Tpetra_CrsMatrix_def.hpp

Go to the documentation of this file.

// @HEADER
// ***********************************************************************
//
//          Tpetra: Templated Linear Algebra Services Package
//                 Copyright (2008) Sandia Corporation
//
// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
// the U.S. Government retains certain rights in this software.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in the
// documentation and/or other materials provided with the distribution.
//
// 3. Neither the name of the Corporation nor the names of the
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// Questions? Contact Michael A. Heroux (maherou@sandia.gov)
//
// ************************************************************************
// @HEADER
 
#ifndef TPETRA_CRSMATRIX_DEF_HPP
#define TPETRA_CRSMATRIX_DEF_HPP
 
 
#include "Tpetra_RowMatrix.hpp"
#include "Tpetra_Import_Util.hpp"
#include "Tpetra_Import_Util2.hpp"
#include "Tpetra_Details_Behavior.hpp"
#include "Tpetra_Details_castAwayConstDualView.hpp"
#include "Tpetra_Details_computeOffsets.hpp"
#include "Tpetra_Details_copyOffsets.hpp"
#include "Tpetra_Details_createMirrorView.hpp"
#include "Tpetra_Details_getDiagCopyWithoutOffsets.hpp"
#include "Tpetra_Details_gathervPrint.hpp"
#include "Tpetra_Details_leftScaleLocalCrsMatrix.hpp"
#include "Tpetra_Details_Profiling.hpp"
#include "Tpetra_Details_rightScaleLocalCrsMatrix.hpp"
 
//#include "Tpetra_Util.hpp" // comes in from Tpetra_CrsGraph_decl.hpp
#include "Teuchos_SerialDenseMatrix.hpp"
#include "KokkosSparse_getDiagCopy.hpp"
#include "Tpetra_Details_copyConvert.hpp"
#include "Tpetra_Details_getEntryOnHost.hpp"
#include "Tpetra_Details_packCrsMatrix.hpp"
#include "Tpetra_Details_unpackCrsMatrixAndCombine.hpp"
#include <memory>
#include <sstream>
#include <typeinfo>
#include <vector>
 
namespace Tpetra {
 
namespace { // (anonymous)
 
  template<class T, class BinaryFunction>
  T atomic_binary_function_update (volatile T* const dest,
                                   const T& inputVal,
                                   BinaryFunction f)
  {
    T oldVal = *dest;
    T assume;
 
    // NOTE (mfh 30 Nov 2015) I do NOT need a fence here for IBM
    // POWER architectures, because 'newval' depends on 'assume',
    // which depends on 'oldVal', which depends on '*dest'.  This
    // sets up a chain of read dependencies that should ensure
    // correct behavior given a sane memory model.
    do {
      assume = oldVal;
      T newVal = f (assume, inputVal);
      oldVal = Kokkos::atomic_compare_exchange (dest, assume, newVal);
    } while (assume != oldVal);
 
    return oldVal;
  }
} // namespace (anonymous)
 
//
// Users must never rely on anything in the Details namespace.
//
namespace Details {
 
template<class Scalar>
struct AbsMax {
  Scalar operator() (const Scalar& x, const Scalar& y) {
    typedef Teuchos::ScalarTraits<Scalar> STS;
    return std::max (STS::magnitude (x), STS::magnitude (y));
  }
};
 
} // namespace Details
} // namespace Tpetra
 
namespace Tpetra {
namespace Classes {
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const Teuchos::RCP<const map_type>& rowMap,
             size_t maxNumEntriesPerRow,
             ProfileType pftype,
             const Teuchos::RCP<Teuchos::ParameterList>& params) :
    dist_object_type (rowMap),
    storageStatus_ (pftype == StaticProfile ?
                    ::Tpetra::Details::STORAGE_1D_UNPACKED :
                    ::Tpetra::Details::STORAGE_2D),
    fillComplete_ (false),
    frobNorm_ (-STM::one ())
  {
    const char tfecfFuncName[] = "CrsMatrix(RCP<const Map>, size_t, "
      "ProfileType[, RCP<ParameterList>]): ";
    Teuchos::RCP<crs_graph_type> graph;
    try {
      graph = Teuchos::rcp (new crs_graph_type (rowMap, maxNumEntriesPerRow,
                                                pftype, params));
    }
    catch (std::exception& e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "CrsGraph constructor (RCP<const Map>, "
         "size_t, ProfileType[, RCP<ParameterList>]) threw an exception: "
         << e.what ());
    }
    // myGraph_ not null means that the matrix owns the graph.  That's
    // different than the const CrsGraph constructor, where the matrix
    // does _not_ own the graph.
    myGraph_ = graph;
    staticGraph_ = myGraph_;
    resumeFill (params);
    checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const Teuchos::RCP<const map_type>& rowMap,
             const Teuchos::ArrayRCP<const size_t>& NumEntriesPerRowToAlloc,
             ProfileType pftype,
             const Teuchos::RCP<Teuchos::ParameterList>& params) :
    dist_object_type (rowMap),
    storageStatus_ (pftype == StaticProfile ?
                    ::Tpetra::Details::STORAGE_1D_UNPACKED :
                    ::Tpetra::Details::STORAGE_2D),
    fillComplete_ (false),
    frobNorm_ (-STM::one ())
  {
    const char tfecfFuncName[] = "CrsMatrix(RCP<const Map>, "
      "ArrayRCP<const size_t>, ProfileType[, RCP<ParameterList>]): ";
    Teuchos::RCP<crs_graph_type> graph;
    try {
      graph = Teuchos::rcp (new crs_graph_type (rowMap, NumEntriesPerRowToAlloc,
                                                pftype, params));
    }
    catch (std::exception &e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "CrsGraph constructor (RCP<const Map>, "
         "ArrayRCP<const size_t>, ProfileType[, RCP<ParameterList>]) threw "
         "an exception: " << e.what ());
    }
    // myGraph_ not null means that the matrix owns the graph.  That's
    // different than the const CrsGraph constructor, where the matrix
    // does _not_ own the graph.
    myGraph_ = graph;
    staticGraph_ = graph;
    resumeFill (params);
    checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const Teuchos::RCP<const map_type>& rowMap,
             const Teuchos::RCP<const map_type>& colMap,
             size_t maxNumEntriesPerRow,
             ProfileType pftype,
             const Teuchos::RCP<Teuchos::ParameterList>& params) :
    dist_object_type (rowMap),
    storageStatus_ (pftype == StaticProfile ?
                    ::Tpetra::Details::STORAGE_1D_UNPACKED :
                    ::Tpetra::Details::STORAGE_2D),
    fillComplete_ (false),
    frobNorm_ (-STM::one ())
  {
    const char tfecfFuncName[] = "CrsMatrix(RCP<const Map>, RCP<const Map>, "
      "size_t, ProfileType[, RCP<ParameterList>]): ";
 
#ifdef HAVE_TPETRA_DEBUG
    // An artifact of debugging something a while back.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! staticGraph_.is_null (), std::logic_error,
      "staticGraph_ is not null at the beginning of the constructor.  "
      "Please report this bug to the Tpetra developers.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! myGraph_.is_null (), std::logic_error,
      "myGraph_ is not null at the beginning of the constructor.  "
      "Please report this bug to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
 
    Teuchos::RCP<crs_graph_type> graph;
    try {
      graph = Teuchos::rcp (new crs_graph_type (rowMap, colMap,
                                                maxNumEntriesPerRow,
                                                pftype, params));
    }
    catch (std::exception &e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "CrsGraph constructor (RCP<const Map>, "
         "RCP<const Map>, size_t, ProfileType[, RCP<ParameterList>]) threw an "
         "exception: " << e.what ());
    }
    // myGraph_ not null means that the matrix owns the graph.  That's
    // different than the const CrsGraph constructor, where the matrix
    // does _not_ own the graph.
    myGraph_ = graph;
    staticGraph_ = myGraph_;
    resumeFill (params);
    checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const Teuchos::RCP<const map_type>& rowMap,
             const Teuchos::RCP<const map_type>& colMap,
             const Teuchos::ArrayRCP<const size_t>& numEntPerRow,
             ProfileType pftype,
             const Teuchos::RCP<Teuchos::ParameterList>& params) :
    dist_object_type (rowMap),
    storageStatus_ (pftype == StaticProfile ?
                    ::Tpetra::Details::STORAGE_1D_UNPACKED :
                    ::Tpetra::Details::STORAGE_2D),
    fillComplete_ (false),
    frobNorm_ (-STM::one ())
  {
    const char tfecfFuncName[] = "CrsMatrix(RCP<const Map>, RCP<const Map>, "
      "ArrayRCP<const size_t>, ProfileType[, RCP<ParameterList>]): ";
    Teuchos::RCP<crs_graph_type> graph;
    try {
      graph = Teuchos::rcp (new crs_graph_type (rowMap, colMap, numEntPerRow,
                                                pftype, params));
    }
    catch (std::exception &e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "CrsGraph constructor (RCP<const Map>, "
         "RCP<const Map>, ArrayRCP<const size_t>, ProfileType[, "
         "RCP<ParameterList>]) threw an exception: " << e.what ());
    }
    // myGraph_ not null means that the matrix owns the graph.  That's
    // different than the const CrsGraph constructor, where the matrix
    // does _not_ own the graph.
    myGraph_ = graph;
    staticGraph_ = graph;
    resumeFill (params);
    checkInternalState ();
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const Teuchos::RCP<const crs_graph_type>& graph,
             const Teuchos::RCP<Teuchos::ParameterList>& /* params */) :
    dist_object_type (graph->getRowMap ()),
    staticGraph_ (graph),
    storageStatus_ (::Tpetra::Details::STORAGE_1D_PACKED),
    fillComplete_ (false),
    frobNorm_ (-STM::one ())
  {
    typedef typename local_matrix_type::values_type values_type;
    const char tfecfFuncName[] = "CrsMatrix(RCP<const CrsGraph>[, "
      "RCP<ParameterList>]): ";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (graph.is_null (), std::runtime_error, "Input graph is null.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! graph->isFillComplete (), std::runtime_error, "Input graph is not "
       "fill complete. You must call fillComplete on the graph before using "
       "it to construct a CrsMatrix.  Note that calling resumeFill on the "
       "graph makes it not fill complete, even if you had previously called "
       "fillComplete.  In that case, you must call fillComplete on the graph "
       "again.");
 
    // The graph is fill complete, so it is locally indexed and has a
    // fixed structure.  This means we can allocate the (1-D) array of
    // values and build the local matrix right now.  Note that the
    // local matrix's number of columns comes from the column Map, not
    // the domain Map.
 
    const size_t numCols = graph->getColMap ()->getNodeNumElements ();
    auto lclGraph = graph->getLocalGraph ();
    const size_t numEnt = lclGraph.entries.extent (0);
    values_type val ("Tpetra::CrsMatrix::val", numEnt);
 
    this->lclMatrix_ = local_matrix_type ("Tpetra::CrsMatrix::lclMatrix_",
                                          numCols, val, lclGraph);
    // FIXME (22 Jun 2016) I would very much like to get rid of
    // k_values1D_ at some point.  I find it confusing to have all
    // these extra references lying around.
    this->k_values1D_ = this->lclMatrix_.values;
 
    checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const Teuchos::RCP<const map_type>& rowMap,
             const Teuchos::RCP<const map_type>& colMap,
             const typename local_matrix_type::row_map_type& rowPointers,
             const typename local_graph_type::entries_type::non_const_type& columnIndices,
             const typename local_matrix_type::values_type& values,
             const Teuchos::RCP<Teuchos::ParameterList>& params) :
    dist_object_type (rowMap),
    storageStatus_ (::Tpetra::Details::STORAGE_1D_PACKED),
    fillComplete_ (false),
    frobNorm_ (-STM::one ())
  {
    using Teuchos::RCP;
    const char tfecfFuncName[] = "Tpetra::CrsMatrix(RCP<const Map>, "
      "RCP<const Map>, ptr, ind, val[, params]): ";
    const char suffix[] = ".  Please report this bug to the Tpetra developers.";
 
    // Check the user's input.  Note that this might throw only on
    // some processes but not others, causing deadlock.  We prefer
    // deadlock due to exceptions to segfaults, because users can
    // catch exceptions.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (values.extent (0) != columnIndices.extent (0),
       std::invalid_argument, "Input arrays don't have matching dimensions.  "
       "values.extent(0) = " << values.extent (0) << " != "
       "columnIndices.extent(0) = " << columnIndices.extent (0) << ".");
#ifdef HAVE_TPETRA_DEBUG
    if (rowPointers.extent (0) != 0) {
      const size_t numEnt =
        ::Tpetra::Details::getEntryOnHost (rowPointers, rowPointers.extent (0) - 1);
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (numEnt != static_cast<size_t> (columnIndices.extent (0)) ||
         numEnt != static_cast<size_t> (values.extent (0)),
         std::invalid_argument, "Last entry of rowPointers says that the matrix"
         " has " << numEnt << " entr" << (numEnt != 1 ? "ies" : "y") << ", but "
         "the dimensions of columnIndices and values don't match this.  "
         "columnIndices.extent(0) = " << columnIndices.extent (0) <<
         " and values.extent(0) = " << values.extent (0) << ".");
    }
#endif // HAVE_TPETRA_DEBUG
 
    RCP<crs_graph_type> graph;
    try {
      graph = Teuchos::rcp (new crs_graph_type (rowMap, colMap, rowPointers,
                                                columnIndices, params));
    }
    catch (std::exception& e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "CrsGraph constructor (RCP<const Map>, "
         "RCP<const Map>, ptr, ind[, params]) threw an exception: "
         << e.what ());
    }
    // The newly created CrsGraph _must_ have a local graph at this
    // point.  We don't really care whether CrsGraph's constructor
    // deep-copies or shallow-copies the input, but the dimensions
    // have to be right.  That's how we tell whether the CrsGraph has
    // a local graph.
    auto lclGraph = graph->getLocalGraph ();
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (lclGraph.row_map.extent (0) != rowPointers.extent (0) ||
       lclGraph.entries.extent (0) != columnIndices.extent (0),
       std::logic_error, "CrsGraph's constructor (rowMap, colMap, ptr, "
       "ind[, params]) did not set the local graph correctly." << suffix);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (lclGraph.entries.extent (0) != values.extent (0),
       std::logic_error, "CrsGraph's constructor (rowMap, colMap, ptr, ind[, "
       "params]) did not set the local graph correctly.  "
       "lclGraph.entries.extent(0) = " << lclGraph.entries.extent (0)
       << " != values.extent(0) = " << values.extent (0) << suffix);
 
    // myGraph_ not null means that the matrix owns the graph.  This
    // is true because the column indices come in as nonconst,
    // implying shared ownership.
    myGraph_ = graph;
    staticGraph_ = graph;
 
    // The graph may not be fill complete yet.  However, it is locally
    // indexed (since we have a column Map) and has a fixed structure
    // (due to the input arrays).  This means we can allocate the
    // (1-D) array of values and build the local matrix right now.
    // Note that the local matrix's number of columns comes from the
    // column Map, not the domain Map.
 
    const size_t numCols = graph->getColMap ()->getNodeNumElements ();
    lclMatrix_ = local_matrix_type ("Tpetra::CrsMatrix::lclMatrix_",
                                    numCols, values, lclGraph);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (lclMatrix_.values.extent (0) != values.extent (0),
       std::logic_error, "Local matrix's constructor did not set the values "
       "correctly.  lclMatrix_.values.extent(0) = " <<
       lclMatrix_.values.extent (0) << " != values.extent(0) = " <<
       values.extent (0) << suffix);
 
    // FIXME (22 Jun 2016) I would very much like to get rid of
    // k_values1D_ at some point.  I find it confusing to have all
    // these extra references lying around.
    this->k_values1D_ = this->lclMatrix_.values;
 
    checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const Teuchos::RCP<const map_type>& rowMap,
             const Teuchos::RCP<const map_type>& colMap,
             const Teuchos::ArrayRCP<size_t>& ptr,
             const Teuchos::ArrayRCP<LocalOrdinal>& ind,
             const Teuchos::ArrayRCP<Scalar>& val,
             const Teuchos::RCP<Teuchos::ParameterList>& params) :
    dist_object_type (rowMap),
    storageStatus_ (::Tpetra::Details::STORAGE_1D_PACKED),
    fillComplete_ (false),
    frobNorm_ (-STM::one ())
  {
    using Kokkos::Compat::getKokkosViewDeepCopy;
    using Teuchos::av_reinterpret_cast;
    using Teuchos::RCP;
    typedef typename local_matrix_type::values_type values_type;
    typedef impl_scalar_type IST;
    const char tfecfFuncName[] = "Tpetra::CrsMatrix(RCP<const Map>, "
      "RCP<const Map>, ptr, ind, val[, params]): ";
 
    RCP<crs_graph_type> graph;
    try {
      graph = Teuchos::rcp (new crs_graph_type (rowMap, colMap, ptr,
                                                ind, params));
    }
    catch (std::exception& e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "CrsGraph constructor (RCP<const Map>, "
         "RCP<const Map>, ArrayRCP<size_t>, ArrayRCP<LocalOrdinal>[, "
         "RCP<ParameterList>]) threw an exception: " << e.what ());
    }
    // myGraph_ not null means that the matrix owns the graph.  This
    // is true because the column indices come in as nonconst,
    // implying shared ownership.
    myGraph_ = graph;
    staticGraph_ = graph;
 
    // The graph may not be fill complete yet.  However, it is locally
    // indexed (since we have a column Map) and has a fixed structure
    // (due to the input arrays).  This means we can allocate the
    // (1-D) array of values and build the local matrix right now.
    // Note that the local matrix's number of columns comes from the
    // column Map, not the domain Map.
 
    // The graph _must_ have a local graph at this point.  We don't
    // really care whether CrsGraph's constructor deep-copies or
    // shallow-copies the input, but the dimensions have to be right.
    // That's how we tell whether the CrsGraph has a local graph.
    auto lclGraph = staticGraph_->getLocalGraph ();
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (static_cast<size_t> (lclGraph.row_map.extent (0)) != static_cast<size_t> (ptr.size ()) ||
       static_cast<size_t> (lclGraph.entries.extent (0)) != static_cast<size_t> (ind.size ()),
       std::logic_error, "CrsGraph's constructor (rowMap, colMap, ptr, "
       "ind[, params]) did not set the local graph correctly.  Please "
       "report this bug to the Tpetra developers.");
 
    const size_t numCols = staticGraph_->getColMap ()->getNodeNumElements ();
    values_type valIn = getKokkosViewDeepCopy<device_type> (av_reinterpret_cast<IST> (val ()));
    this->lclMatrix_ = local_matrix_type ("Tpetra::CrsMatrix::lclMatrix_",
                                          numCols, valIn, lclGraph);
    // FIXME (22 Jun 2016) I would very much like to get rid of
    // k_values1D_ at some point.  I find it confusing to have all
    // these extra references lying around.
    this->k_values1D_ = this->lclMatrix_.values;
 
    checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const Teuchos::RCP<const map_type>& rowMap,
             const Teuchos::RCP<const map_type>& colMap,
             const local_matrix_type& lclMatrix,
             const Teuchos::RCP<Teuchos::ParameterList>& params) :
    dist_object_type (rowMap),
    lclMatrix_ (lclMatrix),
    k_values1D_ (lclMatrix.values),
    storageStatus_ (::Tpetra::Details::STORAGE_1D_PACKED),
    fillComplete_ (true),
    frobNorm_ (-STM::one ())
  {
    const char tfecfFuncName[] = "Tpetra::CrsMatrix(RCP<const Map>, "
      "RCP<const Map>, local_matrix_type[, RCP<ParameterList>]): ";
    Teuchos::RCP<crs_graph_type> graph;
    try {
      graph = Teuchos::rcp (new crs_graph_type (rowMap, colMap,
                                                lclMatrix.graph, params));
    }
    catch (std::exception& e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "CrsGraph constructor (RCP<const Map>, "
         "RCP<const Map>, local_graph_type[, RCP<ParameterList>]) threw an "
         "exception: " << e.what ());
    }
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (!graph->isFillComplete (), std::logic_error, "CrsGraph constructor (RCP"
       "<const Map>, RCP<const Map>, local_graph_type[, RCP<ParameterList>]) "
       "did not produce a fill-complete graph.  Please report this bug to the "
       "Tpetra developers.");
    // myGraph_ not null means that the matrix owns the graph.  This
    // is true because the column indices come in as nonconst through
    // the matrix, implying shared ownership.
    myGraph_ = graph;
    staticGraph_ = graph;
 
    const bool callComputeGlobalConstants = params.get () == nullptr ||
      params->get ("compute global constants", true);
    if (callComputeGlobalConstants) {
      this->computeGlobalConstants ();
    }
 
    // Sanity checks at the end.
#ifdef HAVE_TPETRA_DEBUG
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(isFillActive (), std::logic_error,
      "We're at the end of fillComplete(), but isFillActive() is true.  "
      "Please report this bug to the Tpetra developers.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(! isFillComplete (), std::logic_error,
      "We're at the end of fillComplete(), but isFillComplete() is false.  "
      "Please report this bug to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
    checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  CrsMatrix (const local_matrix_type& lclMatrix,
             const Teuchos::RCP<const map_type>& rowMap,
             const Teuchos::RCP<const map_type>& colMap,
             const Teuchos::RCP<const map_type>& domainMap,
             const Teuchos::RCP<const map_type>& rangeMap,
             const Teuchos::RCP<Teuchos::ParameterList>& params) :
    dist_object_type (rowMap),
    lclMatrix_ (lclMatrix),
    k_values1D_ (lclMatrix.values),
    storageStatus_ (::Tpetra::Details::STORAGE_1D_PACKED),
    fillComplete_ (true),
    frobNorm_ (-STM::one ())
  {
    const char tfecfFuncName[] = "Tpetra::CrsMatrix(RCP<const Map>, "
      "RCP<const Map>, RCP<const Map>, RCP<const Map>, local_matrix_type[, "
      "RCP<ParameterList>]): ";
    Teuchos::RCP<crs_graph_type> graph;
    try {
      graph = Teuchos::rcp (new crs_graph_type (lclMatrix.graph, rowMap, colMap,
                                                domainMap, rangeMap, params));
    }
    catch (std::exception& e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "CrsGraph constructor (RCP<const Map>, "
         "RCP<const Map>, RCP<const Map>, RCP<const Map>, local_graph_type[, "
         "RCP<ParameterList>]) threw an exception: " << e.what ());
    }
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (!graph->isFillComplete (), std::logic_error, "CrsGraph constructor (RCP"
       "<const Map>, RCP<const Map>, RCP<const Map>, RCP<const Map>, local_graph_type[, "
       "RCP<ParameterList>]) did not produce a fill-complete graph.  Please report this "
       "bug to the Tpetra developers.");
    // myGraph_ not null means that the matrix owns the graph.  This
    // is true because the column indices come in as nonconst through
    // the matrix, implying shared ownership.
    myGraph_ = graph;
    staticGraph_ = graph;
 
    const bool callComputeGlobalConstants = params.get () == nullptr ||
      params->get ("compute global constants", true);
    if (callComputeGlobalConstants) {
      this->computeGlobalConstants ();
    }
 
    // Sanity checks at the end.
#ifdef HAVE_TPETRA_DEBUG
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(isFillActive (), std::logic_error,
      "We're at the end of fillComplete(), but isFillActive() is true.  "
      "Please report this bug to the Tpetra developers.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(! isFillComplete (), std::logic_error,
      "We're at the end of fillComplete(), but isFillComplete() is false.  "
      "Please report this bug to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
    checkInternalState ();
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  ~CrsMatrix ()
  {}
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<const Teuchos::Comm<int> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getComm () const {
    return getCrsGraphRef ().getComm ();
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<Node>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNode () const {
    return getCrsGraphRef ().getNode ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  ProfileType
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getProfileType () const {
    return this->getCrsGraphRef ().getProfileType ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isFillComplete () const {
    return fillComplete_;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isFillActive () const {
    return ! fillComplete_;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isStorageOptimized () const {
    return this->getCrsGraphRef ().isStorageOptimized ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isLocallyIndexed () const {
    return getCrsGraphRef ().isLocallyIndexed ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isGloballyIndexed () const {
    return getCrsGraphRef ().isGloballyIndexed ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  hasColMap () const {
    return getCrsGraphRef ().hasColMap ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  global_size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGlobalNumEntries () const {
    return getCrsGraphRef ().getGlobalNumEntries ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNodeNumEntries () const {
    return getCrsGraphRef ().getNodeNumEntries ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  global_size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGlobalNumRows () const {
    return getCrsGraphRef ().getGlobalNumRows ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  global_size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGlobalNumCols () const {
    return getCrsGraphRef ().getGlobalNumCols ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNodeNumRows () const {
    return getCrsGraphRef ().getNodeNumRows ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNodeNumCols () const {
    return getCrsGraphRef ().getNodeNumCols ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  global_size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGlobalNumDiagsImpl () const {
    const crs_graph_type& G = this->getCrsGraphRef ();
    using HDM = ::Tpetra::Details::HasDeprecatedMethods2630_WarningThisClassIsNotForUsers;
    return dynamic_cast<const HDM&> (G).getGlobalNumDiagsImpl ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  global_size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGlobalNumDiags () const {
    return this->getGlobalNumDiagsImpl ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNodeNumDiagsImpl () const {
    const crs_graph_type& G = this->getCrsGraphRef ();
    using HDM = ::Tpetra::Details::HasDeprecatedMethods2630_WarningThisClassIsNotForUsers;
    return dynamic_cast<const HDM&> (G).getNodeNumDiagsImpl ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNodeNumDiags () const {
    return this->getNodeNumDiagsImpl ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNumEntriesInGlobalRow (GlobalOrdinal globalRow) const {
    return getCrsGraphRef ().getNumEntriesInGlobalRow (globalRow);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNumEntriesInLocalRow (LocalOrdinal localRow) const {
    return getCrsGraphRef ().getNumEntriesInLocalRow (localRow);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGlobalMaxNumRowEntries () const {
    return getCrsGraphRef ().getGlobalMaxNumRowEntries ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getNodeMaxNumRowEntries () const {
    return getCrsGraphRef ().getNodeMaxNumRowEntries ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  GlobalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getIndexBase () const {
    return getRowMap ()->getIndexBase ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<const Map<LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getRowMap () const {
    return getCrsGraphRef ().getRowMap ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<const Map<LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getColMap () const {
    return getCrsGraphRef ().getColMap ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<const Map<LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getDomainMap () const {
    return getCrsGraphRef ().getDomainMap ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<const Map<LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getRangeMap () const {
    return getCrsGraphRef ().getRangeMap ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<const RowGraph<LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGraph () const {
    if (staticGraph_ != Teuchos::null) {
      return staticGraph_;
    }
    return myGraph_;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<const CrsGraph<LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getCrsGraph () const {
    if (staticGraph_ != Teuchos::null) {
      return staticGraph_;
    }
    return myGraph_;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  const CrsGraph<LocalOrdinal, GlobalOrdinal, Node>&
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getCrsGraphRef () const {
    if (! this->staticGraph_.is_null ()) {
      return * (this->staticGraph_);
    }
    else {
#ifdef HAVE_TPETRA_DEBUG
      const char tfecfFuncName[] = "getCrsGraphRef: ";
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (this->myGraph_.is_null (), std::logic_error,
         "Both staticGraph_ and myGraph_ are null.  "
         "Please report this bug to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
      return * (this->myGraph_);
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isLowerTriangularImpl () const {
    const crs_graph_type& G = this->getCrsGraphRef ();
    using HDM = ::Tpetra::Details::HasDeprecatedMethods2630_WarningThisClassIsNotForUsers;
    return dynamic_cast<const HDM&> (G).isLowerTriangularImpl ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isLowerTriangular () const {
    return this->isLowerTriangularImpl ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isUpperTriangularImpl () const {
    const crs_graph_type& G = this->getCrsGraphRef ();
    using HDM = ::Tpetra::Details::HasDeprecatedMethods2630_WarningThisClassIsNotForUsers;
    return dynamic_cast<const HDM&> (G).isUpperTriangularImpl ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isUpperTriangular () const {
    return this->isUpperTriangularImpl ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  isStaticGraph () const {
    return myGraph_.is_null ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  hasTransposeApply () const {
    return true;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  supportsRowViews () const {
    return true;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::ArrayRCP<Teuchos::Array<typename CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::impl_scalar_type> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  allocateValues2D ()
  {
    using Teuchos::arcp;
    using Teuchos::Array;
    using Teuchos::ArrayRCP;
    typedef impl_scalar_type IST;
    typedef LocalOrdinal LO;
    const char tfecfFuncName[] = "allocateValues2D: ";
 
    const crs_graph_type& graph = this->getCrsGraphRef ();
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! graph.indicesAreAllocated (), std::runtime_error,
       "Graph indices must be allocated before values.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (graph.getProfileType () != DynamicProfile, std::runtime_error,
       "Graph indices must be allocated in a dynamic profile.");
 
    const LO lclNumRows = graph.getNodeNumRows ();
    Teuchos::ArrayRCP<Teuchos::Array<IST> > values2D (lclNumRows);
    if (! graph.lclInds2D_.is_null ()) {
      for (LO lclRow = 0; lclRow < lclNumRows; ++lclRow) {
        values2D[lclRow].resize (graph.lclInds2D_[lclRow].size ());
      }
    }
    else if (! graph.gblInds2D_.is_null ()) {
      for (LO lclRow = 0; lclRow < lclNumRows; ++lclRow) {
        values2D[lclRow].resize (graph.gblInds2D_[lclRow].size ());
      }
    }
    return values2D;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  allocateValues (ELocalGlobal lg, GraphAllocationStatus gas)
  {
    using ::Tpetra::Details::ProfilingRegion;
    const char tfecfFuncName[] = "allocateValues: ";
    ProfilingRegion regionAllocateValues ("Tpetra::CrsMatrix::allocateValues");
 
#ifdef HAVE_TPETRA_DEBUG
    const char suffix[] = "  Please report this bug to the Tpetra developers.";
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (this->staticGraph_.is_null (), std::logic_error,
       "staticGraph_ is null." << suffix);
 
    // If the graph indices are already allocated, then gas should be
    // GraphAlreadyAllocated.  Otherwise, gas should be
    // GraphNotYetAllocated.
    if ((gas == GraphAlreadyAllocated) != this->staticGraph_->indicesAreAllocated ()) {
      const char err1[] = "The caller has asserted that the graph is ";
      const char err2[] = "already allocated, but the static graph says "
        "that its indices are ";
      const char err3[] = "already allocated.  Please report this bug to "
        "the Tpetra developers.";
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (gas == GraphAlreadyAllocated && ! this->staticGraph_->indicesAreAllocated (),
        std::logic_error, err1 << err2 << "not " << err3);
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (gas != GraphAlreadyAllocated && this->staticGraph_->indicesAreAllocated (),
         std::logic_error, err1 << "not " << err2 << err3);
    }
 
    // If the graph is unallocated, then it had better be a
    // matrix-owned graph.  ("Matrix-owned graph" means that the
    // matrix gets to define the graph structure.  If the CrsMatrix
    // constructor that takes an RCP<const CrsGraph> was used, then
    // the matrix does _not_ own the graph.)
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! this->staticGraph_->indicesAreAllocated () &&
       this->myGraph_.is_null (), std::logic_error,
       "The static graph says that its indices are not allocated, "
       "but the graph is not owned by the matrix." << suffix);
#endif // HAVE_TPETRA_DEBUG
 
    if (gas == GraphNotYetAllocated) {
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (this->myGraph_.is_null (), std::logic_error,
         "gas = GraphNotYetAllocated, but myGraph_ is null." << suffix);
#endif // HAVE_TPETRA_DEBUG
      try {
        this->myGraph_->allocateIndices (lg);
      }
      catch (std::exception& e) {
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (true, std::runtime_error, "CrsGraph::allocateIndices "
           "threw an exception: " << e.what ());
      }
      catch (...) {
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (true, std::runtime_error, "CrsGraph::allocateIndices "
           "threw an exception not a subclass of std::exception.");
      }
    }
 
    // Allocate matrix values.
    if (this->getProfileType () == StaticProfile) {
      // "Static profile" means that the number of matrix entries in
      // each row was fixed at the time the CrsMatrix constructor was
      // called.  This lets us use 1-D storage for the matrix's
      // values.  ("1-D storage" means the same as that used by the
      // three arrays in the compressed sparse row storage format.)
 
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (this->staticGraph_.is_null (), std::logic_error,
         "this->getProfileType() == StaticProfile, but staticGraph_ is null."
         << suffix);
#endif // HAVE_TPETRA_DEBUG
 
      const size_t lclNumRows = this->staticGraph_->getNodeNumRows ();
      typename Graph::local_graph_type::row_map_type k_ptrs =
        this->staticGraph_->k_rowPtrs_;
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (k_ptrs.extent (0) != lclNumRows+1, std::logic_error,
        "With StaticProfile, row offsets array has length "
        << k_ptrs.extent (0) << " != (lclNumRows+1) = "
        << (lclNumRows+1) << ".");
 
      const size_t lclTotalNumEntries =
        ::Tpetra::Details::getEntryOnHost (k_ptrs, lclNumRows);
 
      // Allocate array of (packed???) matrix values.
      typedef typename local_matrix_type::values_type values_type;
      this->k_values1D_ =
        values_type ("Tpetra::CrsMatrix::val", lclTotalNumEntries);
    }
    else {
      // "Dynamic profile" means the number of matrix entries in each
      // row is not fixed and may expand.  Thus, we store the matrix's
      // values in "2-D storage," meaning an array of arrays.  The
      // outer array has as many inner arrays as there are rows in the
      // matrix, and each inner array stores the values in that row.
      this->values2D_ = this->allocateValues2D ();
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getAllValues (Teuchos::ArrayRCP<const size_t>& rowPointers,
                Teuchos::ArrayRCP<const LocalOrdinal>& columnIndices,
                Teuchos::ArrayRCP<const Scalar>& values) const
  {
    using Teuchos::RCP;
    const char tfecfFuncName[] = "getAllValues: ";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      columnIndices.size () != values.size (), std::runtime_error,
      "Requires that columnIndices and values are the same size.");
 
    RCP<const crs_graph_type> relevantGraph = getCrsGraph ();
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      relevantGraph.is_null (), std::runtime_error,
      "Requires that getCrsGraph() is not null.");
    try {
      rowPointers = relevantGraph->getNodeRowPtrs ();
    }
    catch (std::exception &e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
        true, std::runtime_error,
        "Caught exception while calling graph->getNodeRowPtrs(): "
        << e.what ());
    }
    try {
      columnIndices = relevantGraph->getNodePackedIndices ();
    }
    catch (std::exception &e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
        true, std::runtime_error,
        "Caught exception while calling graph->getNodePackedIndices(): "
        << e.what ());
    }
    Teuchos::ArrayRCP<const impl_scalar_type> vals =
      Kokkos::Compat::persistingView (k_values1D_);
    values = Teuchos::arcp_reinterpret_cast<const Scalar> (vals);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  fillLocalGraphAndMatrix (const Teuchos::RCP<Teuchos::ParameterList>& params)
  {
    using ::Tpetra::Details::computeOffsetsFromCounts;
    using ::Tpetra::Details::ProfilingRegion;
    using Kokkos::create_mirror_view;
    using Teuchos::arcp_const_cast;
    using Teuchos::Array;
    using Teuchos::ArrayRCP;
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
    typedef typename local_matrix_type::row_map_type row_map_type;
    typedef typename Graph::local_graph_type::entries_type::non_const_type lclinds_1d_type;
    typedef typename local_matrix_type::values_type values_type;
    ProfilingRegion regionFLGAM ("Tpetra::CrsGraph::fillLocalGraphAndMatrix");
 
#ifdef HAVE_TPETRA_DEBUG
    const char tfecfFuncName[] = "fillLocalGraphAndMatrix (called from "
      "fillComplete or expertStaticFillComplete): ";
#endif // HAVE_TPETRA_DEBUG
 
#ifdef HAVE_TPETRA_DEBUG
    // fillComplete() only calls fillLocalGraphAndMatrix() if the
    // matrix owns the graph, which means myGraph_ is not null.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (myGraph_.is_null (), std::logic_error, "The nonconst graph (myGraph_) "
       "is null.  This means that the matrix has a const (a.k.a. \"static\") "
       "graph.  fillComplete or expertStaticFillComplete should never call "
       "fillLocalGraphAndMatrix in that case.  "
       "Please report this bug to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
 
    const size_t lclNumRows = this->getNodeNumRows ();
 
    // This method's goal is to fill in the three arrays (compressed
    // sparse row format) that define the sparse graph's and matrix's
    // structure, and the sparse matrix's values.
    //
    // Use the nonconst version of row_map_type for k_ptrs,
    // because row_map_type is const and we need to modify k_ptrs here.
    typename row_map_type::non_const_type k_ptrs;
    row_map_type k_ptrs_const;
    lclinds_1d_type k_inds;
    values_type k_vals;
 
    // Get references to the data in myGraph_, so we can modify them
    // as well.  Note that we only call fillLocalGraphAndMatrix() if
    // the matrix owns the graph, which means myGraph_ is not null.
    lclinds_1d_type k_lclInds1D_ = myGraph_->k_lclInds1D_;
 
    typedef decltype (myGraph_->k_numRowEntries_) row_entries_type;
 
    if (getProfileType () == DynamicProfile) {
      // Pack 2-D storage (DynamicProfile) into 1-D packed storage.
      //
      // DynamicProfile means that the matrix's column indices and
      // values are currently stored in a 2-D "unpacked" format, in
      // the arrays-of-arrays myGraph_->lclInds2D_ (for column
      // indices) and values2D_ (for values).  We allocate 1-D storage
      // (k_inds resp. k_vals), and then copy from 2-D storage
      // (lclInds2D_ resp. values2D_) into 1-D storage (k_inds
      // resp. k_vals).
 
      // We're be packing on host.  k_numRowEntries_ lives on host,
      // and computeOffsetsFromCounts accepts a host View for counts,
      // even if offsets is a device View.  (Furthermore, the "host"
      // View may very well live in CudaUVMSpace, so doing this has no
      // penalty, other than requiring synchronization between Cuda
      // and host.  UVM memory gets grumpy if both device and host
      // attempt to access it at the same time without an intervening
      // fence.)
      typename row_entries_type::const_type numRowEnt_h =
        myGraph_->k_numRowEntries_;
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (static_cast<size_t> (numRowEnt_h.extent (0)) != lclNumRows,
         std::logic_error, "(DynamicProfile branch) numRowEnt_h has the "
         "wrong length.  numRowEnt_h.extent(0) = "
         << numRowEnt_h.extent (0) << " != getNodeNumRows() = "
         << lclNumRows << ".");
#endif // HAVE_TPETRA_DEBUG
 
      // We're packing on host (since we can't read Teuchos data
      // structures on device), so let's fill the packed row offsets
      // on host first.
      k_ptrs = typename row_map_type::non_const_type ("Tpetra::CrsGraph::ptr",
                                                      lclNumRows+1);
      typename row_map_type::non_const_type::HostMirror h_ptrs =
        create_mirror_view (k_ptrs);
 
      // Pack the row offsets into k_ptrs, by doing a sum-scan of
      // the array of valid entry counts per row.
      //
      // Return value is the total number of entries in the matrix on
      // the calling process.  It's cheap to compute and useful as a
      // sanity check.
      const size_t lclTotalNumEntries =
        computeOffsetsFromCounts (h_ptrs, numRowEnt_h);
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (static_cast<size_t> (h_ptrs.extent (0)) != lclNumRows + 1,
         std::logic_error, "(DynamicProfile branch) After packing h_ptrs, "
         "h_ptrs.extent(0) = " << h_ptrs.extent (0) << " != "
         "(lclNumRows+1) = " << (lclNumRows+1) << ".");
      {
        const size_t h_ptrs_lastEnt = h_ptrs(lclNumRows); // it's a host View
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (h_ptrs_lastEnt != lclTotalNumEntries, std::logic_error,
           "(DynamicProfile branch) After packing h_ptrs, h_ptrs(lclNumRows="
           << lclNumRows << ") = " << h_ptrs_lastEnt << " != total number "
           "of entries on the calling process = " << lclTotalNumEntries << ".");
      }
#endif // HAVE_TPETRA_DEBUG
 
      // Allocate the arrays of packed column indices and values.
      k_inds = lclinds_1d_type ("Tpetra::CrsGraph::ind", lclTotalNumEntries);
      k_vals = values_type ("Tpetra::CrsMatrix::val", lclTotalNumEntries);
 
      // We need host views of the above, since 2-D storage lives on host.
      typename lclinds_1d_type::HostMirror h_inds = create_mirror_view (k_inds);
      typename values_type::HostMirror h_vals = create_mirror_view (k_vals);
 
      // Pack the column indices and values on the host.
      ArrayRCP<Array<LocalOrdinal> > lclInds2D = myGraph_->lclInds2D_;
      for (size_t row = 0; row < lclNumRows; ++row) {
        const size_t numEnt = numRowEnt_h(row);
        std::copy (lclInds2D[row].begin(),
                   lclInds2D[row].begin() + numEnt,
                   h_inds.data() + h_ptrs(row));
        std::copy (values2D_[row].begin(),
                   values2D_[row].begin() + numEnt,
                   h_vals.data() + h_ptrs(row));
      }
 
      // Copy the packed column indices and values to the device.
      Kokkos::deep_copy (k_inds, h_inds);
      Kokkos::deep_copy (k_vals, h_vals);
      // Copy the packed row offsets to the device too.
      // We didn't actually need them on device before.
      Kokkos::deep_copy (k_ptrs, h_ptrs);
      k_ptrs_const = k_ptrs; // const version of k_ptrs
 
#ifdef HAVE_TPETRA_DEBUG
      // Sanity check of packed row offsets.
      if (k_ptrs.extent (0) != 0) {
        const size_t numOffsets = static_cast<size_t> (k_ptrs.extent (0));
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (numOffsets != lclNumRows + 1, std::logic_error, "(DynamicProfile "
           "branch) After copying into k_ptrs, k_ptrs.extent(0) = " <<
           numOffsets << " != (lclNumRows+1) = " << (lclNumRows+1) << ".");
 
        const auto valToCheck = ::Tpetra::Details::getEntryOnHost (k_ptrs, numOffsets-1);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (valToCheck) != k_vals.extent (0),
          std::logic_error, "(DynamicProfile branch) After packing, k_ptrs("
           << (numOffsets-1) << ") = " << valToCheck << " != "
           "k_vals.extent(0) = " << k_vals.extent (0) << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (valToCheck) != k_inds.extent (0),
          std::logic_error, "(DynamicProfile branch) After packing, k_ptrs("
           << (numOffsets-1) << ") = " << valToCheck << " != "
           "k_inds.extent(0) = " << k_inds.extent (0) << ".");
      }
#endif // HAVE_TPETRA_DEBUG
    }
    else if (getProfileType () == StaticProfile) {
      // StaticProfile means that the matrix's column indices and
      // values are currently stored in a 1-D format, with row offsets
      // in k_rowPtrs_ and local column indices in k_lclInds1D_.
 
      // StaticProfile also means that the graph's array of row
      // offsets must already be allocated.
      typename Graph::local_graph_type::row_map_type curRowOffsets =
        myGraph_->k_rowPtrs_;
 
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (curRowOffsets.extent (0) == 0, std::logic_error,
         "(StaticProfile branch) curRowOffsets.extent(0) == 0.");
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (curRowOffsets.extent (0) != lclNumRows + 1, std::logic_error,
         "(StaticProfile branch) curRowOffsets.extent(0) = "
         << curRowOffsets.extent (0) << " != lclNumRows + 1 = "
         << (lclNumRows + 1) << ".")
      {
        const size_t numOffsets = curRowOffsets.extent (0);
        const auto valToCheck =
          ::Tpetra::Details::getEntryOnHost (curRowOffsets, numOffsets - 1);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (numOffsets != 0 &&
           myGraph_->k_lclInds1D_.extent (0) != valToCheck,
           std::logic_error, "(StaticProfile branch) numOffsets = " <<
           numOffsets << " != 0 and myGraph_->k_lclInds1D_.extent(0) = "
           << myGraph_->k_lclInds1D_.extent (0) << " != curRowOffsets("
           << numOffsets << ") = " << valToCheck << ".");
      }
#endif // HAVE_TPETRA_DEBUG
 
      if (myGraph_->getNodeNumEntries () != myGraph_->getNodeAllocationSize ()) {
        // The matrix's current 1-D storage is "unpacked."  This means
        // the row offsets may differ from what the final row offsets
        // should be.  This could happen, for example, if the user
        // specified StaticProfile in the constructor and set an upper
        // bound on the number of entries per row, but didn't fill all
        // those entries.
#ifdef HAVE_TPETRA_DEBUG
        if (curRowOffsets.extent (0) != 0) {
          const size_t numOffsets =
            static_cast<size_t> (curRowOffsets.extent (0));
          const auto valToCheck =
            ::Tpetra::Details::getEntryOnHost (curRowOffsets, numOffsets-1);
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (static_cast<size_t> (valToCheck) !=
             static_cast<size_t> (k_values1D_.extent (0)),
             std::logic_error, "(StaticProfile unpacked branch) Before "
             "allocating or packing, curRowOffsets(" << (numOffsets-1) << ") = "
             << valToCheck << " != k_values1D_.extent(0)"
             " = " << k_values1D_.extent (0) << ".");
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (static_cast<size_t> (valToCheck) !=
             static_cast<size_t> (myGraph_->k_lclInds1D_.extent (0)),
             std::logic_error, "(StaticProfile unpacked branch) Before "
             "allocating or packing, curRowOffsets(" << (numOffsets-1) << ") = "
             << valToCheck
             << " != myGraph_->k_lclInds1D_.extent(0) = "
             << myGraph_->k_lclInds1D_.extent (0) << ".");
        }
#endif // HAVE_TPETRA_DEBUG
 
        // Pack the row offsets into k_ptrs, by doing a sum-scan of
        // the array of valid entry counts per row.
 
        // Total number of entries in the matrix on the calling
        // process.  We will compute this in the loop below.  It's
        // cheap to compute and useful as a sanity check.
        size_t lclTotalNumEntries = 0;
        // This will be a host view of packed row offsets.
        typename row_map_type::non_const_type::HostMirror h_ptrs;
        {
          // Allocate the packed row offsets array.  We use a nonconst
          // temporary (packedRowOffsets) here, because k_ptrs is
          // const.  We will assign packedRowOffsets to k_ptrs below.
          typename row_map_type::non_const_type
            packedRowOffsets ("Tpetra::CrsGraph::ptr", lclNumRows + 1);
          typename row_entries_type::const_type numRowEnt_h =
            myGraph_->k_numRowEntries_;
          // We're computing offsets on device.  This function can
          // handle numRowEnt_h being a host View.
          lclTotalNumEntries =
            computeOffsetsFromCounts (packedRowOffsets, numRowEnt_h);
          // packedRowOffsets is modifiable; k_ptrs isn't, so we have
          // to use packedRowOffsets in the loop above and assign here.
          k_ptrs = packedRowOffsets;
          k_ptrs_const = k_ptrs;
        }
 
#ifdef HAVE_TPETRA_DEBUG
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (k_ptrs.extent (0)) != lclNumRows + 1,
           std::logic_error,
           "(StaticProfile unpacked branch) After packing k_ptrs, "
           "k_ptrs.extent(0) = " << k_ptrs.extent (0) << " != "
           "lclNumRows+1 = " << (lclNumRows+1) << ".");
        {
          const auto valToCheck = ::Tpetra::Details::getEntryOnHost (k_ptrs, lclNumRows);
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (valToCheck != lclTotalNumEntries, std::logic_error,
             "(StaticProfile unpacked branch) After filling k_ptrs, "
             "k_ptrs(lclNumRows=" << lclNumRows << ") = " << valToCheck
             << " != total number of entries on the calling process = "
             << lclTotalNumEntries << ".");
        }
#endif // HAVE_TPETRA_DEBUG
 
        // Allocate the arrays of packed column indices and values.
        k_inds = lclinds_1d_type ("Tpetra::CrsGraph::ind", lclTotalNumEntries);
        k_vals = values_type ("Tpetra::CrsMatrix::val", lclTotalNumEntries);
 
        // curRowOffsets (myGraph_->k_rowPtrs_) (???), k_lclInds1D_,
        // and k_values1D_ are currently unpacked.  Pack them, using
        // the packed row offsets array k_ptrs that we created above.
        //
        // FIXME (mfh 06 Aug 2014) If "Optimize Storage" is false, we
        // need to keep around the unpacked row offsets, column
        // indices, and values arrays.
 
        // Pack the column indices from unpacked k_lclInds1D_ into
        // packed k_inds.  We will replace k_lclInds1D_ below.
        typedef pack_functor<typename Graph::local_graph_type::entries_type::non_const_type,
          typename Graph::local_graph_type::row_map_type>
          inds_packer_type;
        inds_packer_type indsPacker (k_inds, myGraph_->k_lclInds1D_,
                                     k_ptrs, curRowOffsets);
        typedef typename decltype (k_inds)::execution_space exec_space;
        typedef Kokkos::RangePolicy<exec_space, LocalOrdinal> range_type;
        Kokkos::parallel_for (range_type (0, lclNumRows), indsPacker);
 
        // Pack the values from unpacked k_values1D_ into packed
        // k_vals.  We will replace k_values1D_ below.
        typedef pack_functor<values_type, row_map_type> vals_packer_type;
        vals_packer_type valsPacker (k_vals, this->k_values1D_,
                                     k_ptrs, curRowOffsets);
        Kokkos::parallel_for (range_type (0, lclNumRows), valsPacker);
 
#ifdef HAVE_TPETRA_DEBUG
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (k_ptrs.extent (0) == 0, std::logic_error,
           "(StaticProfile \"Optimize Storage\" = "
           "true branch) After packing, k_ptrs.extent(0) = 0.  This "
           "probably means that k_rowPtrs_ was never allocated.");
        if (k_ptrs.extent (0) != 0) {
          const size_t numOffsets = static_cast<size_t> (k_ptrs.extent (0));
          const auto valToCheck = ::Tpetra::Details::getEntryOnHost (k_ptrs, numOffsets - 1);
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (static_cast<size_t> (valToCheck) != k_vals.extent (0),
             std::logic_error,
             "(StaticProfile \"Optimize Storage\"=true branch) After packing, "
             "k_ptrs(" << (numOffsets-1) << ") = " << valToCheck <<
             " != k_vals.extent(0) = " << k_vals.extent (0) << ".");
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (static_cast<size_t> (valToCheck) != k_inds.extent (0),
             std::logic_error,
             "(StaticProfile \"Optimize Storage\"=true branch) After packing, "
             "k_ptrs(" << (numOffsets-1) << ") = " << valToCheck <<
             " != k_inds.extent(0) = " << k_inds.extent (0) << ".");
        }
#endif // HAVE_TPETRA_DEBUG
      }
      else { // We don't have to pack, so just set the pointers.
        k_ptrs_const = myGraph_->k_rowPtrs_;
        k_inds = myGraph_->k_lclInds1D_;
        k_vals = this->k_values1D_;
 
#ifdef HAVE_TPETRA_DEBUG
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (k_ptrs_const.extent (0) == 0, std::logic_error,
          "(StaticProfile \"Optimize Storage\"=false branch) "
          "k_ptrs_const.extent(0) = 0.  This probably means that "
          "k_rowPtrs_ was never allocated.");
        if (k_ptrs_const.extent (0) != 0) {
          const size_t numOffsets = static_cast<size_t> (k_ptrs_const.extent (0));
          const auto valToCheck = ::Tpetra::Details::getEntryOnHost (k_ptrs_const, numOffsets - 1);
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (static_cast<size_t> (valToCheck) != k_vals.extent (0),
             std::logic_error,
             "(StaticProfile \"Optimize Storage\"=false branch) "
             "k_ptrs_const(" << (numOffsets-1) << ") = " << valToCheck
             << " != k_vals.extent(0) = " << k_vals.extent (0) << ".");
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (static_cast<size_t> (valToCheck) != k_inds.extent (0),
             std::logic_error,
             "(StaticProfile \"Optimize Storage\" = false branch) "
             "k_ptrs_const(" << (numOffsets-1) << ") = " << valToCheck
             << " != k_inds.extent(0) = " << k_inds.extent (0) << ".");
        }
#endif // HAVE_TPETRA_DEBUG
      }
    }
 
#ifdef HAVE_TPETRA_DEBUG
    // Extra sanity checks.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (static_cast<size_t> (k_ptrs_const.extent (0)) != lclNumRows + 1,
       std::logic_error, "After packing, k_ptrs_const.extent(0) = " <<
       k_ptrs_const.extent (0) << " != lclNumRows+1 = " << (lclNumRows+1)
       << ".");
    if (k_ptrs_const.extent (0) != 0) {
      const size_t numOffsets = static_cast<size_t> (k_ptrs_const.extent (0));
      const size_t k_ptrs_const_numOffsetsMinus1 =
        ::Tpetra::Details::getEntryOnHost (k_ptrs_const, numOffsets - 1);
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (k_ptrs_const_numOffsetsMinus1 != k_vals.extent (0),
         std::logic_error, "After packing, k_ptrs_const(" << (numOffsets-1) <<
         ") = " << k_ptrs_const_numOffsetsMinus1 << " != k_vals.extent(0)"
         " = " << k_vals.extent (0) << ".");
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (k_ptrs_const_numOffsetsMinus1 != k_inds.extent (0),
         std::logic_error, "After packing, k_ptrs_const(" << (numOffsets-1) <<
         ") = " << k_ptrs_const_numOffsetsMinus1 << " != k_inds.extent(0)"
         " = " << k_inds.extent (0) << ".");
    }
#endif // HAVE_TPETRA_DEBUG
 
    // May we ditch the old allocations for the packed (and otherwise
    // "optimized") allocations, later in this routine?  Optimize
    // storage if the graph is not static, or if the graph already has
    // optimized storage.
    const bool defaultOptStorage =
      ! isStaticGraph () || staticGraph_->isStorageOptimized ();
    const bool requestOptimizedStorage =
      (! params.is_null () && params->get ("Optimize Storage", defaultOptStorage)) ||
      (params.is_null () && defaultOptStorage);
 
    // The graph has optimized storage when indices are allocated,
    // myGraph_->k_numRowEntries_ is empty, and there are more than
    // zero rows on this process.  It's impossible for the graph to
    // have dynamic profile (getProfileType() == DynamicProfile) and
    // be optimized (isStorageOptimized()).
    if (requestOptimizedStorage) {
      // Free the old, unpacked, unoptimized allocations.
      // Change the graph from dynamic to static allocation profile
 
      // Free graph data structures that are only needed for 2-D or
      // unpacked 1-D storage.
      myGraph_->lclInds2D_ = null; // legacy KokkosClassic 2-D storage
      myGraph_->k_numRowEntries_ = row_entries_type ();
 
      // Free the matrix's 2-D storage.
      this->values2D_ = null;
 
      // Keep the new 1-D packed allocations.
      myGraph_->k_rowPtrs_ = k_ptrs_const;
      myGraph_->k_lclInds1D_ = k_inds;
      this->k_values1D_ = k_vals;
 
      // Whatever graph was before, it's StaticProfile now.
      myGraph_->pftype_ = StaticProfile;
      myGraph_->storageStatus_ = ::Tpetra::Details::STORAGE_1D_PACKED;
      this->storageStatus_ = ::Tpetra::Details::STORAGE_1D_PACKED;
    }
 
    // Make the local graph, using the arrays of row offsets and
    // column indices that we built above.  The local graph should be
    // null, but we delete it first so that any memory can be freed
    // before we allocate the new one.
    //
    // FIXME (mfh 06,28 Aug 2014) It would make more sense for
    // Tpetra::CrsGraph to have a protected method that accepts k_inds
    // and k_ptrs, and creates the local graph lclGraph_.
    myGraph_->lclGraph_ =
      typename Graph::local_graph_type (k_inds, k_ptrs_const);
 
    // Make the local matrix, using the local graph and vals array.
    lclMatrix_ = local_matrix_type ("Tpetra::CrsMatrix::lclMatrix_",
                                    getNodeNumCols (), k_vals,
                                    myGraph_->lclGraph_);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  fillLocalMatrix (const Teuchos::RCP<Teuchos::ParameterList>& params)
  {
    using ::Tpetra::Details::ProfilingRegion;
    using Kokkos::create_mirror_view;
    using Teuchos::ArrayRCP;
    using Teuchos::Array;
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
    typedef LocalOrdinal LO;
    typedef typename Graph::local_graph_type::row_map_type row_map_type;
    typedef typename row_map_type::non_const_type non_const_row_map_type;
    typedef typename local_matrix_type::values_type values_type;
#ifdef HAVE_TPETRA_DEBUG
    const char tfecfFuncName[] = "fillLocalMatrix (called from fillComplete): ";
#endif // HAVE_TPETRA_DEBUG
    ProfilingRegion regionFLM ("Tpetra::CrsMatrix::fillLocalMatrix");
 
    const size_t lclNumRows = getNodeNumRows();
    const map_type& rowMap = * (getRowMap ());
    RCP<node_type> node = rowMap.getNode ();
 
    // The goals of this routine are first, to allocate and fill
    // packed 1-D storage (see below for an explanation) in the vals
    // array, and second, to give vals to the local matrix and
    // finalize the local matrix.  We only need k_ptrs, the packed 1-D
    // row offsets, within the scope of this routine, since we're only
    // filling the local matrix here (use fillLocalGraphAndMatrix() to
    // fill both the graph and the matrix at the same time).
 
    // get data from staticGraph_
    ArrayRCP<Array<LO> > lclInds2D = staticGraph_->lclInds2D_;
    size_t nodeNumEntries   = staticGraph_->getNodeNumEntries ();
    size_t nodeNumAllocated = staticGraph_->getNodeAllocationSize ();
    row_map_type k_rowPtrs_ = staticGraph_->lclGraph_.row_map;
 
    row_map_type k_ptrs; // "packed" row offsets array
    values_type k_vals; // "packed" values array
 
    // May we ditch the old allocations for the packed (and otherwise
    // "optimized") allocations, later in this routine?  Request
    // optimized storage by default.
    bool requestOptimizedStorage = true;
    const bool default_OptimizeStorage =
      ! isStaticGraph () || staticGraph_->isStorageOptimized ();
    if (! params.is_null () && ! params->get ("Optimize Storage", default_OptimizeStorage)) {
      requestOptimizedStorage = false;
    }
    // If we're not allowed to change a static graph, then we can't
    // change the storage of the matrix, either.  This means that if
    // the graph's storage isn't already optimized, we can't optimize
    // the matrix's storage either.  Check and give warning, as
    // appropriate.
    if (! staticGraph_->isStorageOptimized () && requestOptimizedStorage) {
      TPETRA_ABUSE_WARNING(true, std::runtime_error,
        "You requested optimized storage by setting the"
        "\"Optimize Storage\" flag to \"true\" in the parameter list, or by virtue"
        "of default behavior. However, the associated CrsGraph was filled separately"
        "and requested not to optimize storage. Therefore, the CrsMatrix cannot"
        "optimize storage.");
      requestOptimizedStorage = false;
    }
 
    typedef decltype (staticGraph_->k_numRowEntries_) row_entries_type;
 
    if (getProfileType() == DynamicProfile) {
      // Pack 2-D storage (DynamicProfile) into 1-D packed storage.
      //
      // DynamicProfile means that the matrix's values are currently
      // stored in a 2-D "unpacked" format, in the array-of-arrays
      // values2D_.  We allocate 1-D storage and then copy from 2-D
      // storage in values2D_ into 1-D storage in k_vals.  Since we're
      // only allocating the local matrix here, not the local graph,
      // we don't need to keep the row offsets array, but we do need
      // it here temporarily in order to convert to 1-D storage.  (The
      // allocStorage() function needs it.)  We'll free ptrs later in
      // this method.
      //
      // FIXME (mfh 08 Aug 2014) If we're in this method, then the
      // graph should already have packed 1-D storage.  Why can't we
      // just use the graph's current row offsets array?
 
      // Pack the row offsets into k_ptrs, by doing a sum-scan of
      // the array of valid entry counts per row.
      //
      // Total number of entries in the matrix on the calling
      // process.  We will compute this in the loop below.  It's
      // cheap to compute and useful as a sanity check.
      size_t lclTotalNumEntries = 0;
      // This will be a host view of packed row offsets.
      typename non_const_row_map_type::HostMirror h_ptrs;
 
      typename row_entries_type::const_type numRowEnt_h =
        staticGraph_->k_numRowEntries_;
      {
        non_const_row_map_type packedRowOffsets ("Tpetra::CrsGraph::ptr",
                                                 lclNumRows+1);
        // NOTE (mfh 27 Jun 2016) We need h_ptrs on host anyway, so
        // let's just compute offsets on host.
        h_ptrs = create_mirror_view (packedRowOffsets);
        using ::Tpetra::Details::computeOffsetsFromCounts;
        lclTotalNumEntries = computeOffsetsFromCounts (h_ptrs, numRowEnt_h);
        Kokkos::deep_copy (packedRowOffsets, h_ptrs);
        k_ptrs = packedRowOffsets;
      }
 
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (static_cast<size_t> (k_ptrs.extent (0)) != lclNumRows + 1,
         std::logic_error, "In DynamicProfile branch, after packing k_ptrs, "
         "k_ptrs.extent(0) = " << k_ptrs.extent (0) << " != "
         "(lclNumRows+1) = " << (lclNumRows+1) << ".");
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (static_cast<size_t> (h_ptrs.extent (0)) != lclNumRows + 1,
         std::logic_error, "In DynamicProfile branch, after packing h_ptrs, "
         "h_ptrs.extent(0) = " << h_ptrs.extent (0) << " != "
         "(lclNumRows+1) = " << (lclNumRows+1) << ".");
      {
        const auto valToCheck = ::Tpetra::Details::getEntryOnHost (k_ptrs, lclNumRows);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (valToCheck) != lclTotalNumEntries,
           std::logic_error, "(DynamicProfile branch) After packing k_ptrs, "
           "k_ptrs(lclNumRows = " << lclNumRows << ") = " << valToCheck
           << " != total number of entries on the calling process = "
           << lclTotalNumEntries << ".");
      }
#endif // HAVE_TPETRA_DEBUG
 
      // Allocate the array of packed values.
      k_vals = values_type ("Tpetra::CrsMatrix::val", lclTotalNumEntries);
      // We need a host view of the above, since 2-D storage lives on host.
      typename values_type::HostMirror h_vals = create_mirror_view (k_vals);
      // Pack the values on the host.
      for (size_t lclRow = 0; lclRow < lclNumRows; ++lclRow) {
        const size_t numEnt = numRowEnt_h(lclRow);
        std::copy (values2D_[lclRow].begin(),
                   values2D_[lclRow].begin() + numEnt,
                   h_vals.data() + h_ptrs(lclRow));
      }
      // Copy the packed values to the device.
      Kokkos::deep_copy (k_vals, h_vals);
 
#ifdef HAVE_TPETRA_DEBUG
      // Sanity check of packed row offsets.
      if (k_ptrs.extent (0) != 0) {
        const size_t numOffsets = static_cast<size_t> (k_ptrs.extent (0));
        const auto valToCheck =
          ::Tpetra::Details::getEntryOnHost (k_ptrs, numOffsets - 1);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (valToCheck) != k_vals.extent (0),
           std::logic_error, "(DynamicProfile branch) After packing, k_ptrs("
           << (numOffsets-1) << ") = " << valToCheck << " != "
           "k_vals.extent(0) = " << k_vals.extent (0) << ".");
      }
#endif // HAVE_TPETRA_DEBUG
    }
    else if (getProfileType () == StaticProfile) {
      // StaticProfile means that the matrix's values are currently
      // stored in a 1-D format.  However, this format is "unpacked";
      // it doesn't necessarily have the same row offsets as indicated
      // by the ptrs array returned by allocRowPtrs.  This could
      // happen, for example, if the user specified StaticProfile in
      // the constructor and fixed the number of matrix entries in
      // each row, but didn't fill all those entries.
      //
      // As above, we don't need to keep the "packed" row offsets
      // array ptrs here, but we do need it here temporarily, so we
      // have to allocate it.  We'll free ptrs later in this method.
      //
      // Note that this routine checks whether storage has already
      // been packed.  This is a common case for solution of nonlinear
      // PDEs using the finite element method, as long as the
      // structure of the sparse matrix does not change between linear
      // solves.
      if (nodeNumEntries != nodeNumAllocated) {
        // We have to pack the 1-D storage, since the user didn't fill
        // up all requested storage.
        non_const_row_map_type tmpk_ptrs ("Tpetra::CrsGraph::ptr",
                                          lclNumRows+1);
        // Total number of entries in the matrix on the calling
        // process.  We will compute this in the loop below.  It's
        // cheap to compute and useful as a sanity check.
        size_t lclTotalNumEntries = 0;
        k_ptrs = tmpk_ptrs;
        {
          typename row_entries_type::const_type numRowEnt_d =
            staticGraph_->k_numRowEntries_;
          using ::Tpetra::Details::computeOffsetsFromCounts;
          // This function can handle the counts being a host View.
          lclTotalNumEntries = computeOffsetsFromCounts (tmpk_ptrs, numRowEnt_d);
        }
 
        // Allocate the "packed" values array.
        // It has exactly the right number of entries.
        k_vals = values_type ("Tpetra::CrsMatrix::val", lclTotalNumEntries);
 
        // Pack k_values1D_ into k_vals.  We will replace k_values1D_ below.
        typedef pack_functor<values_type, row_map_type> packer_type;
        packer_type valsPacker (k_vals, k_values1D_, tmpk_ptrs, k_rowPtrs_);
 
        typedef typename decltype (k_vals)::execution_space exec_space;
        typedef Kokkos::RangePolicy<exec_space, LocalOrdinal> range_type;
        Kokkos::parallel_for (range_type (0, lclNumRows), valsPacker);
      }
      else { // We don't have to pack, so just set the pointer.
        k_vals = k_values1D_;
      }
    }
 
    // May we ditch the old allocations for the packed one?
    if (requestOptimizedStorage) {
      // The user requested optimized storage, so we can dump the
      // unpacked 2-D and 1-D storage, and keep the packed storage.
      values2D_ = null;
      k_values1D_ = k_vals;
      this->storageStatus_ = ::Tpetra::Details::STORAGE_1D_PACKED;
    }
 
    // Build the local sparse matrix object.  At this point, the local
    // matrix certainly has a column Map.  Remember that the local
    // matrix's number of columns comes from the column Map, not the
    // domain Map.
    lclMatrix_ = local_matrix_type ("Tpetra::CrsMatrix::lclMatrix_",
                                    getColMap ()->getNodeNumElements (),
                                    k_vals,
                                    staticGraph_->getLocalGraph ());
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  insertIndicesAndValues (crs_graph_type& graph,
                          RowInfo& rowInfo,
                          const typename crs_graph_type::SLocalGlobalViews& newInds,
                          const Teuchos::ArrayView<impl_scalar_type>& oldRowVals,
                          const Teuchos::ArrayView<const impl_scalar_type>& newRowVals,
                          const ELocalGlobal lg,
                          const ELocalGlobal I)
  {
    const size_t oldNumEnt = rowInfo.numEntries;
    const size_t numInserted = graph.insertIndices (rowInfo, newInds, lg, I);
 
    // Use of memcpy here works around an issue with GCC >= 4.9.0,
    // that probably relates to scalar_type vs. impl_scalar_type
    // aliasing.  See history of Tpetra_CrsGraph_def.hpp for
    // details; look for GCC_WORKAROUND macro definition.
    if (numInserted > 0) {
      const size_t startOffset = oldNumEnt;
      memcpy (&oldRowVals[startOffset], &newRowVals[0],
              numInserted * sizeof (impl_scalar_type));
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  insertLocalValues (const LocalOrdinal lclRow,
                     const Teuchos::ArrayView<const LocalOrdinal>& indices,
                     const Teuchos::ArrayView<const Scalar>& values)
  {
    using std::endl;
    typedef impl_scalar_type IST;
    const char tfecfFuncName[] = "insertLocalValues: ";
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! this->isFillActive (), std::runtime_error,
       "Fill is not active.  After calling fillComplete, you must call "
       "resumeFill before you may insert entries into the matrix again.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (this->isStaticGraph (), std::runtime_error,
       "Cannot insert indices with static graph; use replaceLocalValues() "
       "instead.");
    // At this point, we know that myGraph_ is nonnull.
    crs_graph_type& graph = * (this->myGraph_);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (graph.colMap_.is_null (), std::runtime_error,
       "Cannot insert local indices without a column map.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (graph.isGloballyIndexed (),
       std::runtime_error, "Graph indices are global; use "
       "insertGlobalValues().");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (values.size () != indices.size (), std::runtime_error,
       "values.size() = " << values.size ()
       << " != indices.size() = " << indices.size () << ".");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      ! graph.rowMap_->isNodeLocalElement (lclRow), std::runtime_error,
      "Local row index " << lclRow << " does not belong to this process.");
 
    if (! graph.indicesAreAllocated ()) {
      this->allocateValues (LocalIndices, GraphNotYetAllocated);
    }
 
    const size_t numEntriesToAdd = static_cast<size_t> (indices.size ());
#ifdef HAVE_TPETRA_DEBUG
    // In a debug build, test whether any of the given column indices
    // are not in the column Map.  Keep track of the invalid column
    // indices so we can tell the user about them.
    {
      using Teuchos::toString;
 
      const map_type& colMap = * (graph.colMap_);
      Teuchos::Array<LocalOrdinal> badColInds;
      bool allInColMap = true;
      for (size_t k = 0; k < numEntriesToAdd; ++k) {
        if (! colMap.isNodeLocalElement (indices[k])) {
          allInColMap = false;
          badColInds.push_back (indices[k]);
        }
      }
      if (! allInColMap) {
        std::ostringstream os;
        os << "You attempted to insert entries in owned row " << lclRow
           << ", at the following column indices: " << toString (indices)
           << "." << endl;
        os << "Of those, the following indices are not in the column Map on "
          "this process: " << toString (badColInds) << "." << endl << "Since "
          "the matrix has a column Map already, it is invalid to insert "
          "entries at those locations.";
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (true, std::invalid_argument, os.str ());
      }
    }
#endif // HAVE_TPETRA_DEBUG
 
    RowInfo rowInfo = graph.getRowInfo (lclRow);
    const size_t curNumEnt = rowInfo.numEntries;
    const size_t newNumEnt = curNumEnt + numEntriesToAdd;
    if (newNumEnt > rowInfo.allocSize) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (this->getProfileType () == StaticProfile, std::runtime_error,
         "New indices exceed statically allocated graph structure.");
      // This must be a nonconst reference, since we'll reallocate.
      Teuchos::Array<IST>& curVals = this->values2D_[lclRow];
      // Make space for the new matrix entries.
      // Teuchos::ArrayRCP::resize automatically copies over values on
      // reallocation.
      graph.lclInds2D_[rowInfo.localRow].resize (newNumEnt);
      curVals.resize (newNumEnt);
      rowInfo.allocSize = newNumEnt; // give rowInfo updated allocSize
    }
    typename crs_graph_type::SLocalGlobalViews indsView;
    indsView.linds = indices;
 
    Teuchos::ArrayView<IST> valsView = this->getViewNonConst (rowInfo);
    Teuchos::ArrayView<const IST> valsIn =
      Teuchos::av_reinterpret_cast<const IST> (values);
    this->insertIndicesAndValues (graph, rowInfo, indsView, valsView,
                                  valsIn, LocalIndices, LocalIndices);
#ifdef HAVE_TPETRA_DEBUG
    const size_t chkNewNumEnt = graph.getNumEntriesInLocalRow (lclRow);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (chkNewNumEnt != newNumEnt, std::logic_error,
      "The row should have " << newNumEnt << " entries after insert, but "
      "instead has " << chkNewNumEnt << ".  Please report this bug to "
      "the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  insertLocalValues (const LocalOrdinal localRow,
                     const LocalOrdinal numEnt,
                     const Scalar vals[],
                     const LocalOrdinal cols[])
  {
    Teuchos::ArrayView<const LocalOrdinal> colsT (cols, numEnt);
    Teuchos::ArrayView<const Scalar> valsT (vals, numEnt);
    this->insertLocalValues (localRow, colsT, valsT);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  insertGlobalValuesImpl (crs_graph_type& graph,
                          RowInfo& rowInfo,
                          const GlobalOrdinal gblColInds[],
                          const impl_scalar_type vals[],
                          const size_t numInputEnt)
  {
    typedef impl_scalar_type IST;
    typedef GlobalOrdinal GO;
    const char tfecfFuncName[] = "insertGlobalValuesImpl: ";
 
#ifdef HAVE_TPETRA_DEBUG
    const size_t origNumEnt = graph.getNumEntriesInLocalRow (rowInfo.localRow);
#endif // HAVE_TPETRA_DEBUG
 
    if (! graph.indicesAreAllocated ()) {
      this->allocateValues (GlobalIndices, GraphNotYetAllocated);
      // mfh 23 Jul 2017: allocateValues invalidates existing
      // getRowInfo results.  Once we get rid of lazy graph
      // allocation, we'll be able to move the getRowInfo call outside
      // of this method.
      rowInfo = graph.getRowInfo (rowInfo.localRow);
    }
 
    const size_t curNumEnt = rowInfo.numEntries;
    const size_t newNumEnt = curNumEnt + numInputEnt;
    if (newNumEnt > rowInfo.allocSize) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (this->getProfileType () == StaticProfile &&
         newNumEnt > rowInfo.allocSize, std::runtime_error,
         "New indices exceed statically allocated graph structure.  "
         "curNumEnt (" << curNumEnt << ") + numInputEnt ("
         << numInputEnt << ") > allocSize (" << rowInfo.allocSize
         << ").");
      // This needs to be a nonconst reference, in case we want to
      // reallocate it.
      Teuchos::Array<IST>& curVals = this->values2D_[rowInfo.localRow];
      // Teuchos::ArrayRCP::resize automatically copies over values on
      // reallocation.
      graph.gblInds2D_[rowInfo.localRow].resize (newNumEnt);
      curVals.resize (newNumEnt);
      rowInfo.allocSize = newNumEnt; // reassign for updated allocSize
    }
 
    using Teuchos::ArrayView;
    typename crs_graph_type::SLocalGlobalViews inputIndsAV;
    inputIndsAV.ginds = ArrayView<const GO> (gblColInds, numInputEnt);
    ArrayView<IST> curValsAV = this->getViewNonConst (rowInfo);
    ArrayView<const IST> inputValsAV (vals, numInputEnt);
 
    const ELocalGlobal curIndexingStatus =
      this->isGloballyIndexed () ? GlobalIndices : LocalIndices;
    // curIndexingStatus == GlobalIndices means the method calls
    // getGlobalViewNonConst() and does direct copying, which should
    // be reasonably fast.  LocalIndices means the method calls the
    // Map's getLocalElement() method once per entry to insert.  This
    // may be slow.
    this->insertIndicesAndValues (graph, rowInfo, inputIndsAV, curValsAV,
                                  inputValsAV, GlobalIndices,
                                  curIndexingStatus);
#ifdef HAVE_TPETRA_DEBUG
    const size_t chkNewNumEnt =
      graph.getNumEntriesInLocalRow (rowInfo.localRow);
    if (chkNewNumEnt != newNumEnt) {
      std::ostringstream os;
      os << std::endl << "newNumEnt = " << newNumEnt
         << " != graph.getNumEntriesInLocalRow(" << rowInfo.localRow
         << ") = " << chkNewNumEnt << "." << std::endl
         << "\torigNumEnt: " << origNumEnt << std::endl
         << "\tnumInputEnt: " << numInputEnt << std::endl
         << "\tgblColInds: [";
      for (size_t k = 0; k < numInputEnt; ++k) {
        os << gblColInds[k];
        if (k + size_t (1) < numInputEnt) {
          os << ",";
        }
      }
      os << "]" << std::endl
         << "\tvals: [";
      for (size_t k = 0; k < numInputEnt; ++k) {
        os << vals[k];
        if (k + size_t (1) < numInputEnt) {
          os << ",";
        }
      }
      os << "]" << std::endl;
 
      if (this->supportsRowViews ()) {
        Teuchos::ArrayView<const Scalar> vals2;
        if (this->isGloballyIndexed ()) {
          Teuchos::ArrayView<const GlobalOrdinal> gblColInds2;
          const GlobalOrdinal gblRow =
            graph.rowMap_->getGlobalElement (rowInfo.localRow);
          if (gblRow == Tpetra::Details::OrdinalTraits<GlobalOrdinal>::invalid ()) {
            os << "Local row index " << rowInfo.localRow << " is invalid!" << std::endl;
          }
          else {
            bool getViewThrew = false;
            try {
              this->getGlobalRowView (gblRow, gblColInds2, vals2);
            }
            catch (std::exception& e) {
              getViewThrew = true;
              os << "getGlobalRowView threw exception:" << std::endl
                 << e.what () << std::endl;
            }
            if (! getViewThrew) {
              os << "\tNew global column indices: "
                 << Teuchos::toString (gblColInds2) << std::endl
                 << "\tNew values: " << Teuchos::toString (vals2) << std::endl;
            }
          }
        }
        else if (this->isLocallyIndexed ()) {
          Teuchos::ArrayView<const LocalOrdinal> lclColInds2;
          this->getLocalRowView (rowInfo.localRow, lclColInds2, vals2);
          os << "\tNew local column indices: " << Teuchos::toString (lclColInds2)
             << std::endl;
          os << "\tNew values: " << Teuchos::toString (vals2) << std::endl;
        }
      }
 
      os << "Please report this bug to the Tpetra developers.";
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::logic_error, os.str ());
    }
#endif // HAVE_TPETRA_DEBUG
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  insertGlobalValues (const GlobalOrdinal gblRow,
                      const Teuchos::ArrayView<const GlobalOrdinal>& indices,
                      const Teuchos::ArrayView<const Scalar>& values)
  {
    using Teuchos::toString;
    using std::endl;
    typedef impl_scalar_type IST;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef Tpetra::Details::OrdinalTraits<LO> OTLO;
    typedef typename Teuchos::ArrayView<const GO>::size_type size_type;
    const char tfecfFuncName[] = "insertGlobalValues: ";
 
#ifdef HAVE_TPETRA_DEBUG
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (values.size () != indices.size (), std::runtime_error,
      "values.size() = " << values.size () << " != indices.size() = "
      << indices.size () << ".");
#endif // HAVE_TPETRA_DEBUG
 
    // getRowMap() is not thread safe, because it increments RCP's
    // reference count.  getCrsGraphRef() is thread safe.
    const map_type& rowMap = * (this->getCrsGraphRef ().rowMap_);
    const LO lclRow = rowMap.getLocalElement (gblRow);
 
    if (lclRow == OTLO::invalid ()) {
      // Input row is _not_ owned by the calling process.
      //
      // See a note (now deleted) from mfh 14 Dec 2012: If input row
      // is not in the row Map, it doesn't matter whether or not the
      // graph is static; the data just get stashed for later use by
      // globalAssemble().
      this->insertNonownedGlobalValues (gblRow, indices, values);
    }
    else { // Input row _is_ owned by the calling process
      if (this->isStaticGraph ()) {
        // Uh oh!  Not allowed to insert into owned rows in that case.
        const int myRank = rowMap.getComm ()->getRank ();
        const int numProcs = rowMap.getComm ()->getSize ();
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (true, std::runtime_error,
           "The matrix was constructed with a constant (\"static\") graph, "
           "yet the given global row index " << gblRow << " is in the row "
           "Map on the calling process (with rank " << myRank << ", of " <<
           numProcs << " process(es)).  In this case, you may not insert "
           "new entries into rows owned by the calling process.");
      }
 
      crs_graph_type& graph = * (this->myGraph_);
      const IST* const inputVals =
        reinterpret_cast<const IST*> (values.getRawPtr ());
      const GO* const inputGblColInds = indices.getRawPtr ();
      const size_t numInputEnt = indices.size ();
      RowInfo rowInfo = graph.getRowInfo (lclRow);
 
      // If the matrix has a column Map, check at this point whether
      // the column indices belong to the column Map.
      //
      // FIXME (mfh 16 May 2013) We may want to consider deferring the
      // test to the CrsGraph method, since it may have to do this
      // anyway.
      if (! graph.colMap_.is_null ()) {
        const map_type& colMap = * (graph.colMap_);
        // In a debug build, keep track of the nonowned ("bad") column
        // indices, so that we can display them in the exception
        // message.  In a release build, just ditch the loop early if
        // we encounter a nonowned column index.
#ifdef HAVE_TPETRA_DEBUG
        Teuchos::Array<GO> badColInds;
#endif // HAVE_TPETRA_DEBUG
        const size_type numEntriesToInsert = indices.size ();
        bool allInColMap = true;
        for (size_type k = 0; k < numEntriesToInsert; ++k) {
          if (! colMap.isNodeGlobalElement (indices[k])) {
            allInColMap = false;
#ifdef HAVE_TPETRA_DEBUG
            badColInds.push_back (indices[k]);
#else
            break;
#endif // HAVE_TPETRA_DEBUG
          }
        }
        if (! allInColMap) {
          std::ostringstream os;
          os << "You attempted to insert entries in owned row " << gblRow
             << ", at the following column indices: " << toString (indices)
             << "." << endl;
#ifdef HAVE_TPETRA_DEBUG
          os << "Of those, the following indices are not in the column Map "
            "on this process: " << toString (badColInds) << "." << endl
             << "Since the matrix has a column Map already, it is invalid "
            "to insert entries at those locations.";
#else
          os << "At least one of those indices is not in the column Map "
            "on this process." << endl << "It is invalid to insert into "
            "columns not in the column Map on the process that owns the "
            "row.";
#endif // HAVE_TPETRA_DEBUG
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (true, std::invalid_argument, os.str ());
        }
      }
 
      this->insertGlobalValuesImpl (graph, rowInfo, inputGblColInds,
                                    inputVals, numInputEnt);
    }
  }
 
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  insertGlobalValues (const GlobalOrdinal globalRow,
                      const LocalOrdinal numEnt,
                      const Scalar vals[],
                      const GlobalOrdinal inds[])
  {
    Teuchos::ArrayView<const GlobalOrdinal> indsT (inds, numEnt);
    Teuchos::ArrayView<const Scalar> valsT (vals, numEnt);
    this->insertGlobalValues (globalRow, indsT, valsT);
  }
 
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  insertGlobalValuesFiltered (const GlobalOrdinal gblRow,
                              const Teuchos::ArrayView<const GlobalOrdinal>& indices,
                              const Teuchos::ArrayView<const Scalar>& values)
  {
    typedef impl_scalar_type IST;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef Tpetra::Details::OrdinalTraits<LO> OTLO;
    const char tfecfFuncName[] = "insertGlobalValuesFiltered: ";
 
#ifdef HAVE_TPETRA_DEBUG
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (values.size () != indices.size (), std::runtime_error,
      "values.size() = " << values.size () << " != indices.size() = "
      << indices.size () << ".");
#endif // HAVE_TPETRA_DEBUG
 
    // getRowMap() is not thread safe, because it increments RCP's
    // reference count.  getCrsGraphRef() is thread safe.
    const map_type& rowMap = * (this->getCrsGraphRef ().rowMap_);
    const LO lclRow = rowMap.getLocalElement (gblRow);
 
    if (lclRow == OTLO::invalid ()) {
      // Input row is _not_ owned by the calling process.
      //
      // See a note (now deleted) from mfh 14 Dec 2012: If input row
      // is not in the row Map, it doesn't matter whether or not the
      // graph is static; the data just get stashed for later use by
      // globalAssemble().
      this->insertNonownedGlobalValues (gblRow, indices, values);
    }
    else { // Input row _is_ owned by the calling process
      if (this->isStaticGraph ()) {
        // Uh oh!  Not allowed to insert into owned rows in that case.
        const int myRank = rowMap.getComm ()->getRank ();
        const int numProcs = rowMap.getComm ()->getSize ();
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (true, std::runtime_error,
           "The matrix was constructed with a constant (\"static\") graph, "
           "yet the given global row index " << gblRow << " is in the row "
           "Map on the calling process (with rank " << myRank << ", of " <<
           numProcs << " process(es)).  In this case, you may not insert "
           "new entries into rows owned by the calling process.");
      }
 
      crs_graph_type& graph = * (this->myGraph_);
      const IST* const inputVals =
        reinterpret_cast<const IST*> (values.getRawPtr ());
      const GO* const inputGblColInds = indices.getRawPtr ();
      const size_t numInputEnt = indices.size ();
      RowInfo rowInfo = graph.getRowInfo (lclRow);
 
      if (! graph.colMap_.is_null ()) { // We have a column Map.
        const map_type& colMap = * (graph.colMap_);
        size_t curOffset = 0;
        while (curOffset < numInputEnt) {
          // Find a sequence of input indices that are in the column
          // Map on the calling process.  Doing a sequence at a time,
          // instead of one at a time, amortizes some overhead.
          size_t endOffset = curOffset;
          for ( ; endOffset < numInputEnt &&
                  colMap.getLocalElement (inputGblColInds[endOffset]) != OTLO::invalid ();
                ++endOffset)
            {}
          // curOffset, endOffset: half-exclusive range of indices in
          // the column Map on the calling process.  If endOffset ==
          // curOffset, the range is empty.
          const LO numIndInSeq = (endOffset - curOffset);
          if (numIndInSeq != 0) {
            this->insertGlobalValuesImpl (graph, rowInfo,
                                          inputGblColInds + curOffset,
                                          inputVals + curOffset,
                                          numIndInSeq);
          }
          // Invariant before the increment line: Either endOffset ==
          // numInputEnt, or inputGblColInds[endOffset] is not in the
          // column Map on the calling process.
#ifdef HAVE_TPETRA_DEBUG
          const bool invariant = endOffset == numInputEnt ||
            colMap.getLocalElement (inputGblColInds[endOffset]) == OTLO::invalid ();
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (! invariant, std::logic_error, std::endl << "Invariant failed!");
#endif // HAVE_TPETRA_DEBUG
          curOffset = endOffset + 1;
        }
      }
      else { // we don't have a column Map.
        this->insertGlobalValuesImpl (graph, rowInfo, inputGblColInds,
                                      inputVals, numInputEnt);
      }
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  replaceLocalValuesImpl (impl_scalar_type rowVals[],
                          const crs_graph_type& graph,
                          const RowInfo& rowInfo,
                          const LocalOrdinal inds[],
                          const impl_scalar_type newVals[],
                          const LocalOrdinal numElts) const
  {
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
 
    const bool sorted = graph.isSorted ();
 
    size_t hint = 0; // Guess for the current index k into rowVals
    LO numValid = 0; // number of valid local column indices
 
    // NOTE (mfh 11 Oct 2015) This method assumes UVM.  More
    // accurately, it assumes that the host execution space can
    // access data in both InputMemorySpace and ValsMemorySpace.
 
    if (graph.isLocallyIndexed ()) {
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getLocalKokkosRowView (rowInfo);
 
      for (LO j = 0; j < numElts; ++j) {
        const LO lclColInd = inds[j];
        const size_t offset =
          KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                       lclColInd, hint, sorted);
        if (offset != rowInfo.numEntries) {
          rowVals[offset] = newVals[j];
          hint = offset + 1;
          ++numValid;
        }
      }
    }
    else if (graph.isGloballyIndexed ()) {
      if (graph.colMap_.is_null ()) {
        return Teuchos::OrdinalTraits<LO>::invalid ();
      }
      const map_type colMap = * (graph.colMap_);
 
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getGlobalKokkosRowView (rowInfo);
 
      for (LO j = 0; j < numElts; ++j) {
        const GO gblColInd = colMap.getGlobalElement (inds[j]);
        if (gblColInd != Teuchos::OrdinalTraits<GO>::invalid ()) {
          const size_t offset =
            KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                         gblColInd, hint, sorted);
          if (offset != rowInfo.numEntries) {
            rowVals[offset] = newVals[j];
            hint = offset + 1;
            ++numValid;
          }
        }
      }
    }
    // NOTE (mfh 26 Jun 2014, 26 Nov 2015) In the current version of
    // CrsGraph and CrsMatrix, it's possible for a matrix (or graph)
    // to be neither locally nor globally indexed on a process.
    // This means that the graph or matrix has no entries on that
    // process.  Epetra also works like this.  It's related to lazy
    // allocation (on first insertion, not at graph / matrix
    // construction).  Lazy allocation will go away because it is
    // not thread scalable.
 
    return numValid;
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  replaceLocalValues (const LocalOrdinal localRow,
                      const Teuchos::ArrayView<const LocalOrdinal>& lclCols,
                      const Teuchos::ArrayView<const Scalar>& vals) const
  {
    typedef LocalOrdinal LO;
 
    const LO numInputEnt = static_cast<LO> (lclCols.size ());
    if (static_cast<LO> (vals.size ()) != numInputEnt) {
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    const LO* const inputInds = lclCols.getRawPtr ();
    const Scalar* const inputVals = vals.getRawPtr ();
    return this->replaceLocalValues (localRow, numInputEnt,
                                     inputVals, inputInds);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  replaceLocalValues (const LocalOrdinal localRow,
                      const LocalOrdinal numEnt,
                      const Scalar inputVals[],
                      const LocalOrdinal inputCols[]) const
  {
    typedef impl_scalar_type IST;
    typedef LocalOrdinal LO;
 
    if (! this->isFillActive () || this->staticGraph_.is_null ()) {
      // Fill must be active and the "nonconst" graph must exist.
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    const crs_graph_type& graph = * (this->staticGraph_);
    const RowInfo rowInfo = graph.getRowInfo (localRow);
 
    if (rowInfo.localRow == Teuchos::OrdinalTraits<size_t>::invalid ()) {
      // The calling process does not own this row, so it is not
      // allowed to modify its values.
      return static_cast<LO> (0);
    }
    auto curRowVals = this->getRowViewNonConst (rowInfo);
    const IST* const inVals = reinterpret_cast<const IST*> (inputVals);
    return this->replaceLocalValuesImpl (curRowVals.data (), graph, rowInfo,
                                         inputCols, inVals, numEnt);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  replaceGlobalValuesImpl (impl_scalar_type rowVals[],
                           const crs_graph_type& graph,
                           const RowInfo& rowInfo,
                           const GlobalOrdinal inds[],
                           const impl_scalar_type newVals[],
                           const LocalOrdinal numElts) const
  {
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
 
    const bool sorted = graph.isSorted ();
 
    size_t hint = 0; // guess at the index's relative offset in the row
    LO numValid = 0; // number of valid input column indices
 
    // NOTE (mfh 11 Oct 2015) This method assumes UVM.  More
    // accurately, it assumes that the host execution space can
    // access data in all the Views.
 
    if (graph.isLocallyIndexed ()) {
      // NOTE (mfh 04 Nov 2015) Dereferencing an RCP or reading its
      // pointer does NOT change its reference count.  Thus, this
      // code is still thread safe.
      if (graph.colMap_.is_null ()) {
        // NO input column indices are valid in this case, since if
        // the column Map is null on the calling process, then the
        // calling process owns no graph entries.
        return numValid;
      }
      const map_type& colMap = * (graph.colMap_);
 
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getLocalKokkosRowView (rowInfo);
      const LO LINV = Teuchos::OrdinalTraits<LO>::invalid ();
      for (LO j = 0; j < numElts; ++j) {
        const LO lclColInd = colMap.getLocalElement (inds[j]);
        if (lclColInd != LINV) {
          const size_t offset =
            KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                         lclColInd, hint, sorted);
          if (offset != rowInfo.numEntries) {
            rowVals[offset] = newVals[j];
            hint = offset + 1;
            numValid++;
          }
        }
      }
    }
    else if (graph.isGloballyIndexed ()) {
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getGlobalKokkosRowView (rowInfo);
 
      for (LO j = 0; j < numElts; ++j) {
        const GO gblColInd = inds[j];
        const size_t offset =
          KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                       gblColInd, hint, sorted);
        if (offset != rowInfo.numEntries) {
          rowVals[offset] = newVals[j];
          hint = offset + 1;
          numValid++;
        }
      }
    }
    // If the graph is neither locally nor globally indexed on the
    // calling process, that means the calling process has no graph
    // entries.  Thus, none of the input column indices are valid.
 
    return numValid;
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  replaceGlobalValues (const GlobalOrdinal globalRow,
                       const Teuchos::ArrayView<const GlobalOrdinal>& inputGblColInds,
                       const Teuchos::ArrayView<const Scalar>& inputVals) const
  {
    typedef LocalOrdinal LO;
 
    const LO numInputEnt = static_cast<LO> (inputGblColInds.size ());
    if (static_cast<LO> (inputVals.size ()) != numInputEnt) {
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    return this->replaceGlobalValues (globalRow, numInputEnt,
                                      inputVals.getRawPtr (),
                                      inputGblColInds.getRawPtr ());
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  replaceGlobalValues (const GlobalOrdinal globalRow,
                       const LocalOrdinal numEnt,
                       const Scalar inputVals[],
                       const GlobalOrdinal inputGblColInds[]) const
  {
    typedef impl_scalar_type IST;
    typedef LocalOrdinal LO;
 
    if (! this->isFillActive () || this->staticGraph_.is_null ()) {
      // Fill must be active and the "nonconst" graph must exist.
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    const crs_graph_type& graph = * (this->staticGraph_);
 
    const RowInfo rowInfo = graph.getRowInfoFromGlobalRowIndex (globalRow);
    if (rowInfo.localRow == Teuchos::OrdinalTraits<size_t>::invalid ()) {
      // The input local row is invalid on the calling process,
      // which means that the calling process summed 0 entries.
      return static_cast<LO> (0);
    }
 
    auto curRowVals = this->getRowViewNonConst (rowInfo);
    const IST* const inVals = reinterpret_cast<const IST*> (inputVals);
    return this->replaceGlobalValuesImpl (curRowVals.data (), graph, rowInfo,
                                          inputGblColInds, inVals, numEnt);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  sumIntoGlobalValuesImpl (impl_scalar_type rowVals[],
                           const crs_graph_type& graph,
                           const RowInfo& rowInfo,
                           const GlobalOrdinal inds[],
                           const impl_scalar_type newVals[],
                           const LocalOrdinal numElts,
                           const bool atomic) const
  {
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
 
    const bool sorted = graph.isSorted ();
 
    size_t hint = 0; // guess at the index's relative offset in the row
    LO numValid = 0; // number of valid input column indices
 
    // NOTE (mfh 11 Oct 2015) This method assumes UVM.  More
    // accurately, it assumes that the host execution space can
    // access data in both InputMemorySpace and ValsMemorySpace.
 
    if (graph.isLocallyIndexed ()) {
      // NOTE (mfh 04 Nov 2015) Dereferencing an RCP or reading its
      // pointer does NOT change its reference count.  Thus, this
      // code is still thread safe.
      if (graph.colMap_.is_null ()) {
        // NO input column indices are valid in this case, since if
        // the column Map is null on the calling process, then the
        // calling process owns no graph entries.
        return numValid;
      }
      const map_type& colMap = * (graph.colMap_);
 
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getLocalKokkosRowView (rowInfo);
      const LO LINV = Teuchos::OrdinalTraits<LO>::invalid ();
 
      for (LO j = 0; j < numElts; ++j) {
        const LO lclColInd = colMap.getLocalElement (inds[j]);
        if (lclColInd != LINV) {
          const size_t offset =
            KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                         lclColInd, hint, sorted);
          if (offset != rowInfo.numEntries) {
            if (atomic) {
              Kokkos::atomic_add (&rowVals[offset], newVals[j]);
            }
            else {
              rowVals[offset] += newVals[j];
            }
            hint = offset + 1;
            numValid++;
          }
        }
      }
    }
    else if (graph.isGloballyIndexed ()) {
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getGlobalKokkosRowView (rowInfo);
 
      for (LO j = 0; j < numElts; ++j) {
        const GO gblColInd = inds[j];
        const size_t offset =
          KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                       gblColInd, hint, sorted);
        if (offset != rowInfo.numEntries) {
          if (atomic) {
            Kokkos::atomic_add (&rowVals[offset], newVals[j]);
          }
          else {
            rowVals[offset] += newVals[j];
          }
          hint = offset + 1;
          numValid++;
        }
      }
    }
    // If the graph is neither locally nor globally indexed on the
    // calling process, that means the calling process has no graph
    // entries.  Thus, none of the input column indices are valid.
 
    return numValid;
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  sumIntoGlobalValues (const GlobalOrdinal gblRow,
                       const Teuchos::ArrayView<const GlobalOrdinal>& inputGblColInds,
                       const Teuchos::ArrayView<const Scalar>& inputVals,
                       const bool atomic)
  {
    typedef LocalOrdinal LO;
 
    const LO numInputEnt = static_cast<LO> (inputGblColInds.size ());
    if (static_cast<LO> (inputVals.size ()) != numInputEnt) {
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    return this->sumIntoGlobalValues (gblRow, numInputEnt,
                                      inputVals.getRawPtr (),
                                      inputGblColInds.getRawPtr (),
                                      atomic);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  sumIntoGlobalValues (const GlobalOrdinal gblRow,
                       const LocalOrdinal numInputEnt,
                       const Scalar inputVals[],
                       const GlobalOrdinal inputGblColInds[],
                       const bool atomic)
  {
    typedef impl_scalar_type IST;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
 
    if (! this->isFillActive () || this->staticGraph_.is_null ()) {
      // Fill must be active and the "nonconst" graph must exist.
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    const crs_graph_type& graph = * (this->staticGraph_);
 
    const RowInfo rowInfo = graph.getRowInfoFromGlobalRowIndex (gblRow);
    if (rowInfo.localRow == Teuchos::OrdinalTraits<size_t>::invalid ()) {
      // mfh 23 Mar 2017, 26 Jul 2017: This branch may not be not
      // thread safe in a debug build, in part because it uses
      // Teuchos::ArrayView, and in part because of the data structure
      // used to stash outgoing entries.
      using Teuchos::ArrayView;
      ArrayView<const GO> inputGblColInds_av (numInputEnt == 0 ? NULL :
                                              inputGblColInds, numInputEnt);
      ArrayView<const Scalar> inputVals_av (numInputEnt == 0 ? NULL :
                                            inputVals, numInputEnt);
      // gblRow is not in the row Map on the calling process, so stash
      // the given entries away in a separate data structure.
      // globalAssemble() (called during fillComplete()) will exchange
      // that data and sum it in using sumIntoGlobalValues().
      this->insertNonownedGlobalValues (gblRow, inputGblColInds_av,
                                        inputVals_av);
      // FIXME (mfh 08 Jul 2014) It's not clear what to return here,
      // since we won't know whether the given indices were valid
      // until globalAssemble (called in fillComplete) is called.
      // That's why insertNonownedGlobalValues doesn't return
      // anything.  Just for consistency, I'll return the number of
      // entries that the user gave us.
      return numInputEnt;
    }
    else { // input row is in the row Map on the calling process
      auto curRowVals = this->getRowViewNonConst (rowInfo);
      const IST* const inVals = reinterpret_cast<const IST*> (inputVals);
      return this->sumIntoGlobalValuesImpl (curRowVals.data (), graph, rowInfo,
                                            inputGblColInds, inVals,
                                            numInputEnt, atomic);
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  transformLocalValues (const LocalOrdinal lclRow,
                        const LocalOrdinal numInputEnt,
                        const impl_scalar_type inputVals[],
                        const LocalOrdinal inputCols[],
                        std::function<impl_scalar_type (const impl_scalar_type&, const impl_scalar_type&) > f,
                        const bool atomic) const
  {
    using Tpetra::Details::OrdinalTraits;
    typedef LocalOrdinal LO;
 
    if (! this->isFillActive () || this->staticGraph_.is_null ()) {
      // Fill must be active and the "nonconst" graph must exist.
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    const crs_graph_type& graph = * (this->staticGraph_);
    const RowInfo rowInfo = graph.getRowInfo (lclRow);
 
    if (rowInfo.localRow == OrdinalTraits<size_t>::invalid ()) {
      // The calling process does not own this row, so it is not
      // allowed to modify its values.
      return static_cast<LO> (0);
    }
    auto curRowVals = this->getRowViewNonConst (rowInfo);
    return this->transformLocalValues (curRowVals.data (), graph,
                                       rowInfo, inputCols, inputVals,
                                       numInputEnt, f, atomic);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  transformGlobalValues (const GlobalOrdinal gblRow,
                         const LocalOrdinal numInputEnt,
                         const impl_scalar_type inputVals[],
                         const GlobalOrdinal inputCols[],
                         std::function<impl_scalar_type (const impl_scalar_type&, const impl_scalar_type&) > f,
                         const bool atomic) const
  {
    using Tpetra::Details::OrdinalTraits;
    typedef LocalOrdinal LO;
 
    if (! this->isFillActive () || this->staticGraph_.is_null ()) {
      // Fill must be active and the "nonconst" graph must exist.
      return OrdinalTraits<LO>::invalid ();
    }
    const crs_graph_type& graph = * (this->staticGraph_);
    const RowInfo rowInfo = graph.getRowInfoFromGlobalRowIndex (gblRow);
 
    if (rowInfo.localRow == OrdinalTraits<size_t>::invalid ()) {
      // The calling process does not own this row, so it is not
      // allowed to modify its values.
      return static_cast<LO> (0);
    }
    auto curRowVals = this->getRowViewNonConst (rowInfo);
    return this->transformGlobalValues (curRowVals.data (), graph,
                                        rowInfo, inputCols, inputVals,
                                        numInputEnt, f, atomic);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  transformLocalValues (impl_scalar_type rowVals[],
                        const crs_graph_type& graph,
                        const RowInfo& rowInfo,
                        const LocalOrdinal inds[],
                        const impl_scalar_type newVals[],
                        const LocalOrdinal numElts,
                        std::function<impl_scalar_type (const impl_scalar_type&, const impl_scalar_type&) > f,
                        const bool atomic) const
  {
    typedef impl_scalar_type ST;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
 
    //if (newVals.extent (0) != inds.extent (0)) {
    // The sizes of the input arrays must match.
    //return Tpetra::Details::OrdinalTraits<LO>::invalid ();
    //}
    //const LO numElts = static_cast<LO> (inds.extent (0));
    const bool sorted = graph.isSorted ();
 
    LO numValid = 0; // number of valid input column indices
    size_t hint = 0; // Guess for the current index k into rowVals
 
    if (graph.isLocallyIndexed ()) {
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getLocalKokkosRowView (rowInfo);
 
      for (LO j = 0; j < numElts; ++j) {
        const LO lclColInd = inds[j];
        const size_t offset =
          KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                       lclColInd, hint, sorted);
        if (offset != rowInfo.numEntries) {
          if (atomic) {
            // NOTE (mfh 30 Nov 2015) The commented-out code is
            // wrong because another thread may have changed
            // rowVals[offset] between those two lines of code.
            //
            //const ST newVal = f (rowVals[offset], newVals[j]);
            //Kokkos::atomic_assign (&rowVals[offset], newVal);
 
            volatile ST* const dest = &rowVals[offset];
            (void) atomic_binary_function_update (dest, newVals[j], f);
          }
          else {
            // use binary function f
            rowVals[offset] = f (rowVals[offset], newVals[j]);
          }
          hint = offset + 1;
          ++numValid;
        }
      }
    }
    else if (graph.isGloballyIndexed ()) {
      // NOTE (mfh 26 Nov 2015) Dereferencing an RCP or reading its
      // pointer does NOT change its reference count.  Thus, this
      // code is still thread safe.
      if (graph.colMap_.is_null ()) {
        // NO input column indices are valid in this case.  Either
        // the column Map hasn't been set yet (so local indices
        // don't exist yet), or the calling process owns no graph
        // entries.
        return numValid;
      }
      const map_type& colMap = * (graph.colMap_);
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getGlobalKokkosRowView (rowInfo);
 
      const GO GINV = Teuchos::OrdinalTraits<GO>::invalid ();
      for (LO j = 0; j < numElts; ++j) {
        const GO gblColInd = colMap.getGlobalElement (inds[j]);
        if (gblColInd != GINV) {
          const size_t offset =
            KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                         gblColInd, hint, sorted);
          if (offset != rowInfo.numEntries) {
            if (atomic) {
              // NOTE (mfh 30 Nov 2015) The commented-out code is
              // wrong because another thread may have changed
              // rowVals[offset] between those two lines of code.
              //
              //const ST newVal = f (rowVals[offset], newVals[j]);
              //Kokkos::atomic_assign (&rowVals[offset], newVal);
 
              volatile ST* const dest = &rowVals[offset];
              (void) atomic_binary_function_update (dest, newVals[j], f);
            }
            else {
              // use binary function f
              rowVals[offset] = f (rowVals[offset], newVals[j]);
            }
            hint = offset + 1;
            numValid++;
          }
        }
      }
    }
    // If the graph is neither locally nor globally indexed on the
    // calling process, that means the calling process has no graph
    // entries.  Thus, none of the input column indices are valid.
 
    return numValid;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  transformGlobalValues (impl_scalar_type rowVals[],
                         const crs_graph_type& graph,
                         const RowInfo& rowInfo,
                         const GlobalOrdinal inds[],
                         const impl_scalar_type newVals[],
                         const LocalOrdinal numElts,
                         std::function<impl_scalar_type (const impl_scalar_type&, const impl_scalar_type&) > f,
                         const bool atomic) const
  {
    typedef impl_scalar_type ST;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
 
    //if (newVals.extent (0) != inds.extent (0)) {
    // The sizes of the input arrays must match.
    //return Tpetra::Details::OrdinalTraits<LO>::invalid ();
    //}
    //const LO numElts = static_cast<LO> (inds.extent (0));
    const bool sorted = graph.isSorted ();
 
    LO numValid = 0; // number of valid input column indices
    size_t hint = 0; // Guess for the current index k into rowVals
 
    if (graph.isGloballyIndexed ()) {
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getGlobalKokkosRowView (rowInfo);
 
      for (LO j = 0; j < numElts; ++j) {
        const GO gblColInd = inds[j];
        const size_t offset =
          KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                       gblColInd, hint, sorted);
        if (offset != rowInfo.numEntries) {
          if (atomic) {
            // NOTE (mfh 30 Nov 2015) The commented-out code is
            // wrong because another thread may have changed
            // rowVals[offset] between those two lines of code.
            //
            //const ST newVal = f (rowVals[offset], newVals[j]);
            //Kokkos::atomic_assign (&rowVals[offset], newVal);
 
            volatile ST* const dest = &rowVals[offset];
            (void) atomic_binary_function_update (dest, newVals[j], f);
          }
          else {
            // use binary function f
            rowVals[offset] = f (rowVals[offset], newVals[j]);
          }
          hint = offset + 1;
          ++numValid;
        }
      }
    }
    else if (graph.isLocallyIndexed ()) {
      // NOTE (mfh 26 Nov 2015) Dereferencing an RCP or reading its
      // pointer does NOT change its reference count.  Thus, this
      // code is still thread safe.
      if (graph.colMap_.is_null ()) {
        // NO input column indices are valid in this case.  Either the
        // column Map hasn't been set yet (so local indices don't
        // exist yet), or the calling process owns no graph entries.
        return numValid;
      }
      const map_type& colMap = * (graph.colMap_);
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getLocalKokkosRowView (rowInfo);
 
      const LO LINV = Teuchos::OrdinalTraits<LO>::invalid ();
      for (LO j = 0; j < numElts; ++j) {
        const LO lclColInd = colMap.getLocalElement (inds[j]);
        if (lclColInd != LINV) {
          const size_t offset =
            KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                         lclColInd, hint, sorted);
          if (offset != rowInfo.numEntries) {
            if (atomic) {
              // NOTE (mfh 30 Nov 2015) The commented-out code is
              // wrong because another thread may have changed
              // rowVals[offset] between those two lines of code.
              //
              //const ST newVal = f (rowVals[offset], newVals[j]);
              //Kokkos::atomic_assign (&rowVals[offset], newVal);
 
              volatile ST* const dest = &rowVals[offset];
              (void) atomic_binary_function_update (dest, newVals[j], f);
            }
            else {
              // use binary function f
              rowVals[offset] = f (rowVals[offset], newVals[j]);
            }
            hint = offset + 1;
            numValid++;
          }
        }
      }
    }
    // If the graph is neither locally nor globally indexed on the
    // calling process, that means the calling process has no graph
    // entries.  Thus, none of the input column indices are valid.
 
    return numValid;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  sumIntoLocalValuesImpl (impl_scalar_type rowVals[],
                          const crs_graph_type& graph,
                          const RowInfo& rowInfo,
                          const LocalOrdinal inds[],
                          const impl_scalar_type newVals[],
                          const LocalOrdinal numElts,
                          const bool atomic) const
  {
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
 
    const bool sorted = graph.isSorted ();
 
    size_t hint = 0; // Guess for the current index k into rowVals
    LO numValid = 0; // number of valid local column indices
 
    // NOTE (mfh 11 Oct 2015) This method assumes UVM.  More
    // accurately, it assumes that the host execution space can
    // access data in both InputMemorySpace and ValsMemorySpace.
 
    if (graph.isLocallyIndexed ()) {
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getLocalKokkosRowView (rowInfo);
 
      for (LO j = 0; j < numElts; ++j) {
        const LO lclColInd = inds[j];
        const size_t offset =
          KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                       lclColInd, hint, sorted);
        if (offset != rowInfo.numEntries) {
          if (atomic) {
            Kokkos::atomic_add (&rowVals[offset], newVals[j]);
          }
          else {
            rowVals[offset] += newVals[j];
          }
          hint = offset + 1;
          ++numValid;
        }
      }
    }
    else if (graph.isGloballyIndexed ()) {
      if (graph.colMap_.is_null ()) {
        return Teuchos::OrdinalTraits<LO>::invalid ();
      }
      const map_type colMap = * (graph.colMap_);
 
      // Get a view of the column indices in the row.  This amortizes
      // the cost of getting the view over all the entries of inds.
      auto colInds = graph.getGlobalKokkosRowView (rowInfo);
 
      for (LO j = 0; j < numElts; ++j) {
        const GO gblColInd = colMap.getGlobalElement (inds[j]);
        if (gblColInd != Teuchos::OrdinalTraits<GO>::invalid ()) {
          const size_t offset =
            KokkosSparse::findRelOffset (colInds, rowInfo.numEntries,
                                         gblColInd, hint, sorted);
          if (offset != rowInfo.numEntries) {
            if (atomic) {
              Kokkos::atomic_add (&rowVals[offset], newVals[j]);
            }
            else {
              rowVals[offset] += newVals[j];
            }
            hint = offset + 1;
            ++numValid;
          }
        }
      }
    }
    // NOTE (mfh 26 Jun 2014, 26 Nov 2015) In the current version of
    // CrsGraph and CrsMatrix, it's possible for a matrix (or graph)
    // to be neither locally nor globally indexed on a process.
    // This means that the graph or matrix has no entries on that
    // process.  Epetra also works like this.  It's related to lazy
    // allocation (on first insertion, not at graph / matrix
    // construction).  Lazy allocation will go away because it is
    // not thread scalable.
 
    return numValid;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  sumIntoLocalValues (const LocalOrdinal localRow,
                      const Teuchos::ArrayView<const LocalOrdinal>& indices,
                      const Teuchos::ArrayView<const Scalar>& values,
                      const bool atomic) const
  {
    typedef LocalOrdinal LO;
 
    const LO numInputEnt = static_cast<LO> (indices.size ());
    if (static_cast<LO> (values.size ()) != numInputEnt) {
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    const LO* const inputInds = indices.getRawPtr ();
    const Scalar* const inputVals = values.getRawPtr ();
    return this->sumIntoLocalValues (localRow, numInputEnt,
                                     inputVals, inputInds, atomic);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  sumIntoLocalValues (const LocalOrdinal localRow,
                      const LocalOrdinal numEnt,
                      const Scalar vals[],
                      const LocalOrdinal cols[],
                      const bool atomic) const
  {
    typedef impl_scalar_type IST;
    typedef LocalOrdinal LO;
 
    if (! this->isFillActive () || this->staticGraph_.is_null ()) {
      // Fill must be active and the "nonconst" graph must exist.
      return Teuchos::OrdinalTraits<LO>::invalid ();
    }
    const crs_graph_type& graph = * (this->staticGraph_);
    const RowInfo rowInfo = graph.getRowInfo (localRow);
 
    if (rowInfo.localRow == Teuchos::OrdinalTraits<size_t>::invalid ()) {
      // The calling process does not own this row, so it is not
      // allowed to modify its values.
      return static_cast<LO> (0);
    }
    auto curRowVals = this->getRowViewNonConst (rowInfo);
    const IST* const inputVals = reinterpret_cast<const IST*> (vals);
    return this->sumIntoLocalValuesImpl (curRowVals.data (), graph, rowInfo,
                                         cols, inputVals, numEnt, atomic);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::ArrayView<const typename CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::impl_scalar_type>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getView (RowInfo rowinfo) const
  {
    using Kokkos::MemoryUnmanaged;
    using Kokkos::View;
    using Teuchos::ArrayView;
    typedef impl_scalar_type ST;
    typedef std::pair<size_t, size_t> range_type;
 
    if (k_values1D_.extent (0) != 0 && rowinfo.allocSize > 0) {
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION(
        rowinfo.offset1D + rowinfo.allocSize > k_values1D_.extent (0),
        std::range_error, "Tpetra::CrsMatrix::getView: Invalid access "
        "to 1-D storage of values." << std::endl << "rowinfo.offset1D (" <<
        rowinfo.offset1D << ") + rowinfo.allocSize (" << rowinfo.allocSize <<
        ") > k_values1D_.extent(0) (" << k_values1D_.extent (0) << ").");
#endif // HAVE_TPETRA_DEBUG
      range_type range (rowinfo.offset1D, rowinfo.offset1D + rowinfo.allocSize);
      typedef View<const ST*, execution_space, MemoryUnmanaged> subview_type;
      // mfh 23 Nov 2015: Don't just create a subview of k_values1D_
      // directly, because that first creates a _managed_ subview,
      // then returns an unmanaged version of that.  That touches the
      // reference count, which costs performance in a measurable way.
      // Instead, we create a temporary unmanaged view, then create
      // the subview from that.
      subview_type sv = Kokkos::subview (subview_type (k_values1D_), range);
      const ST* const sv_raw = (rowinfo.allocSize == 0) ? NULL : sv.data ();
      return ArrayView<const ST> (sv_raw, rowinfo.allocSize);
    }
    else if (values2D_ != Teuchos::null) {
      return values2D_[rowinfo.localRow] ();
    }
    else {
      return ArrayView<impl_scalar_type> ();
    }
  }
 
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getViewRawConst (const impl_scalar_type*& vals,
                   LocalOrdinal& numEnt,
                   const RowInfo& rowinfo) const
  {
    if (k_values1D_.extent (0) != 0 && rowinfo.allocSize > 0) {
#ifdef HAVE_TPETRA_DEBUG
      if (rowinfo.offset1D + rowinfo.allocSize > k_values1D_.extent (0)) {
        vals = NULL;
        numEnt = 0;
        return Teuchos::OrdinalTraits<LocalOrdinal>::invalid ();
      }
#endif // HAVE_TPETRA_DEBUG
      vals = k_values1D_.data () + rowinfo.offset1D;
      numEnt = rowinfo.allocSize;
    }
    else if (! values2D_.is_null ()) {
#ifdef HAVE_TPETRA_DEBUG
      if (rowinfo.localRow >= static_cast<size_t> (values2D_.size ())) {
        vals = NULL;
        numEnt = 0;
        return Teuchos::OrdinalTraits<LocalOrdinal>::invalid ();
      }
#endif // HAVE_TPETRA_DEBUG
      // Use const reference so that we don't update ArrayRCP's
      // reference count, which is not thread safe.
      const auto& curRow = values2D_[rowinfo.localRow];
      vals = curRow.getRawPtr ();
      numEnt = curRow.size ();
    }
    else {
      vals = NULL;
      numEnt = 0;
    }
 
    return static_cast<LocalOrdinal> (0);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getViewRaw (impl_scalar_type*& vals,
              LocalOrdinal& numEnt,
              const RowInfo& rowinfo) const
  {
    const impl_scalar_type* valsConst;
    const LocalOrdinal err = this->getViewRawConst (valsConst, numEnt, rowinfo);
    vals = const_cast<impl_scalar_type*> (valsConst);
    return err;
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Kokkos::View<const typename CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::impl_scalar_type*,
               typename CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::execution_space,
               Kokkos::MemoryUnmanaged>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getRowView (const RowInfo& rowInfo) const
  {
    using Kokkos::MemoryUnmanaged;
    using Kokkos::View;
    typedef impl_scalar_type ST;
    typedef View<const ST*, execution_space, MemoryUnmanaged> subview_type;
    typedef std::pair<size_t, size_t> range_type;
 
    if (k_values1D_.extent (0) != 0 && rowInfo.allocSize > 0) {
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION
        (rowInfo.offset1D + rowInfo.allocSize > this->k_values1D_.extent (0),
         std::range_error, "Tpetra::CrsMatrix::getRowView: Invalid access "
         "to 1-D storage of values.  rowInfo.offset1D ("
         << rowInfo.offset1D << ") + rowInfo.allocSize (" << rowInfo.allocSize
         << ") > this->k_values1D_.extent(0) ("
         << this->k_values1D_.extent (0) << ").");
#endif // HAVE_TPETRA_DEBUG
      range_type range (rowInfo.offset1D, rowInfo.offset1D + rowInfo.allocSize);
      // mfh 23 Nov 2015: Don't just create a subview of k_values1D_
      // directly, because that first creates a _managed_ subview,
      // then returns an unmanaged version of that.  That touches the
      // reference count, which costs performance in a measurable way.
      // Instead, we create a temporary unmanaged view, then create
      // the subview from that.
      return Kokkos::subview (subview_type (this->k_values1D_), range);
    }
    else if (this->values2D_ != Teuchos::null) {
      // Use a reference, so that I don't touch the Teuchos::ArrayView
      // reference count in a debug build.  (It has no reference count
      // in a release build.)  This ensures thread safety.
      auto& rowView = this->values2D_[rowInfo.localRow];
      return subview_type (rowView.getRawPtr (), rowView.size ());
    }
    else {
      return subview_type ();
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Kokkos::View<typename CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::impl_scalar_type*,
               typename CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::execution_space,
               Kokkos::MemoryUnmanaged>
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getRowViewNonConst (const RowInfo& rowInfo) const
  {
    using Kokkos::MemoryUnmanaged;
    using Kokkos::View;
    typedef impl_scalar_type ST;
    typedef View<ST*, execution_space, MemoryUnmanaged> subview_type;
    typedef std::pair<size_t, size_t> range_type;
 
    if (k_values1D_.extent (0) != 0 && rowInfo.allocSize > 0) {
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION
        (rowInfo.offset1D + rowInfo.allocSize > this->k_values1D_.extent (0),
         std::range_error, "Tpetra::CrsMatrix::getRowViewNonConst: Invalid "
         "access to 1-D storage of values.  rowInfo.offset1D ("
         << rowInfo.offset1D << ") + rowInfo.allocSize (" << rowInfo.allocSize
         << ") > this->k_values1D_.extent(0) ("
         << this->k_values1D_.extent (0) << ").");
#endif // HAVE_TPETRA_DEBUG
      range_type range (rowInfo.offset1D, rowInfo.offset1D + rowInfo.allocSize);
      // mfh 23 Nov 2015: Don't just create a subview of k_values1D_
      // directly, because that first creates a _managed_ subview,
      // then returns an unmanaged version of that.  That touches the
      // reference count, which costs performance in a measurable way.
      // Instead, we create a temporary unmanaged view, then create
      // the subview from that.
      return Kokkos::subview (subview_type (this->k_values1D_), range);
    }
    else if (this->values2D_ != Teuchos::null) {
      // Use a reference, so that I don't touch the Teuchos::ArrayView
      // reference count in a debug build.  (It has no reference count
      // in a release build.)  This ensures thread safety.
      auto& rowView = this->values2D_[rowInfo.localRow];
      return subview_type (rowView.getRawPtr (), rowView.size ());
    }
    else {
      return subview_type ();
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::ArrayView<typename CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::impl_scalar_type>
  CrsMatrix<Scalar, LocalOrdinal,GlobalOrdinal, Node>::
  getViewNonConst (const RowInfo& rowinfo) const
  {
    return Teuchos::av_const_cast<impl_scalar_type> (this->getView (rowinfo));
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getLocalRowCopy (LocalOrdinal localRow,
                   const Teuchos::ArrayView<LocalOrdinal>& indices,
                   const Teuchos::ArrayView<Scalar>& values,
                   size_t& numEntries) const
  {
    using Teuchos::ArrayView;
    using Teuchos::av_reinterpret_cast;
    const char tfecfFuncName[] = "getLocalRowCopy: ";
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! this->hasColMap (), std::runtime_error,
       "The matrix does not have a column Map yet.  This means we don't have "
       "local indices for columns yet, so it doesn't make sense to call this "
       "method.  If the matrix doesn't have a column Map yet, you should call "
       "fillComplete on it first.");
 
    const RowInfo rowinfo = staticGraph_->getRowInfo (localRow);
    const size_t theNumEntries = rowinfo.numEntries;
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (static_cast<size_t> (indices.size ()) < theNumEntries ||
       static_cast<size_t> (values.size ()) < theNumEntries,
       std::runtime_error, "Row with local index " << localRow << " has " <<
       theNumEntries << " entry/ies, but indices.size() = " <<
       indices.size () << " and values.size() = " << values.size () << ".");
    numEntries = theNumEntries; // first side effect
 
    if (rowinfo.localRow != Teuchos::OrdinalTraits<size_t>::invalid ()) {
      if (staticGraph_->isLocallyIndexed ()) {
        const LocalOrdinal* curLclInds;
        const impl_scalar_type* curVals;
        LocalOrdinal numSpots; // includes both current entries and extra space
 
        // If we got this far, rowinfo should be correct and should
        // refer to a valid local row.  Thus, these error checks are
        // superfluous, but we retain them in a debug build.
#ifdef HAVE_TPETRA_DEBUG
        int err =
          staticGraph_->getLocalViewRawConst (curLclInds, numSpots, rowinfo);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (err != static_cast<LocalOrdinal> (0), std::logic_error,
           "staticGraph_->getLocalViewRawConst returned nonzero error code "
           << err << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (numSpots) < theNumEntries, std::logic_error,
           "numSpots = " << numSpots << " < theNumEntries = " << theNumEntries
           << ".");
        const LocalOrdinal numSpotsBefore = numSpots;
        err = getViewRawConst (curVals, numSpots, rowinfo);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (err != static_cast<LocalOrdinal> (0), std::logic_error,
           "getViewRaw returned nonzero error code " << err << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (numSpotsBefore != numSpots, std::logic_error,
           "numSpotsBefore = " << numSpotsBefore << " != numSpots = "
           << numSpots << ".");
#else
        (void) staticGraph_->getLocalViewRawConst (curLclInds, numSpots, rowinfo);
        (void) getViewRawConst (curVals, numSpots, rowinfo);
#endif // HAVE_TPETRA_DEBUG
 
        for (size_t j = 0; j < theNumEntries; ++j) {
          values[j] = curVals[j];
          indices[j] = curLclInds[j];
        }
      }
      else if (staticGraph_->isGloballyIndexed ()) {
        // Don't call getColMap(), because it touches RCP's reference count.
        const map_type& colMap = * (staticGraph_->colMap_);
        const GlobalOrdinal* curGblInds;
        const impl_scalar_type* curVals;
        LocalOrdinal numSpots; // includes both current entries and extra space
 
        // If we got this far, rowinfo should be correct and should
        // refer to a valid local row.  Thus, these error checks are
        // superfluous, but we retain them in a debug build.
#ifdef HAVE_TPETRA_DEBUG
        int err =
          staticGraph_->getGlobalViewRawConst (curGblInds, numSpots, rowinfo);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (err != static_cast<LocalOrdinal> (0), std::logic_error,
           "staticGraph_->getGlobalViewRawConst returned nonzero error code "
           << err << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (numSpots) < theNumEntries, std::logic_error,
           "numSpots = " << numSpots << " < theNumEntries = " << theNumEntries
           << ".");
        const LocalOrdinal numSpotsBefore = numSpots;
        err = getViewRawConst (curVals, numSpots, rowinfo);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (err != static_cast<LocalOrdinal> (0), std::logic_error,
           "getViewRawConst returned nonzero error code " << err << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (numSpotsBefore != numSpots, std::logic_error,
           "numSpotsBefore = " << numSpotsBefore << " != numSpots = "
           << numSpots << ".");
#else
        (void) staticGraph_->getGlobalViewRawConst (curGblInds, numSpots, rowinfo);
        (void) getViewRawConst (curVals, numSpots, rowinfo);
#endif //HAVE_TPETRA_DEBUG
 
        for (size_t j = 0; j < theNumEntries; ++j) {
          values[j] = curVals[j];
          indices[j] = colMap.getLocalElement (curGblInds[j]);
        }
      }
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGlobalRowCopy (GlobalOrdinal globalRow,
                    const Teuchos::ArrayView<GlobalOrdinal>& indices,
                    const Teuchos::ArrayView<Scalar>& values,
                    size_t& numEntries) const
  {
    using Teuchos::ArrayView;
    using Teuchos::av_reinterpret_cast;
    const char tfecfFuncName[] = "getGlobalRowCopy: ";
 
    const RowInfo rowinfo =
      staticGraph_->getRowInfoFromGlobalRowIndex (globalRow);
    const size_t theNumEntries = rowinfo.numEntries;
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      static_cast<size_t> (indices.size ()) < theNumEntries ||
      static_cast<size_t> (values.size ()) < theNumEntries,
      std::runtime_error, "Row with global index " << globalRow << " has "
      << theNumEntries << " entry/ies, but indices.size() = " <<
      indices.size () << " and values.size() = " << values.size () << ".");
    numEntries = theNumEntries; // first side effect
 
    if (rowinfo.localRow != Teuchos::OrdinalTraits<size_t>::invalid ()) {
      if (staticGraph_->isLocallyIndexed ()) {
        const map_type& colMap = * (staticGraph_->colMap_);
        const LocalOrdinal* curLclInds;
        const impl_scalar_type* curVals;
        LocalOrdinal numSpots; // includes both current entries and extra space
 
        // If we got this far, rowinfo should be correct and should
        // refer to a valid local row.  Thus, these error checks are
        // superfluous, but we retain them in a debug build.
#ifdef HAVE_TPETRA_DEBUG
        int err =
          staticGraph_->getLocalViewRawConst (curLclInds, numSpots, rowinfo);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (err != static_cast<LocalOrdinal> (0), std::logic_error,
           "staticGraph_->getLocalViewRawConst returned nonzero error code "
           << err << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (numSpots) < theNumEntries, std::logic_error,
           "numSpots = " << numSpots << " < theNumEntries = " << theNumEntries
           << ".");
        const LocalOrdinal numSpotsBefore = numSpots;
        err = getViewRawConst (curVals, numSpots, rowinfo);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (err != static_cast<LocalOrdinal> (0), std::logic_error,
           "getViewRaw returned nonzero error code " << err << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (numSpotsBefore != numSpots, std::logic_error,
           "numSpotsBefore = " << numSpotsBefore << " != numSpots = "
           << numSpots << ".");
#else
        (void) staticGraph_->getLocalViewRawConst (curLclInds, numSpots, rowinfo);
        (void) getViewRawConst (curVals, numSpots, rowinfo);
#endif //HAVE_TPETRA_DEBUG
 
        for (size_t j = 0; j < theNumEntries; ++j) {
          values[j] = curVals[j];
          indices[j] = colMap.getGlobalElement (curLclInds[j]);
        }
      }
      else if (staticGraph_->isGloballyIndexed ()) {
        const GlobalOrdinal* curGblInds;
        const impl_scalar_type* curVals;
        LocalOrdinal numSpots; // includes both current entries and extra space
 
        // If we got this far, rowinfo should be correct and should
        // refer to a valid local row.  Thus, these error checks are
        // superfluous, but we retain them in a debug build.
#ifdef HAVE_TPETRA_DEBUG
        int err =
          staticGraph_->getGlobalViewRawConst (curGblInds, numSpots, rowinfo);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (err != static_cast<LocalOrdinal> (0), std::logic_error,
           "staticGraph_->getGlobalViewRawConst returned nonzero error code "
           << err << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (static_cast<size_t> (numSpots) < theNumEntries, std::logic_error,
           "numSpots = " << numSpots << " < theNumEntries = " << theNumEntries
           << ".");
        const LocalOrdinal numSpotsBefore = numSpots;
        err = getViewRawConst (curVals, numSpots, rowinfo);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (err != static_cast<LocalOrdinal> (0), std::logic_error,
           "getViewRawConst returned nonzero error code " << err << ".");
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (numSpotsBefore != numSpots, std::logic_error,
           "numSpotsBefore = " << numSpotsBefore << " != numSpots = "
           << numSpots << ".");
#else
        (void) staticGraph_->getGlobalViewRawConst (curGblInds, numSpots, rowinfo);
        (void) getViewRawConst (curVals, numSpots, rowinfo);
#endif //HAVE_TPETRA_DEBUG
 
        for (size_t j = 0; j < theNumEntries; ++j) {
          values[j] = curVals[j];
          indices[j] = curGblInds[j];
        }
      }
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getLocalRowView (LocalOrdinal localRow,
                   Teuchos::ArrayView<const LocalOrdinal>& indices,
                   Teuchos::ArrayView<const Scalar>& values) const
  {
    using Teuchos::ArrayView;
    using Teuchos::av_reinterpret_cast;
    typedef LocalOrdinal LO;
    const char tfecfFuncName[] = "getLocalRowView: ";
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      isGloballyIndexed (), std::runtime_error, "The matrix currently stores "
      "its indices as global indices, so you cannot get a view with local "
      "column indices.  If the matrix has a column Map, you may call "
      "getLocalRowCopy() to get local column indices; otherwise, you may get "
      "a view with global column indices by calling getGlobalRowCopy().");
    indices = Teuchos::null;
    values = Teuchos::null;
    const RowInfo rowinfo = staticGraph_->getRowInfo (localRow);
    if (rowinfo.localRow != Teuchos::OrdinalTraits<size_t>::invalid () &&
        rowinfo.numEntries > 0) {
      ArrayView<const LO> indTmp = staticGraph_->getLocalView (rowinfo);
      ArrayView<const Scalar> valTmp =
        av_reinterpret_cast<const Scalar> (this->getView (rowinfo));
      indices = indTmp (0, rowinfo.numEntries);
      values = valTmp (0, rowinfo.numEntries);
    }
 
#ifdef HAVE_TPETRA_DEBUG
    const char suffix[] = ".  This should never happen.  Please report this "
      "bug to the Tpetra developers.";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (static_cast<size_t> (indices.size ()) !=
       static_cast<size_t> (values.size ()), std::logic_error,
       "At the end of this method, for local row " << localRow << ", "
       "indices.size() = " << indices.size () << " != values.size () = "
       << values.size () << suffix);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (static_cast<size_t> (indices.size ()) !=
       static_cast<size_t> (rowinfo.numEntries), std::logic_error,
       "At the end of this method, for local row " << localRow << ", "
       "indices.size() = " << indices.size () << " != rowinfo.numEntries = "
       << rowinfo.numEntries << suffix);
    const size_t expectedNumEntries = getNumEntriesInLocalRow (localRow);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (rowinfo.numEntries != expectedNumEntries, std::logic_error, "At the end "
       "of this method, for local row " << localRow << ", rowinfo.numEntries = "
       << rowinfo.numEntries << " != getNumEntriesInLocalRow(localRow) = " <<
       expectedNumEntries << suffix);
#endif // HAVE_TPETRA_DEBUG
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getLocalRowView (const LocalOrdinal lclRow,
                   LocalOrdinal& numEnt,
                   const impl_scalar_type*& val,
                   const LocalOrdinal*& ind) const
  {
    typedef LocalOrdinal LO;
 
    // Don't call getCrsGraph(), because that modfies an RCP reference
    // count, which is not thread safe.  Checking whether an RCP is
    // null does NOT modify its reference count, and is therefore
    // thread safe.  Note that isGloballyIndexed() calls
    // getCrsGraph(), so we have to go to the graph directly.
    if (staticGraph_.is_null () || staticGraph_->isGloballyIndexed ()) {
      return Tpetra::Details::OrdinalTraits<LO>::invalid ();
    }
    else {
      const RowInfo rowInfo = staticGraph_->getRowInfo (lclRow);
      if (rowInfo.localRow == Tpetra::Details::OrdinalTraits<size_t>::invalid ()) {
        numEnt = 0; // no valid entries in this row on the calling process
        val = NULL;
        ind = NULL;
        // First argument (lclRow) invalid, so make 1 the error code.
        return static_cast<LO> (1);
      }
      else {
        numEnt = static_cast<LO> (rowInfo.numEntries);
        auto lclColInds = staticGraph_->getLocalKokkosRowView (rowInfo);
        ind = lclColInds.data (); // FIXME (mfh 18 Jul 2016) UVM
        const LO err = this->getViewRawConst (val, numEnt, rowInfo);
        return err;
      }
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getLocalRowViewRaw (const LocalOrdinal lclRow,
                      LocalOrdinal& numEnt,
                      const LocalOrdinal*& lclColInds,
                      const Scalar*& vals) const
  {
    const impl_scalar_type* vals_ist = NULL;
    const LocalOrdinal errCode =
      this->getLocalRowView (lclRow, numEnt, vals_ist, lclColInds);
    vals = reinterpret_cast<const Scalar*> (vals_ist);
    return errCode;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getGlobalRowView (GlobalOrdinal globalRow,
                    Teuchos::ArrayView<const GlobalOrdinal>& indices,
                    Teuchos::ArrayView<const Scalar>& values) const
  {
    using Teuchos::ArrayView;
    using Teuchos::av_reinterpret_cast;
    typedef GlobalOrdinal GO;
    const char tfecfFuncName[] = "getGlobalRowView: ";
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      isLocallyIndexed (), std::runtime_error,
      "The matrix is locally indexed, so we cannot return a view of the row "
      "with global column indices.  Use getGlobalRowCopy() instead.");
    indices = Teuchos::null;
    values  = Teuchos::null;
    const RowInfo rowinfo =
      staticGraph_->getRowInfoFromGlobalRowIndex (globalRow);
    if (rowinfo.localRow != Teuchos::OrdinalTraits<size_t>::invalid () &&
        rowinfo.numEntries > 0) {
      ArrayView<const GO> indTmp = staticGraph_->getGlobalView (rowinfo);
      ArrayView<const Scalar> valTmp =
        av_reinterpret_cast<const Scalar> (this->getView (rowinfo));
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (static_cast<size_t> (indTmp.size ()) < rowinfo.numEntries ||
         static_cast<size_t> (valTmp.size ()) < rowinfo.numEntries,
         std::logic_error, std::endl << "rowinfo.numEntries not accurate.  "
         << std::endl << "indTmp.size() = " << indTmp.size ()
         << ", valTmp.size() = " << valTmp.size ()
         << ", rowinfo.numEntries = " << rowinfo.numEntries << ".");
#endif // HAVE_TPETRA_DEBUG
      indices = indTmp (0, rowinfo.numEntries);
      values = valTmp (0, rowinfo.numEntries);
    }
 
#ifdef HAVE_TPETRA_DEBUG
    const char suffix[] = ".  This should never happen.  Please report this "
      "bug to the Tpetra developers.";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (static_cast<size_t> (indices.size ()) !=
       static_cast<size_t> (values.size ()), std::logic_error,
       "At the end of this method, for global row " << globalRow << ", "
       "indices.size() = " << indices.size () << " != values.size () = "
       << values.size () << suffix);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (static_cast<size_t> (indices.size ()) !=
       static_cast<size_t> (rowinfo.numEntries), std::logic_error,
       "At the end of this method, for global row " << globalRow << ", "
       "indices.size() = " << indices.size () << " != rowinfo.numEntries = "
       << rowinfo.numEntries << suffix);
    const size_t expectedNumEntries = getNumEntriesInGlobalRow (globalRow);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (rowinfo.numEntries != expectedNumEntries, std::logic_error, "At the end "
       "of this method, for global row " << globalRow << ", rowinfo.numEntries "
       "= " << rowinfo.numEntries << " != getNumEntriesInGlobalRow(globalRow) ="
       " " << expectedNumEntries << suffix);
#endif // HAVE_TPETRA_DEBUG
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  scale (const Scalar& alpha)
  {
    typedef LocalOrdinal LO;
    typedef typename Teuchos::Array<Scalar>::size_type size_type;
    const char tfecfFuncName[] = "scale: ";
    const impl_scalar_type theAlpha = static_cast<impl_scalar_type> (alpha);
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      ! isFillActive (), std::runtime_error,
      "Fill must be active before you may call this method.  "
      "Please call resumeFill() to make fill active.");
 
    const size_t nlrs = staticGraph_->getNodeNumRows ();
    const size_t numEntries = staticGraph_->getNodeNumEntries ();
    if (! staticGraph_->indicesAreAllocated () ||
        nlrs == 0 || numEntries == 0) {
      // do nothing
    }
    else {
      if (staticGraph_->getProfileType () == StaticProfile) {
        const LO lclNumRows = lclMatrix_.numRows ();
        for (LO lclRow = 0; lclRow < lclNumRows; ++lclRow) {
          auto row_i = lclMatrix_.row (lclRow);
          for (LO k = 0; k < row_i.length; ++k) {
            // FIXME (mfh 02 Jan 2015) This assumes CUDA UVM.
            row_i.value (k) *= theAlpha;
          }
        }
      }
      else if (staticGraph_->getProfileType () == DynamicProfile) {
        for (size_t row = 0; row < nlrs; ++row) {
          const size_type numEnt = getNumEntriesInLocalRow (row);
          Teuchos::ArrayView<impl_scalar_type> rowVals = values2D_[row] ();
          for (size_type k = 0; k < numEnt; ++k) {
            rowVals[k] *= theAlpha;
          }
        }
      }
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  setAllToScalar (const Scalar& alpha)
  {
    const char tfecfFuncName[] = "setAllToScalar: ";
    const impl_scalar_type theAlpha = static_cast<impl_scalar_type> (alpha);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      ! isFillActive (), std::runtime_error,
      "Fill must be active before you may call this method.  "
      "Please call resumeFill() to make fill active.");
 
    // replace all values in the matrix
    // it is easiest to replace all allocated values, instead of replacing only the ones with valid entries
    // however, if there are no valid entries, we can short-circuit
    // furthermore, if the values aren't allocated, we can short-circuit (no entry have been inserted so far)
    const size_t nlrs = staticGraph_->getNodeNumRows();
    const size_t numEntries = staticGraph_->getNodeNumEntries();
    if (! staticGraph_->indicesAreAllocated () || numEntries == 0) {
      // do nothing
    }
    else {
      const ProfileType profType = staticGraph_->getProfileType ();
      if (profType == StaticProfile) {
        // FIXME (mfh 24 Dec 2014) Once CrsMatrix implements DualView
        // semantics, this would be the place to mark memory as
        // modified.
        Kokkos::deep_copy (k_values1D_, theAlpha);
      }
      else if (profType == DynamicProfile) {
        for (size_t row = 0; row < nlrs; ++row) {
          std::fill (values2D_[row].begin (), values2D_[row].end (), theAlpha);
        }
      }
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  setAllValues (const typename local_matrix_type::row_map_type& rowPointers,
                const typename local_graph_type::entries_type::non_const_type& columnIndices,
                const typename local_matrix_type::values_type& values)
  {
    const char tfecfFuncName[] = "setAllValues: ";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (columnIndices.size () != values.size (), std::invalid_argument,
       "columnIndices.size() = " << columnIndices.size () << " != values.size()"
       " = " << values.size () << ".");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (myGraph_.is_null (), std::runtime_error, "myGraph_ must not be null.");
 
    try {
      myGraph_->setAllIndices (rowPointers, columnIndices);
    }
    catch (std::exception &e) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "myGraph_->setAllIndices() threw an "
         "exception: " << e.what ());
    }
    // Make sure that myGraph_ now has a local graph.  It may not be
    // fillComplete yet, so it's important to check.  We don't care
    // whether setAllIndices() did a shallow copy or a deep copy, so a
    // good way to check is to compare dimensions.
    auto lclGraph = myGraph_->getLocalGraph ();
    const size_t numEnt = lclGraph.entries.extent (0);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (lclGraph.row_map.extent (0) != rowPointers.extent (0) ||
       numEnt != static_cast<size_t> (columnIndices.extent (0)),
       std::logic_error, "myGraph_->setAllIndices() did not correctly create "
       "local graph.  Please report this bug to the Tpetra developers.");
 
    const size_t numCols = myGraph_->getColMap ()->getNodeNumElements ();
    this->lclMatrix_ = local_matrix_type ("Tpetra::CrsMatrix::lclMatrix_",
                                          numCols, values, lclGraph);
    // FIXME (22 Jun 2016) I would very much like to get rid of
    // k_values1D_ at some point.  I find it confusing to have all
    // these extra references lying around.
    this->k_values1D_ = this->lclMatrix_.values;
 
    // Storage MUST be packed, since the interface doesn't give any
    // way to indicate any extra space at the end of each row.
    this->storageStatus_ = ::Tpetra::Details::STORAGE_1D_PACKED;
 
    checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  setAllValues (const Teuchos::ArrayRCP<size_t>& ptr,
                const Teuchos::ArrayRCP<LocalOrdinal>& ind,
                const Teuchos::ArrayRCP<Scalar>& val)
  {
    using Kokkos::Compat::getKokkosViewDeepCopy;
    using Teuchos::ArrayRCP;
    using Teuchos::av_reinterpret_cast;
    typedef device_type DT;
    typedef impl_scalar_type IST;
    typedef typename local_matrix_type::row_map_type row_map_type;
    //typedef typename row_map_type::non_const_value_type row_offset_type;
    const char tfecfFuncName[] = "setAllValues(ArrayRCP<size_t>, ArrayRCP<LO>, ArrayRCP<Scalar>): ";
 
    // The row offset type may depend on the execution space.  It may
    // not necessarily be size_t.  If it's not, we need to make a deep
    // copy.  We need to make a deep copy anyway so that Kokkos can
    // own the memory.  Regardless, ptrIn gets the copy.
    typename row_map_type::non_const_type ptrNative ("ptr", ptr.size ());
    Kokkos::View<const size_t*,
      typename row_map_type::array_layout,
      Kokkos::HostSpace,
      Kokkos::MemoryUnmanaged> ptrSizeT (ptr.getRawPtr (), ptr.size ());
    ::Tpetra::Details::copyOffsets (ptrNative, ptrSizeT);
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (ptrNative.extent (0) != ptrSizeT.extent (0),
       std::logic_error, "ptrNative.extent(0) = " <<
       ptrNative.extent (0) << " != ptrSizeT.extent(0) = "
       << ptrSizeT.extent (0) << ".  Please report this bug to the "
       "Tpetra developers.");
 
    auto indIn = getKokkosViewDeepCopy<DT> (ind ());
    auto valIn = getKokkosViewDeepCopy<DT> (av_reinterpret_cast<IST> (val ()));
    this->setAllValues (ptrNative, indIn, valIn);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getLocalDiagOffsets (Teuchos::ArrayRCP<size_t>& offsets) const
  {
    const char tfecfFuncName[] = "getLocalDiagOffsets: ";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (staticGraph_.is_null (), std::runtime_error, "The matrix has no graph.");
 
    // mfh 11 May 2016: We plan to deprecate the ArrayRCP version of
    // this method in CrsGraph too, so don't call it (otherwise build
    // warnings will show up and annoy users).  Instead, copy results
    // in and out, if the memory space requires it.
 
    const size_t lclNumRows = staticGraph_->getNodeNumRows ();
    if (static_cast<size_t> (offsets.size ()) < lclNumRows) {
      offsets.resize (lclNumRows);
    }
 
    // The input ArrayRCP must always be a host pointer.  Thus, if
    // device_type::memory_space is Kokkos::HostSpace, it's OK for us
    // to write to that allocation directly as a Kokkos::View.
    typedef typename device_type::memory_space memory_space;
    if (std::is_same<memory_space, Kokkos::HostSpace>::value) {
      // It is always syntactically correct to assign a raw host
      // pointer to a device View, so this code will compile correctly
      // even if this branch never runs.
      typedef Kokkos::View<size_t*, device_type,
                           Kokkos::MemoryUnmanaged> output_type;
      output_type offsetsOut (offsets.getRawPtr (), lclNumRows);
      staticGraph_->getLocalDiagOffsets (offsetsOut);
    }
    else {
      Kokkos::View<size_t*, device_type> offsetsTmp ("diagOffsets", lclNumRows);
      staticGraph_->getLocalDiagOffsets (offsetsTmp);
      typedef Kokkos::View<size_t*, Kokkos::HostSpace,
                           Kokkos::MemoryUnmanaged> output_type;
      output_type offsetsOut (offsets.getRawPtr (), lclNumRows);
      Kokkos::deep_copy (offsetsOut, offsetsTmp);
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getLocalDiagCopy (Vector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& diag) const
  {
    using Teuchos::ArrayRCP;
    using Teuchos::ArrayView;
    using Teuchos::av_reinterpret_cast;
    const char tfecfFuncName[] = "getLocalDiagCopy (1-arg): ";
    typedef local_ordinal_type LO;
 
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      staticGraph_.is_null (), std::runtime_error,
      "This method requires that the matrix have a graph.");
    auto rowMapPtr = this->getRowMap ();
    if (rowMapPtr.is_null () || rowMapPtr->getComm ().is_null ()) {
      // Processes on which the row Map or its communicator is null
      // don't participate.  Users shouldn't even call this method on
      // those processes.
      return;
    }
    auto colMapPtr = this->getColMap ();
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! this->hasColMap () || colMapPtr.is_null (), std::runtime_error,
       "This method requires that the matrix have a column Map.");
    const map_type& rowMap = * rowMapPtr;
    const map_type& colMap = * colMapPtr;
    const LO myNumRows = static_cast<LO> (this->getNodeNumRows ());
 
#ifdef HAVE_TPETRA_DEBUG
    // isCompatible() requires an all-reduce, and thus this check
    // should only be done in debug mode.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      ! diag.getMap ()->isCompatible (rowMap), std::runtime_error,
      "The input Vector's Map must be compatible with the CrsMatrix's row "
      "Map.  You may check this by using Map's isCompatible method: "
      "diag.getMap ()->isCompatible (A.getRowMap ());");
#endif // HAVE_TPETRA_DEBUG
 
    if (this->isFillComplete ()) {
      diag.template modify<device_type> ();
      const auto D_lcl = diag.template getLocalView<device_type> ();
      // 1-D subview of the first (and only) column of D_lcl.
      const auto D_lcl_1d =
        Kokkos::subview (D_lcl, Kokkos::make_pair (LO (0), myNumRows), 0);
 
      const auto lclRowMap = rowMap.getLocalMap ();
      const auto lclColMap = colMap.getLocalMap ();
      const auto lclMatrix = this->lclMatrix_;
      using ::Tpetra::Details::getDiagCopyWithoutOffsets;
      (void) getDiagCopyWithoutOffsets (D_lcl_1d, lclRowMap,
                                        lclColMap, lclMatrix);
    }
    else {
      using ::Tpetra::Details::getLocalDiagCopyWithoutOffsetsNotFillComplete;
      (void) getLocalDiagCopyWithoutOffsetsNotFillComplete (diag, *this);
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getLocalDiagCopy (Vector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& diag,
                    const Kokkos::View<const size_t*, device_type,
                      Kokkos::MemoryUnmanaged>& offsets) const
  {
    typedef LocalOrdinal LO;
 
#ifdef HAVE_TPETRA_DEBUG
    const char tfecfFuncName[] = "getLocalDiagCopy: ";
    const map_type& rowMap = * (this->getRowMap ());
    // isCompatible() requires an all-reduce, and thus this check
    // should only be done in debug mode.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      ! diag.getMap ()->isCompatible (rowMap), std::runtime_error,
      "The input Vector's Map must be compatible with (in the sense of Map::"
      "isCompatible) the CrsMatrix's row Map.");
#endif // HAVE_TPETRA_DEBUG
 
    // For now, we fill the Vector on the host and sync to device.
    // Later, we may write a parallel kernel that works entirely on
    // device.
    //
    // NOTE (mfh 21 Jan 2016): The host kernel here assumes UVM.  Once
    // we write a device kernel, it will not need to assume UVM.
 
    diag.template modify<device_type> ();
    auto D_lcl = diag.template getLocalView<device_type> ();
    const LO myNumRows = static_cast<LO> (this->getNodeNumRows ());
    // Get 1-D subview of the first (and only) column of D_lcl.
    auto D_lcl_1d =
      Kokkos::subview (D_lcl, Kokkos::make_pair (LO (0), myNumRows), 0);
 
    KokkosSparse::getDiagCopy (D_lcl_1d, offsets, this->lclMatrix_);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getLocalDiagCopy (Vector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& diag,
                    const Teuchos::ArrayView<const size_t>& offsets) const
  {
    typedef LocalOrdinal LO;
    typedef impl_scalar_type IST;
    typedef Vector<Scalar, LocalOrdinal, GlobalOrdinal, Node> vec_type;
    typedef typename vec_type::dual_view_type dual_view_type;
    typedef typename dual_view_type::host_mirror_space::execution_space host_execution_space;
 
#ifdef HAVE_TPETRA_DEBUG
    const char tfecfFuncName[] = "getLocalDiagCopy: ";
    const map_type& rowMap = * (this->getRowMap ());
    // isCompatible() requires an all-reduce, and thus this check
    // should only be done in debug mode.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      ! diag.getMap ()->isCompatible (rowMap), std::runtime_error,
      "The input Vector's Map must be compatible with (in the sense of Map::"
      "isCompatible) the CrsMatrix's row Map.");
#endif // HAVE_TPETRA_DEBUG
 
    // See #1510.  In case diag has already been marked modified on
    // device, we need to clear that flag, since the code below works
    // on host.
    auto diag_dv = diag.getDualView ();
    diag_dv.modified_device () = 0;
 
    // For now, we fill the Vector on the host and sync to device.
    // Later, we may write a parallel kernel that works entirely on
    // device.
    diag.template modify<host_execution_space> ();
    auto lclVecHost = diag.template getLocalView<host_execution_space> ();
    // 1-D subview of the first (and only) column of lclVecHost.
    auto lclVecHost1d = Kokkos::subview (lclVecHost, Kokkos::ALL (), 0);
 
    Kokkos::View<const size_t*, Kokkos::HostSpace,
                 Kokkos::MemoryTraits<Kokkos::Unmanaged> >
      h_offsets (offsets.getRawPtr (), offsets.size ());
    // Find the diagonal entries and put them in lclVecHost1d.
    const LO myNumRows = static_cast<LO> (this->getNodeNumRows ());
    typedef Kokkos::RangePolicy<host_execution_space, LO> policy_type;
    const size_t INV = Tpetra::Details::OrdinalTraits<size_t>::invalid ();
 
    Kokkos::parallel_for (policy_type (0, myNumRows), [&] (const LO& lclRow) {
      lclVecHost1d(lclRow) = STS::zero (); // default value if no diag entry
      if (h_offsets[lclRow] != INV) {
        auto curRow = lclMatrix_.rowConst (lclRow);
        lclVecHost1d(lclRow) = static_cast<IST> (curRow.value(h_offsets[lclRow]));
      }
    });
    diag.template sync<execution_space> (); // sync changes back to device
  }
 
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  leftScale (const Vector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& x)
  {
    using ::Tpetra::Details::ProfilingRegion;
    using Teuchos::ArrayRCP;
    using Teuchos::ArrayView;
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
    using Teuchos::rcpFromRef;
    using LO = local_ordinal_type;
    typedef Vector<Scalar, LocalOrdinal, GlobalOrdinal, Node> vec_type;
    const char tfecfFuncName[] = "leftScale: ";
 
    ProfilingRegion region ("Tpetra::CrsMatrix::leftScale");
 
    RCP<const vec_type> xp;
    if (this->getRangeMap ()->isSameAs (* (x.getMap ()))) {
      // Take from Epetra: If we have a non-trivial exporter, we must
      // import elements that are permuted or are on other processors.
      auto exporter = this->getCrsGraphRef ().getExporter ();
      if (exporter.get () != nullptr) {
        RCP<vec_type> tempVec (new vec_type (this->getRowMap ()));
        tempVec->doImport (x, *exporter, REPLACE); // reverse mode
        xp = tempVec;
      }
      else {
        xp = rcpFromRef (x);
      }
    }
    else if (this->getRowMap ()->isSameAs (* (x.getMap ()))) {
      xp = rcpFromRef (x);
    }
    else {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::invalid_argument, "x's Map must be the same as "
         "either the row Map or the range Map of the CrsMatrix.");
    }
 
    // Check whether A has a valid local matrix.  It might not if it
    // was not created with a local matrix, and if fillComplete has
    // never been called on it before.  A never-initialized (and thus
    // invalid) local matrix has zero rows, because it was default
    // constructed.
    const LO lclNumRows =
      static_cast<LO> (this->getRowMap ()->getNodeNumElements ());
    const bool validLocalMatrix = this->lclMatrix_.numRows () == lclNumRows;
 
    if (validLocalMatrix) {
      using dev_memory_space = typename device_type::memory_space;
      if (xp->template need_sync<dev_memory_space> ()) {
        using Teuchos::rcp_const_cast;
        rcp_const_cast<vec_type> (xp)->template sync<dev_memory_space> ();
      }
      auto x_lcl = xp->template getLocalView<dev_memory_space> ();
      auto x_lcl_1d = Kokkos::subview (x_lcl, Kokkos::ALL (), 0);
      ::Tpetra::Details::leftScaleLocalCrsMatrix (this->lclMatrix_, x_lcl_1d, false, false);
    }
    else {
      execution_space::fence (); // for UVM's sake
 
      ArrayRCP<const Scalar> vectorVals = xp->getData (0);
      ArrayView<impl_scalar_type> rowValues = Teuchos::null;
      for (LocalOrdinal i = 0; i < lclNumRows; ++i) {
        const RowInfo rowinfo = this->staticGraph_->getRowInfo (i);
        rowValues = this->getViewNonConst (rowinfo);
        const impl_scalar_type scaleValue = static_cast<impl_scalar_type> (vectorVals[i]);
        for (size_t j = 0; j < rowinfo.numEntries; ++j) {
          rowValues[j] *= scaleValue;
        }
      }
      execution_space::fence (); // for UVM's sake
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  rightScale (const Vector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& x)
  {
    using ::Tpetra::Details::ProfilingRegion;
    using Teuchos::ArrayRCP;
    using Teuchos::ArrayView;
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
    using Teuchos::rcpFromRef;
    using LO = local_ordinal_type;
    typedef Vector<Scalar, LocalOrdinal, GlobalOrdinal, Node> vec_type;
    const char tfecfFuncName[] = "rightScale: ";
 
    ProfilingRegion region ("Tpetra::CrsMatrix::rightScale");
 
    RCP<const vec_type> xp;
    if (this->getDomainMap ()->isSameAs (* (x.getMap ()))) {
      // Take from Epetra: If we have a non-trivial exporter, we must
      // import elements that are permuted or are on other processors.
      auto importer = this->getCrsGraphRef ().getImporter ();
      if (importer.get () != nullptr) {
        RCP<vec_type> tempVec (new vec_type (this->getColMap ()));
        tempVec->doImport (x, *importer, REPLACE);
        xp = tempVec;
      }
      else {
        xp = rcpFromRef (x);
      }
    }
    else if (this->getColMap ()->isSameAs (* (x.getMap ()))) {
      xp = rcpFromRef (x);
    } else {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (true, std::runtime_error, "x's Map must be the same as "
         "either the domain Map or the column Map of the CrsMatrix.");
    }
 
    // Check whether A has a valid local matrix.  It might not if it
    // was not created with a local matrix, and if fillComplete has
    // never been called on it before.  A never-initialized (and thus
    // invalid) local matrix has zero rows, because it was default
    // constructed.
    const LO lclNumRows =
      static_cast<LO> (this->getRowMap ()->getNodeNumElements ());
    const bool validLocalMatrix = this->lclMatrix_.numRows () == lclNumRows;
 
    if (validLocalMatrix) {
      using dev_memory_space = typename device_type::memory_space;
      if (xp->template need_sync<dev_memory_space> ()) {
        using Teuchos::rcp_const_cast;
        rcp_const_cast<vec_type> (xp)->template sync<dev_memory_space> ();
      }
      auto x_lcl = xp->template getLocalView<dev_memory_space> ();
      auto x_lcl_1d = Kokkos::subview (x_lcl, Kokkos::ALL (), 0);
      ::Tpetra::Details::rightScaleLocalCrsMatrix (this->lclMatrix_, x_lcl_1d, false, false);
    }
    else {
      execution_space::fence (); // for UVM's sake
 
      ArrayRCP<const Scalar> vectorVals = xp->getData (0);
      ArrayView<impl_scalar_type> rowValues = null;
      for (LO i = 0; i < lclNumRows; ++i) {
        const RowInfo rowinfo = this->staticGraph_->getRowInfo (i);
        rowValues = this->getViewNonConst (rowinfo);
        ArrayView<const LO> colInds;
        this->getCrsGraphRef ().getLocalRowView (i, colInds);
        for (size_t j = 0; j < rowinfo.numEntries; ++j) {
          rowValues[j] *= static_cast<impl_scalar_type> (vectorVals[colInds[j]]);
        }
      }
      execution_space::fence (); // for UVM's sake
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  typename CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::mag_type
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getFrobeniusNorm () const
  {
    using Teuchos::ArrayView;
    using Teuchos::outArg;
    using Teuchos::REDUCE_SUM;
    using Teuchos::reduceAll;
    typedef typename Teuchos::ArrayRCP<const impl_scalar_type>::size_type size_type;
 
    // FIXME (mfh 05 Aug 2014) Write a thread-parallel kernel for the
    // local part of this computation.  It could make sense to put
    // this operation in the Kokkos::CrsMatrix.
 
    // check the cache first
    mag_type frobNorm = frobNorm_;
    if (frobNorm == -STM::one ()) {
      mag_type mySum = STM::zero ();
      if (getNodeNumEntries() > 0) {
        if (isStorageOptimized ()) {
          // "Optimized" storage is packed storage.  That means we can
          // iterate in one pass through the 1-D values array.
          const size_type numEntries =
            static_cast<size_type> (getNodeNumEntries ());
          for (size_type k = 0; k < numEntries; ++k) {
            // FIXME (mfh 05 Aug 2014) This assumes UVM.
            const impl_scalar_type val = k_values1D_(k);
            // Note (etp 06 Jan 2015) We need abs() here for composite types
            // (in general, if mag_type is on the left-hand-side, we need
            // abs() on the right-hand-side)
            const mag_type val_abs = STS::abs (val);
            mySum += val_abs * val_abs;
          }
        }
        else {
          const LocalOrdinal numRows =
            static_cast<LocalOrdinal> (this->getNodeNumRows ());
          for (LocalOrdinal r = 0; r < numRows; ++r) {
            const RowInfo rowInfo = myGraph_->getRowInfo (r);
            const size_type numEntries =
              static_cast<size_type> (rowInfo.numEntries);
            ArrayView<const impl_scalar_type> A_r =
              this->getView (rowInfo).view (0, numEntries);
            for (size_type k = 0; k < numEntries; ++k) {
              const impl_scalar_type val = A_r[k];
              const mag_type val_abs = STS::abs (val);
              mySum += val_abs * val_abs;
            }
          }
        }
      }
      mag_type totalSum = STM::zero ();
      reduceAll<int, mag_type> (* (getComm ()), REDUCE_SUM,
                                mySum, outArg (totalSum));
      frobNorm = STM::sqrt (totalSum);
    }
    if (isFillComplete ()) {
      // Only cache the result if the matrix is fill complete.
      // Otherwise, the values might still change.  resumeFill clears
      // the cache.
      frobNorm_ = frobNorm;
    }
    return frobNorm;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  replaceColMap (const Teuchos::RCP<const map_type>& newColMap)
  {
    const char tfecfFuncName[] = "replaceColMap: ";
    // FIXME (mfh 06 Aug 2014) What if the graph is locally indexed?
    // Then replacing the column Map might mean that we need to
    // reindex the column indices.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      myGraph_.is_null (), std::runtime_error,
      "This method does not work if the matrix has a const graph.  The whole "
      "idea of a const graph is that you are not allowed to change it, but "
      "this method necessarily must modify the graph, since the graph owns "
      "the matrix's column Map.");
    myGraph_->replaceColMap (newColMap);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  reindexColumns (crs_graph_type* const graph,
                  const Teuchos::RCP<const map_type>& newColMap,
                  const Teuchos::RCP<const import_type>& newImport,
                  const bool sortEachRow)
  {
    const char tfecfFuncName[] = "reindexColumns: ";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      graph == NULL && myGraph_.is_null (), std::invalid_argument,
      "The input graph is NULL, but the matrix does not own its graph.");
 
    crs_graph_type& theGraph = (graph == NULL) ? *myGraph_ : *graph;
    const bool sortGraph = false; // we'll sort graph & matrix together below
    theGraph.reindexColumns (newColMap, newImport, sortGraph);
    if (sortEachRow && theGraph.isLocallyIndexed () && ! theGraph.isSorted ()) {
      const LocalOrdinal lclNumRows =
        static_cast<LocalOrdinal> (theGraph.getNodeNumRows ());
      for (LocalOrdinal row = 0; row < lclNumRows; ++row) {
        const RowInfo rowInfo = theGraph.getRowInfo (row);
        auto lclColInds = theGraph.getLocalKokkosRowViewNonConst (rowInfo);
        auto vals = this->getRowViewNonConst (rowInfo);
        // FIXME (mfh 09 May 2017) This assumes CUDA UVM, at least for
        // lclColInds, if not also for values.
        sort2 (lclColInds.data (),
               lclColInds.data () + rowInfo.numEntries,
               vals.data ());
      }
      theGraph.indicesAreSorted_ = true;
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  replaceDomainMapAndImporter (const Teuchos::RCP<const map_type>& newDomainMap,
                               Teuchos::RCP<const import_type>& newImporter)
  {
    const char tfecfFuncName[] = "replaceDomainMapAndImporter: ";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      myGraph_.is_null (), std::runtime_error,
      "This method does not work if the matrix has a const graph.  The whole "
      "idea of a const graph is that you are not allowed to change it, but this"
      " method necessarily must modify the graph, since the graph owns the "
      "matrix's domain Map and Import objects.");
    myGraph_->replaceDomainMapAndImporter (newDomainMap, newImporter);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  insertNonownedGlobalValues (const GlobalOrdinal globalRow,
                              const Teuchos::ArrayView<const GlobalOrdinal>& indices,
                              const Teuchos::ArrayView<const Scalar>& values)
  {
    using Teuchos::Array;
    typedef GlobalOrdinal GO;
    typedef typename Array<GO>::size_type size_type;
 
    const size_type numToInsert = indices.size ();
    // Add the new data to the list of nonlocals.
    // This creates the arrays if they don't exist yet.
    std::pair<Array<GO>, Array<Scalar> >& curRow = nonlocals_[globalRow];
    Array<GO>& curRowInds = curRow.first;
    Array<Scalar>& curRowVals = curRow.second;
    const size_type newCapacity = curRowInds.size () + numToInsert;
    curRowInds.reserve (newCapacity);
    curRowVals.reserve (newCapacity);
    for (size_type k = 0; k < numToInsert; ++k) {
      curRowInds.push_back (indices[k]);
      curRowVals.push_back (values[k]);
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  globalAssemble ()
  {
    using ::Tpetra::Details::ProfilingRegion;
    using Teuchos::Comm;
    using Teuchos::outArg;
    using Teuchos::RCP;
    using Teuchos::rcp;
    using Teuchos::REDUCE_MAX;
    using Teuchos::REDUCE_MIN;
    using Teuchos::reduceAll;
    typedef CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> crs_matrix_type;
    //typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef typename Teuchos::Array<GO>::size_type size_type;
    const char tfecfFuncName[] = "globalAssemble: "; // for exception macro
    ProfilingRegion regionGlobalAssemble ("Tpetra::CrsMatrix::globalAssemble");
 
    RCP<const Comm<int> > comm = getComm ();
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! isFillActive (), std::runtime_error, "Fill must be active before "
       "you may call this method.");
 
    const size_t myNumNonlocalRows = nonlocals_.size ();
 
    // If no processes have nonlocal rows, then we don't have to do
    // anything.  Checking this is probably cheaper than constructing
    // the Map of nonlocal rows (see below) and noticing that it has
    // zero global entries.
    {
      const int iHaveNonlocalRows = (myNumNonlocalRows == 0) ? 0 : 1;
      int someoneHasNonlocalRows = 0;
      reduceAll<int, int> (*comm, REDUCE_MAX, iHaveNonlocalRows,
                           outArg (someoneHasNonlocalRows));
      if (someoneHasNonlocalRows == 0) {
        return; // no process has nonlocal rows, so nothing to do
      }
    }
 
    // 1. Create a list of the "nonlocal" rows on each process.  this
    //    requires iterating over nonlocals_, so while we do this,
    //    deduplicate the entries and get a count for each nonlocal
    //    row on this process.
    // 2. Construct a new row Map corresponding to those rows.  This
    //    Map is likely overlapping.  We know that the Map is not
    //    empty on all processes, because the above all-reduce and
    //    return exclude that case.
 
    RCP<const map_type> nonlocalRowMap;
    // Keep this for CrsGraph's constructor, so we can use StaticProfile.
    Teuchos::ArrayRCP<size_t> numEntPerNonlocalRow (myNumNonlocalRows);
    {
      Teuchos::Array<GO> myNonlocalGblRows (myNumNonlocalRows);
      size_type curPos = 0;
      for (auto mapIter = nonlocals_.begin (); mapIter != nonlocals_.end ();
           ++mapIter, ++curPos) {
        myNonlocalGblRows[curPos] = mapIter->first;
        // Get the values and column indices by reference, since we
        // intend to change them in place (that's what "erase" does).
        Teuchos::Array<GO>& gblCols = (mapIter->second).first;
        Teuchos::Array<Scalar>& vals = (mapIter->second).second;
 
        // Sort both arrays jointly, using the column indices as keys,
        // then merge them jointly.  "Merge" here adds values
        // corresponding to the same column indices.  The first 2 args
        // of merge2 are output arguments that work just like the
        // return value of std::unique.
        sort2 (gblCols.begin (), gblCols.end (), vals.begin ());
        typename Teuchos::Array<GO>::iterator gblCols_newEnd;
        typename Teuchos::Array<Scalar>::iterator vals_newEnd;
        merge2 (gblCols_newEnd, vals_newEnd,
                gblCols.begin (), gblCols.end (),
                vals.begin (), vals.end ());
        gblCols.erase (gblCols_newEnd, gblCols.end ());
        vals.erase (vals_newEnd, vals.end ());
        numEntPerNonlocalRow[curPos] = gblCols.size ();
      }
 
      // Currently, Map requires that its indexBase be the global min
      // of all its global indices.  Map won't compute this for us, so
      // we must do it.  If our process has no nonlocal rows, set the
      // "min" to the max possible GO value.  This ensures that if
      // some process has at least one nonlocal row, then it will pick
      // that up as the min.  We know that at least one process has a
      // nonlocal row, since the all-reduce and return at the top of
      // this method excluded that case.
      GO myMinNonlocalGblRow = std::numeric_limits<GO>::max ();
      {
        auto iter = std::min_element (myNonlocalGblRows.begin (),
                                      myNonlocalGblRows.end ());
        if (iter != myNonlocalGblRows.end ()) {
          myMinNonlocalGblRow = *iter;
        }
      }
      GO gblMinNonlocalGblRow = 0;
      reduceAll<int, GO> (*comm, REDUCE_MIN, myMinNonlocalGblRow,
                          outArg (gblMinNonlocalGblRow));
      const GO indexBase = gblMinNonlocalGblRow;
      const global_size_t INV = Teuchos::OrdinalTraits<global_size_t>::invalid ();
      nonlocalRowMap = rcp (new map_type (INV, myNonlocalGblRows (), indexBase, comm));
    }
 
    // 3. Use the values and column indices for each nonlocal row, as
    //    stored in nonlocals_, to construct a CrsMatrix corresponding
    //    to nonlocal rows.  We may use StaticProfile, since we have
    //    exact counts of the number of entries in each nonlocal row.
 
    RCP<crs_matrix_type> nonlocalMatrix =
      rcp (new crs_matrix_type (nonlocalRowMap, numEntPerNonlocalRow,
                                StaticProfile));
    {
      size_type curPos = 0;
      for (auto mapIter = nonlocals_.begin (); mapIter != nonlocals_.end ();
           ++mapIter, ++curPos) {
        const GO gblRow = mapIter->first;
        // Get values & column indices by ref, just to avoid copy.
        Teuchos::Array<GO>& gblCols = (mapIter->second).first;
        Teuchos::Array<Scalar>& vals = (mapIter->second).second;
        //const LO numEnt = static_cast<LO> (numEntPerNonlocalRow[curPos]);
        nonlocalMatrix->insertGlobalValues (gblRow, gblCols (), vals ());
      }
    }
    // There's no need to fill-complete the nonlocals matrix.
    // We just use it as a temporary container for the Export.
 
    // 4. If the original row Map is one to one, then we can Export
    //    directly from nonlocalMatrix into this.  Otherwise, we have
    //    to create a temporary matrix with a one-to-one row Map,
    //    Export into that, then Import from the temporary matrix into
    //    *this.
 
    auto origRowMap = this->getRowMap ();
    const bool origRowMapIsOneToOne = origRowMap->isOneToOne ();
 
    int isLocallyComplete = 1; // true by default
 
    if (origRowMapIsOneToOne) {
      export_type exportToOrig (nonlocalRowMap, origRowMap);
      if (! exportToOrig.isLocallyComplete ()) {
        isLocallyComplete = 0;
      }
      this->doExport (*nonlocalMatrix, exportToOrig, Tpetra::ADD);
      // We're done at this point!
    }
    else {
      // If you ask a Map whether it is one to one, it does some
      // communication and stashes intermediate results for later use
      // by createOneToOne.  Thus, calling createOneToOne doesn't cost
      // much more then the original cost of calling isOneToOne.
      auto oneToOneRowMap = Tpetra::createOneToOne (origRowMap);
      export_type exportToOneToOne (nonlocalRowMap, oneToOneRowMap);
      if (! exportToOneToOne.isLocallyComplete ()) {
        isLocallyComplete = 0;
      }
 
      // Create a temporary matrix with the one-to-one row Map.
      //
      // TODO (mfh 09 Sep 2016, 12 Sep 2016) Estimate # entries in
      // each row, to avoid reallocation during the Export operation.
      crs_matrix_type oneToOneMatrix (oneToOneRowMap, 0);
      // Export from matrix of nonlocals into the temp one-to-one matrix.
      oneToOneMatrix.doExport (*nonlocalMatrix, exportToOneToOne, Tpetra::ADD);
 
      // We don't need the matrix of nonlocals anymore, so get rid of
      // it, to keep the memory high-water mark down.
      nonlocalMatrix = Teuchos::null;
 
      // Import from the one-to-one matrix to the original matrix.
      import_type importToOrig (oneToOneRowMap, origRowMap);
      this->doImport (oneToOneMatrix, importToOrig, Tpetra::ADD);
    }
 
    // It's safe now to clear out nonlocals_, since we've already
    // committed side effects to *this.  The standard idiom for
    // clearing a Container like std::map, is to swap it with an empty
    // Container and let the swapped Container fall out of scope.
    decltype (nonlocals_) newNonlocals;
    std::swap (nonlocals_, newNonlocals);
 
    // FIXME (mfh 12 Sep 2016) I don't like this all-reduce, and I
    // don't like throwing an exception here.  A local return value
    // would likely be more useful to users.  However, if users find
    // themselves exercising nonlocal inserts often, then they are
    // probably novice users who need the help.  See Gibhub Issues
    // #603 and #601 (esp. the latter) for discussion.
 
    int isGloballyComplete = 0; // output argument of reduceAll
    reduceAll<int, int> (*comm, REDUCE_MIN, isLocallyComplete,
                         outArg (isGloballyComplete));
    TEUCHOS_TEST_FOR_EXCEPTION
      (isGloballyComplete != 1, std::runtime_error, "On at least one process, "
       "you called insertGlobalValues with a global row index which is not in "
       "the matrix's row Map on any process in its communicator.");
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  resumeFill (const Teuchos::RCP<Teuchos::ParameterList>& params)
  {
    if (! isStaticGraph ()) { // Don't resume fill of a nonowned graph.
      myGraph_->resumeFill (params);
    }
    clearGlobalConstants ();
    fillComplete_ = false;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  computeGlobalConstants ()
  {
    // This method doesn't do anything.  The analogous method in
    // CrsGraph does actually compute something.
    //
    // Oddly enough, clearGlobalConstants() clears frobNorm_ (by
    // setting it to -1), but computeGlobalConstants() does _not_
    // compute the Frobenius norm; this is done on demand in
    // getFrobeniusNorm(), and the result is cached there.
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  haveGlobalConstants() const {
    return getCrsGraphRef ().haveGlobalConstants ();
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  clearGlobalConstants () {
    // We use -1 to indicate that the Frobenius norm needs to be
    // recomputed, since the values might change between now and the
    // next fillComplete call.
    //
    // Oddly enough, clearGlobalConstants() clears frobNorm_, but
    // computeGlobalConstants() does _not_ compute the Frobenius norm;
    // this is done on demand in getFrobeniusNorm(), and the result is
    // cached there.
    frobNorm_ = -STM::one ();
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  fillComplete (const Teuchos::RCP<Teuchos::ParameterList>& params)
  {
    const char tfecfFuncName[] = "fillComplete(params): ";
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (this->getCrsGraph ().is_null (), std::logic_error,
      "getCrsGraph() returns null.  This should not happen at this point.  "
      "Please report this bug to the Tpetra developers.");
 
    const crs_graph_type& graph = this->getCrsGraphRef ();
    if (this->isStaticGraph () && graph.isFillComplete ()) {
      // If this matrix's graph is fill complete and the user did not
      // supply a domain or range Map, use the graph's domain and
      // range Maps.
      this->fillComplete (graph.getDomainMap (), graph.getRangeMap (), params);
    }
    else { // assume that user's row Map is the domain and range Map
      Teuchos::RCP<const map_type> rangeMap = graph.getRowMap ();
      Teuchos::RCP<const map_type> domainMap = rangeMap;
      this->fillComplete (domainMap, rangeMap, params);
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  fillComplete (const Teuchos::RCP<const map_type>& domainMap,
                const Teuchos::RCP<const map_type>& rangeMap,
                const Teuchos::RCP<Teuchos::ParameterList>& params)
  {
    using ::Tpetra::Details::ProfilingRegion;
    using Teuchos::ArrayRCP;
    using Teuchos::RCP;
    using Teuchos::rcp;
    const char tfecfFuncName[] = "fillComplete: ";
    ProfilingRegion regionFillComplete ("Tpetra::CrsMatrix::fillComplete");
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! this->isFillActive () || this->isFillComplete (), std::runtime_error,
       "Matrix fill state must be active (isFillActive() "
       "must be true) before you may call fillComplete().");
    const int numProcs = this->getComm ()->getSize ();
 
    //
    // Read parameters from the input ParameterList.
    //
 
    // If true, the caller promises that no process did nonlocal
    // changes since the last call to fillComplete.
    bool assertNoNonlocalInserts = false;
    // If true, makeColMap sorts remote GIDs (within each remote
    // process' group).
    bool sortGhosts = true;
 
    if (! params.is_null ()) {
      assertNoNonlocalInserts = params->get ("No Nonlocal Changes",
                                             assertNoNonlocalInserts);
      if (params->isParameter ("sort column map ghost gids")) {
        sortGhosts = params->get ("sort column map ghost gids", sortGhosts);
      }
      else if (params->isParameter ("Sort column Map ghost GIDs")) {
        sortGhosts = params->get ("Sort column Map ghost GIDs", sortGhosts);
      }
    }
    // We also don't need to do global assembly if there is only one
    // process in the communicator.
    const bool needGlobalAssemble = ! assertNoNonlocalInserts && numProcs > 1;
    // This parameter only matters if this matrix owns its graph.
    if (! this->myGraph_.is_null ()) {
      this->myGraph_->sortGhostsAssociatedWithEachProcessor_ = sortGhosts;
    }
 
    if (! this->getCrsGraphRef ().indicesAreAllocated ()) {
      if (this->hasColMap ()) {
        // We have a column Map, so use local indices.
        this->allocateValues (LocalIndices, GraphNotYetAllocated);
      } else {
        // We don't have a column Map, so use global indices.
        this->allocateValues (GlobalIndices, GraphNotYetAllocated);
      }
    }
    // Global assemble, if we need to.  This call only costs a single
    // all-reduce if we didn't need global assembly after all.
    if (needGlobalAssemble) {
      this->globalAssemble ();
    }
    else {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (numProcs == 1 && nonlocals_.size() > 0,
         std::runtime_error, "Cannot have nonlocal entries on a serial run.  "
         "An invalid entry (i.e., with row index not in the row Map) must have "
         "been submitted to the CrsMatrix.");
    }
 
    if (this->isStaticGraph ()) {
      // FIXME (mfh 14 Nov 2016) In order to fix #843, I enable the
      // checks below only in debug mode.  It would be nicer to do a
      // local check, then propagate the error state in a deferred
      // way, whenever communication happens.  That would reduce the
      // cost of checking, to the point where it may make sense to
      // enable it even in release mode.
#ifdef HAVE_TPETRA_DEBUG
      // FIXME (mfh 18 Jun 2014) This check for correctness of the
      // input Maps incurs a penalty of two all-reduces for the
      // otherwise optimal const graph case.
      //
      // We could turn these (max) 2 all-reduces into (max) 1, by
      // fusing them.  We could do this by adding a "locallySameAs"
      // method to Map, which would return one of four states:
      //
      //   a. Certainly globally the same
      //   b. Certainly globally not the same
      //   c. Locally the same
      //   d. Locally not the same
      //
      // The first two states don't require further communication.
      // The latter two states require an all-reduce to communicate
      // globally, but we only need one all-reduce, since we only need
      // to check whether at least one of the Maps is wrong.
      const bool domainMapsMatch =
        this->staticGraph_->getDomainMap ()->isSameAs (*domainMap);
      const bool rangeMapsMatch =
        this->staticGraph_->getRangeMap ()->isSameAs (*rangeMap);
 
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (! domainMapsMatch, std::runtime_error,
         "The CrsMatrix's domain Map does not match the graph's domain Map.  "
         "The graph cannot be changed because it was given to the CrsMatrix "
         "constructor as const.  You can fix this by passing in the graph's "
         "domain Map and range Map to the matrix's fillComplete call.");
 
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (! rangeMapsMatch, std::runtime_error,
         "The CrsMatrix's range Map does not match the graph's range Map.  "
         "The graph cannot be changed because it was given to the CrsMatrix "
         "constructor as const.  You can fix this by passing in the graph's "
         "domain Map and range Map to the matrix's fillComplete call.");
#endif // HAVE_TPETRA_DEBUG
 
      // The matrix does _not_ own the graph, and the graph's
      // structure is already fixed, so just fill the local matrix.
      this->fillLocalMatrix (params);
    }
    else {
      // Set the graph's domain and range Maps.  This will clear the
      // Import if the domain Map has changed (is a different
      // pointer), and the Export if the range Map has changed (is a
      // different pointer).
      this->myGraph_->setDomainRangeMaps (domainMap, rangeMap);
 
      // Make the graph's column Map, if necessary.
      Teuchos::Array<int> remotePIDs (0);
      const bool mustBuildColMap = ! this->hasColMap ();
      if (mustBuildColMap) {
        this->myGraph_->makeColMap (remotePIDs);
      }
 
      // Make indices local, if necessary.  The method won't do
      // anything if the graph is already locally indexed.
      const std::pair<size_t, std::string> makeIndicesLocalResult =
        this->myGraph_->makeIndicesLocal ();
      // TODO (mfh 20 Jul 2017) Instead of throwing here, pass along
      // the error state to makeImportExport or
      // computeGlobalConstants, which may do all-reduces and thus may
      // have the opportunity to communicate that error state.
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (makeIndicesLocalResult.first != 0, std::runtime_error,
         makeIndicesLocalResult.second);
 
      const bool sorted = this->myGraph_->isSorted ();
      const bool merged = this->myGraph_->isMerged ();
      this->sortAndMergeIndicesAndValues (sorted, merged);
 
      // Make Import and Export objects, if they haven't been made
      // already.  If we made a column Map above, reuse information
      // from that process to avoid communiation in the Import setup.
      this->myGraph_->makeImportExport (remotePIDs, mustBuildColMap);
 
      // The matrix _does_ own the graph, so fill the local graph at
      // the same time as the local matrix.
      this->fillLocalGraphAndMatrix (params);
 
      const bool callGraphComputeGlobalConstants = params.get () == nullptr ||
        params->get ("compute global constants", true);
      const bool computeLocalTriangularConstants = params.get () == nullptr ||
        params->get ("compute local triangular constants", true);
      if (callGraphComputeGlobalConstants) {
        this->myGraph_->computeGlobalConstants (computeLocalTriangularConstants);
      }
      else {
        this->myGraph_->computeLocalConstants (computeLocalTriangularConstants);
      }
      this->myGraph_->fillComplete_ = true;
      this->myGraph_->checkInternalState ();
    }
 
    const bool callComputeGlobalConstants = params.get () == nullptr ||
      params->get ("compute global constants", true);
    if (callComputeGlobalConstants) {
      this->computeGlobalConstants ();
    }
 
    // FIXME (mfh 28 Aug 2014) "Preserve Local Graph" bool parameter no longer used.
 
    this->fillComplete_ = true; // Now we're fill complete!
    this->checkInternalState ();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  expertStaticFillComplete (const Teuchos::RCP<const map_type> & domainMap,
                            const Teuchos::RCP<const map_type> & rangeMap,
                            const Teuchos::RCP<const import_type>& importer,
                            const Teuchos::RCP<const export_type>& exporter,
                            const Teuchos::RCP<Teuchos::ParameterList> &params)
  {
#ifdef HAVE_TPETRA_MMM_TIMINGS
    std::string label;
    if(!params.is_null())
      label = params->get("Timer Label",label);
    std::string prefix = std::string("Tpetra ")+ label + std::string(": ");
    using Teuchos::TimeMonitor;
    Teuchos::RCP<Teuchos::TimeMonitor> MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("ESFC-M-Graph"))));
#endif
 
 
    const char tfecfFuncName[] = "expertStaticFillComplete: ";
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC( ! isFillActive() || isFillComplete(),
      std::runtime_error, "Matrix fill state must be active (isFillActive() "
      "must be true) before calling fillComplete().");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      myGraph_.is_null (), std::logic_error, "myGraph_ is null.  This is not allowed.");
 
 
    // We will presume globalAssemble is not needed, so we do the ESFC on the graph
    myGraph_->expertStaticFillComplete (domainMap, rangeMap, importer, exporter,params);
 
    const bool callComputeGlobalConstants = params.get () == nullptr ||
      params->get ("compute global constants", true);
    if (callComputeGlobalConstants) {
#ifdef HAVE_TPETRA_MMM_TIMINGS
      MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("ESFC-M-cGC"))));
#endif
      this->computeGlobalConstants ();
    }
 
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("ESFC-M-fLGAM"))));
#endif
 
    // Fill the local graph and matrix
    fillLocalGraphAndMatrix (params);
 
    // FIXME (mfh 28 Aug 2014) "Preserve Local Graph" bool parameter no longer used.
 
    // Now we're fill complete!
    fillComplete_ = true;
 
    // Sanity checks at the end.
#ifdef HAVE_TPETRA_DEBUG
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(isFillActive(), std::logic_error,
      ": We're at the end of fillComplete(), but isFillActive() is true.  "
      "Please report this bug to the Tpetra developers.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(! isFillComplete(), std::logic_error,
      ": We're at the end of fillComplete(), but isFillActive() is true.  "
      "Please report this bug to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
 
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("ESFC-M-cIS"))));
#endif
 
    checkInternalState();
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  mergeRowIndicesAndValues (crs_graph_type& graph,
                            const RowInfo& rowInfo)
  {
#ifdef HAVE_TPETRA_DEBUG
    const char tfecfFuncName[] = "mergeRowIndicesAndValues: ";
#endif // HAVE_TPETRA_DEBUG
 
    auto rowValues = this->getRowViewNonConst (rowInfo);
    typedef typename std::decay<decltype (rowValues[0]) >::type value_type;
    value_type* rowValueIter = rowValues.data ();
    auto inds_view = graph.getLocalKokkosRowViewNonConst (rowInfo);
 
    // beg,end define a half-exclusive interval over which to iterate.
    LocalOrdinal* beg = inds_view.data ();
    LocalOrdinal* end = inds_view.data () + rowInfo.numEntries;
 
#ifdef HAVE_TPETRA_DEBUG
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (rowInfo.allocSize != static_cast<size_t> (inds_view.extent (0)) ||
       rowInfo.allocSize != static_cast<size_t> (rowValues.extent (0)),
       std::runtime_error, "rowInfo.allocSize = " << rowInfo.allocSize
       << " != inds_view.extent(0) = " << inds_view.extent (0)
       << " || rowInfo.allocSize = " << rowInfo.allocSize
       << " != rowValues.extent(0) = " << rowValues.extent (0) << ".");
#endif // HAVE_TPETRA_DEBUG
 
    LocalOrdinal* newend = beg;
    if (beg != end) {
      LocalOrdinal* cur = beg + 1;
      value_type* vcur = rowValueIter + 1;
      value_type* vend = rowValueIter;
      cur = beg+1;
      while (cur != end) {
        if (*cur != *newend) {
          // new entry; save it
          ++newend;
          ++vend;
          (*newend) = (*cur);
          (*vend) = (*vcur);
        }
        else {
          // old entry; merge it
          //(*vend) = f (*vend, *vcur);
          (*vend) += *vcur;
        }
        ++cur;
        ++vcur;
      }
      ++newend; // one past the last entry, per typical [beg,end) semantics
    }
    const size_t mergedEntries = newend - beg;
    graph.k_numRowEntries_(rowInfo.localRow) = mergedEntries;
    const size_t numDups = rowInfo.numEntries - mergedEntries;
    return numDups;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  sortAndMergeIndicesAndValues (const bool sorted, const bool merged)
  {
    using ::Tpetra::Details::ProfilingRegion;
    typedef LocalOrdinal LO;
    typedef typename Kokkos::View<LO*, device_type>::HostMirror::execution_space
      host_execution_space;
    typedef Kokkos::RangePolicy<host_execution_space, LO> range_type;
    //typedef Kokkos::RangePolicy<Kokkos::Serial, LO> range_type;
    const char tfecfFuncName[] = "sortAndMergeIndicesAndValues: ";
    ProfilingRegion regionSAM ("Tpetra::CrsMatrix::sortAndMergeIndicesAndValues");
 
    if (! sorted || ! merged) {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (this->isStaticGraph (), std::runtime_error, "Cannot sort or merge with "
         "\"static\" (const) graph, since the matrix does not own the graph.");
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (this->myGraph_.is_null (), std::logic_error, "myGraph_ is null, but "
         "this matrix claims ! isStaticGraph().  "
         "Please report this bug to the Tpetra developers.");
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (this->isStorageOptimized (), std::logic_error, "It is invalid to call "
         "this method if the graph's storage has already been optimized.  "
         "Please report this bug to the Tpetra developers.");
 
      crs_graph_type& graph = * (this->myGraph_);
      const LO lclNumRows = static_cast<LO> (this->getNodeNumRows ());
      size_t totalNumDups = 0;
      // FIXME (mfh 10 May 2017) This may assume CUDA UVM.
      Kokkos::parallel_reduce (range_type (0, lclNumRows),
        [this, &graph, sorted, merged] (const LO& lclRow, size_t& numDups) {
          const RowInfo rowInfo = graph.getRowInfo (lclRow);
          if (! sorted) {
            auto lclColInds = graph.getLocalKokkosRowViewNonConst (rowInfo);
            auto vals = this->getRowViewNonConst (rowInfo);
            // FIXME (mfh 09 May 2017) This assumes CUDA UVM, at least
            // for lclColInds, if not also for values.
            sort2 (lclColInds.data (),
                   lclColInds.data () + rowInfo.numEntries,
                   vals.data ());
          }
          if (! merged) {
            numDups += this->mergeRowIndicesAndValues (graph, rowInfo);
          }
        }, totalNumDups);
      if (! sorted) {
        graph.indicesAreSorted_ = true; // we just sorted every row
      }
      if (! merged) {
        graph.noRedundancies_ = true; // we just merged every row
      }
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  applyNonTranspose (const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node> & X_in,
                     MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node> & Y_in,
                     Scalar alpha,
                     Scalar beta) const
  {
    using Tpetra::Details::ProfilingRegion;
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
    using Teuchos::rcp_const_cast;
    using Teuchos::rcpFromRef;
    const Scalar ZERO = Teuchos::ScalarTraits<Scalar>::zero ();
    const Scalar ONE = Teuchos::ScalarTraits<Scalar>::one ();
 
    // mfh 05 Jun 2014: Special case for alpha == 0.  I added this to
    // fix an Ifpack2 test (RILUKSingleProcessUnitTests), which was
    // failing only for the Kokkos refactor version of Tpetra.  It's a
    // good idea regardless to have the bypass.
    if (alpha == ZERO) {
      if (beta == ZERO) {
        Y_in.putScalar (ZERO);
      } else if (beta != ONE) {
        Y_in.scale (beta);
      }
      return;
    }
 
    // It's possible that X is a view of Y or vice versa.  We don't
    // allow this (apply() requires that X and Y not alias one
    // another), but it's helpful to detect and work around this case.
    // We don't try to to detect the more subtle cases (e.g., one is a
    // subview of the other, but their initial pointers differ).  We
    // only need to do this if this matrix's Import is trivial;
    // otherwise, we don't actually apply the operator from X into Y.
 
    RCP<const import_type> importer = this->getGraph ()->getImporter ();
    RCP<const export_type> exporter = this->getGraph ()->getExporter ();
 
    // If beta == 0, then the output MV will be overwritten; none of
    // its entries should be read.  (Sparse BLAS semantics say that we
    // must ignore any Inf or NaN entries in Y_in, if beta is zero.)
    // This matters if we need to do an Export operation; see below.
    const bool Y_is_overwritten = (beta == ZERO);
 
    // We treat the case of a replicated MV output specially.
    const bool Y_is_replicated =
      (! Y_in.isDistributed () && this->getComm ()->getSize () != 1);
 
    // This is part of the special case for replicated MV output.
    // We'll let each process do its thing, but do an all-reduce at
    // the end to sum up the results.  Setting beta=0 on all processes
    // but Proc 0 makes the math work out for the all-reduce.  (This
    // assumes that the replicated data is correctly replicated, so
    // that the data are the same on all processes.)
    if (Y_is_replicated && this->getComm ()->getRank () > 0) {
      beta = ZERO;
    }
 
    // Temporary MV for Import operation.  After the block of code
    // below, this will be an (Imported if necessary) column Map MV
    // ready to give to localMultiply().
    RCP<const MV> X_colMap;
    if (importer.is_null ()) {
      if (! X_in.isConstantStride ()) {
        // Not all sparse mat-vec kernels can handle an input MV with
        // nonconstant stride correctly, so we have to copy it in that
        // case into a constant stride MV.  To make a constant stride
        // copy of X_in, we force creation of the column (== domain)
        // Map MV (if it hasn't already been created, else fetch the
        // cached copy).  This avoids creating a new MV each time.
        RCP<MV> X_colMapNonConst = getColumnMapMultiVector (X_in, true);
        Tpetra::deep_copy (*X_colMapNonConst, X_in);
        X_colMap = rcp_const_cast<const MV> (X_colMapNonConst);
      }
      else {
        // The domain and column Maps are the same, so do the local
        // multiply using the domain Map input MV X_in.
        X_colMap = rcpFromRef (X_in);
      }
    }
    else { // need to Import source (multi)vector
      ProfilingRegion regionImport ("Tpetra::CrsMatrix::apply: Import");
 
      // We're doing an Import anyway, which will copy the relevant
      // elements of the domain Map MV X_in into a separate column Map
      // MV.  Thus, we don't have to worry whether X_in is constant
      // stride.
      RCP<MV> X_colMapNonConst = getColumnMapMultiVector (X_in);
 
      // Import from the domain Map MV to the column Map MV.
      X_colMapNonConst->doImport (X_in, *importer, INSERT);
      X_colMap = rcp_const_cast<const MV> (X_colMapNonConst);
    }
 
    // Temporary MV for doExport (if needed), or for copying a
    // nonconstant stride output MV into a constant stride MV.  This
    // is null if we don't need the temporary MV, that is, if the
    // Export is trivial (null).
    RCP<MV> Y_rowMap = getRowMapMultiVector (Y_in);
 
    // If we have a nontrivial Export object, we must perform an
    // Export.  In that case, the local multiply result will go into
    // the row Map multivector.  We don't have to make a
    // constant-stride version of Y_in in this case, because we had to
    // make a constant stride Y_rowMap MV and do an Export anyway.
    if (! exporter.is_null ()) {
      this->localApply (*X_colMap, *Y_rowMap, Teuchos::NO_TRANS, alpha, ZERO);
      {
        ProfilingRegion regionExport ("Tpetra::CrsMatrix::apply: Export");
 
        // If we're overwriting the output MV Y_in completely (beta ==
        // 0), then make sure that it is filled with zeros before we
        // do the Export.  Otherwise, the ADD combine mode will use
        // data in Y_in, which is supposed to be zero.
        if (Y_is_overwritten) {
          Y_in.putScalar (ZERO);
        }
        else {
          // Scale output MV by beta, so that doExport sums in the
          // mat-vec contribution: Y_in = beta*Y_in + alpha*A*X_in.
          Y_in.scale (beta);
        }
        // Do the Export operation.
        Y_in.doExport (*Y_rowMap, *exporter, ADD);
      }
    }
    else { // Don't do an Export: row Map and range Map are the same.
      //
      // If Y_in does not have constant stride, or if the column Map
      // MV aliases Y_in, then we can't let the kernel write directly
      // to Y_in.  Instead, we have to use the cached row (== range)
      // Map MV as temporary storage.
      //
      // FIXME (mfh 05 Jun 2014) This test for aliasing only tests if
      // the user passed in the same MultiVector for both X and Y.  It
      // won't detect whether one MultiVector views the other.  We
      // should also check the MultiVectors' raw data pointers.
      if (! Y_in.isConstantStride () || X_colMap.getRawPtr () == &Y_in) {
        // Force creating the MV if it hasn't been created already.
        // This will reuse a previously created cached MV.
        Y_rowMap = getRowMapMultiVector (Y_in, true);
 
        // If beta == 0, we don't need to copy Y_in into Y_rowMap,
        // since we're overwriting it anyway.
        if (beta != ZERO) {
          Tpetra::deep_copy (*Y_rowMap, Y_in);
        }
        this->localApply (*X_colMap, *Y_rowMap, Teuchos::NO_TRANS, alpha, beta);
        Tpetra::deep_copy (Y_in, *Y_rowMap);
      }
      else {
        this->localApply (*X_colMap, Y_in, Teuchos::NO_TRANS, alpha, beta);
      }
    }
 
    // If the range Map is a locally replicated Map, sum up
    // contributions from each process.  We set beta = 0 on all
    // processes but Proc 0 initially, so this will handle the scaling
    // factor beta correctly.
    if (Y_is_replicated) {
      ProfilingRegion regionReduce ("Tpetra::CrsMatrix::apply: Reduce Y");
      Y_in.reduce ();
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  applyTranspose (const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& X_in,
                  MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& Y_in,
                  const Teuchos::ETransp mode,
                  Scalar alpha,
                  Scalar beta) const
  {
    using Tpetra::Details::ProfilingRegion;
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
    using Teuchos::rcp_const_cast;
    using Teuchos::rcpFromRef;
    const Scalar ZERO = Teuchos::ScalarTraits<Scalar>::zero ();
 
    // Take shortcuts for alpha == 0.
    if (alpha == ZERO) {
      // Follow the Sparse BLAS convention by ignoring both the matrix
      // and X_in, in this case.
      if (beta == ZERO) {
        // Follow the Sparse BLAS convention by overwriting any Inf or
        // NaN values in Y_in, in this case.
        Y_in.putScalar (ZERO);
      }
      else {
        Y_in.scale (beta);
      }
      return;
    }
 
    const size_t numVectors = X_in.getNumVectors ();
 
    // We don't allow X_in and Y_in to alias one another.  It's hard
    // to check this, because advanced users could create views from
    // raw pointers.  However, if X_in and Y_in reference the same
    // object, we will do the user a favor by copying X into new
    // storage (with a warning).  We only need to do this if we have
    // trivial importers; otherwise, we don't actually apply the
    // operator from X into Y.
    RCP<const import_type> importer = this->getGraph ()->getImporter ();
    RCP<const export_type> exporter = this->getGraph ()->getExporter ();
    // access X indirectly, in case we need to create temporary storage
    RCP<const MV> X;
 
    // some parameters for below
    const bool Y_is_replicated = ! Y_in.isDistributed ();
    const bool Y_is_overwritten = (beta == ZERO);
    if (Y_is_replicated && this->getComm ()->getRank () > 0) {
      beta = ZERO;
    }
 
    // The kernels do not allow input or output with nonconstant stride.
    if (! X_in.isConstantStride () && importer.is_null ()) {
      X = rcp (new MV (X_in, Teuchos::Copy)); // Constant-stride copy of X_in
    } else {
      X = rcpFromRef (X_in); // Reference to X_in
    }
 
    // Set up temporary multivectors for Import and/or Export.
    if (importer != Teuchos::null) {
      if (importMV_ != Teuchos::null && importMV_->getNumVectors() != numVectors) {
        importMV_ = null;
      }
      if (importMV_ == null) {
        importMV_ = rcp (new MV (this->getColMap (), numVectors));
      }
    }
    if (exporter != Teuchos::null) {
      if (exportMV_ != Teuchos::null && exportMV_->getNumVectors() != numVectors) {
        exportMV_ = null;
      }
      if (exportMV_ == null) {
        exportMV_ = rcp (new MV (this->getRowMap (), numVectors));
      }
    }
 
    // If we have a non-trivial exporter, we must import elements that
    // are permuted or are on other processors.
    if (! exporter.is_null ()) {
      ProfilingRegion regionImport ("Tpetra::CrsMatrix::apply (transpose): Import");
      exportMV_->doImport (X_in, *exporter, INSERT);
      X = exportMV_; // multiply out of exportMV_
    }
 
    // If we have a non-trivial importer, we must export elements that
    // are permuted or belong to other processors.  We will compute
    // solution into the to-be-exported MV; get a view.
    if (importer != Teuchos::null) {
      ProfilingRegion regionExport ("Tpetra::CrsMatrix::apply (transpose): Export");
 
      // FIXME (mfh 18 Apr 2015) Temporary fix suggested by Clark
      // Dohrmann on Fri 17 Apr 2015.  At some point, we need to go
      // back and figure out why this helps.  importMV_ SHOULD be
      // completely overwritten in the localMultiply() call below,
      // because beta == ZERO there.
      importMV_->putScalar (ZERO);
      // Do the local computation.
      this->localApply (*X, *importMV_, mode, alpha, ZERO);
      if (Y_is_overwritten) {
        Y_in.putScalar (ZERO);
      } else {
        Y_in.scale (beta);
      }
      Y_in.doExport (*importMV_, *importer, ADD);
    }
    // otherwise, multiply into Y
    else {
      // can't multiply in-situ; can't multiply into non-strided multivector
      //
      // FIXME (mfh 05 Jun 2014) This test for aliasing only tests if
      // the user passed in the same MultiVector for both X and Y.  It
      // won't detect whether one MultiVector views the other.  We
      // should also check the MultiVectors' raw data pointers.
      if (! Y_in.isConstantStride () || X.getRawPtr () == &Y_in) {
        // Make a deep copy of Y_in, into which to write the multiply result.
        MV Y (Y_in, Teuchos::Copy);
        this->localApply (*X, Y, mode, alpha, beta);
        Tpetra::deep_copy (Y_in, Y);
      } else {
        this->localApply (*X, Y_in, mode, alpha, beta);
      }
    }
 
    // If the range Map is a locally replicated map, sum the
    // contributions from each process.  (That's why we set beta=0
    // above for all processes but Proc 0.)
    if (Y_is_replicated) {
      ProfilingRegion regionReduce ("Tpetra::CrsMatrix::apply (transpose): Reduce Y");
      Y_in.reduce ();
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  localApply (const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& X,
              MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>&Y,
              const Teuchos::ETransp mode,
              const Scalar& alpha,
              const Scalar& beta) const
  {
    using Tpetra::Details::ProfilingRegion;
    ProfilingRegion regionLocalApply ("Tpetra::CrsMatrix::localApply");
    this->template localMultiply<Scalar, Scalar> (X, Y, mode, alpha, beta);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  apply (const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node> &X,
         MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node> &Y,
         Teuchos::ETransp mode,
         Scalar alpha,
         Scalar beta) const
  {
    using Tpetra::Details::ProfilingRegion;
    const char fnName[] = "Tpetra::CrsMatrix::apply";
 
    TEUCHOS_TEST_FOR_EXCEPTION
      (! isFillComplete (), std::runtime_error,
       fnName << ": Cannot call apply() until fillComplete() "
       "has been called.");
 
    if (mode == Teuchos::NO_TRANS) {
      ProfilingRegion regionNonTranspose (fnName);
      this->applyNonTranspose (X, Y, alpha, beta);
    }
    else {
      ProfilingRegion regionTranspose ("Tpetra::CrsMatrix::apply (transpose)");
 
      //Thyra was implicitly assuming that Y gets set to zero / or is overwritten
      //when bets==0. This was not the case with transpose in a multithreaded
      //environment where a multiplication with subsequent atomic_adds is used
      //since 0 is effectively not special cased. Doing the explicit set to zero here
      //This catches cases where Y is nan or inf.
      const Scalar ZERO = Teuchos::ScalarTraits<Scalar>::zero ();
      if (beta == ZERO) {
        Y.putScalar (ZERO);
      }
      this->applyTranspose (X, Y, mode, alpha, beta);
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  gaussSeidel (const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& B,
               MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& X,
               const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& D,
               const Scalar& dampingFactor,
               const ESweepDirection direction,
               const int numSweeps) const
  {
    reorderedGaussSeidel (B, X, D, Teuchos::null, dampingFactor, direction, numSweeps);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  reorderedGaussSeidel (const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& B,
                        MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& X,
                        const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& D,
                        const Teuchos::ArrayView<LocalOrdinal>& rowIndices,
                        const Scalar& dampingFactor,
                        const ESweepDirection direction,
                        const int numSweeps) const
  {
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
    using Teuchos::rcp_const_cast;
    using Teuchos::rcpFromRef;
    typedef Scalar ST;
 
    TEUCHOS_TEST_FOR_EXCEPTION(
      isFillComplete() == false, std::runtime_error,
      "Tpetra::CrsMatrix::gaussSeidel: cannot call this method until "
      "fillComplete() has been called.");
    TEUCHOS_TEST_FOR_EXCEPTION(
      numSweeps < 0,
      std::invalid_argument,
      "Tpetra::CrsMatrix::gaussSeidel: The number of sweeps must be , "
      "nonnegative but you provided numSweeps = " << numSweeps << " < 0.");
 
    // Translate from global to local sweep direction.
    // While doing this, validate the input.
    ESweepDirection localDirection;
    if (direction == Forward) {
      localDirection = Forward;
    }
    else if (direction == Backward) {
      localDirection = Backward;
    }
    else if (direction == Symmetric) {
      // We'll control local sweep direction manually.
      localDirection = Forward;
    }
    else {
      TEUCHOS_TEST_FOR_EXCEPTION(true, std::invalid_argument,
        "Tpetra::CrsMatrix::gaussSeidel: The 'direction' enum does not have "
        "any of its valid values: Forward, Backward, or Symmetric.");
    }
 
    if (numSweeps == 0) {
      return; // Nothing to do.
    }
 
    // We don't need the Export object because this method assumes
    // that the row, domain, and range Maps are the same.  We do need
    // the Import object, if there is one, though.
    RCP<const import_type> importer = this->getGraph()->getImporter();
    RCP<const export_type> exporter = this->getGraph()->getExporter();
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! exporter.is_null (), std::runtime_error,
      "Tpetra's gaussSeidel implementation requires that the row, domain, "
      "and range Maps be the same.  This cannot be the case, because the "
      "matrix has a nontrivial Export object.");
 
    RCP<const map_type> domainMap = this->getDomainMap ();
    RCP<const map_type> rangeMap = this->getRangeMap ();
    RCP<const map_type> rowMap = this->getGraph ()->getRowMap ();
    RCP<const map_type> colMap = this->getGraph ()->getColMap ();
 
#ifdef HAVE_TEUCHOS_DEBUG
    {
      // The relation 'isSameAs' is transitive.  It's also a
      // collective, so we don't have to do a "shared" test for
      // exception (i.e., a global reduction on the test value).
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! X.getMap ()->isSameAs (*domainMap),
        std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidel requires that the input "
        "multivector X be in the domain Map of the matrix.");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! B.getMap ()->isSameAs (*rangeMap),
        std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidel requires that the input "
        "B be in the range Map of the matrix.");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! D.getMap ()->isSameAs (*rowMap),
        std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidel requires that the input "
        "D be in the row Map of the matrix.");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! rowMap->isSameAs (*rangeMap),
        std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidel requires that the row Map and the "
        "range Map be the same (in the sense of Tpetra::Map::isSameAs).");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! domainMap->isSameAs (*rangeMap),
        std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidel requires that the domain Map and "
        "the range Map of the matrix be the same.");
    }
#else
    // Forestall any compiler warnings for unused variables.
    (void) rangeMap;
    (void) rowMap;
#endif // HAVE_TEUCHOS_DEBUG
 
    // If B is not constant stride, copy it into a constant stride
    // multivector.  We'l handle the right-hand side B first and deal
    // with X right before the sweeps, to improve locality of the
    // first sweep.  (If the problem is small enough, then that will
    // hopefully keep more of the entries of X in cache.  This
    // optimizes for the typical case of a small number of sweeps.)
    RCP<const MV> B_in;
    if (B.isConstantStride()) {
      B_in = rcpFromRef (B);
    }
    else {
      // The range Map and row Map are the same in this case, so we
      // can use the (possibly cached) row Map multivector to store a
      // constant stride copy of B.  We don't have to copy back, since
      // Gauss-Seidel won't modify B.
      RCP<MV> B_in_nonconst = getRowMapMultiVector (B, true);
      deep_copy (*B_in_nonconst, B); // Copy from B into B_in(_nonconst).
      B_in = rcp_const_cast<const MV> (B_in_nonconst);
 
      TPETRA_EFFICIENCY_WARNING(
        ! B.isConstantStride (),
        std::runtime_error,
        "gaussSeidel: The current implementation of the Gauss-Seidel kernel "
        "requires that X and B both have constant stride.  Since B does not "
        "have constant stride, we had to make a copy.  This is a limitation of "
        "the current implementation and not your fault, but we still report it "
        "as an efficiency warning for your information.");
    }
 
    // If X is not constant stride, copy it into a constant stride
    // multivector.  Also, make the column Map multivector X_colMap,
    // and its domain Map view X_domainMap.  (X actually must be a
    // domain Map view of a column Map multivector; exploit this, if X
    // has constant stride.)
 
    RCP<MV> X_domainMap;
    RCP<MV> X_colMap;
    bool copiedInput = false;
 
    if (importer.is_null ()) { // Domain and column Maps are the same.
      if (X.isConstantStride ()) {
        X_domainMap = rcpFromRef (X);
        X_colMap = X_domainMap;
        copiedInput = false;
      }
      else {
        // Get a temporary column Map multivector, make a domain Map
        // view of it, and copy X into the domain Map view.  We have
        // to copy here because we won't be doing Import operations.
        X_colMap = getColumnMapMultiVector (X, true);
        X_domainMap = X_colMap; // Domain and column Maps are the same.
        deep_copy (*X_domainMap, X); // Copy X into the domain Map view.
        copiedInput = true;
        TPETRA_EFFICIENCY_WARNING(
          ! X.isConstantStride (), std::runtime_error,
          "Tpetra::CrsMatrix::gaussSeidel: The current implementation of the "
          "Gauss-Seidel kernel requires that X and B both have constant "
          "stride.  Since X does not have constant stride, we had to make a "
          "copy.  This is a limitation of the current implementation and not "
          "your fault, but we still report it as an efficiency warning for "
          "your information.");
      }
    }
    else { // We will be doing Import operations in the sweeps.
      if (X.isConstantStride ()) {
        X_domainMap = rcpFromRef (X);
        // This kernel assumes that X is a domain Map view of a column
        // Map multivector.  We will only check if this is valid if
        // the CMake configure Teuchos_ENABLE_DEBUG is ON.
        X_colMap = X_domainMap->offsetViewNonConst (colMap, 0);
 
        // FIXME (mfh 19 Mar 2013) Do we need to fill the remote
        // entries of X_colMap with zeros?  Do we need to fill all of
        // X_domainMap initially with zeros?  Ifpack
        // (Ifpack_PointRelaxation.cpp, line 906) creates an entirely
        // new MultiVector each time.
 
        // Do the first Import for the first sweep.  This simplifies
        // the logic in the sweeps.
        X_colMap->doImport (X, *importer, INSERT);
        copiedInput = false;
      }
      else {
        // Get a temporary column Map multivector X_colMap, and make a
        // domain Map view X_domainMap of it.  Instead of copying, we
        // do an Import from X into X_domainMap.  This saves us a
        // copy, since the Import has to copy the data anyway.
        X_colMap = getColumnMapMultiVector (X, true);
        X_domainMap = X_colMap->offsetViewNonConst (domainMap, 0);
        X_colMap->doImport (X, *importer, INSERT);
        copiedInput = true;
        TPETRA_EFFICIENCY_WARNING(
          ! X.isConstantStride (), std::runtime_error,
          "Tpetra::CrsMatrix::gaussSeidel: The current implementation of the "
          "Gauss-Seidel kernel requires that X and B both have constant stride.  "
          "Since X does not have constant stride, we had to make a copy.  "
          "This is a limitation of the current implementation and not your fault, "
          "but we still report it as an efficiency warning for your information.");
      }
    }
 
    for (int sweep = 0; sweep < numSweeps; ++sweep) {
      if (! importer.is_null () && sweep > 0) {
        // We already did the first Import for the zeroth sweep.
        X_colMap->doImport (*X_domainMap, *importer, INSERT);
      }
 
      // Do local Gauss-Seidel.
      if (direction != Symmetric) {
        if (rowIndices.is_null ()) {
          this->template localGaussSeidel<ST, ST> (*B_in, *X_colMap, D,
                                                   dampingFactor,
                                                   localDirection);
        }
        else {
          this->template reorderedLocalGaussSeidel<ST, ST> (*B_in, *X_colMap,
                                                            D, rowIndices,
                                                            dampingFactor,
                                                            localDirection);
        }
      }
      else { // direction == Symmetric
        const bool doImportBetweenDirections = false;
        if (rowIndices.is_null ()) {
          this->template localGaussSeidel<ST, ST> (*B_in, *X_colMap, D,
                                                   dampingFactor,
                                                   Forward);
          // mfh 18 Mar 2013: Aztec's implementation of "symmetric
          // Gauss-Seidel" does _not_ do an Import between the forward
          // and backward sweeps.  This makes sense, because Aztec
          // considers "symmetric Gauss-Seidel" a subdomain solver.
          if (doImportBetweenDirections) {
            // Communicate again before the Backward sweep.
            if (! importer.is_null ()) {
              X_colMap->doImport (*X_domainMap, *importer, INSERT);
            }
          }
          this->template localGaussSeidel<ST, ST> (*B_in, *X_colMap, D,
                                                   dampingFactor,
                                                   Backward);
        }
        else {
          this->template reorderedLocalGaussSeidel<ST, ST> (*B_in, *X_colMap,
                                                            D, rowIndices,
                                                            dampingFactor,
                                                            Forward);
          if (doImportBetweenDirections) {
            // Communicate again before the Backward sweep.
            if (! importer.is_null ()) {
              X_colMap->doImport (*X_domainMap, *importer, INSERT);
            }
          }
          this->template reorderedLocalGaussSeidel<ST, ST> (*B_in, *X_colMap,
                                                            D, rowIndices,
                                                            dampingFactor,
                                                            Backward);
        }
      }
    }
 
    if (copiedInput) {
      deep_copy (X, *X_domainMap); // Copy back from X_domainMap to X.
    }
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  gaussSeidelCopy (MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& X,
                   const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& B,
                   const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& D,
                   const Scalar& dampingFactor,
                   const ESweepDirection direction,
                   const int numSweeps,
                   const bool zeroInitialGuess) const
  {
    reorderedGaussSeidelCopy (X, B, D, Teuchos::null, dampingFactor, direction,
                              numSweeps, zeroInitialGuess);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  reorderedGaussSeidelCopy (MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& X,
                            const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& B,
                            const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& D,
                            const Teuchos::ArrayView<LocalOrdinal>& rowIndices,
                            const Scalar& dampingFactor,
                            const ESweepDirection direction,
                            const int numSweeps,
                            const bool zeroInitialGuess) const
  {
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
    using Teuchos::rcpFromRef;
    using Teuchos::rcp_const_cast;
    typedef Scalar ST;
    const char prefix[] = "Tpetra::CrsMatrix::(reordered)gaussSeidelCopy: ";
    const Scalar ZERO = Teuchos::ScalarTraits<Scalar>::zero ();
 
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! isFillComplete (), std::runtime_error,
      prefix << "The matrix is not fill complete.");
    TEUCHOS_TEST_FOR_EXCEPTION(
      numSweeps < 0, std::invalid_argument,
      prefix << "The number of sweeps must be nonnegative, "
      "but you provided numSweeps = " << numSweeps << " < 0.");
 
    // Translate from global to local sweep direction.
    // While doing this, validate the input.
    ESweepDirection localDirection;
    if (direction == Forward) {
      localDirection = Forward;
    }
    else if (direction == Backward) {
      localDirection = Backward;
    }
    else if (direction == Symmetric) {
      // We'll control local sweep direction manually.
      localDirection = Forward;
    }
    else {
      TEUCHOS_TEST_FOR_EXCEPTION(
        true, std::invalid_argument,
        prefix << "The 'direction' enum does not have any of its valid "
        "values: Forward, Backward, or Symmetric.");
    }
 
    if (numSweeps == 0) {
      return;
    }
 
    RCP<const import_type> importer = this->getGraph ()->getImporter ();
    RCP<const export_type> exporter = this->getGraph ()->getExporter ();
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! exporter.is_null (), std::runtime_error,
      "This method's implementation currently requires that the matrix's row, "
      "domain, and range Maps be the same.  This cannot be the case, because "
      "the matrix has a nontrivial Export object.");
 
    RCP<const map_type> domainMap = this->getDomainMap ();
    RCP<const map_type> rangeMap = this->getRangeMap ();
    RCP<const map_type> rowMap = this->getGraph ()->getRowMap ();
    RCP<const map_type> colMap = this->getGraph ()->getColMap ();
 
#ifdef HAVE_TEUCHOS_DEBUG
    {
      // The relation 'isSameAs' is transitive.  It's also a
      // collective, so we don't have to do a "shared" test for
      // exception (i.e., a global reduction on the test value).
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! X.getMap ()->isSameAs (*domainMap), std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidelCopy requires that the input "
        "multivector X be in the domain Map of the matrix.");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! B.getMap ()->isSameAs (*rangeMap), std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidelCopy requires that the input "
        "B be in the range Map of the matrix.");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! D.getMap ()->isSameAs (*rowMap), std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidelCopy requires that the input "
        "D be in the row Map of the matrix.");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! rowMap->isSameAs (*rangeMap), std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidelCopy requires that the row Map and the "
        "range Map be the same (in the sense of Tpetra::Map::isSameAs).");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! domainMap->isSameAs (*rangeMap), std::runtime_error,
        "Tpetra::CrsMatrix::gaussSeidelCopy requires that the domain Map and "
        "the range Map of the matrix be the same.");
    }
#else
    // Forestall any compiler warnings for unused variables.
    (void) rangeMap;
    (void) rowMap;
#endif // HAVE_TEUCHOS_DEBUG
 
    // Fetch a (possibly cached) temporary column Map multivector
    // X_colMap, and a domain Map view X_domainMap of it.  Both have
    // constant stride by construction.  We know that the domain Map
    // must include the column Map, because our Gauss-Seidel kernel
    // requires that the row Map, domain Map, and range Map are all
    // the same, and that each process owns all of its own diagonal
    // entries of the matrix.
 
    RCP<MV> X_colMap;
    RCP<MV> X_domainMap;
    bool copyBackOutput = false;
    if (importer.is_null ()) {
      if (X.isConstantStride ()) {
        X_colMap = rcpFromRef (X);
        X_domainMap = rcpFromRef (X);
        // Column Map and domain Map are the same, so there are no
        // remote entries.  Thus, if we are not setting the initial
        // guess to zero, we don't have to worry about setting remote
        // entries to zero, even though we are not doing an Import in
        // this case.
        if (zeroInitialGuess) {
          X_colMap->putScalar (ZERO);
        }
        // No need to copy back to X at end.
      }
      else { // We must copy X into a constant stride multivector.
        // Just use the cached column Map multivector for that.
        // force=true means fill with zeros, so no need to fill
        // remote entries (not in domain Map) with zeros.
        X_colMap = getColumnMapMultiVector (X, true);
        // X_domainMap is always a domain Map view of the column Map
        // multivector.  In this case, the domain and column Maps are
        // the same, so X_domainMap _is_ X_colMap.
        X_domainMap = X_colMap;
        if (! zeroInitialGuess) { // Don't copy if zero initial guess
          try {
            deep_copy (*X_domainMap , X); // Copy X into constant stride MV
          } catch (std::exception& e) {
            std::ostringstream os;
            os << "Tpetra::CrsMatrix::reorderedGaussSeidelCopy: "
              "deep_copy(*X_domainMap, X) threw an exception: "
               << e.what () << ".";
            TEUCHOS_TEST_FOR_EXCEPTION(true, std::runtime_error, e.what ());
          }
        }
        copyBackOutput = true; // Don't forget to copy back at end.
        TPETRA_EFFICIENCY_WARNING(
          ! X.isConstantStride (),
          std::runtime_error,
          "gaussSeidelCopy: The current implementation of the Gauss-Seidel "
          "kernel requires that X and B both have constant stride.  Since X "
          "does not have constant stride, we had to make a copy.  This is a "
          "limitation of the current implementation and not your fault, but we "
          "still report it as an efficiency warning for your information.");
      }
    }
    else { // Column Map and domain Map are _not_ the same.
      X_colMap = getColumnMapMultiVector (X);
      X_domainMap = X_colMap->offsetViewNonConst (domainMap, 0);
 
#ifdef HAVE_TPETRA_DEBUG
      auto X_colMap_host_view =
        X_colMap->template getLocalView<Kokkos::HostSpace> ();
      auto X_domainMap_host_view =
        X_domainMap->template getLocalView<Kokkos::HostSpace> ();
 
      if (X_colMap->getLocalLength () != 0 && X_domainMap->getLocalLength ()) {
        TEUCHOS_TEST_FOR_EXCEPTION
          (X_colMap_host_view.data () != X_domainMap_host_view.data (),
           std::logic_error, "Tpetra::CrsMatrix::gaussSeidelCopy: Pointer to "
           "start of column Map view of X is not equal to pointer to start of "
           "(domain Map view of) X.  This may mean that Tpetra::MultiVector::"
           "offsetViewNonConst is broken.  "
           "Please report this bug to the Tpetra developers.");
      }
 
      TEUCHOS_TEST_FOR_EXCEPTION(
        X_colMap_host_view.extent (0) < X_domainMap_host_view.extent (0) ||
        X_colMap->getLocalLength () < X_domainMap->getLocalLength (),
        std::logic_error, "Tpetra::CrsMatrix::gaussSeidelCopy: "
        "X_colMap has fewer local rows than X_domainMap.  "
        "X_colMap_host_view.extent(0) = " << X_colMap_host_view.extent (0)
        << ", X_domainMap_host_view.extent(0) = "
        << X_domainMap_host_view.extent (0)
        << ", X_colMap->getLocalLength() = " << X_colMap->getLocalLength ()
        << ", and X_domainMap->getLocalLength() = "
        << X_domainMap->getLocalLength ()
        << ".  This means that Tpetra::MultiVector::offsetViewNonConst "
        "is broken.  Please report this bug to the Tpetra developers.");
 
      TEUCHOS_TEST_FOR_EXCEPTION(
        X_colMap->getNumVectors () != X_domainMap->getNumVectors (),
        std::logic_error, "Tpetra::CrsMatrix::gaussSeidelCopy: "
        "X_colMap has a different number of columns than X_domainMap.  "
        "X_colMap->getNumVectors() = " << X_colMap->getNumVectors ()
        << " != X_domainMap->getNumVectors() = "
        << X_domainMap->getNumVectors ()
        << ".  This means that Tpetra::MultiVector::offsetViewNonConst "
        "is broken.  Please report this bug to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
 
      if (zeroInitialGuess) {
        // No need for an Import, since we're filling with zeros.
        X_colMap->putScalar (ZERO);
      } else {
        // We could just copy X into X_domainMap.  However, that
        // wastes a copy, because the Import also does a copy (plus
        // communication).  Since the typical use case for
        // Gauss-Seidel is a small number of sweeps (2 is typical), we
        // don't want to waste that copy.  Thus, we do the Import
        // here, and skip the first Import in the first sweep.
        // Importing directly from X effects the copy into X_domainMap
        // (which is a view of X_colMap).
        X_colMap->doImport (X, *importer, INSERT);
      }
      copyBackOutput = true; // Don't forget to copy back at end.
    } // if column and domain Maps are (not) the same
 
    // The Gauss-Seidel / SOR kernel expects multivectors of constant
    // stride.  X_colMap is by construction, but B might not be.  If
    // it's not, we have to make a copy.
    RCP<const MV> B_in;
    if (B.isConstantStride ()) {
      B_in = rcpFromRef (B);
    }
    else {
      // Range Map and row Map are the same in this case, so we can
      // use the cached row Map multivector to store a constant stride
      // copy of B.
      RCP<MV> B_in_nonconst = getRowMapMultiVector (B, true);
      try {
        deep_copy (*B_in_nonconst, B);
      } catch (std::exception& e) {
        std::ostringstream os;
        os << "Tpetra::CrsMatrix::reorderedGaussSeidelCopy: "
          "deep_copy(*B_in_nonconst, B) threw an exception: "
           << e.what () << ".";
        TEUCHOS_TEST_FOR_EXCEPTION(true, std::runtime_error, e.what ());
      }
      B_in = rcp_const_cast<const MV> (B_in_nonconst);
 
      TPETRA_EFFICIENCY_WARNING(
        ! B.isConstantStride (),
        std::runtime_error,
        "gaussSeidelCopy: The current implementation requires that B have "
        "constant stride.  Since B does not have constant stride, we had to "
        "copy it into a separate constant-stride multivector.  This is a "
        "limitation of the current implementation and not your fault, but we "
        "still report it as an efficiency warning for your information.");
    }
 
    for (int sweep = 0; sweep < numSweeps; ++sweep) {
      if (! importer.is_null () && sweep > 0) {
        // We already did the first Import for the zeroth sweep above,
        // if it was necessary.
        X_colMap->doImport (*X_domainMap, *importer, INSERT);
      }
 
      // Do local Gauss-Seidel.
      if (direction != Symmetric) {
        if (rowIndices.is_null ()) {
          this->template localGaussSeidel<ST, ST> (*B_in, *X_colMap, D,
                                                   dampingFactor,
                                                   localDirection);
        }
        else {
          this->template reorderedLocalGaussSeidel<ST, ST> (*B_in, *X_colMap,
                                                            D, rowIndices,
                                                            dampingFactor,
                                                            localDirection);
        }
      }
      else { // direction == Symmetric
        if (rowIndices.is_null ()) {
          this->template localGaussSeidel<ST, ST> (*B_in, *X_colMap, D,
                                                   dampingFactor,
                                                   Forward);
          // mfh 18 Mar 2013: Aztec's implementation of "symmetric
          // Gauss-Seidel" does _not_ do an Import between the forward
          // and backward sweeps.  This makes symmetric Gauss-Seidel a
          // symmetric preconditioner if the matrix A is symmetric.  We
          // imitate Aztec's behavior here.
          this->template localGaussSeidel<ST, ST> (*B_in, *X_colMap, D,
                                                   dampingFactor,
                                                   Backward);
        }
        else {
          this->template reorderedLocalGaussSeidel<ST, ST> (*B_in, *X_colMap,
                                                            D, rowIndices,
                                                            dampingFactor,
                                                            Forward);
          this->template reorderedLocalGaussSeidel<ST, ST> (*B_in, *X_colMap,
                                                            D, rowIndices,
                                                            dampingFactor,
                                                            Backward);
 
        }
      }
    }
 
    if (copyBackOutput) {
      try {
        deep_copy (X , *X_domainMap); // Copy result back into X.
      } catch (std::exception& e) {
        TEUCHOS_TEST_FOR_EXCEPTION(
          true, std::runtime_error, prefix << "deep_copy(X, *X_domainMap) "
          "threw an exception: " << e.what ());
      }
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  template<class T>
  Teuchos::RCP<CrsMatrix<T, LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  convert () const
  {
    using Teuchos::RCP;
    typedef CrsMatrix<T, LocalOrdinal, GlobalOrdinal, Node> output_matrix_type;
    const char tfecfFuncName[] = "convert: ";
 
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! this->isFillComplete (), std::runtime_error, "This matrix (the source "
       "of the conversion) is not fill complete.  You must first call "
       "fillComplete() (possibly with the domain and range Map) without an "
       "intervening call to resumeFill(), before you may call this method.");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (! this->isStaticGraph (), std::logic_error, "This matrix (the source "
       "of the conversion) claims to be fill complete, but does not have a "
       "static (i.e., constant) graph.  Please report this bug to the Tpetra "
       "developers.");
 
    RCP<output_matrix_type> newMatrix
      (new output_matrix_type (this->getCrsGraph ()));
    // Copy old values into new values.  impl_scalar_type and T may
    // differ, so we can't use Kokkos::deep_copy.
    ::Tpetra::Details::copyConvert (newMatrix->lclMatrix_.values,
                                    this->lclMatrix_.values);
    // Since newmat has a static (const) graph, the graph already has
    // a column Map, and Import and Export objects already exist (if
    // applicable).  Thus, calling fillComplete is cheap.
    newMatrix->fillComplete (this->getDomainMap (), this->getRangeMap ());
 
    return newMatrix;
  }
 
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  checkInternalState () const
  {
#ifdef HAVE_TPETRA_DEBUG
    const char tfecfFuncName[] = "checkInternalState: ";
    const char err[] = "Internal state is not consistent.  "
      "Please report this bug to the Tpetra developers.";
 
    // This version of the graph (RCP<const crs_graph_type>) must
    // always be nonnull.
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      staticGraph_.is_null (),
      std::logic_error, err);
    // myGraph == null means that the matrix has a const ("static")
    // graph.  Otherwise, the matrix has a dynamic graph (it owns its
    // graph).
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      ! myGraph_.is_null () && myGraph_ != staticGraph_,
      std::logic_error, err);
    // if matrix is fill complete, then graph must be fill complete
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      isFillComplete () && ! staticGraph_->isFillComplete (),
      std::logic_error, err << "  Specifically, the matrix is fill complete, "
      "but its graph is NOT fill complete.");
    // if matrix is storage optimized, it should have a 1D allocation
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      isStorageOptimized () && ! values2D_.is_null (),
      std::logic_error, err);
    // if matrix/graph are static profile, then 2D allocation should not be present
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      getProfileType() == StaticProfile && values2D_ != Teuchos::null,
      std::logic_error, err);
    // if matrix/graph are dynamic profile, then 1D allocation should not be present
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      getProfileType() == DynamicProfile && k_values1D_.extent (0) > 0,
      std::logic_error, err);
    // if values are allocated and they are non-zero in number, then
    // one of the allocations should be present
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      staticGraph_->indicesAreAllocated () &&
      staticGraph_->getNodeAllocationSize() > 0 &&
      staticGraph_->getNodeNumRows() > 0
      && values2D_.is_null () &&
      k_values1D_.extent (0) == 0,
      std::logic_error, err);
    // we cannot have both a 1D and 2D allocation
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
      k_values1D_.extent (0) > 0 && values2D_ != Teuchos::null,
      std::logic_error, err << "  Specifically, k_values1D_ is allocated (has "
      "size " << k_values1D_.extent (0) << " > 0) and values2D_ is also "
      "allocated.  CrsMatrix is not suppose to have both a 1-D and a 2-D "
      "allocation at the same time.");
#endif
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  std::string
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  description () const
  {
    std::ostringstream os;
 
    os << "Tpetra::CrsMatrix (Kokkos refactor): {";
    if (this->getObjectLabel () != "") {
      os << "Label: \"" << this->getObjectLabel () << "\", ";
    }
    if (isFillComplete ()) {
      os << "isFillComplete: true"
         << ", global dimensions: [" << getGlobalNumRows () << ", "
         << getGlobalNumCols () << "]"
         << ", global number of entries: " << getGlobalNumEntries ()
         << "}";
    }
    else {
      os << "isFillComplete: false"
         << ", global dimensions: [" << getGlobalNumRows () << ", "
         << getGlobalNumCols () << "]}";
    }
    return os.str ();
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  describe (Teuchos::FancyOStream &out,
            const Teuchos::EVerbosityLevel verbLevel) const
  {
    using std::endl;
    using std::setw;
    using Teuchos::ArrayView;
    using Teuchos::Comm;
    using Teuchos::RCP;
    using Teuchos::TypeNameTraits;
    using Teuchos::VERB_DEFAULT;
    using Teuchos::VERB_NONE;
    using Teuchos::VERB_LOW;
    using Teuchos::VERB_MEDIUM;
    using Teuchos::VERB_HIGH;
    using Teuchos::VERB_EXTREME;
 
    const Teuchos::EVerbosityLevel vl = (verbLevel == VERB_DEFAULT) ? VERB_LOW : verbLevel;
 
    if (vl == VERB_NONE) {
      return; // Don't print anything at all
    }
    // By convention, describe() always begins with a tab.
    Teuchos::OSTab tab0 (out);
 
    RCP<const Comm<int> > comm = this->getComm();
    const int myRank = comm->getRank();
    const int numProcs = comm->getSize();
    size_t width = 1;
    for (size_t dec=10; dec<getGlobalNumRows(); dec *= 10) {
      ++width;
    }
    width = std::max<size_t> (width, static_cast<size_t> (11)) + 2;
 
    //    none: print nothing
    //     low: print O(1) info from node 0
    //  medium: print O(P) info, num entries per process
    //    high: print O(N) info, num entries per row
    // extreme: print O(NNZ) info: print indices and values
    //
    // for medium and higher, print constituent objects at specified verbLevel
    if (myRank == 0) {
      out << "Tpetra::CrsMatrix (Kokkos refactor):" << endl;
    }
    Teuchos::OSTab tab1 (out);
 
    if (myRank == 0) {
      if (this->getObjectLabel () != "") {
        out << "Label: \"" << this->getObjectLabel () << "\", ";
      }
      {
        out << "Template parameters:" << endl;
        Teuchos::OSTab tab2 (out);
        out << "Scalar: " << TypeNameTraits<Scalar>::name () << endl
            << "LocalOrdinal: " << TypeNameTraits<LocalOrdinal>::name () << endl
            << "GlobalOrdinal: " << TypeNameTraits<GlobalOrdinal>::name () << endl
            << "Node: " << TypeNameTraits<Node>::name () << endl;
      }
      if (isFillComplete()) {
        out << "isFillComplete: true" << endl
            << "Global dimensions: [" << getGlobalNumRows () << ", "
            << getGlobalNumCols () << "]" << endl
            << "Global number of entries: " << getGlobalNumEntries () << endl
            << endl << "Global max number of entries in a row: "
            << getGlobalMaxNumRowEntries () << endl;
      }
      else {
        out << "isFillComplete: false" << endl
            << "Global dimensions: [" << getGlobalNumRows () << ", "
            << getGlobalNumCols () << "]" << endl;
      }
    }
 
    if (vl < VERB_MEDIUM) {
      return; // all done!
    }
 
    // Describe the row Map.
    if (myRank == 0) {
      out << endl << "Row Map:" << endl;
    }
    if (getRowMap ().is_null ()) {
      if (myRank == 0) {
        out << "null" << endl;
      }
    }
    else {
      if (myRank == 0) {
        out << endl;
      }
      getRowMap ()->describe (out, vl);
    }
 
    // Describe the column Map.
    if (myRank == 0) {
      out << "Column Map: ";
    }
    if (getColMap ().is_null ()) {
      if (myRank == 0) {
        out << "null" << endl;
      }
    } else if (getColMap () == getRowMap ()) {
      if (myRank == 0) {
        out << "same as row Map" << endl;
      }
    } else {
      if (myRank == 0) {
        out << endl;
      }
      getColMap ()->describe (out, vl);
    }
 
    // Describe the domain Map.
    if (myRank == 0) {
      out << "Domain Map: ";
    }
    if (getDomainMap ().is_null ()) {
      if (myRank == 0) {
        out << "null" << endl;
      }
    } else if (getDomainMap () == getRowMap ()) {
      if (myRank == 0) {
        out << "same as row Map" << endl;
      }
    } else if (getDomainMap () == getColMap ()) {
      if (myRank == 0) {
        out << "same as column Map" << endl;
      }
    } else {
      if (myRank == 0) {
        out << endl;
      }
      getDomainMap ()->describe (out, vl);
    }
 
    // Describe the range Map.
    if (myRank == 0) {
      out << "Range Map: ";
    }
    if (getRangeMap ().is_null ()) {
      if (myRank == 0) {
        out << "null" << endl;
      }
    } else if (getRangeMap () == getDomainMap ()) {
      if (myRank == 0) {
        out << "same as domain Map" << endl;
      }
    } else if (getRangeMap () == getRowMap ()) {
      if (myRank == 0) {
        out << "same as row Map" << endl;
      }
    } else {
      if (myRank == 0) {
        out << endl;
      }
      getRangeMap ()->describe (out, vl);
    }
 
    // O(P) data
    for (int curRank = 0; curRank < numProcs; ++curRank) {
      if (myRank == curRank) {
        out << "Process rank: " << curRank << endl;
        Teuchos::OSTab tab2 (out);
        if (! staticGraph_->indicesAreAllocated ()) {
          out << "Graph indices not allocated" << endl;
        }
        else {
          out << "Number of allocated entries: "
              << staticGraph_->getNodeAllocationSize () << endl;
        }
        out << "Number of entries: " << getNodeNumEntries () << endl
            << "Max number of entries per row: " << getNodeMaxNumRowEntries ()
            << endl;
      }
      // Give output time to complete by executing some barriers.
      comm->barrier ();
      comm->barrier ();
      comm->barrier ();
    }
 
    if (vl < VERB_HIGH) {
      return; // all done!
    }
 
    // O(N) and O(NNZ) data
    for (int curRank = 0; curRank < numProcs; ++curRank) {
      if (myRank == curRank) {
        out << std::setw(width) << "Proc Rank"
            << std::setw(width) << "Global Row"
            << std::setw(width) << "Num Entries";
        if (vl == VERB_EXTREME) {
          out << std::setw(width) << "(Index,Value)";
        }
        out << endl;
        for (size_t r = 0; r < getNodeNumRows (); ++r) {
          const size_t nE = getNumEntriesInLocalRow(r);
          GlobalOrdinal gid = getRowMap()->getGlobalElement(r);
          out << std::setw(width) << myRank
              << std::setw(width) << gid
              << std::setw(width) << nE;
          if (vl == VERB_EXTREME) {
            if (isGloballyIndexed()) {
              ArrayView<const GlobalOrdinal> rowinds;
              ArrayView<const Scalar> rowvals;
              getGlobalRowView (gid, rowinds, rowvals);
              for (size_t j = 0; j < nE; ++j) {
                out << " (" << rowinds[j]
                    << ", " << rowvals[j]
                    << ") ";
              }
            }
            else if (isLocallyIndexed()) {
              ArrayView<const LocalOrdinal> rowinds;
              ArrayView<const Scalar> rowvals;
              getLocalRowView (r, rowinds, rowvals);
              for (size_t j=0; j < nE; ++j) {
                out << " (" << getColMap()->getGlobalElement(rowinds[j])
                    << ", " << rowvals[j]
                    << ") ";
              }
            } // globally or locally indexed
          } // vl == VERB_EXTREME
          out << endl;
        } // for each row r on this process
      } // if (myRank == curRank)
 
      // Give output time to complete
      comm->barrier ();
      comm->barrier ();
      comm->barrier ();
    } // for each process p
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  checkSizes (const SrcDistObject& source)
  {
    // It's not clear what kind of compatibility checks on sizes can
    // be performed here.  Epetra_CrsGraph doesn't check any sizes for
    // compatibility.
 
    // Currently, the source object must be a RowMatrix with the same
    // four template parameters as the target CrsMatrix.  We might
    // relax this requirement later.
    typedef RowMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> row_matrix_type;
    const row_matrix_type* srcRowMat =
      dynamic_cast<const row_matrix_type*> (&source);
    return (srcRowMat != NULL);
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  bool
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  useNewInterface ()
  {
    return true;
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  copyAndPermuteImpl (const RowMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>& srcMat,
                      const size_t numSameIDs,
                      const LocalOrdinal permuteToLIDs[],
                      const LocalOrdinal permuteFromLIDs[],
                      const size_t numPermutes)
  {
    using Tpetra::Details::ProfilingRegion;
    using Teuchos::Array;
    using Teuchos::ArrayView;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
#ifdef HAVE_TPETRA_DEBUG
    // Method name string for TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC.
    const char tfecfFuncName[] = "copyAndPermuteImpl: ";
#endif // HAVE_TPETRA_DEBUG
    ProfilingRegion regionCAP ("Tpetra::CrsMatrix::copyAndPermuteImpl");
 
    const bool sourceIsLocallyIndexed = srcMat.isLocallyIndexed ();
    //
    // Copy the first numSame row from source to target (this matrix).
    // This involves copying rows corresponding to LIDs [0, numSame-1].
    //
    const map_type& srcRowMap = * (srcMat.getRowMap ());
    Array<GO> rowInds;
    Array<Scalar> rowVals;
    const LO numSameIDs_as_LID = static_cast<LO> (numSameIDs);
    for (LO sourceLID = 0; sourceLID < numSameIDs_as_LID; ++sourceLID) {
      // Global ID for the current row index in the source matrix.
      // The first numSameIDs GIDs in the two input lists are the
      // same, so sourceGID == targetGID in this case.
      const GO sourceGID = srcRowMap.getGlobalElement (sourceLID);
      const GO targetGID = sourceGID;
 
      // Input views for the combineGlobalValues() call below.
      ArrayView<const GO> rowIndsConstView;
      ArrayView<const Scalar> rowValsConstView;
 
      if (sourceIsLocallyIndexed) {
        const size_t rowLength = srcMat.getNumEntriesInGlobalRow (sourceGID);
        if (rowLength > static_cast<size_t> (rowInds.size())) {
          rowInds.resize (rowLength);
          rowVals.resize (rowLength);
        }
        // Resizing invalidates an Array's views, so we must make new
        // ones, even if rowLength hasn't changed.
        ArrayView<GO> rowIndsView = rowInds.view (0, rowLength);
        ArrayView<Scalar> rowValsView = rowVals.view (0, rowLength);
 
        // The source matrix is locally indexed, so we have to get a
        // copy.  Really it's the GIDs that have to be copied (because
        // they have to be converted from LIDs).
        size_t checkRowLength = 0;
        srcMat.getGlobalRowCopy (sourceGID, rowIndsView, rowValsView, checkRowLength);
 
#ifdef HAVE_TPETRA_DEBUG
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(rowLength != checkRowLength,
          std::logic_error, "For global row index " << sourceGID << ", the source"
          " matrix's getNumEntriesInGlobalRow() method returns a row length of "
          << rowLength << ", but the getGlobalRowCopy() method reports that "
          "the row length is " << checkRowLength << ".  Please report this bug "
          "to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
 
        rowIndsConstView = rowIndsView.view (0, rowLength);
        rowValsConstView = rowValsView.view (0, rowLength);
      }
      else { // source matrix is globally indexed.
        srcMat.getGlobalRowView (sourceGID, rowIndsConstView, rowValsConstView);
      }
 
      // Combine the data into the target matrix.
      if (this->isStaticGraph ()) {
        // Applying a permutation to a matrix with a static graph
        // means REPLACE-ing entries.
        combineGlobalValues (targetGID, rowIndsConstView, rowValsConstView, REPLACE);
      }
      else {
        // Applying a permutation to a matrix with a dynamic graph
        // means INSERT-ing entries.  This has the same effect as
        // ADD, if the target graph already has an entry there.
        combineGlobalValues (targetGID, rowIndsConstView, rowValsConstView, INSERT);
      }
    } // For each of the consecutive source and target IDs that are the same
 
    //
    // Permute the remaining rows.
    //
    const map_type& tgtRowMap = * (this->getRowMap ());
    for (size_t p = 0; p < numPermutes; ++p) {
      const GO sourceGID = srcRowMap.getGlobalElement (permuteFromLIDs[p]);
      const GO targetGID = tgtRowMap.getGlobalElement (permuteToLIDs[p]);
 
      // Input views for the combineGlobalValues() call below.
      ArrayView<const GO> rowIndsConstView;
      ArrayView<const Scalar> rowValsConstView;
 
      if (sourceIsLocallyIndexed) {
        const size_t rowLength = srcMat.getNumEntriesInGlobalRow (sourceGID);
        if (rowLength > static_cast<size_t> (rowInds.size ())) {
          rowInds.resize (rowLength);
          rowVals.resize (rowLength);
        }
        // Resizing invalidates an Array's views, so we must make new
        // ones, even if rowLength hasn't changed.
        ArrayView<GO> rowIndsView = rowInds.view (0, rowLength);
        ArrayView<Scalar> rowValsView = rowVals.view (0, rowLength);
 
        // The source matrix is locally indexed, so we have to get a
        // copy.  Really it's the GIDs that have to be copied (because
        // they have to be converted from LIDs).
        size_t checkRowLength = 0;
        srcMat.getGlobalRowCopy (sourceGID, rowIndsView, rowValsView, checkRowLength);
 
#ifdef HAVE_TPETRA_DEBUG
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(rowLength != checkRowLength,
          std::logic_error, "For the source matrix's global row index "
          << sourceGID << ", the source matrix's getNumEntriesInGlobalRow() "
          "method returns a row length of " << rowLength << ", but the "
          "getGlobalRowCopy() method reports that the row length is "
          << checkRowLength << ".  Please report this bug to the Tpetra "
          "developers.");
#endif // HAVE_TPETRA_DEBUG
 
        rowIndsConstView = rowIndsView.view (0, rowLength);
        rowValsConstView = rowValsView.view (0, rowLength);
      }
      else {
        srcMat.getGlobalRowView (sourceGID, rowIndsConstView, rowValsConstView);
      }
 
      // Combine the data into the target matrix.
      if (isStaticGraph()) {
        this->combineGlobalValues (targetGID, rowIndsConstView,
                                   rowValsConstView, REPLACE);
      }
      else {
        this->combineGlobalValues (targetGID, rowIndsConstView,
                                   rowValsConstView, INSERT);
      }
    } // For each ID to permute
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  copyAndPermuteNew (const SrcDistObject& srcObj,
                     const size_t numSameIDs,
                     const Kokkos::DualView<const local_ordinal_type*, device_type>& permuteToLIDs,
                     const Kokkos::DualView<const local_ordinal_type*, device_type>& permuteFromLIDs)
  {
    using Tpetra::Details::castAwayConstDualView;
    using Tpetra::Details::dualViewStatusToString;
    using Tpetra::Details::ProfilingRegion;
    using std::endl;
    typedef Kokkos::HostSpace host_mem_space;
    typedef typename device_type::memory_space dev_mem_space;
    // Method name string for TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC.
    const char tfecfFuncName[] = "copyAndPermuteNew: ";
    ProfilingRegion regionCAP ("Tpetra::CrsMatrix::copyAndPermuteNew");
 
    // mfh 18 Oct 2017: Set TPETRA_VERBOSE to true for copious debug
    // output to std::cerr on every MPI process.  This is unwise for
    // runs with large numbers of MPI processes.
    const bool verbose = ::Tpetra::Details::Behavior::verbose ();
    std::unique_ptr<std::string> prefix;
    if (verbose) {
      int myRank = 0;
      auto map = this->getMap ();
      if (! map.is_null ()) {
        auto comm = map->getComm ();
        if (! comm.is_null ()) {
          myRank = comm->getRank ();
        }
      }
 
      // Restrict pfxStrm to inner scope to reduce high-water memory usage.
      prefix = [myRank] () {
        std::ostringstream pfxStrm;
        pfxStrm << "(Proc " << myRank << ") ";
        return std::unique_ptr<std::string> (new std::string (pfxStrm.str ()));
      } ();
      std::ostringstream os;
      os << *prefix << "Tpetra::CrsMatrix::copyAndPermuteNew: " << endl
         << *prefix << "  "
         << dualViewStatusToString (permuteToLIDs, "permuteToLIDs") << endl
         << *prefix << "  "
         << dualViewStatusToString (permuteFromLIDs, "permuteFromLIDs") << endl;
      std::cerr << os.str ();
    }
 
    const auto numPermute = permuteToLIDs.extent (0);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (numPermute != permuteFromLIDs.extent (0),
       std::invalid_argument, "permuteToLIDs.extent(0) = "
       << numPermute << "!= permuteFromLIDs.extent(0) = "
       << permuteFromLIDs.extent (0) << ".");
 
    // We want to keep permuteToLIDs and permuteFromLIDs on device, if
    // possible, but respect their current placement.  This is because
    // DistObject might use their placement to decide where to pack
    // and/or unpack.
    const bool permuteToLIDs_sync_back =
      permuteToLIDs.modified_device () >= permuteToLIDs.modified_host ();
    auto permuteToLIDs_nc = castAwayConstDualView (permuteToLIDs);
    permuteToLIDs_nc.template sync<host_mem_space> ();
    auto permuteToLIDs_h = permuteToLIDs.template view<host_mem_space> ();
 
    const bool permuteFromLIDs_sync_back =
      permuteFromLIDs.modified_device () >= permuteFromLIDs.modified_host ();
    auto permuteFromLIDs_nc = castAwayConstDualView (permuteFromLIDs);
    permuteFromLIDs_nc.template sync<host_mem_space> ();
    auto permuteFromLIDs_h = permuteFromLIDs.template view<host_mem_space> ();
 
    if (verbose) {
      std::ostringstream os;
      os << *prefix << "permuteToLIDs_sync_back: "
         << (permuteToLIDs_sync_back ? "true" : "false") << ", "
         << "permuteFromLIDs_sync_back: "
         << (permuteFromLIDs_sync_back ? "true" : "false") << endl;
      std::cerr << os.str ();
    }
 
    // This dynamic cast should succeed, because we've already tested
    // it in checkSizes().
    typedef ::Tpetra::RowMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> RMT;
    const RMT& srcMat = dynamic_cast<const RMT&> (srcObj);
 
    this->copyAndPermuteImpl (srcMat, numSameIDs, permuteToLIDs_h.data (),
                              permuteFromLIDs_h.data (), numPermute);
 
    if (permuteToLIDs_sync_back) {
      permuteToLIDs_nc.template sync<dev_mem_space> ();
    }
    if (permuteFromLIDs_sync_back) {
      permuteFromLIDs_nc.template sync<dev_mem_space> ();
    }
 
    if (verbose) {
      std::ostringstream os;
      os << *prefix << "copyAndPermuteNew: after:" << endl
         << *prefix << "  "
         << dualViewStatusToString (permuteToLIDs, "permuteToLIDs") << endl
         << *prefix << "  "
         << dualViewStatusToString (permuteFromLIDs, "permuteFromLIDs") << endl;
      std::cerr << os.str ();
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  packAndPrepareNew (const SrcDistObject& source,
                     const Kokkos::DualView<const local_ordinal_type*, device_type>& exportLIDs,
                     Kokkos::DualView<char*, buffer_device_type>& exports,
                     const Kokkos::DualView<size_t*, buffer_device_type>& numPacketsPerLID,
                     size_t& constantNumPackets,
                     Distributor& distor)
  {
    using Tpetra::Details::dualViewStatusToString;
    using Tpetra::Details::ProfilingRegion;
    using Teuchos::outArg;
    using Teuchos::REDUCE_MAX;
    using Teuchos::reduceAll;
    using std::endl;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    const char tfecfFuncName[] = "packAndPrepareNew: ";
    ProfilingRegion regionPAP ("Tpetra::CrsMatrix::packAndPrepareNew");
 
    const bool debug = ::Tpetra::Details::Behavior::debug ();
    const bool verbose = ::Tpetra::Details::Behavior::verbose ();
 
    // Processes on which the communicator is null should not participate.
    Teuchos::RCP<const Teuchos::Comm<int> > pComm = this->getComm ();
    if (pComm.is_null ()) {
      return;
    }
    const Teuchos::Comm<int>& comm = *pComm;
    const int myRank = comm.getSize ();
 
    std::unique_ptr<std::string> prefix;
    if (verbose) {
      // Restrict pfxStrm to inner scope to reduce high-water memory usage.
      prefix = [myRank] () {
        std::ostringstream pfxStrm;
        pfxStrm << "(Proc " << myRank << ") ";
        return std::unique_ptr<std::string> (new std::string (pfxStrm.str ()));
      } ();
 
      std::ostringstream os;
      os << *prefix << "Tpetra::CrsMatrix::packAndPrepareNew: " << endl
         << *prefix << "  "
         << dualViewStatusToString (exportLIDs, "exportLIDs")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (exports, "exports")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (numPacketsPerLID, "numPacketsPerLID")
         << endl;
      std::cerr << os.str ();
    }
 
    // Attempt to cast the source object to CrsMatrix.  If successful,
    // use the source object's packNew() method to pack its data for
    // communication.  Otherwise, attempt to cast to RowMatrix; if
    // successful, use the source object's pack() method.  Otherwise,
    // the source object doesn't have the right type.
    //
    // FIXME (mfh 30 Jun 2013, 11 Sep 2017) We don't even need the
    // RowMatrix to have the same Node type.  Unfortunately, we don't
    // have a way to ask if the RowMatrix is "a RowMatrix with any
    // Node type," since RowMatrix doesn't have a base class.  A
    // hypothetical RowMatrixBase<Scalar, LO, GO> class, which does
    // not currently exist, would satisfy this requirement.
    //
    // Why RowMatrixBase<Scalar, LO, GO>?  The source object's Scalar
    // type doesn't technically need to match the target object's
    // Scalar type, so we could just have RowMatrixBase<LO, GO>.  LO
    // and GO need not be the same, as long as there is no overflow of
    // the indices.  However, checking for index overflow is global
    // and therefore undesirable.
 
    std::ostringstream msg; // for collecting error messages
    int lclBad = 0; // to be set below
 
    typedef CrsMatrix<Scalar, LO, GO, Node> crs_matrix_type;
    const crs_matrix_type* srcCrsMat =
      dynamic_cast<const crs_matrix_type*> (&source);
    if (srcCrsMat != NULL) {
      if (verbose) {
        std::ostringstream os;
        os << *prefix << "Source matrix same (CrsMatrix) type as target; "
          "calling packNew" << endl;
        std::cerr << os.str ();
      }
      try {
        srcCrsMat->packNew (exportLIDs, exports, numPacketsPerLID,
                            constantNumPackets, distor);
      }
      catch (std::exception& e) {
        lclBad = 1;
        msg << "Proc " << myRank << ": " << e.what () << std::endl;
      }
    }
    else {
      using Tpetra::Details::castAwayConstDualView;
      using Kokkos::HostSpace;
      using Kokkos::subview;
      typedef Kokkos::DualView<char*, buffer_device_type> exports_type;
      typedef Kokkos::pair<size_t, size_t> range_type;
 
      if (verbose) {
        std::ostringstream os;
        os << *prefix << "Source matrix NOT same (CrsMatrix) type as target"
           << endl;
        std::cerr << os.str ();
      }
 
      typedef RowMatrix<Scalar, LO, GO, Node> row_matrix_type;
      const row_matrix_type* srcRowMat =
        dynamic_cast<const row_matrix_type*> (&source);
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (srcRowMat == NULL, std::invalid_argument,
         "The source object of the Import or Export operation is neither a "
         "CrsMatrix (with the same template parameters as the target object), "
         "nor a RowMatrix (with the same first four template parameters as the "
         "target object).");
 
      // For the RowMatrix case, we need to convert from
      // Kokkos::DualView to Teuchos::Array*.  This doesn't need to be
      // so terribly efficient, since packing a non-CrsMatrix
      // RowMatrix for Import/Export into a CrsMatrix is not a
      // critical case.  Thus, we may allocate Teuchos::Array objects
      // here and copy to and from Kokkos::*View.
 
      // Sync exportLIDs to host, and view its host data as a Teuchos::ArrayView.
      {
        auto exportLIDs_nc = castAwayConstDualView (exportLIDs);
        exportLIDs_nc.template sync<HostSpace> ();
      }
      auto exportLIDs_h = exportLIDs.template view<HostSpace> ();
      Teuchos::ArrayView<const LO> exportLIDs_av (exportLIDs_h.data (),
                                                  exportLIDs_h.size ());
 
      // pack() will allocate exports_a as needed.  We'll copy back
      // into exports (after (re)allocating exports if needed) below.
      Teuchos::Array<char> exports_a;
 
      // View exportLIDs' host data as a Teuchos::ArrayView.  We don't
      // need to sync, since we're doing write-only access, but we do
      // need to mark the DualView as modified on host.
      {
        auto numPacketsPerLID_nc = numPacketsPerLID; // const DV& -> DV
        numPacketsPerLID_nc.modified_device() = 0; // write-only host access
        numPacketsPerLID_nc.modified_host() = 1;
      }
      auto numPacketsPerLID_h = numPacketsPerLID.template view<HostSpace> ();
      Teuchos::ArrayView<size_t> numPacketsPerLID_av (numPacketsPerLID_h.data (),
                                                      numPacketsPerLID_h.size ());
 
      // Invoke RowMatrix's legacy pack() interface, using above
      // Teuchos::Array* objects.
      try {
        srcRowMat->pack (exportLIDs_av, exports_a, numPacketsPerLID_av,
                         constantNumPackets, distor);
      }
      catch (std::exception& e) {
        lclBad = 1;
        msg << "Proc " << myRank << ": " << e.what () << std::endl;
      }
 
      // Allocate 'exports', and copy exports_a back into it.
      const size_t newAllocSize = static_cast<size_t> (exports_a.size ());
      if (static_cast<size_t> (exports.extent (0)) < newAllocSize) {
        const std::string oldLabel = exports.d_view.label ();
        const std::string newLabel = (oldLabel == "") ? "exports" : oldLabel;
        exports = exports_type (newLabel, newAllocSize);
      }
      // It's safe to assume that we're working on host anyway, so
      // just keep exports sync'd to host.
      exports.modified_device() = 0; // ignore current device contents
      exports.modified_host() = 1;
 
      auto exports_h = exports.template view<HostSpace> ();
      auto exports_h_sub = subview (exports_h, range_type (0, newAllocSize));
 
      // Kokkos::deep_copy needs a Kokkos::View input, so turn
      // exports_a into a nonowning Kokkos::View first before copying.
      typedef typename exports_type::t_host::execution_space HES;
      typedef Kokkos::Device<HES, HostSpace> host_device_type;
      Kokkos::View<const char*, host_device_type>
        exports_a_kv (exports_a.getRawPtr (), newAllocSize);
      Kokkos::deep_copy (exports_h_sub, exports_a_kv);
    }
 
    if (debug) {
      int gblBad = 0; // output argument; to be set below
      reduceAll<int, int> (comm, REDUCE_MAX, lclBad, outArg (gblBad));
      if (gblBad != 0) {
        Tpetra::Details::gathervPrint (std::cerr, msg.str (), comm);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (true, std::logic_error, "packNew() or pack() threw an exception on "
           "one or more participating processes.");
      }
    }
    else {
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (lclBad != 0, std::logic_error, "packNew threw an exception on one "
         "or more participating processes.  Here is this process' error "
         "message: " << msg.str ());
    }
 
    if (verbose) {
      std::ostringstream os;
      os << *prefix << "packAndPrepareNew: Done!" << endl
         << *prefix << "  "
         << dualViewStatusToString (exportLIDs, "exportLIDs")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (exports, "exports")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (numPacketsPerLID, "numPacketsPerLID")
         << endl;
      std::cerr << os.str ();
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  packRow (char exports[],
           const size_t offset,
           const size_t numEnt,
           const GlobalOrdinal gidsIn[],
           const impl_scalar_type valsIn[],
           const size_t numBytesPerValue) const
  {
    using Kokkos::View;
    using Kokkos::subview;
    using Tpetra::Details::PackTraits;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef impl_scalar_type ST;
    typedef typename View<int*, device_type>::HostMirror::execution_space HES;
 
    if (numEnt == 0) {
      // Empty rows always take zero bytes, to ensure sparsity.
      return 0;
    }
 
    const GO gid = 0; // packValueCount wants this
    const LO numEntLO = static_cast<size_t> (numEnt);
 
    const size_t numEntBeg = offset;
    const size_t numEntLen = PackTraits<LO, HES>::packValueCount (numEntLO);
    const size_t gidsBeg = numEntBeg + numEntLen;
    const size_t gidsLen = numEnt * PackTraits<GO, HES>::packValueCount (gid);
    const size_t valsBeg = gidsBeg + gidsLen;
    const size_t valsLen = numEnt * numBytesPerValue;
 
    char* const numEntOut = exports + numEntBeg;
    char* const gidsOut = exports + gidsBeg;
    char* const valsOut = exports + valsBeg;
 
    size_t numBytesOut = 0;
    int errorCode = 0;
    numBytesOut += PackTraits<LO, HES>::packValue (numEntOut, numEntLO);
 
    {
      Kokkos::pair<int, size_t> p;
      p = PackTraits<GO, HES>::packArray (gidsOut, gidsIn, numEnt);
      errorCode += p.first;
      numBytesOut += p.second;
 
      p = PackTraits<ST, HES>::packArray (valsOut, valsIn, numEnt);
      errorCode += p.first;
      numBytesOut += p.second;
    }
 
    const size_t expectedNumBytes = numEntLen + gidsLen + valsLen;
    TEUCHOS_TEST_FOR_EXCEPTION
      (numBytesOut != expectedNumBytes, std::logic_error, "packRow: "
       "numBytesOut = " << numBytesOut << " != expectedNumBytes = "
       << expectedNumBytes << ".");
    TEUCHOS_TEST_FOR_EXCEPTION
      (errorCode != 0, std::runtime_error, "packRow: "
       "PackTraits::packArray returned a nonzero error code");
 
    return numBytesOut;
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  size_t
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  unpackRow (GlobalOrdinal gidsOut[],
             impl_scalar_type valsOut[],
             const char imports[],
             const size_t offset,
             const size_t numBytes,
             const size_t numEnt,
             const size_t numBytesPerValue)
  {
    using Kokkos::View;
    using Kokkos::subview;
    using Tpetra::Details::PackTraits;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef impl_scalar_type ST;
    typedef typename View<int*, device_type>::HostMirror::execution_space HES;
 
    if (numBytes == 0) {
      // Rows with zero bytes should always have zero entries.
      if (numEnt != 0) {
        const int myRank = this->getMap ()->getComm ()->getRank ();
        TEUCHOS_TEST_FOR_EXCEPTION
          (true, std::logic_error, "(Proc " << myRank << ") CrsMatrix::"
           "unpackRow: The number of bytes to unpack numBytes=0, but the "
           "number of entries to unpack (as reported by numPacketsPerLID) "
           "for this row numEnt=" << numEnt << " != 0.");
      }
      return 0;
    }
 
    if (numEnt == 0 && numBytes != 0) {
      const int myRank = this->getMap ()->getComm ()->getRank ();
      TEUCHOS_TEST_FOR_EXCEPTION
        (true, std::logic_error, "(Proc " << myRank << ") CrsMatrix::"
         "unpackRow: The number of entries to unpack (as reported by "
         "numPacketsPerLID) numEnt=0, but the number of bytes to unpack "
         "numBytes=" << numBytes << " != 0.");
    }
 
    const GO gid = 0; // packValueCount wants this
    const LO lid = 0; // packValueCount wants this
 
    const size_t numEntBeg = offset;
    const size_t numEntLen = PackTraits<LO, HES>::packValueCount (lid);
    const size_t gidsBeg = numEntBeg + numEntLen;
    const size_t gidsLen = numEnt * PackTraits<GO, HES>::packValueCount (gid);
    const size_t valsBeg = gidsBeg + gidsLen;
    const size_t valsLen = numEnt * numBytesPerValue;
 
    const char* const numEntIn = imports + numEntBeg;
    const char* const gidsIn = imports + gidsBeg;
    const char* const valsIn = imports + valsBeg;
 
    size_t numBytesOut = 0;
    int errorCode = 0;
    LO numEntOut;
    numBytesOut += PackTraits<LO, HES>::unpackValue (numEntOut, numEntIn);
    if (static_cast<size_t> (numEntOut) != numEnt ||
        numEntOut == static_cast<LO> (0)) {
      const int myRank = this->getMap ()->getComm ()->getRank ();
      std::ostringstream os;
      os << "(Proc " << myRank << ") CrsMatrix::unpackRow: ";
      bool firstErrorCondition = false;
      if (static_cast<size_t> (numEntOut) != numEnt) {
        os << "Number of entries from numPacketsPerLID numEnt=" << numEnt
           << " does not equal number of entries unpacked from imports "
          "buffer numEntOut=" << numEntOut << ".";
        firstErrorCondition = true;
      }
      if (numEntOut == static_cast<LO> (0)) {
        if (firstErrorCondition) {
          os << "  Also, ";
        }
        os << "Number of entries unpacked from imports buffer numEntOut=0, "
          "but number of bytes to unpack for this row numBytes=" << numBytes
           << " != 0.  This should never happen, since packRow should only "
          "ever pack rows with a nonzero number of entries.  In this case, "
          "the number of entries from numPacketsPerLID is numEnt=" << numEnt
           << ".";
      }
      TEUCHOS_TEST_FOR_EXCEPTION(true, std::logic_error, os.str ());
    }
 
    {
      Kokkos::pair<int, size_t> p;
      p = PackTraits<GO, HES>::unpackArray (gidsOut, gidsIn, numEnt);
      errorCode += p.first;
      numBytesOut += p.second;
 
      p = PackTraits<ST, HES>::unpackArray (valsOut, valsIn, numEnt);
      errorCode += p.first;
      numBytesOut += p.second;
    }
 
    TEUCHOS_TEST_FOR_EXCEPTION
      (numBytesOut != numBytes, std::logic_error, "unpackRow: numBytesOut = "
       << numBytesOut << " != numBytes = " << numBytes << ".");
 
    const size_t expectedNumBytes = numEntLen + gidsLen + valsLen;
    TEUCHOS_TEST_FOR_EXCEPTION
      (numBytesOut != expectedNumBytes, std::logic_error, "unpackRow: "
       "numBytesOut = " << numBytesOut << " != expectedNumBytes = "
       << expectedNumBytes << ".");
 
    TEUCHOS_TEST_FOR_EXCEPTION
      (errorCode != 0, std::runtime_error, "unpackRow: "
       "PackTraits::unpackArray returned a nonzero error code");
 
    return numBytesOut;
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  allocatePackSpaceNew (Kokkos::DualView<char*, buffer_device_type>& exports,
                        size_t& totalNumEntries,
                        const Kokkos::DualView<const local_ordinal_type*, device_type>& exportLIDs) const
  {
    using Tpetra::Details::dualViewStatusToString;
    using std::endl;
    typedef impl_scalar_type IST;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    //const char tfecfFuncName[] = "allocatePackSpaceNew: ";
 
    // mfh 18 Oct 2017: Set TPETRA_VERBOSE to true for copious debug
    // output to std::cerr on every MPI process.  This is unwise for
    // runs with large numbers of MPI processes.
    const bool verbose = ::Tpetra::Details::Behavior::verbose ();
    std::unique_ptr<std::string> prefix;
    if (verbose) {
      int myRank = 0;
      auto map = this->getMap ();
      if (! map.is_null ()) {
        auto comm = map->getComm ();
        if (! comm.is_null ()) {
          myRank = comm->getRank ();
        }
      }
      // Restrict pfxStrm to inner scope to reduce high-water memory usage.
      prefix = [myRank] () {
        std::ostringstream pfxStrm;
        pfxStrm << "(Proc " << myRank << ") ";
        return std::unique_ptr<std::string> (new std::string (pfxStrm.str ()));
      } ();
 
      std::ostringstream os;
      os << *prefix << "Tpetra::CrsMatrix::allocatePackSpaceNew: Before:"
         << endl
         << *prefix << "  "
         << dualViewStatusToString (exports, "exports")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (exportLIDs, "exportLIDs")
         << endl;
      std::cerr << os.str ();
    }
 
    // The number of export LIDs must fit in LocalOrdinal, assuming
    // that the LIDs are distinct and valid on the calling process.
    const LO numExportLIDs = static_cast<LO> (exportLIDs.extent (0));
 
    // We need to access exportLIDs on host, but Kokkos forbids
    // sync'ing of a DualView of const.  We won't modify the entries,
    // so it's fair to leave it const, except for sync'ing it.
    {
      Kokkos::DualView<local_ordinal_type*, device_type> exportLIDs_nc =
        Tpetra::Details::castAwayConstDualView (exportLIDs);
      exportLIDs_nc.template sync<Kokkos::HostSpace> ();
    }
    auto exportLIDs_h = exportLIDs.template view<Kokkos::HostSpace> ();
 
    // Count the total number of matrix entries to send.
    totalNumEntries = 0;
    for (LO i = 0; i < numExportLIDs; ++i) {
      const LO lclRow = exportLIDs_h[i];
      size_t curNumEntries = this->getNumEntriesInLocalRow (lclRow);
      // FIXME (mfh 25 Jan 2015) We should actually report invalid row
      // indices as an error.  Just consider them nonowned for now.
      if (curNumEntries == Teuchos::OrdinalTraits<size_t>::invalid ()) {
        curNumEntries = 0;
      }
      totalNumEntries += curNumEntries;
    }
 
    // FIXME (mfh 24 Feb 2013, 24 Mar 2017) This code is only correct
    // if sizeof(IST) is a meaningful representation of the amount of
    // data in a Scalar instance.  (LO and GO are always built-in
    // integer types.)
    //
    // Allocate the exports array.  It does NOT need padding for
    // alignment, since we use memcpy to write to / read from send /
    // receive buffers.
    const size_t allocSize =
      static_cast<size_t> (numExportLIDs) * sizeof (LO) +
      totalNumEntries * (sizeof (IST) + sizeof (GO));
    if (static_cast<size_t> (exports.extent (0)) < allocSize) {
      typedef Kokkos::DualView<char*, buffer_device_type> exports_type;
 
      const std::string oldLabel = exports.d_view.label ();
      const std::string newLabel = (oldLabel == "") ? "exports" : oldLabel;
      exports = exports_type (newLabel, allocSize);
    }
 
    if (verbose) {
      std::ostringstream os;
      os << *prefix << "Tpetra::CrsMatrix::allocatePackSpaceNew: After:"
         << endl
         << *prefix << "  "
         << dualViewStatusToString (exports, "exports")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (exportLIDs, "exportLIDs")
         << endl;
      std::cerr << os.str ();
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  packNew (const Kokkos::DualView<const local_ordinal_type*, device_type>& exportLIDs,
           Kokkos::DualView<char*, buffer_device_type>& exports,
           const Kokkos::DualView<size_t*, buffer_device_type>& numPacketsPerLID,
           size_t& constantNumPackets,
           Distributor& dist) const
  {
    // The call to packNew in packAndPrepareNew catches and handles any exceptions.
    if (this->isStaticGraph ()) {
      using ::Tpetra::Details::packCrsMatrixNew;
      packCrsMatrixNew (*this, exports, numPacketsPerLID, exportLIDs,
                        constantNumPackets, dist);
    }
    else {
      this->packNonStaticNew (exportLIDs, exports, numPacketsPerLID,
                              constantNumPackets, dist);
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  packNonStaticNew (const Kokkos::DualView<const local_ordinal_type*, device_type>& exportLIDs,
                    Kokkos::DualView<char*, buffer_device_type>& exports,
                    const Kokkos::DualView<size_t*, buffer_device_type>& numPacketsPerLID,
                    size_t& constantNumPackets,
                    Distributor& distor) const
  {
    using Kokkos::View;
    using Tpetra::Details::dualViewStatusToString;
    using Tpetra::Details::PackTraits;
    using Tpetra::Details::create_mirror_view_from_raw_host_array;
    using std::endl;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef impl_scalar_type ST;
    typedef typename View<int*, device_type>::HostMirror::execution_space HES;
    const char tfecfFuncName[] = "packNonStaticNew: ";
 
    // mfh 18 Oct 2017: Set TPETRA_VERBOSE to true for copious debug
    // output to std::cerr on every MPI process.  This is unwise for
    // runs with large numbers of MPI processes.
    const bool verbose = ::Tpetra::Details::Behavior::verbose ();
    std::unique_ptr<std::string> prefix;
    if (verbose) {
      int myRank = 0;
      auto map = this->getMap ();
      if (! map.is_null ()) {
        auto comm = map->getComm ();
        if (! comm.is_null ()) {
          myRank = comm->getRank ();
        }
      }
      // Restrict pfxStrm to inner scope to reduce high-water memory usage.
      prefix = [myRank] () {
        std::ostringstream pfxStrm;
        pfxStrm << "(Proc " << myRank << ") ";
        return std::unique_ptr<std::string> (new std::string (pfxStrm.str ()));
      } ();
 
      std::ostringstream os;
      os << *prefix << "Tpetra::CrsMatrix::packNonStaticNew:" << endl;
      std::cerr << os.str ();
    }
 
    const size_t numExportLIDs = static_cast<size_t> (exportLIDs.extent (0));
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (numExportLIDs != static_cast<size_t> (numPacketsPerLID.extent (0)),
       std::invalid_argument, "exportLIDs.size() = " << numExportLIDs
       << " != numPacketsPerLID.size() = " << numPacketsPerLID.extent (0)
       << ".");
 
    // Setting this to zero tells the caller to expect a possibly
    // different ("nonconstant") number of packets per local index
    // (i.e., a possibly different number of entries per row).
    constantNumPackets = 0;
 
    // The pack buffer 'exports' enters this method possibly
    // unallocated.  Do the first two parts of "Count, allocate, fill,
    // compute."
    size_t totalNumEntries = 0;
    this->allocatePackSpaceNew (exports, totalNumEntries, exportLIDs);
    const size_t bufSize = static_cast<size_t> (exports.extent (0));
 
    // Write-only host access
    exports.modified_device() = 0;
    exports.modified_host() = 1;
    auto exports_h = exports.template view<Kokkos::HostSpace> ();
    if (verbose) {
      std::ostringstream os;
      os << *prefix << "After marking exports as modified on host, "
         << dualViewStatusToString (exports, "exports") << endl;
      std::cerr << os.str ();
    }
 
    // Read-only host access
    auto exportLIDs_h = exportLIDs.template view<Kokkos::HostSpace> ();
 
    // Write-only host access
    numPacketsPerLID.modified_device() = 0;
    numPacketsPerLID.modified_host() = 1;
    auto numPacketsPerLID_h = numPacketsPerLID.template view<Kokkos::HostSpace> ();
 
    // Compute the number of "packets" (in this case, bytes) per
    // export LID (in this case, local index of the row to send), and
    // actually pack the data.
    size_t offset = 0; // current index into 'exports' array.
    for (size_t i = 0; i < numExportLIDs; ++i) {
      const LO lclRow = exportLIDs_h[i];
 
      size_t numEnt;
      numEnt = this->getNumEntriesInLocalRow (lclRow);
 
      // Only pack this row's data if it has a nonzero number of
      // entries.  We can do this because receiving processes get the
      // number of packets, and will know that zero packets means zero
      // entries.
      if (numEnt == 0) {
        numPacketsPerLID_h[i] = 0;
        continue;
      }
 
      // Temporary buffer for global column indices.
      View<GO*, HES> gidsIn_k;
      {
        GO gid = 0;
        gidsIn_k = PackTraits<GO, HES>::allocateArray(gid, numEnt, "gids");
      }
 
      Teuchos::ArrayView<const Scalar> valsIn;
      if (this->isLocallyIndexed ()) {
        // If the matrix is locally indexed on the calling process, we
        // have to use its column Map (which it _must_ have in this
        // case) to convert to global indices.
        Teuchos::ArrayView<const LO> lidsIn;
        this->getLocalRowView (lclRow, lidsIn, valsIn);
        const map_type& colMap = * (this->getColMap ());
        for (size_t k = 0; k < numEnt; ++k) {
          gidsIn_k[k] = colMap.getGlobalElement (lidsIn[k]);
        }
      }
      else if (this->isGloballyIndexed ()) {
        // If the matrix is globally indexed on the calling process,
        // then we can use the column indices directly.  However, we
        // have to get the global row index.  The calling process must
        // have a row Map, since otherwise it shouldn't be participating
        // in packing operations.
        Teuchos::ArrayView<const GO> gblIndView;;
        const map_type& rowMap = * (this->getRowMap ());
        const GO gblRow = rowMap.getGlobalElement (lclRow);
        this->getGlobalRowView (gblRow, gblIndView, valsIn);
        for (size_t k = 0; k < numEnt; ++k) {
          gidsIn_k[k] = gblIndView[k];
        }
      }
      // mfh 11 Sep 2017: Currently, if the matrix is neither globally
      // nor locally indexed, then it has no entries.  Therefore,
      // there is nothing to pack.  No worries!
 
      typename HES::device_type outputDevice;
      auto valsIn_k =
        create_mirror_view_from_raw_host_array (outputDevice,
                                                reinterpret_cast<const ST*> (valsIn.getRawPtr ()),
                                                valsIn.size (),
                                                true, "valsIn");
      const size_t numBytesPerValue =
        PackTraits<ST,HES>::packValueCount (valsIn[0]);
      const size_t numBytes =
        this->packRow (exports_h.data (), offset, numEnt, gidsIn_k.data (),
                       valsIn_k.data (), numBytesPerValue);
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (offset > bufSize || offset + numBytes > bufSize, std::logic_error,
         "First invalid offset into 'exports' pack buffer at index i = " << i
         << ".  exportLIDs_h[i]: " << exportLIDs_h[i] << ", bufSize: " <<
         bufSize << ", offset: " << offset << ", numBytes: " << numBytes <<
         ".");
      // numPacketsPerLID_h[i] is the number of "packets" in the
      // current local row i.  Packet=char (really "byte") so use the
      // number of bytes of the packed data for that row.
      numPacketsPerLID_h[i] = numBytes;
      offset += numBytes;
    }
 
    if (verbose) {
      std::ostringstream os;
      os << *prefix << "Tpetra::CrsMatrix::packNonStaticNew: After:" << endl
         << *prefix << "  "
         << dualViewStatusToString (exports, "exports")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (exportLIDs, "exportLIDs")
         << endl;
      std::cerr << os.str ();
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  LocalOrdinal
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  combineGlobalValuesRaw (const LocalOrdinal lclRow,
                          const LocalOrdinal numEnt,
                          const impl_scalar_type vals[],
                          const GlobalOrdinal cols[],
                          const Tpetra::CombineMode combineMode)
  {
    typedef GlobalOrdinal GO;
    //const char tfecfFuncName[] = "combineGlobalValuesRaw: ";
 
    // mfh 23 Mar 2017: This branch is not thread safe in a debug
    // build, due to use of Teuchos::ArrayView; see #229.
    const GO gblRow = this->myGraph_->rowMap_->getGlobalElement (lclRow);
    Teuchos::ArrayView<const GO> cols_av (numEnt == 0 ? NULL : cols, numEnt);
    Teuchos::ArrayView<const Scalar> vals_av (numEnt == 0 ? NULL : reinterpret_cast<const Scalar*> (vals), numEnt);
 
    // FIXME (mfh 23 Mar 2017) This is a work-around for less common
    // combine modes.  combineGlobalValues throws on error; it does
    // not return an error code.  Thus, if it returns, it succeeded.
    this->combineGlobalValues (gblRow, cols_av, vals_av, combineMode);
    return numEnt;
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  combineGlobalValues (const GlobalOrdinal globalRowIndex,
                       const Teuchos::ArrayView<const GlobalOrdinal>& columnIndices,
                       const Teuchos::ArrayView<const Scalar>& values,
                       const Tpetra::CombineMode combineMode)
  {
    const char tfecfFuncName[] = "combineGlobalValues: ";
 
    if (isStaticGraph ()) {
      // INSERT doesn't make sense for a static graph, since you
      // aren't allowed to change the structure of the graph.
      // However, all the other combine modes work.
      if (combineMode == ADD) {
        sumIntoGlobalValues (globalRowIndex, columnIndices, values);
      }
      else if (combineMode == REPLACE) {
        replaceGlobalValues (globalRowIndex, columnIndices, values);
      }
      else if (combineMode == ABSMAX) {
        using ::Tpetra::Details::AbsMax;
        AbsMax<Scalar> f;
        this->template transformGlobalValues<AbsMax<Scalar> > (globalRowIndex,
                                                               columnIndices,
                                                               values, f);
      }
      else if (combineMode == INSERT) {
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
          isStaticGraph () && combineMode == INSERT, std::invalid_argument,
          "INSERT combine mode is not allowed if the matrix has a static graph "
          "(i.e., was constructed with the CrsMatrix constructor that takes a "
          "const CrsGraph pointer).");
      }
      else {
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
          true, std::logic_error, "Invalid combine mode; should never get "
          "here!  Please report this bug to the Tpetra developers.");
      }
    }
    else { // The matrix has a dynamic graph.
      if (combineMode == ADD || combineMode == INSERT) {
        // For a dynamic graph, all incoming column indices are
        // inserted into the target graph.  Duplicate indices will
        // have their values summed.  In this context, ADD and INSERT
        // are equivalent.  We need to call insertGlobalValues()
        // anyway if the column indices don't yet exist in this row,
        // so we just call insertGlobalValues() for both cases.
        try {
          this->insertGlobalValuesFiltered (globalRowIndex, columnIndices,
                                            values);
        }
        catch (std::exception& e) {
          TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
            (true, std::runtime_error, std::endl
             << "insertGlobalValuesFiltered(" << globalRowIndex << ", "
             << std::endl << Teuchos::toString (columnIndices) << ", "
             << std::endl << Teuchos::toString (values)
             << ") threw an exception: " << std::endl << e.what ());
        }
      }
      // FIXME (mfh 14 Mar 2012):
      //
      // Implementing ABSMAX or REPLACE for a dynamic graph would
      // require modifying assembly to attach a possibly different
      // combine mode to each inserted (i, j, A_ij) entry.  For
      // example, consider two different Export operations to the same
      // target CrsMatrix, the first with ABSMAX combine mode and the
      // second with REPLACE.  This isn't a common use case, so we
      // won't mess with it for now.
      else if (combineMode == ABSMAX) {
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
          ! isStaticGraph () && combineMode == ABSMAX, std::logic_error,
          "ABSMAX combine mode when the matrix has a dynamic graph is not yet "
          "implemented.");
      }
      else if (combineMode == REPLACE) {
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
          ! isStaticGraph () && combineMode == REPLACE, std::logic_error,
          "REPLACE combine mode when the matrix has a dynamic graph is not yet "
          "implemented.");
      }
      else {
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC(
          true, std::logic_error, "Should never get here!  Please report this "
          "bug to the Tpetra developers.");
      }
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  unpackAndCombineNew (const Kokkos::DualView<const local_ordinal_type*, device_type>& importLIDs,
                       const Kokkos::DualView<const char*, buffer_device_type>& imports,
                       const Kokkos::DualView<const size_t*, buffer_device_type>& numPacketsPerLID,
                       const size_t constantNumPackets,
                       Distributor& distor,
                       const CombineMode combineMode)
  {
    using Tpetra::Details::dualViewStatusToString;
    using Tpetra::Details::ProfilingRegion;
    using std::endl;
    const char tfecfFuncName[] = "unpackAndCombineNew: ";
    ProfilingRegion regionUAC ("Tpetra::CrsMatrix::unpackAndCombineNew");
 
    const bool debug = ::Tpetra::Details::Behavior::debug ();
    const bool verbose = ::Tpetra::Details::Behavior::verbose ();
    constexpr int numValidModes = 5;
    const CombineMode validModes[numValidModes] =
      {ADD, REPLACE, ABSMAX, INSERT, ZERO};
    const char* validModeNames[numValidModes] =
      {"ADD", "REPLACE", "ABSMAX", "INSERT", "ZERO"};
 
    std::unique_ptr<std::string> prefix;
    int myRank = 0;
    if (verbose) {
      auto map = this->getMap ();
      if (! map.is_null ()) {
        auto comm = map->getComm ();
        if (! comm.is_null ()) {
          myRank = comm->getRank ();
        }
      }
      prefix = [myRank] () {
        std::ostringstream pfxStrm;
        pfxStrm << "(Proc " << myRank << ") ";
        return std::unique_ptr<std::string> (new std::string (pfxStrm.str ()));
      } ();
 
      std::ostringstream os;
      os << *prefix << "Tpetra::CrsMatrix::unpackAndCombineNew: " << endl
         << *prefix << "  "
         << dualViewStatusToString (importLIDs, "importLIDs")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (imports, "imports")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (numPacketsPerLID, "numPacketsPerLID")
         << endl
         << *prefix << "  constantNumPackets: " << constantNumPackets
         << endl
         << *prefix << "  combineMode: " << combineModeToString (combineMode)
         << endl;
      std::cerr << os.str ();
    }
 
    if (debug) {
      if (std::find (validModes, validModes+numValidModes, combineMode) ==
          validModes+numValidModes) {
        std::ostringstream os;
        os << "Invalid combine mode.  Valid modes are {";
        for (int k = 0; k < numValidModes; ++k) {
          os << validModeNames[k];
          if (k < numValidModes - 1) {
            os << ", ";
          }
        }
        os << "}.";
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (true, std::invalid_argument, os.str ());
      }
    }
 
    if (combineMode == ZERO) {
      return; // nothing to do
    }
 
    if (debug) {
      using Teuchos::reduceAll;
      std::unique_ptr<std::ostringstream> msg (new std::ostringstream ());
      int lclBad = 0;
      try {
        this->unpackAndCombineNewImpl (importLIDs, imports, numPacketsPerLID,
                                       constantNumPackets, distor, combineMode);
      } catch (std::exception& e) {
        lclBad = 1;
        *msg << e.what ();
      }
      int gblBad = 0;
      const Teuchos::Comm<int>& comm = * (this->getComm ());
      reduceAll<int, int> (comm, Teuchos::REDUCE_MAX,
                           lclBad, Teuchos::outArg (gblBad));
      if (gblBad != 0) {
        // mfh 22 Oct 2017: 'prefix' might be null, since it is only
        // initialized in a debug build.  Thus, we get the process
        // rank again here.  This is an error message, so the small
        // run-time cost doesn't matter.  See #1887.
        std::ostringstream os;
        os << "(Proc " << comm.getRank () << ") " << msg->str () << endl;
        msg = std::unique_ptr<std::ostringstream> (new std::ostringstream ());
        ::Tpetra::Details::gathervPrint (*msg, os.str (), comm);
        TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
          (true, std::logic_error, std::endl << "unpackAndCombineNewImpl() "
           "threw an exception on one or more participating processes: "
           << endl << msg->str ());
      }
    }
    else {
      this->unpackAndCombineNewImpl (importLIDs, imports, numPacketsPerLID,
                                     constantNumPackets, distor, combineMode);
    }
 
    if (verbose) {
      std::ostringstream os;
      os << *prefix << "unpackAndCombineNew: Done!" << endl
         << *prefix << "  "
         << dualViewStatusToString (importLIDs, "importLIDs")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (imports, "imports")
         << endl
         << *prefix << "  "
         << dualViewStatusToString (numPacketsPerLID, "numPacketsPerLID")
         << endl;
      std::cerr << os.str ();
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  unpackAndCombineNewImpl (const Kokkos::DualView<const LocalOrdinal*, device_type>& importLIDs,
                           const Kokkos::DualView<const char*, buffer_device_type>& imports,
                           const Kokkos::DualView<const size_t*, buffer_device_type>& numPacketsPerLID,
                           const size_t constantNumPackets,
                           Distributor & distor,
                           const CombineMode combineMode,
                           const bool atomic)
  {
    // Exception are caught and handled upstream, so we just call the
    // implementations directly.
    if (this->isStaticGraph ()) {
      using ::Tpetra::Details::unpackCrsMatrixAndCombineNew;
      unpackCrsMatrixAndCombineNew (*this, imports, numPacketsPerLID,
                                    importLIDs, constantNumPackets,
                                    distor, combineMode, atomic);
    }
    else {
      this->unpackAndCombineNewImplNonStatic (importLIDs, imports,
                                              numPacketsPerLID,
                                              constantNumPackets,
                                              distor, combineMode);
    }
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  unpackAndCombineNewImplNonStatic (const Kokkos::DualView<const LocalOrdinal*, device_type>& importLIDs,
                                    const Kokkos::DualView<const char*, buffer_device_type>& imports,
                                    const Kokkos::DualView<const size_t*, buffer_device_type>& numPacketsPerLID,
                                    const size_t constantNumPackets,
                                    Distributor& distor,
                                    const CombineMode combineMode)
  {
    using Kokkos::View;
    using Kokkos::subview;
    using Kokkos::MemoryUnmanaged;
    using Tpetra::Details::castAwayConstDualView;
    using Tpetra::Details::create_mirror_view_from_raw_host_array;
    using Tpetra::Details::PackTraits;
    using std::endl;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef impl_scalar_type ST;
    typedef typename Teuchos::ArrayView<const LO>::size_type size_type;
    typedef typename View<int*, device_type>::HostMirror::execution_space HES;
    typedef std::pair<typename View<int*, HES>::size_type,
                      typename View<int*, HES>::size_type> pair_type;
    typedef View<GO*, HES, MemoryUnmanaged> gids_out_type;
    typedef View<ST*, HES, MemoryUnmanaged> vals_out_type;
    const char tfecfFuncName[] = "unpackAndCombineNewImplNonStatic: ";
 
    // mfh 18 Oct 2017: Set TPETRA_VERBOSE to true for copious debug
    // output to std::cerr on every MPI process.  This is unwise for
    // runs with large numbers of MPI processes.
    const bool verbose = ::Tpetra::Details::Behavior::verbose ();
    std::unique_ptr<std::string> prefix;
    if (verbose) {
      int myRank = 0;
      auto map = this->getMap ();
      if (! map.is_null ()) {
        auto comm = map->getComm ();
        if (! comm.is_null ()) {
          myRank = comm->getRank ();
        }
      }
      // Restrict pfxStrm to inner scope to reduce high-water memory usage.
      prefix = [myRank] () {
        std::ostringstream pfxStrm;
        pfxStrm << "(Proc " << myRank << ") ";
        return std::unique_ptr<std::string> (new std::string (pfxStrm.str ()));
      } ();
 
      std::ostringstream os;
      os << *prefix << "Tpetra::CrsMatrix::unpackAndCombineNewImplNonStatic:"
         << endl; // we've already printed statuses of DualViews
      std::cerr << os.str ();
    }
 
    const size_type numImportLIDs = importLIDs.extent (0);
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (numImportLIDs != static_cast<size_type> (numPacketsPerLID.extent (0)),
       std::invalid_argument, "importLIDs.size() = " << numImportLIDs
       << " != numPacketsPerLID.size() = " << numPacketsPerLID.extent (0)
       << ".");
    TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
      (combineMode != ADD && combineMode != INSERT && combineMode != REPLACE &&
       combineMode != ABSMAX && combineMode != ZERO, std::invalid_argument,
       "Invalid CombineMode value " << combineMode << ".  Valid "
       << "values include ADD, INSERT, REPLACE, ABSMAX, and ZERO.");
    if (combineMode == ZERO || numImportLIDs == 0) {
      return; // nothing to do; no need to combine entries
    }
 
    // We're unpacking on host.  This is read-only host access of imports.
    {
      auto imports_nc = castAwayConstDualView (imports);
      imports_nc.template sync<Kokkos::HostSpace> ();
    }
    auto imports_h = imports.template view<Kokkos::HostSpace> ();
 
    // Read-only host access.
    {
      auto numPacketsPerLID_nc = castAwayConstDualView (numPacketsPerLID);
      numPacketsPerLID_nc.template sync<Kokkos::HostSpace> ();
    }
    auto numPacketsPerLID_h = numPacketsPerLID.template view<Kokkos::HostSpace> ();
 
    // Read-only host access.
    {
      auto importLIDs_nc = castAwayConstDualView (importLIDs);
      importLIDs_nc.template sync<Kokkos::HostSpace> ();
    }
    auto importLIDs_h = importLIDs.template view<Kokkos::HostSpace> ();
 
    size_t numBytesPerValue;
    {
      // FIXME (mfh 17 Feb 2015, tjf 2 Aug 2017) What do I do about Scalar types
      // with run-time size?  We already assume that all entries in both the
      // source and target matrices have the same size.  If the calling process
      // owns at least one entry in either matrix, we can use that entry to set
      // the size.  However, it is possible that the calling process owns no
      // entries.  In that case, we're in trouble.  One way to fix this would be
      // for each row's data to contain the run-time size.  This is only
      // necessary if the size is not a compile-time constant.
      Scalar val;
      numBytesPerValue = PackTraits<ST, HES>::packValueCount (val);
    }
 
    // Determine the maximum number of entries in any one row
    size_t offset = 0;
    size_t maxRowNumEnt = 0;
    for (size_type i = 0; i < numImportLIDs; ++i) {
      const size_t numBytes = numPacketsPerLID_h[i];
      if (numBytes == 0) {
        continue; // empty buffer for that row means that the row is empty
      }
      // We need to unpack a nonzero number of entries for this row.
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (offset + numBytes > static_cast<size_t> (imports_h.extent (0)),
         std::logic_error, "At local row index importLIDs_h[i=" << i << "]="
         << importLIDs_h[i] << ", offset (=" << offset << ") + numBytes (="
         << numBytes << ") > imports_h.extent(0)="
         << imports_h.extent (0) << ".");
#endif // HAVE_TPETRA_DEBUG
 
      LO numEntLO = 0;
 
#ifdef HAVE_TPETRA_DEBUG
      const size_t theNumBytes = PackTraits<LO, HES>::packValueCount (numEntLO);
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (theNumBytes > numBytes, std::logic_error, "theNumBytes = "
         << theNumBytes << " > numBytes = " << numBytes << ".");
#endif // HAVE_TPETRA_DEBUG
 
      const char* const inBuf = imports_h.data () + offset;
      const size_t actualNumBytes =
        PackTraits<LO, HES>::unpackValue (numEntLO, inBuf);
 
#ifdef HAVE_TPETRA_DEBUG
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (actualNumBytes > numBytes, std::logic_error, "At i = " << i
         << ", actualNumBytes=" << actualNumBytes
         << " > numBytes=" << numBytes << ".");
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (numEntLO == 0, std::logic_error, "At local row index importLIDs_h[i="
         << i << "]=" << importLIDs_h[i] << ", the number of entries read "
         "from the packed data is numEntLO=" << numEntLO << ", but numBytes="
         << numBytes << " != 0.");
#else
      (void) actualNumBytes;
#endif // HAVE_TPETRA_DEBUG
 
      maxRowNumEnt = std::max (static_cast<size_t> (numEntLO), maxRowNumEnt);
      offset += numBytes;
    }
 
    // Temporary space to cache incoming global column indices and
    // values.  Column indices come in as global indices, in case the
    // source object's column Map differs from the target object's
    // (this's) column Map.
    View<GO*, HES> gblColInds;
    View<LO*, HES> lclColInds;
    View<ST*, HES> vals;
    {
      GO gid = 0;
      LO lid = 0;
      // FIXME (mfh 17 Feb 2015, tjf 2 Aug 2017) What do I do about Scalar types
      // with run-time size?  We already assume that all entries in both the
      // source and target matrices have the same size.  If the calling process
      // owns at least one entry in either matrix, we can use that entry to set
      // the size.  However, it is possible that the calling process owns no
      // entries.  In that case, we're in trouble.  One way to fix this would be
      // for each row's data to contain the run-time size.  This is only
      // necessary if the size is not a compile-time constant.
      Scalar val;
      gblColInds = PackTraits<GO, HES>::allocateArray (gid, maxRowNumEnt, "gids");
      lclColInds = PackTraits<LO, HES>::allocateArray (lid, maxRowNumEnt, "lids");
      vals = PackTraits<ST, HES>::allocateArray (val, maxRowNumEnt, "vals");
    }
 
    offset = 0;
    for (size_type i = 0; i < numImportLIDs; ++i) {
      const size_t numBytes = numPacketsPerLID_h[i];
      if (numBytes == 0) {
        continue; // empty buffer for that row means that the row is empty
      }
      LO numEntLO = 0;
      const char* const inBuf = imports_h.data () + offset;
      const size_t actualNumBytes = PackTraits<LO, HES>::unpackValue (numEntLO, inBuf);
      (void) actualNumBytes;
 
      const size_t numEnt = static_cast<size_t>(numEntLO);;
      const LO lclRow = importLIDs_h[i];
 
      gids_out_type gidsOut = subview (gblColInds, pair_type (0, numEnt));
      vals_out_type valsOut = subview (vals, pair_type (0, numEnt));
 
      const size_t numBytesOut =
        unpackRow (gidsOut.data (), valsOut.data (), imports_h.data (),
                   offset, numBytes, numEnt, numBytesPerValue);
      TEUCHOS_TEST_FOR_EXCEPTION_CLASS_FUNC
        (numBytes != numBytesOut, std::logic_error, "At i = " << i << ", "
         << "numBytes = " << numBytes << " != numBytesOut = " << numBytesOut
         << ".");
 
      const ST* const valsRaw = const_cast<const ST*> (valsOut.data ());
      const GO* const gidsRaw = const_cast<const GO*> (gidsOut.data ());
      this->combineGlobalValuesRaw (lclRow, numEnt, valsRaw, gidsRaw, combineMode);
 
      // Don't update offset until current LID has succeeded.
      offset += numBytes;
    } // for each import LID i
  }
 
  template<class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getColumnMapMultiVector (const MV& X_domainMap,
                           const bool force) const
  {
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
 
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! this->hasColMap (), std::runtime_error, "Tpetra::CrsMatrix::getColumn"
      "MapMultiVector: You may only call this method if the matrix has a "
      "column Map.  If the matrix does not yet have a column Map, you should "
      "first call fillComplete (with domain and range Map if necessary).");
 
    // If the graph is not fill complete, then the Import object (if
    // one should exist) hasn't been constructed yet.
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! this->getGraph ()->isFillComplete (), std::runtime_error, "Tpetra::"
      "CrsMatrix::getColumnMapMultiVector: You may only call this method if "
      "this matrix's graph is fill complete.");
 
    const size_t numVecs = X_domainMap.getNumVectors ();
    RCP<const import_type> importer = this->getGraph ()->getImporter ();
    RCP<const map_type> colMap = this->getColMap ();
 
    RCP<MV> X_colMap; // null by default
 
    // If the Import object is trivial (null), then we don't need a
    // separate column Map multivector.  Just return null in that
    // case.  The caller is responsible for knowing not to use the
    // returned null pointer.
    //
    // If the Import is nontrivial, then we do need a separate
    // column Map multivector for the Import operation.  Check in
    // that case if we have to (re)create the column Map
    // multivector.
    if (! importer.is_null () || force) {
      if (importMV_.is_null () || importMV_->getNumVectors () != numVecs) {
        X_colMap = rcp (new MV (colMap, numVecs));
 
        // Cache the newly created multivector for later reuse.
        importMV_ = X_colMap;
      }
      else { // Yay, we can reuse the cached multivector!
        X_colMap = importMV_;
        // mfh 09 Jan 2013: We don't have to fill with zeros first,
        // because the Import uses INSERT combine mode, which overwrites
        // existing entries.
        //
        //X_colMap->putScalar (ZERO);
      }
    }
    return X_colMap;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  getRowMapMultiVector (const MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node>& Y_rangeMap,
                        const bool force) const
  {
    using Teuchos::null;
    using Teuchos::RCP;
    using Teuchos::rcp;
 
    // If the graph is not fill complete, then the Export object (if
    // one should exist) hasn't been constructed yet.
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! this->getGraph ()->isFillComplete (), std::runtime_error, "Tpetra::"
      "CrsMatrix::getRowMapMultiVector: You may only call this method if this "
      "matrix's graph is fill complete.");
 
    const size_t numVecs = Y_rangeMap.getNumVectors ();
    RCP<const export_type> exporter = this->getGraph ()->getExporter ();
    // Every version of the constructor takes either a row Map, or a
    // graph (all of whose constructors take a row Map).  Thus, the
    // matrix always has a row Map.
    RCP<const map_type> rowMap = this->getRowMap ();
 
    RCP<MV> Y_rowMap; // null by default
 
    // If the Export object is trivial (null), then we don't need a
    // separate row Map multivector.  Just return null in that case.
    // The caller is responsible for knowing not to use the returned
    // null pointer.
    //
    // If the Export is nontrivial, then we do need a separate row
    // Map multivector for the Export operation.  Check in that case
    // if we have to (re)create the row Map multivector.
    if (! exporter.is_null () || force) {
      if (exportMV_.is_null () || exportMV_->getNumVectors () != numVecs) {
        Y_rowMap = rcp (new MV (rowMap, numVecs));
        exportMV_ = Y_rowMap; // Cache the newly created MV for later reuse.
      }
      else { // Yay, we can reuse the cached multivector!
        Y_rowMap = exportMV_;
      }
    }
    return Y_rowMap;
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  removeEmptyProcessesInPlace (const Teuchos::RCP<const map_type>& newMap)
  {
    TEUCHOS_TEST_FOR_EXCEPTION(
      myGraph_.is_null (), std::logic_error, "Tpetra::CrsMatrix::"
      "removeEmptyProcessesInPlace: This method does not work when the matrix "
      "was created with a constant graph (that is, when it was created using "
      "the version of its constructor that takes an RCP<const CrsGraph>).  "
      "This is because the matrix is not allowed to modify the graph in that "
      "case, but removing empty processes requires modifying the graph.");
    myGraph_->removeEmptyProcessesInPlace (newMap);
    // Even though CrsMatrix's row Map (as returned by getRowMap())
    // comes from its CrsGraph, CrsMatrix still implements DistObject,
    // so we also have to change the DistObject's Map.
    this->map_ = this->getRowMap ();
    // In the nonconst graph case, staticGraph_ is just a const
    // pointer to myGraph_.  This assignment is probably redundant,
    // but it doesn't hurt.
    staticGraph_ = Teuchos::rcp_const_cast<const Graph> (myGraph_);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  Teuchos::RCP<RowMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> >
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  add (const Scalar& alpha,
       const RowMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>& A,
       const Scalar& beta,
       const Teuchos::RCP<const map_type>& domainMap,
       const Teuchos::RCP<const map_type>& rangeMap,
       const Teuchos::RCP<Teuchos::ParameterList>& params) const
  {
    using Teuchos::Array;
    using Teuchos::ArrayRCP;
    using Teuchos::ArrayView;
    using Teuchos::ParameterList;
    using Teuchos::RCP;
    using Teuchos::rcp;
    using Teuchos::rcp_implicit_cast;
    using Teuchos::sublist;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef RowMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> row_matrix_type;
    typedef CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> crs_matrix_type;
 
    const crs_matrix_type& B = *this; // a convenient abbreviation
    const Scalar ZERO = Teuchos::ScalarTraits<Scalar>::zero ();
    const Scalar ONE = Teuchos::ScalarTraits<Scalar>::one ();
 
    // If the user didn't supply a domain or range Map, then try to
    // get one from B first (if it has them), then from A (if it has
    // them).  If we don't have any domain or range Maps, scold the
    // user.
    RCP<const map_type> A_domainMap = A.getDomainMap ();
    RCP<const map_type> A_rangeMap = A.getRangeMap ();
    RCP<const map_type> B_domainMap = B.getDomainMap ();
    RCP<const map_type> B_rangeMap = B.getRangeMap ();
 
    RCP<const map_type> theDomainMap = domainMap;
    RCP<const map_type> theRangeMap = rangeMap;
 
    if (domainMap.is_null ()) {
      if (B_domainMap.is_null ()) {
        TEUCHOS_TEST_FOR_EXCEPTION(
          A_domainMap.is_null (), std::invalid_argument,
          "Tpetra::CrsMatrix::add: If neither A nor B have a domain Map, "
          "then you must supply a nonnull domain Map to this method.");
        theDomainMap = A_domainMap;
      } else {
        theDomainMap = B_domainMap;
      }
    }
    if (rangeMap.is_null ()) {
      if (B_rangeMap.is_null ()) {
        TEUCHOS_TEST_FOR_EXCEPTION(
          A_rangeMap.is_null (), std::invalid_argument,
          "Tpetra::CrsMatrix::add: If neither A nor B have a range Map, "
          "then you must supply a nonnull range Map to this method.");
        theRangeMap = A_rangeMap;
      } else {
        theRangeMap = B_rangeMap;
      }
    }
 
#ifdef HAVE_TPETRA_DEBUG
    // In a debug build, check that A and B have matching domain and
    // range Maps, if they have domain and range Maps at all.  (If
    // they aren't fill complete, then they may not yet have them.)
    if (! A_domainMap.is_null () && ! A_rangeMap.is_null ()) {
      if (! B_domainMap.is_null () && ! B_rangeMap.is_null ()) {
        TEUCHOS_TEST_FOR_EXCEPTION(
          ! B_domainMap->isSameAs (*A_domainMap), std::invalid_argument,
          "Tpetra::CrsMatrix::add: The input RowMatrix A must have a domain Map "
          "which is the same as (isSameAs) this RowMatrix's domain Map.");
        TEUCHOS_TEST_FOR_EXCEPTION(
          ! B_rangeMap->isSameAs (*A_rangeMap), std::invalid_argument,
          "Tpetra::CrsMatrix::add: The input RowMatrix A must have a range Map "
          "which is the same as (isSameAs) this RowMatrix's range Map.");
        TEUCHOS_TEST_FOR_EXCEPTION(
          ! domainMap.is_null () && ! domainMap->isSameAs (*B_domainMap),
          std::invalid_argument,
          "Tpetra::CrsMatrix::add: The input domain Map must be the same as "
          "(isSameAs) this RowMatrix's domain Map.");
        TEUCHOS_TEST_FOR_EXCEPTION(
          ! rangeMap.is_null () && ! rangeMap->isSameAs (*B_rangeMap),
          std::invalid_argument,
          "Tpetra::CrsMatrix::add: The input range Map must be the same as "
          "(isSameAs) this RowMatrix's range Map.");
      }
    }
    else if (! B_domainMap.is_null () && ! B_rangeMap.is_null ()) {
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! domainMap.is_null () && ! domainMap->isSameAs (*B_domainMap),
        std::invalid_argument,
        "Tpetra::CrsMatrix::add: The input domain Map must be the same as "
        "(isSameAs) this RowMatrix's domain Map.");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! rangeMap.is_null () && ! rangeMap->isSameAs (*B_rangeMap),
        std::invalid_argument,
        "Tpetra::CrsMatrix::add: The input range Map must be the same as "
        "(isSameAs) this RowMatrix's range Map.");
    }
    else {
      TEUCHOS_TEST_FOR_EXCEPTION(
        domainMap.is_null () || rangeMap.is_null (), std::invalid_argument,
        "Tpetra::CrsMatrix::add: If neither A nor B have a domain and range "
        "Map, then you must supply a nonnull domain and range Map to this "
        "method.");
    }
#endif // HAVE_TPETRA_DEBUG
 
    // What parameters do we pass to C's constructor?  Do we call
    // fillComplete on C after filling it?  And if so, what parameters
    // do we pass to C's fillComplete call?
    bool callFillComplete = true;
    RCP<ParameterList> constructorSublist;
    RCP<ParameterList> fillCompleteSublist;
    if (! params.is_null ()) {
      callFillComplete = params->get ("Call fillComplete", callFillComplete);
      constructorSublist = sublist (params, "Constructor parameters");
      fillCompleteSublist = sublist (params, "fillComplete parameters");
    }
 
    RCP<const map_type> A_rowMap = A.getRowMap ();
    RCP<const map_type> B_rowMap = B.getRowMap ();
    RCP<const map_type> C_rowMap = B_rowMap; // see discussion in documentation
    RCP<crs_matrix_type> C; // The result matrix.
 
    // If A and B's row Maps are the same, we can compute an upper
    // bound on the number of entries in each row of C, before
    // actually computing the sum.  A reasonable upper bound is the
    // sum of the two entry counts in each row.  If we choose this as
    // the actual per-row upper bound, we can use static profile.
    if (A_rowMap->isSameAs (*B_rowMap)) {
      const LO localNumRows = static_cast<LO> (A_rowMap->getNodeNumElements ());
      ArrayRCP<size_t> C_maxNumEntriesPerRow (localNumRows, 0);
 
      // Get the number of entries in each row of A.
      if (alpha != ZERO) {
        for (LO localRow = 0; localRow < localNumRows; ++localRow) {
          const size_t A_numEntries = A.getNumEntriesInLocalRow (localRow);
          C_maxNumEntriesPerRow[localRow] += A_numEntries;
        }
      }
      // Get the number of entries in each row of B.
      if (beta != ZERO) {
        for (LO localRow = 0; localRow < localNumRows; ++localRow) {
          const size_t B_numEntries = B.getNumEntriesInLocalRow (localRow);
          C_maxNumEntriesPerRow[localRow] += B_numEntries;
        }
      }
      // Construct the result matrix C.
      if (constructorSublist.is_null ()) {
        C = rcp (new crs_matrix_type (C_rowMap, C_maxNumEntriesPerRow,
                                      StaticProfile));
      } else {
        C = rcp (new crs_matrix_type (C_rowMap, C_maxNumEntriesPerRow,
                                      StaticProfile, constructorSublist));
      }
      // Since A and B have the same row Maps, we could add them
      // together all at once and merge values before we call
      // insertGlobalValues.  However, we don't really need to, since
      // we've already allocated enough space in each row of C for C
      // to do the merge itself.
    }
    else { // the row Maps of A and B are not the same
      // Construct the result matrix C.
      if (constructorSublist.is_null ()) {
        C = rcp (new crs_matrix_type (C_rowMap, 0, DynamicProfile));
      } else {
        C = rcp (new crs_matrix_type (C_rowMap, 0, DynamicProfile,
                                      constructorSublist));
      }
    }
 
#ifdef HAVE_TPETRA_DEBUG
    TEUCHOS_TEST_FOR_EXCEPTION(C.is_null (), std::logic_error,
      "Tpetra::RowMatrix::add: C should not be null at this point.  "
      "Please report this bug to the Tpetra developers.");
#endif // HAVE_TPETRA_DEBUG
    //
    // Compute C = alpha*A + beta*B.
    //
    Array<GO> ind;
    Array<Scalar> val;
 
    if (alpha != ZERO) {
      const LO A_localNumRows = static_cast<LO> (A_rowMap->getNodeNumElements ());
      for (LO localRow = 0; localRow < A_localNumRows; ++localRow) {
        size_t A_numEntries = A.getNumEntriesInLocalRow (localRow);
        const GO globalRow = A_rowMap->getGlobalElement (localRow);
        if (A_numEntries > static_cast<size_t> (ind.size ())) {
          ind.resize (A_numEntries);
          val.resize (A_numEntries);
        }
        ArrayView<GO> indView = ind (0, A_numEntries);
        ArrayView<Scalar> valView = val (0, A_numEntries);
        A.getGlobalRowCopy (globalRow, indView, valView, A_numEntries);
 
        if (alpha != ONE) {
          for (size_t k = 0; k < A_numEntries; ++k) {
            valView[k] *= alpha;
          }
        }
        C->insertGlobalValues (globalRow, indView, valView);
      }
    }
 
    if (beta != ZERO) {
      const LO B_localNumRows = static_cast<LO> (B_rowMap->getNodeNumElements ());
      for (LO localRow = 0; localRow < B_localNumRows; ++localRow) {
        size_t B_numEntries = B.getNumEntriesInLocalRow (localRow);
        const GO globalRow = B_rowMap->getGlobalElement (localRow);
        if (B_numEntries > static_cast<size_t> (ind.size ())) {
          ind.resize (B_numEntries);
          val.resize (B_numEntries);
        }
        ArrayView<GO> indView = ind (0, B_numEntries);
        ArrayView<Scalar> valView = val (0, B_numEntries);
        B.getGlobalRowCopy (globalRow, indView, valView, B_numEntries);
 
        if (beta != ONE) {
          for (size_t k = 0; k < B_numEntries; ++k) {
            valView[k] *= beta;
          }
        }
        C->insertGlobalValues (globalRow, indView, valView);
      }
    }
 
    if (callFillComplete) {
      if (fillCompleteSublist.is_null ()) {
        C->fillComplete (theDomainMap, theRangeMap);
      } else {
        C->fillComplete (theDomainMap, theRangeMap, fillCompleteSublist);
      }
    }
    return rcp_implicit_cast<row_matrix_type> (C);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  transferAndFillComplete (Teuchos::RCP<CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> > & destMat,
                           const ::Tpetra::Details::Transfer<LocalOrdinal, GlobalOrdinal, Node>& rowTransfer,
                           const Teuchos::RCP<const ::Tpetra::Details::Transfer<LocalOrdinal, GlobalOrdinal, Node> > & domainTransfer,
                           const Teuchos::RCP<const map_type>& domainMap,
                           const Teuchos::RCP<const map_type>& rangeMap,
                           const Teuchos::RCP<Teuchos::ParameterList>& params) const
  {
    using Tpetra::Details::getArrayViewFromDualView;
    using Tpetra::Details::packCrsMatrixWithOwningPIDs;
    using Tpetra::Details::unpackAndCombineWithOwningPIDsCount;
    using Tpetra::Details::unpackAndCombineIntoCrsArrays;
    using Teuchos::ArrayRCP;
    using Teuchos::ArrayView;
    using Teuchos::Comm;
    using Teuchos::ParameterList;
    using Teuchos::RCP;
    typedef LocalOrdinal LO;
    typedef GlobalOrdinal GO;
    typedef node_type NT;
    typedef CrsMatrix<Scalar, LO, GO, NT> this_type;
    typedef Vector<int, LO, GO, NT> IntVectorType;
 
#ifdef HAVE_TPETRA_MMM_TIMINGS
    std::string label;
    if(!params.is_null())
      label = params->get("Timer Label",label);
    std::string prefix = std::string("Tpetra ")+ label + std::string(": ");
    using Teuchos::TimeMonitor;
    Teuchos::RCP<Teuchos::TimeMonitor> MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC Pack-1"))));
#endif
 
    // Make sure that the input argument rowTransfer is either an
    // Import or an Export.  Import and Export are the only two
    // subclasses of Transfer that we defined, but users might
    // (unwisely, for now at least) decide to implement their own
    // subclasses.  Exclude this possibility.
    const import_type* xferAsImport = dynamic_cast<const import_type*> (&rowTransfer);
    const export_type* xferAsExport = dynamic_cast<const export_type*> (&rowTransfer);
    TEUCHOS_TEST_FOR_EXCEPTION(
      xferAsImport == NULL && xferAsExport == NULL, std::invalid_argument,
      "Tpetra::CrsMatrix::transferAndFillComplete: The 'rowTransfer' input "
      "argument must be either an Import or an Export, and its template "
      "parameters must match the corresponding template parameters of the "
      "CrsMatrix.");
 
    // Make sure that the input argument domainTransfer is either an
    // Import or an Export.  Import and Export are the only two
    // subclasses of Transfer that we defined, but users might
    // (unwisely, for now at least) decide to implement their own
    // subclasses.  Exclude this possibility.
    Teuchos::RCP<const import_type> xferDomainAsImport = Teuchos::rcp_dynamic_cast<const import_type> (domainTransfer);
    Teuchos::RCP<const export_type> xferDomainAsExport = Teuchos::rcp_dynamic_cast<const export_type> (domainTransfer);
 
    if(! domainTransfer.is_null()) {
      TEUCHOS_TEST_FOR_EXCEPTION(
         (xferDomainAsImport.is_null() && xferDomainAsExport.is_null()), std::invalid_argument,
        "Tpetra::CrsMatrix::transferAndFillComplete: The 'domainTransfer' input "
        "argument must be either an Import or an Export, and its template "
        "parameters must match the corresponding template parameters of the "
        "CrsMatrix.");
 
      TEUCHOS_TEST_FOR_EXCEPTION(
         ( xferAsImport != NULL || ! xferDomainAsImport.is_null() ) &&
         (( xferAsImport != NULL &&   xferDomainAsImport.is_null() ) ||
          ( xferAsImport == NULL && ! xferDomainAsImport.is_null() )), std::invalid_argument,
         "Tpetra::CrsMatrix::transferAndFillComplete: The 'rowTransfer' and 'domainTransfer' input "
         "arguments must be of the same type (either Import or Export).");
 
      TEUCHOS_TEST_FOR_EXCEPTION(
         ( xferAsExport != NULL || ! xferDomainAsExport.is_null() ) &&
         (( xferAsExport != NULL &&   xferDomainAsExport.is_null() ) ||
          ( xferAsExport == NULL && ! xferDomainAsExport.is_null() )), std::invalid_argument,
         "Tpetra::CrsMatrix::transferAndFillComplete: The 'rowTransfer' and 'domainTransfer' input "
         "arguments must be of the same type (either Import or Export).");
    } // domainTransfer != null
 
 
    // FIXME (mfh 15 May 2014) Wouldn't communication still be needed,
    // if the source Map is not distributed but the target Map is?
    const bool communication_needed = rowTransfer.getSourceMap ()->isDistributed ();
 
    //
    // Get the caller's parameters
    //
 
    bool reverseMode = false; // Are we in reverse mode?
    bool restrictComm = false; // Do we need to restrict the communicator?
    RCP<ParameterList> matrixparams; // parameters for the destination matrix
    if (! params.is_null ()) {
      reverseMode = params->get ("Reverse Mode", reverseMode);
      restrictComm = params->get ("Restrict Communicator", restrictComm);
      matrixparams = sublist (params, "CrsMatrix");
    }
 
    // Get the new domain and range Maps.  We need some of them for
    // error checking, now that we have the reverseMode parameter.
    RCP<const map_type> MyRowMap = reverseMode ?
      rowTransfer.getSourceMap () : rowTransfer.getTargetMap ();
    RCP<const map_type> MyColMap; // create this below
    RCP<const map_type> MyDomainMap = ! domainMap.is_null () ?
      domainMap : getDomainMap ();
    RCP<const map_type> MyRangeMap = ! rangeMap.is_null () ?
      rangeMap : getRangeMap ();
    RCP<const map_type> BaseRowMap = MyRowMap;
    RCP<const map_type> BaseDomainMap = MyDomainMap;
 
    // If the user gave us a nonnull destMat, then check whether it's
    // "pristine."  That means that it has no entries.
    //
    // FIXME (mfh 15 May 2014) If this is not true on all processes,
    // then this exception test may hang.  It would be better to
    // forward an error flag to the next communication phase.
    if (! destMat.is_null ()) {
      // FIXME (mfh 15 May 2014): The Epetra idiom for checking
      // whether a graph or matrix has no entries on the calling
      // process, is that it is neither locally nor globally indexed.
      // This may change eventually with the Kokkos refactor version
      // of Tpetra, so it would be better just to check the quantity
      // of interest directly.  Note that with the Kokkos refactor
      // version of Tpetra, asking for the total number of entries in
      // a graph or matrix that is not fill complete might require
      // computation (kernel launch), since it is not thread scalable
      // to update a count every time an entry is inserted.
      const bool NewFlag = ! destMat->getGraph ()->isLocallyIndexed () &&
        ! destMat->getGraph ()->isGloballyIndexed ();
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! NewFlag, std::invalid_argument, "Tpetra::CrsMatrix::"
        "transferAndFillComplete: The input argument 'destMat' is only allowed "
        "to be nonnull, if its graph is empty (neither locally nor globally "
        "indexed).");
      // FIXME (mfh 15 May 2014) At some point, we want to change
      // graphs and matrices so that their DistObject Map
      // (this->getMap()) may differ from their row Map.  This will
      // make redistribution for 2-D distributions more efficient.  I
      // hesitate to change this check, because I'm not sure how much
      // the code here depends on getMap() and getRowMap() being the
      // same.
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! destMat->getRowMap ()->isSameAs (*MyRowMap), std::invalid_argument,
        "Tpetra::CrsMatrix::transferAndFillComplete: The (row) Map of the "
        "input argument 'destMat' is not the same as the (row) Map specified "
        "by the input argument 'rowTransfer'.");
      TEUCHOS_TEST_FOR_EXCEPTION(
        ! destMat->checkSizes (*this), std::invalid_argument,
        "Tpetra::CrsMatrix::transferAndFillComplete: You provided a nonnull "
        "destination matrix, but checkSizes() indicates that it is not a legal "
        "legal target for redistribution from the source matrix (*this).  This "
        "may mean that they do not have the same dimensions.");
    }
 
    // If forward mode (the default), then *this's (row) Map must be
    // the same as the source Map of the Transfer.  If reverse mode,
    // then *this's (row) Map must be the same as the target Map of
    // the Transfer.
    //
    // FIXME (mfh 15 May 2014) At some point, we want to change graphs
    // and matrices so that their DistObject Map (this->getMap()) may
    // differ from their row Map.  This will make redistribution for
    // 2-D distributions more efficient.  I hesitate to change this
    // check, because I'm not sure how much the code here depends on
    // getMap() and getRowMap() being the same.
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! (reverseMode || getRowMap ()->isSameAs (*rowTransfer.getSourceMap ())),
      std::invalid_argument, "Tpetra::CrsMatrix::transferAndFillComplete: "
      "rowTransfer->getSourceMap() must match this->getRowMap() in forward mode.");
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! (! reverseMode || getRowMap ()->isSameAs (*rowTransfer.getTargetMap ())),
      std::invalid_argument, "Tpetra::CrsMatrix::transferAndFillComplete: "
      "rowTransfer->getTargetMap() must match this->getRowMap() in reverse mode.");
 
    // checks for domainTransfer
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! xferDomainAsImport.is_null() && ! xferDomainAsImport->getTargetMap()->isSameAs(*domainMap),
      std::invalid_argument,
      "Tpetra::CrsMatrix::transferAndFillComplete: The target map of the 'domainTransfer' input "
      "argument must be the same as the rebalanced domain map 'domainMap'");
 
    TEUCHOS_TEST_FOR_EXCEPTION(
      ! xferDomainAsExport.is_null() && ! xferDomainAsExport->getSourceMap()->isSameAs(*domainMap),
      std::invalid_argument,
      "Tpetra::CrsMatrix::transferAndFillComplete: The source map of the 'domainTransfer' input "
      "argument must be the same as the rebalanced domain map 'domainMap'");
 
    // The basic algorithm here is:
    //
    // 1. Call the moral equivalent of "distor.do" to handle the import.
    // 2. Copy all the Imported and Copy/Permuted data into the raw
    //    CrsMatrix / CrsGraphData pointers, still using GIDs.
    // 3. Call an optimized version of MakeColMap that avoids the
    //    Directory lookups (since the importer knows who owns all the
    //    GIDs) AND reindexes to LIDs.
    // 4. Call expertStaticFillComplete()
 
    // Get information from the Importer
    const size_t NumSameIDs = rowTransfer.getNumSameIDs();
    ArrayView<const LO> ExportLIDs = reverseMode ?
      rowTransfer.getRemoteLIDs () : rowTransfer.getExportLIDs ();
    ArrayView<const LO> RemoteLIDs = reverseMode ?
      rowTransfer.getExportLIDs () : rowTransfer.getRemoteLIDs ();
    ArrayView<const LO> PermuteToLIDs = reverseMode ?
      rowTransfer.getPermuteFromLIDs () : rowTransfer.getPermuteToLIDs ();
    ArrayView<const LO> PermuteFromLIDs = reverseMode ?
      rowTransfer.getPermuteToLIDs () : rowTransfer.getPermuteFromLIDs ();
    Distributor& Distor = rowTransfer.getDistributor ();
 
    // Owning PIDs
    Teuchos::Array<int> SourcePids;
    Teuchos::Array<int> TargetPids;
    int MyPID = getComm ()->getRank ();
 
    // Temp variables for sub-communicators
    RCP<const map_type> ReducedRowMap, ReducedColMap,
      ReducedDomainMap, ReducedRangeMap;
    RCP<const Comm<int> > ReducedComm;
 
    // If the user gave us a null destMat, then construct the new
    // destination matrix.  We will replace its column Map later.
    if (destMat.is_null ()) {
      destMat = rcp (new this_type (MyRowMap, 0, StaticProfile, matrixparams));
    }
 
    /***************************************************/
    /***** 1) First communicator restriction phase ****/
    /***************************************************/
    if (restrictComm) {
      ReducedRowMap = MyRowMap->removeEmptyProcesses ();
      ReducedComm = ReducedRowMap.is_null () ?
        Teuchos::null :
        ReducedRowMap->getComm ();
      destMat->removeEmptyProcessesInPlace (ReducedRowMap);
 
      ReducedDomainMap = MyRowMap.getRawPtr () == MyDomainMap.getRawPtr () ?
        ReducedRowMap :
        MyDomainMap->replaceCommWithSubset (ReducedComm);
      ReducedRangeMap = MyRowMap.getRawPtr () == MyRangeMap.getRawPtr () ?
        ReducedRowMap :
        MyRangeMap->replaceCommWithSubset (ReducedComm);
 
      // Reset the "my" maps
      MyRowMap    = ReducedRowMap;
      MyDomainMap = ReducedDomainMap;
      MyRangeMap  = ReducedRangeMap;
 
      // Update my PID, if we've restricted the communicator
      if (! ReducedComm.is_null ()) {
        MyPID = ReducedComm->getRank ();
      }
      else {
        MyPID = -2; // For debugging
      }
    }
    else {
      ReducedComm = MyRowMap->getComm ();
    }
 
    /***************************************************/
    /***** 2) From Tpera::DistObject::doTransfer() ****/
    /***************************************************/
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC ImportSetup"))));
#endif
    // Get the owning PIDs
    RCP<const import_type> MyImporter = getGraph ()->getImporter ();
 
    // check whether domain maps of source matrix and base domain map is the same
    bool bSameDomainMap = BaseDomainMap->isSameAs (*getDomainMap ());
 
    if (! restrictComm && ! MyImporter.is_null () && bSameDomainMap ) {
      // Same domain map as source matrix
      //
      // NOTE: This won't work for restrictComm (because the Import
      // doesn't know the restricted PIDs), though writing an
      // optimized version for that case would be easy (Import an
      // IntVector of the new PIDs).  Might want to add this later.
      Import_Util::getPids (*MyImporter, SourcePids, false);
    }
    else if (restrictComm && ! MyImporter.is_null () && bSameDomainMap) {
      // Same domain map as source matrix (restricted communicator)
      // We need one import from the domain to the column map
      IntVectorType SourceDomain_pids(getDomainMap (),true);
      IntVectorType SourceCol_pids(getColMap());
      // SourceDomain_pids contains the restricted pids
      SourceDomain_pids.putScalar(MyPID);
 
      SourceCol_pids.doImport (SourceDomain_pids, *MyImporter, INSERT);
      SourcePids.resize (getColMap ()->getNodeNumElements ());
      SourceCol_pids.get1dCopy (SourcePids ());
    }
    else if (MyImporter.is_null () && bSameDomainMap) {
      // Matrix has no off-process entries
      SourcePids.resize (getColMap ()->getNodeNumElements ());
      SourcePids.assign (getColMap ()->getNodeNumElements (), MyPID);
    }
    else if ( ! MyImporter.is_null () &&
              ! domainTransfer.is_null () ) {
      // general implementation for rectangular matrices with
      // domain map different than SourceMatrix domain map.
      // User has to provide a DomainTransfer object. We need
      // to communications (import/export)
 
      // TargetDomain_pids lives on the rebalanced new domain map
      IntVectorType TargetDomain_pids (domainMap);
      TargetDomain_pids.putScalar (MyPID);
 
      // SourceDomain_pids lives on the non-rebalanced old domain map
      IntVectorType SourceDomain_pids (getDomainMap ());
 
      // SourceCol_pids lives on the non-rebalanced old column map
      IntVectorType SourceCol_pids (getColMap ());
 
      if (! reverseMode && ! xferDomainAsImport.is_null() ) {
        SourceDomain_pids.doExport (TargetDomain_pids, *xferDomainAsImport, INSERT);
      }
      else if (reverseMode && ! xferDomainAsExport.is_null() ) {
        SourceDomain_pids.doExport (TargetDomain_pids, *xferDomainAsExport, INSERT);
      }
      else if (! reverseMode && ! xferDomainAsExport.is_null() ) {
        SourceDomain_pids.doImport (TargetDomain_pids, *xferDomainAsExport, INSERT);
      }
      else if (reverseMode && ! xferDomainAsImport.is_null() ) {
        SourceDomain_pids.doImport (TargetDomain_pids, *xferDomainAsImport, INSERT);
      }
      else {
        TEUCHOS_TEST_FOR_EXCEPTION(
          true, std::logic_error, "Tpetra::CrsMatrix::"
          "transferAndFillComplete: Should never get here!  "
          "Please report this bug to a Tpetra developer.");
      }
      SourceCol_pids.doImport (SourceDomain_pids, *MyImporter, INSERT);
      SourcePids.resize (getColMap ()->getNodeNumElements ());
      SourceCol_pids.get1dCopy (SourcePids ());
    }
    else if (BaseDomainMap->isSameAs (*BaseRowMap) &&
             getDomainMap ()->isSameAs (*getRowMap ())) {
      // We can use the rowTransfer + SourceMatrix's Import to find out who owns what.
      IntVectorType TargetRow_pids (domainMap);
      IntVectorType SourceRow_pids (getRowMap ());
      IntVectorType SourceCol_pids (getColMap ());
 
      TargetRow_pids.putScalar (MyPID);
      if (! reverseMode && xferAsImport != NULL) {
        SourceRow_pids.doExport (TargetRow_pids, *xferAsImport, INSERT);
      }
      else if (reverseMode && xferAsExport != NULL) {
        SourceRow_pids.doExport (TargetRow_pids, *xferAsExport, INSERT);
      }
      else if (! reverseMode && xferAsExport != NULL) {
        SourceRow_pids.doImport (TargetRow_pids, *xferAsExport, INSERT);
      }
      else if (reverseMode && xferAsImport != NULL) {
        SourceRow_pids.doImport (TargetRow_pids, *xferAsImport, INSERT);
      }
      else {
        TEUCHOS_TEST_FOR_EXCEPTION(
          true, std::logic_error, "Tpetra::CrsMatrix::"
          "transferAndFillComplete: Should never get here!  "
          "Please report this bug to a Tpetra developer.");
      }
      SourceCol_pids.doImport (SourceRow_pids, *MyImporter, INSERT);
      SourcePids.resize (getColMap ()->getNodeNumElements ());
      SourceCol_pids.get1dCopy (SourcePids ());
    }
    else {
      TEUCHOS_TEST_FOR_EXCEPTION(
        true, std::invalid_argument, "Tpetra::CrsMatrix::"
        "transferAndFillComplete: This method only allows either domainMap == "
        "getDomainMap (), or (domainMap == rowTransfer.getTargetMap () and "
        "getDomainMap () == getRowMap ()).");
    }
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC Pack-2"))));
#endif
 
    // Tpetra-specific stuff
    size_t constantNumPackets = destMat->constantNumberOfPackets ();
    if (constantNumPackets == 0) {
      destMat->reallocArraysForNumPacketsPerLid (ExportLIDs.size (),
                                                 RemoteLIDs.size ());
    }
    else {
      // There are a constant number of packets per element.  We
      // already know (from the number of "remote" (incoming)
      // elements) how many incoming elements we expect, so we can
      // resize the buffer accordingly.
      const size_t rbufLen = RemoteLIDs.size() * constantNumPackets;
      destMat->reallocImportsIfNeeded (rbufLen);
    }
 
    // Pack & Prepare w/ owning PIDs
#ifdef HAVE_TPETRA_DEBUG
    {
      using Teuchos::outArg;
      using Teuchos::REDUCE_MAX;
      using Teuchos::reduceAll;
      using std::cerr;
      using std::endl;
      RCP<const Teuchos::Comm<int> > comm = this->getComm ();
      const int myRank = comm->getRank ();
      const int numProcs = comm->getSize ();
 
      std::ostringstream os;
      int lclErr = 0;
      try {
        // packAndPrepare* methods modify numExportPacketsPerLID_.
        destMat->numExportPacketsPerLID_.template modify<Kokkos::HostSpace> ();
        Teuchos::ArrayView<size_t> numExportPacketsPerLID =
          getArrayViewFromDualView (destMat->numExportPacketsPerLID_);
        packCrsMatrixWithOwningPIDs (*this, destMat->exports_,
                                     numExportPacketsPerLID, ExportLIDs,
                                     SourcePids, constantNumPackets, Distor);
      }
      catch (std::exception& e) {
        os << "Proc " << myRank << ": " << e.what ();
        lclErr = 1;
      }
      int gblErr = 0;
      if (! comm.is_null ()) {
        reduceAll<int, int> (*comm, REDUCE_MAX, lclErr, outArg (gblErr));
      }
      if (gblErr != 0) {
        if (myRank == 0) {
          cerr << "packCrsMatrixWithOwningPIDs threw an exception: " << endl;
        }
        std::ostringstream err;
        for (int r = 0; r < numProcs; ++r) {
          if (r == myRank && lclErr != 0) {
            cerr << os.str () << endl;
          }
          comm->barrier ();
          comm->barrier ();
          comm->barrier ();
        }
 
        TEUCHOS_TEST_FOR_EXCEPTION(
          true, std::logic_error, "packCrsMatrixWithOwningPIDs threw an "
          "exception.");
      }
    }
 
#else
    {
      // packAndPrepare* methods modify numExportPacketsPerLID_.
      destMat->numExportPacketsPerLID_.template modify<Kokkos::HostSpace> ();
      Teuchos::ArrayView<size_t> numExportPacketsPerLID =
        getArrayViewFromDualView (destMat->numExportPacketsPerLID_);
      packCrsMatrixWithOwningPIDs (*this, destMat->exports_,
                                   numExportPacketsPerLID, ExportLIDs,
                                   SourcePids, constantNumPackets, Distor);
    }
#endif // HAVE_TPETRA_DEBUG
 
    // Do the exchange of remote data.
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC Transfer"))));
#endif
 
    if (communication_needed) {
      if (reverseMode) {
        if (constantNumPackets == 0) { // variable number of packets per LID
          // Make sure that host has the latest version, since we're
          // using the version on host.  If host has the latest
          // version, syncing to host does nothing.
          destMat->numExportPacketsPerLID_.template sync<Kokkos::HostSpace> ();
          Teuchos::ArrayView<const size_t> numExportPacketsPerLID =
            getArrayViewFromDualView (destMat->numExportPacketsPerLID_);
          destMat->numImportPacketsPerLID_.template sync<Kokkos::HostSpace> ();
          Teuchos::ArrayView<size_t> numImportPacketsPerLID =
            getArrayViewFromDualView (destMat->numImportPacketsPerLID_);
          Distor.doReversePostsAndWaits (numExportPacketsPerLID, 1,
                                         numImportPacketsPerLID);
          size_t totalImportPackets = 0;
          for (Array_size_type i = 0; i < numImportPacketsPerLID.size (); ++i) {
            totalImportPackets += numImportPacketsPerLID[i];
          }
 
          // Reallocation MUST go before setting the modified flag,
          // because it may clear out the flags.
          destMat->reallocImportsIfNeeded (totalImportPackets);
          destMat->imports_.template modify<Kokkos::HostSpace> ();
          Teuchos::ArrayView<char> hostImports =
            getArrayViewFromDualView (destMat->imports_);
          // This is a legacy host pack/unpack path, so use the host
          // version of exports_.
          destMat->exports_.template sync<Kokkos::HostSpace> ();
          Teuchos::ArrayView<const char> hostExports =
            getArrayViewFromDualView (destMat->exports_);
          Distor.doReversePostsAndWaits (hostExports,
                                         numExportPacketsPerLID,
                                         hostImports,
                                         numImportPacketsPerLID);
        }
        else { // constant number of packets per LI
          destMat->imports_.template modify<Kokkos::HostSpace> ();
          Teuchos::ArrayView<char> hostImports =
            getArrayViewFromDualView (destMat->imports_);
          // This is a legacy host pack/unpack path, so use the host
          // version of exports_.
          destMat->exports_.template sync<Kokkos::HostSpace> ();
          Teuchos::ArrayView<const char> hostExports =
            getArrayViewFromDualView (destMat->exports_);
          Distor.doReversePostsAndWaits (hostExports,
                                         constantNumPackets,
                                         hostImports);
        }
      }
      else { // forward mode (the default)
        if (constantNumPackets == 0) { // variable number of packets per LID
          // Make sure that host has the latest version, since we're
          // using the version on host.  If host has the latest
          // version, syncing to host does nothing.
          destMat->numExportPacketsPerLID_.template sync<Kokkos::HostSpace> ();
          Teuchos::ArrayView<const size_t> numExportPacketsPerLID =
            getArrayViewFromDualView (destMat->numExportPacketsPerLID_);
          destMat->numImportPacketsPerLID_.template sync<Kokkos::HostSpace> ();
          Teuchos::ArrayView<size_t> numImportPacketsPerLID =
            getArrayViewFromDualView (destMat->numImportPacketsPerLID_);
          Distor.doPostsAndWaits (numExportPacketsPerLID, 1,
                                  numImportPacketsPerLID);
          size_t totalImportPackets = 0;
          for (Array_size_type i = 0; i < numImportPacketsPerLID.size (); ++i) {
            totalImportPackets += numImportPacketsPerLID[i];
          }
 
          // Reallocation MUST go before setting the modified flag,
          // because it may clear out the flags.
          destMat->reallocImportsIfNeeded (totalImportPackets);
          destMat->imports_.template modify<Kokkos::HostSpace> ();
          Teuchos::ArrayView<char> hostImports =
            getArrayViewFromDualView (destMat->imports_);
          // This is a legacy host pack/unpack path, so use the host
          // version of exports_.
          destMat->exports_.template sync<Kokkos::HostSpace> ();
          Teuchos::ArrayView<const char> hostExports =
            getArrayViewFromDualView (destMat->exports_);
          Distor.doPostsAndWaits (hostExports,
                                  numExportPacketsPerLID,
                                  hostImports,
                                  numImportPacketsPerLID);
        }
        else { // constant number of packets per LID
          destMat->imports_.template modify<Kokkos::HostSpace> ();
          Teuchos::ArrayView<char> hostImports =
            getArrayViewFromDualView (destMat->imports_);
          // This is a legacy host pack/unpack path, so use the host
          // version of exports_.
          destMat->exports_.template sync<Kokkos::HostSpace> ();
          Teuchos::ArrayView<const char> hostExports =
            getArrayViewFromDualView (destMat->exports_);
          Distor.doPostsAndWaits (hostExports,
                                  constantNumPackets,
                                  hostImports);
        }
      }
    }
 
    /*********************************************************************/
    /**** 3) Copy all of the Same/Permute/Remote data into CSR_arrays ****/
    /*********************************************************************/
 
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC Unpack-1"))));
#endif
 
    // Backwards compatibility measure.  We'll use this again below.
    destMat->numImportPacketsPerLID_.template sync<Kokkos::HostSpace> ();
    Teuchos::ArrayView<const size_t> numImportPacketsPerLID =
      getArrayViewFromDualView (destMat->numImportPacketsPerLID_);
    destMat->imports_.template sync<Kokkos::HostSpace> ();
    Teuchos::ArrayView<const char> hostImports =
      getArrayViewFromDualView (destMat->imports_);
    size_t mynnz =
      unpackAndCombineWithOwningPIDsCount (*this, RemoteLIDs, hostImports,
                                           numImportPacketsPerLID,
                                           constantNumPackets, Distor, INSERT,
                                           NumSameIDs, PermuteToLIDs, PermuteFromLIDs);
    size_t N = BaseRowMap->getNodeNumElements ();
 
    // Allocations
    ArrayRCP<size_t> CSR_rowptr(N+1);
    ArrayRCP<GO> CSR_colind_GID;
    ArrayRCP<LO> CSR_colind_LID;
    ArrayRCP<Scalar> CSR_vals;
    CSR_colind_GID.resize (mynnz);
    CSR_vals.resize (mynnz);
 
    // If LO and GO are the same, we can reuse memory when
    // converting the column indices from global to local indices.
    if (typeid (LO) == typeid (GO)) {
      CSR_colind_LID = Teuchos::arcp_reinterpret_cast<LO> (CSR_colind_GID);
    }
    else {
      CSR_colind_LID.resize (mynnz);
    }
 
    // FIXME (mfh 15 May 2014) Why can't we abstract this out as an
    // unpackAndCombine method on a "CrsArrays" object?  This passing
    // in a huge list of arrays is icky.  Can't we have a bit of an
    // abstraction?  Implementing a concrete DistObject subclass only
    // takes five methods.
    unpackAndCombineIntoCrsArrays (*this, RemoteLIDs, hostImports,
                                   numImportPacketsPerLID, constantNumPackets,
                                   Distor, INSERT, NumSameIDs, PermuteToLIDs,
                                   PermuteFromLIDs, N, mynnz, MyPID,
                                   CSR_rowptr (), CSR_colind_GID (),
                                   Teuchos::av_reinterpret_cast<impl_scalar_type> (CSR_vals ()),
                                   SourcePids (), TargetPids);
 
    /**************************************************************/
    /**** 4) Call Optimized MakeColMap w/ no Directory Lookups ****/
    /**************************************************************/
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC Unpack-2"))));
#endif
    // Call an optimized version of makeColMap that avoids the
    // Directory lookups (since the Import object knows who owns all
    // the GIDs).
    Teuchos::Array<int> RemotePids;
    Import_Util::lowCommunicationMakeColMapAndReindex (CSR_rowptr (),
                                                       CSR_colind_LID (),
                                                       CSR_colind_GID (),
                                                       BaseDomainMap,
                                                       TargetPids, RemotePids,
                                                       MyColMap);
 
    /*******************************************************/
    /**** 4) Second communicator restriction phase      ****/
    /*******************************************************/
    if (restrictComm) {
      ReducedColMap = (MyRowMap.getRawPtr () == MyColMap.getRawPtr ()) ?
        ReducedRowMap :
        MyColMap->replaceCommWithSubset (ReducedComm);
      MyColMap = ReducedColMap; // Reset the "my" maps
    }
 
    // Replace the col map
    destMat->replaceColMap (MyColMap);
 
    // Short circuit if the processor is no longer in the communicator
    //
    // NOTE: Epetra replaces modifies all "removed" processes so they
    // have a dummy (serial) Map that doesn't touch the original
    // communicator.  Duplicating that here might be a good idea.
    if (ReducedComm.is_null ()) {
      return;
    }
 
    /***************************************************/
    /**** 5) Sort                                   ****/
    /***************************************************/
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC Unpack-3"))));
#endif
    if ((! reverseMode && xferAsImport != NULL) ||
        (reverseMode && xferAsExport != NULL)) {
      Import_Util::sortCrsEntries (CSR_rowptr (),
                                   CSR_colind_LID (),
                                   CSR_vals ());
    }
    else if ((! reverseMode && xferAsExport != NULL) ||
             (reverseMode && xferAsImport != NULL)) {
      Import_Util::sortAndMergeCrsEntries (CSR_rowptr (),
                                           CSR_colind_LID (),
                                           CSR_vals ());
      if (CSR_rowptr[N] != mynnz) {
        CSR_colind_LID.resize (CSR_rowptr[N]);
        CSR_vals.resize (CSR_rowptr[N]);
      }
    }
    else {
      TEUCHOS_TEST_FOR_EXCEPTION(
        true, std::logic_error, "Tpetra::CrsMatrix::"
        "transferAndFillComplete: Should never get here!  "
        "Please report this bug to a Tpetra developer.");
    }
    /***************************************************/
    /**** 6) Reset the colmap and the arrays        ****/
    /***************************************************/
 
    // Call constructor for the new matrix (restricted as needed)
    //
    // NOTE (mfh 15 May 2014) This should work fine for the Kokkos
    // refactor version of CrsMatrix, though it reserves the right to
    // make a deep copy of the arrays.
    destMat->setAllValues (CSR_rowptr, CSR_colind_LID, CSR_vals);
 
    /***************************************************/
    /**** 7) Build Importer & Call ESFC             ****/
    /***************************************************/
    // Pre-build the importer using the existing PIDs
    Teuchos::ParameterList esfc_params;
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC CreateImporter"))));
#endif
    RCP<import_type> MyImport = rcp (new import_type (MyDomainMap, MyColMap, RemotePids));
#ifdef HAVE_TPETRA_MMM_TIMINGS
    MM = Teuchos::rcp(new TimeMonitor(*TimeMonitor::getNewTimer(prefix + std::string("TAFC ESFC"))));
 
    esfc_params.set("Timer Label",prefix + std::string("TAFC"));
#endif
    if(!params.is_null())
      esfc_params.set("compute global constants",params->get("compute global constants",true));
 
    destMat->expertStaticFillComplete (MyDomainMap, MyRangeMap, MyImport,Teuchos::null,rcp(&esfc_params,false));
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  importAndFillComplete (Teuchos::RCP<CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> >& destMatrix,
                         const import_type& importer,
                         const Teuchos::RCP<const map_type>& domainMap,
                         const Teuchos::RCP<const map_type>& rangeMap,
                         const Teuchos::RCP<Teuchos::ParameterList>& params) const
  {
    transferAndFillComplete (destMatrix, importer, Teuchos::null, domainMap, rangeMap, params);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  importAndFillComplete (Teuchos::RCP<CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> >& destMatrix,
                         const import_type& rowImporter,
                         const import_type& domainImporter,
                         const Teuchos::RCP<const map_type>& domainMap,
                         const Teuchos::RCP<const map_type>& rangeMap,
                         const Teuchos::RCP<Teuchos::ParameterList>& params) const
  {
    transferAndFillComplete (destMatrix, rowImporter, Teuchos::rcpFromRef(domainImporter), domainMap, rangeMap, params);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  exportAndFillComplete (Teuchos::RCP<CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> >& destMatrix,
                         const export_type& exporter,
                         const Teuchos::RCP<const map_type>& domainMap,
                         const Teuchos::RCP<const map_type>& rangeMap,
                         const Teuchos::RCP<Teuchos::ParameterList>& params) const
  {
    transferAndFillComplete (destMatrix, exporter, Teuchos::null, domainMap, rangeMap, params);
  }
 
  template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
  void
  CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
  exportAndFillComplete (Teuchos::RCP<CrsMatrix<Scalar, LocalOrdinal, GlobalOrdinal, Node> >& destMatrix,
                         const export_type& rowExporter,
                         const export_type& domainExporter,
                         const Teuchos::RCP<const map_type>& domainMap,
                         const Teuchos::RCP<const map_type>& rangeMap,
                         const Teuchos::RCP<Teuchos::ParameterList>& params) const
  {
    transferAndFillComplete (destMatrix, rowExporter, Teuchos::rcpFromRef(domainExporter), domainMap, rangeMap, params);
  }
 
} // namespace Classes
} // namespace Tpetra
 
//
// Explicit instantiation macro
//
// Must be expanded from within the Tpetra namespace!
//
 
#define TPETRA_CRSMATRIX_MATRIX_INSTANT(SCALAR,LO,GO,NODE) \
  \
  namespace Classes { \
    template class CrsMatrix< SCALAR , LO , GO , NODE >; \
  } \
  template Teuchos::RCP< CrsMatrix< SCALAR , LO , GO , NODE > >   \
                CrsMatrix< SCALAR , LO , GO , NODE >::convert< SCALAR > () const;
 
#define TPETRA_CRSMATRIX_CONVERT_INSTANT(SO,SI,LO,GO,NODE) \
  \
  template Teuchos::RCP< CrsMatrix< SO , LO , GO , NODE > >   \
                CrsMatrix< SI , LO , GO , NODE >::convert< SO > () const;
 
#define TPETRA_CRSMATRIX_IMPORT_AND_FILL_COMPLETE_INSTANT(SCALAR, LO, GO, NODE) \
  template<>                                                                        \
  Teuchos::RCP<CrsMatrix<SCALAR, LO, GO, NODE> >                        \
  importAndFillCompleteCrsMatrix (const Teuchos::RCP<const CrsMatrix<SCALAR, LO, GO, NODE> >& sourceMatrix, \
                                  const Import<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type>& importer, \
                                  const Teuchos::RCP<const Map<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,      \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,     \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type> >& domainMap, \
                                  const Teuchos::RCP<const Map<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,      \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,     \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type> >& rangeMap,  \
                                                               const Teuchos::RCP<Teuchos::ParameterList>& params);
 
#define TPETRA_CRSMATRIX_IMPORT_AND_FILL_COMPLETE_INSTANT_TWO(SCALAR, LO, GO, NODE) \
  template<>                                                                        \
  Teuchos::RCP<CrsMatrix<SCALAR, LO, GO, NODE> >                        \
  importAndFillCompleteCrsMatrix (const Teuchos::RCP<const CrsMatrix<SCALAR, LO, GO, NODE> >& sourceMatrix, \
                                  const Import<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type>& rowImporter, \
                                  const Import<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type>& domainImporter, \
                                  const Teuchos::RCP<const Map<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,      \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,     \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type> >& domainMap, \
                                  const Teuchos::RCP<const Map<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,      \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,     \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type> >& rangeMap,  \
                                                               const Teuchos::RCP<Teuchos::ParameterList>& params);
 
 
#define TPETRA_CRSMATRIX_EXPORT_AND_FILL_COMPLETE_INSTANT(SCALAR, LO, GO, NODE) \
  template<>                                                                        \
  Teuchos::RCP<CrsMatrix<SCALAR, LO, GO, NODE> >                        \
  exportAndFillCompleteCrsMatrix (const Teuchos::RCP<const CrsMatrix<SCALAR, LO, GO, NODE> >& sourceMatrix, \
                                  const Export<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type>& exporter, \
                                  const Teuchos::RCP<const Map<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,      \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,     \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type> >& domainMap, \
                                  const Teuchos::RCP<const Map<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,      \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,     \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type> >& rangeMap,  \
                                                               const Teuchos::RCP<Teuchos::ParameterList>& params);
 
#define TPETRA_CRSMATRIX_EXPORT_AND_FILL_COMPLETE_INSTANT_TWO(SCALAR, LO, GO, NODE) \
  template<>                                                                        \
  Teuchos::RCP<CrsMatrix<SCALAR, LO, GO, NODE> >                        \
  exportAndFillCompleteCrsMatrix (const Teuchos::RCP<const CrsMatrix<SCALAR, LO, GO, NODE> >& sourceMatrix, \
                                  const Export<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type>& rowExporter, \
                                  const Export<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,  \
                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type>& domainExporter, \
                                  const Teuchos::RCP<const Map<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,      \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,     \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type> >& domainMap, \
                                  const Teuchos::RCP<const Map<CrsMatrix<SCALAR, LO, GO, NODE>::local_ordinal_type,      \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::global_ordinal_type,     \
                                                               CrsMatrix<SCALAR, LO, GO, NODE>::node_type> >& rangeMap,  \
                                                               const Teuchos::RCP<Teuchos::ParameterList>& params);
 
 
#define TPETRA_CRSMATRIX_INSTANT(SCALAR, LO, GO ,NODE)                    \
  TPETRA_CRSMATRIX_MATRIX_INSTANT(SCALAR, LO, GO, NODE)                   \
  TPETRA_CRSMATRIX_IMPORT_AND_FILL_COMPLETE_INSTANT(SCALAR, LO, GO, NODE) \
  TPETRA_CRSMATRIX_EXPORT_AND_FILL_COMPLETE_INSTANT(SCALAR, LO, GO, NODE) \
  TPETRA_CRSMATRIX_IMPORT_AND_FILL_COMPLETE_INSTANT_TWO(SCALAR, LO, GO, NODE) \
  TPETRA_CRSMATRIX_EXPORT_AND_FILL_COMPLETE_INSTANT_TWO(SCALAR, LO, GO, NODE)
 
#endif // TPETRA_CRSMATRIX_DEF_HPP