Tpetra_Details_CooMatrix.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_DETAILS_COOMATRIX_HPP
#define TPETRA_DETAILS_COOMATRIX_HPP


#include "TpetraCore_config.h"
#include "Tpetra_Details_PackTriples.hpp"
#include "Tpetra_DistObject.hpp"
#include "Tpetra_Details_gathervPrint.hpp"
#include "Teuchos_TypeNameTraits.hpp"

#include <initializer_list>
#include <map>
#include <memory>
#include <string>
#include <type_traits>
#include <vector>


namespace Tpetra {
namespace Details {

// Implementation details of Tpetra::Details.
// So, users REALLY should not use anything in here.
namespace Impl {

template<class IndexType>
struct CooGraphEntry {
  IndexType row;
  IndexType col;
};

template<class IndexType>
struct CompareCooGraphEntries {
  bool
  operator () (const CooGraphEntry<IndexType>& x,
               const CooGraphEntry<IndexType>& y) const
  {
    return (x.row < y.row) || (x.row == y.row && x.col < y.col);
  }
};

template<class SC, class GO>
class CooMatrixImpl {
private:
  typedef std::map<CooGraphEntry<GO>, SC,
                   CompareCooGraphEntries<GO> > entries_type;

  entries_type entries_;

public:
  typedef char packet_type;

  CooMatrixImpl () = default;

  void
  sumIntoGlobalValue (const GO gblRowInd,
                      const GO gblColInd,
                      const SC& val)
  {
    // There's no sense in worrying about the insertion hint, since
    // indices may be all over the place.  Users make no promises of
    // sorting or locality of input.
    CooGraphEntry<GO> ij {gblRowInd, gblColInd};
    auto result = this->entries_.insert ({ij, val});
    if (! result.second) { // already in the map
      result.first->second += val; // sum in the new value
    }
  }

  void
  sumIntoGlobalValues (const GO gblRowInds[],
                       const GO gblColInds[],
                       const SC vals[],
                       const std::size_t numEnt)
  {
    for (std::size_t k = 0; k < numEnt; ++k) {
      // There's no sense in worrying about the insertion hint, since
      // indices may be all over the place.  Users make no promises of
      // sorting or locality of input.
      CooGraphEntry<GO> ij {gblRowInds[k], gblColInds[k]};
      const SC val = vals[k];
      auto result = this->entries_.insert ({ij, val});
      if (! result.second) { // already in the map
        result.first->second += val; // sum in the new value
      }
    }
  }

  std::size_t
  getLclNumEntries () const
  {
    return this->entries_.size ();
  }

  void
  forAllEntries (std::function<void (const GO, const GO, const SC&)> f) const
  {
    for (auto iter = this->entries_.begin ();
         iter != this->entries_.end (); ++iter) {
      f (iter->first.row, iter->first.col, iter->second);
    }
  }

  void
  mergeIntoRow (const GO tgtGblRow,
                const CooMatrixImpl<SC, GO>& src,
                const GO srcGblRow)
  {
    // const char prefix[] =
    //   "Tpetra::Details::Impl::CooMatrixImpl::mergeIntoRow: ";

    entries_type& tgtEntries = this->entries_;
    const entries_type& srcEntries = src.entries_;

    // Find all entries with the given global row index.  The min GO
    // value is guaranteed to be the least possible column index, so
    // beg1 is a lower bound for all (row index, column index) pairs.
    // Lower bound is inclusive; upper bound is exclusive.
    auto srcBeg = srcEntries.lower_bound ({srcGblRow, std::numeric_limits<GO>::min ()});
    auto srcEnd = srcEntries.upper_bound ({srcGblRow+1, std::numeric_limits<GO>::min ()});
    auto tgtBeg = tgtEntries.lower_bound ({tgtGblRow, std::numeric_limits<GO>::min ()});
    auto tgtEnd = tgtEntries.upper_bound ({tgtGblRow+1, std::numeric_limits<GO>::min ()});

    // Don't waste time iterating over the current row of *this, if
    // the current row of src is empty.
    if (srcBeg != srcEnd) {
      for (auto tgtCur = tgtBeg; tgtCur != tgtEnd; ++tgtCur) {
        auto srcCur = srcBeg;
        while (srcCur != srcEnd && srcCur->first.col < tgtCur->first.col) {
          ++srcCur;
        }
        // At this point, one of the following is true:
        //
        // 1. srcCur == srcEnd, thus nothing more to insert.
        // 2. srcCur->first.col > tgtCur->first.col, thus the current
        //    row of src has no entry matching tgtCur->first, so we
        //    have to insert it.  Insertion does not invalidate
        //    tgtEntries iterators, and neither srcEntries nor
        //    tgtEntries have duplicates, so this is safe.
        // 3. srcCur->first.col == tgtCur->first.col, so add the two entries.
        if (srcCur != srcEnd) {
          if (srcCur->first.col == tgtCur->first.col) {
            tgtCur->second += srcCur->second;
          }
          else {
            // tgtCur is (optimally) right before where we want to put
            // the new entry, since srcCur points to the first entry
            // in srcEntries whose column index is greater than that
            // of the entry to which tgtCur points.
            (void) tgtEntries.insert (tgtCur, *srcCur);
          }
        } // if srcCur != srcEnd
      } // for each entry in the current row (lclRow) of *this
    } // if srcBeg != srcEnd
  }

  int
  countPackRow (int& numPackets,
                const GO gblRow,
                const ::Teuchos::Comm<int>& comm,
                std::ostream* errStrm = NULL) const
  {
    using ::Tpetra::Details::countPackTriples;
    using ::Tpetra::Details::countPackTriplesCount;
    using std::endl;
    const char prefix[] = "Tpetra::Details::Impl::CooMatrixImpl::countPackRow: ";
#ifdef HAVE_TPETRACORE_MPI
    int errCode = MPI_SUCCESS;
#else
    int errCode = 0;
#endif // HAVE_TPETRACORE_MPI

    // Count the number of entries in the given row.
    const GO minGO = std::numeric_limits<GO>::min ();
    auto beg = this->entries_.lower_bound ({gblRow, minGO});
    auto end = this->entries_.upper_bound ({gblRow+1, minGO});
    int numEnt = 0;
    for (auto iter = beg; iter != end; ++iter) {
      ++numEnt;
      if (numEnt == std::numeric_limits<int>::max ()) {
        if (errStrm != NULL) {
          *errStrm << prefix << "In (global) row " << gblRow << ", the number "
            "of entries thus far, numEnt = " << numEnt << ", has reached the "
            "maximum int value.  We need to store this count as int for MPI's "
            "sake." << endl;
        }
        return -1;
      }
    }

    int numPacketsForCount = 0; // output argument of countPackTriplesCount
    {
      const int errCode =
        countPackTriplesCount (comm, numPacketsForCount, errStrm);
      if (errCode != 0) {
        if (errStrm != NULL) {
          *errStrm << prefix << "countPackTriplesCount "
            "returned errCode = " << errCode << " != 0." << endl;
        }
        return errCode;
      }
      if (numPacketsForCount < 0) {
        if (errStrm != NULL) {
          *errStrm << prefix << "countPackTriplesCount returned "
            "numPacketsForCount = " << numPacketsForCount << " < 0." << endl;
        }
        return -1;
      }
    }

    int numPacketsForTriples = 0; // output argument of countPackTriples
    {
      const int errCode = countPackTriples<SC, GO> (numEnt, comm, numPacketsForTriples);
      TEUCHOS_TEST_FOR_EXCEPTION
        (errCode != 0, std::runtime_error, prefix << "countPackTriples "
         "returned errCode = " << errCode << " != 0.");
      TEUCHOS_TEST_FOR_EXCEPTION
        (numPacketsForTriples < 0, std::logic_error, prefix << "countPackTriples "
         "returned numPacketsForTriples = " << numPacketsForTriples << " < 0.");
    }

    numPackets = numPacketsForCount + numPacketsForTriples;
    return errCode;
  }

  void
  packRow (packet_type outBuf[],
           const int outBufSize,
           int& outBufCurPos, // in/out argument
           const ::Teuchos::Comm<int>& comm,
           std::vector<GO>& gblRowInds,
           std::vector<GO>& gblColInds,
           std::vector<SC>& vals,
           const GO gblRow) const
  {
    using ::Tpetra::Details::packTriples;
    using ::Tpetra::Details::packTriplesCount;
    const char prefix[] = "Tpetra::Details::Impl::CooMatrixImpl::packRow: ";

    const GO minGO = std::numeric_limits<GO>::min ();
    auto beg = this->entries_.lower_bound ({gblRow, minGO});
    auto end = this->entries_.upper_bound ({gblRow+1, minGO});

    // This doesn't actually deallocate.  Only swapping with an empty
    // std::vector does that.
    gblRowInds.resize (0);
    gblColInds.resize (0);
    vals.resize (0);

    int numEnt = 0;
    for (auto iter = beg; iter != end; ++iter) {
      gblRowInds.push_back (iter->first.row);
      gblColInds.push_back (iter->first.col);
      vals.push_back (iter->second);
      ++numEnt;
      TEUCHOS_TEST_FOR_EXCEPTION
        (numEnt == std::numeric_limits<int>::max (), std::runtime_error, prefix
         << "In (global) row " << gblRow << ", the number of entries thus far, "
         "numEnt = " << numEnt << ", has reached the maximum int value.  "
         "We need to store this count as int for MPI's sake.");
    }

    {
      const int errCode =
        packTriplesCount (numEnt, outBuf, outBufSize, outBufCurPos, comm);
      TEUCHOS_TEST_FOR_EXCEPTION
        (errCode != 0, std::runtime_error, prefix
         << "In (global) row " << gblRow << ", packTriplesCount returned "
         "errCode = " << errCode << " != 0.");
    }
    {
      const int errCode =
        packTriples (gblRowInds.data (),
                     gblColInds.data (),
                     vals.data (),
                     numEnt,
                     outBuf,
                     outBufSize,
                     outBufCurPos, // in/out argument
                     comm);
      TEUCHOS_TEST_FOR_EXCEPTION
        (errCode != 0, std::runtime_error, prefix << "In (global) row "
         << gblRow << ", packTriples returned errCode = " << errCode
         << " != 0.");
    }
  }

  GO
  getMyGlobalRowIndices (std::vector<GO>& rowInds) const
  {
    rowInds.clear ();

    GO lclMinRowInd = std::numeric_limits<GO>::max (); // compute local min
    for (typename entries_type::const_iterator iter = this->entries_.begin ();
         iter != this->entries_.end (); ++iter) {
      const GO rowInd = iter->first.row;
      if (rowInd < lclMinRowInd) {
        lclMinRowInd = rowInd;
      }
      if (rowInds.empty () || rowInds.back () != rowInd) {
        rowInds.push_back (rowInd); // don't insert duplicates
      }
    }
    return lclMinRowInd;
  }

  template<class OffsetType>
  void
  buildCrs (std::vector<OffsetType>& rowOffsets,
            GO gblColInds[],
            SC vals[]) const
  {
    static_assert (std::is_integral<OffsetType>::value,
                   "OffsetType must be a built-in integer type.");

    // clear() doesn't generally free storage; it just resets the
    // length.  Thus, we reuse any existing storage here.
    rowOffsets.clear ();

    const std::size_t numEnt = this->getLclNumEntries ();
    if (numEnt == 0) {
      rowOffsets.push_back (0);
    }
    else {
      typename entries_type::const_iterator iter = this->entries_.begin ();
      GO prevGblRowInd = iter->first.row;

      OffsetType k = 0;
      for ( ; iter != this->entries_.end (); ++iter, ++k) {
        const GO gblRowInd = iter->first.row;
        if (k == 0 || gblRowInd != prevGblRowInd) {
          // The row offsets array always has at least one entry.  The
          // first entry is always zero, and the last entry is always
          // the number of matrix entries.
          rowOffsets.push_back (k);
          prevGblRowInd = gblRowInd;
        }
        gblColInds[k] = iter->first.col;

        static_assert (std::is_same<typename std::decay<decltype (iter->second)>::type, SC>::value,
                       "The type of iter->second != SC.");
        vals[k] = iter->second;
      }
      rowOffsets.push_back (static_cast<OffsetType> (numEnt));
    }
  }

  template<class OffsetType, class LO>
  void
  buildLocallyIndexedCrs (std::vector<OffsetType>& rowOffsets,
                          LO lclColInds[],
                          SC vals[],
                          std::function<LO (const GO)> gblToLcl) const
  {
    static_assert (std::is_integral<OffsetType>::value,
                   "OffsetType must be a built-in integer type.");
    static_assert (std::is_integral<LO>::value,
                   "LO must be a built-in integer type.");

    // clear() doesn't generally free storage; it just resets the
    // length.  Thus, we reuse any existing storage here.
    rowOffsets.clear ();

    const std::size_t numEnt = this->getLclNumEntries ();
    if (numEnt == 0) {
      rowOffsets.push_back (0);
    }
    else {
      typename entries_type::const_iterator iter = this->entries_.begin ();
      GO prevGblRowInd = iter->first.row;

      OffsetType k = 0;
      for ( ; iter != this->entries_.end (); ++iter, ++k) {
        const GO gblRowInd = iter->first.row;
        if (k == 0 || gblRowInd != prevGblRowInd) {
          // The row offsets array always has at least one entry.  The
          // first entry is always zero, and the last entry is always
          // the number of matrix entries.
          rowOffsets.push_back (k);
          prevGblRowInd = gblRowInd;
        }
        lclColInds[k] = gblToLcl (iter->first.col);
        vals[k] = iter->second;
      }
      rowOffsets.push_back (static_cast<OffsetType> (numEnt));
    }
  }
};

} // namespace Impl

template<class SC,
         class LO = ::Tpetra::DistObject<char>::local_ordinal_type,
         class GO = ::Tpetra::DistObject<char>::global_ordinal_type,
         class NT = ::Tpetra::DistObject<char>::node_type>
class CooMatrix : public ::Tpetra::DistObject<char, LO, GO, NT> {
public:
  typedef char packet_type;
  typedef SC scalar_type;
  typedef LO local_ordinal_type;
  typedef GO global_ordinal_type;
  typedef NT node_type;
  typedef typename NT::device_type device_type;
  typedef ::Tpetra::Map<local_ordinal_type, global_ordinal_type, node_type> map_type;

private:
  typedef ::Tpetra::DistObject<packet_type, local_ordinal_type,
                               global_ordinal_type, node_type> base_type;

  Impl::CooMatrixImpl<SC, GO> impl_;

public:
  CooMatrix () :
    base_type (::Teuchos::null),
    localError_ (new bool (false)),
    errs_ (new std::shared_ptr<std::ostringstream> ()) // ptr to a null ptr
  {}

  CooMatrix (const ::Teuchos::RCP<const map_type>& map) :
    base_type (map),
    localError_ (new bool (false)),
    errs_ (new std::shared_ptr<std::ostringstream> ()) // ptr to a null ptr
  {}

  virtual ~CooMatrix () {}

  void
  sumIntoGlobalValue (const GO gblRowInd,
                      const GO gblColInd,
                      const SC& val)
  {
    this->impl_.sumIntoGlobalValue (gblRowInd, gblColInd, val);
  }

  void
  sumIntoGlobalValues (const GO gblRowInds[],
                       const GO gblColInds[],
                       const SC vals[],
                       const std::size_t numEnt)
  {
    this->impl_.sumIntoGlobalValues (gblRowInds, gblColInds, vals, numEnt);
  }

  void
  sumIntoGlobalValues (std::initializer_list<GO> gblRowInds,
                       std::initializer_list<GO> gblColInds,
                       std::initializer_list<SC> vals,
                       const std::size_t numEnt)
  {
    this->impl_.sumIntoGlobalValues (gblRowInds.begin (), gblColInds.begin (),
                                     vals.begin (), numEnt);
  }

  std::size_t
  getLclNumEntries () const
  {
    return this->impl_.getLclNumEntries ();
  }

  template<class OffsetType>
  void
  buildCrs (::Kokkos::View<OffsetType*, device_type>& rowOffsets,
            ::Kokkos::View<GO*, device_type>& gblColInds,
            ::Kokkos::View<typename ::Kokkos::Details::ArithTraits<SC>::val_type*, device_type>& vals)
  {
    static_assert (std::is_integral<OffsetType>::value,
                   "OffsetType must be a built-in integer type.");
    using ::Kokkos::create_mirror_view;
    using ::Kokkos::deep_copy;
    using ::Kokkos::View;
    typedef typename ::Kokkos::Details::ArithTraits<SC>::val_type ISC;

    const std::size_t numEnt = this->getLclNumEntries ();

    gblColInds = View<GO*, device_type> ("gblColInds", numEnt);
    vals = View<ISC*, device_type> ("vals", numEnt);
    auto gblColInds_h = create_mirror_view (gblColInds);
    auto vals_h = create_mirror_view (vals);

    std::vector<std::size_t> rowOffsetsSV;
    this->impl_.buildCrs (rowOffsetsSV,
                          gblColInds_h.ptr_on_device (),
                          vals_h.ptr_on_device ());
    rowOffsets =
      View<OffsetType*, device_type> ("rowOffsets", rowOffsetsSV.size ());
    typename View<OffsetType*, device_type>::HostMirror
      rowOffsets_h (rowOffsetsSV.data (), rowOffsetsSV.size ());
    deep_copy (rowOffsets, rowOffsets_h);

    deep_copy (gblColInds, gblColInds_h);
    deep_copy (vals, vals_h);
  }

  void
  fillComplete (const ::Teuchos::RCP<const ::Teuchos::Comm<int> >& comm)
  {
    if (comm.is_null ()) {
      this->map_ = ::Teuchos::null;
    }
    else {
      this->map_ = this->buildMap (comm);
    }
  }

  void
  fillComplete ()
  {
    TEUCHOS_TEST_FOR_EXCEPTION
      (this->getMap ().is_null (), std::runtime_error, "Tpetra::Details::"
       "CooMatrix::fillComplete: This object does not yet have a Map.  "
       "You must call the version of fillComplete "
       "that takes an input communicator.");
    this->fillComplete (this->getMap ()->getComm ());
  }

  bool localError () const {
    return *localError_;
  }

  std::string errorMessages () const {
    return ((*errs_).get () == NULL) ? std::string ("") : (*errs_)->str ();
  }

private:
  std::shared_ptr<bool> localError_;

  std::shared_ptr<std::shared_ptr<std::ostringstream> > errs_;

  std::ostream&
  markLocalErrorAndGetStream ()
  {
    * (this->localError_) = true;
    if ((*errs_).get () == NULL) {
      *errs_ = std::shared_ptr<std::ostringstream> (new std::ostringstream ());
    }
    return **errs_;
  }

public:
  virtual std::string description () const {
    using Teuchos::TypeNameTraits;

    std::ostringstream os;
    os << "\"Tpetra::Details::CooMatrix\": {"
       << "SC: " << TypeNameTraits<SC>::name ()
       << ", LO: " << TypeNameTraits<LO>::name ()
       << ", GO: " << TypeNameTraits<GO>::name ()
       << ", NT: " << TypeNameTraits<NT>::name ();
    if (this->getObjectLabel () != "") {
      os << ", Label: \"" << this->getObjectLabel () << "\"";
    }
    os << ", Has Map: " << (this->map_.is_null () ? "false" : "true")
       << "}";
    return os.str ();
  }

  virtual void
  describe (Teuchos::FancyOStream& out,
            const Teuchos::EVerbosityLevel verbLevel =
            Teuchos::Describable::verbLevel_default) const
  {
    using ::Tpetra::Details::gathervPrint;
    using ::Teuchos::EVerbosityLevel;
    using ::Teuchos::OSTab;
    using ::Teuchos::TypeNameTraits;
    using ::Teuchos::VERB_DEFAULT;
    using ::Teuchos::VERB_LOW;
    using ::Teuchos::VERB_MEDIUM;
    using std::endl;

    const auto vl = (verbLevel == VERB_DEFAULT) ? VERB_LOW : verbLevel;

    auto comm = this->getMap ().is_null () ?
      Teuchos::null : this->getMap ()->getComm ();
    const int myRank = comm.is_null () ? 0 : comm->getRank ();
    // const int numProcs = comm.is_null () ? 1 : comm->getSize ();

    if (vl != Teuchos::VERB_NONE) {
      // Convention is to start describe() implementations with a tab.
      OSTab tab0 (out);
      if (myRank == 0) {
        out << "\"Tpetra::Details::CooMatrix\":" << endl;
      }
      OSTab tab1 (out);
      if (myRank == 0) {
        out << "Template parameters:" << endl;
        {
          OSTab tab2 (out);
          out << "SC: " << TypeNameTraits<SC>::name () << endl
              << "LO: " << TypeNameTraits<LO>::name () << endl
              << "GO: " << TypeNameTraits<GO>::name () << endl
              << "NT: " << TypeNameTraits<NT>::name () << endl;
        }
        if (this->getObjectLabel () != "") {
          out << "Label: \"" << this->getObjectLabel () << "\"" << endl;
        }
        out << "Has Map: " << (this->map_.is_null () ? "false" : "true") << endl;
      } // if myRank == 0

      // Describe the Map, if it exists.
      if (! this->map_.is_null ()) {
        if (myRank == 0) {
          out << "Map:" << endl;
        }
        OSTab tab2 (out);
        this->map_->describe (out, vl);
      }

      // At verbosity > VERB_LOW, each process prints something.
      if (vl > VERB_LOW) {
        std::ostringstream os;
        os << "Process " << myRank << ":" << endl;
        //OSTab tab3 (os);

        const std::size_t numEnt = this->impl_.getLclNumEntries ();
        os << " Local number of entries: " << numEnt << endl;
        if (vl > VERB_MEDIUM) {
          os << " Local entries:" << endl;
          //OSTab tab4 (os);
          this->impl_.forAllEntries ([&os] (const GO row, const GO col, const SC& val) {
              os << "  {"
                 << "row: " << row
                 << ", col: " << col
                 << ", val: " << val
                 << "}" << endl;
            });
        }
        gathervPrint (out, os.str (), *comm);
      }
    } // vl != Teuchos::VERB_NONE
  }

private:
  Teuchos::RCP<const map_type>
  buildMap (const ::Teuchos::RCP<const ::Teuchos::Comm<int> >& comm)
  {
    using ::Teuchos::outArg;
    using ::Teuchos::rcp;
    using ::Teuchos::REDUCE_MIN;
    using ::Teuchos::reduceAll;
    typedef ::Tpetra::global_size_t GST;
    //const char prefix[] = "Tpetra::Details::CooMatrix::buildMap: ";

    // Processes where comm is null don't participate in the Map.
    if (comm.is_null ()) {
      return ::Teuchos::null;
    }

    // mfh 17 Jan 2017: We just happen to use row indices, because
    // that's what Tpetra::CrsMatrix currently uses.  That's probably
    // not the best thing to use, but it's not bad for commonly
    // encountered matrices.  A better more general approach might be
    // to hash (row index, column index) pairs to a global index.  One
    // could make that unique by doing a parallel scan at map
    // construction time.

    std::vector<GO> rowInds;
    const GO lclMinGblRowInd = this->impl_.getMyGlobalRowIndices (rowInds);

    // Compute the globally min row index for the "index base."
    GO gblMinGblRowInd = 0; // output argument
    reduceAll<int, GO> (*comm, REDUCE_MIN, lclMinGblRowInd,
                        outArg (gblMinGblRowInd));
    const GO indexBase = gblMinGblRowInd;
    const GST INV = Tpetra::Details::OrdinalTraits<GST>::invalid ();
    return rcp (new map_type (INV, rowInds.data (), rowInds.size (),
                              indexBase, comm));
  }

protected:
  virtual size_t constantNumberOfPackets () const {
    return static_cast<size_t> (0);
  }

  virtual bool
  checkSizes (const ::Tpetra::SrcDistObject& source)
  {
    using std::endl;
    typedef CooMatrix<SC, LO, GO, NT> this_type;
    const char prefix[] = "Tpetra::Details::CooMatrix::checkSizes: ";

    const this_type* src = dynamic_cast<const this_type* > (&source);

    if (src == NULL) {
      std::ostream& err = markLocalErrorAndGetStream ();
      err << prefix << "The source object of the Import or Export "
        "must be a CooMatrix with the same template parameters as the "
        "target object." << endl;
    }
    else if (this->map_.is_null ()) {
      std::ostream& err = markLocalErrorAndGetStream ();
      err << prefix << "The target object of the Import or Export "
        "must be a CooMatrix with a nonnull Map." << endl;
    }
    return ! (* (this->localError_));
  }

  virtual bool useNewInterface () { return true; }

  virtual void
  copyAndPermuteNew (const ::Tpetra::SrcDistObject& sourceObject,
                     const size_t numSameIDs,
                     const ::Kokkos::DualView<const LO*, device_type>& permuteToLIDs,
                     const ::Kokkos::DualView<const LO*, device_type>& permuteFromLIDs)
  {
    using std::endl;
    typedef CooMatrix<SC, LO, GO, NT> this_type;
    const char prefix[] = "Tpetra::Details::CooMatrix::copyAndPermuteNew: ";

    // There's no communication in this method, so it's safe just to
    // return on error.

    if (* (this->localError_)) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "The target object of the Import or Export is "
        "already in an error state." << endl;
      return;
    }

    const this_type* src = dynamic_cast<const this_type*> (&sourceObject);
    if (src == NULL) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "Input argument 'sourceObject' is not a CooMatrix."
          << endl;
      return;
    }

    const size_t numPermuteIDs =
      static_cast<size_t> (permuteToLIDs.dimension_0 ());
    if (numPermuteIDs != static_cast<size_t> (permuteFromLIDs.dimension_0 ())) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "permuteToLIDs.dimension_0() = "
          << numPermuteIDs << " != permuteFromLIDs.dimension_0() = "
          << permuteFromLIDs.dimension_0 () << "." << endl;
      return;
    }
    if (sizeof (int) <= sizeof (size_t) &&
        numPermuteIDs > static_cast<size_t> (std::numeric_limits<int>::max ())) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "numPermuteIDs = " << numPermuteIDs
          << ", a size_t, overflows int." << endl;
      return;
    }

    // TODO (mfh 19 Jan 2017) Check that permuteToLIDs and
    // permuteFromLIDs are sync'd to host.

    // Even though this is an std::set, we can start with numSameIDs
    // just by iterating through the first entries of the set.

    if (sizeof (int) <= sizeof (size_t) &&
        numSameIDs > static_cast<size_t> (std::numeric_limits<int>::max ())) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "numSameIDs = " << numSameIDs
          << ", a size_t, overflows int." << endl;
      return;
    }

    //
    // Copy in entries from any initial rows with the same global row indices.
    //
    const LO numSame = static_cast<int> (numSameIDs);
    // Count of local row indices encountered here with invalid global
    // row indices.  If nonzero, something went wrong.  If something
    // did go wrong, we'll defer responding until the end of this
    // method, so we can print as much useful info as possible.
    LO numInvalidSameRows = 0;
    for (LO lclRow = 0; lclRow < numSame; ++lclRow) {
      // All numSame initial rows have the same global index in both
      // source and target Maps, so we only need to convert to global
      // once.
      const GO gblRow = this->map_->getGlobalElement (lclRow);
      if (gblRow == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
        ++numInvalidSameRows;
        continue;
      }
      else {
        this->impl_.mergeIntoRow (gblRow, src->impl_, gblRow);
      }
    }

    //
    // Copy in entries from remaining rows that are permutations, that
    // is, that live in both the source and target Maps, but aren't
    // included in the "same" list (see above).
    //
    const LO numPermute = static_cast<int> (numPermuteIDs);
    // Count of local "from" row indices encountered here with invalid
    // global row indices.  If nonzero, something went wrong.  If
    // something did go wrong, we'll defer responding until the end of
    // this method, so we can print as much useful info as possible.
    LO numInvalidRowsFrom = 0;
    // Count of local "to" row indices encountered here with invalid
    // global row indices.  If nonzero, something went wrong.  If
    // something did go wrong, we'll defer responding until the end of
    // this method, so we can print as much useful info as possible.
    LO numInvalidRowsTo = 0;
    for (LO k = 0; k < numPermute; ++k) {
      const LO lclRowFrom = permuteFromLIDs.h_view(k);
      const LO lclRowTo = permuteToLIDs.h_view(k);
      const GO gblRowFrom = src->map_->getGlobalElement (lclRowFrom);
      const GO gblRowTo = this->map_->getGlobalElement (lclRowTo);

      bool bothConversionsValid = true;
      if (gblRowFrom == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
        ++numInvalidRowsFrom;
        bothConversionsValid = false;
      }
      if (gblRowTo == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
        ++numInvalidRowsTo;
        bothConversionsValid = false;
      }
      if (bothConversionsValid) {
        this->impl_.mergeIntoRow (gblRowTo, src->impl_, gblRowFrom);
      }
    }

    // Print info if any errors occurred.
    if (numInvalidSameRows != 0 || numInvalidRowsFrom != 0 || numInvalidRowsTo != 0) {
      // Collect and print all the invalid input row indices, for the
      // "same," "from," and "to" lists.
      std::vector<std::pair<LO, GO> > invalidSameRows;
      invalidSameRows.reserve (numInvalidSameRows);
      std::vector<std::pair<LO, GO> > invalidRowsFrom;
      invalidRowsFrom.reserve (numInvalidRowsFrom);
      std::vector<std::pair<LO, GO> > invalidRowsTo;
      invalidRowsTo.reserve (numInvalidRowsTo);

      for (LO lclRow = 0; lclRow < numSame; ++lclRow) {
        // All numSame initial rows have the same global index in both
        // source and target Maps, so we only need to convert to global
        // once.
        const GO gblRow = this->map_->getGlobalElement (lclRow);
        if (gblRow == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
          invalidSameRows.push_back ({lclRow, gblRow});
        }
      }

      for (LO k = 0; k < numPermute; ++k) {
        const LO lclRowFrom = permuteFromLIDs.h_view(k);
        const LO lclRowTo = permuteToLIDs.h_view(k);
        const GO gblRowFrom = src->map_->getGlobalElement (lclRowFrom);
        const GO gblRowTo = this->map_->getGlobalElement (lclRowTo);

        if (gblRowFrom == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
          invalidRowsFrom.push_back ({lclRowFrom, gblRowFrom});
        }
        if (gblRowTo == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
          invalidRowsTo.push_back ({lclRowTo, gblRowTo});
        }
      }

      std::ostringstream os;
      if (numInvalidSameRows != 0) {
        os << "Invalid permute \"same\" (local, global) index pairs: [";
        for (std::size_t k = 0; k < invalidSameRows.size (); ++k) {
          const auto& p = invalidSameRows[k];
          os << "(" << p.first << "," << p.second << ")";
          if (k + 1 < invalidSameRows.size ()) {
            os << ", ";
          }
        }
        os << "]" << std::endl;
      }
      if (numInvalidRowsFrom != 0) {
        os << "Invalid permute \"from\" (local, global) index pairs: [";
        for (std::size_t k = 0; k < invalidRowsFrom.size (); ++k) {
          const auto& p = invalidRowsFrom[k];
          os << "(" << p.first << "," << p.second << ")";
          if (k + 1 < invalidRowsFrom.size ()) {
            os << ", ";
          }
        }
        os << "]" << std::endl;
      }
      if (numInvalidRowsTo != 0) {
        os << "Invalid permute \"to\" (local, global) index pairs: [";
        for (std::size_t k = 0; k < invalidRowsTo.size (); ++k) {
          const auto& p = invalidRowsTo[k];
          os << "(" << p.first << "," << p.second << ")";
          if (k + 1 < invalidRowsTo.size ()) {
            os << ", ";
          }
        }
        os << "]" << std::endl;
      }

      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << os.str ();
      return;
    }
  }

  virtual void
  packAndPrepareNew (const ::Tpetra::SrcDistObject& sourceObject,
                     const ::Kokkos::DualView<const local_ordinal_type*, device_type>& exportLIDs,
                     ::Kokkos::DualView<packet_type*, device_type>& exports,
                     const ::Kokkos::DualView<size_t*, device_type>& numPacketsPerLID,
                     size_t& constantNumPackets,
                     ::Tpetra::Distributor& /* distor */)
  {
    using ::Teuchos::Comm;
    using ::Teuchos::RCP;
    using std::endl;
    typedef CooMatrix<SC, LO, GO, NT> this_type;
    typedef ::Kokkos::DualView<packet_type*, device_type> out_buf_dv_type;
    const char prefix[] = "Tpetra::Details::CooMatrix::packAndPrepareNew: ";
    const char suffix[] = "  This should never happen.  "
      "Please report this bug to the Tpetra developers.";

    // Tell the caller that different rows may have different numbers
    // of matrix entries.
    constantNumPackets = 0;

    const this_type* src = dynamic_cast<const this_type*> (&sourceObject);
    if (src == NULL) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "Input argument 'sourceObject' is not a CooMatrix."
          << endl;
    }
    else if (* (src->localError_)) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "The source (input) object of the Import or Export "
        "is already in an error state on this process."
          << endl;
    }
    else if (* (this->localError_)) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "The target (output, \"this\") object of the Import "
        "or Export is already in an error state on this process." << endl;
    }
    // Respond to detected error(s) by resizing 'exports' to zero (so
    // we won't be tempted to read it later), and filling
    // numPacketsPerLID with zeros.
    if (* (this->localError_)) {
      // Resize 'exports' to zero, so we won't be tempted to read it.
      exports = out_buf_dv_type ("CooMatrix exports", 0);
      // Trick to get around const DualView& being const.
      {
        using ::Kokkos::DualView;
        DualView<size_t*, device_type> numPacketsPerLID_tmp = numPacketsPerLID;
        numPacketsPerLID_tmp.template sync<Kokkos::HostSpace> ();
        numPacketsPerLID_tmp.template modify<Kokkos::HostSpace> ();
      }
      // Fill numPacketsPerLID with zeros.
      ::Kokkos::deep_copy (numPacketsPerLID.h_view, static_cast<size_t> (0));
      return;
    }

    // TODO (mfh 28 Jan 2017) pack source object's data, NOT *this's data!

    const size_t numExports = exportLIDs.dimension_0 ();
    if (numExports == 0) {
      exports = out_buf_dv_type (exports.h_view.label (), 0);
      return; // nothing to send
    }
    RCP<const Comm<int> > comm = src->getMap ().is_null () ?
      Teuchos::null : src->getMap ()->getComm ();
    if (comm.is_null () || comm->getSize () == 1) {
      if (numExports != static_cast<size_t> (0)) {
        std::ostream& err = this->markLocalErrorAndGetStream ();
        err << prefix << "The input communicator is either null or "
          "has only one process, but numExports = " << numExports << " != 0."
            << suffix << endl;
        return;
      }
      // Don't go into the rest of this method unless there are
      // actually processes other than the calling process.  This is
      // because the pack and unpack functions only have nonstub
      // implementations if building with MPI.
      return;
    }

    // Trick to get around const DualView& being const.
    {
      using ::Kokkos::DualView;
      DualView<size_t*, device_type> numPacketsPerLID_tmp = numPacketsPerLID;
      numPacketsPerLID_tmp.template sync<Kokkos::HostSpace> ();
      numPacketsPerLID_tmp.template modify<Kokkos::HostSpace> ();
    }

    int totalNumPackets = 0;
    size_t numInvalidRowInds = 0;
    std::ostringstream errStrm; // current loop iteration's error messages
    for (size_t k = 0; k < numExports; ++k) {
      const LO lclRow = exportLIDs.h_view[k];
      // We're packing the source object's data, so we need to use the
      // source object's Map to convert from local to global indices.
      const GO gblRow = src->map_->getGlobalElement (lclRow);
      if (gblRow == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
        // Mark the error later; just count for now.
        ++numInvalidRowInds;
        numPacketsPerLID.h_view[k] = 0;
        continue;
      }

      // Count the number of bytes needed to pack the current row of
      // the source object.
      int numPackets = 0;
      const int errCode =
        src->impl_.countPackRow (numPackets, gblRow, *comm, &errStrm);
      if (errCode != 0) {
        std::ostream& err = this->markLocalErrorAndGetStream ();
        err << prefix << errStrm.str () << endl;
        numPacketsPerLID.h_view[k] = 0;
        continue;
      }

      // Make sure that the total number of packets fits in int.
      // MPI requires this.
      const long long newTotalNumPackets =
        static_cast<long long> (totalNumPackets) +
        static_cast<long long> (numPackets);
      if (newTotalNumPackets >
          static_cast<long long> (std::numeric_limits<int>::max ())) {
        std::ostream& err = this->markLocalErrorAndGetStream ();
        err << prefix << "The new total number of packets "
            << newTotalNumPackets << " does not fit in int." << endl;
        // At this point, we definitely cannot continue.  In order to
        // leave the output arguments in a rational state, we zero out
        // all remaining entries of numPacketsPerLID before returning.
        for (size_t k2 = k; k2 < numExports; ++k2) {
          numPacketsPerLID.h_view[k2] = 0;
        }
        return;
      }
      numPacketsPerLID.h_view[k] = static_cast<size_t> (numPackets);
      totalNumPackets = static_cast<int> (newTotalNumPackets);
    }

    // If we found invalid row indices in exportLIDs, go back,
    // collect, and report them.
    if (numInvalidRowInds != 0) {
      std::vector<std::pair<LO, GO> > invalidRowInds;
      for (size_t k = 0; k < numExports; ++k) {
        const LO lclRow = exportLIDs.h_view[k];
        // We're packing the source object's data, so we need to use
        // the source object's Map to convert from local to global
        // indices.
        const GO gblRow = src->map_->getGlobalElement (lclRow);
        if (gblRow == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
          invalidRowInds.push_back ({lclRow, gblRow});
        }
      }
      std::ostringstream os;
      os << prefix << "We found " << numInvalidRowInds << " invalid row ind"
         << (numInvalidRowInds != static_cast<size_t> (1) ? "ices" : "ex")
         << " out of " << numExports << " in exportLIDs.  Here is the list "
         << " of invalid row indices: [";
      for (size_t k = 0; k < invalidRowInds.size (); ++k) {
        os << "(LID: " << invalidRowInds[k].first << ", GID: "
           << invalidRowInds[k].second << ")";
        if (k + 1 < invalidRowInds.size ()) {
          os << ", ";
        }
      }
      os << "].";

      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << os.str () << std::endl;
      return;
    }

    if (static_cast<int> (exports.dimension_0 ()) != totalNumPackets) {
      exports = out_buf_dv_type ("CooMatrix exports", totalNumPackets);
      exports.template sync<Kokkos::HostSpace> (); // make sure alloc'd on host
    }
    exports.template modify<Kokkos::HostSpace> ();

    // FIXME (mfh 17 Jan 2017) packTriples wants three arrays, not a
    // single array of structs.  For now, we just copy.
    std::vector<GO> gblRowInds;
    std::vector<GO> gblColInds;
    std::vector<SC> vals;

    int outBufCurPos = 0;
    packet_type* outBuf = exports.h_view.ptr_on_device ();
    for (size_t k = 0; k < numExports; ++k) {
      const LO lclRow = exportLIDs.h_view[k];
      // We're packing the source object's data, so we need to use the
      // source object's Map to convert from local to global indices.
      const GO gblRow = src->map_->getGlobalElement (lclRow);
      // Pack the current row of the source object.
      src->impl_.packRow (outBuf, totalNumPackets, outBufCurPos, *comm,
                          gblRowInds, gblColInds, vals, gblRow);
    }

    // Keep 'exports' modified on host.
    //exports.template sync<typename device_type::memory_space> ();
  }

  virtual void
  unpackAndCombineNew (const ::Kokkos::DualView<const local_ordinal_type*, device_type>& importLIDs,
                       const ::Kokkos::DualView<const packet_type*, device_type>& imports,
                       const ::Kokkos::DualView<const size_t*, device_type>& numPacketsPerLID,
                       const size_t /* constantNumPackets */, // we don't actually use this; we assume this is 0
                       ::Tpetra::Distributor& /* distor */,
                       const ::Tpetra::CombineMode /* CM */)
  {
    using ::Kokkos::DualView;
    using ::Teuchos::Comm;
    using ::Teuchos::RCP;
    using std::endl;
    //typedef ::Kokkos::DualView<const packet_type*, device_type> in_buf_dv_type;
    typedef typename device_type::memory_space DMS;
    typedef ::Kokkos::HostSpace HMS;
    const char prefix[] = "Tpetra::Details::CooMatrix::unpackAndCombineNew: ";
    const char suffix[] = "  This should never happen.  "
      "Please report this bug to the Tpetra developers.";

    const std::size_t numImports = importLIDs.dimension_0 ();
    if (numImports == 0) {
      return; // nothing to receive
    }
    else if (imports.dimension_0 () == 0) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      typename DualView<const LO*, device_type>::t_host importLIDs_h;
      {
        if (importLIDs.template need_sync<HMS> ()) {
          // Device version of the DualView is the most recently updated.
          // It's forbidden to sync a DualView<const T>, so we must copy.
          auto importLIDs_d = importLIDs.template view<DMS> ();
          typedef typename decltype (importLIDs_h)::non_const_type HVNC;
          HVNC importLIDs_h_nc (importLIDs_d.label (), importLIDs.dimension_0 ());
          Kokkos::deep_copy (importLIDs_h_nc, importLIDs_d);
          importLIDs_h = importLIDs_h_nc;
        }
        else { // host version of the DualView is up-to-date
          importLIDs_h = importLIDs.h_view; // importLIDs.template view<HMS> ();
        }
      }
      err << prefix << "importLIDs.dimension_0() = " << numImports << " != 0, "
          << "but imports.dimension_0() = 0.  This doesn't make sense, because "
          << "for every incoming LID, CooMatrix packs at least the count of "
          << "triples associated with that LID, even if the count is zero.  "
          << "importLIDs = [";
      for (std::size_t k = 0; k < numImports; ++k) {
        err << importLIDs.h_view[k];
        if (k + 1 < numImports) {
          err << ", ";
        }
      }
      err << "].  " << suffix << endl;
      return;
    }

    RCP<const Comm<int> > comm = this->getMap ().is_null () ?
      Teuchos::null : this->getMap ()->getComm ();
    if (comm.is_null () || comm->getSize () == 1) {
      if (numImports != static_cast<size_t> (0)) {
        std::ostream& err = this->markLocalErrorAndGetStream ();
        err << prefix << "The input communicator is either null or "
          "has only one process, but numImports = " << numImports << " != 0."
            << suffix << endl;
        return;
      }
      // Don't go into the rest of this method unless there are
      // actually processes other than the calling process.  This is
      // because the pack and unpack functions only have nonstub
      // implementations if building with MPI.
      return;
    }

    // Make sure that the length of 'imports' fits in int.
    // This is ultimately an MPI restriction.
    if (static_cast<size_t> (imports.dimension_0 ()) >
        static_cast<size_t> (std::numeric_limits<int>::max ())) {
      std::ostream& err = this->markLocalErrorAndGetStream ();
      err << prefix << "imports.dimension_0() = "
          << imports.dimension_0 () << " does not fit in int." << endl;
      return;
    }
    const int inBufSize = static_cast<int> (imports.dimension_0 ());

    // It's forbidden to sync a DualView<const T>, so if the input
    // DualView are not in sync on host, we must make copies.

    typename std::decay<decltype (importLIDs)>::type::t_host importLIDs_h;
    typename std::decay<decltype (imports)>::type::t_host imports_h;
    typename std::decay<decltype (numPacketsPerLID)>::type::t_host numPacketsPerLID_h;
    {
      if (importLIDs.template need_sync<HMS> ()) {
        // Device version of the DualView is the most recently updated.
        auto importLIDs_d = importLIDs.template view<DMS> ();
        typedef typename decltype (importLIDs_h)::non_const_type HVNC;
        HVNC importLIDs_h_nc (importLIDs_d.label (),
                              importLIDs.dimension_0 ());
        Kokkos::deep_copy (importLIDs_h_nc, importLIDs_d);
        importLIDs_h = importLIDs_h_nc;
      }
      else { // host version of the DualView is up-to-date
        importLIDs_h = importLIDs.h_view; // importLIDs.template view<HMS> ();
      }

      if (imports.template need_sync<HMS> ()) {
        // Device version of the DualView is the most recently updated.
        auto imports_d = imports.template view<DMS> ();
        typedef typename decltype (imports_h)::non_const_type HVNC;
        HVNC imports_h_nc (imports_d.label (),
                           imports.dimension_0 ());
        Kokkos::deep_copy (imports_h_nc, imports_d);
        imports_h = imports_h_nc;
      }
      else { // host version of the DualView is up-to-date
        imports_h = imports.h_view; // imports.template view<HMS> ();
      }

      if (numPacketsPerLID.template need_sync<HMS> ()) {
        // Device version of the DualView is the most recently updated.
        auto numPacketsPerLID_d = numPacketsPerLID.template view<DMS> ();
        typedef typename decltype (numPacketsPerLID_h)::non_const_type HVNC;
        HVNC numPacketsPerLID_h_nc (numPacketsPerLID_d.label (),
                                    numPacketsPerLID.dimension_0 ());
        Kokkos::deep_copy (numPacketsPerLID_h_nc, numPacketsPerLID_d);
        numPacketsPerLID_h = numPacketsPerLID_h_nc;
      }
      else { // host version of the DualView is up-to-date
        numPacketsPerLID_h = numPacketsPerLID.h_view; // numPacketsPerLID.template view<HMS> ();
      }
    }

    // FIXME (mfh 17,24 Jan 2017) packTriples wants three arrays, not a
    // single array of structs.  For now, we just copy.
    std::vector<GO> gblRowInds;
    std::vector<GO> gblColInds;
    std::vector<SC> vals;

    const packet_type* inBuf = imports_h.ptr_on_device ();
    int inBufCurPos = 0;
    size_t numInvalidRowInds = 0;
    int errCode = 0;
    std::ostringstream errStrm; // for unpack* error output.
    for (size_t k = 0; k < numImports; ++k) {
      const LO lclRow = importLIDs_h(k);
      const GO gblRow = this->map_->getGlobalElement (lclRow);
      if (gblRow == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
        ++numInvalidRowInds;
        continue;
      }

      // Remember where we were, so we don't overrun the buffer
      // length.  inBufCurPos is an in/out argument of unpackTriples*.
      const int origInBufCurPos = inBufCurPos;

      int numEnt = 0; // output argument of unpackTriplesCount
      errCode = unpackTriplesCount (inBuf, inBufSize, inBufCurPos,
                                    numEnt, *comm, &errStrm);
      if (errCode != 0 || numEnt < 0 || inBufCurPos < origInBufCurPos) {
        std::ostream& err = this->markLocalErrorAndGetStream ();

        err << prefix << "In unpack loop, k=" << k << ": ";
        if (errCode != 0) {
          err << "  unpackTriplesCount returned errCode = " << errCode
              << " != 0." << endl;
        }
        if (numEnt < 0) {
          err << "  unpackTriplesCount returned errCode = 0, but numEnt = "
              << numEnt << " < 0." << endl;
        }
        if (inBufCurPos < origInBufCurPos) {
          err << "  After unpackTriplesCount, inBufCurPos = " << inBufCurPos
              << " < origInBufCurPos = " << origInBufCurPos << "." << endl;
        }
        err << "  unpackTriplesCount report: " << errStrm.str () << endl;
        err << suffix << endl;

        // We only continue in a debug build, because the above error
        // messages could consume too much memory and cause an
        // out-of-memory error, without actually printing.  Printing
        // everything is useful in a debug build, but not so much in a
        // release build.
#ifdef HAVE_TPETRA_DEBUG
        // Clear out the current error stream, so we don't accumulate
        // over loop iterations.
        errStrm.str ("");
        continue;
#else
        return;
#endif // HAVE_TPETRA_DEBUG
      }

      // FIXME (mfh 17,24 Jan 2017) packTriples wants three arrays,
      // not a single array of structs.  For now, we just copy.
      gblRowInds.resize (numEnt);
      gblColInds.resize (numEnt);
      vals.resize (numEnt);

      errCode = unpackTriples (inBuf, inBufSize, inBufCurPos,
                               gblRowInds.data (), gblColInds.data (),
                               vals.data (), numEnt, *comm, &errStrm);
      if (errCode != 0) {
        std::ostream& err = this->markLocalErrorAndGetStream ();
        err << prefix << "unpackTriples returned errCode = "
            << errCode << " != 0.  It reports: " << errStrm.str ()
            << endl;
        // We only continue in a debug build, because the above error
        // messages could consume too much memory and cause an
        // out-of-memory error, without actually printing.  Printing
        // everything is useful in a debug build, but not so much in a
        // release build.
#ifdef HAVE_TPETRA_DEBUG
        // Clear out the current error stream, so we don't accumulate
        // over loop iterations.
        errStrm.str ("");
        continue;
#else
        return;
#endif // HAVE_TPETRA_DEBUG
      }
      this->sumIntoGlobalValues (gblRowInds.data (), gblColInds.data (),
                                 vals.data (), numEnt);
    }

    // If we found invalid row indices in exportLIDs, go back,
    // collect, and report them.
    if (numInvalidRowInds != 0) {
      // Mark the error now, before we possibly run out of memory.
      // The latter could raise an exception (e.g., std::bad_alloc),
      // but at least we would get the error state right.
      std::ostream& err = this->markLocalErrorAndGetStream ();

      std::vector<std::pair<LO, GO> > invalidRowInds;
      for (size_t k = 0; k < numImports; ++k) {
        const LO lclRow = importLIDs_h(k);
        const GO gblRow = this->map_->getGlobalElement (lclRow);
        if (gblRow == ::Tpetra::Details::OrdinalTraits<GO>::invalid ()) {
          invalidRowInds.push_back ({lclRow, gblRow});
        }
      }

      err << prefix << "We found " << numInvalidRowInds << " invalid row ind"
          << (numInvalidRowInds != static_cast<size_t> (1) ? "ices" : "ex")
          << " out of " << numImports << " in importLIDs.  Here is the list "
          << " of invalid row indices: [";
      for (size_t k = 0; k < invalidRowInds.size (); ++k) {
        err << "(LID: " << invalidRowInds[k].first << ", GID: "
            << invalidRowInds[k].second << ")";
        if (k + 1 < invalidRowInds.size ()) {
          err << ", ";
        }
      }
      err << "].";
      return;
    }
  }
};

} // namespace Details
} // namespace Tpetra

#endif // TPETRA_DETAILS_COOMATRIX_HPP