toroidal_average.hpp

Go to the documentation of this file.

#ifndef TOROIDAL_AVERAGE_HPP
#define TOROIDAL_AVERAGE_HPP

#include "my_mirror_view.hpp"
#include "domain_decomposition.hpp"

// Does a copy to the MPI memory space if different from the View's memory space
template<class T>
void toroidal_average_in_place(const DomainDecomposition<DeviceType>& pol_decomp, T& f){
    // No operation if n_planes is 1
    if(pol_decomp.mpi.n_intpl_ranks==1) return;

    // Cast to 1D with an unmanaged view for a generic element-wise operation
    View<double*,CLayout, typename T::device_type, Kokkos::MemoryTraits<Kokkos::Unmanaged>> f_1D(f.data(), f.size());

    using MPIHostType = HostType; // Do this MPI_Allreduce on Host; GPU-aware MPI consistently hangs or aborts here

    // Copy to host if on GPU and GPU-aware MPI is not available
    auto f_mpi_d = my_mirror_scratch_view(f_1D,MPIHostType());
    auto f_mpi = my_mirror_view(f_1D, MPIHostType(), f_mpi_d);
    mirror_copy(f_mpi, f_1D);

    /*** Sum all planes ***/
    MPI_Allreduce(MPI_IN_PLACE, f_mpi.data(), f_mpi.size(), MPI_DOUBLE, MPI_SUM, pol_decomp.mpi.intpl_comm);

    // Copy back
    mirror_copy(f_1D, f_mpi);

    /*** Normalize by n planes ***/
    // Invert first so that the loop has a multiplication rather than a division
    double inv_n_planes = 1.0/pol_decomp.mpi.n_intpl_ranks;

    // Do the normalization
    Kokkos::parallel_for("toroidal_average", Kokkos::RangePolicy<typename T::execution_space>(0, f.size()), KOKKOS_LAMBDA(const int i){
        f_1D(i) *= inv_n_planes;
    });
    Kokkos::fence();
}

#endif