Update of /cvsroot/boost/boost/boost/parallel/mpi/collectives
In directory 
sc8-pr-cvs3.sourceforge.net:/tmp/cvs-serv657/boost/parallel/mpi/collectives

Added Files:
        all_gather.hpp all_reduce.hpp all_to_all.hpp broadcast.hpp 
        gather.hpp reduce.hpp scan.hpp scatter.hpp 
Log Message:
Import Boost.MPI with the beginnings of a BBv2-based build system

--- NEW FILE: all_gather.hpp ---
// Copyright (C) 2005-2006 Douglas Gregor <doug.gregor -at- gmail.com>.

// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)

// Message Passing Interface 1.1 -- Section 4.7. Gather-to-all
#ifndef BOOST_PARALLEL_MPI_ALL_GATHER_HPP
#define BOOST_PARALLEL_MPI_ALL_GATHER_HPP

#include <boost/parallel/mpi/exception.hpp>
#include <boost/parallel/mpi/datatype.hpp>
#include <vector>
#include <boost/serialization/vector.hpp>

// all_gather falls back to gather+broadcast in some cases
#include <boost/parallel/mpi/collectives/broadcast.hpp>
#include <boost/parallel/mpi/collectives/gather.hpp>

namespace boost { namespace parallel { namespace mpi {

namespace detail {
  // We're all-gathering for a type that has an associated MPI
  // datatype, so we'll use MPI_Gather to do all of the work.
  template<typename T>
  void
  all_gather_impl(const communicator& comm, const T& value,
                  std::vector<T>& values, mpl::true_)
  {
    values.resize(comm.size());

    MPI_Datatype type = boost::parallel::mpi::get_mpi_datatype<T>();
    BOOST_MPI_CHECK_RESULT(MPI_Allgather,
                           (const_cast<T*>(&value), 1, type,
                            &values[0], 1, type, comm));
  }

  // We're all-gathering for a type that has no associated MPI. So,
  // we'll do a manual gather followed by a broadcast.
  template<typename T>
  void
  all_gather_impl(const communicator& comm, const T& value,
                  std::vector<T>& values, mpl::false_)
  {
    gather(comm, value, values, 0);
    broadcast(comm, values, 0);
  }
} // end namespace detail

template<typename T>
void
all_gather(const communicator& comm, const T& value, std::vector<T>& values)
{
  detail::all_gather_impl(comm, value, values, is_mpi_datatype<T>());
}

} } } // end namespace boost::parallel::mpi

#endif // BOOST_PARALLEL_MPI_ALL_GATHER_HPP

--- NEW FILE: all_reduce.hpp ---
// Copyright (C) 2005-2006 Douglas Gregor <doug.gregor -at- gmail.com>
// Copyright (C) 2004 The Trustees of Indiana University

// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)

//   Authors: Douglas Gregor
//            Andrew Lumsdaine

// Message Passing Interface 1.1 -- Section 4.9.1. Reduce
#ifndef BOOST_PARALLEL_MPI_ALL_REDUCE_HPP
#define BOOST_PARALLEL_MPI_ALL_REDUCE_HPP

// All-reduce falls back to reduce() + broadcast() in some cases.
#include <boost/parallel/mpi/collectives/broadcast.hpp>
#include <boost/parallel/mpi/collectives/reduce.hpp>

namespace boost { namespace parallel { namespace mpi {

namespace detail {
  /**********************************************************************
   * Simple reduction with MPI_Allreduce                                *
   **********************************************************************/
  // We are reducing for a type that has an associated MPI
  // datatype and operation, so we'll use MPI_Allreduce directly.
  template<typename T, typename Op>
  void
  all_reduce_impl(const communicator& comm, const T& in_value, T& out_value,
                  Op op, mpl::true_ /*is_mpi_op*/,
                  mpl::true_ /*is_mpi_datatype*/)
  {
    BOOST_MPI_CHECK_RESULT(MPI_Allreduce,
                           (const_cast<T*>(&in_value), &out_value, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            is_mpi_op<Op, T>::op(), comm));
  }

  /**********************************************************************
   * User-defined reduction with MPI_Allreduce                          *
   **********************************************************************/
  // We are reducing at the root for a type that has an associated MPI
  // datatype but with a custom operation. We'll use MPI_Reduce
  // directly, but we'll need to create an MPI_Op manually.
  template<typename T, typename Op>
  void
  all_reduce_impl(const communicator& comm, const T& in_value, T& out_value,
                  Op op, mpl::false_ /*is_mpi_op*/,
                  mpl::true_ /*is_mpi_datatype*/)
  {
    user_op<Op, T> mpi_op(op);
    BOOST_MPI_CHECK_RESULT(MPI_Allreduce,
                           (const_cast<T*>(&in_value), &out_value, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            mpi_op.get_mpi_op(), comm));
  }

  /**********************************************************************
   * User-defined, tree-based reduction for non-MPI data types          *
   **********************************************************************/
  // We are reducing at the root for a type that has no associated MPI
  // datatype and operation, so we'll use a simple tree-based
  // algorithm.
  template<typename T, typename Op>
  void
  all_reduce_impl(const communicator& comm, const T& in_value, T& out_value,
                  Op op, mpl::false_ /*is_mpi_op*/,
                  mpl::false_ /*is_mpi_datatype*/)
  {
    reduce(comm, in_value, out_value, op, 0);
    broadcast(comm, out_value, 0);
  }
} // end namespace detail


template<typename T, typename Op>
void
all_reduce(const communicator& comm, const T& in_value, T& out_value, Op op)
{
  detail::all_reduce_impl(comm, in_value, out_value, op,
                          is_mpi_op<Op, T>(), is_mpi_datatype<T>());
}

template<typename T, typename Op>
T all_reduce(const communicator& comm, const T& in_value, Op op)
{
  T result;
  all_reduce(comm, in_value, result, op);
  return result;
}

} } } // end namespace boost::parallel::mpi

#endif // BOOST_PARALLEL_MPI_ALL_REDUCE_HPP

--- NEW FILE: all_to_all.hpp ---
// Copyright (C) 2005, 2006 Douglas Gregor.

// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)

// Message Passing Interface 1.1 -- Section 4.8. All-to-all
#ifndef BOOST_PARALLEL_MPI_ALL_TO_ALL_HPP
#define BOOST_PARALLEL_MPI_ALL_TO_ALL_HPP

#include <boost/parallel/mpi/exception.hpp>
#include <boost/parallel/mpi/datatype.hpp>
#include <vector>
#include <boost/parallel/mpi/packed_oarchive.hpp>
#include <boost/parallel/mpi/packed_iarchive.hpp>
#include <boost/parallel/mpi/communicator.hpp>
#include <boost/parallel/mpi/environment.hpp>
#include <boost/assert.hpp>
#include <boost/parallel/mpi/collectives_fwd.hpp>
#include <boost/parallel/mpi/allocator.hpp>

namespace boost { namespace parallel { namespace mpi {


template<typename T>
void
all_to_all(const communicator& comm, const std::vector<T>& in_values,
           std::vector<T>& out_values);
           
namespace detail {
  // We're performaing an all-to-all with a type that has an
  // associated MPI datatype, so we'll use MPI_Alltoall to do all of
  // the work.
  template<typename T>
  void
  all_to_all_impl(const communicator& comm, const std::vector<T>& in_values,
                  std::vector<T>& out_values, mpl::true_)
  {
    MPI_Datatype type = get_mpi_datatype<T>();
    out_values.resize(comm.size());
    BOOST_MPI_CHECK_RESULT(MPI_Alltoall,
                           (const_cast<T*>(&in_values.front()), 1, type,
                            &out_values.front(), 1, type, comm));
  }

  // We're performing an all-to-all with a type that does not have an
  // associated MPI datatype, so we'll need to serialize
  // it. Unfortunately, this means that we cannot use MPI_Alltoall, so
  // we'll just have to send individual messages to the other
  // processes.
  template<typename T>
  void
  all_to_all_impl(const communicator& comm, const std::vector<T>& in_values,
                  std::vector<T>& out_values, mpl::false_)
  {
    int size = comm.size();
    int rank = comm.rank();

    out_values.resize(comm.size());

    // The amount of data to be sent to each process
    std::vector<int> send_sizes(size);

    // The displacements for each outgoing value.
    std::vector<int> send_disps(size);

    // The buffer that will store all of the outgoing values
    std::vector<char, allocator<char> > outgoing;

    // Pack the buffer with all of the outgoing values.
    for (int dest = 0; dest < size; ++dest) {
      // Keep track of the displacements
      send_disps[dest] = outgoing.size();

      // Our own value will never be transmitted, so don't pack it.
      if (dest != rank) {
        packed_oarchive oa(comm, outgoing);
        oa << in_values[dest];
      }

      // Keep track of the sizes
      send_sizes[dest] = outgoing.size() - send_disps[dest];
    }

    // Determine how much data each process will receive.
    std::vector<int> recv_sizes(size);
    all_to_all(comm, send_sizes, recv_sizes);

    // Prepare a buffer to receive the incoming data.
    std::vector<int> recv_disps(size);
    int sum = 0;
    for (int src = 0; src < size; ++src) {
      recv_disps[src] = sum;
      sum += recv_sizes[src];
    }
    std::vector<char, allocator<char> > incoming(sum > 0? sum : 1);

    // Transmit the actual data
    BOOST_MPI_CHECK_RESULT(MPI_Alltoallv,
                           (&outgoing[0], &send_sizes[0],
                            &send_disps[0], MPI_PACKED,
                            &incoming[0], &recv_sizes[0],
                            &recv_disps[0], MPI_PACKED,
                            comm));

    // Deserialize data from the iarchive
    for (int src = 0; src < size; ++src) {
      if (src == rank) out_values[src] = in_values[src];
      else {
        packed_iarchive ia(comm, incoming, boost::archive::no_header,
                           recv_disps[src]);
        ia >> out_values[src];
      }
    }
  }
} // end namespace detail

template<typename T>
void
all_to_all(const communicator& comm, const std::vector<T>& in_values,
           std::vector<T>& out_values)
{
  BOOST_ASSERT((int)in_values.size() == comm.size());
  detail::all_to_all_impl(comm, in_values, out_values, is_mpi_datatype<T>());
}

} } } // end namespace boost::parallel::mpi

#endif // BOOST_PARALLEL_MPI_ALL_TO_ALL_HPP

--- NEW FILE: broadcast.hpp ---
// Copyright (C) 2005, 2006 Douglas Gregor <doug.gregor -at- gmail.com>.

// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)

// Message Passing Interface 1.1 -- Section 4.4. Broadcast
#ifndef BOOST_PARALLEL_MPI_BROADCAST_HPP
#define BOOST_PARALLEL_MPI_BROADCAST_HPP

#include <boost/parallel/mpi/collectives_fwd.hpp>
#include <boost/parallel/mpi/exception.hpp>
#include <boost/parallel/mpi/datatype.hpp>
#include <boost/parallel/mpi/communicator.hpp>

namespace boost { namespace parallel { namespace mpi {
/************************************************************************
 * Specializations                                                      *
 ************************************************************************/

/**
 * INTERNAL ONLY
 */
template<>
void
broadcast<const packed_oarchive>(const communicator& comm,
                                 const packed_oarchive& oa,
                                 int root);

/**
 * INTERNAL ONLY
 */
template<>
void
broadcast<packed_oarchive>(const communicator& comm, packed_oarchive& oa,
                           int root);

/**
 * INTERNAL ONLY
 */
template<>
void
broadcast<packed_iarchive>(const communicator& comm, packed_iarchive& ia,
                           int root);

/**
 * INTERNAL ONLY
 */
template<>
void
broadcast<const packed_skeleton_oarchive>(const communicator& comm,
                                          const packed_skeleton_oarchive& oa,
                                          int root);

/**
 * INTERNAL ONLY
 */
template<>
void
broadcast<packed_skeleton_oarchive>(const communicator& comm,
                                    packed_skeleton_oarchive& oa, int root);

/**
 * INTERNAL ONLY
 */
template<>
void
broadcast<packed_skeleton_iarchive>(const communicator& comm,
                                    packed_skeleton_iarchive& ia, int root);

/**
 * INTERNAL ONLY
 */
template<>
void broadcast<content>(const communicator& comm, content& c, int root);

/**
 * INTERNAL ONLY
 */
template<>
void broadcast<const content>(const communicator& comm, const content& c,
                              int root);

/************************************************************************
 * broadcast() implementation                                           *
 ************************************************************************/
namespace detail {
  // We're sending a type that has an associated MPI datatype, so
  // we'll use MPI_Bcast to do all of the work.
  template<typename T>
  void broadcast_impl(const communicator& comm, T& value, int root, mpl::true_)
  {
    BOOST_MPI_CHECK_RESULT(MPI_Bcast,
                           (&value, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            root, MPI_Comm(comm)));
  }

  // We're sending a type that does not have an associated MPI
  // datatype, so we'll need to serialize it. Unfortunately, this
  // means that we cannot use MPI_Bcast, so we'll just send from the
  // root to everyone else.
  template<typename T>
  void
  broadcast_impl(const communicator& comm, T& value, int root, mpl::false_)
  {
    if (comm.rank() == root) {
      packed_oarchive oa(comm);
      oa << value;
      broadcast(comm, oa, root);
    } else {
      packed_iarchive ia(comm);
      broadcast(comm, ia, root);
      ia >> value;
    }
  }
} // end namespace detail

template<typename T>
void broadcast(const communicator& comm, T& value, int root = 0)
{
  detail::broadcast_impl(comm, value, root, is_mpi_datatype<T>());
}

} } } // end namespace boost::parallel::mpi

// If the user has already included skeleton_and_content.hpp, include
// the code to broadcast skeletons and content.
#ifdef BOOST_PARALLEL_MPI_SKELETON_AND_CONTENT_HPP
#  include <boost/parallel/mpi/detail/broadcast_sc.hpp>
#endif

#endif // BOOST_PARALLEL_MPI_BROADCAST_HPP

--- NEW FILE: gather.hpp ---
// Copyright (C) 2005, 2006 Douglas Gregor.

// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)

// Message Passing Interface 1.1 -- Section 4.5. Gather
#ifndef BOOST_PARALLEL_MPI_GATHER_HPP
#define BOOST_PARALLEL_MPI_GATHER_HPP

#include <boost/parallel/mpi/exception.hpp>
#include <boost/parallel/mpi/datatype.hpp>
#include <vector>
#include <boost/parallel/mpi/packed_oarchive.hpp>
#include <boost/parallel/mpi/packed_iarchive.hpp>
#include <boost/parallel/mpi/detail/point_to_point.hpp>
#include <boost/parallel/mpi/communicator.hpp>
#include <boost/parallel/mpi/environment.hpp>
#include <boost/assert.hpp>

namespace boost { namespace parallel { namespace mpi {

namespace detail {
  // We're gathering at the root for a type that has an associated MPI
  // datatype, so we'll use MPI_Gather to do all of the work.
  template<typename T>
  void
  gather_impl(const communicator& comm, const T& value, std::vector<T>& values,
              int root, mpl::true_)
  {
    values.resize(comm.size());

    MPI_Datatype type = get_mpi_datatype<T>();
    BOOST_MPI_CHECK_RESULT(MPI_Gather,
                           (const_cast<T*>(&value), 1, type,
                            &values[0], 1, type, root, comm));
  }

  // We're gathering from a non-root for a type that has an associated MPI
  // datatype, so we'll use MPI_Gather to do all of the work.
  template<typename T>
  void
  gather_impl(const communicator& comm, const T& value, int root, mpl::true_)
  {
    MPI_Datatype type = get_mpi_datatype<T>();
    BOOST_MPI_CHECK_RESULT(MPI_Gather,
                           (const_cast<T*>(&value), 1, type,
                            0, 1, type, root, comm));
  }

  // We're gathering at the root for a type that does not have an
  // associated MPI datatype, so we'll need to serialize
  // it. Unfortunately, this means that we cannot use MPI_Gather, so
  // we'll just have all of the non-root nodes send individual
  // messages to the root.
  template<typename T>
  void
  gather_impl(const communicator& comm, const T& value, std::vector<T>& values,
              int root, mpl::false_)
  {
    int tag = environment::collectives_tag();
    int size = comm.size();

    // We know we'll get "size" values
    values.clear();
    values.reserve(size);

    MPI_Status status;
    for (int src = 0; src < size; ++src) {
      if (src == root) {
        // Our own value will never be transmitted: just copy it.
        values.push_back(value);
      } else {
        // Receive archive
        packed_iarchive ia(comm);
        detail::packed_archive_recv(comm, src, tag, ia, status);

        // Unpack the data
        T value_from_src;
        ia >> value_from_src;

        // Save the result
        values.push_back(value_from_src);
      }
    }
  }

  // We're gathering at a non-root for a type that does not have an
  // associated MPI datatype, so we'll need to serialize
  // it. Unfortunately, this means that we cannot use MPI_Gather, so
  // we'll just have all of the non-root nodes send individual
  // messages to the root.
  template<typename T>
  void
  gather_impl(const communicator& comm, const T& value, int root, mpl::false_)
  {
    int tag = environment::collectives_tag();

    packed_oarchive oa(comm);
    oa << value;
    detail::packed_archive_send(comm, root, tag, oa);
  }
} // end namespace detail

template<typename T>
void
gather(const communicator& comm, const T& value, std::vector<T>& values,
       int root)
{
  if (comm.rank() == root)
    detail::gather_impl(comm, value, values, root, is_mpi_datatype<T>());
  else
    detail::gather_impl(comm, value, root, is_mpi_datatype<T>());
}

template<typename T>
void gather(const communicator& comm, const T& value, int root)
{
  BOOST_ASSERT(comm.rank() != root);
  detail::gather_impl(comm, value, root, is_mpi_datatype<T>());
}

} } } // end namespace boost::parallel::mpi

#endif // BOOST_PARALLEL_MPI_GATHER_HPP

--- NEW FILE: reduce.hpp ---
// Copyright (C) 2005-2006 Douglas Gregor <[EMAIL PROTECTED]>.
// Copyright (C) 2004 The Trustees of Indiana University

// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)

//   Authors: Douglas Gregor
//            Andrew Lumsdaine

// Message Passing Interface 1.1 -- Section 4.9.1. Reduce
#ifndef BOOST_PARALLEL_MPI_REDUCE_HPP
#define BOOST_PARALLEL_MPI_REDUCE_HPP

#include <boost/parallel/mpi/exception.hpp>
#include <boost/parallel/mpi/datatype.hpp>

// For (de-)serializing sends and receives
#include <boost/parallel/mpi/packed_oarchive.hpp>
#include <boost/parallel/mpi/packed_iarchive.hpp>

// For packed_[io]archive sends and receives
#include <boost/parallel/mpi/detail/point_to_point.hpp>

#include <boost/parallel/mpi/communicator.hpp>
#include <boost/parallel/mpi/environment.hpp>
#include <boost/parallel/mpi/detail/computation_tree.hpp>
#include <boost/parallel/mpi/operations.hpp>
#include <algorithm>
#include <exception>
#include <boost/assert.hpp>

namespace boost { namespace parallel { namespace mpi {


/************************************************************************
 * Implementation details                                               *
 ************************************************************************/
namespace detail {
  /**********************************************************************
   * Simple reduction with MPI_Reduce                                   *
   **********************************************************************/
  // We are reducing at the root for a type that has an associated MPI
  // datatype and operation, so we'll use MPI_Reduce directly.
  template<typename T, typename Op>
  void
  reduce_impl(const communicator& comm, const T& in_value, T& out_value, Op op,
              int root, mpl::true_ /*is_mpi_op*/,
              mpl::true_/*is_mpi_datatype*/)
  {
    BOOST_MPI_CHECK_RESULT(MPI_Reduce,
                           (const_cast<T*>(&in_value), &out_value, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            is_mpi_op<Op, T>::op(), root, comm));
  }

  // We are reducing to the root for a type that has an associated MPI
  // datatype and operation, so we'll use MPI_Reduce directly.
  template<typename T, typename Op>
  void
  reduce_impl(const communicator& comm, const T& in_value, Op op, int root,
              mpl::true_ /*is_mpi_op*/, mpl::true_/*is_mpi_datatype*/)
  {
    BOOST_MPI_CHECK_RESULT(MPI_Reduce,
                           (const_cast<T*>(&in_value), 0, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            is_mpi_op<Op, T>::op(), root, comm));
  }

  // We are reducing at the root for a type that has an associated MPI
  // datatype but with a custom operation. We'll use MPI_Reduce
  // directly, but we'll need to create an MPI_Op manually.
  template<typename T, typename Op>
  void
  reduce_impl(const communicator& comm, const T& in_value, T& out_value,
              Op op, int root, mpl::false_ /*is_mpi_op*/,
              mpl::true_/*is_mpi_datatype*/)
  {
    user_op<Op, T> mpi_op(op);
    BOOST_MPI_CHECK_RESULT(MPI_Reduce,
                           (const_cast<T*>(&in_value), &out_value, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            mpi_op.get_mpi_op(), root, comm));
  }

  /**********************************************************************
   * User-defined reduction with MPI_Reduce                             *
   **********************************************************************/

  // We are reducing to the root for a type that has an associated MPI
  // datatype but with a custom operation. We'll use MPI_Reduce
  // directly, but we'll need to create an MPI_Op manually.
  template<typename T, typename Op>
  void
  reduce_impl(const communicator& comm, const T& in_value, Op op, int root,
              mpl::false_ /*is_mpi_op*/, mpl::true_/*is_mpi_datatype*/)
  {
    user_op<Op, T> mpi_op(op);
    BOOST_MPI_CHECK_RESULT(MPI_Reduce,
                           (const_cast<T*>(&in_value), 0, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            mpi_op.get_mpi_op(), root, comm));
  }

  /**********************************************************************
   * User-defined, tree-based reduction for non-MPI data types          *
   **********************************************************************/

  // Commutative reduction
  template<typename T, typename Op>
  void
  tree_reduce_impl(const communicator& comm, const T& in_value, T& out_value,
                   Op op, int root, mpl::true_ /*is_commutative*/)
  {
    out_value = in_value;

    int size = comm.size();
    int rank = comm.rank();

    // The computation tree we will use.
    detail::computation_tree tree(rank, size, root);

    int tag = environment::collectives_tag();

    MPI_Status status;
    int children = 0;
    for (int child = tree.child_begin();
         children < tree.branching_factor() && child != root;
         ++children, child = (child + 1) % size) {
      // Receive archive
      packed_iarchive ia(comm);
      detail::packed_archive_recv(comm, child, tag, ia, status);

      T incoming;
      ia >> incoming;

      out_value = op(out_value, incoming);
    }

    // For non-roots, send the result to the parent.
    if (tree.parent() != rank) {
      packed_oarchive oa(comm);
      oa << out_value;
      detail::packed_archive_send(comm, tree.parent(), tag, oa);
    }
  }

  // Commutative reduction from a non-root.
  template<typename T, typename Op>
  void
  tree_reduce_impl(const communicator& comm, const T& in_value, Op op,
                   int root, mpl::true_ /*is_commutative*/)
  {
    T result;
    detail::tree_reduce_impl(comm, in_value, result, op, root, mpl::true_());
  }

  // Non-commutative reduction
  template<typename T, typename Op>
  void
  tree_reduce_impl(const communicator& comm, const T& in_value, T& out_value,
                   Op op, int root, mpl::false_ /*is_commutative*/)
  {
    int tag = environment::collectives_tag();

    int left_child = root / 2;
    int right_child = (root + comm.size()) / 2;

    MPI_Status status;
    if (left_child != root) {
      // Receive value from the left child and merge it with the value
      // we had incoming.
      packed_iarchive ia(comm);
      detail::packed_archive_recv(comm, left_child, tag, ia, status);
      T incoming;
      ia >> incoming;
      out_value = op(incoming, in_value);
    } else {
      // There was no left value, so copy our incoming value.
      out_value = in_value;
    }

    if (right_child != root) {
      // Receive value from the right child and merge it with the
      // value we had incoming.
      packed_iarchive ia(comm);
      detail::packed_archive_recv(comm, right_child, tag, ia, status);
      T incoming;
      ia >> incoming;
      out_value = op(out_value, incoming);
    }
  }

  // Non-commutative reduction from a non-root.
  template<typename T, typename Op>
  void
  tree_reduce_impl(const communicator& comm, const T& in_value, Op op,
                   int root, mpl::false_ /*is_commutative*/)
  {
    int size = comm.size();
    int rank = comm.rank();

    int tag = environment::collectives_tag();

    // Determine our parents and children in the commutative binary
    // computation tree.
    int grandparent = root;
    int parent = root;
    int left_bound = 0;
    int right_bound = size;
    int left_child, right_child;
    do {
      left_child = (left_bound + parent) / 2;
      right_child = (parent + right_bound) / 2;

      if (rank < parent) {
        // Go left.
        grandparent = parent;
        right_bound = parent;
        parent = left_child;
      } else if (rank > parent) {
        // Go right.
        grandparent = parent;
        left_bound = parent + 1;
        parent = right_child;
      } else {
        // We've found the parent/
        break;
      }
    } while (true);

    // Our parent is the grandparent of our children. This is a slight
    // abuse of notation, but it makes the send-to-parent below make
    // more sense.
    parent = grandparent;

    MPI_Status status;
    T out_value;
    if (left_child != rank) {
      // Receive value from the left child and merge it with the value
      // we had incoming.
      packed_iarchive ia(comm);
      detail::packed_archive_recv(comm, left_child, tag, ia, status);
      T incoming;
      ia >> incoming;
      out_value = op(incoming, in_value);
    } else {
      // There was no left value, so copy our incoming value.
      out_value = in_value;
    }

    if (right_child != rank) {
      // Receive value from the right child and merge it with the
      // value we had incoming.
      packed_iarchive ia(comm);
      detail::packed_archive_recv(comm, right_child, tag, ia, status);
      T incoming;
      ia >> incoming;
      out_value = op(out_value, incoming);
    }

    // Send the combined value to our parent.
    packed_oarchive oa(comm);
    oa << out_value;
    detail::packed_archive_send(comm, parent, tag, oa);
  }

  // We are reducing at the root for a type that has no associated MPI
  // datatype and operation, so we'll use a simple tree-based
  // algorithm.
  template<typename T, typename Op>
  void
  reduce_impl(const communicator& comm, const T& in_value, T& out_value, Op op,
              int root, mpl::false_ /*is_mpi_op*/,
              mpl::false_/*is_mpi_datatype*/)
  {
    detail::tree_reduce_impl(comm, in_value, out_value, op, root,
                             is_commutative<Op, T>());
  }

  // We are reducing to the root for a type that has no associated MPI
  // datatype and operation, so we'll use a simple tree-based
  // algorithm.
  template<typename T, typename Op>
  void
  reduce_impl(const communicator& comm, const T& in_value, Op op, int root,
              mpl::false_ /*is_mpi_op*/, mpl::false_/*is_mpi_datatype*/)
  {
    detail::tree_reduce_impl(comm, in_value, op, root,
                             is_commutative<Op, T>());
  }
} // end namespace detail


template<typename T, typename Op>
void
reduce(const communicator& comm, const T& in_value, T& out_value, Op op,
       int root)
{
  if (comm.rank() == root)
    detail::reduce_impl(comm, in_value, out_value, op, root,
                        is_mpi_op<Op, T>(), is_mpi_datatype<T>());
  else
    detail::reduce_impl(comm, in_value, op, root,
                        is_mpi_op<Op, T>(), is_mpi_datatype<T>());
}

template<typename T, typename Op>
void reduce(const communicator& comm, const T& in_value, Op op, int root)
{
  BOOST_ASSERT(comm.rank() != root);

  detail::reduce_impl(comm, in_value, op, root,
                      is_mpi_op<Op, T>(), is_mpi_datatype<T>());
}

} } } // end namespace boost::parallel::mpi

#endif // BOOST_PARALLEL_MPI_REDUCE_HPP

--- NEW FILE: scan.hpp ---
// Copyright (C) 2005-2006 Douglas Gregor <[EMAIL PROTECTED]>.
// Copyright (C) 2004 The Trustees of Indiana University

// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)

//   Authors: Douglas Gregor
//            Andrew Lumsdaine

// Message Passing Interface 1.1 -- Section 4.9.1. Scan
#ifndef BOOST_PARALLEL_MPI_SCAN_HPP
#define BOOST_PARALLEL_MPI_SCAN_HPP

#include <boost/parallel/mpi/exception.hpp>
#include <boost/parallel/mpi/datatype.hpp>

// For (de-)serializing sends and receives
#include <boost/parallel/mpi/packed_oarchive.hpp>
#include <boost/parallel/mpi/packed_iarchive.hpp>

// For packed_[io]archive sends and receives
#include <boost/parallel/mpi/detail/point_to_point.hpp>

#include <boost/parallel/mpi/communicator.hpp>
#include <boost/parallel/mpi/environment.hpp>
#include <boost/parallel/mpi/detail/computation_tree.hpp>
#include <boost/parallel/mpi/operations.hpp>
#include <algorithm>
#include <exception>
#include <boost/assert.hpp>

namespace boost { namespace parallel { namespace mpi {


/************************************************************************
 * Implementation details                                               *
 ************************************************************************/
namespace detail {
  /**********************************************************************
   * Simple prefix reduction with MPI_Scan                              *
   **********************************************************************/

  // We are performing prefix reduction for a type that has an
  // associated MPI datatype and operation, so we'll use MPI_Scan
  // directly.
  template<typename T, typename Op>
  void
  scan_impl(const communicator& comm, const T& in_value, T& out_value, Op op,
            mpl::true_ /*is_mpi_op*/, mpl::true_ /*is_mpi_datatype*/)
  {
    BOOST_MPI_CHECK_RESULT(MPI_Scan,
                           (const_cast<T*>(&in_value), &out_value, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            is_mpi_op<Op, T>::op(), comm));
  }

  /**********************************************************************
   * User-defined prefix reduction with MPI_Scan                        *
   **********************************************************************/

  // We are performing prefix reduction for a type that has an
  // associated MPI datatype but with a custom operation. We'll use
  // MPI_Scan directly, but we'll need to create an MPI_Op manually.
  template<typename T, typename Op>
  void
  scan_impl(const communicator& comm, const T& in_value, T& out_value,
            Op op, mpl::false_ /*is_mpi_op*/, mpl::true_ /*is_mpi_datatype*/)
  {
    user_op<Op, T> mpi_op(op);
    BOOST_MPI_CHECK_RESULT(MPI_Scan,
                           (const_cast<T*>(&in_value), &out_value, 1,
                            boost::parallel::mpi::get_mpi_datatype<T>(),
                            mpi_op.get_mpi_op(), comm));
  }

  /**********************************************************************
   * User-defined, tree-based reduction for non-MPI data types          *
   **********************************************************************/

  template<typename T, typename Op>
  void
  upper_lower_scan(const communicator& comm, const T& in_value, 
                   T& out_value, Op& op, int lower, int upper)
  {
    int tag = environment::collectives_tag();
    int rank = comm.rank();

    if (lower + 1 == upper) {
      out_value = in_value;
    } else {
      int middle = (lower + upper) / 2;
      
      if (rank < middle) {
        // Lower half
        upper_lower_scan(comm, in_value, out_value, op, lower, middle);

        // If we're the last process in the lower half, send our value
        // to everyone in the upper half.
        if (rank == middle - 1) {
          for (int p = middle; p < upper; ++p)
            comm.send(p, tag, out_value);
        }
      } else {
        // Upper half
        upper_lower_scan(comm, in_value, out_value, op, middle, upper);

        // Receive value from the last process in the lower half.
        T left_value;
        comm.recv(middle - 1, tag, left_value);

        // Combine value that came from the left with our value
        out_value = op(left_value, out_value);
      }
    }
  }

  // We are performing prefix reduction for a type that has no
  // associated MPI datatype and operation, so we'll use a simple
  // upper/lower algorithm.
  template<typename T, typename Op>
  inline void
  scan_impl(const communicator& comm, const T& in_value, T& out_value, Op op,
            mpl::false_ /*is_mpi_op*/, mpl::false_/*is_mpi_datatype*/)
  {
    upper_lower_scan(comm, in_value, out_value, op, 0, comm.size());
  }
} // end namespace detail


template<typename T, typename Op>
inline void
scan(const communicator& comm, const T& in_value, T& out_value, Op op)
{
  detail::scan_impl(comm, in_value, out_value, op, 
                    is_mpi_op<Op, T>(), is_mpi_datatype<T>());
}

template<typename T, typename Op>
inline T
scan(const communicator& comm, const T& in_value, Op op)
{
  T out_value;
  detail::scan_impl(comm, in_value, out_value, op, 
                    is_mpi_op<Op, T>(), is_mpi_datatype<T>());
  return out_value;
}

} } } // end namespace boost::parallel::mpi

#endif // BOOST_PARALLEL_MPI_SCAN_HPP

--- NEW FILE: scatter.hpp ---
// Copyright (C) 2005, 2006 Douglas Gregor.

// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)

// Message Passing Interface 1.1 -- Section 4.6. Scatter
#ifndef BOOST_PARALLEL_MPI_SCATTER_HPP
#define BOOST_PARALLEL_MPI_SCATTER_HPP

#include <boost/parallel/mpi/exception.hpp>
#include <boost/parallel/mpi/datatype.hpp>
#include <vector>
#include <boost/parallel/mpi/packed_oarchive.hpp>
#include <boost/parallel/mpi/packed_iarchive.hpp>
#include <boost/parallel/mpi/detail/point_to_point.hpp>
#include <boost/parallel/mpi/communicator.hpp>
#include <boost/parallel/mpi/environment.hpp>
#include <boost/assert.hpp>

namespace boost { namespace parallel { namespace mpi {

namespace detail {
  // We're scattering from the root for a type that has an associated MPI
  // datatype, so we'll use MPI_Scatter to do all of the work.
  template<typename T>
  void
  scatter_impl(const communicator& comm, const std::vector<T>& values,
               T& value, int root, mpl::true_)
  {
    MPI_Datatype type = get_mpi_datatype<T>();
    BOOST_MPI_CHECK_RESULT(MPI_Scatter,
                           (const_cast<T*>(&values.front()), 1, type,
                            &value, 1, type, root, comm));
  }

  // We're scattering from a non-root for a type that has an associated MPI
  // datatype, so we'll use MPI_Scatter to do all of the work.
  template<typename T>
  void
  scatter_impl(const communicator& comm, T& value, int root, mpl::true_)
  {
    MPI_Datatype type = get_mpi_datatype<T>();
    BOOST_MPI_CHECK_RESULT(MPI_Scatter,
                           (0, 1, type,
                            &value, 1, type,
                            root, comm));
  }

  // We're scattering from the root for a type that does not have an
  // associated MPI datatype, so we'll need to serialize
  // it. Unfortunately, this means that we cannot use MPI_Scatter, so
  // we'll just have the root send individual messages to the other
  // processes.
  template<typename T>
  void
  scatter_impl(const communicator& comm, const std::vector<T>& values,
               T& value, int root, mpl::false_)
  {
    int tag = environment::collectives_tag();
    int size = comm.size();

    for (int dest = 0; dest < size; ++dest) {
      if (dest == root) {
        // Our own value will never be transmitted: just copy it.
        value = values[dest];
      } else {
        // Send archive
        packed_oarchive oa(comm);
        oa << values[dest];
        detail::packed_archive_send(comm, dest, tag, oa);
      }
    }
  }

  // We're scattering to a non-root for a type that does not have an
  // associated MPI datatype, so we'll need to de-serialize
  // it. Unfortunately, this means that we cannot use MPI_Scatter, so
  // we'll just have all of the non-root nodes send individual
  // messages to the root.
  template<typename T>
  void
  scatter_impl(const communicator& comm, T& value, int root, mpl::false_)
  {
    int tag = environment::collectives_tag();

    packed_iarchive ia(comm);
    MPI_Status status;
    detail::packed_archive_recv(comm, root, tag, ia, status);
    ia >> value;
  }
} // end namespace detail

template<typename T>
void
scatter(const communicator& comm, const std::vector<T>& values, T& value,
        int root)
{
  if (comm.rank() == root)
    detail::scatter_impl(comm, values, value, root, is_mpi_datatype<T>());
  else
    detail::scatter_impl(comm, value, root, is_mpi_datatype<T>());
}

template<typename T>
void scatter(const communicator& comm, T& value, int root)
{
  BOOST_ASSERT(comm.rank() != root);
  detail::scatter_impl(comm, value, root, is_mpi_datatype<T>());
}

} } } // end namespace boost::parallel::mpi

#endif // BOOST_PARALLEL_MPI_SCATTER_HPP


-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys - and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV
_______________________________________________
Boost-cvs mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/boost-cvs

Reply via email to