portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-25 11:12:02 +01:00
Code Releases Activity

Hadrons: remove the use of OpenMP reductions

Browse Source
This commit is contained in:
Antonin Portelli
2018-11-16 20:00:29 +00:00
parent 2ac57370f1
commit 8b007b5c24
3 changed files with 128 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
Hadrons/A2AMatrix.hpp
View File

@ -167,33 +167,52 @@ public:
template <typename C, typename MatLeft, typename MatRight>
static inline void accTrMul(C &acc, const MatLeft &a, const MatRight &b)
{
int nThreads = GridThread::GetThreads();
std::vector<C> tacc(nThreads, 0.);
if ((MatLeft::Options == Eigen::RowMajor) and
(MatRight::Options == Eigen::ColMajor))
{
parallel_for_reduce(ComplexPlus, acc) (unsigned int r = 0; r < a.rows(); ++r)
parallel_for (int thr = 0; thr < nThreads; ++thr)
{
int rt, nr;
GridThread::GetWork(a.rows(), thr, nr, rt);
for (unsigned int r = rt; r < nr + rt; ++r)
{
#ifdef USE_MKL
ComplexD tmp;
dotuRow(tmp, r, a, b);
acc += tmp;
C tmp;
dotuRow(tmp, r, a, b);
tacc[thr] += tmp;
#else
acc += a.row(r).conjugate().dot(b.col(r));
tacc[thr] += a.row(r).conjugate().dot(b.col(r));
#endif
}
}
}
else
{
parallel_for_reduce(ComplexPlus, acc) (unsigned int c = 0; c < a.cols(); ++c)
parallel_for (int thr = 0; thr < nThreads; ++thr)
{
int ct, nc;
GridThread::GetWork(a.cols(), thr, nc, ct);
for (unsigned int c = ct; c < nc + ct; ++c)
{
#ifdef USE_MKL
ComplexD tmp;
dotuCol(tmp, c, a, b);
acc += tmp;
C tmp;
dotuCol(tmp, c, a, b);
tacc[thr] += tmp;
#else
acc += a.col(c).conjugate().dot(b.row(c));
tacc[thr] += a.col(c).conjugate().dot(b.row(c));
#endif
}
}
}
for (int thr = 0; thr < nThreads; ++thr)
{
acc += tacc[thr];
}
}
template <typename MatLeft, typename MatRight>