From 595555c739072b04bdac505f760f34012fcf3b9c Mon Sep 17 00:00:00 2001 From: Denis Davydov Date: Wed, 21 Feb 2018 14:09:11 +0100 Subject: [PATCH] add more parameters to LA::distributed::BlockVector::mmult() change meaning of arguments in BlockVector::mmult() --- doc/news/changes/minor/20180210DenisDavydov | 2 +- .../deal.II/lac/la_parallel_block_vector.h | 12 +- .../lac/la_parallel_block_vector.templates.h | 39 ++-- source/lac/la_parallel_block_vector.inst.in | 4 +- tests/mpi/parallel_block_vector_08.cc | 2 +- tests/mpi/parallel_block_vector_09.cc | 2 +- tests/mpi/parallel_block_vector_12.cc | 181 ++++++++++++++++++ ..._vector_12.with_p4est=true.mpirun=4.output | 4 + 8 files changed, 221 insertions(+), 25 deletions(-) create mode 100644 tests/mpi/parallel_block_vector_12.cc create mode 100644 tests/mpi/parallel_block_vector_12.with_p4est=true.mpirun=4.output diff --git a/doc/news/changes/minor/20180210DenisDavydov b/doc/news/changes/minor/20180210DenisDavydov index ddda9265bb..27f476c8ea 100644 --- a/doc/news/changes/minor/20180210DenisDavydov +++ b/doc/news/changes/minor/20180210DenisDavydov @@ -1,4 +1,4 @@ -New: add parallel::distributed::BlockVector::mmult(const BlockVector &, const FullMatrixType &) +New: add parallel::distributed::BlockVector::mmult(BlockVector &, const FullMatrixType &, const NumberType, const NumberType) const and parallel::distributed::BlockVector::multivector_inner_product_with_metric(const FullMatrixType &, const BlockVector &V, const bool) const to operate on multivectors with a metric tensor.
diff --git a/include/deal.II/lac/la_parallel_block_vector.h b/include/deal.II/lac/la_parallel_block_vector.h index 31d644831a..9d0c60b3ff 100644 --- a/include/deal.II/lac/la_parallel_block_vector.h +++ b/include/deal.II/lac/la_parallel_block_vector.h @@ -491,16 +491,18 @@ namespace LinearAlgebra /** * Set each block of this vector as follows: - * $U^i = \sum_{j} V_j A^{ji}$ where $U^i$ - * and $V_j$ indicate the $i$th block (not element) of $U$ and the - * $j$th block of $V$, respectively. + * $V^i = s V^i + b \sum_{j} U_j A^{ji}$ where $V^i$ + * and $U_j$ indicate the $i$th block (not element) of $V$ and the + * $j$th block of $U$, respectively. * * Obviously, this function can only be used if all blocks of both vectors * are of the same size. */ template - void mmult(const BlockVector &V, - const FullMatrixType &matrix); + void mmult(BlockVector &V, + const FullMatrixType &matrix, + const Number s = Number(0.), + const Number b = Number(1.)) const; /** * Add @p a to all components. Note that @p a is a scalar not a vector. 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 0090e56c5c..df6ad24706 100644 --- a/include/deal.II/lac/la_parallel_block_vector.templates.h +++ b/include/deal.II/lac/la_parallel_block_vector.templates.h @@ -835,9 +835,9 @@ namespace LinearAlgebra return; Assert (matrix.m() == m, - dealii::ExcDimensionMismatch(matrix.m(),m)); + ExcDimensionMismatch(matrix.m(),m)); Assert (matrix.n() == n, - dealii::ExcDimensionMismatch(matrix.n(),n)); + ExcDimensionMismatch(matrix.n(),n)); // reset the matrix matrix = typename FullMatrixType::value_type(0.0); @@ -846,7 +846,7 @@ namespace LinearAlgebra if (symmetric) { Assert (m == n, - dealii::ExcDimensionMismatch(m,n)); + ExcDimensionMismatch(m,n)); for (unsigned int i = 0; i < m; i++) for (unsigned int j = i; j < n; j++) @@ -888,14 +888,14 @@ namespace LinearAlgebra return res; Assert (matrix.m() == m, - dealii::ExcDimensionMismatch(matrix.m(),m)); + ExcDimensionMismatch(matrix.m(),m)); Assert (matrix.n() == n, - dealii::ExcDimensionMismatch(matrix.n(),n)); + ExcDimensionMismatch(matrix.n(),n)); if (symmetric) { Assert (m == n, - dealii::ExcDimensionMismatch(m,n)); + ExcDimensionMismatch(m,n)); for (unsigned int i = 0; i < m; i++) { @@ -919,11 +919,13 @@ namespace LinearAlgebra template template void - BlockVector::mmult(const BlockVector &V, - const FullMatrixType &matrix) + BlockVector::mmult(BlockVector &V, + const FullMatrixType &matrix, + const Number s, + const Number b) const { - const unsigned int n = this->n_blocks(); - const unsigned int m = V.n_blocks(); + const unsigned int m = this->n_blocks(); + const unsigned int n = V.n_blocks(); // in case one vector is empty and the second one is not, the // FullMatrix resized to (m,n) will have 0 both in m() and n() @@ -933,14 +935,21 @@ namespace LinearAlgebra return; Assert (matrix.m() == m, - dealii::ExcDimensionMismatch(matrix.m(),m)); + ExcDimensionMismatch(matrix.m(),m)); Assert (matrix.n() == n, - dealii::ExcDimensionMismatch(matrix.n(),n)); + ExcDimensionMismatch(matrix.n(),n)); - (*this) = Number(); for (unsigned int i = 0; i < n; i++) - for (unsigned int j = 0; j < m; j++) - this->block(i).add (matrix(j,i), V.block(j)); + { + // 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)); + for (unsigned int j = 0 ; j < k; j++) + V.block(i).add (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)); + } } diff --git a/source/lac/la_parallel_block_vector.inst.in b/source/lac/la_parallel_block_vector.inst.in index 657e150417..b87cb3e604 100644 --- a/source/lac/la_parallel_block_vector.inst.in +++ b/source/lac/la_parallel_block_vector.inst.in @@ -36,10 +36,10 @@ for (SCALAR : REAL_AND_COMPLEX_SCALARS) BlockVector::multivector_inner_product_with_metric(const LAPACKFullMatrix &, const BlockVector &V, const bool) const; template void - BlockVector::mmult(const BlockVector &V,const FullMatrix &); + BlockVector::mmult(BlockVector &V,const FullMatrix &, const SCALAR, const SCALAR) const; template void - BlockVector::mmult(const BlockVector &V,const LAPACKFullMatrix &); + BlockVector::mmult(BlockVector &V,const LAPACKFullMatrix &, const SCALAR, const SCALAR) const; \} \} } diff --git a/tests/mpi/parallel_block_vector_08.cc b/tests/mpi/parallel_block_vector_08.cc index bccb635727..8c2d4d3d69 100644 --- a/tests/mpi/parallel_block_vector_08.cc +++ b/tests/mpi/parallel_block_vector_08.cc @@ -129,7 +129,7 @@ void test (const unsigned int n_blocks = 5) for (unsigned int j = 0; j < n_blocks; ++j) metric(i,j) = 0.3 + (i*3 + j*7); - left.mmult(right,metric); + right.mmult(left,metric); for (unsigned int i = 0; i < n_blocks; ++i) for (unsigned int j = 0; j < n_blocks; ++j) diff --git a/tests/mpi/parallel_block_vector_09.cc b/tests/mpi/parallel_block_vector_09.cc index a66bc3780e..f4bdf5e683 100644 --- a/tests/mpi/parallel_block_vector_09.cc +++ b/tests/mpi/parallel_block_vector_09.cc @@ -139,7 +139,7 @@ void test (const unsigned int n_blocks = 5) } const double res = left.multivector_inner_product_with_metric(metric, right); - left2.mmult(left, metric); + left.mmult(left2, metric); const double res2 = left2 * right; const double diff_norm = std::abs(res - res2); diff --git a/tests/mpi/parallel_block_vector_12.cc b/tests/mpi/parallel_block_vector_12.cc new file mode 100644 index 0000000000..4518331011 --- /dev/null +++ b/tests/mpi/parallel_block_vector_12.cc @@ -0,0 +1,181 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2017 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE at +// the top level of the deal.II distribution. +// +// --------------------------------------------------------------------- + + + +// this BlockVector::mmult(const BlockVector &V,const FullMatrixType &matrix, s,b) +// also for vectors of different number of blocks. +// Triangulation and Mass operator are the same as in matrix_free/mass_operator_01.cc + +#include "../tests.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include + + + + +template +void test (const unsigned int n = 5, const unsigned int m = 3) +{ + typedef double number; + + parallel::distributed::Triangulation tria (MPI_COMM_WORLD); + GridGenerator::hyper_cube (tria); + tria.refine_global(1); + typename Triangulation::active_cell_iterator + cell = tria.begin_active (), + endc = tria.end(); + cell = tria.begin_active (); + for (; cell!=endc; ++cell) + if (cell->is_locally_owned()) + if (cell->center().norm()<0.2) + cell->set_refine_flag(); + tria.execute_coarsening_and_refinement(); + if (dim < 3 && fe_degree < 2) + tria.refine_global(2); + else + tria.refine_global(1); + if (tria.begin(tria.n_levels()-1)->is_locally_owned()) + tria.begin(tria.n_levels()-1)->set_refine_flag(); + if (tria.last()->is_locally_owned()) + tria.last()->set_refine_flag(); + tria.execute_coarsening_and_refinement(); + cell = tria.begin_active (); + for (unsigned int i=0; i<10-3*dim; ++i) + { + cell = tria.begin_active (); + unsigned int counter = 0; + for (; cell!=endc; ++cell, ++counter) + if (cell->is_locally_owned()) + if (counter % (7-i) == 0) + cell->set_refine_flag(); + tria.execute_coarsening_and_refinement(); + } + + FE_Q fe (fe_degree); + DoFHandler dof (tria); + dof.distribute_dofs(fe); + + IndexSet owned_set = dof.locally_owned_dofs(); + IndexSet relevant_set; + DoFTools::extract_locally_relevant_dofs (dof, relevant_set); + + ConstraintMatrix constraints (relevant_set); + DoFTools::make_hanging_node_constraints(dof, constraints); + VectorTools::interpolate_boundary_values (dof, 0, Functions::ZeroFunction(), + constraints); + constraints.close(); + + std::shared_ptr > mf_data(new MatrixFree ()); + { + const QGauss<1> quad (fe_degree+2); + typename MatrixFree::AdditionalData data; + data.tasks_parallel_scheme = + MatrixFree::AdditionalData::none; + data.tasks_block_size = 7; + mf_data->reinit (dof, constraints, quad, data); + } + + MatrixFreeOperators::MassOperator > mf; + mf.initialize(mf_data); + mf.compute_diagonal(); + + LinearAlgebra::distributed::BlockVector left(n), right(m), left2(n); + for (unsigned int b = 0; b < n; ++b) + { + mf_data->initialize_dof_vector (left.block(b)); + mf_data->initialize_dof_vector (left2.block(b)); + left.block(b) = 0.; + left2.block(b) = 0.; + for (unsigned int i=0; i(); + } + } + + for (unsigned int b = 0; b < m; ++b) + { + mf_data->initialize_dof_vector (right.block(b)); + right.block(b) = 0.; + for (unsigned int i=0; i(); + } + } + + + FullMatrix metric(m,n); + for (unsigned int i = 0; i < m; ++i) + for (unsigned int j = 0; j < n; ++j) + metric(i,j) = 0.3 + (i*3 + j*7); + + right.mmult(left,metric); // L = RM + right.mmult(left,metric,2.,3.); // L = 2L + 3 RM = 5RM + + for (unsigned int i = 0; i < n; ++i) + for (unsigned int j = 0; j < m; ++j) + left2.block(i).add(5.*metric(j,i), right.block(j)); + + left2.add(-1., left); + + const double diff_norm = left2.linfty_norm(); + deallog << "Norm of difference: " << diff_norm << std::endl; +} + + +int main (int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization (argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process (MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test<2,1>(5,3); + test<2,1>(3,3); + test<2,1>(3,5); + } + else + { + test<2,1>(5,3); + test<2,1>(3,3); + test<2,1>(3,5); + } +} diff --git a/tests/mpi/parallel_block_vector_12.with_p4est=true.mpirun=4.output b/tests/mpi/parallel_block_vector_12.with_p4est=true.mpirun=4.output new file mode 100644 index 0000000000..48c41a9ae5 --- /dev/null +++ b/tests/mpi/parallel_block_vector_12.with_p4est=true.mpirun=4.output @@ -0,0 +1,4 @@ + +DEAL:0::Norm of difference: 0.000 +DEAL:0::Norm of difference: 0.000 +DEAL:0::Norm of difference: 0.000 -- 2.39.5