From: Denis Davydov Date: Sat, 30 Sep 2017 21:36:42 +0000 (+0200) Subject: add LA::d::BlockVector::multivector_inner_product() X-Git-Tag: v9.0.0-rc1~1000^2~2 X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=3bb5e0da3ac9508356090e824ea399262ca873bb;p=dealii.git add LA::d::BlockVector::multivector_inner_product() --- diff --git a/doc/news/changes/minor/20170930DenisDavydov b/doc/news/changes/minor/20170930DenisDavydov new file mode 100644 index 0000000000..125ce686b4 --- /dev/null +++ b/doc/news/changes/minor/20170930DenisDavydov @@ -0,0 +1,4 @@ +New: add LinearAlgebra::distributed::BlockVector::multivector_inner_product() +to perform inner product between each block of the two block vectors. +
+(Denis Davydov, 2017/09/30) diff --git a/include/deal.II/lac/la_parallel_block_vector.h b/include/deal.II/lac/la_parallel_block_vector.h index 11b0101221..37006d8407 100644 --- a/include/deal.II/lac/la_parallel_block_vector.h +++ b/include/deal.II/lac/la_parallel_block_vector.h @@ -447,6 +447,24 @@ namespace LinearAlgebra */ virtual Number operator* (const VectorSpaceVector &V) const override; + /** + * Calculate the scalar product between each block of this vector and @p V + * and store the result in a full matrix @p matrix. This function + * computes the result by forming $A_{ij}=U_i \cdot V_j$ where $U_i$ + * and $V_i$ indicate the $i$th block (not element!) of $U$ and the + * $j$th block of $V$, respectively. If @p symmetric is + * true, it is assumed that inner product results in a + * square symmetric matrix and almost half of the scalar products can be avoided. + * + * Obviously, this function can only be used if each block in this + * object and @p V are of the same size. + * + * @note Internally, a single global reduction will be called to + * accumulate scalar product between locally owned degrees of freedom. + */ + template + void multivector_inner_product(FullMatrixType &matrix, const BlockVector &V, const bool symmetric = false) 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 1200f1fcbf..64c1afca1b 100644 --- a/include/deal.II/lac/la_parallel_block_vector.templates.h +++ b/include/deal.II/lac/la_parallel_block_vector.templates.h @@ -779,6 +779,56 @@ namespace LinearAlgebra MemoryConsumption::memory_consumption (this->components)); } + + + template + template + void + BlockVector::multivector_inner_product(FullMatrixType &matrix, + const BlockVector &V, + const bool symmetric) const + { + 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() + // which is how TableBase::reinit() works as of deal.ii@8.5.0. + // Since in this case there is nothing to do anyway -- return immediately. + if (n==0 || m==0) + return; + + Assert (matrix.m() == m, + dealii::ExcDimensionMismatch(matrix.m(),m)); + Assert (matrix.n() == n, + dealii::ExcDimensionMismatch(matrix.n(),n)); + + // reset the matrix + matrix = Number(); + + if (symmetric) + { + Assert (m == n, + dealii::ExcDimensionMismatch(m,n)); + + for (unsigned int i = 0; i < m; i++) + for (unsigned int j = i; j < n; j++) + matrix(i,j) = this->block(i).inner_product_local(V.block(j)); + + for (unsigned int i = 0; i < m; i++) + for (unsigned int j = i+1; j < n; j++) + matrix(j,i) = matrix(i,j); + } + else + { + for (unsigned int i = 0; i < m; ++i) + for (unsigned int j = 0; j < n; ++j) + matrix(i,j) = this->block(i).inner_product_local(V.block(j)); + } + + Utilities::MPI::sum(matrix, this->block(0).get_mpi_communicator(), matrix); + } + } // end of namespace distributed } // end of namespace parallel diff --git a/source/lac/la_parallel_block_vector.cc b/source/lac/la_parallel_block_vector.cc index 56fc41983d..9e8b8397e0 100644 --- a/source/lac/la_parallel_block_vector.cc +++ b/source/lac/la_parallel_block_vector.cc @@ -16,6 +16,9 @@ #include #include +#include +#include + DEAL_II_NAMESPACE_OPEN diff --git a/source/lac/la_parallel_block_vector.inst.in b/source/lac/la_parallel_block_vector.inst.in index 5fab92fb70..41ecec5924 100644 --- a/source/lac/la_parallel_block_vector.inst.in +++ b/source/lac/la_parallel_block_vector.inst.in @@ -22,6 +22,12 @@ for (SCALAR : REAL_SCALARS) namespace distributed \{ template class BlockVector; + template + void + BlockVector::multivector_inner_product(FullMatrix &, const BlockVector &V, const bool) const; + template + void + BlockVector::multivector_inner_product(LAPACKFullMatrix &, const BlockVector &V, const bool) const; \} \} } diff --git a/tests/mpi/parallel_block_vector_05.cc b/tests/mpi/parallel_block_vector_05.cc new file mode 100644 index 0000000000..4078094d9a --- /dev/null +++ b/tests/mpi/parallel_block_vector_05.cc @@ -0,0 +1,200 @@ +// --------------------------------------------------------------------- +// +// 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::scalar_product(). +// 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_blocks = 5) +{ + 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_blocks), right(n_blocks); + for (unsigned int b = 0; b < n_blocks; ++b) + { + mf_data->initialize_dof_vector (left.block(b)); + mf_data->initialize_dof_vector (right.block(b)); + + for (unsigned int i=0; i product(n_blocks,n_blocks), reference(n_blocks,n_blocks); + + deallog << "Symmetric:" << std::endl; + + left.multivector_inner_product(product, right, true); + + for (unsigned int i = 0; i < n_blocks; ++i) + for (unsigned int j = 0; j < n_blocks; ++j) + reference(i,j) = left.block(i) * right.block(j); + + product.add(-1.,reference); + + const double diff_norm = product.frobenius_norm(); + + deallog << "Norm of difference: " << diff_norm << std::endl; + } + + // Non-symmetric case + { + const unsigned int n_blocks_2 = n_blocks+3; + LinearAlgebra::distributed::BlockVector left2(n_blocks_2); + for (unsigned int b = 0; b < left2.n_blocks(); ++b) + { + mf_data->initialize_dof_vector (left2.block(b)); + for (unsigned int i=0; i product2(n_blocks_2,n_blocks), reference2(n_blocks_2,n_blocks); + + deallog << "Non-Symmetric:" << std::endl; + + left2.multivector_inner_product(product2, right, false); + + for (unsigned int i = 0; i < n_blocks_2; ++i) + for (unsigned int j = 0; j < n_blocks; ++j) + reference2(i,j) = left2.block(i) * right.block(j); + + product2.add(-1.,reference2); + + const double diff_norm2 = product2.frobenius_norm(); + + deallog << "Norm of difference: " << diff_norm2 << 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>(); + } + else + { + test<2,1>(); + } +} + diff --git a/tests/mpi/parallel_block_vector_05.with_p4est=true.mpirun=1.output b/tests/mpi/parallel_block_vector_05.with_p4est=true.mpirun=1.output new file mode 100644 index 0000000000..6063f6f07e --- /dev/null +++ b/tests/mpi/parallel_block_vector_05.with_p4est=true.mpirun=1.output @@ -0,0 +1,5 @@ + +DEAL:0::Symmetric: +DEAL:0::Norm of difference: 0.000 +DEAL:0::Non-Symmetric: +DEAL:0::Norm of difference: 0.000 diff --git a/tests/mpi/parallel_block_vector_05.with_p4est=true.mpirun=4.output b/tests/mpi/parallel_block_vector_05.with_p4est=true.mpirun=4.output new file mode 100644 index 0000000000..6063f6f07e --- /dev/null +++ b/tests/mpi/parallel_block_vector_05.with_p4est=true.mpirun=4.output @@ -0,0 +1,5 @@ + +DEAL:0::Symmetric: +DEAL:0::Norm of difference: 0.000 +DEAL:0::Non-Symmetric: +DEAL:0::Norm of difference: 0.000