portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
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
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
Grid/threads/Threads.h
View File

@ -38,14 +38,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#ifdef GRID_OMP #ifdef GRID_OMP
#include <omp.h> #include <omp.h>
// complex reductions
#pragma omp declare reduction(ComplexPlus:Grid::ComplexD, Grid::vComplexD, Grid::ComplexF, Grid::vComplexF: omp_out += omp_in)
#define PARALLEL_FOR_LOOP _Pragma("omp parallel for schedule(static)") #define PARALLEL_FOR_LOOP _Pragma("omp parallel for schedule(static)")
#define PARALLEL_FOR_LOOP_INTERN _Pragma("omp for schedule(static)") #define PARALLEL_FOR_LOOP_INTERN _Pragma("omp for schedule(static)")
#define PARALLEL_FOR_REDUCE_HELPER0(x) #x
#define PARALLEL_FOR_REDUCE_HELPER1(op, var) PARALLEL_FOR_REDUCE_HELPER0(omp parallel for schedule(static) reduction(op:var))
#define PARALLEL_FOR_LOOP_REDUCE(op, var) _Pragma(PARALLEL_FOR_REDUCE_HELPER1(op, var))
#define PARALLEL_NESTED_LOOP2 _Pragma("omp parallel for collapse(2)") #define PARALLEL_NESTED_LOOP2 _Pragma("omp parallel for collapse(2)")
#define PARALLEL_NESTED_LOOP5 _Pragma("omp parallel for collapse(5)") #define PARALLEL_NESTED_LOOP5 _Pragma("omp parallel for collapse(5)")
#define PARALLEL_REGION _Pragma("omp parallel") #define PARALLEL_REGION _Pragma("omp parallel")
@ -63,7 +57,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define parallel_region PARALLEL_REGION #define parallel_region PARALLEL_REGION
#define parallel_for PARALLEL_FOR_LOOP for #define parallel_for PARALLEL_FOR_LOOP for
#define parallel_for_internal PARALLEL_FOR_LOOP_INTERN for #define parallel_for_internal PARALLEL_FOR_LOOP_INTERN for
#define parallel_for_reduce(op, var) PARALLEL_FOR_LOOP_REDUCE(op, var) for
#define parallel_for_nest2 PARALLEL_NESTED_LOOP2 for #define parallel_for_nest2 PARALLEL_NESTED_LOOP2 for
#define parallel_for_nest5 PARALLEL_NESTED_LOOP5 for #define parallel_for_nest5 PARALLEL_NESTED_LOOP5 for

35
Hadrons/A2AMatrix.hpp
View File

@ -167,34 +167,53 @@ public:
template <typename C, typename MatLeft, typename MatRight> template <typename C, typename MatLeft, typename MatRight>
static inline void accTrMul(C &acc, const MatLeft &a, const MatRight &b) 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 if ((MatLeft::Options == Eigen::RowMajor) and
(MatRight::Options == Eigen::ColMajor)) (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 #ifdef USE_MKL
ComplexD tmp; C tmp;
dotuRow(tmp, r, a, b); dotuRow(tmp, r, a, b);
acc += tmp; tacc[thr] += tmp;
#else #else
acc += a.row(r).conjugate().dot(b.col(r)); tacc[thr] += a.row(r).conjugate().dot(b.col(r));
#endif #endif
} }
} }
}
else 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 #ifdef USE_MKL
ComplexD tmp; C tmp;
dotuCol(tmp, c, a, b); dotuCol(tmp, c, a, b);
acc += tmp; tacc[thr] += tmp;
#else #else
acc += a.col(c).conjugate().dot(b.row(c)); tacc[thr] += a.col(c).conjugate().dot(b.row(c));
#endif #endif
} }
} }
} }
for (int thr = 0; thr < nThreads; ++thr)
{
acc += tacc[thr];
}
}
template <typename MatLeft, typename MatRight> template <typename MatLeft, typename MatRight>
static inline double accTrMulFlops(const MatLeft &a, const MatRight &b) static inline double accTrMulFlops(const MatLeft &a, const MatRight &b)

112
Hadrons/Utilities/ContractorBenchmark.cc
View File

@ -199,23 +199,49 @@ void fullTrBenchmark(const unsigned int ni, const unsigned int nj, const unsigne
trBenchmark("Naive loop rows first", left, right, ref, trBenchmark("Naive loop rows first", left, right, ref,
[](ComplexD &res, const MatLeft &a, const MatRight &b) [](ComplexD &res, const MatLeft &a, const MatRight &b)
{ {
int nThreads = GridThread::GetThreads();
std::vector<ComplexD> tres(nThreads, 0.);
res = 0.; res = 0.;
auto nr = a.rows(), nc = a.cols(); parallel_for (int thr = 0; thr < nThreads; ++thr)
parallel_for_reduce(ComplexPlus, res) (unsigned int i = 0; i < nr; ++i) {
int rt, nr;
auto nc = a.cols();
GridThread::GetWork(a.rows(), thr, nr, rt);
for (unsigned int i = rt; i < nr + rt; ++i)
for (unsigned int j = 0; j < nc; ++j) for (unsigned int j = 0; j < nc; ++j)
{ {
res += a(i, j)*b(j, i); tres[thr] += a(i, j)*b(j, i);
}
}
for (int thr = 0; thr < nThreads; ++thr)
{
res += tres[thr];
} }
}); });
trBenchmark("Naive loop cols first", left, right, ref, trBenchmark("Naive loop cols first", left, right, ref,
[](ComplexD &res, const MatLeft &a, const MatRight &b) [](ComplexD &res, const MatLeft &a, const MatRight &b)
{ {
int nThreads = GridThread::GetThreads();
std::vector<ComplexD> tres(nThreads, 0.);
res = 0.; res = 0.;
auto nr = a.rows(), nc = a.cols(); parallel_for (int thr = 0; thr < nThreads; ++thr)
parallel_for_reduce(ComplexPlus, res) (unsigned int j = 0; j < nc; ++j) {
int ct, nc;
auto nr = a.rows();
GridThread::GetWork(a.cols(), thr, nc, ct);
for (unsigned int j = ct; j < nc + ct; ++j)
for (unsigned int i = 0; i < nr; ++i) for (unsigned int i = 0; i < nr; ++i)
{ {
res += a(i, j)*b(j, i); tres[thr] += a(i, j)*b(j, i);
}
}
for (int thr = 0; thr < nThreads; ++thr)
{
res += tres[thr];
} }
}); });
trBenchmark("Eigen tr(A*B)", left, right, ref, trBenchmark("Eigen tr(A*B)", left, right, ref,
@ -226,21 +252,45 @@ void fullTrBenchmark(const unsigned int ni, const unsigned int nj, const unsigne
trBenchmark("Eigen row-wise dot", left, right, ref, trBenchmark("Eigen row-wise dot", left, right, ref,
[](ComplexD &res, const MatLeft &a, const MatRight &b) [](ComplexD &res, const MatLeft &a, const MatRight &b)
{ {
res = 0.; int nThreads = GridThread::GetThreads();
std::vector<ComplexD> tres(nThreads, 0.);
parallel_for_reduce(ComplexPlus, res) (unsigned int r = 0; r < a.rows(); ++r) res = 0.;
parallel_for (int thr = 0; thr < nThreads; ++thr)
{ {
res += a.row(r).conjugate().dot(b.col(r)); int rt, nr;
GridThread::GetWork(a.rows(), thr, nr, rt);
for (unsigned int i = rt; i < nr + rt; ++i)
{
tres[thr] += a.row(i).conjugate().dot(b.col(i));
}
}
for (int thr = 0; thr < nThreads; ++thr)
{
res += tres[thr];
} }
}); });
trBenchmark("Eigen col-wise dot", left, right, ref, trBenchmark("Eigen col-wise dot", left, right, ref,
[](ComplexD &res, const MatLeft &a, const MatRight &b) [](ComplexD &res, const MatLeft &a, const MatRight &b)
{ {
res = 0.; int nThreads = GridThread::GetThreads();
std::vector<ComplexD> tres(nThreads, 0.);
parallel_for_reduce(ComplexPlus, res) (unsigned int c = 0; c < a.cols(); ++c) res = 0.;
parallel_for (int thr = 0; thr < nThreads; ++thr)
{ {
res += a.col(c).conjugate().dot(b.row(c)); int ct, nc;
GridThread::GetWork(a.cols(), thr, nc, ct);
for (unsigned int j = ct; j < nc + ct; ++j)
{
tres[thr] += a.col(j).conjugate().dot(b.row(j));
}
}
for (int thr = 0; thr < nThreads; ++thr)
{
res += tres[thr];
} }
}); });
trBenchmark("Eigen Hadamard", left, right, ref, trBenchmark("Eigen Hadamard", left, right, ref,
@ -252,27 +302,49 @@ void fullTrBenchmark(const unsigned int ni, const unsigned int nj, const unsigne
trBenchmark("MKL row-wise zdotu", left, right, ref, trBenchmark("MKL row-wise zdotu", left, right, ref,
[](ComplexD &res, const MatLeft &a, const MatRight &b) [](ComplexD &res, const MatLeft &a, const MatRight &b)
{ {
res = 0.; int nThreads = GridThread::GetThreads();
std::vector<ComplexD> tres(nThreads, 0.);
parallel_for_reduce(ComplexPlus, res) (unsigned int r = 0; r < a.rows(); ++r) res = 0.;
parallel_for (int thr = 0; thr < nThreads; ++thr)
{ {
ComplexD tmp; ComplexD tmp;
int rt, nr;
zdotuRow(tmp, r, a, b); GridThread::GetWork(a.rows(), thr, nr, rt);
res += tmp; for (unsigned int i = rt; i < nr + rt; ++i)
{
zdotuRow(tmp, i, a, b);
tres[thr] += tmp;
}
}
for (int thr = 0; thr < nThreads; ++thr)
{
res += tres[thr];
} }
}); });
trBenchmark("MKL col-wise zdotu", left, right, ref, trBenchmark("MKL col-wise zdotu", left, right, ref,
[](ComplexD &res, const MatLeft &a, const MatRight &b) [](ComplexD &res, const MatLeft &a, const MatRight &b)
{ {
res = 0.; int nThreads = GridThread::GetThreads();
std::vector<ComplexD> tres(nThreads, 0.);
parallel_for_reduce(ComplexPlus, res) (unsigned int c = 0; c < a.cols(); ++c) res = 0.;
parallel_for (int thr = 0; thr < nThreads; ++thr)
{ {
ComplexD tmp; ComplexD tmp;
int ct, nc;
zdotuCol(tmp, c, a, b); GridThread::GetWork(a.cols(), thr, nc, ct);
res += tmp; for (unsigned int j = ct; j < nc + ct; ++j)
{
zdotuCol(tmp, j, a, b);
tres[thr] += tmp;
}
}
for (int thr = 0; thr < nThreads; ++thr)
{
res += tres[thr];
} }
}); });
#endif #endif