libtrilinos-zoltan2-dev 12.10.1-3 / usr / include / trilinos / Zoltan2_TpetraRowMatrixAdapter.hpp

// @HEADER
//
// ***********************************************************************
//
//   Zoltan2: A package of combinatorial algorithms for scientific computing
//                  Copyright 2012 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 Karen Devine      (kddevin@sandia.gov)
//                    Erik Boman        (egboman@sandia.gov)
//                    Siva Rajamanickam (srajama@sandia.gov)
//
// ***********************************************************************
//
// @HEADER

/*! \file Zoltan2_TpetraRowMatrixAdapter.hpp
    \brief Defines the TpetraRowMatrixAdapter class.
*/

#ifndef _ZOLTAN2_TPETRAROWMATRIXADAPTER_HPP_
#define _ZOLTAN2_TPETRAROWMATRIXADAPTER_HPP_

#include <Zoltan2_MatrixAdapter.hpp>
#include <Zoltan2_StridedData.hpp>
#include <Zoltan2_PartitioningHelpers.hpp>

#include <Tpetra_RowMatrix.hpp>

namespace Zoltan2 {

//////////////////////////////////////////////////////////////////////////////
/*!  \brief Provides access for Zoltan2 to Tpetra::RowMatrix data.

    The \c scalar_t type, representing user data such as matrix values, is
    used by Zoltan2 for weights, coordinates, part sizes and
    quality metrics.
    Some User types (like Tpetra::RowMatrix) have an inherent scalar type,
    and some
    (like Tpetra::RowGraph) do not.  For such objects, the scalar type is
    set by Zoltan2 to \c float.  If you wish to change it to double, set
    the second template parameter to \c double.

*/

template <typename User, typename UserCoord=User>
  class TpetraRowMatrixAdapter : public MatrixAdapter<User,UserCoord> {
public:

#ifndef DOXYGEN_SHOULD_SKIP_THIS
  typedef typename InputTraits<User>::scalar_t scalar_t;
  typedef typename InputTraits<User>::lno_t    lno_t;
  typedef typename InputTraits<User>::gno_t    gno_t;
  typedef typename InputTraits<User>::part_t   part_t;
  typedef typename InputTraits<User>::node_t   node_t;
  typedef User user_t;
  typedef UserCoord userCoord_t;
#endif

  /*! \brief Destructor
   */
  ~TpetraRowMatrixAdapter() { }

  /*! \brief Constructor   
   *    \param inmatrix The user's Tpetra RowMatrix object 
   *    \param nWeightsPerRow If row weights will be provided in setRowWeights(),
   *        the set \c nWeightsPerRow to the number of weights per row.
   */
  TpetraRowMatrixAdapter(const RCP<const User> &inmatrix,
                         int nWeightsPerRow=0);

  /*! \brief Specify a weight for each entity of the primaryEntityType.
   *    \param weightVal A pointer to the weights for this index.
   *    \stride          A stride to be used in reading the values.  The
   *        index \c idx weight for entity \k should be found at
   *        <tt>weightVal[k*stride]</tt>.
   *    \param idx  A value between zero and one less that the \c nWeightsPerRow 
   *                  argument to the constructor.
   *
   * The order of weights should correspond to the order of the primary 
   * entity type; see, e.g.,  setRowWeights below.
   */

  void setWeights(const scalar_t *weightVal, int stride, int idx = 0);

  /*! \brief Specify a weight for each row.
   *    \param weightVal A pointer to the weights for this index.
   *    \stride          A stride to be used in reading the values.  The
   *        index \c idx weight for row \k should be found at
   *        <tt>weightVal[k*stride]</tt>.
   *    \param idx  A value between zero and one less that the \c nWeightsPerRow 
   *                  argument to the constructor.
   *
   * The order of weights should correspond to the order of rows
   * returned by
   *   \code
   *       theMatrix->getRowMap()->getNodeElementList();
   *   \endcode
   */

  void setRowWeights(const scalar_t *weightVal, int stride, int idx = 0);

  /*! \brief Specify an index for which the weight should be
              the degree of the entity
   *    \param idx Zoltan2 will use the entity's 
   *         degree as the entity weight for index \c idx.
   */
  void setWeightIsDegree(int idx);

  /*! \brief Specify an index for which the row weight should be
              the global number of nonzeros in the row
   *    \param idx Zoltan2 will use the global number of nonzeros in a row
   *         as the row weight for index \c idx.
   */
  void setRowWeightIsNumberOfNonZeros(int idx);

  ////////////////////////////////////////////////////
  // The MatrixAdapter interface.
  ////////////////////////////////////////////////////

  size_t getLocalNumRows() const { 
    return matrix_->getNodeNumRows();
  }

  size_t getLocalNumColumns() const { 
    return matrix_->getNodeNumCols();
  }

  size_t getLocalNumEntries() const {
    return matrix_->getNodeNumEntries();
  }

  bool CRSViewAvailable() const { return true; }

  void getRowIDsView(const gno_t *&rowIds) const 
  {
    ArrayView<const gno_t> rowView = rowMap_->getNodeElementList();
    rowIds = rowView.getRawPtr();
  }

  void getCRSView(const lno_t *&offsets, const gno_t *&colIds) const
  {
    offsets = offset_.getRawPtr();
    colIds = columnIds_.getRawPtr();
  }

  void getCRSView(const lno_t *&offsets, const gno_t *&colIds,
                    const scalar_t *&values) const
  {
    offsets = offset_.getRawPtr();
    colIds = columnIds_.getRawPtr();
    values = values_.getRawPtr();
  }


  int getNumWeightsPerRow() const { return nWeightsPerRow_; }

  void getRowWeightsView(const scalar_t *&weights, int &stride,
                           int idx = 0) const
  {
    env_->localInputAssertion(__FILE__, __LINE__,
      "invalid weight index",
      idx >= 0 && idx < nWeightsPerRow_, BASIC_ASSERTION);
    size_t length;
    rowWeights_[idx].getStridedList(length, weights, stride);
  }

  bool useNumNonzerosAsRowWeight(int idx) const { return numNzWeight_[idx];}

  template <typename Adapter>
    void applyPartitioningSolution(const User &in, User *&out,
         const PartitioningSolution<Adapter> &solution) const;

  template <typename Adapter>
    void applyPartitioningSolution(const User &in, RCP<User> &out,
         const PartitioningSolution<Adapter> &solution) const;

private:

  RCP<Environment> env_;    // for error messages, etc.

  RCP<const User> matrix_;
  RCP<const Tpetra::Map<lno_t, gno_t, node_t> > rowMap_;
  RCP<const Tpetra::Map<lno_t, gno_t, node_t> > colMap_;
  lno_t base_;
  ArrayRCP<lno_t> offset_;
  ArrayRCP<gno_t> columnIds_;  // TODO:  KDD Is it necessary to copy and store
  ArrayRCP<scalar_t> values_;  // TODO:  the matrix here?  Would prefer views.

  int nWeightsPerRow_;
  ArrayRCP<StridedData<lno_t, scalar_t> > rowWeights_;
  ArrayRCP<bool> numNzWeight_;

  bool mayHaveDiagonalEntries;

  RCP<User> doMigration(const User &from, size_t numLocalRows,
                        const gno_t *myNewRows) const;
};

/////////////////////////////////////////////////////////////////
// Definitions
/////////////////////////////////////////////////////////////////

template <typename User, typename UserCoord>
  TpetraRowMatrixAdapter<User,UserCoord>::TpetraRowMatrixAdapter(
    const RCP<const User> &inmatrix, int nWeightsPerRow):
      env_(rcp(new Environment)),
      matrix_(inmatrix), rowMap_(), colMap_(), base_(),
      offset_(), columnIds_(),
      nWeightsPerRow_(nWeightsPerRow), rowWeights_(), numNzWeight_(),
      mayHaveDiagonalEntries(true)
{
  typedef StridedData<lno_t,scalar_t> input_t;

  rowMap_ = matrix_->getRowMap();
  colMap_ = matrix_->getColMap();
  base_ = rowMap_->getIndexBase();

  size_t nrows = matrix_->getNodeNumRows();
  size_t nnz = matrix_->getNodeNumEntries();
  size_t maxnumentries = 
         matrix_->getNodeMaxNumRowEntries(); // Diff from CrsMatrix
 
  offset_.resize(nrows+1, 0);
  columnIds_.resize(nnz);
  values_.resize(nnz);
  ArrayRCP<lno_t> indices(maxnumentries); // Diff from CrsMatrix
  ArrayRCP<scalar_t> nzs(maxnumentries);  // Diff from CrsMatrix
  lno_t next = 0;
  for (size_t i=0; i < nrows; i++){
    lno_t row = i + base_;
    matrix_->getLocalRowCopy(row, indices(), nzs(), nnz); // Diff from CrsMatrix
    for (size_t j=0; j < nnz; j++){
      values_[next] = nzs[j];
      // TODO - this will be slow
      //   Is it possible that global columns ids might be stored in order?
      columnIds_[next++] = colMap_->getGlobalElement(indices[j]);
    }
    offset_[i+1] = offset_[i] + nnz;
  } 

  if (nWeightsPerRow_ > 0){
    rowWeights_ = arcp(new input_t [nWeightsPerRow_], 0, nWeightsPerRow_, true);
    numNzWeight_ = arcp(new bool [nWeightsPerRow_], 0, nWeightsPerRow_, true);
    for (int i=0; i < nWeightsPerRow_; i++)
      numNzWeight_[i] = false;
  }
}

////////////////////////////////////////////////////////////////////////////
template <typename User, typename UserCoord>
  void TpetraRowMatrixAdapter<User,UserCoord>::setWeights(
    const scalar_t *weightVal, int stride, int idx)
{
  if (this->getPrimaryEntityType() == MATRIX_ROW)
    setRowWeights(weightVal, stride, idx);
  else {
    // TODO:  Need to allow weights for columns and/or nonzeros
    std::ostringstream emsg;
    emsg << __FILE__ << "," << __LINE__
         << " error:  setWeights not yet supported for"
         << " columns or nonzeros."
         << std::endl;
    throw std::runtime_error(emsg.str());
  }
}

////////////////////////////////////////////////////////////////////////////
template <typename User, typename UserCoord>
  void TpetraRowMatrixAdapter<User,UserCoord>::setRowWeights(
    const scalar_t *weightVal, int stride, int idx)
{
  typedef StridedData<lno_t,scalar_t> input_t;
  env_->localInputAssertion(__FILE__, __LINE__,
    "invalid row weight index",
    idx >= 0 && idx < nWeightsPerRow_, BASIC_ASSERTION);
  size_t nvtx = getLocalNumRows();
  ArrayRCP<const scalar_t> weightV(weightVal, 0, nvtx*stride, false);
  rowWeights_[idx] = input_t(weightV, stride);
}

////////////////////////////////////////////////////////////////////////////
template <typename User, typename UserCoord>
  void TpetraRowMatrixAdapter<User,UserCoord>::setWeightIsDegree(
    int idx)
{
  if (this->getPrimaryEntityType() == MATRIX_ROW)
    setRowWeightIsNumberOfNonZeros(idx);
  else {
    // TODO:  Need to allow weights for columns and/or nonzeros
    std::ostringstream emsg;
    emsg << __FILE__ << "," << __LINE__
         << " error:  setWeightIsNumberOfNonZeros not yet supported for"
         << " columns" << std::endl;
    throw std::runtime_error(emsg.str());
  }
}

////////////////////////////////////////////////////////////////////////////
template <typename User, typename UserCoord>
  void TpetraRowMatrixAdapter<User,UserCoord>::setRowWeightIsNumberOfNonZeros(
    int idx)
{
  env_->localInputAssertion(__FILE__, __LINE__,
    "invalid row weight index",
    idx >= 0 && idx < nWeightsPerRow_, BASIC_ASSERTION);

  numNzWeight_[idx] = true;
}

////////////////////////////////////////////////////////////////////////////
template <typename User, typename UserCoord>
  template <typename Adapter>
    void TpetraRowMatrixAdapter<User,UserCoord>::applyPartitioningSolution(
      const User &in, User *&out, 
      const PartitioningSolution<Adapter> &solution) const
{ 
  // Get an import list (rows to be received)
  size_t numNewRows;
  ArrayRCP<gno_t> importList;
  try{
    numNewRows = Zoltan2::getImportList<Adapter,
                                        TpetraRowMatrixAdapter<User,UserCoord> >
                                       (solution, this, importList);
  }
  Z2_FORWARD_EXCEPTIONS;

  // Move the rows, creating a new matrix.
  RCP<User> outPtr = doMigration(in, numNewRows, importList.getRawPtr());
  out = outPtr.get();
  outPtr.release();
}

////////////////////////////////////////////////////////////////////////////
template <typename User, typename UserCoord>
  template <typename Adapter>
    void TpetraRowMatrixAdapter<User,UserCoord>::applyPartitioningSolution(
      const User &in, RCP<User> &out, 
      const PartitioningSolution<Adapter> &solution) const
{ 
  // Get an import list (rows to be received)
  size_t numNewRows;
  ArrayRCP<gno_t> importList;
  try{
    numNewRows = Zoltan2::getImportList<Adapter,
                                        TpetraRowMatrixAdapter<User,UserCoord> >
                                       (solution, this, importList);
  }
  Z2_FORWARD_EXCEPTIONS;

  // Move the rows, creating a new matrix.
  out = doMigration(in, numNewRows, importList.getRawPtr());
}


////////////////////////////////////////////////////////////////////////////
template < typename User, typename UserCoord>
RCP<User> TpetraRowMatrixAdapter<User,UserCoord>::doMigration(
  const User &from,
  size_t numLocalRows, 
  const gno_t *myNewRows
) const
{
  typedef Tpetra::Map<lno_t, gno_t, node_t> map_t;
  typedef Tpetra::CrsMatrix<scalar_t, lno_t, gno_t, node_t> tcrsmatrix_t;

  // We cannot create a Tpetra::RowMatrix, unless the underlying type is 
  // something we know (like Tpetra::CrsMatrix).
  // If the underlying type is something different, the user probably doesn't 
  // want a Tpetra::CrsMatrix back, so we throw an error.

  // Try to cast "from" matrix to a TPetra::CrsMatrix
  // If that fails we throw an error.
  // We could cast as a ref which will throw std::bad_cast but with ptr
  // approach it might be clearer what's going on here
  const tcrsmatrix_t *pCrsMatrix = dynamic_cast<const tcrsmatrix_t *>(&from);

  if(!pCrsMatrix) {
    throw std::logic_error("TpetraRowMatrixAdapter cannot migrate data for "
                           "your RowMatrix; it can migrate data only for "
                           "Tpetra::CrsMatrix.  "
                           "You can inherit from TpetraRowMatrixAdapter and "
                           "implement migration for your RowMatrix.");
  }

  lno_t base = 0;

  // source map
  const RCP<const map_t> &smap = from.getRowMap();
  gno_t numGlobalRows = smap->getGlobalNumElements();

  // target map
  ArrayView<const gno_t> rowList(myNewRows, numLocalRows);
  const RCP<const Teuchos::Comm<int> > &comm = from.getComm();
  RCP<const map_t> tmap = rcp(new map_t(numGlobalRows, rowList, base, comm));

  // importer
  Tpetra::Import<lno_t, gno_t, node_t> importer(smap, tmap);

  // target matrix
  // Chris Siefert proposed using the following to make migration 
  // more efficient.  
  // By default, the Domain and Range maps are the same as in "from".
  // As in the original code, we instead set them both to tmap.  
  // The assumption is a square matrix.
  // TODO:  what about rectangular matrices?  
  // TODO:  Should choice of domain/range maps be an option to this function?

  // KDD 3/7/16:  disabling Chris' new code to avoid dashboard failures;
  // KDD 3/7/16:  can re-enable when issue #114 is fixed.
  // KDD 3/7/16:  when re-enable CSIEFERT code, can comment out
  // KDD 3/7/16:  "Original way" code.
  // CSIEFERT RCP<tcrsmatrix_t> M;
  // CSIEFERT from.importAndFillComplete(M, importer, tmap, tmap);

  // Original way we did it:
  //
  int oldNumElts = smap->getNodeNumElements();
  int newNumElts = numLocalRows;

  // number of non zeros in my new rows
  typedef Tpetra::Vector<scalar_t, lno_t, gno_t, node_t> vector_t;
  vector_t numOld(smap);  // TODO These vectors should have scalar=size_t,
  vector_t numNew(tmap);  // but ETI does not yet support that.
  for (int lid=0; lid < oldNumElts; lid++){
    numOld.replaceGlobalValue(smap->getGlobalElement(lid),
      scalar_t(from.getNumEntriesInLocalRow(lid)));
  }
  numNew.doImport(numOld, importer, Tpetra::INSERT);

  // TODO Could skip this copy if could declare vector with scalar=size_t.
  ArrayRCP<size_t> nnz(newNumElts);
  if (newNumElts > 0){
    ArrayRCP<scalar_t> ptr = numNew.getDataNonConst(0);
    for (int lid=0; lid < newNumElts; lid++){
      nnz[lid] = static_cast<size_t>(ptr[lid]);
    }
  }

  RCP<tcrsmatrix_t> M = rcp(new tcrsmatrix_t(tmap, nnz,
                                               Tpetra::StaticProfile));
  M->doImport(from, importer, Tpetra::INSERT);
  M->fillComplete();

  // End of original way we did it.
  return Teuchos::rcp_dynamic_cast<User>(M);
}

}  //namespace Zoltan2
  
#endif