plane_field_gatherer.hpp

Go to the documentation of this file.

#ifndef PLANE_FIELD_GATHERER_HPP
#define PLANE_FIELD_GATHERER_HPP

#include "domain_decomposition.hpp"
#include "grid.hpp"
#include "my_subview.hpp"
#include "get_potential_grad.hpp"

class PlaneFieldGatherer{
    const DomainDecomposition<DeviceType> pol_decomp;
    const Grid<DeviceType> grid;

    bool gather_subset;
    bool gather_near_field;
    int near_field_pid;
    int nplanes;
    int nnode;

    int tmp_nphi;
    View<Field<VarType::Scalar,PIT_GLOBAL>**, CLayout,HostType> tmp_s; // for scalar fields
    View<Field<VarType::Vector,PIT_GLOBAL>**, CLayout,HostType> tmp_v; // for vector fields

    View<Field<VarType::Scalar,PIT_GLOBAL>*, CLayout,HostType> tmp_s_full; // for scalar fields
    View<Field<VarType::Vector,PIT_GLOBAL>*, CLayout,HostType> tmp_v_full; // for vector fields

    /* USE_GPU without USE_GPU_AWARE_MPI requires that a temporary array be allocated
     */
    int choose_tmp_nphi(){
#if defined(USE_GPU) && !defined(USE_GPU_AWARE_MPI)
        return nplanes;
#else
        return 0;
#endif
    }

    /* // Series of host MPI_Igatherv to populate phi_ff. An MPI_Allgatherv or other op (or intercommunicator?) might be possible/better.
     */
    template<typename T, typename FT>
    void allgather_to_local_ranks(View<FT*, CLayout,HostType>& tmp_full,T* destination){
#ifdef USE_MPI
        int one_obj_in_dbl = sizeof(FT)/8; // from bytes to doubles

        View<int*, HostType> recvcounts(NoInit("recvcounts"), pol_decomp.mpi.n_intpl_ranks);
        View<int*, HostType> displs(NoInit("displs"),pol_decomp.mpi.n_intpl_ranks);
        std::vector<MPI_Request> rrequests(pol_decomp.mpi.n_intpl_ranks);

        // Loop over ranks in global inter-planar communicator, each taking turns as root
        for(int i_root=0; i_root<pol_decomp.mpi.n_intpl_ranks; i_root++){
            int root_local_rank = pol_decomp.field_decomp.map_from_global_intpl(i_root); // will range from e.g. 0 to 7
            int root_near_field_pid = pol_decomp.field_decomp.find_domain_owner(i_root, grid.nplanes, pol_decomp.node_offset, grid.nnode);

            // Gather from the ranks that hold the assigned planes, or from the ranks that hold the near field planes if gathering the near field
            int source_rank = gather_near_field ? root_near_field_pid : root_local_rank;

            int root_local_first_plane = pol_decomp.field_decomp.all_first_plane(source_rank);
            int root_local_last_plane = pol_decomp.field_decomp.all_last_plane(source_rank);
            int root_local_nplanes = positive_modulo(root_local_last_plane - root_local_first_plane,grid.nplanes) + 1;

            int root_local_first_node = pol_decomp.field_decomp.all_first_node(source_rank);
            int root_local_nnode = pol_decomp.field_decomp.all_last_node(source_rank) - pol_decomp.field_decomp.all_first_node(source_rank) + 1;

            int size = root_local_nnode * one_obj_in_dbl;

            // Now determine where each contributing rank will send its contribution
            // Initialize to zero
            Kokkos::deep_copy(recvcounts, 0);
            Kokkos::deep_copy(displs, 0);

            // Note that the data arrives OUT OF ORDER to handle the case where the ghost planes wrap around plane 0
            for(int i=0; i<root_local_nplanes; i++){
                int contributor = (i+root_local_first_plane)%grid.nplanes;
                recvcounts(contributor) = size;
                displs(contributor) = i*size;
                //if(i<displs.size()-1) displs(i+1) = displs(i) + recvcounts(contributor);
            }

            int my_size = recvcounts(pol_decomp.mpi.my_intpl_rank);

            MPI_Igatherv(tmp_full.data()+root_local_first_node, my_size, MPI_DOUBLE, destination, recvcounts.data(), displs.data(), MPI_DOUBLE, i_root, pol_decomp.mpi.intpl_comm, &(rrequests[i_root]));
        }
        // Wait for Igatherv to complete
        for(int i_root=0; i_root<pol_decomp.mpi.n_intpl_ranks; i_root++) MPI_Wait(&(rrequests[i_root]), MPI_STATUS_IGNORE);
#endif
    }

    /* Copy the irho==0 index of a rho_ff field
     */
    template<typename FT>
    void copy_to_tmp_full(View<FT*, CLayout,HostType>& tmp_full, const View<FT**, CLayout,HostType>& rho_ff_h){
        constexpr int ZERO_GYRO = 0;

        // Check that the host array is allocated
        if(rho_ff_h.extent(0) != grid.nnode || rho_ff_h.extent(1)<1) exit_XGC("\nError in gather_phi_ff_on_device: expected host view of size (nnode,nrho)\n");

        // Copy this rank's contribution to tmp_full
        Kokkos::parallel_for("gather_field_info", Kokkos::RangePolicy<HostExSpace>( 0, grid.nnode), [=] (const int inode){
            tmp_full(inode)=rho_ff_h(inode,ZERO_GYRO);
        });
    }

    /* Copy a ff field
     */
    template<typename FT>
    void copy_to_tmp_full(View<FT*, CLayout,HostType>& tmp_full, const View<FT*, CLayout,HostType>& ff_h){
        // Check that the host array is allocated
        if(ff_h.extent(0) != grid.nnode) exit_XGC("\nError in gather_phi_ff_on_device: expected host view of size (nnode,nrho)\n");

        // Copy this rank's contribution to tmp_full
        Kokkos::parallel_for("gather_field_info", Kokkos::RangePolicy<HostExSpace>( 0, grid.nnode), [=] (const int inode){
            tmp_full(inode)=ff_h(inode);
        });
    }

    public:

    template<typename GT0, typename GT, typename GT2>
    void calculate_phi_ff_on_device(const Simulation<DeviceType>& sml, const Grid<DeviceType>& grid, const MagneticField<DeviceType>& magnetic_field, Smoothing& smoothing, GetPotentialGradTemp<DeviceType, DeviceType>& tmp, const View<Field<VarType::Scalar,PhiInterpType::None>*,CLayout,HostType>& dpot_h, const GT0& pot0_h, GT& phi_ff, GT2& pot_phi_ff, bool potential_is_requested=true, bool use_field00=false, bool ignore_poloidal_dpot=false){

        // Need 3 ghost planes for this algorithm
        int n_initial_ghost_planes = 2;
        int n_final_ghost_planes = 1;
        int n_ghost_planes = n_initial_ghost_planes + n_final_ghost_planes;

        GPTLstart("PRE_GATHER_POT");

        // Allocate phi_ff device view
        phi_ff = GT("gfpack_view",nplanes, nnode);

        // Make an unmanaged view to reuse the phi_ff allocation:
        auto* ptr_to_end_of_phi_ff = phi_ff.data()+phi_ff.size();
        int size_of_dpot_phi = (nplanes+n_ghost_planes)*nnode;
        double* dpot_phi_addr = (double*)(ptr_to_end_of_phi_ff) - size_of_dpot_phi;
        View<double**,CLayout,DeviceType, Kokkos::MemoryTraits<Kokkos::Unmanaged>> dpot_phi(dpot_phi_addr, nplanes+n_ghost_planes, nnode);
        GPTLstop("PRE_GATHER_POT");

        GPTLstart("GATHER_POT");
#ifdef USE_MPI
        // Gather the potential from all planes, leaving room at the beginning of the array for two ghost planes

        // Mirror views for when GPU-aware MPI is not available
        auto pot_tmp = my_mirror_view(dpot_h, MPIDeviceType());
        mirror_copy(pot_tmp, dpot_h);

        auto dpot_phi_MPI_d = my_mirror_scratch_view(dpot_phi,MPIDeviceType());
        auto dpot_phi_tmp = my_mirror_view(dpot_phi, MPIDeviceType(), dpot_phi_MPI_d);
        auto destination = dpot_phi_tmp.data() + nnode*n_initial_ghost_planes; // 2 initial ghost planes
        MPI_Allgather(pot_tmp.data(), nnode, MPI_DOUBLE, destination, nnode, MPI_DOUBLE, pol_decomp.mpi.intpl_comm);
        mirror_copy(dpot_phi, dpot_phi_tmp); // Send to device
#endif
        GPTLstop("GATHER_POT");

        // Incoming view input_field has size (nplanes+3): 2 ghost planes at the beginning, 1 at the end

        // If the ghost planes wrap around, copy them
        // e.g. with 8 planes named Plane 0-7:
        // dpot_phi(0) is a copy of Plane 6
        // dpot_phi(1) is a copy of Plane 7
        // dpot_phi(10) is a copy of Plane 0
        int nplanes_l = nplanes;
        Kokkos::parallel_for("set_up_ghost_planes", Kokkos::RangePolicy<ExSpace>(0,grid.nnode), KOKKOS_LAMBDA( const int i ){
            dpot_phi(0,i) = dpot_phi(nplanes_l-2 + n_initial_ghost_planes, i);
            dpot_phi(1,i) = dpot_phi(nplanes_l-1 + n_initial_ghost_planes, i);
            dpot_phi(nplanes_l + n_initial_ghost_planes,i) = dpot_phi(0 + n_initial_ghost_planes, i);
        });
        Kokkos::fence();

        /**** Get potential gradient ****/
        GPTLstart("CALC_E_FROM_POT");

        GPTLstart("GET_POT_GRAD_PHI");
        // Allocate phi_ff device view
        if(potential_is_requested) pot_phi_ff = GT2("gfpack_potview",nplanes, nnode);

        auto gpg_field_args = GetPotGradFieldArgs<DeviceType, DeviceType, vec2d_if_axisym<PIT_GLOBAL>(), PIT_GLOBAL, TorType::MultiplePlanes, KinType::DriftKin>
                                            (dpot_phi, ignore_poloidal_dpot);

        // Set up E00
        if(use_field00){
            gpg_field_args.field00 = Field00<DeviceType>(pot0_h, sml.grad_psitheta);
            gpg_field_args.field00.calculate_gradient(grid, tmp.grad_matrices);
        }

        // Request potential and/or gradient outputs
        if(potential_is_requested) gpg_field_args.request_potential(pot_phi_ff);
        gpg_field_args.request_gradient(phi_ff);

        // Get requested fields
        get_field_grad(sml, grid, pol_decomp, magnetic_field, smoothing, gpg_field_args, tmp);

        GPTLstop("GET_POT_GRAD_PHI");
        GPTLstop("CALC_E_FROM_POT");
    }

    private:

    template<typename T_h, typename FT>
    void gather_phi_ff_on_device(View<FT**, CLayout,HostType>& tmp, View<FT*, CLayout,HostType>& tmp_full, const T_h& rho_ff_h, View<FT**, CLayout,DeviceType>& phi_ff){
#if defined(USE_GPU) && !defined(USE_GPU_AWARE_MPI)
        constexpr bool gpu_without_gpu_aware_MPI = true;
#else
        constexpr bool gpu_without_gpu_aware_MPI = false;
#endif

        // Copy this rank's contribution to tmp_full
        copy_to_tmp_full(tmp_full, rho_ff_h);

        // Allocate phi_ff device view
        phi_ff = View<FT**, CLayout,DeviceType>("gfpack_view",nplanes, nnode);

        // If using GPUs without GPU-aware MPI, need to do MPI gather into a temporary host allocation
        auto destination = (gpu_without_gpu_aware_MPI ? tmp.data() : phi_ff.data());

        // Gather contributions from each rank; if we're just gathering a subset of the field, the logic is more complicated
        if(gather_subset){
            allgather_to_local_ranks(tmp_full, destination);
        }else{
#ifdef USE_MPI
            int one_obj_in_dbl = sizeof(FT)/8; // from bytes to doubles
            int sz=nnode * one_obj_in_dbl;
            MPI_Allgather(tmp_full.data(), sz, MPI_DOUBLE, destination, sz, MPI_DOUBLE, pol_decomp.mpi.intpl_comm);
#else
            // Used for mini_appKernel; just replicate the rho field for now
            if(gpu_without_gpu_aware_MPI)
                for(int i=0; i<grid.nplanes; i++) Kokkos::deep_copy(my_subview(tmp, i), tmp_full);
            else
                for(int i=0; i<grid.nplanes; i++) Kokkos::deep_copy(my_subview(phi_ff, i), tmp_full);
#endif
        }
        // If using GPUs without GPU-aware MPI, copy from the temporary allocation to device memory
        if(gpu_without_gpu_aware_MPI){
            Kokkos::deep_copy(phi_ff, tmp);
        }
    }

    template<typename FT>
    inline View<FT**, CLayout,HostType>& which_tmp();

    template<typename FT>
    inline View<FT*, CLayout,HostType>& which_tmp_full();

    public:

    PlaneFieldGatherer(const DomainDecomposition<DeviceType>& pol_decomp, const Grid<DeviceType>& grid, bool near_field=false)
      : pol_decomp(pol_decomp),
        grid(grid),
        gather_subset(pol_decomp.decompose_fields),
        gather_near_field(near_field),
        near_field_pid((gather_subset && near_field) ? pol_decomp.field_decomp.find_domain_owner(pol_decomp.plane_index, grid.nplanes, pol_decomp.node_offset, grid.nnode) : 0),
        // If using field_decomposition, we don't copy the entire field
        nplanes(     gather_subset ? (near_field ? pol_decomp.field_decomp.all_n_planes   (near_field_pid, grid.nplanes) : pol_decomp.field_decomp.n_planes   ) : grid.nplanes),
        nnode(       gather_subset ? (near_field ? pol_decomp.field_decomp.all_n_nodes    (near_field_pid)               : pol_decomp.field_decomp.n_nodes    ) : grid.nnode),
        tmp_nphi(choose_tmp_nphi()),
        tmp_s("tmp_s", tmp_nphi, nnode),
        tmp_v("tmp_v", tmp_nphi, nnode),
        tmp_s_full("tmp_s_full", grid.nnode),
        tmp_v_full("tmp_v_full", grid.nnode)
    {}

    // Interface to use correct tmp
    template<typename T_h, typename FT>
    void gather_phi_ff_on_device(const T_h& rho_ff_h, View<FT**, CLayout,DeviceType>& phi_ff){
        gather_phi_ff_on_device(which_tmp<FT>(), which_tmp_full<FT>(), rho_ff_h, phi_ff);
    }
};

template<>
inline View<Field<VarType::Scalar,PIT_GLOBAL>**, CLayout,HostType>& PlaneFieldGatherer::which_tmp(){
    return tmp_s;
}

template<>
inline View<Field<VarType::Vector,PIT_GLOBAL>**, CLayout,HostType>& PlaneFieldGatherer::which_tmp(){
    return tmp_v;
}

template<>
inline View<Field<VarType::Scalar,PIT_GLOBAL>*, CLayout,HostType>& PlaneFieldGatherer::which_tmp_full(){
    return tmp_s_full;
}

template<>
inline View<Field<VarType::Vector,PIT_GLOBAL>*, CLayout,HostType>& PlaneFieldGatherer::which_tmp_full(){
    return tmp_v_full;
}

#endif