portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-09 21:50:45 +01:00
Code Releases Activity

revert changes due to performance penalty in Wilson using MPI

Browse Source
This commit is contained in:
nmeyer-ur 2020-06-02 13:08:57 +02:00
parent 7bee4ebb54
commit 5050833b42
1 changed files with 11 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
Grid/simd/Grid_a64fx-fixedsize.h
View File

@ -419,7 +419,6 @@ struct Conj{
} }
}; };
/*
struct TimesMinusI{ struct TimesMinusI{
// Complex float // Complex float
inline vecf operator()(vecf a, vecf b){ inline vecf operator()(vecf a, vecf b){
@ -442,11 +441,14 @@ struct TimesMinusI{
return svneg_m(a_v, pg_odd, a_v); return svneg_m(a_v, pg_odd, a_v);
} }
}; };
*/
// alternative implementation using fcadd // alternative implementation using fcadd
// this is not optimal because we have op1 = op2 + TimesMinusI(op3) etc // this is not optimal because we have op1 = op2 + TimesMinusI(op3) etc
// ideally we have AddTimesMinusI(op1,op2,op3) // ideally we have AddTimesMinusI(op1,op2,op3)
//
// makes performance worse in Benchmark_wilson using MPI
// increases halogtime and gathertime
/*
struct TimesMinusI{ struct TimesMinusI{
// Complex float // Complex float
inline vecf operator()(vecf a, vecf b){ inline vecf operator()(vecf a, vecf b){
@ -463,8 +465,8 @@ struct TimesMinusI{
return svcadd_x(pg1, z_v, a, 270); return svcadd_x(pg1, z_v, a, 270);
} }
}; };
*/
/*
struct TimesI{ struct TimesI{
// Complex float // Complex float
inline vecf operator()(vecf a, vecf b){ inline vecf operator()(vecf a, vecf b){
@ -487,11 +489,15 @@ struct TimesI{
return svneg_m(a_v, pg_even, a_v); return svneg_m(a_v, pg_even, a_v);
} }
}; };
*/
// alternative implementation using fcadd // alternative implementation using fcadd
// this is not optimal because we have op1 = op2 + TimesI(op3) etc // this is not optimal because we have op1 = op2 + TimesI(op3) etc
// ideally we have AddTimesI(op1,op2,op3) // ideally we have AddTimesI(op1,op2,op3)
//
// makes performance worse in Benchmark_wilson using MPI
// increases halogtime and gathertime
/*
struct TimesI{ struct TimesI{
// Complex float // Complex float
inline vecf operator()(vecf a, vecf b){ inline vecf operator()(vecf a, vecf b){
@ -508,7 +514,7 @@ struct TimesI{
return svcadd_x(pg1, z_v, a, 90); return svcadd_x(pg1, z_v, a, 90);
} }
}; };
*/
struct PrecisionChange { struct PrecisionChange {