DenseMatrix/OpenMP/Morpheus_Convert_Impl.hpp

 
#ifndef MORPHEUS_DENSEMATRIX_OPENMP_CONVERT_IMPL_HPP
#define MORPHEUS_DENSEMATRIX_OPENMP_CONVERT_IMPL_HPP
 
#include <Morpheus_Macros.hpp>
#if defined(MORPHEUS_ENABLE_OPENMP)
 
#include <Morpheus_Exceptions.hpp>
#include <Morpheus_SpaceTraits.hpp>
#include <Morpheus_FormatTraits.hpp>
#include <Morpheus_FormatTags.hpp>
#include <Morpheus_Spaces.hpp>
 
namespace Morpheus {
namespace Impl {
 
template <typename ExecSpace, typename SourceType, typename DestinationType>
void convert(
    const SourceType& src, DestinationType& dst,
    typename std::enable_if<
        Morpheus::is_dense_matrix_format_container_v<SourceType> &&
        Morpheus::is_dense_matrix_format_container_v<DestinationType> &&
        Morpheus::has_custom_backend_v<ExecSpace> &&
        Morpheus::has_openmp_execution_space_v<ExecSpace> &&
        Morpheus::has_access_v<ExecSpace, SourceType, DestinationType>>::type* =
        nullptr) {
  using size_type = typename SourceType::size_type;
 
  MORPHEUS_ASSERT((dst.nrows() >= src.nrows()) && (dst.ncols() >= src.ncols()),
                  "Destination matrix must have equal or larger shape to the "
                  "source matrix");
 
#pragma omp parallel for collapse(2)
  for (size_type i = 0; i < src.nrows(); i++) {
    for (size_type j = 0; j < src.ncols(); j++) {
      dst(i, j) = src(i, j);
    }
  }
}
 
template <typename ExecSpace, typename SourceType, typename DestinationType>
void convert(
    const SourceType& src, DestinationType& dst,
    typename std::enable_if<
        Morpheus::is_dense_matrix_format_container_v<SourceType> &&
        Morpheus::is_dense_vector_format_container_v<DestinationType> &&
        Morpheus::has_custom_backend_v<ExecSpace> &&
        Morpheus::has_openmp_execution_space_v<ExecSpace> &&
        Morpheus::has_access_v<ExecSpace, SourceType, DestinationType>>::type* =
        nullptr) {
  using size_type = typename SourceType::size_type;
 
  dst.resize(src.nrows() * src.ncols());
 
#pragma omp parallel for collapse(2)
  for (size_type i = 0; i < src.nrows(); i++) {
    for (size_type j = 0; j < src.ncols(); j++) {
      size_type idx = i * src.ncols() + j;
      dst(idx)      = src(i, j);
    }
  }
}
 
template <typename ExecSpace, typename SourceType, typename DestinationType>
void convert(
    const SourceType&, DestinationType&,
    typename std::enable_if<
        Morpheus::is_dense_matrix_format_container_v<SourceType> &&
        Morpheus::is_coo_matrix_format_container_v<DestinationType> &&
        Morpheus::has_custom_backend_v<ExecSpace> &&
        Morpheus::has_openmp_execution_space_v<ExecSpace> &&
        Morpheus::has_access_v<ExecSpace, SourceType, DestinationType>>::type* =
        nullptr) {
  throw Morpheus::NotImplementedException("convert<Kokkos::OpenMP>");
}
 
template <typename ExecSpace, typename SourceType, typename DestinationType>
void convert(
    const SourceType& src, DestinationType& dst,
    typename std::enable_if<
        Morpheus::is_coo_matrix_format_container_v<SourceType> &&
        Morpheus::is_dense_matrix_format_container_v<DestinationType> &&
        Morpheus::has_custom_backend_v<ExecSpace> &&
        Morpheus::has_openmp_execution_space_v<ExecSpace> &&
        Morpheus::has_access_v<ExecSpace, SourceType, DestinationType>>::type* =
        nullptr) {
  using size_type = typename SourceType::size_type;
 
  dst.resize(src.nrows(), src.ncols());
 
#pragma omp parallel for
  for (size_type n = 0; n < src.nnnz(); n++) {
    size_type i = src.crow_indices(n);
    size_type j = src.ccolumn_indices(n);
    dst(i, j)   = src.cvalues(n);
  }
}
 
}  // namespace Impl
}  // namespace Morpheus
 
#endif  // MORPHEUS_ENABLE_OPENMP
#endif  // MORPHEUS_DENSEMATRIX_OPENMP_CONVERT_IMPL_HPP