GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-13 12:19:12 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add multi-threading for sparse-row-major * dense-row-major

Browse Source
This commit is contained in:
Gael Guennebaud 2018-04-25 10:14:48 +02:00
parent 2f3287da7d
commit 8810baaed4
1 changed files with 30 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
Eigen/src/SparseCore/SparseDenseProduct.h
View File

@ -88,10 +88,11 @@ struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, A
typedef typename internal::remove_all<SparseLhsType>::type Lhs; typedef typename internal::remove_all<SparseLhsType>::type Lhs;
typedef typename internal::remove_all<DenseRhsType>::type Rhs; typedef typename internal::remove_all<DenseRhsType>::type Rhs;
typedef typename internal::remove_all<DenseResType>::type Res; typedef typename internal::remove_all<DenseResType>::type Res;
typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator; typedef evaluator<Lhs> LhsEval;
typedef typename LhsEval::InnerIterator LhsInnerIterator;
static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha)
{ {
evaluator<Lhs> lhsEval(lhs); LhsEval lhsEval(lhs);
for(Index c=0; c<rhs.cols(); ++c) for(Index c=0; c<rhs.cols(); ++c)
{ {
for(Index j=0; j<lhs.outerSize(); ++j) for(Index j=0; j<lhs.outerSize(); ++j)
@ -111,16 +112,37 @@ struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, t
typedef typename internal::remove_all<SparseLhsType>::type Lhs; typedef typename internal::remove_all<SparseLhsType>::type Lhs;
typedef typename internal::remove_all<DenseRhsType>::type Rhs; typedef typename internal::remove_all<DenseRhsType>::type Rhs;
typedef typename internal::remove_all<DenseResType>::type Res; typedef typename internal::remove_all<DenseResType>::type Res;
typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator; typedef evaluator<Lhs> LhsEval;
typedef typename LhsEval::InnerIterator LhsInnerIterator;
static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha)
{ {
evaluator<Lhs> lhsEval(lhs); Index n = lhs.rows();
for(Index j=0; j<lhs.outerSize(); ++j) LhsEval lhsEval(lhs);
#ifdef EIGEN_HAS_OPENMP
Eigen::initParallel();
Index threads = Eigen::nbThreads();
// This 20000 threshold has been found experimentally on 2D and 3D Poisson problems.
// It basically represents the minimal amount of work to be done to be worth it.
if(threads>1 && lhsEval.nonZerosEstimate()*rhs.cols() > 20000)
{ {
typename Res::RowXpr res_j(res.row(j)); #pragma omp parallel for schedule(dynamic,(n+threads*4-1)/(threads*4)) num_threads(threads)
for(LhsInnerIterator it(lhsEval,j); it ;++it) for(Index i=0; i<n; ++i)
res_j += (alpha*it.value()) * rhs.row(it.index()); processRow(lhsEval,rhs,res,alpha,i);
} }
else
#endif
{
for(Index i=0; i<n; ++i)
processRow(lhsEval, rhs, res, alpha, i);
}
}
static void processRow(const LhsEval& lhsEval, const DenseRhsType& rhs, Res& res, const typename Res::Scalar& alpha, Index i)
{
typename Res::RowXpr res_i(res.row(i));
for(LhsInnerIterator it(lhsEval,i); it ;++it)
res_i += (alpha*it.value()) * rhs.row(it.index());
} }
}; };