trilinos_sparsity_pattern.cc

Go to the documentation of this file.

// ------------------------------------------------------------------------
//
// SPDX-License-Identifier: LGPL-2.1-or-later
// Copyright (C) 2008 - 2025 by the deal.II authors
//
// This file is part of the deal.II library.
//
// Part of the source code is dual licensed under Apache-2.0 WITH
// LLVM-exception OR LGPL-2.1-or-later. Detailed license information
// governing the source code and code contributions can be found in
// LICENSE.md and CONTRIBUTING.md at the top level directory of deal.II.
//
// ------------------------------------------------------------------------
 
#include <deal.II/lac/trilinos_index_access.h>
#include <deal.II/lac/trilinos_sparsity_pattern.h>
 
#ifdef DEAL_II_WITH_TRILINOS
 
#  include <deal.II/base/mpi.h>
#  include <deal.II/base/trilinos_utilities.h>
 
#  include <deal.II/lac/dynamic_sparsity_pattern.h>
#  include <deal.II/lac/sparsity_pattern.h>
 
DEAL_II_DISABLE_EXTRA_DIAGNOSTICS
#  include <Epetra_Export.h>
DEAL_II_ENABLE_EXTRA_DIAGNOSTICS
 
#  include <limits>
 
DEAL_II_NAMESPACE_OPEN
 
namespace TrilinosWrappers
{
  namespace SparsityPatternIterators
  {
    void
    Accessor::visit_present_row()
    {
      // if we are asked to visit the past-the-end line, then simply
      // release all our caches and go on with life
      if (this->a_row == sparsity_pattern->n_rows() ||
          (sparsity_pattern->in_local_range(this->a_row) == false))
        {
          colnum_cache.reset();
          return;
        }
 
      // otherwise first flush Trilinos caches if necessary
      if (!sparsity_pattern->is_compressed())
        sparsity_pattern->compress();
 
      colnum_cache = std::make_shared<std::vector<size_type>>(
        sparsity_pattern->row_length(this->a_row));
 
      if (colnum_cache->size() > 0)
        {
          // get a representation of the present row
          int       ncols;
          const int ierr = sparsity_pattern->graph->ExtractGlobalRowCopy(
            this->a_row,
            colnum_cache->size(),
            ncols,
            reinterpret_cast<TrilinosWrappers::types::int_type *>(
              const_cast<size_type *>(colnum_cache->data())));
          AssertThrow(ierr == 0, ExcTrilinosError(ierr));
          AssertThrow(static_cast<std::vector<size_type>::size_type>(ncols) ==
                        colnum_cache->size(),
                      ExcInternalError());
        }
    }
  } // namespace SparsityPatternIterators
 
 
  // The constructor is actually the
  // only point where we have to check
  // whether we build a serial or a
  // parallel Trilinos matrix.
  // Actually, it does not even matter
  // how many threads there are, but
  // only if we use an MPI compiler or
  // a standard compiler. So, even one
  // thread on a configuration with
  // MPI will still get a parallel
  // interface.
  SparsityPattern::SparsityPattern()
  {
    column_space_map =
      std::make_unique<Epetra_Map>(TrilinosWrappers::types::int_type(0),
                                   TrilinosWrappers::types::int_type(0),
                                   Utilities::Trilinos::comm_self());
    graph = std::make_unique<Epetra_FECrsGraph>(View,
                                                *column_space_map,
                                                *column_space_map,
                                                0);
    graph->FillComplete();
  }
 
 
 
  SparsityPattern::SparsityPattern(const size_type m,
                                   const size_type n,
                                   const size_type n_entries_per_row)
  {
    reinit(m, n, n_entries_per_row);
  }
 
 
 
  SparsityPattern::SparsityPattern(
    const size_type               m,
    const size_type               n,
    const std::vector<size_type> &n_entries_per_row)
  {
    reinit(m, n, n_entries_per_row);
  }
 
 
 
  SparsityPattern::SparsityPattern(SparsityPattern &&other) noexcept
    : SparsityPatternBase(std::move(other))
    , column_space_map(std::move(other.column_space_map))
    , graph(std::move(other.graph))
    , nonlocal_graph(std::move(other.nonlocal_graph))
  {}
 
 
 
  // Copy function only works if the
  // sparsity pattern is empty.
  SparsityPattern::SparsityPattern(const SparsityPattern &input_sparsity)
    : SparsityPatternBase(input_sparsity)
    , column_space_map(new Epetra_Map(TrilinosWrappers::types::int_type(0),
                                      TrilinosWrappers::types::int_type(0),
                                      Utilities::Trilinos::comm_self()))
    , graph(
        new Epetra_FECrsGraph(View, *column_space_map, *column_space_map, 0))
  {
    Assert(input_sparsity.n_rows() == 0,
           ExcMessage(
             "Copy constructor only works for empty sparsity patterns."));
  }
 
 
 
  SparsityPattern::SparsityPattern(const IndexSet &parallel_partitioning,
                                   const MPI_Comm  communicator,
                                   const size_type n_entries_per_row)
  {
    reinit(parallel_partitioning,
           parallel_partitioning,
           communicator,
           n_entries_per_row);
  }
 
 
 
  SparsityPattern::SparsityPattern(
    const IndexSet               &parallel_partitioning,
    const MPI_Comm                communicator,
    const std::vector<size_type> &n_entries_per_row)
  {
    reinit(parallel_partitioning,
           parallel_partitioning,
           communicator,
           n_entries_per_row);
  }
 
 
 
  SparsityPattern::SparsityPattern(const IndexSet &row_parallel_partitioning,
                                   const IndexSet &col_parallel_partitioning,
                                   const MPI_Comm  communicator,
                                   const size_type n_entries_per_row)
  {
    reinit(row_parallel_partitioning,
           col_parallel_partitioning,
           communicator,
           n_entries_per_row);
  }
 
 
 
  SparsityPattern::SparsityPattern(
    const IndexSet               &row_parallel_partitioning,
    const IndexSet               &col_parallel_partitioning,
    const MPI_Comm                communicator,
    const std::vector<size_type> &n_entries_per_row)
  {
    reinit(row_parallel_partitioning,
           col_parallel_partitioning,
           communicator,
           n_entries_per_row);
  }
 
 
 
  SparsityPattern::SparsityPattern(const IndexSet &row_parallel_partitioning,
                                   const IndexSet &col_parallel_partitioning,
                                   const IndexSet &writable_rows,
                                   const MPI_Comm  communicator,
                                   const size_type n_max_entries_per_row)
  {
    reinit(row_parallel_partitioning,
           col_parallel_partitioning,
           writable_rows,
           communicator,
           n_max_entries_per_row);
  }
 
 
 
  void
  SparsityPattern::reinit(const size_type m,
                          const size_type n,
                          const size_type n_entries_per_row)
  {
    reinit(complete_index_set(m),
           complete_index_set(n),
           MPI_COMM_SELF,
           n_entries_per_row);
  }
 
 
 
  void
  SparsityPattern::reinit(const size_type               m,
                          const size_type               n,
                          const std::vector<size_type> &n_entries_per_row)
  {
    reinit(complete_index_set(m),
           complete_index_set(n),
           MPI_COMM_SELF,
           n_entries_per_row);
  }
 
 
 
  namespace
  {
    using size_type = SparsityPattern::size_type;
 
    void
    reinit_sp(const Epetra_Map                   &row_map,
              const Epetra_Map                   &col_map,
              const size_type                     n_entries_per_row,
              std::unique_ptr<Epetra_Map>        &column_space_map,
              std::unique_ptr<Epetra_FECrsGraph> &graph,
              std::unique_ptr<Epetra_CrsGraph>   &nonlocal_graph)
    {
      Assert(row_map.IsOneToOne(),
             ExcMessage("Row map must be 1-to-1, i.e., no overlap between "
                        "the maps of different processors."));
      Assert(col_map.IsOneToOne(),
             ExcMessage("Column map must be 1-to-1, i.e., no overlap between "
                        "the maps of different processors."));
 
      nonlocal_graph.reset();
      graph.reset();
      column_space_map = std::make_unique<Epetra_Map>(col_map);
 
      AssertThrow((TrilinosWrappers::max_my_gid(row_map) -
                   TrilinosWrappers::min_my_gid(row_map)) *
                      std::uint64_t(n_entries_per_row) <
                    static_cast<std::uint64_t>(std::numeric_limits<int>::max()),
                  ExcMessage("The TrilinosWrappers use Epetra internally which "
                             "uses 'signed int' to represent local indices. "
                             "Therefore, only 2,147,483,647 nonzero matrix "
                             "entries can be stored on a single process, "
                             "but you are requesting more than that. "
                             "If possible, use more MPI processes."));
 
 
      // for more than one processor, need to specify only row map first and
      // let the matrix entries decide about the column map (which says which
      // columns are present in the matrix, not to be confused with the
      // col_map that tells how the domain dofs of the matrix will be
      // distributed). for only one processor, we can directly assign the
      // columns as well. If we use a recent Trilinos version, we can also
      // require building a non-local graph which gives us thread-safe
      // initialization.
      if (row_map.Comm().NumProc() > 1)
        graph = std::make_unique<Epetra_FECrsGraph>(
          Copy, row_map, n_entries_per_row, false
          // TODO: Check which new Trilinos version supports this...
          // Remember to change tests/trilinos/assemble_matrix_parallel_07, too.
          // #if DEAL_II_TRILINOS_VERSION_GTE(11,14,0)
          //                 , true
          // #endif
        );
      else
        graph = std::make_unique<Epetra_FECrsGraph>(
          Copy, row_map, col_map, n_entries_per_row, false);
    }
 
 
 
    void
    reinit_sp(const Epetra_Map                   &row_map,
              const Epetra_Map                   &col_map,
              const std::vector<size_type>       &n_entries_per_row,
              std::unique_ptr<Epetra_Map>        &column_space_map,
              std::unique_ptr<Epetra_FECrsGraph> &graph,
              std::unique_ptr<Epetra_CrsGraph>   &nonlocal_graph)
    {
      Assert(row_map.IsOneToOne(),
             ExcMessage("Row map must be 1-to-1, i.e., no overlap between "
                        "the maps of different processors."));
      Assert(col_map.IsOneToOne(),
             ExcMessage("Column map must be 1-to-1, i.e., no overlap between "
                        "the maps of different processors."));
 
      // release memory before reallocation
      nonlocal_graph.reset();
      graph.reset();
      AssertDimension(n_entries_per_row.size(),
                      TrilinosWrappers::n_global_elements(row_map));
 
      column_space_map = std::make_unique<Epetra_Map>(col_map);
 
      // Translate the vector of row lengths into one that only stores
      // those entries that related to the locally stored rows of the matrix:
      std::vector<int> local_entries_per_row(
        TrilinosWrappers::max_my_gid(row_map) -
        TrilinosWrappers::min_my_gid(row_map));
      for (unsigned int i = 0; i < local_entries_per_row.size(); ++i)
        local_entries_per_row[i] =
          n_entries_per_row[TrilinosWrappers::min_my_gid(row_map) + i];
 
      AssertThrow(std::accumulate(local_entries_per_row.begin(),
                                  local_entries_per_row.end(),
                                  std::uint64_t(0)) <
                    static_cast<std::uint64_t>(std::numeric_limits<int>::max()),
                  ExcMessage("The TrilinosWrappers use Epetra internally which "
                             "uses 'signed int' to represent local indices. "
                             "Therefore, only 2,147,483,647 nonzero matrix "
                             "entries can be stored on a single process, "
                             "but you are requesting more than that. "
                             "If possible, use more MPI processes."));
 
      if (row_map.Comm().NumProc() > 1)
        graph = std::make_unique<Epetra_FECrsGraph>(
          Copy, row_map, local_entries_per_row.data(), false
          // TODO: Check which new Trilinos version supports this...
          // Remember to change tests/trilinos/assemble_matrix_parallel_07, too.
          // #if DEAL_II_TRILINOS_VERSION_GTE(11,14,0)
          //                , true
          // #endif
        );
      else
        graph = std::make_unique<Epetra_FECrsGraph>(
          Copy, row_map, col_map, local_entries_per_row.data(), false);
    }
 
 
 
    template <typename SparsityPatternType>
    void
    reinit_sp(const Epetra_Map                   &row_map,
              const Epetra_Map                   &col_map,
              const SparsityPatternType          &sp,
              const bool                          exchange_data,
              std::unique_ptr<Epetra_Map>        &column_space_map,
              std::unique_ptr<Epetra_FECrsGraph> &graph,
              std::unique_ptr<Epetra_CrsGraph>   &nonlocal_graph)
    {
      nonlocal_graph.reset();
      graph.reset();
 
      AssertDimension(sp.n_rows(),
                      TrilinosWrappers::n_global_elements(row_map));
      AssertDimension(sp.n_cols(),
                      TrilinosWrappers::n_global_elements(col_map));
 
      column_space_map = std::make_unique<Epetra_Map>(col_map);
 
      Assert(row_map.LinearMap() == true,
             ExcMessage(
               "This function only works if the row map is contiguous."));
 
      const size_type first_row = TrilinosWrappers::min_my_gid(row_map),
                      last_row  = TrilinosWrappers::max_my_gid(row_map) + 1;
      std::vector<int> n_entries_per_row(last_row - first_row);
 
      // Trilinos wants the row length as an int this is hopefully never going
      // to be a problem.
      for (size_type row = first_row; row < last_row; ++row)
        n_entries_per_row[row - first_row] =
          static_cast<int>(sp.row_length(row));
 
      AssertThrow(std::accumulate(n_entries_per_row.begin(),
                                  n_entries_per_row.end(),
                                  std::uint64_t(0)) <
                    static_cast<std::uint64_t>(std::numeric_limits<int>::max()),
                  ExcMessage("The TrilinosWrappers use Epetra internally which "
                             "uses 'signed int' to represent local indices. "
                             "Therefore, only 2,147,483,647 nonzero matrix "
                             "entries can be stored on a single process, "
                             "but you are requesting more than that. "
                             "If possible, use more MPI processes."));
 
      if (row_map.Comm().NumProc() > 1)
        graph = std::make_unique<Epetra_FECrsGraph>(Copy,
                                                    row_map,
                                                    n_entries_per_row.data(),
                                                    false);
      else
        graph = std::make_unique<Epetra_FECrsGraph>(
          Copy, row_map, col_map, n_entries_per_row.data(), false);
 
      AssertDimension(sp.n_rows(), n_global_rows(*graph));
 
      std::vector<TrilinosWrappers::types::int_type> row_indices;
 
      for (size_type row = first_row; row < last_row; ++row)
        {
          const TrilinosWrappers::types::int_type row_length =
            sp.row_length(row);
          if (row_length == 0)
            continue;
 
          row_indices.resize(row_length, -1);
          {
            typename SparsityPatternType::iterator p = sp.begin(row);
            // avoid incrementing p over the end of the current row because
            // it is slow for DynamicSparsityPattern in parallel
            for (int col = 0; col < row_length;)
              {
                row_indices[col++] = p->column();
                if (col < row_length)
                  ++p;
              }
          }
          if (exchange_data == false)
            graph->Epetra_CrsGraph::InsertGlobalIndices(row,
                                                        row_length,
                                                        row_indices.data());
          else
            {
              // Include possibility to exchange data since
              // DynamicSparsityPattern is able to do so
              const TrilinosWrappers::types::int_type trilinos_row = row;
              graph->InsertGlobalIndices(1,
                                         &trilinos_row,
                                         row_length,
                                         row_indices.data());
            }
        }
 
      // TODO A dynamic_cast fails here, this is suspicious.
      const auto &range_map =
        static_cast<const Epetra_Map &>(graph->RangeMap()); // NOLINT
      int ierr = graph->GlobalAssemble(*column_space_map, range_map, true);
      AssertThrow(ierr == 0, ExcTrilinosError(ierr));
 
      ierr = graph->OptimizeStorage();
      AssertThrow(ierr == 0, ExcTrilinosError(ierr));
    }
  } // namespace
 
 
 
  void
  SparsityPattern::reinit(const IndexSet &parallel_partitioning,
                          const MPI_Comm  communicator,
                          const size_type n_entries_per_row)
  {
    SparsityPatternBase::resize(parallel_partitioning.size(),
                                parallel_partitioning.size());
    Epetra_Map map =
      parallel_partitioning.make_trilinos_map(communicator, false);
    reinit_sp(
      map, map, n_entries_per_row, column_space_map, graph, nonlocal_graph);
  }
 
 
 
  void
  SparsityPattern::reinit(const IndexSet               &parallel_partitioning,
                          const MPI_Comm                communicator,
                          const std::vector<size_type> &n_entries_per_row)
  {
    SparsityPatternBase::resize(parallel_partitioning.size(),
                                parallel_partitioning.size());
    Epetra_Map map =
      parallel_partitioning.make_trilinos_map(communicator, false);
    reinit_sp(
      map, map, n_entries_per_row, column_space_map, graph, nonlocal_graph);
  }
 
 
 
  void
  SparsityPattern::reinit(const IndexSet &row_parallel_partitioning,
                          const IndexSet &col_parallel_partitioning,
                          const MPI_Comm  communicator,
                          const size_type n_entries_per_row)
  {
    SparsityPatternBase::resize(row_parallel_partitioning.size(),
                                col_parallel_partitioning.size());
    Epetra_Map row_map =
      row_parallel_partitioning.make_trilinos_map(communicator, false);
    Epetra_Map col_map =
      col_parallel_partitioning.make_trilinos_map(communicator, false);
    reinit_sp(row_map,
              col_map,
              n_entries_per_row,
              column_space_map,
              graph,
              nonlocal_graph);
  }
 
 
 
  void
  SparsityPattern::reinit(const IndexSet &row_parallel_partitioning,
                          const IndexSet &col_parallel_partitioning,
                          const MPI_Comm  communicator,
                          const std::vector<size_type> &n_entries_per_row)
  {
    SparsityPatternBase::resize(row_parallel_partitioning.size(),
                                col_parallel_partitioning.size());
    Epetra_Map row_map =
      row_parallel_partitioning.make_trilinos_map(communicator, false);
    Epetra_Map col_map =
      col_parallel_partitioning.make_trilinos_map(communicator, false);
    reinit_sp(row_map,
              col_map,
              n_entries_per_row,
              column_space_map,
              graph,
              nonlocal_graph);
  }
 
 
 
  void
  SparsityPattern::reinit(const IndexSet &row_parallel_partitioning,
                          const IndexSet &col_parallel_partitioning,
                          const IndexSet &writable_rows,
                          const MPI_Comm  communicator,
                          const size_type n_entries_per_row)
  {
    SparsityPatternBase::resize(row_parallel_partitioning.size(),
                                col_parallel_partitioning.size());
    Epetra_Map row_map =
      row_parallel_partitioning.make_trilinos_map(communicator, false);
    Epetra_Map col_map =
      col_parallel_partitioning.make_trilinos_map(communicator, false);
    reinit_sp(row_map,
              col_map,
              n_entries_per_row,
              column_space_map,
              graph,
              nonlocal_graph);
 
    IndexSet nonlocal_partitioner = writable_rows;
    AssertDimension(nonlocal_partitioner.size(),
                    row_parallel_partitioning.size());
    if constexpr (running_in_debug_mode())
      {
        {
          IndexSet tmp = writable_rows & row_parallel_partitioning;
          Assert(tmp == row_parallel_partitioning,
                 ExcMessage(
                   "The set of writable rows passed to this method does not "
                   "contain the locally owned rows, which is not allowed."));
        }
      }
    nonlocal_partitioner.subtract_set(row_parallel_partitioning);
    if (Utilities::MPI::n_mpi_processes(communicator) > 1)
      {
        Epetra_Map nonlocal_map =
          nonlocal_partitioner.make_trilinos_map(communicator, true);
        nonlocal_graph =
          std::make_unique<Epetra_CrsGraph>(Copy, nonlocal_map, 0);
      }
    else
      Assert(nonlocal_partitioner.n_elements() == 0, ExcInternalError());
  }
 
 
 
  template <typename SparsityPatternType>
  void
  SparsityPattern::reinit(
    const IndexSet            &row_parallel_partitioning,
    const IndexSet            &col_parallel_partitioning,
    const SparsityPatternType &nontrilinos_sparsity_pattern,
    const MPI_Comm             communicator,
    const bool                 exchange_data)
  {
    SparsityPatternBase::resize(row_parallel_partitioning.size(),
                                col_parallel_partitioning.size());
    Epetra_Map row_map =
      row_parallel_partitioning.make_trilinos_map(communicator, false);
    Epetra_Map col_map =
      col_parallel_partitioning.make_trilinos_map(communicator, false);
    reinit_sp(row_map,
              col_map,
              nontrilinos_sparsity_pattern,
              exchange_data,
              column_space_map,
              graph,
              nonlocal_graph);
  }
 
 
 
  template <typename SparsityPatternType>
  void
  SparsityPattern::reinit(
    const IndexSet            &parallel_partitioning,
    const SparsityPatternType &nontrilinos_sparsity_pattern,
    const MPI_Comm             communicator,
    const bool                 exchange_data)
  {
    AssertDimension(nontrilinos_sparsity_pattern.n_rows(),
                    parallel_partitioning.size());
    AssertDimension(nontrilinos_sparsity_pattern.n_cols(),
                    parallel_partitioning.size());
    SparsityPatternBase::resize(parallel_partitioning.size(),
                                parallel_partitioning.size());
    Epetra_Map map =
      parallel_partitioning.make_trilinos_map(communicator, false);
    reinit_sp(map,
              map,
              nontrilinos_sparsity_pattern,
              exchange_data,
              column_space_map,
              graph,
              nonlocal_graph);
  }
 
 
 
  SparsityPattern &
  SparsityPattern::operator=(const SparsityPattern &)
  {
    DEAL_II_NOT_IMPLEMENTED();
    return *this;
  }
 
 
 
  void
  SparsityPattern::copy_from(const SparsityPattern &sp)
  {
    SparsityPatternBase::resize(sp.n_rows(), sp.n_cols());
    column_space_map = std::make_unique<Epetra_Map>(*sp.column_space_map);
    graph            = std::make_unique<Epetra_FECrsGraph>(*sp.graph);
 
    if (sp.nonlocal_graph.get() != nullptr)
      nonlocal_graph = std::make_unique<Epetra_CrsGraph>(*sp.nonlocal_graph);
    else
      nonlocal_graph.reset();
  }
 
 
 
  template <typename SparsityPatternType>
  void
  SparsityPattern::copy_from(const SparsityPatternType &sp)
  {
    SparsityPatternBase::resize(sp.n_rows(), sp.n_cols());
    const Epetra_Map rows(TrilinosWrappers::types::int_type(sp.n_rows()),
                          0,
                          Utilities::Trilinos::comm_self());
    const Epetra_Map columns(TrilinosWrappers::types::int_type(sp.n_cols()),
                             0,
                             Utilities::Trilinos::comm_self());
 
    reinit_sp(
      rows, columns, sp, false, column_space_map, graph, nonlocal_graph);
  }
 
 
 
  void
  SparsityPattern::clear()
  {
    SparsityPatternBase::resize(0, 0);
    // When we clear the matrix, reset
    // the pointer and generate an
    // empty sparsity pattern.
    column_space_map =
      std::make_unique<Epetra_Map>(TrilinosWrappers::types::int_type(0),
                                   TrilinosWrappers::types::int_type(0),
                                   Utilities::Trilinos::comm_self());
    graph = std::make_unique<Epetra_FECrsGraph>(View,
                                                *column_space_map,
                                                *column_space_map,
                                                0);
    graph->FillComplete();
 
    nonlocal_graph.reset();
  }
 
 
 
  void
  SparsityPattern::compress()
  {
    int ierr;
    Assert(column_space_map.get(), ExcInternalError());
    if (nonlocal_graph.get() != nullptr)
      {
        if (nonlocal_graph->IndicesAreGlobal() == false &&
            nonlocal_graph->RowMap().NumMyElements() > 0 &&
            n_global_elements(*column_space_map) > 0)
          {
            // Insert dummy element at (row, column) that corresponds to row 0
            // in local index counting.
            TrilinosWrappers::types::int_type row =
              TrilinosWrappers::global_index(nonlocal_graph->RowMap(), 0);
            TrilinosWrappers::types::int_type column = 0;
 
            // in case we have a square sparsity pattern, add the entry on the
            // diagonal
            if (TrilinosWrappers::n_global_elements(*column_space_map) ==
                TrilinosWrappers::n_global_elements(graph->RangeMap()))
              column = row;
            // if not, take a column index that we have ourselves since we
            // know for sure it is there (and it will not create spurious
            // messages to many ranks like putting index 0 on many processors)
            else if (column_space_map->NumMyElements() > 0)
              column = TrilinosWrappers::global_index(*column_space_map, 0);
            ierr = nonlocal_graph->InsertGlobalIndices(row, 1, &column);
            AssertThrow(ierr == 0, ExcTrilinosError(ierr));
          }
        Assert(nonlocal_graph->RowMap().NumMyElements() == 0 ||
                 n_global_elements(*column_space_map) == 0 ||
                 nonlocal_graph->IndicesAreGlobal() == true,
               ExcInternalError());
 
        ierr =
          nonlocal_graph->FillComplete(*column_space_map, graph->RangeMap());
        AssertThrow(ierr >= 0, ExcTrilinosError(ierr));
        ierr = nonlocal_graph->OptimizeStorage();
        AssertThrow(ierr >= 0, ExcTrilinosError(ierr));
        Epetra_Export exporter(nonlocal_graph->RowMap(), graph->RowMap());
        ierr = graph->Export(*nonlocal_graph, exporter, Add);
        AssertThrow(ierr == 0, ExcTrilinosError(ierr));
        ierr = graph->FillComplete(*column_space_map, graph->RangeMap());
        AssertThrow(ierr == 0, ExcTrilinosError(ierr));
      }
    else
      {
        // TODO A dynamic_cast fails here, this is suspicious.
        const auto &range_map =
          static_cast<const Epetra_Map &>(graph->RangeMap()); // NOLINT
        ierr = graph->GlobalAssemble(*column_space_map, range_map, true);
        AssertThrow(ierr == 0, ExcTrilinosError(ierr));
      }
 
    try
      {
        ierr = graph->OptimizeStorage();
      }
    catch (const int error_code)
      {
        AssertThrow(
          false,
          ExcMessage(
            "The Epetra_CrsGraph::OptimizeStorage() function "
            "has thrown an error with code " +
            std::to_string(error_code) +
            ". You will have to look up the exact meaning of this error "
            "in the Trilinos source code, but oftentimes, this function "
            "throwing an error indicates that you are trying to allocate "
            "more than 2,147,483,647 nonzero entries in the sparsity "
            "pattern on the local process; this will not work because "
            "Epetra indexes entries with a simple 'signed int'."));
      }
    AssertThrow(ierr == 0, ExcTrilinosError(ierr));
 
    // Check consistency between the sizes set at the beginning and what
    // Trilinos stores:
    using namespace deal_II_exceptions::internals;
    Assert(compare_for_equality(n_rows(), n_global_rows(*graph)),
           ExcInternalError());
    Assert(compare_for_equality(n_cols(), n_global_cols(*graph)),
           ExcInternalError());
  }
 
 
 
  bool
  SparsityPattern::row_is_stored_locally(const size_type i) const
  {
    return graph->RowMap().LID(
             static_cast<TrilinosWrappers::types::int_type>(i)) != -1;
  }
 
 
 
  bool
  SparsityPattern::exists(const size_type i, const size_type j) const
  {
    if (!row_is_stored_locally(i))
      {
        return false;
      }
    else
      {
        // Extract local indices in
        // the matrix.
        int trilinos_i =
              graph->LRID(static_cast<TrilinosWrappers::types::int_type>(i)),
            trilinos_j =
              graph->LCID(static_cast<TrilinosWrappers::types::int_type>(j));
 
        // Check whether the matrix
        // already is transformed to
        // local indices.
        if (graph->Filled() == false)
          {
            int nnz_present = graph->NumGlobalIndices(i);
            int nnz_extracted;
            TrilinosWrappers::types::int_type *col_indices;
 
            // Generate the view and make
            // sure that we have not generated
            // an error.
            // TODO: trilinos_i is the local row index -> it is an int but
            // ExtractGlobalRowView requires trilinos_i to be the global row
            // index and thus it should be a long long int
            int ierr = graph->ExtractGlobalRowView(trilinos_i,
                                                   nnz_extracted,
                                                   col_indices);
            Assert(ierr == 0, ExcTrilinosError(ierr));
            Assert(nnz_present == nnz_extracted,
                   ExcDimensionMismatch(nnz_present, nnz_extracted));
 
            // Search the index
            const std::ptrdiff_t local_col_index =
              std::find(col_indices, col_indices + nnz_present, trilinos_j) -
              col_indices;
 
            if (local_col_index == nnz_present)
              return false;
          }
        else
          {
            // Prepare pointers for extraction
            // of a view of the row.
            int  nnz_present = graph->NumGlobalIndices(i);
            int  nnz_extracted;
            int *col_indices;
 
            // Generate the view and make
            // sure that we have not generated
            // an error.
            int ierr =
              graph->ExtractMyRowView(trilinos_i, nnz_extracted, col_indices);
            Assert(ierr == 0, ExcTrilinosError(ierr));
 
            Assert(nnz_present == nnz_extracted,
                   ExcDimensionMismatch(nnz_present, nnz_extracted));
 
            // Search the index
            const std::ptrdiff_t local_col_index =
              std::find(col_indices, col_indices + nnz_present, trilinos_j) -
              col_indices;
 
            if (local_col_index == nnz_present)
              return false;
          }
      }
 
    return true;
  }
 
 
 
  SparsityPattern::size_type
  SparsityPattern::bandwidth() const
  {
    size_type local_b = 0;
    for (int i = 0; i < static_cast<int>(local_size()); ++i)
      {
        int *indices;
        int  num_entries;
        graph->ExtractMyRowView(i, num_entries, indices);
        for (unsigned int j = 0; j < static_cast<unsigned int>(num_entries);
             ++j)
          {
            if (static_cast<size_type>(std::abs(i - indices[j])) > local_b)
              local_b = std::abs(i - indices[j]);
          }
      }
 
    TrilinosWrappers::types::int_type global_b = 0;
    graph->Comm().MaxAll(reinterpret_cast<TrilinosWrappers::types::int_type *>(
                           &local_b),
                         &global_b,
                         1);
    return static_cast<size_type>(global_b);
  }
 
 
 
  unsigned int
  SparsityPattern::local_size() const
  {
    return graph->NumMyRows();
  }
 
 
 
  std::pair<SparsityPattern::size_type, SparsityPattern::size_type>
  SparsityPattern::local_range() const
  {
    const size_type begin = TrilinosWrappers::min_my_gid(graph->RowMap());
    const size_type end   = TrilinosWrappers::max_my_gid(graph->RowMap()) + 1;
 
    return {begin, end};
  }
 
 
 
  std::uint64_t
  SparsityPattern::n_nonzero_elements() const
  {
    return n_global_entries(*graph);
  }
 
 
 
  unsigned int
  SparsityPattern::max_entries_per_row() const
  {
    return graph->MaxNumIndices();
  }
 
 
 
  SparsityPattern::size_type
  SparsityPattern::row_length(const size_type row) const
  {
    Assert(row < n_rows(), ExcInternalError());
 
    // Get a representation of the where the present row is located on
    // the current processor
    TrilinosWrappers::types::int_type local_row =
      graph->LRID(static_cast<TrilinosWrappers::types::int_type>(row));
 
    // On the processor who owns this row, we'll have a non-negative
    // value for `local_row` and can ask for the length of the row.
    if (local_row >= 0)
      return graph->NumMyIndices(local_row);
    else
      return static_cast<size_type>(-1);
  }
 
 
 
  void
  SparsityPattern::add_row_entries(const size_type                  &row,
                                   const ArrayView<const size_type> &columns,
                                   const bool indices_are_sorted)
  {
    add_entries(row, columns.begin(), columns.end(), indices_are_sorted);
  }
 
 
 
  const Epetra_Map &
  SparsityPattern::domain_partitioner() const
  {
    // TODO A dynamic_cast fails here, this is suspicious.
    const auto &domain_map =
      static_cast<const Epetra_Map &>(graph->DomainMap()); // NOLINT
    return domain_map;
  }
 
 
 
  const Epetra_Map &
  SparsityPattern::range_partitioner() const
  {
    // TODO A dynamic_cast fails here, this is suspicious.
    const auto &range_map =
      static_cast<const Epetra_Map &>(graph->RangeMap()); // NOLINT
    return range_map;
  }
 
 
 
  MPI_Comm
  SparsityPattern::get_mpi_communicator() const
  {
    const Epetra_MpiComm *mpi_comm =
      dynamic_cast<const Epetra_MpiComm *>(&graph->RangeMap().Comm());
    Assert(mpi_comm != nullptr, ExcInternalError());
    return mpi_comm->Comm();
  }
 
 
 
  void
  SparsityPattern::write_ascii()
  {
    DEAL_II_NOT_IMPLEMENTED();
  }
 
 
 
  // As of now, no particularly neat
  // output is generated in case of
  // multiple processors.
  void
  SparsityPattern::print(std::ostream &out,
                         const bool    write_extended_trilinos_info) const
  {
    if (write_extended_trilinos_info)
      out << *graph;
    else
      {
        int *indices;
        int  num_entries;
 
        for (int i = 0; i < graph->NumMyRows(); ++i)
          {
            graph->ExtractMyRowView(i, num_entries, indices);
            for (int j = 0; j < num_entries; ++j)
              out << "(" << TrilinosWrappers::global_index(graph->RowMap(), i)
                  << ","
                  << TrilinosWrappers::global_index(graph->ColMap(), indices[j])
                  << ") " << std::endl;
          }
      }
 
    AssertThrow(out.fail() == false, ExcIO());
  }
 
 
 
  void
  SparsityPattern::print_gnuplot(std::ostream &out) const
  {
    Assert(graph->Filled() == true, ExcInternalError());
    for (::types::global_dof_index row = 0; row < local_size(); ++row)
      {
        int *indices;
        int  num_entries;
        graph->ExtractMyRowView(row, num_entries, indices);
 
        Assert(num_entries >= 0, ExcInternalError());
        // avoid sign comparison warning
        const ::types::global_dof_index num_entries_ = num_entries;
        for (::types::global_dof_index j = 0; j < num_entries_; ++j)
          // while matrix entries are usually
          // written (i,j), with i vertical and
          // j horizontal, gnuplot output is
          // x-y, that is we have to exchange
          // the order of output
          out << static_cast<int>(
                   TrilinosWrappers::global_index(graph->ColMap(), indices[j]))
              << " "
              << -static_cast<int>(
                   TrilinosWrappers::global_index(graph->RowMap(), row))
              << std::endl;
      }
 
    AssertThrow(out.fail() == false, ExcIO());
  }
 
  // TODO: Implement!
  std::size_t
  SparsityPattern::memory_consumption() const
  {
    DEAL_II_NOT_IMPLEMENTED();
    return 0;
  }
 
 
#  ifndef DOXYGEN
  // explicit instantiations
  //
  template void
  SparsityPattern::copy_from(const ::SparsityPattern &);
  template void
  SparsityPattern::copy_from(const ::DynamicSparsityPattern &);
 
  template void
  SparsityPattern::reinit(const IndexSet &,
                          const ::SparsityPattern &,
                          const MPI_Comm,
                          bool);
  template void
  SparsityPattern::reinit(const IndexSet &,
                          const ::DynamicSparsityPattern &,
                          const MPI_Comm,
                          bool);
 
 
  template void
  SparsityPattern::reinit(const IndexSet &,
                          const IndexSet &,
                          const ::SparsityPattern &,
                          const MPI_Comm,
                          bool);
  template void
  SparsityPattern::reinit(const IndexSet &,
                          const IndexSet &,
                          const ::DynamicSparsityPattern &,
                          const MPI_Comm,
                          bool);
#  endif
 
} // namespace TrilinosWrappers
 
DEAL_II_NAMESPACE_CLOSE
 
#endif // DEAL_II_WITH_TRILINOS