208, -1;
|
|
mov.b32 %f301, %r209;
|
|
add.f32 %f302, %f300, 0f37000000;
|
|
selp.f32 %f74, %f302, %f300, %p39;
|
|
selp.f32 %f303, %f301, %f297, %p39;
|
|
mov.f32 %f304, 0f3FB8AA3B;
|
|
mul.rn.f32 %f305, %f303, %f304;
|
|
cvt.rzi.f32.f32 %f306, %f305;
|
|
abs.f32 %f307, %f306;
|
|
setp.gt.f32 %p40, %f307, 0f42FC0000;
|
|
mov.b32 %r210, %f306;
|
|
and.b32 %r211, %r210, -2147483648;
|
|
or.b32 %r212, %r211, 1123811328;
|
|
mov.b32 %f308, %r212;
|
|
selp.f32 %f309, %f308, %f306, %p40;
|
|
mov.f32 %f310, 0fBF317218;
|
|
fma.rn.f32 %f311, %f309, %f310, %f303;
|
|
mov.f32 %f312, 0f3102E308;
|
|
fma.rn.f32 %f313, %f309, %f312, %f311;
|
|
mul.f32 %f314, %f313, 0f3FB8AA3B;
|
|
add.f32 %f315, %f309, 0f4B40007F;
|
|
mov.b32 %r213, %f315;
|
|
shl.b32 %r214, %r213, 23;
|
|
mov.b32 %f316, %r214;
|
|
ex2.approx.ftz.f32 %f317, %f314;
|
|
mul.f32 %f75, %f317, %f316;
|
|
setp.eq.f32 %p41, %f75, 0f7F800000;
|
|
mov.f32 %f469, 0f7F800000;
|
|
@%p41 bra $L__BB0_40;
|
|
|
|
fma.rn.f32 %f469, %f75, %f74, %f75;
|
|
|
|
$L__BB0_40:
|
|
setp.lt.f32 %p42, %f66, 0f00000000;
|
|
setp.eq.f32 %p43, %f71, 0f3F800000;
|
|
and.pred %p1, %p42, %p43;
|
|
setp.eq.f32 %p44, %f66, 0f00000000;
|
|
@%p44 bra $L__BB0_44;
|
|
bra.uni $L__BB0_41;
|
|
|
|
$L__BB0_44:
|
|
add.f32 %f321, %f66, %f66;
|
|
mov.b32 %r217, %f321;
|
|
selp.b32 %r218, %r217, 0, %p43;
|
|
or.b32 %r219, %r218, 2139095040;
|
|
setp.gt.f32 %p48, %f67, 0f80000000;
|
|
selp.b32 %r220, %r219, %r218, %p48;
|
|
mov.b32 %f471, %r220;
|
|
bra.uni $L__BB0_45;
|
|
|
|
$L__BB0_35:
|
|
add.f32 %f233, %f66, 0fBF800000;
|
|
fma.rn.f32 %f234, %f67, %f233, 0f3F800000;
|
|
cvt.sat.f32.f32 %f472, %f234;
|
|
bra.uni $L__BB0_53;
|
|
|
|
$L__BB0_37:
|
|
add.f32 %f235, %f66, 0fBF800000;
|
|
fma.rn.f32 %f236, %f235, %f69, 0f3F800000;
|
|
cvt.sat.f32.f32 %f472, %f236;
|
|
bra.uni $L__BB0_53;
|
|
|
|
$L__BB0_76:
|
|
mov.f32 %f446, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f446;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f446;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f446;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs50, 0;
|
|
st.global.v4.u16 [%rd19], {%rs47, %rs48, %rs49, %rs50};
|
|
|
|
$L__BB0_77:
|
|
and.b32 %r260, %r49, 8;
|
|
setp.eq.s32 %p83, %r260, 0;
|
|
@%p83 bra $L__BB0_79;
|
|
|
|
ld.const.u64 %rd69, [params+192];
|
|
cvta.to.global.u64 %rd70, %rd69;
|
|
ld.const.u32 %r261, [params+184];
|
|
mad.lo.s32 %r262, %r261, %r4, %r3;
|
|
mov.f32 %f447, 0f00000000;
|
|
cvt.rzi.u32.f32 %r263, %f447;
|
|
mul.wide.u32 %rd71, %r262, 2;
|
|
add.s64 %rd72, %rd70, %rd71;
|
|
mov.u16 %rs51, 0;
|
|
cvt.u16.u32 %rs52, %r263;
|
|
st.global.v2.u8 [%rd72], {%rs52, %rs51};
|
|
|
|
$L__BB0_79:
|
|
and.b32 %r264, %r49, 64;
|
|
setp.eq.s32 %p84, %r264, 0;
|
|
@%p84 bra $L__BB0_81;
|
|
|
|
ld.const.u64 %rd73, [params+208];
|
|
cvta.to.global.u64 %rd74, %rd73;
|
|
ld.const.u32 %r265, [params+200];
|
|
mad.lo.s32 %r266, %r265, %r4, %r3;
|
|
mul.wide.u32 %rd75, %r266, 4;
|
|
add.s64 %rd76, %rd74, %rd75;
|
|
mov.u16 %rs53, 0;
|
|
st.global.v4.u8 [%rd76], {%rs53, %rs53, %rs53, %rs53};
|
|
bra.uni $L__BB0_81;
|
|
|
|
$L__BB0_41:
|
|
mov.b32 %r215, %f469;
|
|
xor.b32 %r216, %r215, -2147483648;
|
|
mov.b32 %f318, %r216;
|
|
selp.f32 %f471, %f318, %f469, %p1;
|
|
setp.geu.f32 %p45, %f66, 0f00000000;
|
|
@%p45 bra $L__BB0_45;
|
|
|
|
cvt.rzi.f32.f32 %f319, %f69;
|
|
setp.eq.f32 %p46, %f319, %f69;
|
|
@%p46 bra $L__BB0_45;
|
|
|
|
mov.f32 %f471, 0f7FFFFFFF;
|
|
|
|
$L__BB0_45:
|
|
add.f32 %f322, %f72, %f73;
|
|
mov.b32 %r221, %f322;
|
|
setp.lt.s32 %p49, %r221, 2139095040;
|
|
@%p49 bra $L__BB0_52;
|
|
|
|
setp.gtu.f32 %p50, %f72, 0f7F800000;
|
|
setp.gtu.f32 %p51, %f73, 0f7F800000;
|
|
or.pred %p52, %p50, %p51;
|
|
@%p52 bra $L__BB0_51;
|
|
bra.uni $L__BB0_47;
|
|
|
|
$L__BB0_51:
|
|
sub.f32 %f471, %f66, %f67;
|
|
bra.uni $L__BB0_52;
|
|
|
|
$L__BB0_47:
|
|
setp.eq.f32 %p53, %f73, 0f7F800000;
|
|
@%p53 bra $L__BB0_50;
|
|
bra.uni $L__BB0_48;
|
|
|
|
$L__BB0_50:
|
|
setp.gt.f32 %p56, %f72, 0f3F800000;
|
|
selp.b32 %r225, 2139095040, 0, %p56;
|
|
xor.b32 %r226, %r225, 2139095040;
|
|
setp.gt.f32 %p57, %f67, 0f80000000;
|
|
selp.b32 %r227, %r226, %r225, %p57;
|
|
mov.b32 %f323, %r227;
|
|
setp.eq.f32 %p58, %f66, 0fBF800000;
|
|
selp.f32 %f471, 0f3F800000, %f323, %p58;
|
|
bra.uni $L__BB0_52;
|
|
|
|
$L__BB0_48:
|
|
setp.neu.f32 %p54, %f72, 0f7F800000;
|
|
@%p54 bra $L__BB0_52;
|
|
|
|
setp.le.f32 %p55, %f67, 0f80000000;
|
|
selp.b32 %r222, 2139095040, 0, %p55;
|
|
or.b32 %r223, %r222, -2147483648;
|
|
selp.b32 %r224, %r223, %r222, %p1;
|
|
mov.b32 %f471, %r224;
|
|
|
|
$L__BB0_52:
|
|
setp.eq.f32 %p59, %f67, 0f80000000;
|
|
setp.eq.f32 %p60, %f66, 0f3F800000;
|
|
or.pred %p61, %p59, %p60;
|
|
selp.f32 %f324, 0f3F800000, %f471, %p61;
|
|
cvt.sat.f32.f32 %f472, %f324;
|
|
|
|
$L__BB0_53:
|
|
ld.const.u32 %r48, [params+104];
|
|
and.b32 %r228, %r48, 4;
|
|
setp.eq.s32 %p62, %r228, 0;
|
|
@%p62 bra $L__BB0_57;
|
|
|
|
ld.const.u32 %r229, [params+108];
|
|
setp.eq.s32 %p63, %r229, 0;
|
|
ld.const.u64 %rd56, [params+224];
|
|
cvta.to.global.u64 %rd57, %rd56;
|
|
ld.const.u32 %r230, [params+216];
|
|
mad.lo.s32 %r231, %r230, %r4, %r3;
|
|
mul.wide.u32 %rd58, %r231, 8;
|
|
add.s64 %rd16, %rd57, %rd58;
|
|
@%p63 bra $L__BB0_56;
|
|
|
|
ld.global.v4.u16 {%rs22, %rs23, %rs24, %rs25}, [%rd16];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f325, %rs22;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f326, %rs23;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f327, %rs24;}
|
|
|
|
// end inline asm
|
|
add.f32 %f328, %f472, %f325;
|
|
add.f32 %f329, %f472, %f326;
|
|
add.f32 %f330, %f472, %f327;
|
|
mov.f32 %f331, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs20, %f330;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs19, %f329;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs18, %f328;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f331;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd16], {%rs18, %rs19, %rs20, %rs21};
|
|
bra.uni $L__BB0_57;
|
|
|
|
$L__BB0_56:
|
|
mov.f32 %f335, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f335;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f472;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f472;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f472;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd16], {%rs26, %rs27, %rs28, %rs29};
|
|
|
|
$L__BB0_57:
|
|
and.b32 %r232, %r48, 8;
|
|
setp.eq.s32 %p64, %r232, 0;
|
|
@%p64 bra $L__BB0_71;
|
|
|
|
ld.const.u64 %rd59, [params+192];
|
|
cvta.to.global.u64 %rd17, %rd59;
|
|
ld.const.u32 %r233, [params+184];
|
|
mad.lo.s32 %r234, %r233, %r4, %r3;
|
|
cvt.u64.u32 %rd18, %r234;
|
|
mov.f32 %f337, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f338, %f337;
|
|
add.f32 %f339, %f338, %f338;
|
|
mov.f32 %f340, 0f3EE8BA2E;
|
|
sub.f32 %f341, %f340, %f339;
|
|
abs.f32 %f87, %f341;
|
|
abs.f32 %f88, %f472;
|
|
setp.lt.f32 %p65, %f88, 0f00800000;
|
|
mul.f32 %f342, %f88, 0f4B800000;
|
|
selp.f32 %f343, %f342, %f88, %p65;
|
|
selp.f32 %f344, 0fC3170000, 0fC2FE0000, %p65;
|
|
mov.b32 %r235, %f343;
|
|
and.b32 %r236, %r235, 8388607;
|
|
or.b32 %r237, %r236, 1065353216;
|
|
mov.b32 %f345, %r237;
|
|
shr.u32 %r238, %r235, 23;
|
|
cvt.rn.f32.u32 %f346, %r238;
|
|
add.f32 %f347, %f344, %f346;
|
|
setp.gt.f32 %p66, %f345, 0f3FB504F3;
|
|
mul.f32 %f348, %f345, 0f3F000000;
|
|
add.f32 %f349, %f347, 0f3F800000;
|
|
selp.f32 %f350, %f349, %f347, %p66;
|
|
selp.f32 %f351, %f348, %f345, %p66;
|
|
add.f32 %f352, %f351, 0fBF800000;
|
|
add.f32 %f353, %f351, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f354, %f353;
|
|
add.f32 %f355, %f352, %f352;
|
|
mul.f32 %f356, %f355, %f354;
|
|
mul.f32 %f357, %f356, %f356;
|
|
mov.f32 %f358, 0f3C4CAF63;
|
|
mov.f32 %f359, 0f3B18F0FE;
|
|
fma.rn.f32 %f360, %f359, %f357, %f358;
|
|
mov.f32 %f361, 0f3DAAAABD;
|
|
fma.rn.f32 %f362, %f360, %f357, %f361;
|
|
mul.rn.f32 %f363, %f362, %f357;
|
|
mul.rn.f32 %f364, %f363, %f356;
|
|
sub.f32 %f365, %f352, %f356;
|
|
add.f32 %f366, %f365, %f365;
|
|
neg.f32 %f367, %f356;
|
|
fma.rn.f32 %f368, %f367, %f352, %f366;
|
|
mul.rn.f32 %f369, %f354, %f368;
|
|
add.f32 %f370, %f364, %f356;
|
|
sub.f32 %f371, %f356, %f370;
|
|
add.f32 %f372, %f364, %f371;
|
|
add.f32 %f373, %f369, %f372;
|
|
add.f32 %f374, %f370, %f373;
|
|
sub.f32 %f375, %f370, %f374;
|
|
add.f32 %f376, %f373, %f375;
|
|
mov.f32 %f377, 0f3F317200;
|
|
mul.rn.f32 %f378, %f350, %f377;
|
|
mov.f32 %f379, 0f35BFBE8E;
|
|
mul.rn.f32 %f380, %f350, %f379;
|
|
add.f32 %f381, %f378, %f374;
|
|
sub.f32 %f382, %f378, %f381;
|
|
add.f32 %f383, %f374, %f382;
|
|
add.f32 %f384, %f376, %f383;
|
|
add.f32 %f385, %f380, %f384;
|
|
add.f32 %f386, %f381, %f385;
|
|
sub.f32 %f387, %f381, %f386;
|
|
add.f32 %f388, %f385, %f387;
|
|
mul.rn.f32 %f389, %f340, %f386;
|
|
neg.f32 %f390, %f389;
|
|
fma.rn.f32 %f391, %f340, %f386, %f390;
|
|
fma.rn.f32 %f392, %f340, %f388, %f391;
|
|
mov.f32 %f393, 0f00000000;
|
|
fma.rn.f32 %f394, %f393, %f386, %f392;
|
|
add.rn.f32 %f395, %f389, %f394;
|
|
neg.f32 %f396, %f395;
|
|
add.rn.f32 %f397, %f389, %f396;
|
|
add.rn.f32 %f398, %f397, %f394;
|
|
mov.b32 %r239, %f395;
|
|
setp.eq.s32 %p67, %r239, 1118925336;
|
|
add.s32 %r240, %r239, -1;
|
|
mov.b32 %f399, %r240;
|
|
add.f32 %f400, %f398, 0f37000000;
|
|
selp.f32 %f89, %f400, %f398, %p67;
|
|
selp.f32 %f401, %f399, %f395, %p67;
|
|
mov.f32 %f402, 0f3FB8AA3B;
|
|
mul.rn.f32 %f403, %f401, %f402;
|
|
cvt.rzi.f32.f32 %f404, %f403;
|
|
abs.f32 %f405, %f404;
|
|
setp.gt.f32 %p68, %f405, 0f42FC0000;
|
|
mov.b32 %r241, %f404;
|
|
and.b32 %r242, %r241, -2147483648;
|
|
or.b32 %r243, %r242, 1123811328;
|
|
mov.b32 %f406, %r243;
|
|
selp.f32 %f407, %f406, %f404, %p68;
|
|
mov.f32 %f408, 0fBF317218;
|
|
fma.rn.f32 %f409, %f407, %f408, %f401;
|
|
mov.f32 %f410, 0f3102E308;
|
|
fma.rn.f32 %f411, %f407, %f410, %f409;
|
|
mul.f32 %f412, %f411, 0f3FB8AA3B;
|
|
add.f32 %f413, %f407, 0f4B40007F;
|
|
mov.b32 %r244, %f413;
|
|
shl.b32 %r245, %r244, 23;
|
|
mov.b32 %f414, %r245;
|
|
ex2.approx.ftz.f32 %f415, %f412;
|
|
mul.f32 %f90, %f415, %f414;
|
|
setp.eq.f32 %p69, %f90, 0f7F800000;
|
|
mov.f32 %f473, 0f7F800000;
|
|
@%p69 bra $L__BB0_60;
|
|
|
|
fma.rn.f32 %f473, %f90, %f89, %f90;
|
|
|
|
$L__BB0_60:
|
|
setp.lt.f32 %p70, %f472, 0f00000000;
|
|
setp.eq.f32 %p71, %f87, 0f3F800000;
|
|
and.pred %p2, %p70, %p71;
|
|
setp.eq.f32 %p72, %f472, 0f00000000;
|
|
@%p72 bra $L__BB0_64;
|
|
bra.uni $L__BB0_61;
|
|
|
|
$L__BB0_64:
|
|
add.f32 %f420, %f472, %f472;
|
|
selp.f32 %f475, %f420, 0f00000000, %p71;
|
|
bra.uni $L__BB0_65;
|
|
|
|
$L__BB0_61:
|
|
mov.b32 %r246, %f473;
|
|
xor.b32 %r247, %r246, -2147483648;
|
|
mov.b32 %f416, %r247;
|
|
selp.f32 %f475, %f416, %f473, %p2;
|
|
setp.geu.f32 %p73, %f472, 0f00000000;
|
|
@%p73 bra $L__BB0_65;
|
|
|
|
mov.f32 %f417, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f418, %f417;
|
|
setp.eq.f32 %p74, %f418, 0f3EE8BA2E;
|
|
@%p74 bra $L__BB0_65;
|
|
|
|
mov.f32 %f475, 0f7FFFFFFF;
|
|
|
|
$L__BB0_65:
|
|
add.f32 %f421, %f88, 0f3EE8BA2E;
|
|
mov.b32 %r248, %f421;
|
|
setp.lt.s32 %p76, %r248, 2139095040;
|
|
@%p76 bra $L__BB0_70;
|
|
|
|
setp.gtu.f32 %p77, %f88, 0f7F800000;
|
|
@%p77 bra $L__BB0_69;
|
|
bra.uni $L__BB0_67;
|
|
|
|
$L__BB0_69:
|
|
add.f32 %f475, %f472, 0f3EE8BA2E;
|
|
bra.uni $L__BB0_70;
|
|
|
|
$L__BB0_67:
|
|
setp.neu.f32 %p78, %f88, 0f7F800000;
|
|
@%p78 bra $L__BB0_70;
|
|
|
|
selp.f32 %f475, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
$L__BB0_70:
|
|
mul.f32 %f422, %f475, 0f437F0000;
|
|
setp.eq.f32 %p79, %f472, 0f3F800000;
|
|
selp.f32 %f423, 0f437F0000, %f422, %p79;
|
|
cvt.rzi.u32.f32 %r249, %f423;
|
|
shl.b64 %rd60, %rd18, 1;
|
|
add.s64 %rd61, %rd17, %rd60;
|
|
cvt.u16.u32 %rs30, %r249;
|
|
mov.u16 %rs31, 255;
|
|
st.global.v2.u8 [%rd61], {%rs30, %rs31};
|
|
|
|
$L__BB0_71:
|
|
ld.const.u32 %r267, [params+104];
|
|
and.b32 %r250, %r267, 64;
|
|
setp.eq.s32 %p80, %r250, 0;
|
|
@%p80 bra $L__BB0_81;
|
|
|
|
mul.f32 %f424, %f457, %f457;
|
|
fma.rn.f32 %f425, %f458, %f458, %f424;
|
|
fma.rn.f32 %f426, %f456, %f456, %f425;
|
|
sqrt.rn.f32 %f427, %f426;
|
|
rcp.rn.f32 %f428, %f427;
|
|
mul.f32 %f429, %f458, %f428;
|
|
mul.f32 %f430, %f457, %f428;
|
|
mul.f32 %f431, %f456, %f428;
|
|
ld.const.u64 %rd62, [params+208];
|
|
cvta.to.global.u64 %rd63, %rd62;
|
|
ld.const.u32 %r251, [params+200];
|
|
mad.lo.s32 %r252, %r251, %r4, %r3;
|
|
fma.rn.f32 %f432, %f429, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f433, %f432, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r253, %f433;
|
|
fma.rn.f32 %f434, %f430, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f435, %f434, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r254, %f435;
|
|
fma.rn.f32 %f436, %f431, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f437, %f436, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r255, %f437;
|
|
mul.wide.u32 %rd64, %r252, 4;
|
|
add.s64 %rd65, %rd63, %rd64;
|
|
cvt.u16.u32 %rs32, %r255;
|
|
cvt.u16.u32 %rs33, %r254;
|
|
cvt.u16.u32 %rs34, %r253;
|
|
mov.u16 %rs35, 255;
|
|
st.global.v4.u8 [%rd65], {%rs34, %rs33, %rs32, %rs35};
|
|
|
|
$L__BB0_81:
|
|
ret;
|
|
|
|
}
|
|
|