view release on metacpan or  search on metacpan

src/boost/mpi/nonblocking.hpp  view on Meta::CPAN

#include <boost/optional.hpp>
#include <utility> // for std::pair
#include <algorithm> // for iter_swap, reverse
#include <boost/static_assert.hpp>
#include <boost/mpi/request.hpp>
#include <boost/mpi/status.hpp>
#include <boost/mpi/exception.hpp>

namespace boost { namespace mpi {

/** 
 *  @brief Wait until any non-blocking request has completed.
 *
 *  This routine takes in a set of requests stored in the iterator
 *  range @c [first,last) and waits until any of these requests has
 *  been completed. It provides functionality equivalent to 
 *  @c MPI_Waitany.
 *
 *  @param first The iterator that denotes the beginning of the
 *  sequence of request objects.
 *
 *  @param last The iterator that denotes the end of the sequence of
 *  request objects. This may not be equal to @c first.
 *
 *  @returns A pair containing the status object that corresponds to
 *  the completed operation and the iterator referencing the completed
 *  request.
 */
template<typename ForwardIterator>
std::pair<status, ForwardIterator> 
wait_any(ForwardIterator first, ForwardIterator last)
{
  using std::advance;

  BOOST_ASSERT(first != last);
  
  typedef typename std::iterator_traits<ForwardIterator>::difference_type
    difference_type;

  bool all_trivial_requests = true;
  difference_type n = 0;
  ForwardIterator current = first;
  while (true) {
    // Check if we have found a completed request. If so, return it.
    if (optional<status> result = current->test())
      return std::make_pair(*result, current);

    // Check if this request (and all others before it) are "trivial"
    // requests, e.g., they can be represented with a single
    // MPI_Request.
    all_trivial_requests = 
      all_trivial_requests
      && !current->m_handler 
      && current->m_requests[1] == MPI_REQUEST_NULL;

    // Move to the next request.
    ++n;
    if (++current == last) {
      // We have reached the end of the list. If all requests thus far
      // have been trivial, we can call MPI_Waitany directly, because
      // it may be more efficient than our busy-wait semantics.
      if (all_trivial_requests) {
        std::vector<MPI_Request> requests;
        requests.reserve(n);
        for (current = first; current != last; ++current)
          requests.push_back(current->m_requests[0]);

        // Let MPI wait until one of these operations completes.
        int index;
        status stat;
        BOOST_MPI_CHECK_RESULT(MPI_Waitany, 
                               (n, &requests[0], &index, &stat.m_status));

        // We don't have a notion of empty requests or status objects,
        // so this is an error.
        if (index == MPI_UNDEFINED)
          boost::throw_exception(exception("MPI_Waitany", MPI_ERR_REQUEST));

        // Find the iterator corresponding to the completed request.
        current = first;
        advance(current, index);
        current->m_requests[0] = requests[index];
        return std::make_pair(stat, current);
      }

      // There are some nontrivial requests, so we must continue our
      // busy waiting loop.
      n = 0;
      current = first;
      all_trivial_requests = true;
    }
  }

  // We cannot ever get here
  BOOST_ASSERT(false);
}

/** 
 *  @brief Test whether any non-blocking request has completed.
 *
 *  This routine takes in a set of requests stored in the iterator
 *  range @c [first,last) and tests whether any of these requests has
 *  been completed. This routine is similar to @c wait_any, but will
 *  not block waiting for requests to completed. It provides
 *  functionality equivalent to @c MPI_Testany.
 *
 *  @param first The iterator that denotes the beginning of the
 *  sequence of request objects.
 *
 *  @param last The iterator that denotes the end of the sequence of
 *  request objects. 
 *
 *  @returns If any outstanding requests have completed, a pair
 *  containing the status object that corresponds to the completed
 *  operation and the iterator referencing the completed
 *  request. Otherwise, an empty @c optional<>.
 */
template<typename ForwardIterator>
optional<std::pair<status, ForwardIterator> >
test_any(ForwardIterator first, ForwardIterator last)
{
  for (ForwardIterator current = first; first != last; ++first) {
    // Check if we have found a completed request. If so, return it.
    if (optional<status> result = current->test())
      return std::make_pair(*result, current);
  }

  // We found nothing
  return optional<std::pair<status, ForwardIterator> >();
}

/** 
 *  @brief Wait until all non-blocking requests have completed.
 *
 *  This routine takes in a set of requests stored in the iterator
 *  range @c [first,last) and waits until all of these requests have
 *  been completed. It provides functionality equivalent to 
 *  @c MPI_Waitall.
 *
 *  @param first The iterator that denotes the beginning of the
 *  sequence of request objects.
 *
 *  @param last The iterator that denotes the end of the sequence of
 *  request objects. 
 *
 *  @param out If provided, an output iterator through which the
 *  status of each request will be emitted. The @c status objects are

src/boost/mpi/nonblocking.hpp  view on Meta::CPAN

  using std::advance;

  if (first == last)
    return std::make_pair(out, first);
  
  typedef typename std::iterator_traits<BidirectionalIterator>::difference_type
    difference_type;

  bool all_trivial_requests = true;
  difference_type n = 0;
  BidirectionalIterator current = first;
  BidirectionalIterator start_of_completed = last;
  while (true) {
    // Check if we have found a completed request. 
    if (optional<status> result = current->test()) {
      using std::iter_swap;

      // Emit the resulting status object
      *out++ = *result;

      // We're expanding the set of completed requests
      --start_of_completed;

      if (current == start_of_completed) {
        // If we have hit the end of the list of pending
        // requests. Finish up by fixing the order of the completed
        // set to match the order in which we emitted status objects,
        // then return.
        std::reverse(start_of_completed, last);
        return std::make_pair(out, start_of_completed);
      }

      // Swap the request we just completed with the last request that
      // has not yet been tested.
      iter_swap(current, start_of_completed);

      continue;
    }

    // Check if this request (and all others before it) are "trivial"
    // requests, e.g., they can be represented with a single
    // MPI_Request.
    all_trivial_requests = 
      all_trivial_requests
      && !current->m_handler 
      && current->m_requests[1] == MPI_REQUEST_NULL;

    // Move to the next request.
    ++n;
    if (++current == start_of_completed) {
      if (start_of_completed != last) {
        // We have satisfied some requests. Make the order of the
        // completed requests match that of the status objects we've
        // already emitted and we're done.
        std::reverse(start_of_completed, last);
        return std::make_pair(out, start_of_completed);
      }

      // We have reached the end of the list. If all requests thus far
      // have been trivial, we can call MPI_Waitsome directly, because
      // it may be more efficient than our busy-wait semantics.
      if (all_trivial_requests) {
        std::vector<MPI_Request> requests;
        std::vector<int> indices(n);
        std::vector<MPI_Status> stats(n);
        requests.reserve(n);
        for (current = first; current != last; ++current)
          requests.push_back(current->m_requests[0]);

        // Let MPI wait until some of these operations complete.
        int num_completed;
        BOOST_MPI_CHECK_RESULT(MPI_Waitsome, 
                               (n, &requests[0], &num_completed, &indices[0],
                                &stats[0]));

        // Translate the index-based result of MPI_Waitsome into a
        // partitioning on the requests.
        int current_offset = 0;
        current = first;
        for (int index = 0; index < num_completed; ++index, ++out) {
          using std::iter_swap;

          // Move "current" to the request object at this index
          advance(current, indices[index] - current_offset);
          current_offset = indices[index];

          // Emit the status object
          status stat;
          stat.m_status = stats[index];
          *out = stat;

          // Finish up the request and swap it into the "completed
          // requests" partition.
          current->m_requests[0] = requests[indices[index]];
          --start_of_completed;
          iter_swap(current, start_of_completed);
        }

        // We have satisfied some requests. Make the order of the
        // completed requests match that of the status objects we've
        // already emitted and we're done.
        std::reverse(start_of_completed, last);
        return std::make_pair(out, start_of_completed);
      }

      // There are some nontrivial requests, so we must continue our
      // busy waiting loop.
      n = 0;
      current = first;
    }
  }

  // We cannot ever get here
  BOOST_ASSERT(false);
}

/**
 *  \overload
 */
template<typename BidirectionalIterator>
BidirectionalIterator
wait_some(BidirectionalIterator first, BidirectionalIterator last)
{
  using std::advance;

  if (first == last)
    return first;
  
  typedef typename std::iterator_traits<BidirectionalIterator>::difference_type
    difference_type;

  bool all_trivial_requests = true;
  difference_type n = 0;
  BidirectionalIterator current = first;
  BidirectionalIterator start_of_completed = last;
  while (true) {
    // Check if we have found a completed request. 
    if (optional<status> result = current->test()) {
      using std::iter_swap;

      // We're expanding the set of completed requests
      --start_of_completed;

      // If we have hit the end of the list of pending requests, we're
      // done.
      if (current == start_of_completed)
        return start_of_completed;

      // Swap the request we just completed with the last request that
      // has not yet been tested.
      iter_swap(current, start_of_completed);

      continue;
    }

    // Check if this request (and all others before it) are "trivial"
    // requests, e.g., they can be represented with a single
    // MPI_Request.
    all_trivial_requests = 
      all_trivial_requests
      && !current->m_handler 
      && current->m_requests[1] == MPI_REQUEST_NULL;

    // Move to the next request.
    ++n;
    if (++current == start_of_completed) {
        // If we have satisfied some requests, we're done.
      if (start_of_completed != last)
        return start_of_completed;

      // We have reached the end of the list. If all requests thus far
      // have been trivial, we can call MPI_Waitsome directly, because
      // it may be more efficient than our busy-wait semantics.
      if (all_trivial_requests) {
        std::vector<MPI_Request> requests;
        std::vector<int> indices(n);
        requests.reserve(n);
        for (current = first; current != last; ++current)
          requests.push_back(current->m_requests[0]);

        // Let MPI wait until some of these operations complete.
        int num_completed;
        BOOST_MPI_CHECK_RESULT(MPI_Waitsome, 
                               (n, &requests[0], &num_completed, &indices[0],
                                MPI_STATUSES_IGNORE));

        // Translate the index-based result of MPI_Waitsome into a
        // partitioning on the requests.
        int current_offset = 0;
        current = first;
        for (int index = 0; index < num_completed; ++index) {
          using std::iter_swap;

          // Move "current" to the request object at this index
          advance(current, indices[index] - current_offset);
          current_offset = indices[index];

          // Finish up the request and swap it into the "completed
          // requests" partition.
          current->m_requests[0] = requests[indices[index]];
          --start_of_completed;
          iter_swap(current, start_of_completed);
        }

        // We have satisfied some requests, so we are done.
        return start_of_completed;
      }

      // There are some nontrivial requests, so we must continue our
      // busy waiting loop.
      n = 0;
      current = first;
    }
  }

  // We cannot ever get here
  BOOST_ASSERT(false);
}

/** 
 *  @brief Test whether some non-blocking requests have completed.
 *
 *  This routine takes in a set of requests stored in the iterator
 *  range @c [first,last) and tests to see if any of the requests has
 *  completed. It completes all of the requests it can, partitioning
 *  the input sequence into pending requests followed by completed
 *  requests. If an output iterator is provided, @c status objects
 *  will be emitted for each of the completed requests. This routine
 *  is similar to @c wait_some, but does not wait until any requests
 *  have completed. This routine provides functionality equivalent to
 *  @c MPI_Testsome.
 *
 *  @param first The iterator that denotes the beginning of the
 *  sequence of request objects.
 *
 *  @param last The iterator that denotes the end of the sequence of
 *  request objects. This may not be equal to @c first.
 *
 *  @param out If provided, the @c status objects corresponding to
 *  completed requests will be emitted through this output iterator.

 *  @returns If the @p out parameter was provided, a pair containing
 *  the output iterator @p out after all of the @c status objects have
 *  been written through it and an iterator referencing the first
 *  completed request. If no @p out parameter was provided, only the
 *  iterator referencing the first completed request will be emitted.
 */
template<typename BidirectionalIterator, typename OutputIterator>
std::pair<OutputIterator, BidirectionalIterator> 
test_some(BidirectionalIterator first, BidirectionalIterator last,
          OutputIterator out)
{
  BidirectionalIterator current = first;
  BidirectionalIterator start_of_completed = last;
  while (current != start_of_completed) {
    // Check if we have found a completed request. 
    if (optional<status> result = current->test()) {
      using std::iter_swap;

      // Emit the resulting status object
      *out++ = *result;

      // We're expanding the set of completed requests
      --start_of_completed;

      // Swap the request we just completed with the last request that
      // has not yet been tested.
      iter_swap(current, start_of_completed);

      continue;