partition.h

// Copyright (C) 2020 Garth N. Wells
//
// This file is part of DOLFINx (https://www.fenicsproject.org)
//
// SPDX-License-Identifier:    LGPL-3.0-or-later
 
#pragma once
 
#include <algorithm>
#include <cstdint>
#include <dolfinx/common/MPI.h>
#include <dolfinx/common/Timer.h>
#include <dolfinx/graph/AdjacencyList.h>
#include <functional>
#include <mpi.h>
#include <utility>
#include <vector>
 
namespace dolfinx::graph
{
 
using partition_fn = std::function<graph::AdjacencyList<std::int32_t>(
    MPI_Comm comm, int nparts, const AdjacencyList<std::int64_t>& local_graph,
    std::int32_t num_ghost_nodes, bool ghosting)>;
 
AdjacencyList<std::int32_t>
partition_graph(const MPI_Comm comm, int nparts,
                const AdjacencyList<std::int64_t>& local_graph,
                std::int32_t num_ghost_nodes, bool ghosting);
 
namespace build
{
std::tuple<graph::AdjacencyList<std::int64_t>, std::vector<int>,
           std::vector<std::int64_t>, std::vector<int>>
distribute(MPI_Comm comm, const graph::AdjacencyList<std::int64_t>& list,
           const graph::AdjacencyList<std::int32_t>& destinations);
 
std::vector<std::int64_t>
compute_ghost_indices(MPI_Comm comm,
                      const std::vector<std::int64_t>& global_indices,
                      const std::vector<int>& ghost_owners);
 
template <typename T>
xt::xtensor<T, 2> distribute_data(MPI_Comm comm,
                                  const std::vector<std::int64_t>& indices,
                                  const xt::xtensor<T, 2>& x);
 
std::vector<std::int64_t>
compute_local_to_global_links(const graph::AdjacencyList<std::int64_t>& global,
                              const graph::AdjacencyList<std::int32_t>& local);
 
std::vector<std::int32_t>
compute_local_to_local(const std::vector<std::int64_t>& local0_to_global,
                       const std::vector<std::int64_t>& local1_to_global);
} // namespace build
 
//---------------------------------------------------------------------------
// Implementation
//---------------------------------------------------------------------------
template <typename T>
xt::xtensor<T, 2>
build::distribute_data(MPI_Comm comm, const std::vector<std::int64_t>& indices,
                       const xt::xtensor<T, 2>& x)
{
  common::Timer timer("Fetch float data from remote processes");
 
  const std::int64_t num_points_local = x.shape(0);
  const int size = dolfinx::MPI::size(comm);
  const int rank = dolfinx::MPI::rank(comm);
  std::vector<std::int64_t> global_sizes(size);
  MPI_Allgather(&num_points_local, 1, MPI_INT64_T, global_sizes.data(), 1,
                MPI_INT64_T, comm);
  std::vector<std::int64_t> global_offsets(size + 1, 0);
  std::partial_sum(global_sizes.begin(), global_sizes.end(),
                   global_offsets.begin() + 1);
 
  // Build index data requests
  std::vector<int> number_index_send(size, 0);
  std::vector<int> index_owner(indices.size());
  std::vector<int> index_order(indices.size());
  std::iota(index_order.begin(), index_order.end(), 0);
  std::sort(index_order.begin(), index_order.end(),
            [&indices](int a, int b) { return (indices[a] < indices[b]); });
 
  int p = 0;
  for (std::size_t i = 0; i < index_order.size(); ++i)
  {
    int j = index_order[i];
    while (indices[j] >= global_offsets[p + 1])
      ++p;
    index_owner[j] = p;
    number_index_send[p]++;
  }
 
  // Compute send displacements
  std::vector<int> disp_index_send(size + 1, 0);
  std::partial_sum(number_index_send.begin(), number_index_send.end(),
                   disp_index_send.begin() + 1);
 
  // Pack global index send data
  std::vector<std::int64_t> indices_send(disp_index_send.back());
  std::vector<int> disp_tmp = disp_index_send;
  for (std::size_t i = 0; i < indices.size(); ++i)
  {
    const int owner = index_owner[i];
    indices_send[disp_tmp[owner]++] = indices[i];
  }
 
  // Send/receive number of indices to communicate to each process
  std::vector<int> number_index_recv(size);
  MPI_Alltoall(number_index_send.data(), 1, MPI_INT, number_index_recv.data(),
               1, MPI_INT, comm);
 
  // Compute receive displacements
  std::vector<int> disp_index_recv(size + 1, 0);
  std::partial_sum(number_index_recv.begin(), number_index_recv.end(),
                   disp_index_recv.begin() + 1);
 
  // Send/receive global indices
  std::vector<std::int64_t> indices_recv(disp_index_recv.back());
  MPI_Alltoallv(indices_send.data(), number_index_send.data(),
                disp_index_send.data(), MPI_INT64_T, indices_recv.data(),
                number_index_recv.data(), disp_index_recv.data(), MPI_INT64_T,
                comm);
 
  assert(x.shape(1) != 0);
  // Pack point data to send back (transpose)
  xt::xtensor<T, 2> x_return({indices_recv.size(), x.shape(1)});
  for (int p = 0; p < size; ++p)
  {
    for (int i = disp_index_recv[p]; i < disp_index_recv[p + 1]; ++i)
    {
      const std::int32_t index_local = indices_recv[i] - global_offsets[rank];
      assert(index_local >= 0);
      for (std::size_t j = 0; j < x.shape(1); ++j)
        x_return(i, j) = x(index_local, j);
    }
  }
 
  MPI_Datatype compound_type;
  MPI_Type_contiguous(x.shape(1), dolfinx::MPI::mpi_type<T>(), &compound_type);
  MPI_Type_commit(&compound_type);
 
  // Send back point data
  xt::xtensor<T, 2> my_x(
      {static_cast<std::size_t>(disp_index_send.back()), x.shape(1)});
  MPI_Alltoallv(x_return.data(), number_index_recv.data(),
                disp_index_recv.data(), compound_type, my_x.data(),
                number_index_send.data(), disp_index_send.data(), compound_type,
                comm);
 
  return my_x;
}
 
} // namespace dolfinx::graph