chunk_sparsity_pattern.cc

// ---------------------------------------------------------------------
//
// Copyright (C) 2008 - 2019 by the deal.II authors
//
// This file is part of the deal.II library.
//
// The deal.II library is free software; you can use it, redistribute
// it, and/or modify it under the terms of the GNU Lesser General
// Public License as published by the Free Software Foundation; either
// version 2.1 of the License, or (at your option) any later version.
// The full text of the license can be found in the file LICENSE.md at
// the top level directory of deal.II.
//
// ---------------------------------------------------------------------


#include <deal.II/lac/chunk_sparsity_pattern.h>
#include <deal.II/lac/dynamic_sparsity_pattern.h>
#include <deal.II/lac/full_matrix.h>


DEAL_II_NAMESPACE_OPEN


ChunkSparsityPattern::ChunkSparsityPattern()
{
  reinit(0, 0, 0, 1);
}



ChunkSparsityPattern::ChunkSparsityPattern(const ChunkSparsityPattern &s)
  : Subscriptor()
  , chunk_size(s.chunk_size)
  , sparsity_pattern(s.sparsity_pattern)
{
  Assert(s.rows == 0 && s.cols == 0,
         ExcMessage(
           "This constructor can only be called if the provided argument "
           "is the sparsity pattern for an empty matrix. This constructor can "
           "not be used to copy-construct a non-empty sparsity pattern."));

  reinit(0, 0, 0, chunk_size);
}



ChunkSparsityPattern::ChunkSparsityPattern(const size_type m,
                                           const size_type n,
                                           const size_type max_per_row,
                                           const size_type chunk_size)
{
  Assert(chunk_size > 0, ExcInvalidNumber(chunk_size));

  reinit(m, n, max_per_row, chunk_size);
}



ChunkSparsityPattern::ChunkSparsityPattern(
  const size_type               m,
  const size_type               n,
  const std::vector<size_type> &row_lengths,
  const size_type               chunk_size)
{
  Assert(chunk_size > 0, ExcInvalidNumber(chunk_size));

  reinit(m, n, row_lengths, chunk_size);
}



ChunkSparsityPattern::ChunkSparsityPattern(const size_type n,
                                           const size_type max_per_row,
                                           const size_type chunk_size)
{
  reinit(n, n, max_per_row, chunk_size);
}



ChunkSparsityPattern::ChunkSparsityPattern(
  const size_type               m,
  const std::vector<size_type> &row_lengths,
  const size_type               chunk_size)
{
  Assert(chunk_size > 0, ExcInvalidNumber(chunk_size));

  reinit(m, m, row_lengths, chunk_size);
}



ChunkSparsityPattern &
ChunkSparsityPattern::operator=(const ChunkSparsityPattern &s)
{
  Assert(s.rows == 0 && s.cols == 0,
         ExcMessage(
           "This operator can only be called if the provided argument "
           "is the sparsity pattern for an empty matrix. This operator can "
           "not be used to copy a non-empty sparsity pattern."));

  Assert(rows == 0 && cols == 0,
         ExcMessage("This operator can only be called if the current object is "
                    "empty."));

  // perform the checks in the underlying object as well
  sparsity_pattern = s.sparsity_pattern;

  return *this;
}



void
ChunkSparsityPattern::reinit(const size_type m,
                             const size_type n,
                             const size_type max_per_row,
                             const size_type chunk_size)
{
  Assert(chunk_size > 0, ExcInvalidNumber(chunk_size));

  // simply map this function to the other @p{reinit} function
  const std::vector<size_type> row_lengths(m, max_per_row);
  reinit(m, n, row_lengths, chunk_size);
}



void
ChunkSparsityPattern::reinit(const size_type                   m,
                             const size_type                   n,
                             const ArrayView<const size_type> &row_lengths,
                             const size_type                   chunk_size)
{
  Assert(row_lengths.size() == m, ExcInvalidNumber(m));
  Assert(chunk_size > 0, ExcInvalidNumber(chunk_size));

  rows = m;
  cols = n;

  this->chunk_size = chunk_size;

  // pass down to the necessary information to the underlying object. we need
  // to calculate how many chunks we need: we need to round up (m/chunk_size)
  // and (n/chunk_size). rounding up in integer arithmetic equals
  // ((m+chunk_size-1)/chunk_size):
  const size_type m_chunks = (m + chunk_size - 1) / chunk_size,
                  n_chunks = (n + chunk_size - 1) / chunk_size;

  // compute the maximum number of chunks in each row. the passed array
  // denotes the number of entries in each row of the big matrix -- in the
  // worst case, these are all in independent chunks, so we have to calculate
  // it as follows (as an example: let chunk_size==2, row_lengths={2,2,...},
  // and entries in row zero at columns {0,2} and for row one at {4,6} -->
  // we'll need 4 chunks for the first chunk row!) :
  std::vector<unsigned int> chunk_row_lengths(m_chunks, 0);
  for (size_type i = 0; i < m; ++i)
    chunk_row_lengths[i / chunk_size] += row_lengths[i];

  // for the case that the reduced sparsity pattern optimizes the diagonal but
  // the actual sparsity pattern does not, need to take one more entry in the
  // row to fit the user-required entry
  if (m != n && m_chunks == n_chunks)
    for (unsigned int i = 0; i < m_chunks; ++i)
      ++chunk_row_lengths[i];

  sparsity_pattern.reinit(m_chunks, n_chunks, chunk_row_lengths);
}



void
ChunkSparsityPattern::compress()
{
  sparsity_pattern.compress();
}



template <typename SparsityPatternType>
void
ChunkSparsityPattern::copy_from(const SparsityPatternType &dsp,
                                const size_type            chunk_size)
{
  Assert(chunk_size > 0, ExcInvalidNumber(chunk_size));
  this->chunk_size = chunk_size;
  rows             = dsp.n_rows();
  cols             = dsp.n_cols();

  // simple case: just use the given sparsity pattern
  if (chunk_size == 1)
    {
      sparsity_pattern.copy_from(dsp);
      return;
    }

  // create a temporary compressed sparsity pattern that collects all entries
  // from the input sparsity pattern and then initialize the underlying small
  // sparsity pattern
  const size_type m_chunks = (dsp.n_rows() + chunk_size - 1) / chunk_size,
                  n_chunks = (dsp.n_cols() + chunk_size - 1) / chunk_size;
  DynamicSparsityPattern temporary_sp(m_chunks, n_chunks);

  for (size_type row = 0; row < dsp.n_rows(); ++row)
    {
      const size_type reduced_row = row / chunk_size;

      // TODO: This could be made more efficient if we cached the
      // previous column and only called add() if the previous and the
      // current column lead to different chunk columns
      for (typename SparsityPatternType::iterator col_num = dsp.begin(row);
           col_num != dsp.end(row);
           ++col_num)
        temporary_sp.add(reduced_row, col_num->column() / chunk_size);
    }

  sparsity_pattern.copy_from(temporary_sp);
}



template <typename number>
void
ChunkSparsityPattern::copy_from(const FullMatrix<number> &matrix,
                                const size_type           chunk_size)
{
  Assert(chunk_size > 0, ExcInvalidNumber(chunk_size));

  // count number of entries per row, then initialize the underlying sparsity
  // pattern. remember to also allocate space for the diagonal entry (if that
  // hasn't happened yet) if m==n since we always allocate that for diagonal
  // matrices
  std::vector<size_type> entries_per_row(matrix.m(), 0);
  for (size_type row = 0; row < matrix.m(); ++row)
    {
      for (size_type col = 0; col < matrix.n(); ++col)
        if (matrix(row, col) != 0)
          ++entries_per_row[row];

      if ((matrix.m() == matrix.n()) && (matrix(row, row) == 0))
        ++entries_per_row[row];
    }

  reinit(matrix.m(), matrix.n(), entries_per_row, chunk_size);

  // then actually fill it
  for (size_type row = 0; row < matrix.m(); ++row)
    for (size_type col = 0; col < matrix.n(); ++col)
      if (matrix(row, col) != 0)
        add(row, col);

  // finally compress
  compress();
}



void
ChunkSparsityPattern::reinit(const size_type               m,
                             const size_type               n,
                             const std::vector<size_type> &row_lengths,
                             const size_type               chunk_size)
{
  Assert(chunk_size > 0, ExcInvalidNumber(chunk_size));

  reinit(m, n, make_array_view(row_lengths), chunk_size);
}



namespace internal
{
  namespace
  {
    template <typename SparsityPatternType>
    void
    copy_sparsity(const SparsityPatternType &src, SparsityPattern &dst)
    {
      dst.copy_from(src);
    }

    void
    copy_sparsity(const SparsityPattern &src, SparsityPattern &dst)
    {
      dst = src;
    }
  } // namespace
} // namespace internal



template <typename Sparsity>
void
ChunkSparsityPattern::create_from(const size_type m,
                                  const size_type n,
                                  const Sparsity &sparsity_pattern_for_chunks,
                                  const size_type chunk_size_in,
                                  const bool)
{
  Assert(m > (sparsity_pattern_for_chunks.n_rows() - 1) * chunk_size_in &&
           m <= sparsity_pattern_for_chunks.n_rows() * chunk_size_in,
         ExcMessage("Number of rows m is not compatible with chunk size "
                    "and number of rows in sparsity pattern for the chunks."));
  Assert(n > (sparsity_pattern_for_chunks.n_cols() - 1) * chunk_size_in &&
           n <= sparsity_pattern_for_chunks.n_cols() * chunk_size_in,
         ExcMessage(
           "Number of columns m is not compatible with chunk size "
           "and number of columns in sparsity pattern for the chunks."));

  internal::copy_sparsity(sparsity_pattern_for_chunks, sparsity_pattern);
  chunk_size = chunk_size_in;
  rows       = m;
  cols       = n;
}



bool
ChunkSparsityPattern::empty() const
{
  return sparsity_pattern.empty();
}



ChunkSparsityPattern::size_type
ChunkSparsityPattern::max_entries_per_row() const
{
  return sparsity_pattern.max_entries_per_row() * chunk_size;
}



void
ChunkSparsityPattern::add(const size_type i, const size_type j)
{
  Assert(i < rows, ExcInvalidIndex(i, rows));
  Assert(j < cols, ExcInvalidIndex(j, cols));

  sparsity_pattern.add(i / chunk_size, j / chunk_size);
}


bool
ChunkSparsityPattern::exists(const size_type i, const size_type j) const
{
  Assert(i < rows, ExcIndexRange(i, 0, rows));
  Assert(j < cols, ExcIndexRange(j, 0, cols));

  return sparsity_pattern.exists(i / chunk_size, j / chunk_size);
}



void
ChunkSparsityPattern::symmetrize()
{
  // matrix must be square. note that the for some matrix sizes, the current
  // sparsity pattern may not be square even if the underlying sparsity
  // pattern is (e.g. a 10x11 matrix with chunk_size 4)
  Assert(rows == cols, ExcNotQuadratic());

  sparsity_pattern.symmetrize();
}



ChunkSparsityPattern::size_type
ChunkSparsityPattern::row_length(const size_type i) const
{
  Assert(i < rows, ExcIndexRange(i, 0, rows));

  // find out if we did padding and if this row is affected by it
  if (n_cols() % chunk_size == 0)
    return sparsity_pattern.row_length(i / chunk_size) * chunk_size;
  else
    // if columns don't align, then just iterate over all chunks and see
    // what this leads to
    {
      SparsityPattern::const_iterator p =
                                        sparsity_pattern.begin(i / chunk_size),
                                      end =
                                        sparsity_pattern.end(i / chunk_size);
      unsigned int n = 0;
      for (; p != end; ++p)
        if (p->column() != sparsity_pattern.n_cols() - 1)
          n += chunk_size;
        else
          n += (n_cols() % chunk_size);
      return n;
    }
}



ChunkSparsityPattern::size_type
ChunkSparsityPattern::n_nonzero_elements() const
{
  if ((n_rows() % chunk_size == 0) && (n_cols() % chunk_size == 0))
    return (sparsity_pattern.n_nonzero_elements() * chunk_size * chunk_size);
  else
    // some of the chunks reach beyond the extent of this matrix. this
    // requires a somewhat more complicated computations, in particular if the
    // columns don't align
    {
      if ((n_rows() % chunk_size != 0) && (n_cols() % chunk_size == 0))
        {
          // columns align with chunks, but not rows
          size_type n =
            sparsity_pattern.n_nonzero_elements() * chunk_size * chunk_size;
          n -= (sparsity_pattern.n_rows() * chunk_size - n_rows()) *
               sparsity_pattern.row_length(sparsity_pattern.n_rows() - 1) *
               chunk_size;
          return n;
        }

      else
        {
          // if columns don't align, then just iterate over all chunks and see
          // what this leads to. follow the advice in the documentation of the
          // sparsity pattern iterators to do the loop over individual rows,
          // rather than all elements
          size_type n = 0;

          for (size_type row = 0; row < sparsity_pattern.n_rows(); ++row)
            {
              SparsityPattern::const_iterator p = sparsity_pattern.begin(row);
              for (; p != sparsity_pattern.end(row); ++p)
                if ((row != sparsity_pattern.n_rows() - 1) &&
                    (p->column() != sparsity_pattern.n_cols() - 1))
                  n += chunk_size * chunk_size;
                else if ((row == sparsity_pattern.n_rows() - 1) &&
                         (p->column() != sparsity_pattern.n_cols() - 1))
                  // last chunk row, but not last chunk column. only a smaller
                  // number (n_rows % chunk_size) of rows actually exist
                  n += (n_rows() % chunk_size) * chunk_size;
                else if ((row != sparsity_pattern.n_rows() - 1) &&
                         (p->column() == sparsity_pattern.n_cols() - 1))
                  // last chunk column, but not row
                  n += (n_cols() % chunk_size) * chunk_size;
                else
                  // bottom right chunk
                  n += (n_cols() % chunk_size) * (n_rows() % chunk_size);
            }

          return n;
        }
    }
}



void
ChunkSparsityPattern::print(std::ostream &out) const
{
  Assert((sparsity_pattern.rowstart != nullptr) &&
           (sparsity_pattern.colnums != nullptr),
         ExcEmptyObject());

  AssertThrow(out, ExcIO());

  for (size_type i = 0; i < sparsity_pattern.rows; ++i)
    for (size_type d = 0; (d < chunk_size) && (i * chunk_size + d < n_rows());
         ++d)
      {
        out << '[' << i * chunk_size + d;
        for (size_type j = sparsity_pattern.rowstart[i];
             j < sparsity_pattern.rowstart[i + 1];
             ++j)
          if (sparsity_pattern.colnums[j] != sparsity_pattern.invalid_entry)
            for (size_type e = 0;
                 ((e < chunk_size) &&
                  (sparsity_pattern.colnums[j] * chunk_size + e < n_cols()));
                 ++e)
              out << ',' << sparsity_pattern.colnums[j] * chunk_size + e;
        out << ']' << std::endl;
      }

  AssertThrow(out, ExcIO());
}



void
ChunkSparsityPattern::print_gnuplot(std::ostream &out) const
{
  Assert((sparsity_pattern.rowstart != nullptr) &&
           (sparsity_pattern.colnums != nullptr),
         ExcEmptyObject());

  AssertThrow(out, ExcIO());

  // for each entry in the underlying sparsity pattern, repeat everything
  // chunk_size x chunk_size times
  for (size_type i = 0; i < sparsity_pattern.rows; ++i)
    for (size_type j = sparsity_pattern.rowstart[i];
         j < sparsity_pattern.rowstart[i + 1];
         ++j)
      if (sparsity_pattern.colnums[j] != sparsity_pattern.invalid_entry)
        for (size_type d = 0;
             ((d < chunk_size) &&
              (sparsity_pattern.colnums[j] * chunk_size + d < n_cols()));
             ++d)
          for (size_type e = 0;
               (e < chunk_size) && (i * chunk_size + e < n_rows());
               ++e)
            // while matrix entries are usually written (i,j), with i vertical
            // and j horizontal, gnuplot output is x-y, that is we have to
            // exchange the order of output
            out << sparsity_pattern.colnums[j] * chunk_size + d << " "
                << -static_cast<signed int>(i * chunk_size + e) << std::endl;

  AssertThrow(out, ExcIO());
}



ChunkSparsityPattern::size_type
ChunkSparsityPattern::bandwidth() const
{
  // calculate the bandwidth from that of the underlying sparsity
  // pattern. note that even if the bandwidth of that is zero, then the
  // bandwidth of the chunky pattern is chunk_size-1, if it is 1 then the
  // chunky pattern has chunk_size+(chunk_size-1), etc
  //
  // we'll cut it off at max(n(),m())
  return std::min(sparsity_pattern.bandwidth() * chunk_size + (chunk_size - 1),
                  std::max(n_rows(), n_cols()));
}



bool
ChunkSparsityPattern::stores_only_added_elements() const
{
  if (chunk_size == 1)
    return sparsity_pattern.stores_only_added_elements();
  else
    return false;
}



void
ChunkSparsityPattern::block_write(std::ostream &out) const
{
  AssertThrow(out, ExcIO());

  // first the simple objects, bracketed in [...]
  out << '[' << rows << ' ' << cols << ' ' << chunk_size << ' ' << "][";
  // then the underlying sparsity pattern
  sparsity_pattern.block_write(out);
  out << ']';

  AssertThrow(out, ExcIO());
}



void
ChunkSparsityPattern::block_read(std::istream &in)
{
  AssertThrow(in, ExcIO());

  char c;

  // first read in simple data
  in >> c;
  AssertThrow(c == '[', ExcIO());
  in >> rows >> cols >> chunk_size;

  in >> c;
  AssertThrow(c == ']', ExcIO());
  in >> c;
  AssertThrow(c == '[', ExcIO());

  // then read the underlying sparsity pattern
  sparsity_pattern.block_read(in);

  in >> c;
  AssertThrow(c == ']', ExcIO());
}



std::size_t
ChunkSparsityPattern::memory_consumption() const
{
  return (sizeof(*this) + sparsity_pattern.memory_consumption());
}



// explicit instantiations
template void
ChunkSparsityPattern::copy_from<DynamicSparsityPattern>(
  const DynamicSparsityPattern &,
  const size_type);
template void
ChunkSparsityPattern::create_from<SparsityPattern>(const size_type,
                                                   const size_type,
                                                   const SparsityPattern &,
                                                   const size_type,
                                                   const bool);
template void
ChunkSparsityPattern::create_from<DynamicSparsityPattern>(
  const size_type,
  const size_type,
  const DynamicSparsityPattern &,
  const size_type,
  const bool);
template void
ChunkSparsityPattern::copy_from<float>(const FullMatrix<float> &,
                                       const size_type);
template void
ChunkSparsityPattern::copy_from<double>(const FullMatrix<double> &,
                                        const size_type);

DEAL_II_NAMESPACE_CLOSE