Scatterer.h

// Copyright (C) 2022 Igor Baratta and Garth N. Wells
//
// This file is part of DOLFINx (https://www.fenicsproject.org)
//
// SPDX-License-Identifier:    LGPL-3.0-or-later
 
#pragma once
 
#include "IndexMap.h"
#include "MPI.h"
#include "sort.h"
#include <algorithm>
#include <functional>
#include <memory>
#include <mpi.h>
#include <numeric>
#include <span>
#include <type_traits>
#include <vector>
 
using namespace dolfinx;
 
namespace dolfinx::common
{
template <class Allocator = std::allocator<std::int32_t>>
class Scatterer
{
public:
  using allocator_type = Allocator;
 
  enum class type
  {
    neighbor, // use MPI neighborhood collectives
    p2p       // use MPI Isend/Irecv for communication
  };
 
  Scatterer(const IndexMap& map, int bs, const Allocator& alloc = Allocator())
      : _bs(bs), _remote_inds(0, alloc), _local_inds(0, alloc),
        _src(map.src().begin(), map.src().end()),
        _dest(map.dest().begin(), map.dest().end())
  {
    if (dolfinx::MPI::size(map.comm()) == 1)
      return;
 
    // Check that src and dest ranks are unique and sorted
    assert(std::is_sorted(_src.begin(), _src.end()));
    assert(std::is_sorted(_dest.begin(), _dest.end()));
 
    // Create communicators with directed edges:
    // (0) owner -> ghost,
    // (1) ghost -> owner
    MPI_Comm comm0;
    MPI_Dist_graph_create_adjacent(
        map.comm(), _src.size(), _src.data(), MPI_UNWEIGHTED, _dest.size(),
        _dest.data(), MPI_UNWEIGHTED, MPI_INFO_NULL, false, &comm0);
    _comm0 = dolfinx::MPI::Comm(comm0, false);
 
    MPI_Comm comm1;
    MPI_Dist_graph_create_adjacent(
        map.comm(), _dest.size(), _dest.data(), MPI_UNWEIGHTED, _src.size(),
        _src.data(), MPI_UNWEIGHTED, MPI_INFO_NULL, false, &comm1);
    _comm1 = dolfinx::MPI::Comm(comm1, false);
 
    // Build permutation array that sorts ghost indices by owning rank
    std::span owners = map.owners();
    std::vector<std::int32_t> perm(owners.size());
    std::iota(perm.begin(), perm.end(), 0);
    dolfinx::argsort_radix<std::int32_t>(owners, perm);
 
    // Sort (i) ghost indices and (ii) ghost index owners by rank
    // (using perm array)
    std::span ghosts = map.ghosts();
    std::vector<int> owners_sorted(owners.size());
    std::vector<std::int64_t> ghosts_sorted(owners.size());
    std::transform(perm.begin(), perm.end(), owners_sorted.begin(),
                   [&owners](auto idx) { return owners[idx]; });
    std::transform(perm.begin(), perm.end(), ghosts_sorted.begin(),
                   [&ghosts](auto idx) { return ghosts[idx]; });
 
    // For data associated with ghost indices, packed by owning
    // (neighbourhood) rank, compute sizes and displacements. I.e.,
    // when sending ghost index data from this rank to the owning
    // ranks, disp[i] is the first entry in the buffer sent to
    // neighbourhood rank i, and disp[i + 1] - disp[i] is the number
    // of values sent to rank i.
    _sizes_remote.resize(_src.size(), 0);
    _displs_remote.resize(_src.size() + 1, 0);
    std::vector<std::int32_t>::iterator begin = owners_sorted.begin();
    for (std::size_t i = 0; i < _src.size(); i++)
    {
      auto upper = std::upper_bound(begin, owners_sorted.end(), _src[i]);
      int num_ind = std::distance(begin, upper);
      _displs_remote[i + 1] = _displs_remote[i] + num_ind;
      _sizes_remote[i] = num_ind;
      begin = upper;
    }
 
    // For data associated with owned indices that are ghosted by
    // other ranks, compute the size and displacement arrays. When
    // sending data associated with ghost indices to the owner, these
    // size and displacement arrays are for the receive buffer.
 
    // Compute sizes and displacements of local data (how many local
    // elements to be sent/received grouped by neighbors)
    _sizes_local.resize(_dest.size());
    _displs_local.resize(_sizes_local.size() + 1);
    _sizes_remote.reserve(1);
    _sizes_local.reserve(1);
    MPI_Neighbor_alltoall(_sizes_remote.data(), 1, MPI_INT32_T,
                          _sizes_local.data(), 1, MPI_INT32_T, _comm1.comm());
    std::partial_sum(_sizes_local.begin(), _sizes_local.end(),
                     std::next(_displs_local.begin()));
 
    assert((std::int32_t)ghosts_sorted.size() == _displs_remote.back());
    assert((std::int32_t)ghosts_sorted.size() == _displs_remote.back());
 
    // Send ghost global indices to owning rank, and receive owned
    // indices that are ghosts on other ranks
    std::vector<std::int64_t> recv_buffer(_displs_local.back(), 0);
    MPI_Neighbor_alltoallv(ghosts_sorted.data(), _sizes_remote.data(),
                           _displs_remote.data(), MPI_INT64_T,
                           recv_buffer.data(), _sizes_local.data(),
                           _displs_local.data(), MPI_INT64_T, _comm1.comm());
 
    const std::array<std::int64_t, 2> range = map.local_range();
#ifndef NDEBUG
    // Check that all received indice are within the owned range
    std::for_each(recv_buffer.begin(), recv_buffer.end(), [range](auto idx)
                  { assert(idx >= range[0] and idx < range[1]); });
#endif
 
    // Scale sizes and displacements by block size
    {
      auto rescale = [](auto& x, int bs)
      {
        std::transform(x.begin(), x.end(), x.begin(),
                       [bs](auto e) { return e *= bs; });
      };
      rescale(_sizes_local, bs);
      rescale(_displs_local, bs);
      rescale(_sizes_remote, bs);
      rescale(_displs_remote, bs);
    }
 
    // Expand local indices using block size and convert it from
    // global to local numbering
    _local_inds = std::vector<std::int32_t, allocator_type>(
        recv_buffer.size() * _bs, alloc);
    std::int64_t offset = range[0] * _bs;
    for (std::size_t i = 0; i < recv_buffer.size(); i++)
      for (int j = 0; j < _bs; j++)
        _local_inds[i * _bs + j] = (recv_buffer[i] * _bs + j) - offset;
 
    // Expand remote indices using block size
    _remote_inds
        = std::vector<std::int32_t, allocator_type>(perm.size() * _bs, alloc);
    for (std::size_t i = 0; i < perm.size(); i++)
      for (int j = 0; j < _bs; j++)
        _remote_inds[i * _bs + j] = perm[i] * _bs + j;
  }
 
  template <typename T>
  void scatter_fwd_begin(std::span<const T> send_buffer,
                         std::span<T> recv_buffer,
                         std::span<MPI_Request> requests,
                         Scatterer::type type = type::neighbor) const
  {
    // Return early if there are no incoming or outgoing edges
    if (_sizes_local.empty() and _sizes_remote.empty())
      return;
 
    switch (type)
    {
    case type::neighbor:
    {
      assert(requests.size() == std::size_t(1));
      MPI_Ineighbor_alltoallv(
          send_buffer.data(), _sizes_local.data(), _displs_local.data(),
          dolfinx::MPI::mpi_type<T>(), recv_buffer.data(), _sizes_remote.data(),
          _displs_remote.data(), dolfinx::MPI::mpi_type<T>(), _comm0.comm(),
          requests.data());
      break;
    }
    case type::p2p:
    {
      assert(requests.size() == _dest.size() + _src.size());
      for (std::size_t i = 0; i < _src.size(); i++)
      {
        MPI_Irecv(recv_buffer.data() + _displs_remote[i], _sizes_remote[i],
                  dolfinx::MPI::mpi_type<T>(), _src[i], MPI_ANY_TAG,
                  _comm0.comm(), &requests[i]);
      }
 
      for (std::size_t i = 0; i < _dest.size(); i++)
      {
        MPI_Isend(send_buffer.data() + _displs_local[i], _sizes_local[i],
                  dolfinx::MPI::mpi_type<T>(), _dest[i], 0, _comm0.comm(),
                  &requests[i + _src.size()]);
      }
      break;
    }
    default:
      throw std::runtime_error("Scatter::type not recognized");
    }
  }
 
  void scatter_fwd_end(std::span<MPI_Request> requests) const
  {
    // Return early if there are no incoming or outgoing edges
    if (_sizes_local.empty() and _sizes_remote.empty())
      return;
 
    // Wait for communication to complete
    MPI_Waitall(requests.size(), requests.data(), MPI_STATUS_IGNORE);
  }
 
  template <typename T, typename F>
    requires std::is_invocable_v<F, std::span<const T>,
                                 std::span<const std::int32_t>, std::span<T>>
  void scatter_fwd_begin(std::span<const T> local_data,
                         std::span<T> local_buffer, std::span<T> remote_buffer,
                         F pack_fn, std::span<MPI_Request> requests,
                         Scatterer::type type = type::neighbor) const
  {
    assert(local_buffer.size() == _local_inds.size());
    assert(remote_buffer.size() == _remote_inds.size());
    pack_fn(local_data, _local_inds, local_buffer);
    scatter_fwd_begin(std::span<const T>(local_buffer), remote_buffer, requests,
                      type);
  }
 
  template <typename T, typename F>
    requires std::is_invocable_v<F, std::span<const T>,
                                 std::span<const std::int32_t>, std::span<T>,
                                 std::function<T(T, T)>>
  void scatter_fwd_end(std::span<const T> remote_buffer,
                       std::span<T> remote_data, F unpack_fn,
                       std::span<MPI_Request> requests) const
  {
    assert(remote_buffer.size() == _remote_inds.size());
    assert(remote_data.size() == _remote_inds.size());
    scatter_fwd_end(requests);
    unpack_fn(remote_buffer, _remote_inds, remote_data,
              [](T /*a*/, T b) { return b; });
  }
 
  template <typename T>
  void scatter_fwd(std::span<const T> local_data,
                   std::span<T> remote_data) const
  {
    std::vector<MPI_Request> requests(1, MPI_REQUEST_NULL);
    std::vector<T> local_buffer(local_buffer_size(), 0);
    std::vector<T> remote_buffer(remote_buffer_size(), 0);
    auto pack_fn = [](auto&& in, auto&& idx, auto&& out)
    {
      for (std::size_t i = 0; i < idx.size(); ++i)
        out[i] = in[idx[i]];
    };
    scatter_fwd_begin(local_data, std::span<T>(local_buffer),
                      std::span<T>(remote_buffer), pack_fn,
                      std::span<MPI_Request>(requests));
 
    auto unpack_fn = [](auto&& in, auto&& idx, auto&& out, auto op)
    {
      for (std::size_t i = 0; i < idx.size(); ++i)
        out[idx[i]] = op(out[idx[i]], in[i]);
    };
 
    scatter_fwd_end(std::span<const T>(remote_buffer), remote_data, unpack_fn,
                    std::span<MPI_Request>(requests));
  }
 
  template <typename T>
  void scatter_rev_begin(std::span<const T> send_buffer,
                         std::span<T> recv_buffer,
                         std::span<MPI_Request> requests,
                         Scatterer::type type = type::neighbor) const
  {
    // Return early if there are no incoming or outgoing edges
    if (_sizes_local.empty() and _sizes_remote.empty())
      return;
 
    // // Send and receive data
 
    switch (type)
    {
    case type::neighbor:
    {
      assert(requests.size() == 1);
      MPI_Ineighbor_alltoallv(send_buffer.data(), _sizes_remote.data(),
                              _displs_remote.data(), MPI::mpi_type<T>(),
                              recv_buffer.data(), _sizes_local.data(),
                              _displs_local.data(), MPI::mpi_type<T>(),
                              _comm1.comm(), &requests[0]);
      break;
    }
    case type::p2p:
    {
      assert(requests.size() == _dest.size() + _src.size());
      // Start non-blocking send from this process to ghost owners.
      for (std::size_t i = 0; i < _dest.size(); i++)
      {
        MPI_Irecv(recv_buffer.data() + _displs_local[i], _sizes_local[i],
                  dolfinx::MPI::mpi_type<T>(), _dest[i], MPI_ANY_TAG,
                  _comm0.comm(), &requests[i]);
      }
 
      // Start non-blocking receive from neighbor process for which an owned
      // index is a ghost.
      for (std::size_t i = 0; i < _src.size(); i++)
      {
        MPI_Isend(send_buffer.data() + _displs_remote[i], _sizes_remote[i],
                  dolfinx::MPI::mpi_type<T>(), _src[i], 0, _comm0.comm(),
                  &requests[i + _dest.size()]);
      }
      break;
    }
    default:
      throw std::runtime_error("Scatter::type not recognized");
    }
  }
 
  void scatter_rev_end(std::span<MPI_Request> request) const
  {
    // Return early if there are no incoming or outgoing edges
    if (_sizes_local.empty() and _sizes_remote.empty())
      return;
 
    // Wait for communication to complete
    MPI_Waitall(request.size(), request.data(), MPI_STATUS_IGNORE);
  }
 
  template <typename T, typename F>
    requires std::is_invocable_v<F, std::span<const T>,
                                 std::span<const std::int32_t>, std::span<T>>
  void scatter_rev_begin(std::span<const T> remote_data,
                         std::span<T> remote_buffer, std::span<T> local_buffer,
                         F pack_fn, std::span<MPI_Request> request,
                         Scatterer::type type = type::neighbor) const
  {
    assert(local_buffer.size() == _local_inds.size());
    assert(remote_buffer.size() == _remote_inds.size());
    pack_fn(remote_data, _remote_inds, remote_buffer);
    scatter_rev_begin(std::span<const T>(remote_buffer), local_buffer, request,
                      type);
  }
 
  template <typename T, typename F, typename BinaryOp>
    requires std::is_invocable_v<F, std::span<const T>,
                                 std::span<const std::int32_t>, std::span<T>,
                                 BinaryOp>
             and std::is_invocable_r_v<T, BinaryOp, T, T>
  void scatter_rev_end(std::span<const T> local_buffer, std::span<T> local_data,
                       F unpack_fn, BinaryOp op, std::span<MPI_Request> request)
  {
    assert(local_buffer.size() == _local_inds.size());
    if (!_local_inds.empty())
    {
      assert(*std::max_element(_local_inds.begin(), _local_inds.end())
             < std::int32_t(local_data.size()));
    }
    scatter_rev_end(request);
    unpack_fn(local_buffer, _local_inds, local_data, op);
  }
 
  template <typename T, typename BinaryOp>
  void scatter_rev(std::span<T> local_data, std::span<const T> remote_data,
                   BinaryOp op)
  {
    std::vector<T> local_buffer(local_buffer_size(), 0);
    std::vector<T> remote_buffer(remote_buffer_size(), 0);
    auto pack_fn = [](auto&& in, auto&& idx, auto&& out)
    {
      for (std::size_t i = 0; i < idx.size(); ++i)
        out[i] = in[idx[i]];
    };
    auto unpack_fn = [](auto&& in, auto&& idx, auto&& out, auto op)
    {
      for (std::size_t i = 0; i < idx.size(); ++i)
        out[idx[i]] = op(out[idx[i]], in[i]);
    };
    std::vector<MPI_Request> request(1, MPI_REQUEST_NULL);
    scatter_rev_begin(remote_data, std::span<T>(remote_buffer),
                      std::span<T>(local_buffer), pack_fn,
                      std::span<MPI_Request>(request));
    scatter_rev_end(std::span<const T>(local_buffer), local_data, unpack_fn, op,
                    std::span<MPI_Request>(request));
  }
 
  std::int32_t local_buffer_size() const noexcept { return _local_inds.size(); }
 
  std::int32_t remote_buffer_size() const noexcept
  {
    return _remote_inds.size();
  }
 
  const std::vector<std::int32_t>& local_indices() const noexcept
  {
    return _local_inds;
  }
 
  const std::vector<std::int32_t>& remote_indices() const noexcept
  {
    return _remote_inds;
  }
 
  int bs() const noexcept { return _bs; }
 
  std::vector<MPI_Request> create_request_vector(Scatterer::type type
                                                 = type::neighbor)
  {
    std::vector<MPI_Request> requests;
    switch (type)
    {
    case type::neighbor:
      requests = {MPI_REQUEST_NULL};
      break;
    case type::p2p:
      requests.resize(_dest.size() + _src.size(), MPI_REQUEST_NULL);
      break;
    default:
      throw std::runtime_error("Scatter::type not recognized");
    }
    return requests;
  }
 
private:
  // Block size
  int _bs;
 
  // Communicator where the source ranks own the indices in the callers
  // halo, and the destination ranks 'ghost' indices owned by the
  // caller. I.e.,
  // - in-edges (src) are from ranks that own my ghosts
  // - out-edges (dest) go to ranks that 'ghost' my owned indices
  dolfinx::MPI::Comm _comm0{MPI_COMM_NULL};
 
  // Communicator where the source ranks have ghost indices that are
  // owned by the caller, and the destination ranks are the owners of
  // indices in the callers halo region. I.e.,
  // - in-edges (src) are from ranks that 'ghost' my owned indices
  // - out-edges (dest) are to the owning ranks of my ghost indices
  dolfinx::MPI::Comm _comm1{MPI_COMM_NULL};
 
  // Permutation indices used to pack and unpack ghost data (remote)
  std::vector<std::int32_t, allocator_type> _remote_inds;
 
  // Number of remote indices (ghosts) for each neighbor process
  std::vector<int> _sizes_remote;
 
  // Displacements of remote data for mpi scatter and gather
  std::vector<int> _displs_remote;
 
  // Permutation indices used to pack and unpack local shared data
  // (owned indices that are shared with other processes). Indices are
  // grouped by neighbor process.
  std::vector<std::int32_t, allocator_type> _local_inds;
 
  // Number of local shared indices per neighbor process
  std::vector<int> _sizes_local;
 
  // Displacements of local data for mpi scatter and gather
  std::vector<int> _displs_local;
 
  // Set of ranks that own ghosts
  // FIXME: Should we store the index map instead?
  std::vector<int> _src;
 
  // Set of ranks ghost owned indices
  // FIXME: Should we store the index map instead?
  std::vector<int> _dest;
};
} // namespace dolfinx::common