portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-18 07:47:06 +01:00
Code Releases Activity

Serialisation in malloc fixed

Browse Source
This commit is contained in:
paboyle
2016-11-29 22:27:55 +00:00
parent 2f92b4860b
commit bd0430b34f
5 changed files with 128 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
lib/qcd/action/fermion/CayleyFermion5Dvec.cc
View File

@ -71,6 +71,7 @@ void CayleyFermion5D<Impl>::M5D(const FermionField &psi,
chi.checkerboard=psi.checkerboard;
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type * u_p = (scalar_type *)&u[0];
@ -86,6 +87,8 @@ void CayleyFermion5D<Impl>::M5D(const FermionField &psi,
d_p[ss] = diag[s];
}}
M5Dcalls++;
M5Dtime-=usecond();
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=LLs){ // adds LLs
@ -115,6 +118,7 @@ PARALLEL_FOR_LOOP
}
}
M5Dtime+=usecond();
}
template<class Impl>
@ -154,6 +158,8 @@ void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi,
d_p[ss] = diag[s];
}}
M5Dcalls++;
M5Dtime-=usecond();
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=LLs){ // adds LLs
@ -183,6 +189,7 @@ PARALLEL_FOR_LOOP
}
}
M5Dtime+=usecond();
}
template<class Impl>
@ -250,13 +257,11 @@ void CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField
}
}
MooeeInvCalls++;
MooeeInvTime-=usecond();
// Dynamic allocate on stack to get per thread without serialised heap acces
PARALLEL_FOR_LOOP
for(auto site=0;site<vol;site++){
// SiteHalfSpinor *SitePplus =(SiteHalfSpinor *) alloca(LLs*sizeof(SiteHalfSpinor));
// SiteHalfSpinor *SitePminus=(SiteHalfSpinor *) alloca(LLs*sizeof(SiteHalfSpinor));
// SiteSpinor *SiteChi =(SiteSpinor *) alloca(LLs*sizeof(SiteSpinor));
#pragma omp parallel
{
Vector<SiteHalfSpinor> SitePplus(LLs);
Vector<SiteHalfSpinor> SitePminus(LLs);
@ -267,6 +272,9 @@ PARALLEL_FOR_LOOP
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
#pragma omp for
for(auto site=0;site<vol;site++){
for(int s=0;s<LLs;s++){
int lex = s+LLs*site;
spProj5p(SitePplus[s] ,psi[lex]);
@ -294,6 +302,8 @@ PARALLEL_FOR_LOOP
chi[lex] = SiteChi[s]*0.5;
}
}
}
MooeeInvTime+=usecond();
}
INSTANTIATE_DPERP(DomainWallVec5dImplD);