2252 lines
60 KiB
Plaintext
2252 lines
60 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<156>;
|
|
.reg .b16 %rs<166>;
|
|
.reg .f32 %f<1167>;
|
|
.reg .b32 %r<364>;
|
|
.reg .f64 %fd<5>;
|
|
.reg .b64 %rd<122>;
|
|
|
|
|
|
mov.u64 %SPL, __local_depot0;
|
|
add.u64 %rd1, %SPL, 0;
|
|
// begin inline asm
|
|
call (%r53), _optix_get_launch_index_x, ();
|
|
// end inline asm
|
|
// begin inline asm
|
|
call (%r54), _optix_get_launch_index_y, ();
|
|
// end inline asm
|
|
ld.const.u32 %r3, [params+540];
|
|
shl.b32 %r4, %r3, 1;
|
|
ld.const.u64 %rd27, [params+400];
|
|
cvta.to.global.u64 %rd28, %rd27;
|
|
ld.const.u32 %r59, [params+392];
|
|
mad.lo.s32 %r60, %r59, %r54, %r53;
|
|
mul.wide.u32 %rd29, %r60, 4;
|
|
add.s64 %rd2, %rd28, %rd29;
|
|
ld.global.v2.u8 {%rs7, %rs165}, [%rd2];
|
|
or.b16 %rs9, %rs7, %rs165;
|
|
and.b16 %rs10, %rs9, 255;
|
|
setp.eq.s16 %p7, %rs10, 0;
|
|
@%p7 bra $L__BB0_2;
|
|
|
|
ld.global.u8 %rs164, [%rd2+2];
|
|
bra.uni $L__BB0_3;
|
|
|
|
$L__BB0_2:
|
|
ld.global.u8 %rs164, [%rd2+2];
|
|
setp.eq.s16 %p8, %rs164, 0;
|
|
mov.f32 %f1092, 0f00000000;
|
|
mov.u16 %rs165, 0;
|
|
mov.f32 %f1093, %f1092;
|
|
mov.f32 %f1094, %f1092;
|
|
@%p8 bra $L__BB0_4;
|
|
|
|
$L__BB0_3:
|
|
cvt.rn.f32.u16 %f228, %rs7;
|
|
div.rn.f32 %f229, %f228, 0f437F0000;
|
|
fma.rn.f32 %f230, %f229, 0f40000000, 0fBF800000;
|
|
and.b16 %rs13, %rs165, 255;
|
|
cvt.rn.f32.u16 %f231, %rs13;
|
|
div.rn.f32 %f232, %f231, 0f437F0000;
|
|
fma.rn.f32 %f233, %f232, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f234, %rs164;
|
|
div.rn.f32 %f235, %f234, 0f437F0000;
|
|
fma.rn.f32 %f236, %f235, 0f40000000, 0fBF800000;
|
|
mul.f32 %f237, %f233, %f233;
|
|
fma.rn.f32 %f238, %f230, %f230, %f237;
|
|
fma.rn.f32 %f239, %f236, %f236, %f238;
|
|
sqrt.rn.f32 %f240, %f239;
|
|
rcp.rn.f32 %f241, %f240;
|
|
mul.f32 %f1094, %f241, %f236;
|
|
mul.f32 %f1093, %f241, %f233;
|
|
mul.f32 %f1092, %f230, %f241;
|
|
|
|
$L__BB0_4:
|
|
ld.const.v2.u32 {%r61, %r62}, [params];
|
|
add.s32 %r5, %r61, %r53;
|
|
add.s32 %r6, %r62, %r54;
|
|
setp.eq.f32 %p9, %f1092, 0f00000000;
|
|
setp.eq.f32 %p10, %f1093, 0f00000000;
|
|
and.pred %p11, %p9, %p10;
|
|
setp.eq.f32 %p12, %f1094, 0f00000000;
|
|
and.pred %p13, %p12, %p11;
|
|
@%p13 bra $L__BB0_128;
|
|
bra.uni $L__BB0_5;
|
|
|
|
$L__BB0_128:
|
|
ld.const.u32 %r51, [params+104];
|
|
and.b32 %r316, %r51, 1;
|
|
setp.eq.b32 %p145, %r316, 1;
|
|
mov.pred %p146, 0;
|
|
xor.pred %p147, %p145, %p146;
|
|
not.pred %p148, %p147;
|
|
@%p148 bra $L__BB0_130;
|
|
|
|
ld.const.u64 %rd90, [params+144];
|
|
cvta.to.global.u64 %rd91, %rd90;
|
|
ld.const.u32 %r317, [params+136];
|
|
mad.lo.s32 %r318, %r317, %r6, %r5;
|
|
mul.wide.u32 %rd92, %r318, 4;
|
|
add.s64 %rd93, %rd91, %rd92;
|
|
mov.u16 %rs95, 0;
|
|
st.global.v4.u8 [%rd93], {%rs95, %rs95, %rs95, %rs95};
|
|
|
|
$L__BB0_130:
|
|
and.b32 %r319, %r51, 4;
|
|
setp.eq.s32 %p149, %r319, 0;
|
|
@%p149 bra $L__BB0_132;
|
|
|
|
ld.const.u64 %rd94, [params+224];
|
|
cvta.to.global.u64 %rd95, %rd94;
|
|
ld.const.u32 %r320, [params+216];
|
|
mad.lo.s32 %r321, %r320, %r6, %r5;
|
|
mov.f32 %f1023, 0f00000000;
|
|
mul.wide.u32 %rd96, %r321, 8;
|
|
add.s64 %rd97, %rd95, %rd96;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f1023;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f1023;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs96, %f1023;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs99, 0;
|
|
st.global.v4.u16 [%rd97], {%rs96, %rs97, %rs98, %rs99};
|
|
|
|
$L__BB0_132:
|
|
and.b32 %r322, %r51, 16;
|
|
setp.eq.s32 %p150, %r322, 0;
|
|
@%p150 bra $L__BB0_134;
|
|
|
|
ld.const.u64 %rd98, [params+240];
|
|
cvta.to.global.u64 %rd99, %rd98;
|
|
ld.const.u32 %r323, [params+232];
|
|
mad.lo.s32 %r324, %r323, %r6, %r5;
|
|
mov.f32 %f1026, 0f00000000;
|
|
mul.wide.u32 %rd100, %r324, 8;
|
|
add.s64 %rd101, %rd99, %rd100;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs102, %f1026;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f1026;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f1026;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs103, 0;
|
|
st.global.v4.u16 [%rd101], {%rs100, %rs101, %rs102, %rs103};
|
|
|
|
$L__BB0_134:
|
|
ld.const.u32 %r52, [params+108];
|
|
setp.eq.s32 %p151, %r52, 0;
|
|
ld.const.u64 %rd102, [params+256];
|
|
cvta.to.global.u64 %rd103, %rd102;
|
|
ld.const.u32 %r325, [params+248];
|
|
mad.lo.s32 %r326, %r325, %r6, %r5;
|
|
mul.wide.u32 %rd104, %r326, 8;
|
|
add.s64 %rd22, %rd103, %rd104;
|
|
@%p151 bra $L__BB0_136;
|
|
|
|
ld.global.v4.u16 {%rs110, %rs111, %rs112, %rs113}, [%rd22];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1027, %rs110;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1028, %rs111;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1029, %rs112;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1030, %f1027, 0f00000000;
|
|
add.f32 %f1031, %f1028, 0f00000000;
|
|
add.f32 %f1032, %f1029, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f1032;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f1031;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f1030;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs114, 0;
|
|
st.global.v4.u16 [%rd22], {%rs107, %rs108, %rs109, %rs114};
|
|
bra.uni $L__BB0_137;
|
|
|
|
$L__BB0_5:
|
|
ld.const.u64 %rd30, [params+496];
|
|
cvta.to.global.u64 %rd31, %rd30;
|
|
ld.const.u32 %r65, [params+488];
|
|
mad.lo.s32 %r66, %r65, %r54, %r53;
|
|
mul.wide.u32 %rd32, %r66, 8;
|
|
add.s64 %rd33, %rd31, %rd32;
|
|
ld.global.v4.u16 {%rs19, %rs20, %rs21, %rs22}, [%rd33];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f242, %rs19;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f243, %rs20;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f244, %rs21;}
|
|
|
|
// end inline asm
|
|
ld.const.u64 %rd34, [params+432];
|
|
cvta.to.global.u64 %rd35, %rd34;
|
|
ld.const.u32 %r67, [params+424];
|
|
mad.lo.s32 %r68, %r67, %r54, %r53;
|
|
mul.wide.u32 %rd36, %r68, 12;
|
|
add.s64 %rd37, %rd35, %rd36;
|
|
ld.global.f32 %f261, [%rd37];
|
|
mul.f32 %f262, %f261, 0f3456BF95;
|
|
ld.global.f32 %f263, [%rd37+4];
|
|
mul.f32 %f264, %f263, 0f3456BF95;
|
|
ld.global.f32 %f265, [%rd37+8];
|
|
mul.f32 %f266, %f265, 0f3456BF95;
|
|
abs.f32 %f267, %f1092;
|
|
div.rn.f32 %f268, %f262, %f267;
|
|
abs.f32 %f269, %f1093;
|
|
div.rn.f32 %f270, %f264, %f269;
|
|
abs.f32 %f271, %f1094;
|
|
div.rn.f32 %f272, %f266, %f271;
|
|
abs.f32 %f273, %f268;
|
|
abs.f32 %f274, %f270;
|
|
abs.f32 %f275, %f272;
|
|
mov.f32 %f276, 0f38D1B717;
|
|
max.f32 %f277, %f273, %f276;
|
|
max.f32 %f278, %f274, %f276;
|
|
max.f32 %f279, %f275, %f276;
|
|
fma.rn.f32 %f13, %f1092, %f277, %f261;
|
|
fma.rn.f32 %f14, %f1093, %f278, %f263;
|
|
fma.rn.f32 %f15, %f1094, %f279, %f265;
|
|
setp.gt.f32 %p14, %f267, %f271;
|
|
neg.f32 %f280, %f1093;
|
|
selp.f32 %f281, %f280, 0f00000000, %p14;
|
|
mov.f32 %f1124, 0f00000000;
|
|
neg.f32 %f282, %f1094;
|
|
selp.f32 %f283, %f1092, %f282, %p14;
|
|
selp.f32 %f284, 0f00000000, %f1093, %p14;
|
|
mul.f32 %f285, %f283, %f283;
|
|
fma.rn.f32 %f286, %f281, %f281, %f285;
|
|
fma.rn.f32 %f287, %f284, %f284, %f286;
|
|
sqrt.rn.f32 %f288, %f287;
|
|
rcp.rn.f32 %f289, %f288;
|
|
mul.f32 %f16, %f281, %f289;
|
|
mul.f32 %f17, %f283, %f289;
|
|
mul.f32 %f18, %f284, %f289;
|
|
ld.const.u64 %rd38, [params+128];
|
|
cvta.to.global.u64 %rd39, %rd38;
|
|
ld.const.u32 %r69, [params+120];
|
|
mad.lo.s32 %r70, %r69, %r54, %r53;
|
|
mul.wide.u32 %rd40, %r70, 4;
|
|
add.s64 %rd3, %rd39, %rd40;
|
|
setp.lt.s32 %p15, %r3, 1;
|
|
mov.f32 %f1123, %f1124;
|
|
mov.f32 %f1122, %f1124;
|
|
mov.f32 %f1121, %f1124;
|
|
mov.f32 %f1120, %f1124;
|
|
mov.f32 %f1119, %f1124;
|
|
mov.f32 %f1118, %f1124;
|
|
mov.f32 %f1117, %f1124;
|
|
mov.f32 %f1116, %f1124;
|
|
mov.f32 %f1115, %f1124;
|
|
mov.f32 %f1114, %f1124;
|
|
mov.f32 %f1113, %f1124;
|
|
mov.f32 %f1112, %f1124;
|
|
mov.f32 %f1111, %f1124;
|
|
mov.f32 %f1110, %f1124;
|
|
@%p15 bra $L__BB0_34;
|
|
|
|
cvt.rn.f32.s32 %f305, %r4;
|
|
rcp.rn.f32 %f19, %f305;
|
|
ld.global.u32 %r356, [%rd3];
|
|
mul.f32 %f20, %f13, 0f3456BF95;
|
|
mul.f32 %f21, %f14, 0f3456BF95;
|
|
mul.f32 %f22, %f15, 0f3456BF95;
|
|
ld.const.u64 %rd4, [params+96];
|
|
mul.f32 %f306, %f1092, %f17;
|
|
mul.f32 %f307, %f1093, %f16;
|
|
sub.f32 %f23, %f307, %f306;
|
|
mul.f32 %f308, %f1094, %f16;
|
|
mul.f32 %f309, %f1092, %f18;
|
|
sub.f32 %f24, %f309, %f308;
|
|
mul.f32 %f310, %f1093, %f18;
|
|
mul.f32 %f311, %f1094, %f17;
|
|
sub.f32 %f25, %f311, %f310;
|
|
mov.u32 %r71, 0;
|
|
add.s64 %rd5, %rd1, 24;
|
|
mov.u64 %rd41, __cudart_i2opi_f;
|
|
abs.f32 %f376, %f21;
|
|
abs.f32 %f377, %f20;
|
|
max.f32 %f378, %f377, %f376;
|
|
abs.f32 %f379, %f22;
|
|
max.f32 %f380, %f378, %f379;
|
|
mov.u32 %r353, %r71;
|
|
|
|
$L__BB0_7:
|
|
mov.u32 %r355, %r71;
|
|
|
|
$L__BB0_8:
|
|
cvt.rn.f32.s32 %f1069, %r353;
|
|
mad.lo.s32 %r73, %r356, 1664525, 1013904223;
|
|
and.b32 %r74, %r73, 16777215;
|
|
cvt.rn.f32.u32 %f312, %r74;
|
|
fma.rn.f32 %f313, %f312, 0f33800000, %f1069;
|
|
mul.f32 %f314, %f19, %f313;
|
|
mad.lo.s32 %r356, %r73, 1664525, 1013904223;
|
|
and.b32 %r75, %r356, 16777215;
|
|
cvt.rn.f32.u32 %f315, %r75;
|
|
cvt.rn.f32.s32 %f316, %r355;
|
|
fma.rn.f32 %f317, %f315, 0f33800000, %f316;
|
|
mul.f32 %f318, %f19, %f317;
|
|
fma.rn.f32 %f319, %f19, %f313, %f314;
|
|
mov.f32 %f320, 0f3F800000;
|
|
sub.f32 %f57, %f320, %f319;
|
|
mul.f32 %f321, %f57, %f57;
|
|
sub.f32 %f322, %f320, %f321;
|
|
mov.f32 %f323, 0f00000000;
|
|
max.f32 %f324, %f323, %f322;
|
|
sqrt.rn.f32 %f58, %f324;
|
|
mul.f32 %f59, %f318, 0f40C90FDB;
|
|
mul.f32 %f325, %f59, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r360, %f325;
|
|
cvt.rn.f32.s32 %f326, %r360;
|
|
mov.f32 %f327, 0fBFC90FDA;
|
|
fma.rn.f32 %f328, %f326, %f327, %f59;
|
|
mov.f32 %f329, 0fB3A22168;
|
|
fma.rn.f32 %f330, %f326, %f329, %f328;
|
|
mov.f32 %f331, 0fA7C234C5;
|
|
fma.rn.f32 %f1128, %f326, %f331, %f330;
|
|
abs.f32 %f61, %f59;
|
|
setp.ltu.f32 %p16, %f61, 0f47CE4780;
|
|
mov.f32 %f1125, %f1128;
|
|
@%p16 bra $L__BB0_16;
|
|
|
|
setp.eq.f32 %p17, %f61, 0f7F800000;
|
|
@%p17 bra $L__BB0_15;
|
|
bra.uni $L__BB0_10;
|
|
|
|
$L__BB0_15:
|
|
mov.f32 %f334, 0f00000000;
|
|
mul.rn.f32 %f1125, %f59, %f334;
|
|
mov.u32 %r360, 0;
|
|
bra.uni $L__BB0_16;
|
|
|
|
$L__BB0_10:
|
|
mov.b32 %r14, %f59;
|
|
bfe.u32 %r77, %r14, 23, 8;
|
|
mov.u64 %rd119, 0;
|
|
mov.u32 %r357, 0;
|
|
mov.u64 %rd117, %rd1;
|
|
mov.u64 %rd118, %rd41;
|
|
|
|
$L__BB0_11:
|
|
.pragma "nounroll";
|
|
mov.b32 %r336, %f59;
|
|
shl.b32 %r335, %r336, 8;
|
|
or.b32 %r334, %r335, -2147483648;
|
|
ld.global.nc.u32 %r79, [%rd118];
|
|
mad.wide.u32 %rd43, %r79, %r334, %rd119;
|
|
shr.u64 %rd119, %rd43, 32;
|
|
st.local.u32 [%rd117], %rd43;
|
|
add.s64 %rd118, %rd118, 4;
|
|
add.s64 %rd117, %rd117, 4;
|
|
add.s32 %r357, %r357, 1;
|
|
setp.ne.s32 %p18, %r357, 6;
|
|
@%p18 bra $L__BB0_11;
|
|
|
|
add.s32 %r341, %r77, -128;
|
|
mov.b32 %r340, %f59;
|
|
bfe.u32 %r339, %r340, 23, 8;
|
|
add.s32 %r338, %r339, -128;
|
|
shr.u32 %r337, %r338, 5;
|
|
st.local.u32 [%rd5], %rd119;
|
|
mov.u32 %r81, 6;
|
|
sub.s32 %r82, %r81, %r337;
|
|
mul.wide.s32 %rd44, %r82, 4;
|
|
add.s64 %rd45, %rd1, %rd44;
|
|
ld.local.u32 %r358, [%rd45];
|
|
ld.local.u32 %r359, [%rd45+-4];
|
|
and.b32 %r23, %r338, 31;
|
|
setp.eq.s32 %p19, %r23, 0;
|
|
@%p19 bra $L__BB0_14;
|
|
|
|
mov.b32 %r348, %f59;
|
|
bfe.u32 %r347, %r348, 23, 8;
|
|
add.s32 %r346, %r347, -128;
|
|
shr.u32 %r345, %r346, 5;
|
|
mov.u32 %r344, 4;
|
|
sub.s32 %r343, %r344, %r345;
|
|
mov.u32 %r83, 32;
|
|
sub.s32 %r84, %r83, %r23;
|
|
shr.u32 %r85, %r359, %r84;
|
|
shl.b32 %r86, %r358, %r23;
|
|
add.s32 %r358, %r85, %r86;
|
|
mul.wide.s32 %rd46, %r343, 4;
|
|
add.s64 %rd47, %rd1, %rd46;
|
|
ld.local.u32 %r87, [%rd47];
|
|
shr.u32 %r88, %r87, %r84;
|
|
shl.b32 %r89, %r359, %r23;
|
|
add.s32 %r359, %r88, %r89;
|
|
|
|
$L__BB0_14:
|
|
mov.b32 %r342, %f59;
|
|
and.b32 %r90, %r342, -2147483648;
|
|
shr.u32 %r91, %r359, 30;
|
|
shl.b32 %r92, %r358, 2;
|
|
or.b32 %r93, %r91, %r92;
|
|
shr.u32 %r94, %r93, 31;
|
|
shr.u32 %r95, %r358, 30;
|
|
add.s32 %r96, %r94, %r95;
|
|
neg.s32 %r97, %r96;
|
|
setp.eq.s32 %p20, %r90, 0;
|
|
selp.b32 %r360, %r96, %r97, %p20;
|
|
setp.ne.s32 %p21, %r94, 0;
|
|
xor.b32 %r98, %r90, -2147483648;
|
|
selp.b32 %r99, %r98, %r90, %p21;
|
|
selp.b32 %r100, -1, 0, %p21;
|
|
xor.b32 %r101, %r93, %r100;
|
|
shl.b32 %r102, %r359, 2;
|
|
xor.b32 %r103, %r102, %r100;
|
|
cvt.u64.u32 %rd48, %r101;
|
|
cvt.u64.u32 %rd49, %r103;
|
|
bfi.b64 %rd50, %rd48, %rd49, 32, 32;
|
|
cvt.rn.f64.s64 %fd1, %rd50;
|
|
mul.f64 %fd2, %fd1, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f332, %fd2;
|
|
setp.eq.s32 %p22, %r99, 0;
|
|
neg.f32 %f333, %f332;
|
|
selp.f32 %f1125, %f332, %f333, %p22;
|
|
|
|
$L__BB0_16:
|
|
add.s32 %r30, %r360, 1;
|
|
and.b32 %r31, %r30, 1;
|
|
setp.eq.s32 %p23, %r31, 0;
|
|
selp.f32 %f65, %f1125, 0f3F800000, %p23;
|
|
mul.rn.f32 %f66, %f1125, %f1125;
|
|
mov.f32 %f1126, 0fB94D4153;
|
|
@%p23 bra $L__BB0_18;
|
|
|
|
mov.f32 %f336, 0fBAB607ED;
|
|
mov.f32 %f337, 0f37CBAC00;
|
|
fma.rn.f32 %f1126, %f337, %f66, %f336;
|
|
|
|
$L__BB0_18:
|
|
selp.f32 %f338, 0f3C0885E4, 0f3D2AAABB, %p23;
|
|
fma.rn.f32 %f339, %f1126, %f66, %f338;
|
|
selp.f32 %f340, 0fBE2AAAA8, 0fBEFFFFFF, %p23;
|
|
fma.rn.f32 %f341, %f339, %f66, %f340;
|
|
mov.f32 %f342, 0f00000000;
|
|
fma.rn.f32 %f343, %f66, %f65, %f342;
|
|
fma.rn.f32 %f1127, %f341, %f343, %f65;
|
|
and.b32 %r105, %r30, 2;
|
|
setp.eq.s32 %p25, %r105, 0;
|
|
@%p25 bra $L__BB0_20;
|
|
|
|
mov.f32 %f345, 0fBF800000;
|
|
fma.rn.f32 %f1127, %f1127, %f345, %f342;
|
|
|
|
$L__BB0_20:
|
|
mul.f32 %f1070, %f59, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r363, %f1070;
|
|
setp.ltu.f32 %p155, %f61, 0f47CE4780;
|
|
@%p155 bra $L__BB0_28;
|
|
|
|
setp.eq.f32 %p27, %f61, 0f7F800000;
|
|
@%p27 bra $L__BB0_27;
|
|
bra.uni $L__BB0_22;
|
|
|
|
$L__BB0_27:
|
|
mov.f32 %f348, 0f00000000;
|
|
mul.rn.f32 %f1128, %f59, %f348;
|
|
mov.u32 %r363, 0;
|
|
bra.uni $L__BB0_28;
|
|
|
|
$L__BB0_22:
|
|
mov.b32 %r32, %f59;
|
|
bfe.u32 %r106, %r32, 23, 8;
|
|
add.s32 %r33, %r106, -128;
|
|
shl.b32 %r107, %r32, 8;
|
|
or.b32 %r34, %r107, -2147483648;
|
|
shr.u32 %r35, %r33, 5;
|
|
mov.u64 %rd120, 0;
|
|
mov.u64 %rd121, %rd120;
|
|
|
|
$L__BB0_23:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd53, %rd120, 2;
|
|
mov.u64 %rd54, __cudart_i2opi_f;
|
|
add.s64 %rd55, %rd54, %rd53;
|
|
ld.global.nc.u32 %r108, [%rd55];
|
|
mad.wide.u32 %rd56, %r108, %r34, %rd121;
|
|
shr.u64 %rd121, %rd56, 32;
|
|
add.s64 %rd57, %rd1, %rd53;
|
|
st.local.u32 [%rd57], %rd56;
|
|
cvt.u32.u64 %r109, %rd120;
|
|
add.s32 %r110, %r109, 1;
|
|
cvt.s64.s32 %rd120, %r110;
|
|
setp.ne.s32 %p28, %r110, 6;
|
|
@%p28 bra $L__BB0_23;
|
|
|
|
st.local.u32 [%rd5], %rd121;
|
|
mov.u32 %r111, 4;
|
|
sub.s32 %r36, %r111, %r35;
|
|
mov.u32 %r112, 6;
|
|
sub.s32 %r113, %r112, %r35;
|
|
mul.wide.s32 %rd58, %r113, 4;
|
|
add.s64 %rd59, %rd1, %rd58;
|
|
ld.local.u32 %r361, [%rd59];
|
|
ld.local.u32 %r362, [%rd59+-4];
|
|
and.b32 %r39, %r33, 31;
|
|
setp.eq.s32 %p29, %r39, 0;
|
|
@%p29 bra $L__BB0_26;
|
|
|
|
mov.u32 %r114, 32;
|
|
sub.s32 %r115, %r114, %r39;
|
|
shr.u32 %r116, %r362, %r115;
|
|
shl.b32 %r117, %r361, %r39;
|
|
add.s32 %r361, %r116, %r117;
|
|
mul.wide.s32 %rd60, %r36, 4;
|
|
add.s64 %rd61, %rd1, %rd60;
|
|
ld.local.u32 %r118, [%rd61];
|
|
shr.u32 %r119, %r118, %r115;
|
|
shl.b32 %r120, %r362, %r39;
|
|
add.s32 %r362, %r119, %r120;
|
|
|
|
$L__BB0_26:
|
|
and.b32 %r121, %r32, -2147483648;
|
|
shr.u32 %r122, %r362, 30;
|
|
shl.b32 %r123, %r361, 2;
|
|
or.b32 %r124, %r122, %r123;
|
|
shr.u32 %r125, %r124, 31;
|
|
shr.u32 %r126, %r361, 30;
|
|
add.s32 %r127, %r125, %r126;
|
|
neg.s32 %r128, %r127;
|
|
setp.eq.s32 %p30, %r121, 0;
|
|
selp.b32 %r363, %r127, %r128, %p30;
|
|
setp.ne.s32 %p31, %r125, 0;
|
|
xor.b32 %r129, %r121, -2147483648;
|
|
selp.b32 %r130, %r129, %r121, %p31;
|
|
selp.b32 %r131, -1, 0, %p31;
|
|
xor.b32 %r132, %r124, %r131;
|
|
shl.b32 %r133, %r362, 2;
|
|
xor.b32 %r134, %r133, %r131;
|
|
cvt.u64.u32 %rd62, %r132;
|
|
cvt.u64.u32 %rd63, %r134;
|
|
bfi.b64 %rd64, %rd62, %rd63, 32, 32;
|
|
cvt.rn.f64.s64 %fd3, %rd64;
|
|
mul.f64 %fd4, %fd3, 0d3BF921FB54442D19;
|
|
cvt.rn.f32.f64 %f346, %fd4;
|
|
setp.eq.s32 %p32, %r130, 0;
|
|
neg.f32 %f347, %f346;
|
|
selp.f32 %f1128, %f346, %f347, %p32;
|
|
|
|
$L__BB0_28:
|
|
mul.f32 %f75, %f58, %f1127;
|
|
and.b32 %r46, %r363, 1;
|
|
setp.eq.s32 %p33, %r46, 0;
|
|
selp.f32 %f76, %f1128, 0f3F800000, %p33;
|
|
mul.rn.f32 %f77, %f1128, %f1128;
|
|
mov.f32 %f1129, 0fB94D4153;
|
|
@%p33 bra $L__BB0_30;
|
|
|
|
mov.f32 %f350, 0fBAB607ED;
|
|
mov.f32 %f351, 0f37CBAC00;
|
|
fma.rn.f32 %f1129, %f351, %f77, %f350;
|
|
|
|
$L__BB0_30:
|
|
selp.f32 %f352, 0f3C0885E4, 0f3D2AAABB, %p33;
|
|
fma.rn.f32 %f353, %f1129, %f77, %f352;
|
|
selp.f32 %f354, 0fBE2AAAA8, 0fBEFFFFFF, %p33;
|
|
fma.rn.f32 %f355, %f353, %f77, %f354;
|
|
mov.f32 %f356, 0f00000000;
|
|
fma.rn.f32 %f357, %f77, %f76, %f356;
|
|
fma.rn.f32 %f1130, %f355, %f357, %f76;
|
|
and.b32 %r136, %r363, 2;
|
|
setp.eq.s32 %p35, %r136, 0;
|
|
@%p35 bra $L__BB0_32;
|
|
|
|
mov.f32 %f359, 0fBF800000;
|
|
fma.rn.f32 %f1130, %f1130, %f359, %f356;
|
|
|
|
$L__BB0_32:
|
|
mul.f32 %f369, %f58, %f1130;
|
|
mul.f32 %f370, %f16, %f369;
|
|
mul.f32 %f371, %f17, %f369;
|
|
mul.f32 %f372, %f18, %f369;
|
|
fma.rn.f32 %f373, %f25, %f75, %f370;
|
|
fma.rn.f32 %f374, %f24, %f75, %f371;
|
|
fma.rn.f32 %f375, %f23, %f75, %f372;
|
|
fma.rn.f32 %f363, %f1092, %f57, %f373;
|
|
fma.rn.f32 %f364, %f1093, %f57, %f374;
|
|
fma.rn.f32 %f365, %f1094, %f57, %f375;
|
|
mov.f32 %f381, 0f38D1B717;
|
|
max.f32 %f366, %f380, %f381;
|
|
mov.f32 %f367, 0f6C4ECB8F;
|
|
mov.u32 %r170, 1;
|
|
mov.u32 %r173, 2;
|
|
mov.u32 %r175, 4;
|
|
mov.u32 %r178, 1065353216;
|
|
mov.u32 %r179, 2139095039;
|
|
mov.u32 %r207, 0;
|
|
// begin inline asm
|
|
call(%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,%r166,%r167,%r168),_optix_trace_typed_32,(%r207,%rd4,%f13,%f14,%f15,%f363,%f364,%f365,%f366,%f367,%f356,%r170,%r207,%r207,%r173,%r207,%r175,%r178,%r178,%r178,%r179,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207,%r207);
|
|
// end inline asm
|
|
mov.b32 %f382, %r137;
|
|
mov.b32 %f383, %r138;
|
|
mov.b32 %f384, %r139;
|
|
mov.b32 %f385, %r140;
|
|
setp.ltu.f32 %p36, %f385, 0f00000000;
|
|
selp.f32 %f386, %f384, 0f00000000, %p36;
|
|
selp.f32 %f387, %f383, 0f00000000, %p36;
|
|
selp.f32 %f388, %f382, 0f00000000, %p36;
|
|
mul.f32 %f389, %f1093, %f364;
|
|
fma.rn.f32 %f390, %f1092, %f363, %f389;
|
|
fma.rn.f32 %f391, %f1094, %f365, %f390;
|
|
fma.rn.f32 %f1118, %f363, %f388, %f1118;
|
|
fma.rn.f32 %f1117, %f363, %f387, %f1117;
|
|
fma.rn.f32 %f1116, %f363, %f386, %f1116;
|
|
fma.rn.f32 %f1121, %f364, %f388, %f1121;
|
|
fma.rn.f32 %f1120, %f364, %f387, %f1120;
|
|
fma.rn.f32 %f1119, %f364, %f386, %f1119;
|
|
fma.rn.f32 %f1124, %f365, %f388, %f1124;
|
|
fma.rn.f32 %f1123, %f365, %f387, %f1123;
|
|
fma.rn.f32 %f1122, %f365, %f386, %f1122;
|
|
add.f32 %f1115, %f1115, %f388;
|
|
add.f32 %f1114, %f1114, %f387;
|
|
add.f32 %f1113, %f1113, %f386;
|
|
cvt.sat.f32.f32 %f392, %f391;
|
|
fma.rn.f32 %f1112, %f392, %f388, %f1112;
|
|
fma.rn.f32 %f1111, %f392, %f387, %f1111;
|
|
fma.rn.f32 %f1110, %f392, %f386, %f1110;
|
|
add.s32 %r355, %r355, 1;
|
|
setp.lt.s32 %p37, %r355, %r4;
|
|
@%p37 bra $L__BB0_8;
|
|
|
|
add.s32 %r353, %r353, 1;
|
|
setp.lt.s32 %p38, %r353, %r4;
|
|
@%p38 bra $L__BB0_7;
|
|
|
|
$L__BB0_34:
|
|
mov.f32 %f1146, 0f00000000;
|
|
mul.lo.s32 %r208, %r4, %r4;
|
|
cvt.rn.f32.s32 %f396, %r208;
|
|
rcp.rn.f32 %f397, %f396;
|
|
mul.f32 %f398, %f397, %f1112;
|
|
mul.f32 %f399, %f397, %f1111;
|
|
mul.f32 %f400, %f397, %f1110;
|
|
mul.f32 %f113, %f397, %f1115;
|
|
mul.f32 %f114, %f397, %f1114;
|
|
mul.f32 %f115, %f397, %f1113;
|
|
mul.f32 %f116, %f397, %f1118;
|
|
mul.f32 %f117, %f397, %f1117;
|
|
mul.f32 %f118, %f397, %f1116;
|
|
mul.f32 %f119, %f397, %f1121;
|
|
mul.f32 %f120, %f397, %f1120;
|
|
mul.f32 %f121, %f397, %f1119;
|
|
mul.f32 %f122, %f397, %f1124;
|
|
mul.f32 %f123, %f397, %f1123;
|
|
mul.f32 %f124, %f397, %f1122;
|
|
fma.rn.f32 %f401, %f397, %f1112, %f398;
|
|
fma.rn.f32 %f402, %f397, %f1111, %f399;
|
|
fma.rn.f32 %f403, %f397, %f1110, %f400;
|
|
ld.const.u64 %rd66, [params+504];
|
|
cvt.rn.f32.u32 %f404, %r6;
|
|
cvt.rn.f32.u32 %f405, %r5;
|
|
tex.2d.v4.f32.f32 {%f406, %f407, %f408, %f409}, [%rd66, {%f405, %f404}];
|
|
mul.f32 %f125, %f401, %f406;
|
|
mul.f32 %f126, %f402, %f407;
|
|
mul.f32 %f127, %f403, %f408;
|
|
add.f32 %f128, %f242, %f125;
|
|
add.f32 %f129, %f243, %f126;
|
|
add.f32 %f130, %f244, %f127;
|
|
abs.f32 %f410, %f128;
|
|
setp.gtu.f32 %p39, %f410, 0f7F800000;
|
|
mov.f32 %f1147, %f1146;
|
|
mov.f32 %f1148, %f1146;
|
|
@%p39 bra $L__BB0_38;
|
|
|
|
mov.f32 %f1146, 0f00000000;
|
|
abs.f32 %f414, %f129;
|
|
setp.gtu.f32 %p40, %f414, 0f7F800000;
|
|
@%p40 bra $L__BB0_38;
|
|
|
|
mov.f32 %f1146, 0f00000000;
|
|
abs.f32 %f418, %f130;
|
|
setp.gtu.f32 %p41, %f418, 0f7F800000;
|
|
mov.f32 %f1147, %f1146;
|
|
mov.f32 %f1148, %f1146;
|
|
@%p41 bra $L__BB0_38;
|
|
|
|
mov.f32 %f1146, %f128;
|
|
mov.f32 %f1147, %f129;
|
|
mov.f32 %f1148, %f130;
|
|
|
|
$L__BB0_38:
|
|
ld.const.u32 %r49, [params+104];
|
|
and.b32 %r209, %r49, 1;
|
|
setp.eq.b32 %p42, %r209, 1;
|
|
mov.pred %p43, 0;
|
|
xor.pred %p44, %p42, %p43;
|
|
not.pred %p45, %p44;
|
|
@%p45 bra $L__BB0_112;
|
|
|
|
mov.f32 %f423, 0f3EE66666;
|
|
abs.f32 %f141, %f1146;
|
|
setp.lt.f32 %p46, %f141, 0f00800000;
|
|
mul.f32 %f425, %f141, 0f4B800000;
|
|
selp.f32 %f426, %f425, %f141, %p46;
|
|
selp.f32 %f427, 0fC3170000, 0fC2FE0000, %p46;
|
|
mov.b32 %r210, %f426;
|
|
and.b32 %r211, %r210, 8388607;
|
|
or.b32 %r212, %r211, 1065353216;
|
|
mov.b32 %f428, %r212;
|
|
shr.u32 %r213, %r210, 23;
|
|
cvt.rn.f32.u32 %f429, %r213;
|
|
add.f32 %f430, %f427, %f429;
|
|
setp.gt.f32 %p47, %f428, 0f3FB504F3;
|
|
mul.f32 %f431, %f428, 0f3F000000;
|
|
add.f32 %f432, %f430, 0f3F800000;
|
|
selp.f32 %f433, %f432, %f430, %p47;
|
|
selp.f32 %f434, %f431, %f428, %p47;
|
|
add.f32 %f435, %f434, 0fBF800000;
|
|
add.f32 %f436, %f434, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f437, %f436;
|
|
add.f32 %f438, %f435, %f435;
|
|
mul.f32 %f439, %f438, %f437;
|
|
mul.f32 %f440, %f439, %f439;
|
|
mov.f32 %f441, 0f3C4CAF63;
|
|
mov.f32 %f442, 0f3B18F0FE;
|
|
fma.rn.f32 %f443, %f442, %f440, %f441;
|
|
mov.f32 %f444, 0f3DAAAABD;
|
|
fma.rn.f32 %f445, %f443, %f440, %f444;
|
|
mul.rn.f32 %f446, %f445, %f440;
|
|
mul.rn.f32 %f447, %f446, %f439;
|
|
sub.f32 %f448, %f435, %f439;
|
|
add.f32 %f449, %f448, %f448;
|
|
neg.f32 %f450, %f439;
|
|
fma.rn.f32 %f451, %f450, %f435, %f449;
|
|
mul.rn.f32 %f452, %f437, %f451;
|
|
add.f32 %f453, %f447, %f439;
|
|
sub.f32 %f454, %f439, %f453;
|
|
add.f32 %f455, %f447, %f454;
|
|
add.f32 %f456, %f452, %f455;
|
|
add.f32 %f457, %f453, %f456;
|
|
sub.f32 %f458, %f453, %f457;
|
|
add.f32 %f459, %f456, %f458;
|
|
mov.f32 %f460, 0f3F317200;
|
|
mul.rn.f32 %f461, %f433, %f460;
|
|
mov.f32 %f462, 0f35BFBE8E;
|
|
mul.rn.f32 %f463, %f433, %f462;
|
|
add.f32 %f464, %f461, %f457;
|
|
sub.f32 %f465, %f461, %f464;
|
|
add.f32 %f466, %f457, %f465;
|
|
add.f32 %f467, %f459, %f466;
|
|
add.f32 %f468, %f463, %f467;
|
|
add.f32 %f469, %f464, %f468;
|
|
sub.f32 %f470, %f464, %f469;
|
|
add.f32 %f471, %f468, %f470;
|
|
mul.rn.f32 %f472, %f423, %f469;
|
|
neg.f32 %f473, %f472;
|
|
fma.rn.f32 %f474, %f423, %f469, %f473;
|
|
fma.rn.f32 %f475, %f423, %f471, %f474;
|
|
mov.f32 %f476, 0f00000000;
|
|
fma.rn.f32 %f477, %f476, %f469, %f475;
|
|
add.rn.f32 %f478, %f472, %f477;
|
|
neg.f32 %f479, %f478;
|
|
add.rn.f32 %f480, %f472, %f479;
|
|
add.rn.f32 %f481, %f480, %f477;
|
|
mov.b32 %r214, %f478;
|
|
setp.eq.s32 %p48, %r214, 1118925336;
|
|
add.s32 %r215, %r214, -1;
|
|
mov.b32 %f482, %r215;
|
|
add.f32 %f483, %f481, 0f37000000;
|
|
selp.f32 %f142, %f483, %f481, %p48;
|
|
selp.f32 %f484, %f482, %f478, %p48;
|
|
mov.f32 %f485, 0f3FB8AA3B;
|
|
mul.rn.f32 %f486, %f484, %f485;
|
|
cvt.rzi.f32.f32 %f487, %f486;
|
|
abs.f32 %f488, %f487;
|
|
setp.gt.f32 %p49, %f488, 0f42FC0000;
|
|
mov.b32 %r216, %f487;
|
|
and.b32 %r217, %r216, -2147483648;
|
|
or.b32 %r218, %r217, 1123811328;
|
|
mov.b32 %f489, %r218;
|
|
selp.f32 %f490, %f489, %f487, %p49;
|
|
mov.f32 %f491, 0fBF317218;
|
|
fma.rn.f32 %f492, %f490, %f491, %f484;
|
|
mov.f32 %f493, 0f3102E308;
|
|
fma.rn.f32 %f494, %f490, %f493, %f492;
|
|
mul.f32 %f495, %f494, 0f3FB8AA3B;
|
|
add.f32 %f496, %f490, 0f4B40007F;
|
|
mov.b32 %r219, %f496;
|
|
shl.b32 %r220, %r219, 23;
|
|
mov.b32 %f497, %r220;
|
|
ex2.approx.ftz.f32 %f498, %f495;
|
|
mul.f32 %f143, %f498, %f497;
|
|
setp.eq.f32 %p50, %f143, 0f7F800000;
|
|
mov.f32 %f1149, 0f7F800000;
|
|
@%p50 bra $L__BB0_41;
|
|
|
|
fma.rn.f32 %f1149, %f143, %f142, %f143;
|
|
|
|
$L__BB0_41:
|
|
mov.f32 %f1076, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f1075, %f1076;
|
|
add.f32 %f1074, %f1075, %f1075;
|
|
mov.f32 %f1073, 0f3EE66666;
|
|
sub.f32 %f1072, %f1073, %f1074;
|
|
abs.f32 %f1071, %f1072;
|
|
setp.lt.f32 %p51, %f1146, 0f00000000;
|
|
setp.eq.f32 %p52, %f1071, 0f3F800000;
|
|
and.pred %p1, %p51, %p52;
|
|
setp.eq.f32 %p53, %f1146, 0f00000000;
|
|
@%p53 bra $L__BB0_45;
|
|
bra.uni $L__BB0_42;
|
|
|
|
$L__BB0_45:
|
|
add.f32 %f503, %f1146, %f1146;
|
|
selp.f32 %f1151, %f503, 0f00000000, %p52;
|
|
bra.uni $L__BB0_46;
|
|
|
|
$L__BB0_136:
|
|
mov.f32 %f1035, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs117, %f1035;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f1035;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f1035;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs118, 0;
|
|
st.global.v4.u16 [%rd22], {%rs115, %rs116, %rs117, %rs118};
|
|
|
|
$L__BB0_137:
|
|
ld.const.u64 %rd105, [params+272];
|
|
cvta.to.global.u64 %rd106, %rd105;
|
|
ld.const.u32 %r327, [params+264];
|
|
mad.lo.s32 %r328, %r327, %r6, %r5;
|
|
mul.wide.u32 %rd107, %r328, 8;
|
|
add.s64 %rd23, %rd106, %rd107;
|
|
@%p151 bra $L__BB0_139;
|
|
|
|
ld.global.v4.u16 {%rs125, %rs126, %rs127, %rs128}, [%rd23];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1036, %rs125;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1037, %rs126;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1038, %rs127;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1039, %f1036, 0f00000000;
|
|
add.f32 %f1040, %f1037, 0f00000000;
|
|
add.f32 %f1041, %f1038, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f1041;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f1040;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f1039;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs129, 0;
|
|
st.global.v4.u16 [%rd23], {%rs122, %rs123, %rs124, %rs129};
|
|
bra.uni $L__BB0_140;
|
|
|
|
$L__BB0_139:
|
|
mov.f32 %f1044, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs132, %f1044;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs131, %f1044;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs130, %f1044;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs133, 0;
|
|
st.global.v4.u16 [%rd23], {%rs130, %rs131, %rs132, %rs133};
|
|
|
|
$L__BB0_140:
|
|
ld.const.u64 %rd108, [params+288];
|
|
cvta.to.global.u64 %rd109, %rd108;
|
|
ld.const.u32 %r329, [params+280];
|
|
mad.lo.s32 %r330, %r329, %r6, %r5;
|
|
mul.wide.u32 %rd110, %r330, 8;
|
|
add.s64 %rd24, %rd109, %rd110;
|
|
@%p151 bra $L__BB0_142;
|
|
|
|
ld.global.v4.u16 {%rs140, %rs141, %rs142, %rs143}, [%rd24];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1045, %rs140;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1046, %rs141;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1047, %rs142;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1048, %f1045, 0f00000000;
|
|
add.f32 %f1049, %f1046, 0f00000000;
|
|
add.f32 %f1050, %f1047, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs139, %f1050;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f1049;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f1048;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs144, 0;
|
|
st.global.v4.u16 [%rd24], {%rs137, %rs138, %rs139, %rs144};
|
|
bra.uni $L__BB0_143;
|
|
|
|
$L__BB0_142:
|
|
mov.f32 %f1053, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs147, %f1053;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f1053;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs145, %f1053;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs148, 0;
|
|
st.global.v4.u16 [%rd24], {%rs145, %rs146, %rs147, %rs148};
|
|
|
|
$L__BB0_143:
|
|
ld.const.u64 %rd111, [params+304];
|
|
cvta.to.global.u64 %rd112, %rd111;
|
|
ld.const.u32 %r331, [params+296];
|
|
mad.lo.s32 %r332, %r331, %r6, %r5;
|
|
mul.wide.u32 %rd113, %r332, 8;
|
|
add.s64 %rd25, %rd112, %rd113;
|
|
@%p151 bra $L__BB0_145;
|
|
|
|
ld.global.v4.u16 {%rs155, %rs156, %rs157, %rs158}, [%rd25];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1054, %rs155;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1055, %rs156;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1056, %rs157;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1057, %f1054, 0f00000000;
|
|
add.f32 %f1058, %f1055, 0f00000000;
|
|
add.f32 %f1059, %f1056, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs154, %f1059;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs153, %f1058;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f1057;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs159, 0;
|
|
st.global.v4.u16 [%rd25], {%rs152, %rs153, %rs154, %rs159};
|
|
bra.uni $L__BB0_146;
|
|
|
|
$L__BB0_145:
|
|
mov.f32 %f1062, 0f00000000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs162, %f1062;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs161, %f1062;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs160, %f1062;}
|
|
|
|
// end inline asm
|
|
mov.u16 %rs163, 0;
|
|
st.global.v4.u16 [%rd25], {%rs160, %rs161, %rs162, %rs163};
|
|
bra.uni $L__BB0_146;
|
|
|
|
$L__BB0_42:
|
|
mov.b32 %r221, %f1149;
|
|
xor.b32 %r222, %r221, -2147483648;
|
|
mov.b32 %f499, %r222;
|
|
selp.f32 %f1151, %f499, %f1149, %p1;
|
|
setp.geu.f32 %p54, %f1146, 0f00000000;
|
|
@%p54 bra $L__BB0_46;
|
|
|
|
mov.f32 %f500, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f501, %f500;
|
|
setp.eq.f32 %p55, %f501, 0f3EE66666;
|
|
@%p55 bra $L__BB0_46;
|
|
|
|
mov.f32 %f1151, 0f7FFFFFFF;
|
|
|
|
$L__BB0_46:
|
|
abs.f32 %f1077, %f1146;
|
|
add.f32 %f504, %f1077, 0f3EE66666;
|
|
mov.b32 %r223, %f504;
|
|
setp.lt.s32 %p57, %r223, 2139095040;
|
|
@%p57 bra $L__BB0_51;
|
|
|
|
abs.f32 %f1078, %f1146;
|
|
setp.gtu.f32 %p58, %f1078, 0f7F800000;
|
|
@%p58 bra $L__BB0_50;
|
|
bra.uni $L__BB0_48;
|
|
|
|
$L__BB0_50:
|
|
add.f32 %f1151, %f1146, 0f3EE66666;
|
|
bra.uni $L__BB0_51;
|
|
|
|
$L__BB0_48:
|
|
abs.f32 %f1079, %f1146;
|
|
setp.neu.f32 %p59, %f1079, 0f7F800000;
|
|
@%p59 bra $L__BB0_51;
|
|
|
|
selp.f32 %f1151, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
$L__BB0_51:
|
|
setp.eq.f32 %p60, %f1146, 0f3F800000;
|
|
selp.f32 %f152, 0f3F800000, %f1151, %p60;
|
|
abs.f32 %f153, %f1147;
|
|
setp.lt.f32 %p61, %f153, 0f00800000;
|
|
mul.f32 %f506, %f153, 0f4B800000;
|
|
selp.f32 %f507, %f506, %f153, %p61;
|
|
selp.f32 %f508, 0fC3170000, 0fC2FE0000, %p61;
|
|
mov.b32 %r224, %f507;
|
|
and.b32 %r225, %r224, 8388607;
|
|
or.b32 %r226, %r225, 1065353216;
|
|
mov.b32 %f509, %r226;
|
|
shr.u32 %r227, %r224, 23;
|
|
cvt.rn.f32.u32 %f510, %r227;
|
|
add.f32 %f511, %f508, %f510;
|
|
setp.gt.f32 %p62, %f509, 0f3FB504F3;
|
|
mul.f32 %f512, %f509, 0f3F000000;
|
|
add.f32 %f513, %f511, 0f3F800000;
|
|
selp.f32 %f514, %f513, %f511, %p62;
|
|
selp.f32 %f515, %f512, %f509, %p62;
|
|
add.f32 %f516, %f515, 0fBF800000;
|
|
add.f32 %f517, %f515, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f518, %f517;
|
|
add.f32 %f519, %f516, %f516;
|
|
mul.f32 %f520, %f519, %f518;
|
|
mul.f32 %f521, %f520, %f520;
|
|
mov.f32 %f522, 0f3C4CAF63;
|
|
mov.f32 %f523, 0f3B18F0FE;
|
|
fma.rn.f32 %f524, %f523, %f521, %f522;
|
|
mov.f32 %f525, 0f3DAAAABD;
|
|
fma.rn.f32 %f526, %f524, %f521, %f525;
|
|
mul.rn.f32 %f527, %f526, %f521;
|
|
mul.rn.f32 %f528, %f527, %f520;
|
|
sub.f32 %f529, %f516, %f520;
|
|
add.f32 %f530, %f529, %f529;
|
|
neg.f32 %f531, %f520;
|
|
fma.rn.f32 %f532, %f531, %f516, %f530;
|
|
mul.rn.f32 %f533, %f518, %f532;
|
|
add.f32 %f534, %f528, %f520;
|
|
sub.f32 %f535, %f520, %f534;
|
|
add.f32 %f536, %f528, %f535;
|
|
add.f32 %f537, %f533, %f536;
|
|
add.f32 %f538, %f534, %f537;
|
|
sub.f32 %f539, %f534, %f538;
|
|
add.f32 %f540, %f537, %f539;
|
|
mov.f32 %f541, 0f3F317200;
|
|
mul.rn.f32 %f542, %f514, %f541;
|
|
mov.f32 %f543, 0f35BFBE8E;
|
|
mul.rn.f32 %f544, %f514, %f543;
|
|
add.f32 %f545, %f542, %f538;
|
|
sub.f32 %f546, %f542, %f545;
|
|
add.f32 %f547, %f538, %f546;
|
|
add.f32 %f548, %f540, %f547;
|
|
add.f32 %f549, %f544, %f548;
|
|
add.f32 %f550, %f545, %f549;
|
|
sub.f32 %f551, %f545, %f550;
|
|
add.f32 %f552, %f549, %f551;
|
|
mov.f32 %f553, 0f3EE66666;
|
|
mul.rn.f32 %f554, %f553, %f550;
|
|
neg.f32 %f555, %f554;
|
|
fma.rn.f32 %f556, %f553, %f550, %f555;
|
|
fma.rn.f32 %f557, %f553, %f552, %f556;
|
|
mov.f32 %f558, 0f00000000;
|
|
fma.rn.f32 %f559, %f558, %f550, %f557;
|
|
add.rn.f32 %f560, %f554, %f559;
|
|
neg.f32 %f561, %f560;
|
|
add.rn.f32 %f562, %f554, %f561;
|
|
add.rn.f32 %f563, %f562, %f559;
|
|
mov.b32 %r228, %f560;
|
|
setp.eq.s32 %p63, %r228, 1118925336;
|
|
add.s32 %r229, %r228, -1;
|
|
mov.b32 %f564, %r229;
|
|
add.f32 %f565, %f563, 0f37000000;
|
|
selp.f32 %f154, %f565, %f563, %p63;
|
|
selp.f32 %f566, %f564, %f560, %p63;
|
|
mov.f32 %f567, 0f3FB8AA3B;
|
|
mul.rn.f32 %f568, %f566, %f567;
|
|
cvt.rzi.f32.f32 %f569, %f568;
|
|
abs.f32 %f570, %f569;
|
|
setp.gt.f32 %p64, %f570, 0f42FC0000;
|
|
mov.b32 %r230, %f569;
|
|
and.b32 %r231, %r230, -2147483648;
|
|
or.b32 %r232, %r231, 1123811328;
|
|
mov.b32 %f571, %r232;
|
|
selp.f32 %f572, %f571, %f569, %p64;
|
|
mov.f32 %f573, 0fBF317218;
|
|
fma.rn.f32 %f574, %f572, %f573, %f566;
|
|
mov.f32 %f575, 0f3102E308;
|
|
fma.rn.f32 %f576, %f572, %f575, %f574;
|
|
mul.f32 %f577, %f576, 0f3FB8AA3B;
|
|
add.f32 %f578, %f572, 0f4B40007F;
|
|
mov.b32 %r233, %f578;
|
|
shl.b32 %r234, %r233, 23;
|
|
mov.b32 %f579, %r234;
|
|
ex2.approx.ftz.f32 %f580, %f577;
|
|
mul.f32 %f155, %f580, %f579;
|
|
setp.eq.f32 %p65, %f155, 0f7F800000;
|
|
mov.f32 %f1152, 0f7F800000;
|
|
@%p65 bra $L__BB0_53;
|
|
|
|
fma.rn.f32 %f1152, %f155, %f154, %f155;
|
|
|
|
$L__BB0_53:
|
|
setp.lt.f32 %p66, %f1147, 0f00000000;
|
|
and.pred %p2, %p66, %p52;
|
|
setp.eq.f32 %p68, %f1147, 0f00000000;
|
|
@%p68 bra $L__BB0_57;
|
|
bra.uni $L__BB0_54;
|
|
|
|
$L__BB0_57:
|
|
add.f32 %f585, %f1147, %f1147;
|
|
selp.f32 %f1154, %f585, 0f00000000, %p52;
|
|
bra.uni $L__BB0_58;
|
|
|
|
$L__BB0_54:
|
|
mov.b32 %r235, %f1152;
|
|
xor.b32 %r236, %r235, -2147483648;
|
|
mov.b32 %f581, %r236;
|
|
selp.f32 %f1154, %f581, %f1152, %p2;
|
|
setp.geu.f32 %p69, %f1147, 0f00000000;
|
|
@%p69 bra $L__BB0_58;
|
|
|
|
mov.f32 %f582, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f583, %f582;
|
|
setp.eq.f32 %p70, %f583, 0f3EE66666;
|
|
@%p70 bra $L__BB0_58;
|
|
|
|
mov.f32 %f1154, 0f7FFFFFFF;
|
|
|
|
$L__BB0_58:
|
|
abs.f32 %f1080, %f1147;
|
|
add.f32 %f586, %f1080, 0f3EE66666;
|
|
mov.b32 %r237, %f586;
|
|
setp.lt.s32 %p72, %r237, 2139095040;
|
|
@%p72 bra $L__BB0_63;
|
|
|
|
abs.f32 %f1081, %f1147;
|
|
setp.gtu.f32 %p73, %f1081, 0f7F800000;
|
|
@%p73 bra $L__BB0_62;
|
|
bra.uni $L__BB0_60;
|
|
|
|
$L__BB0_62:
|
|
add.f32 %f1154, %f1147, 0f3EE66666;
|
|
bra.uni $L__BB0_63;
|
|
|
|
$L__BB0_60:
|
|
abs.f32 %f1082, %f1147;
|
|
setp.neu.f32 %p74, %f1082, 0f7F800000;
|
|
@%p74 bra $L__BB0_63;
|
|
|
|
selp.f32 %f1154, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
$L__BB0_63:
|
|
setp.eq.f32 %p75, %f1147, 0f3F800000;
|
|
selp.f32 %f164, 0f3F800000, %f1154, %p75;
|
|
abs.f32 %f165, %f1148;
|
|
setp.lt.f32 %p76, %f165, 0f00800000;
|
|
mul.f32 %f588, %f165, 0f4B800000;
|
|
selp.f32 %f589, %f588, %f165, %p76;
|
|
selp.f32 %f590, 0fC3170000, 0fC2FE0000, %p76;
|
|
mov.b32 %r238, %f589;
|
|
and.b32 %r239, %r238, 8388607;
|
|
or.b32 %r240, %r239, 1065353216;
|
|
mov.b32 %f591, %r240;
|
|
shr.u32 %r241, %r238, 23;
|
|
cvt.rn.f32.u32 %f592, %r241;
|
|
add.f32 %f593, %f590, %f592;
|
|
setp.gt.f32 %p77, %f591, 0f3FB504F3;
|
|
mul.f32 %f594, %f591, 0f3F000000;
|
|
add.f32 %f595, %f593, 0f3F800000;
|
|
selp.f32 %f596, %f595, %f593, %p77;
|
|
selp.f32 %f597, %f594, %f591, %p77;
|
|
add.f32 %f598, %f597, 0fBF800000;
|
|
add.f32 %f599, %f597, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f600, %f599;
|
|
add.f32 %f601, %f598, %f598;
|
|
mul.f32 %f602, %f601, %f600;
|
|
mul.f32 %f603, %f602, %f602;
|
|
mov.f32 %f604, 0f3C4CAF63;
|
|
mov.f32 %f605, 0f3B18F0FE;
|
|
fma.rn.f32 %f606, %f605, %f603, %f604;
|
|
mov.f32 %f607, 0f3DAAAABD;
|
|
fma.rn.f32 %f608, %f606, %f603, %f607;
|
|
mul.rn.f32 %f609, %f608, %f603;
|
|
mul.rn.f32 %f610, %f609, %f602;
|
|
sub.f32 %f611, %f598, %f602;
|
|
add.f32 %f612, %f611, %f611;
|
|
neg.f32 %f613, %f602;
|
|
fma.rn.f32 %f614, %f613, %f598, %f612;
|
|
mul.rn.f32 %f615, %f600, %f614;
|
|
add.f32 %f616, %f610, %f602;
|
|
sub.f32 %f617, %f602, %f616;
|
|
add.f32 %f618, %f610, %f617;
|
|
add.f32 %f619, %f615, %f618;
|
|
add.f32 %f620, %f616, %f619;
|
|
sub.f32 %f621, %f616, %f620;
|
|
add.f32 %f622, %f619, %f621;
|
|
mov.f32 %f623, 0f3F317200;
|
|
mul.rn.f32 %f624, %f596, %f623;
|
|
mov.f32 %f625, 0f35BFBE8E;
|
|
mul.rn.f32 %f626, %f596, %f625;
|
|
add.f32 %f627, %f624, %f620;
|
|
sub.f32 %f628, %f624, %f627;
|
|
add.f32 %f629, %f620, %f628;
|
|
add.f32 %f630, %f622, %f629;
|
|
add.f32 %f631, %f626, %f630;
|
|
add.f32 %f632, %f627, %f631;
|
|
sub.f32 %f633, %f627, %f632;
|
|
add.f32 %f634, %f631, %f633;
|
|
mov.f32 %f635, 0f3EE66666;
|
|
mul.rn.f32 %f636, %f635, %f632;
|
|
neg.f32 %f637, %f636;
|
|
fma.rn.f32 %f638, %f635, %f632, %f637;
|
|
fma.rn.f32 %f639, %f635, %f634, %f638;
|
|
mov.f32 %f640, 0f00000000;
|
|
fma.rn.f32 %f641, %f640, %f632, %f639;
|
|
add.rn.f32 %f642, %f636, %f641;
|
|
neg.f32 %f643, %f642;
|
|
add.rn.f32 %f644, %f636, %f643;
|
|
add.rn.f32 %f645, %f644, %f641;
|
|
mov.b32 %r242, %f642;
|
|
setp.eq.s32 %p78, %r242, 1118925336;
|
|
add.s32 %r243, %r242, -1;
|
|
mov.b32 %f646, %r243;
|
|
add.f32 %f647, %f645, 0f37000000;
|
|
selp.f32 %f166, %f647, %f645, %p78;
|
|
selp.f32 %f648, %f646, %f642, %p78;
|
|
mov.f32 %f649, 0f3FB8AA3B;
|
|
mul.rn.f32 %f650, %f648, %f649;
|
|
cvt.rzi.f32.f32 %f651, %f650;
|
|
abs.f32 %f652, %f651;
|
|
setp.gt.f32 %p79, %f652, 0f42FC0000;
|
|
mov.b32 %r244, %f651;
|
|
and.b32 %r245, %r244, -2147483648;
|
|
or.b32 %r246, %r245, 1123811328;
|
|
mov.b32 %f653, %r246;
|
|
selp.f32 %f654, %f653, %f651, %p79;
|
|
mov.f32 %f655, 0fBF317218;
|
|
fma.rn.f32 %f656, %f654, %f655, %f648;
|
|
mov.f32 %f657, 0f3102E308;
|
|
fma.rn.f32 %f658, %f654, %f657, %f656;
|
|
mul.f32 %f659, %f658, 0f3FB8AA3B;
|
|
add.f32 %f660, %f654, 0f4B40007F;
|
|
mov.b32 %r247, %f660;
|
|
shl.b32 %r248, %r247, 23;
|
|
mov.b32 %f661, %r248;
|
|
ex2.approx.ftz.f32 %f662, %f659;
|
|
mul.f32 %f167, %f662, %f661;
|
|
setp.eq.f32 %p80, %f167, 0f7F800000;
|
|
mov.f32 %f1155, 0f7F800000;
|
|
@%p80 bra $L__BB0_65;
|
|
|
|
fma.rn.f32 %f1155, %f167, %f166, %f167;
|
|
|
|
$L__BB0_65:
|
|
setp.lt.f32 %p81, %f1148, 0f00000000;
|
|
and.pred %p3, %p81, %p52;
|
|
setp.eq.f32 %p83, %f1148, 0f00000000;
|
|
@%p83 bra $L__BB0_69;
|
|
bra.uni $L__BB0_66;
|
|
|
|
$L__BB0_69:
|
|
add.f32 %f667, %f1148, %f1148;
|
|
selp.f32 %f1157, %f667, 0f00000000, %p52;
|
|
bra.uni $L__BB0_70;
|
|
|
|
$L__BB0_66:
|
|
mov.b32 %r249, %f1155;
|
|
xor.b32 %r250, %r249, -2147483648;
|
|
mov.b32 %f663, %r250;
|
|
selp.f32 %f1157, %f663, %f1155, %p3;
|
|
setp.geu.f32 %p84, %f1148, 0f00000000;
|
|
@%p84 bra $L__BB0_70;
|
|
|
|
mov.f32 %f664, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f665, %f664;
|
|
setp.eq.f32 %p85, %f665, 0f3EE66666;
|
|
@%p85 bra $L__BB0_70;
|
|
|
|
mov.f32 %f1157, 0f7FFFFFFF;
|
|
|
|
$L__BB0_70:
|
|
abs.f32 %f1083, %f1148;
|
|
add.f32 %f668, %f1083, 0f3EE66666;
|
|
mov.b32 %r251, %f668;
|
|
setp.lt.s32 %p87, %r251, 2139095040;
|
|
@%p87 bra $L__BB0_75;
|
|
|
|
abs.f32 %f1084, %f1148;
|
|
setp.gtu.f32 %p88, %f1084, 0f7F800000;
|
|
@%p88 bra $L__BB0_74;
|
|
bra.uni $L__BB0_72;
|
|
|
|
$L__BB0_74:
|
|
add.f32 %f1157, %f1148, 0f3EE66666;
|
|
bra.uni $L__BB0_75;
|
|
|
|
$L__BB0_72:
|
|
abs.f32 %f1085, %f1148;
|
|
setp.neu.f32 %p89, %f1085, 0f7F800000;
|
|
@%p89 bra $L__BB0_75;
|
|
|
|
selp.f32 %f1157, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
$L__BB0_75:
|
|
setp.eq.f32 %p90, %f1148, 0f3F800000;
|
|
mov.f32 %f670, 0f3F800000;
|
|
selp.f32 %f671, 0f3F800000, %f1157, %p90;
|
|
min.f32 %f672, %f152, %f670;
|
|
mov.f32 %f673, 0f00000000;
|
|
max.f32 %f176, %f673, %f672;
|
|
min.f32 %f674, %f164, %f670;
|
|
max.f32 %f177, %f673, %f674;
|
|
min.f32 %f675, %f671, %f670;
|
|
max.f32 %f178, %f673, %f675;
|
|
mov.f32 %f679, 0f3ED55555;
|
|
abs.f32 %f180, %f176;
|
|
setp.lt.f32 %p91, %f180, 0f00800000;
|
|
mul.f32 %f681, %f180, 0f4B800000;
|
|
selp.f32 %f682, %f681, %f180, %p91;
|
|
selp.f32 %f683, 0fC3170000, 0fC2FE0000, %p91;
|
|
mov.b32 %r254, %f682;
|
|
and.b32 %r255, %r254, 8388607;
|
|
or.b32 %r256, %r255, 1065353216;
|
|
mov.b32 %f684, %r256;
|
|
shr.u32 %r257, %r254, 23;
|
|
cvt.rn.f32.u32 %f685, %r257;
|
|
add.f32 %f686, %f683, %f685;
|
|
setp.gt.f32 %p92, %f684, 0f3FB504F3;
|
|
mul.f32 %f687, %f684, 0f3F000000;
|
|
add.f32 %f688, %f686, 0f3F800000;
|
|
selp.f32 %f689, %f688, %f686, %p92;
|
|
selp.f32 %f690, %f687, %f684, %p92;
|
|
add.f32 %f691, %f690, 0fBF800000;
|
|
add.f32 %f692, %f690, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f693, %f692;
|
|
add.f32 %f694, %f691, %f691;
|
|
mul.f32 %f695, %f694, %f693;
|
|
mul.f32 %f696, %f695, %f695;
|
|
mov.f32 %f697, 0f3C4CAF63;
|
|
mov.f32 %f698, 0f3B18F0FE;
|
|
fma.rn.f32 %f699, %f698, %f696, %f697;
|
|
mov.f32 %f700, 0f3DAAAABD;
|
|
fma.rn.f32 %f701, %f699, %f696, %f700;
|
|
mul.rn.f32 %f702, %f701, %f696;
|
|
mul.rn.f32 %f703, %f702, %f695;
|
|
sub.f32 %f704, %f691, %f695;
|
|
add.f32 %f705, %f704, %f704;
|
|
neg.f32 %f706, %f695;
|
|
fma.rn.f32 %f707, %f706, %f691, %f705;
|
|
mul.rn.f32 %f708, %f693, %f707;
|
|
add.f32 %f709, %f703, %f695;
|
|
sub.f32 %f710, %f695, %f709;
|
|
add.f32 %f711, %f703, %f710;
|
|
add.f32 %f712, %f708, %f711;
|
|
add.f32 %f713, %f709, %f712;
|
|
sub.f32 %f714, %f709, %f713;
|
|
add.f32 %f715, %f712, %f714;
|
|
mov.f32 %f716, 0f3F317200;
|
|
mul.rn.f32 %f717, %f689, %f716;
|
|
mov.f32 %f718, 0f35BFBE8E;
|
|
mul.rn.f32 %f719, %f689, %f718;
|
|
add.f32 %f720, %f717, %f713;
|
|
sub.f32 %f721, %f717, %f720;
|
|
add.f32 %f722, %f713, %f721;
|
|
add.f32 %f723, %f715, %f722;
|
|
add.f32 %f724, %f719, %f723;
|
|
add.f32 %f725, %f720, %f724;
|
|
sub.f32 %f726, %f720, %f725;
|
|
add.f32 %f727, %f724, %f726;
|
|
mul.rn.f32 %f728, %f679, %f725;
|
|
neg.f32 %f729, %f728;
|
|
fma.rn.f32 %f730, %f679, %f725, %f729;
|
|
fma.rn.f32 %f731, %f679, %f727, %f730;
|
|
fma.rn.f32 %f732, %f673, %f725, %f731;
|
|
add.rn.f32 %f733, %f728, %f732;
|
|
neg.f32 %f734, %f733;
|
|
add.rn.f32 %f735, %f728, %f734;
|
|
add.rn.f32 %f736, %f735, %f732;
|
|
mov.b32 %r258, %f733;
|
|
setp.eq.s32 %p93, %r258, 1118925336;
|
|
add.s32 %r259, %r258, -1;
|
|
mov.b32 %f737, %r259;
|
|
add.f32 %f738, %f736, 0f37000000;
|
|
selp.f32 %f181, %f738, %f736, %p93;
|
|
selp.f32 %f739, %f737, %f733, %p93;
|
|
mov.f32 %f740, 0f3FB8AA3B;
|
|
mul.rn.f32 %f741, %f739, %f740;
|
|
cvt.rzi.f32.f32 %f742, %f741;
|
|
abs.f32 %f743, %f742;
|
|
setp.gt.f32 %p94, %f743, 0f42FC0000;
|
|
mov.b32 %r260, %f742;
|
|
and.b32 %r261, %r260, -2147483648;
|
|
or.b32 %r262, %r261, 1123811328;
|
|
mov.b32 %f744, %r262;
|
|
selp.f32 %f745, %f744, %f742, %p94;
|
|
mov.f32 %f746, 0fBF317218;
|
|
fma.rn.f32 %f747, %f745, %f746, %f739;
|
|
mov.f32 %f748, 0f3102E308;
|
|
fma.rn.f32 %f749, %f745, %f748, %f747;
|
|
mul.f32 %f750, %f749, 0f3FB8AA3B;
|
|
add.f32 %f751, %f745, 0f4B40007F;
|
|
mov.b32 %r263, %f751;
|
|
shl.b32 %r264, %r263, 23;
|
|
mov.b32 %f752, %r264;
|
|
ex2.approx.ftz.f32 %f753, %f750;
|
|
mul.f32 %f182, %f753, %f752;
|
|
setp.eq.f32 %p95, %f182, 0f7F800000;
|
|
mov.f32 %f1158, 0f7F800000;
|
|
@%p95 bra $L__BB0_77;
|
|
|
|
fma.rn.f32 %f1158, %f182, %f181, %f182;
|
|
|
|
$L__BB0_77:
|
|
mov.f32 %f1091, 0f3E555555;
|
|
cvt.rzi.f32.f32 %f1090, %f1091;
|
|
add.f32 %f1089, %f1090, %f1090;
|
|
mov.f32 %f1088, 0f3ED55555;
|
|
sub.f32 %f1087, %f1088, %f1089;
|
|
abs.f32 %f1086, %f1087;
|
|
setp.lt.f32 %p96, %f176, 0f00000000;
|
|
setp.eq.f32 %p97, %f1086, 0f3F800000;
|
|
and.pred %p4, %p96, %p97;
|
|
setp.eq.f32 %p98, %f176, 0f00000000;
|
|
@%p98 bra $L__BB0_81;
|
|
bra.uni $L__BB0_78;
|
|
|
|
$L__BB0_81:
|
|
add.f32 %f758, %f176, %f176;
|
|
selp.f32 %f1160, %f758, 0f00000000, %p97;
|
|
bra.uni $L__BB0_82;
|
|
|
|
$L__BB0_78:
|
|
mov.b32 %r265, %f1158;
|
|
xor.b32 %r266, %r265, -2147483648;
|
|
mov.b32 %f754, %r266;
|
|
selp.f32 %f1160, %f754, %f1158, %p4;
|
|
setp.geu.f32 %p99, %f176, 0f00000000;
|
|
@%p99 bra $L__BB0_82;
|
|
|
|
mov.f32 %f755, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f756, %f755;
|
|
setp.eq.f32 %p100, %f756, 0f3ED55555;
|
|
@%p100 bra $L__BB0_82;
|
|
|
|
mov.f32 %f1160, 0f7FFFFFFF;
|
|
|
|
$L__BB0_82:
|
|
add.f32 %f759, %f180, 0f3ED55555;
|
|
mov.b32 %r267, %f759;
|
|
setp.lt.s32 %p102, %r267, 2139095040;
|
|
@%p102 bra $L__BB0_87;
|
|
|
|
setp.gtu.f32 %p103, %f180, 0f7F800000;
|
|
@%p103 bra $L__BB0_86;
|
|
bra.uni $L__BB0_84;
|
|
|
|
$L__BB0_86:
|
|
add.f32 %f1160, %f176, 0f3ED55555;
|
|
bra.uni $L__BB0_87;
|
|
|
|
$L__BB0_84:
|
|
setp.neu.f32 %p104, %f180, 0f7F800000;
|
|
@%p104 bra $L__BB0_87;
|
|
|
|
selp.f32 %f1160, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
$L__BB0_87:
|
|
abs.f32 %f191, %f177;
|
|
setp.lt.f32 %p105, %f191, 0f00800000;
|
|
mul.f32 %f761, %f191, 0f4B800000;
|
|
selp.f32 %f762, %f761, %f191, %p105;
|
|
selp.f32 %f763, 0fC3170000, 0fC2FE0000, %p105;
|
|
mov.b32 %r268, %f762;
|
|
and.b32 %r269, %r268, 8388607;
|
|
or.b32 %r270, %r269, 1065353216;
|
|
mov.b32 %f764, %r270;
|
|
shr.u32 %r271, %r268, 23;
|
|
cvt.rn.f32.u32 %f765, %r271;
|
|
add.f32 %f766, %f763, %f765;
|
|
setp.gt.f32 %p106, %f764, 0f3FB504F3;
|
|
mul.f32 %f767, %f764, 0f3F000000;
|
|
add.f32 %f768, %f766, 0f3F800000;
|
|
selp.f32 %f769, %f768, %f766, %p106;
|
|
selp.f32 %f770, %f767, %f764, %p106;
|
|
add.f32 %f771, %f770, 0fBF800000;
|
|
add.f32 %f772, %f770, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f773, %f772;
|
|
add.f32 %f774, %f771, %f771;
|
|
mul.f32 %f775, %f774, %f773;
|
|
mul.f32 %f776, %f775, %f775;
|
|
mov.f32 %f777, 0f3C4CAF63;
|
|
mov.f32 %f778, 0f3B18F0FE;
|
|
fma.rn.f32 %f779, %f778, %f776, %f777;
|
|
mov.f32 %f780, 0f3DAAAABD;
|
|
fma.rn.f32 %f781, %f779, %f776, %f780;
|
|
mul.rn.f32 %f782, %f781, %f776;
|
|
mul.rn.f32 %f783, %f782, %f775;
|
|
sub.f32 %f784, %f771, %f775;
|
|
add.f32 %f785, %f784, %f784;
|
|
neg.f32 %f786, %f775;
|
|
fma.rn.f32 %f787, %f786, %f771, %f785;
|
|
mul.rn.f32 %f788, %f773, %f787;
|
|
add.f32 %f789, %f783, %f775;
|
|
sub.f32 %f790, %f775, %f789;
|
|
add.f32 %f791, %f783, %f790;
|
|
add.f32 %f792, %f788, %f791;
|
|
add.f32 %f793, %f789, %f792;
|
|
sub.f32 %f794, %f789, %f793;
|
|
add.f32 %f795, %f792, %f794;
|
|
mov.f32 %f796, 0f3F317200;
|
|
mul.rn.f32 %f797, %f769, %f796;
|
|
mov.f32 %f798, 0f35BFBE8E;
|
|
mul.rn.f32 %f799, %f769, %f798;
|
|
add.f32 %f800, %f797, %f793;
|
|
sub.f32 %f801, %f797, %f800;
|
|
add.f32 %f802, %f793, %f801;
|
|
add.f32 %f803, %f795, %f802;
|
|
add.f32 %f804, %f799, %f803;
|
|
add.f32 %f805, %f800, %f804;
|
|
sub.f32 %f806, %f800, %f805;
|
|
add.f32 %f807, %f804, %f806;
|
|
mov.f32 %f808, 0f3ED55555;
|
|
mul.rn.f32 %f809, %f808, %f805;
|
|
neg.f32 %f810, %f809;
|
|
fma.rn.f32 %f811, %f808, %f805, %f810;
|
|
fma.rn.f32 %f812, %f808, %f807, %f811;
|
|
mov.f32 %f813, 0f00000000;
|
|
fma.rn.f32 %f814, %f813, %f805, %f812;
|
|
add.rn.f32 %f815, %f809, %f814;
|
|
neg.f32 %f816, %f815;
|
|
add.rn.f32 %f817, %f809, %f816;
|
|
add.rn.f32 %f818, %f817, %f814;
|
|
mov.b32 %r272, %f815;
|
|
setp.eq.s32 %p107, %r272, 1118925336;
|
|
add.s32 %r273, %r272, -1;
|
|
mov.b32 %f819, %r273;
|
|
add.f32 %f820, %f818, 0f37000000;
|
|
selp.f32 %f192, %f820, %f818, %p107;
|
|
selp.f32 %f821, %f819, %f815, %p107;
|
|
mov.f32 %f822, 0f3FB8AA3B;
|
|
mul.rn.f32 %f823, %f821, %f822;
|
|
cvt.rzi.f32.f32 %f824, %f823;
|
|
abs.f32 %f825, %f824;
|
|
setp.gt.f32 %p108, %f825, 0f42FC0000;
|
|
mov.b32 %r274, %f824;
|
|
and.b32 %r275, %r274, -2147483648;
|
|
or.b32 %r276, %r275, 1123811328;
|
|
mov.b32 %f826, %r276;
|
|
selp.f32 %f827, %f826, %f824, %p108;
|
|
mov.f32 %f828, 0fBF317218;
|
|
fma.rn.f32 %f829, %f827, %f828, %f821;
|
|
mov.f32 %f830, 0f3102E308;
|
|
fma.rn.f32 %f831, %f827, %f830, %f829;
|
|
mul.f32 %f832, %f831, 0f3FB8AA3B;
|
|
add.f32 %f833, %f827, 0f4B40007F;
|
|
mov.b32 %r277, %f833;
|
|
shl.b32 %r278, %r277, 23;
|
|
mov.b32 %f834, %r278;
|
|
ex2.approx.ftz.f32 %f835, %f832;
|
|
mul.f32 %f193, %f835, %f834;
|
|
setp.eq.f32 %p109, %f193, 0f7F800000;
|
|
mov.f32 %f1161, 0f7F800000;
|
|
@%p109 bra $L__BB0_89;
|
|
|
|
fma.rn.f32 %f1161, %f193, %f192, %f193;
|
|
|
|
$L__BB0_89:
|
|
setp.lt.f32 %p110, %f177, 0f00000000;
|
|
and.pred %p5, %p110, %p97;
|
|
setp.eq.f32 %p112, %f177, 0f00000000;
|
|
@%p112 bra $L__BB0_93;
|
|
bra.uni $L__BB0_90;
|
|
|
|
$L__BB0_93:
|
|
add.f32 %f840, %f177, %f177;
|
|
selp.f32 %f1163, %f840, 0f00000000, %p97;
|
|
bra.uni $L__BB0_94;
|
|
|
|
$L__BB0_90:
|
|
mov.b32 %r279, %f1161;
|
|
xor.b32 %r280, %r279, -2147483648;
|
|
mov.b32 %f836, %r280;
|
|
selp.f32 %f1163, %f836, %f1161, %p5;
|
|
setp.geu.f32 %p113, %f177, 0f00000000;
|
|
@%p113 bra $L__BB0_94;
|
|
|
|
mov.f32 %f837, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f838, %f837;
|
|
setp.eq.f32 %p114, %f838, 0f3ED55555;
|
|
@%p114 bra $L__BB0_94;
|
|
|
|
mov.f32 %f1163, 0f7FFFFFFF;
|
|
|
|
$L__BB0_94:
|
|
add.f32 %f841, %f191, 0f3ED55555;
|
|
mov.b32 %r281, %f841;
|
|
setp.lt.s32 %p116, %r281, 2139095040;
|
|
@%p116 bra $L__BB0_99;
|
|
|
|
setp.gtu.f32 %p117, %f191, 0f7F800000;
|
|
@%p117 bra $L__BB0_98;
|
|
bra.uni $L__BB0_96;
|
|
|
|
$L__BB0_98:
|
|
add.f32 %f1163, %f177, 0f3ED55555;
|
|
bra.uni $L__BB0_99;
|
|
|
|
$L__BB0_96:
|
|
setp.neu.f32 %p118, %f191, 0f7F800000;
|
|
@%p118 bra $L__BB0_99;
|
|
|
|
selp.f32 %f1163, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
$L__BB0_99:
|
|
abs.f32 %f202, %f178;
|
|
setp.lt.f32 %p119, %f202, 0f00800000;
|
|
mul.f32 %f843, %f202, 0f4B800000;
|
|
selp.f32 %f844, %f843, %f202, %p119;
|
|
selp.f32 %f845, 0fC3170000, 0fC2FE0000, %p119;
|
|
mov.b32 %r282, %f844;
|
|
and.b32 %r283, %r282, 8388607;
|
|
or.b32 %r284, %r283, 1065353216;
|
|
mov.b32 %f846, %r284;
|
|
shr.u32 %r285, %r282, 23;
|
|
cvt.rn.f32.u32 %f847, %r285;
|
|
add.f32 %f848, %f845, %f847;
|
|
setp.gt.f32 %p120, %f846, 0f3FB504F3;
|
|
mul.f32 %f849, %f846, 0f3F000000;
|
|
add.f32 %f850, %f848, 0f3F800000;
|
|
selp.f32 %f851, %f850, %f848, %p120;
|
|
selp.f32 %f852, %f849, %f846, %p120;
|
|
add.f32 %f853, %f852, 0fBF800000;
|
|
add.f32 %f854, %f852, 0f3F800000;
|
|
rcp.approx.ftz.f32 %f855, %f854;
|
|
add.f32 %f856, %f853, %f853;
|
|
mul.f32 %f857, %f856, %f855;
|
|
mul.f32 %f858, %f857, %f857;
|
|
mov.f32 %f859, 0f3C4CAF63;
|
|
mov.f32 %f860, 0f3B18F0FE;
|
|
fma.rn.f32 %f861, %f860, %f858, %f859;
|
|
mov.f32 %f862, 0f3DAAAABD;
|
|
fma.rn.f32 %f863, %f861, %f858, %f862;
|
|
mul.rn.f32 %f864, %f863, %f858;
|
|
mul.rn.f32 %f865, %f864, %f857;
|
|
sub.f32 %f866, %f853, %f857;
|
|
add.f32 %f867, %f866, %f866;
|
|
neg.f32 %f868, %f857;
|
|
fma.rn.f32 %f869, %f868, %f853, %f867;
|
|
mul.rn.f32 %f870, %f855, %f869;
|
|
add.f32 %f871, %f865, %f857;
|
|
sub.f32 %f872, %f857, %f871;
|
|
add.f32 %f873, %f865, %f872;
|
|
add.f32 %f874, %f870, %f873;
|
|
add.f32 %f875, %f871, %f874;
|
|
sub.f32 %f876, %f871, %f875;
|
|
add.f32 %f877, %f874, %f876;
|
|
mov.f32 %f878, 0f3F317200;
|
|
mul.rn.f32 %f879, %f851, %f878;
|
|
mov.f32 %f880, 0f35BFBE8E;
|
|
mul.rn.f32 %f881, %f851, %f880;
|
|
add.f32 %f882, %f879, %f875;
|
|
sub.f32 %f883, %f879, %f882;
|
|
add.f32 %f884, %f875, %f883;
|
|
add.f32 %f885, %f877, %f884;
|
|
add.f32 %f886, %f881, %f885;
|
|
add.f32 %f887, %f882, %f886;
|
|
sub.f32 %f888, %f882, %f887;
|
|
add.f32 %f889, %f886, %f888;
|
|
mov.f32 %f890, 0f3ED55555;
|
|
mul.rn.f32 %f891, %f890, %f887;
|
|
neg.f32 %f892, %f891;
|
|
fma.rn.f32 %f893, %f890, %f887, %f892;
|
|
fma.rn.f32 %f894, %f890, %f889, %f893;
|
|
mov.f32 %f895, 0f00000000;
|
|
fma.rn.f32 %f896, %f895, %f887, %f894;
|
|
add.rn.f32 %f897, %f891, %f896;
|
|
neg.f32 %f898, %f897;
|
|
add.rn.f32 %f899, %f891, %f898;
|
|
add.rn.f32 %f900, %f899, %f896;
|
|
mov.b32 %r286, %f897;
|
|
setp.eq.s32 %p121, %r286, 1118925336;
|
|
add.s32 %r287, %r286, -1;
|
|
mov.b32 %f901, %r287;
|
|
add.f32 %f902, %f900, 0f37000000;
|
|
selp.f32 %f203, %f902, %f900, %p121;
|
|
selp.f32 %f903, %f901, %f897, %p121;
|
|
mov.f32 %f904, 0f3FB8AA3B;
|
|
mul.rn.f32 %f905, %f903, %f904;
|
|
cvt.rzi.f32.f32 %f906, %f905;
|
|
abs.f32 %f907, %f906;
|
|
setp.gt.f32 %p122, %f907, 0f42FC0000;
|
|
mov.b32 %r288, %f906;
|
|
and.b32 %r289, %r288, -2147483648;
|
|
or.b32 %r290, %r289, 1123811328;
|
|
mov.b32 %f908, %r290;
|
|
selp.f32 %f909, %f908, %f906, %p122;
|
|
mov.f32 %f910, 0fBF317218;
|
|
fma.rn.f32 %f911, %f909, %f910, %f903;
|
|
mov.f32 %f912, 0f3102E308;
|
|
fma.rn.f32 %f913, %f909, %f912, %f911;
|
|
mul.f32 %f914, %f913, 0f3FB8AA3B;
|
|
add.f32 %f915, %f909, 0f4B40007F;
|
|
mov.b32 %r291, %f915;
|
|
shl.b32 %r292, %r291, 23;
|
|
mov.b32 %f916, %r292;
|
|
ex2.approx.ftz.f32 %f917, %f914;
|
|
mul.f32 %f204, %f917, %f916;
|
|
setp.eq.f32 %p123, %f204, 0f7F800000;
|
|
mov.f32 %f1164, 0f7F800000;
|
|
@%p123 bra $L__BB0_101;
|
|
|
|
fma.rn.f32 %f1164, %f204, %f203, %f204;
|
|
|
|
$L__BB0_101:
|
|
setp.lt.f32 %p124, %f178, 0f00000000;
|
|
and.pred %p6, %p124, %p97;
|
|
setp.eq.f32 %p126, %f178, 0f00000000;
|
|
@%p126 bra $L__BB0_105;
|
|
bra.uni $L__BB0_102;
|
|
|
|
$L__BB0_105:
|
|
add.f32 %f922, %f178, %f178;
|
|
selp.f32 %f1166, %f922, 0f00000000, %p97;
|
|
bra.uni $L__BB0_106;
|
|
|
|
$L__BB0_102:
|
|
mov.b32 %r293, %f1164;
|
|
xor.b32 %r294, %r293, -2147483648;
|
|
mov.b32 %f918, %r294;
|
|
selp.f32 %f1166, %f918, %f1164, %p6;
|
|
setp.geu.f32 %p127, %f178, 0f00000000;
|
|
@%p127 bra $L__BB0_106;
|
|
|
|
mov.f32 %f919, 0f3ED55555;
|
|
cvt.rzi.f32.f32 %f920, %f919;
|
|
setp.eq.f32 %p128, %f920, 0f3ED55555;
|
|
@%p128 bra $L__BB0_106;
|
|
|
|
mov.f32 %f1166, 0f7FFFFFFF;
|
|
|
|
$L__BB0_106:
|
|
add.f32 %f923, %f202, 0f3ED55555;
|
|
mov.b32 %r295, %f923;
|
|
setp.lt.s32 %p130, %r295, 2139095040;
|
|
@%p130 bra $L__BB0_111;
|
|
|
|
setp.gtu.f32 %p131, %f202, 0f7F800000;
|
|
@%p131 bra $L__BB0_110;
|
|
bra.uni $L__BB0_108;
|
|
|
|
$L__BB0_110:
|
|
add.f32 %f1166, %f178, 0f3ED55555;
|
|
bra.uni $L__BB0_111;
|
|
|
|
$L__BB0_108:
|
|
setp.neu.f32 %p132, %f202, 0f7F800000;
|
|
@%p132 bra $L__BB0_111;
|
|
|
|
selp.f32 %f1166, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
$L__BB0_111:
|
|
ld.const.u64 %rd116, [params+144];
|
|
cvta.to.global.u64 %rd115, %rd116;
|
|
ld.const.u32 %r352, [params+136];
|
|
mad.lo.s32 %r351, %r352, %r6, %r5;
|
|
cvt.u64.u32 %rd114, %r351;
|
|
fma.rn.f32 %f924, %f1160, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p133, %f176, 0f3F800000;
|
|
mov.f32 %f925, 0f3F800000;
|
|
selp.f32 %f926, 0f3F7FFFFF, %f924, %p133;
|
|
mul.f32 %f927, %f176, 0f414EB852;
|
|
setp.lt.f32 %p134, %f176, 0f3B4D2E1C;
|
|
selp.f32 %f928, %f927, %f926, %p134;
|
|
fma.rn.f32 %f929, %f1163, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p135, %f177, 0f3F800000;
|
|
selp.f32 %f930, 0f3F7FFFFF, %f929, %p135;
|
|
mul.f32 %f931, %f177, 0f414EB852;
|
|
setp.lt.f32 %p136, %f177, 0f3B4D2E1C;
|
|
selp.f32 %f932, %f931, %f930, %p136;
|
|
fma.rn.f32 %f933, %f1166, 0f3F870A3D, 0fBD6147AE;
|
|
setp.eq.f32 %p137, %f178, 0f3F800000;
|
|
selp.f32 %f934, 0f3F7FFFFF, %f933, %p137;
|
|
mul.f32 %f935, %f178, 0f414EB852;
|
|
setp.lt.f32 %p138, %f178, 0f3B4D2E1C;
|
|
selp.f32 %f936, %f935, %f934, %p138;
|
|
min.f32 %f937, %f928, %f925;
|
|
mov.f32 %f938, 0f00000000;
|
|
max.f32 %f939, %f938, %f937;
|
|
mul.f32 %f940, %f939, 0f43800000;
|
|
cvt.rzi.u32.f32 %r296, %f940;
|
|
min.u32 %r297, %r296, 255;
|
|
min.f32 %f941, %f932, %f925;
|
|
max.f32 %f942, %f938, %f941;
|
|
mul.f32 %f943, %f942, 0f43800000;
|
|
cvt.rzi.u32.f32 %r298, %f943;
|
|
min.u32 %r299, %r298, 255;
|
|
min.f32 %f944, %f936, %f925;
|
|
max.f32 %f945, %f938, %f944;
|
|
mul.f32 %f946, %f945, 0f43800000;
|
|
cvt.rzi.u32.f32 %r300, %f946;
|
|
min.u32 %r301, %r300, 255;
|
|
shl.b64 %rd68, %rd114, 2;
|
|
add.s64 %rd69, %rd115, %rd68;
|
|
cvt.u16.u32 %rs23, %r301;
|
|
cvt.u16.u32 %rs24, %r299;
|
|
cvt.u16.u32 %rs25, %r297;
|
|
mov.u16 %rs26, 255;
|
|
st.global.v4.u8 [%rd69], {%rs25, %rs24, %rs23, %rs26};
|
|
|
|
$L__BB0_112:
|
|
ld.const.u32 %r349, [params+104];
|
|
and.b32 %r302, %r349, 4;
|
|
setp.eq.s32 %p139, %r302, 0;
|
|
@%p139 bra $L__BB0_114;
|
|
|
|
ld.const.u64 %rd70, [params+224];
|
|
cvta.to.global.u64 %rd71, %rd70;
|
|
ld.const.u32 %r303, [params+216];
|
|
mad.lo.s32 %r304, %r303, %r6, %r5;
|
|
mov.f32 %f950, 0f3F800000;
|
|
mul.wide.u32 %rd72, %r304, 8;
|
|
add.s64 %rd73, %rd71, %rd72;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f950;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f1148;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f1147;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1146;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd73], {%rs27, %rs28, %rs29, %rs30};
|
|
|
|
$L__BB0_114:
|
|
ld.const.u32 %r350, [params+104];
|
|
and.b32 %r305, %r350, 16;
|
|
setp.eq.s32 %p140, %r305, 0;
|
|
@%p140 bra $L__BB0_116;
|
|
|
|
ld.const.u64 %rd74, [params+240];
|
|
cvta.to.global.u64 %rd75, %rd74;
|
|
ld.const.u32 %r306, [params+232];
|
|
mad.lo.s32 %r307, %r306, %r6, %r5;
|
|
mov.f32 %f954, 0f3F800000;
|
|
mul.wide.u32 %rd76, %r307, 8;
|
|
add.s64 %rd77, %rd75, %rd76;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f954;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f127;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f126;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f125;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd77], {%rs31, %rs32, %rs33, %rs34};
|
|
|
|
$L__BB0_116:
|
|
mul.f32 %f213, %f113, 0f3F000000;
|
|
mul.f32 %f214, %f114, 0f3F000000;
|
|
mul.f32 %f215, %f115, 0f3F000000;
|
|
ld.const.u32 %r50, [params+108];
|
|
setp.eq.s32 %p141, %r50, 0;
|
|
ld.const.u64 %rd78, [params+256];
|
|
cvta.to.global.u64 %rd79, %rd78;
|
|
ld.const.u32 %r308, [params+248];
|
|
mad.lo.s32 %r309, %r308, %r6, %r5;
|
|
mul.wide.u32 %rd80, %r309, 8;
|
|
add.s64 %rd18, %rd79, %rd80;
|
|
@%p141 bra $L__BB0_118;
|
|
|
|
ld.global.v4.u16 {%rs42, %rs43, %rs44, %rs45}, [%rd18];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f955, %rs42;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f956, %rs43;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f957, %rs44;}
|
|
|
|
// end inline asm
|
|
add.f32 %f958, %f213, %f955;
|
|
add.f32 %f959, %f214, %f956;
|
|
add.f32 %f960, %f215, %f957;
|
|
mov.f32 %f961, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f960;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f959;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f958;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f961;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd18], {%rs38, %rs39, %rs40, %rs41};
|
|
bra.uni $L__BB0_119;
|
|
|
|
$L__BB0_118:
|
|
mov.f32 %f965, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f965;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f215;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f214;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f213;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd18], {%rs46, %rs47, %rs48, %rs49};
|
|
|
|
$L__BB0_119:
|
|
mov.f32 %f966, 0f34000000;
|
|
max.f32 %f967, %f213, %f966;
|
|
mul.f32 %f968, %f116, 0f3F000000;
|
|
div.rn.f32 %f969, %f968, %f967;
|
|
max.f32 %f970, %f214, %f966;
|
|
mul.f32 %f971, %f117, 0f3F000000;
|
|
div.rn.f32 %f972, %f971, %f970;
|
|
max.f32 %f973, %f215, %f966;
|
|
mul.f32 %f974, %f118, 0f3F000000;
|
|
div.rn.f32 %f975, %f974, %f973;
|
|
fma.rn.f32 %f216, %f969, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f217, %f972, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f218, %f975, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f976, %f119, 0f3F000000;
|
|
div.rn.f32 %f977, %f976, %f967;
|
|
mul.f32 %f978, %f120, 0f3F000000;
|
|
div.rn.f32 %f979, %f978, %f970;
|
|
mul.f32 %f980, %f121, 0f3F000000;
|
|
div.rn.f32 %f981, %f980, %f973;
|
|
fma.rn.f32 %f219, %f977, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f220, %f979, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f221, %f981, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f982, %f122, 0f3F000000;
|
|
div.rn.f32 %f983, %f982, %f967;
|
|
mul.f32 %f984, %f123, 0f3F000000;
|
|
div.rn.f32 %f985, %f984, %f970;
|
|
mul.f32 %f986, %f124, 0f3F000000;
|
|
div.rn.f32 %f987, %f986, %f973;
|
|
fma.rn.f32 %f222, %f983, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f223, %f985, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f224, %f987, 0f3F000000, 0f3F000000;
|
|
ld.const.u64 %rd81, [params+272];
|
|
cvta.to.global.u64 %rd82, %rd81;
|
|
ld.const.u32 %r310, [params+264];
|
|
mad.lo.s32 %r311, %r310, %r6, %r5;
|
|
mul.wide.u32 %rd83, %r311, 8;
|
|
add.s64 %rd19, %rd82, %rd83;
|
|
@%p141 bra $L__BB0_121;
|
|
|
|
ld.global.v4.u16 {%rs57, %rs58, %rs59, %rs60}, [%rd19];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f988, %rs57;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f989, %rs58;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f990, %rs59;}
|
|
|
|
// end inline asm
|
|
add.f32 %f991, %f216, %f988;
|
|
add.f32 %f992, %f217, %f989;
|
|
add.f32 %f993, %f218, %f990;
|
|
mov.f32 %f994, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f993;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f992;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f991;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs56, %f994;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd19], {%rs53, %rs54, %rs55, %rs56};
|
|
bra.uni $L__BB0_122;
|
|
|
|
$L__BB0_121:
|
|
mov.f32 %f998, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f998;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f218;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f217;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f216;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd19], {%rs61, %rs62, %rs63, %rs64};
|
|
|
|
$L__BB0_122:
|
|
ld.const.u64 %rd84, [params+288];
|
|
cvta.to.global.u64 %rd85, %rd84;
|
|
ld.const.u32 %r312, [params+280];
|
|
mad.lo.s32 %r313, %r312, %r6, %r5;
|
|
mul.wide.u32 %rd86, %r313, 8;
|
|
add.s64 %rd20, %rd85, %rd86;
|
|
@%p141 bra $L__BB0_124;
|
|
|
|
ld.global.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd20];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f999, %rs72;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1000, %rs73;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1001, %rs74;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1002, %f219, %f999;
|
|
add.f32 %f1003, %f220, %f1000;
|
|
add.f32 %f1004, %f221, %f1001;
|
|
mov.f32 %f1005, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f1004;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1003;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f1002;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f1005;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd20], {%rs68, %rs69, %rs70, %rs71};
|
|
bra.uni $L__BB0_125;
|
|
|
|
$L__BB0_124:
|
|
mov.f32 %f1009, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f1009;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f221;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f220;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f219;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd20], {%rs76, %rs77, %rs78, %rs79};
|
|
|
|
$L__BB0_125:
|
|
ld.const.u64 %rd87, [params+304];
|
|
cvta.to.global.u64 %rd88, %rd87;
|
|
ld.const.u32 %r314, [params+296];
|
|
mad.lo.s32 %r315, %r314, %r6, %r5;
|
|
mul.wide.u32 %rd89, %r315, 8;
|
|
add.s64 %rd21, %rd88, %rd89;
|
|
@%p141 bra $L__BB0_127;
|
|
|
|
ld.global.v4.u16 {%rs87, %rs88, %rs89, %rs90}, [%rd21];
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1010, %rs87;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1011, %rs88;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.f32.f16 %f1012, %rs89;}
|
|
|
|
// end inline asm
|
|
add.f32 %f1013, %f222, %f1010;
|
|
add.f32 %f1014, %f223, %f1011;
|
|
add.f32 %f1015, %f224, %f1012;
|
|
mov.f32 %f1016, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f1015;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs84, %f1014;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1013;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f1016;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd21], {%rs83, %rs84, %rs85, %rs86};
|
|
bra.uni $L__BB0_146;
|
|
|
|
$L__BB0_127:
|
|
mov.f32 %f1020, 0f3F800000;
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f1020;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f224;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f223;}
|
|
|
|
// end inline asm
|
|
// begin inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f222;}
|
|
|
|
// end inline asm
|
|
st.global.v4.u16 [%rd21], {%rs91, %rs92, %rs93, %rs94};
|
|
|
|
$L__BB0_146:
|
|
ret;
|
|
|
|
}
|
|
|