From f7145b973c170a749067a57279e0f80811856950 Mon Sep 17 00:00:00 2001 From: Denis Davydov Date: Mon, 2 Apr 2018 09:13:05 +0200 Subject: [PATCH] introduce LA::d::Vector::add_local() and sadd_local() to avoid to many communication in LA::d::BlockVector::mmult() --- .../lac/la_parallel_block_vector.templates.h | 15 +++++++-- include/deal.II/lac/la_parallel_vector.h | 14 +++++++++ .../lac/la_parallel_vector.templates.h | 31 ++++++++++++++++--- 3 files changed, 52 insertions(+), 8 deletions(-) diff --git a/include/deal.II/lac/la_parallel_block_vector.templates.h b/include/deal.II/lac/la_parallel_block_vector.templates.h index df6ad24706..cb5a2e34b1 100644 --- a/include/deal.II/lac/la_parallel_block_vector.templates.h +++ b/include/deal.II/lac/la_parallel_block_vector.templates.h @@ -944,11 +944,20 @@ namespace LinearAlgebra // below we make this work gracefully for identity-like matrices in // which case the two loops over j won't do any work as A(j,i)==0 const unsigned int k = std::min(i,m-1); - V.block(i).sadd(s,matrix(k,i)*b, this->block(k)); + V.block(i).sadd_local(s,matrix(k,i)*b, this->block(k)); for (unsigned int j = 0 ; j < k; j++) - V.block(i).add (matrix(j,i)*b, this->block(j)); + V.block(i).add_local (matrix(j,i)*b, this->block(j)); for (unsigned int j = k+1; j < m; j++) - V.block(i).add (matrix(j,i)*b, this->block(j)); + V.block(i).add_local (matrix(j,i)*b, this->block(j)); + } + + if (V.block(0).vector_is_ghosted) + { + for (unsigned int i = 0; i < n; i++) + Assert (V.block(i).vector_is_ghosted, + ExcMessage("All blocks should be either in ghosted state or not.")); + + V.update_ghost_values(); } } diff --git a/include/deal.II/lac/la_parallel_vector.h b/include/deal.II/lac/la_parallel_vector.h index ec5b5677ca..00179449cf 100644 --- a/include/deal.II/lac/la_parallel_vector.h +++ b/include/deal.II/lac/la_parallel_vector.h @@ -1060,6 +1060,20 @@ namespace LinearAlgebra << "that this vector can access."); private: + + /** + * Simple addition of a multiple of a vector, i.e. *this += a*V + * without MPI communication. + */ + void add_local(const Number a, const VectorSpaceVector &V); + + /** + * Scaling and simple addition of a multiple of a vector, i.e. *this = + * s*(*this)+a*V without MPI communication. + */ + void sadd_local(const Number s, const Number a, + const VectorSpaceVector &V); + /** * Local part of all_zero(). */ diff --git a/include/deal.II/lac/la_parallel_vector.templates.h b/include/deal.II/lac/la_parallel_vector.templates.h index 6a1a0c53f9..d9ba0a6621 100644 --- a/include/deal.II/lac/la_parallel_vector.templates.h +++ b/include/deal.II/lac/la_parallel_vector.templates.h @@ -839,8 +839,8 @@ namespace LinearAlgebra template void - Vector::add (const Number a, - const VectorSpaceVector &vv) + Vector::add_local (const Number a, + const VectorSpaceVector &vv) { // Downcast. Throws an exception if invalid. Assert(dynamic_cast *>(&vv)!=nullptr, @@ -857,6 +857,16 @@ namespace LinearAlgebra internal::VectorOperations::Vectorization_add_av vector_add(values.get(), v.values.get(), a); internal::VectorOperations::parallel_for(vector_add, 0, partitioner->local_size(), thread_loop_partitioner); + } + + + + template + void + Vector::add (const Number a, + const VectorSpaceVector &vv) + { + add_local(a,vv); if (vector_is_ghosted) update_ghost_values(); @@ -930,9 +940,9 @@ namespace LinearAlgebra template void - Vector::sadd (const Number x, - const Number a, - const VectorSpaceVector &vv) + Vector::sadd_local (const Number x, + const Number a, + const VectorSpaceVector &vv) { // Downcast. Throws an exception if invalid. Assert(dynamic_cast *>(&vv)!=nullptr, @@ -946,6 +956,17 @@ namespace LinearAlgebra internal::VectorOperations::Vectorization_sadd_xav vector_sadd(values.get(), v.values.get(), a, x); internal::VectorOperations::parallel_for(vector_sadd, 0, partitioner->local_size(), thread_loop_partitioner); + } + + + + template + void + Vector::sadd (const Number x, + const Number a, + const VectorSpaceVector &vv) + { + sadd_local(x,a,vv); if (vector_is_ghosted) update_ghost_values(); -- 2.39.5