libtrilinos-tpetra-dev 12.12.1-5 / usr / include / trilinos / Tsqr_CacheBlocker.hpp

//@HEADER
// ************************************************************************
//
//          Kokkos: Node API and Parallel Node Kernels
//              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 __TSQR_CacheBlocker_hpp
#define __TSQR_CacheBlocker_hpp

#include <Tsqr_CacheBlockingStrategy.hpp>
#include <Tsqr_MatView.hpp>
#include <Tsqr_Util.hpp>

#include <iterator>
#include <sstream>
#include <stdexcept>

namespace TSQR {

  /// \class CacheBlocker
  /// \brief Break a tall skinny matrix by rows into cache blocks.
  /// \author Mark Hoemmen
  ///
  /// A CacheBlocker uses a particular cache blocking strategy to
  /// partition an nrows by ncols matrix by rows into cache blocks.
  /// The entries in a cache block may be stored contiguously, or as
  /// non-contiguous partitions of a matrix stored conventionally (in
  /// column-major order).
  ///
  /// The CacheBlocker blocks any matrix with the same number of rows
  /// in the same way, regardless of the number of columns (the cache
  /// blocking strategy's number of columns is set on construction).
  /// This is useful for TSQR's apply() routine, which requires that
  /// the output matrix C be blocked in the same way as the input
  /// matrix Q (in which the Q factor is stored implicitly).
  template<class Ordinal, class Scalar>
  class CacheBlocker {
  private:
    typedef MatView<Ordinal, Scalar> mat_view_type;
    typedef ConstMatView<Ordinal, Scalar> const_mat_view_type;

    void
    validate ()
    {
      if (nrows_cache_block_ < ncols_)
        {
          std::ostringstream os;
          os << "The typical cache block size is too small.  Only "
             << nrows_cache_block_ << " rows fit, but every cache block needs "
            "at least as many rows as the number of columns " << ncols_
             << " in the matrix.";
          throw std::logic_error (os.str());
        }
    }

  public:
    /// \brief Constructor
    ///
    /// \param num_rows Number of rows in the matrix to block.
    /// \param num_cols Number of columns in the matrix to block.
    /// \param strategy Cache blocking strategy object (passed by copy).
    ///
    /// \note The CacheBlocker's number of columns may differ from the
    ///   number of columns associated with the cache blocking
    ///   strategy.  The strategy uses a fixed number of columns for
    ///   all matrices with the same number of rows, so that it blocks
    ///   all such matrices in the same way (at the same row indices).
    ///   This is useful for TSQR's apply() and explicit_Q() methods.
    CacheBlocker (const Ordinal num_rows,
                  const Ordinal num_cols,
                  const CacheBlockingStrategy<Ordinal, Scalar>& strategy) :
      nrows_ (num_rows),
      ncols_ (num_cols),
      strategy_ (strategy),
      nrows_cache_block_ (strategy_.cache_block_num_rows (ncols()))
    {
      validate ();
    }

    //! Default constructor, so that CacheBlocker is DefaultConstructible.
    CacheBlocker () :
      nrows_ (0),
      ncols_ (0),
      nrows_cache_block_ (strategy_.cache_block_num_rows (ncols()))
    {}

    //! Copy constructor
    CacheBlocker (const CacheBlocker& rhs) :
      nrows_ (rhs.nrows()),
      ncols_ (rhs.ncols()),
      strategy_ (rhs.strategy_),
      nrows_cache_block_ (rhs.nrows_cache_block_)
    {}

    //! Assignment operator
    CacheBlocker& operator= (const CacheBlocker& rhs) {
      nrows_ = rhs.nrows();
      ncols_ = rhs.ncols();
      strategy_ = rhs.strategy_;
      nrows_cache_block_ = rhs.nrows_cache_block_;
      return *this;
    }

    //! Cache size hint (in bytes).
    size_t cache_size_hint () const { return strategy_.cache_size_hint(); }

    //! Number of rows in the matrix to block.
    Ordinal nrows () const { return nrows_; }

    //! Number of columns in the matrix to block.
    Ordinal ncols () const { return ncols_; }

    /// \brief Split A in place into [A_top; A_rest].
    ///
    /// Return the topmost cache block A_top of A, and modify A in
    /// place to be the "rest" of the matrix A_rest.
    ///
    /// \param A [in/out] On input: view of the matrix to split.
    ///   On output: the "rest" of the matrix.  If there is only
    ///   one cache block, A_top contains all of the matrix and
    ///   A is empty on output.
    /// \param contiguous_cache_blocks [in] Whether cache blocks in
    ///   the matrix A are stored contiguously (default is false).
    ///
    /// \return View of the topmost cache block A_top.
    ///
    /// \note The number of rows in A_top depends on the number of
    ///   columns with which this CacheBlocker was set up (rather than
    ///   the number of columns in A, which may not be the same).  The
    ///   idea is to have the number and distribution of rows in the
    ///   cache blocks be the same as the original nrows() by ncols()
    ///   matrix with which this CacheBlocker was initialized.
    template< class MatrixViewType >
    MatrixViewType
    split_top_block (MatrixViewType& A, const bool contiguous_cache_blocks) const
    {
      typedef typename MatrixViewType::ordinal_type ordinal_type;
      const ordinal_type nrows_top =
        strategy_.top_block_split_nrows (A.nrows(), ncols(),
                                         nrows_cache_block());
      // split_top() sets A to A_rest, and returns A_top.
      return A.split_top (nrows_top, contiguous_cache_blocks);
    }

    /// \brief View of the topmost cache block of A.
    ///
    /// The matrix view A is copied so the view itself won't be modified.
    ///
    /// \param A [in] View of the matrix to block.
    /// \param contiguous_cache_blocks [in] Whether cache blocks in
    ///   the matrix A are stored contiguously (default is false).
    ///
    /// \return View of the topmost cache block of A.
    template< class MatrixViewType >
    MatrixViewType
    top_block (const MatrixViewType& A, const bool contiguous_cache_blocks) const
    {
      typedef typename MatrixViewType::ordinal_type ordinal_type;
      // Ignore the number of columns in A, since we want to block all
      // matrices using the same cache blocking strategy.
      const ordinal_type nrows_top =
        strategy_.top_block_split_nrows (A.nrows(), ncols(),
                                         nrows_cache_block());
      MatrixViewType A_copy (A);
      return A_copy.split_top (nrows_top, contiguous_cache_blocks);
    }

    /// \brief Split A in place into [A_rest; A_bot].
    ///
    /// Return the bottommost cache block A_bot of A, and modify A in
    /// place to be the "rest" of the matrix A_rest.
    ///
    /// \param A [in/out] On input: view of the matrix to split.  On
    ///   output: the "rest" of the matrix.  If there is only one
    ///   cache block, A_bot contains all of the matrix and A is empty
    ///   on output.
    /// \param contiguous_cache_blocks [in] Whether cache blocks in
    ///   the matrix A are stored contiguously (default is false).
    ///
    /// \return View of the bottommost cache block A_bot.
    ///
    template< class MatrixViewType >
    MatrixViewType
    split_bottom_block (MatrixViewType& A, const bool contiguous_cache_blocks) const
    {
      typedef typename MatrixViewType::ordinal_type ordinal_type;
      // Ignore the number of columns in A, since we want to block all
      // matrices using the same cache blocking strategy.
      const ordinal_type nrows_bottom =
        strategy_.bottom_block_split_nrows (A.nrows(), ncols(),
                                            nrows_cache_block());
      // split_bottom() sets A to A_rest, and returns A_bot.
      return A.split_bottom (nrows_bottom, contiguous_cache_blocks);
    }

    /// \brief Fill the matrix A with zeros, respecting cache blocks.
    ///
    /// A specialization of this method for a particular
    /// MatrixViewType will only compile if MatrixViewType has a
    /// method "fill(const Scalar)" or "fill(const Scalar&)".  The
    /// intention is that the method be non-const and that it fill in
    /// the entries of the matrix with Scalar(0).
    ///
    /// \param A [in/out] View of the matrix to fill with zeros.
    ///
    /// \param contiguous_cache_blocks [in] Whether the cache blocks
    ///   in A are stored contiguously.
    ///
    template<class MatrixViewType>
    void
    fill_with_zeros (MatrixViewType A,
                     const bool contiguous_cache_blocks) const
    {
      // Note: if the cache blocks are stored contiguously, A.lda()
      // won't be the correct leading dimension of A, but it won't
      // matter: we only ever operate on A_cur here, and A_cur's
      // leading dimension is set correctly by split_top_block().
      while (! A.empty())
        {
          // This call modifies the matrix view A, but that's OK since
          // we passed the input view by copy, not by reference.
          MatrixViewType A_cur = split_top_block (A, contiguous_cache_blocks);
          A_cur.fill (Scalar(0));
        }
    }

    /// \brief Fill the matrix A with zeros, respecting cache blocks.
    ///
    /// This version of the method takes a raw pointer and matrix
    /// dimensions, rather than a matrix view object.  If
    /// contiguous_cache_blocks==false, the matrix is stored either in
    /// column-major order with leading dimension lda; else, the
    /// matrix is stored in cache blocks, with each cache block's
    /// entries stored contiguously in column-major order.
    ///
    /// \param num_rows [in] Number of rows in the matrix A.
    /// \param num_cols [in] Number of columns in the matrix A.
    /// \param A [out] The matrix to fill with zeros.
    /// \param lda [in] Leading dimension (a.k.a. stride) of the
    ///   matrix A.
    /// \param contiguous_cache_blocks [in] Whether the cache blocks
    ///   in A are stored contiguously.
    void
    fill_with_zeros (const Ordinal num_rows,
                     const Ordinal num_cols,
                     Scalar A[],
                     const Ordinal lda,
                     const bool contiguous_cache_blocks) const
    {
      // We say "A_rest" because it points to the remaining part of
      // the matrix left to process; at the beginning, the "remaining"
      // part is the whole matrix, but that will change as the
      // algorithm progresses.
      //
      // Note: if the cache blocks are stored contiguously, lda won't
      // be the correct leading dimension of A, but it won't matter:
      // we only ever operate on A_cur here, and A_cur's leading
      // dimension is set correctly by A_rest.split_top().
      mat_view_type A_rest (num_rows, num_cols, A, lda);

      while (! A_rest.empty())
        {
          // This call modifies A_rest.
          mat_view_type A_cur = split_top_block (A_rest, contiguous_cache_blocks);
          A_cur.fill (Scalar(0));
        }
    }

    /// \brief Cache-block the given A_in matrix into A_out.
    ///
    /// Given an nrows by ncols (with nrows >= ncols) matrix A_in,
    /// stored in column-major order with leading dimension lda_in (>=
    /// nrows), copy it into A_out in a cache-blocked row block
    /// format.  Each cache block is a matrix in column-major order,
    /// and the elements of a cache block are stored consecutively in
    /// A_out.  The number of rows in each cache block depends on the
    /// cache-blocking strategy that this CacheBlocker uses.
    ///
    /// \param num_rows [in] Total number of rows in the matrices A_in and A_out
    /// \param num_cols [in] Number of columns in the matrices A_in and A_out
    /// \param A_out [out] nrows*ncols contiguous storage into which to write
    ///   the cache-blocked output matrix.
    /// \param A_in [in] nrows by ncols matrix, stored in column-major
    ///   order with leading dimension lda_in >= nrows
    /// \param lda_in [in] Leading dimension of the matrix A_in
    void
    cache_block (const Ordinal num_rows,
                 const Ordinal num_cols,
                 Scalar A_out[],
                 const Scalar A_in[],
                 const Ordinal lda_in) const
    {
      // We say "*_rest" because it points to the remaining part of
      // the matrix left to cache block; at the beginning, the
      // "remaining" part is the whole matrix, but that will change as
      // the algorithm progresses.
      const_mat_view_type A_in_rest (num_rows, num_cols, A_in, lda_in);
      // Leading dimension doesn't matter since A_out will be cache blocked.
      mat_view_type A_out_rest (num_rows, num_cols, A_out, lda_in);

      while (! A_in_rest.empty())
        {
          if (A_out_rest.empty())
            throw std::logic_error("A_out_rest is empty, but A_in_rest is not");

          // This call modifies A_in_rest.
          const_mat_view_type A_in_cur = split_top_block (A_in_rest, false);

          // This call modifies A_out_rest.
          mat_view_type A_out_cur = split_top_block (A_out_rest, true);

          copy_matrix (A_in_cur.nrows(), num_cols, A_out_cur.get(),
                       A_out_cur.lda(), A_in_cur.get(), A_in_cur.lda());
        }
    }

    //! "Un"-cache-block the given A_in matrix into A_out.
    void
    un_cache_block (const Ordinal num_rows,
                    const Ordinal num_cols,
                    Scalar A_out[],
                    const Ordinal lda_out,
                    const Scalar A_in[]) const
    {
      // We say "*_rest" because it points to the remaining part of
      // the matrix left to cache block; at the beginning, the
      // "remaining" part is the whole matrix, but that will change as
      // the algorithm progresses.
      //
      // Leading dimension doesn't matter since A_in is cache blocked.
      const_mat_view_type A_in_rest (num_rows, num_cols, A_in, lda_out);
      mat_view_type A_out_rest (num_rows, num_cols, A_out, lda_out);

      while (! A_in_rest.empty())
        {
          if (A_out_rest.empty())
            throw std::logic_error("A_out_rest is empty, but A_in_rest is not");

          // This call modifies A_in_rest.
          const_mat_view_type A_in_cur = split_top_block (A_in_rest, true);

          // This call modifies A_out_rest.
          mat_view_type A_out_cur = split_top_block (A_out_rest, false);

          copy_matrix (A_in_cur.nrows(), num_cols, A_out_cur.get(),
                       A_out_cur.lda(), A_in_cur.get(), A_in_cur.lda());
        }
    }

    /// \brief Return the cache block with index \c cache_block_index.
    ///
    /// \param A [in] The original matrix.
    /// \param cache_block_index [in] Zero-based index of the cache block.
    ///   If the index is out of bounds, silently return an empty matrix
    ///   view.
    /// \param contiguous_cache_blocks [in] Whether cache blocks are
    ///   stored contiguously.
    ///
    /// \return Cache block of A with the given index, or an empty
    ///   matrix view if the index is out of bounds.
    ///
    /// \note This method is templated on MatrixViewType, so that it
    ///   works with any matrix view type.
    template<class MatrixViewType>
    MatrixViewType
    get_cache_block (MatrixViewType A,
                     const typename MatrixViewType::ordinal_type cache_block_index,
                     const bool contiguous_cache_blocks) const
    {
      typedef typename MatrixViewType::ordinal_type ordinal_type;

      // Total number of cache blocks.
      const ordinal_type num_cache_blocks =
        strategy_.num_cache_blocks (A.nrows(), A.ncols(), nrows_cache_block());

      if (cache_block_index >= num_cache_blocks)
        return MatrixViewType (0, 0, NULL, 0); // empty

      // result[0] = starting row index of the cache block
      // result[1] = number of rows in the cache block
      // result[2] = pointer offset (A.get() + result[2])
      // result[3] = leading dimension (a.k.a. stride) of the cache block
      std::vector<Ordinal> result =
        strategy_.cache_block_details (cache_block_index, A.nrows(), A.ncols(),
                                       A.lda(), nrows_cache_block(),
                                       contiguous_cache_blocks);
      if (result[1] == 0)
        // For some reason, the cache block is empty.
        return MatrixViewType (0, 0, NULL, 0);

      // We expect that ordinal_type is signed, so adding signed
      // (ordinal_type) to unsigned (pointer) may raise compiler
      // warnings.
      return MatrixViewType (result[1], A.ncols(),
                             A.get() + static_cast<size_t>(result[2]),
                             result[3]);
    }

    /// \brief Equality operator.
    ///
    /// Two cache blockers are "equal" if they correspond to matrices
    /// with the same dimensions (number of rows and number of
    /// columns), and if their cache blocking strategies are equal.
    bool
    operator== (const CacheBlockingStrategy<Ordinal, Scalar>& rhs) const
    {
      return nrows() == rhs.nrows() &&
        ncols() == rhs.ncols() &&
        strategy_ == rhs.strategy_;
    }

  private:
    //! Number of rows in the matrix to block.
    Ordinal nrows_;

    //! Number of columns in the matrix to block.
    Ordinal ncols_;

    //! Strategy used to break the matrix into cache blocks.
    CacheBlockingStrategy<Ordinal, Scalar> strategy_;

    /// \brief Number of rows in a "typical" cache block.
    ///
    /// We could instead use the strategy object to recompute this
    /// quantity each time, but we choose to cache the computed value
    /// here.  For an explanation of "typical," see the documentation
    /// of \c nrows_cache_block().
    Ordinal nrows_cache_block_;

    /// \brief Number of rows in a "typical" cache block.
    ///
    /// For an explanation of "typical," see the documentation of
    /// CacheBlockingStrategy.  In brief, some cache blocks may have
    /// more rows (up to but not including nrows_cache_block() +
    /// ncols() rows), and some may have less (but no less than
    /// ncols() rows).
    size_t nrows_cache_block () const { return nrows_cache_block_; }
  };


  /// \class CacheBlockRangeIterator
  /// \brief Bidirectional iterator over a contiguous range of cache blocks.
  /// \author Mark Hoemmen
  ///
  /// "Contiguous range of cache blocks" means that the indices of the
  /// cache blocks, as interpreted by the CacheBlocker object, are
  /// contiguous.
  template<class MatrixViewType>
  class CacheBlockRangeIterator :
    public std::iterator<std::forward_iterator_tag, MatrixViewType>
  {
  public:
    typedef MatrixViewType view_type;
    typedef typename MatrixViewType::ordinal_type ordinal_type;
    typedef typename MatrixViewType::scalar_type scalar_type;

    /// \brief Default constructor.
    ///
    /// \note To implementers: We only implement a default constructor
    ///   because all iterators (e.g., TrivialIterator) must be
    ///   DefaultConstructible.
    CacheBlockRangeIterator () :
      A_ (0, 0, NULL, 0),
      curInd_ (0),
      reverse_ (false),
      contiguousCacheBlocks_ (false)
    {}

    /// \brief Standard constructor.
    ///
    /// \param A [in] View of the matrix over whose cache block(s) to
    ///   iterate.
    /// \param strategy [in] Cache blocking strategy for a matrix with
    ///   the same number of rows as the matrix A.
    /// \param currentIndex [in] The iterator's current cache block index.
    /// \param reverse [in] Whether to iterate over the cache blocks
    ///   in reverse order of their indices.
    /// \param contiguousCacheBlocks [in] Whether cache blocks in the
    ///   matrix A are stored contiguously.
    CacheBlockRangeIterator (const MatrixViewType& A,
                             const CacheBlockingStrategy<ordinal_type, scalar_type>& strategy,
                             const ordinal_type currentIndex,
                             const bool reverse,
                             const bool contiguousCacheBlocks) :
      A_ (A),
      blocker_ (A_.nrows(), A_.ncols(), strategy),
      curInd_ (currentIndex),
      reverse_ (reverse),
      contiguousCacheBlocks_ (contiguousCacheBlocks)
    {}

    //! Copy constructor.
    CacheBlockRangeIterator (const CacheBlockRangeIterator& rhs) :
      A_ (rhs.A_),
      blocker_ (rhs.blocker_),
      curInd_ (rhs.curInd_),
      reverse_ (rhs.reverse_),
      contiguousCacheBlocks_ (rhs.contiguousCacheBlocks_)
    {}

    //! Assignment operator.
    CacheBlockRangeIterator& operator= (const CacheBlockRangeIterator& rhs)
    {
      A_ = rhs.A_;
      blocker_ = rhs.blocker_;
      curInd_ = rhs.curInd_;
      reverse_ = rhs.reverse_;
      contiguousCacheBlocks_ = rhs.contiguousCacheBlocks_;
      return *this;
    }

    //! Prefix increment operator.
    CacheBlockRangeIterator& operator++() {
      if (reverse_)
        --curInd_;
      else
        ++curInd_;
      return *this;
    }

    /// \brief Postfix increment operator.
    ///
    /// This may be less efficient than prefix operator++, since the
    /// postfix operator has to make a copy of the iterator before
    /// modifying it.
    CacheBlockRangeIterator operator++(int) {
      CacheBlockRangeIterator retval (*this);
      operator++();
      return retval;
    }

    /// \brief Equality operator.
    ///
    /// Equality of cache block range iterators only tests the cache
    /// block index, not reverse-ness.  This means we can compare a
    /// reverse-direction iterator with a forward-direction iterator,
    /// and vice versa.
    bool operator== (const CacheBlockRangeIterator& rhs) {
      // Not correct, but fast.  Should return false for different A_
      // or different blocker_.
      return curInd_ == rhs.curInd_;
    }

    //! Inequality operator.
    bool operator!= (const CacheBlockRangeIterator& rhs) {
      // Not correct, but fast.  Should return false for different A_
      // or different blocker_.
      return curInd_ != rhs.curInd_;
    }

    /// \brief A view of the current cache block.
    ///
    /// If the current cache block index is invalid, this returns an
    /// empty cache block (that is, calling empty() on the returned
    /// view returns true).
    MatrixViewType operator*() const {
      return blocker_.get_cache_block (A_, curInd_, contiguousCacheBlocks_);
    }

  private:
    MatrixViewType A_;
    CacheBlocker<ordinal_type, scalar_type> blocker_;
    ordinal_type curInd_;
    bool reverse_;
    bool contiguousCacheBlocks_;
  };

  /// \class CacheBlockRange
  /// \brief Collection of cache blocks with a contiguous range of indices.
  /// \author Mark Hoemmen
  ///
  /// We mean "collection" in the C++ sense: you can iterate over the
  /// elements using iterators.  The iterators are valid only when the
  /// CacheBlockRange is in scope, just like the iterators of
  /// std::vector.
  ///
  /// CacheBlockRange is useful for \c KokkosNodeTsqr, in particular
  /// for \c FactorFirstPass and \c ApplyFirstPass.  Sequential TSQR's
  /// factorization is forward iteration over the collection, and
  /// applying the Q factor or computing the explicit Q factor is
  /// iteration in the reverse direction (decreasing cache block
  /// index).
  ///
  /// This class is templated so that it works with any matrix view.
  template<class MatrixViewType>
  class CacheBlockRange {
  public:
    typedef MatrixViewType view_type;
    typedef typename MatrixViewType::ordinal_type ordinal_type;
    typedef typename MatrixViewType::scalar_type scalar_type;

    /// \typedef iterator
    /// \brief Type of an iterator over the range of cache blocks.
    typedef CacheBlockRangeIterator<MatrixViewType> iterator;

    /// \brief Constructor
    ///
    /// \param A [in] View of the matrix to factor.
    /// \param strategy [in] Cache blocking strategy to use (copied
    ///   on input).
    /// \param startIndex [in] Starting index of the cache block
    ///   sequence.
    /// \param endIndex [in] Ending index (exclusive) of the cache
    ///   block sequence.  Precondition: startIndex <= endIndex.  If
    ///   startIndex == endIndex, the sequence is empty.
    /// \param contiguousCacheBlocks [in] Whether cache blocks in the
    ///   matrix A are stored contiguously.
    CacheBlockRange (MatrixViewType A,
                     const CacheBlockingStrategy<ordinal_type, scalar_type>& strategy,
                     const ordinal_type startIndex,
                     const ordinal_type endIndex,
                     const bool contiguousCacheBlocks) :
      A_ (A),
      startIndex_ (startIndex),
      endIndex_ (endIndex),
      strategy_ (strategy),
      contiguousCacheBlocks_ (contiguousCacheBlocks)
    {}

    bool empty() const {
      return startIndex_ >= endIndex_;
    }

    iterator begin() const {
      return iterator (A_, strategy_, startIndex_, false, contiguousCacheBlocks_);
    }

    iterator end() const {
      return iterator (A_, strategy_, endIndex_, false, contiguousCacheBlocks_);
    }

    iterator rbegin() const {
      return iterator (A_, strategy_, endIndex_-1, true, contiguousCacheBlocks_);
    }

    iterator rend() const {
      // Think about it: rbegin() == rend() means that rbegin() is invalid
      // and shouldn't be dereferenced.  rend() should never be dereferenced.
      return iterator (A_, strategy_, startIndex_-1, true, contiguousCacheBlocks_);
    }

    private:
      //! View of the matrix.
      MatrixViewType A_;

      /// \brief Starting index of the range of cache blocks.
      ///
      /// We always have startIndex_ <= endIndex_.  Reverse-order
      /// iteration is indicated by the iterator's reverse_ member
      /// datum.
      ordinal_type startIndex_;

      /// \brief Ending index (exclusive) of the range of cache blocks.
      ///
      /// See the documentation of startIndex_ for its invariant.
      ordinal_type endIndex_;

      //! Cache blocking strategy for a matrix with the same number of rows as A_.
      CacheBlockingStrategy<ordinal_type, scalar_type> strategy_;

      //! Whether the cache blocks of the matrix A_ are stored contiguously.
      bool contiguousCacheBlocks_;
    };


} // namespace TSQR


#endif // __TSQR_CacheBlocker_hpp