DenseVector/OpenMP/Morpheus_Scan_Impl.hpp

 
#ifndef MORPHEUS_DENSEVECTOR_OPENMP_SCAN_IMPL_HPP
#define MORPHEUS_DENSEVECTOR_OPENMP_SCAN_IMPL_HPP
 
#include <Morpheus_Macros.hpp>
#if defined(MORPHEUS_ENABLE_OPENMP)
 
#include <Morpheus_SpaceTraits.hpp>
#include <Morpheus_FormatTraits.hpp>
#include <Morpheus_FormatTags.hpp>
#include <Morpheus_Spaces.hpp>
 
namespace Morpheus {
namespace Impl {
 
template <typename ExecSpace, typename Vector>
void inclusive_scan(const Vector& in, Vector& out,
                    typename Vector::index_type size,
                    typename Vector::index_type start,
                    typename std::enable_if_t<
                        Morpheus::is_dense_vector_format_container_v<Vector> &&
                        Morpheus::has_custom_backend_v<ExecSpace> &&
                        Morpheus::has_openmp_space_v<ExecSpace> &&
                        Morpheus::has_access_v<ExecSpace, Vector>>* = nullptr) {
  using index_type = typename Vector::index_type;
  using value_type = typename Vector::value_type;
 
#if _OPENMP >= 201511
  value_type initial = 0;
// TODO: Scan semantics require OpenMP5
#pragma omp simd reduction(inscan, + : initial)
  for (index_type i = start; i < size; i++) {
    initial += in[i];
#pragma omp scan inclusive(initial)
    out[i] = initial;
  }
#else
  static_assert("Requires OpenMP4.5 and above");
#endif
}
 
template <typename ExecSpace, typename Vector>
void exclusive_scan(const Vector& in, Vector& out,
                    typename Vector::index_type size,
                    typename Vector::index_type start,
                    typename std::enable_if_t<
                        Morpheus::is_dense_vector_format_container_v<Vector> &&
                        Morpheus::has_custom_backend_v<ExecSpace> &&
                        Morpheus::has_openmp_space_v<ExecSpace> &&
                        Morpheus::has_access_v<ExecSpace, Vector>>* = nullptr) {
  using index_type = typename Vector::index_type;
  using value_type = typename Vector::value_type;
 
#if _OPENMP >= 201511
  if (size > 0) {
    value_type initial = 0;
    // TODO: Scan semantics require OpenMP5
#pragma omp simd reduction(inscan, + : initial)
    for (index_type i = start; i < size - 1; i++) {
      out[i] = initial;
#pragma omp scan exclusive(initial)
      initial += in[i];
    }
    out[size - 1] = initial;
  }
#else
  static_assert("Requires OpenMP4.5 and above");
#endif
}
 
}  // namespace Impl
}  // namespace Morpheus
 
#endif  // MORPHEUS_ENABLE_OPENMP
#endif  // MORPHEUS_DENSEVECTOR_OPENMP_SCAN_IMPL_HPP