libdune-common-dev 2.3.1-1 / usr / include / dune / common / parallel / collectivecommunication.hh

// -*- tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
// vi: set et ts=4 sw=2 sts=2:
#ifndef DUNE_COLLECTIVECOMMUNICATION_HH
#define DUNE_COLLECTIVECOMMUNICATION_HH
/*!
   \file
   \brief Implements an utility class that provides
   collective communication methods for sequential programs.

   \ingroup ParallelCommunication
 */
#include <iostream>
#include <complex>
#include <algorithm>

#include <dune/common/exceptions.hh>

/*! \defgroup ParallelCommunication Parallel Communication
   \ingroup Common

   \brief Abstractions for paralle computing

   Dune offers an abstraction to the basic methods of parallel
   communication. It allows to switch parallel features on and off,
   without changing the code. This is done using either CollectiveCommunication
   or MPICollectiveCommunication.

 */

/*!
   \file
   \brief An abstraction to the basic methods of parallel communication,
     following the message-passing paradigm.
   \ingroup ParallelCommunication
 */

namespace Dune
{

  /* define some type that definitely differs from MPI_Comm */
  struct No_Comm {};


  /*! @brief Collective communication interface and sequential default implementation

     CollectiveCommunication offers an abstraction to the basic methods
     of parallel communication, following the message-passing
     paradigm. It allows to switch parallel features on and off, without
     changing the code. Currently only MPI and sequential code are
     supported.

     A CollectiveCommunication object is returned by all grids (also
     the sequential ones) in order to allow code to be written in
     a transparent way for sequential and parallel grids.

     This class provides a default implementation for sequential grids.
     The number of processes involved is 1, any sum, maximum, etc. returns
     just its input argument and so on.

     In specializations one can implement the real thing using appropriate
     communication functions, e.g. there exists an implementation using
     the Message Passing %Interface (MPI), see Dune::CollectiveCommunication<MPI_Comm>.

     Moreover, the communication subsystem used by an implementation
     is not visible in the interface, i.e. Dune grid implementations
     are not restricted to MPI.

     \ingroup ParallelCommunication
   */
  template<typename C>
  class CollectiveCommunication
  {
  public:
    //! Construct default object
    CollectiveCommunication()
    {}
    CollectiveCommunication (const C&)
    {}

    //! Return rank, is between 0 and size()-1
    int rank () const
    {
      return 0;
    }

    //! Number of processes in set, is greater than 0
    int size () const
    {
      return 1;
    }

    /** @brief  Compute the sum of the argument over all processes and
            return the result in every process. Assumes that T has an operator+
     */
    template<typename T>
    T sum (T& in) const     // MPI does not know about const :-(
    {
      return in;
    }

    /** @brief Compute the sum over all processes for each component of an array and return the result
            in every process. Assumes that T has an operator+
     */
    template<typename T>
    int sum (T* inout, int len) const
    {
      return 0;
    }

    /** @brief  Compute the product of the argument over all processes and
            return the result in every process. Assumes that T has an operator*
     */
    template<typename T>
    T prod (T& in) const     // MPI does not know about const :-(
    {
      return in;
    }

    /** @brief Compute the product over all processes
            for each component of an array and return the result
            in every process. Assumes that T has an operator*
     */
    template<typename T>
    int prod (T* inout, int len) const
    {
      return 0;
    }

    /** @brief  Compute the minimum of the argument over all processes and
            return the result in every process. Assumes that T has an operator<
     */
    template<typename T>
    T min (T& in) const     // MPI does not know about const :-(
    {
      return in;
    }

    /** @brief Compute the minimum over all processes
            for each component of an array and return the result
            in every process. Assumes that T has an operator<
     */
    template<typename T>
    int min (T* inout, int len) const
    {
      return 0;
    }

    /** @brief  Compute the maximum of the argument over all processes and
            return the result in every process. Assumes that T has an operator<
     */
    template<typename T>
    T max (T& in) const     // MPI does not know about const :-(
    {
      return in;
    }

    /** @brief Compute the maximum over all processes
            for each component of an array and return the result
            in every process. Assumes that T has an operator<
     */
    template<typename T>
    int max (T* inout, int len) const
    {
      return 0;
    }

    /** @brief Wait until all processes have arrived at this point in the program.
     */
    int barrier () const
    {
      return 0;
    }

    /** @brief Distribute an array from the process with rank root to all other processes
     */
    template<typename T>
    int broadcast (T* inout, int len, int root) const
    {
      return 0;
    }

    /** @brief  Gather arrays on root task.
     *
     * Each process sends its in array of length len to the root process
     * (including the root itself). In the root process these arrays are stored in rank
     * order in the out array which must have size len * number of processes.
     * @param[in] in The send buffer with the data to send.
     * @param[out] out The buffer to store the received data in. Might have length zero on non-root
     *                  tasks.
     * @param[in] len The number of elements to send on each task.
     * @param[out] root The root task that gathers the data.
     */
    template<typename T>
    int gather (T* in, T* out, int len, int root) const     // note out must have same size as in
    {
      for (int i=0; i<len; i++)
        out[i] = in[i];
      return 0;
    }

    /** @brief Scatter array from a root to all other task.
     *
     * The root process sends the elements with index from k*len to (k+1)*len-1 in its array to
     * task k, which stores it at index 0 to len-1.
     * @param[in] send The array to scatter. Might have length zero on non-root
     *                  tasks.
     * @param[out] recv The buffer to store the received data in. Upon completion of the
     *                 method each task will have same data stored there as the one in
     *                 send buffer of the root task before.
     * @param[in] len The number of elements in the recv buffer.
     * @param[out] root The root task that gathers the data.
     */
    template<typename T>
    int scatter (T* send, T* recv, int len, int root) const // note out must have same size as in
    {
      for (int i=0; i<len; i++)
        recv[i] = send[i];
      return 0;
    }

    /**
     * @brief Gathers data from all tasks and distribute it to all.
     *
     * The block of data sent from the  jth  process  is  received  by  every
     *  process and placed in the jth block of the buffer recvbuf.
     *
     * @param[in] sbuf The buffer with the data to send. Has to be the same for
     *                 each task.
     * @param[in] count The number of elements to send by any process.
     * @param[out] rbuf The receive buffer for the data. Has to be of size
     *  notasks*count, with notasks being the number of tasks in the communicator.
     */
    template<typename T>
    int allgather(T* sbuf, int count, T* rbuf) const
    {
      for(T* end=sbuf+count; sbuf < end; ++sbuf, ++rbuf)
        *rbuf=*sbuf;
      return 0;
    }

    /**
     * @brief Compute something over all processes
     * for each component of an array and return the result
     * in every process.
     *
     * The template parameter BinaryFunction is the type of
     * the binary function to use for the computation
     *
     * @param inout The array to compute on.
     * @param len The number of components in the array
     */
    template<typename BinaryFunction, typename Type>
    int allreduce(Type* inout, int len) const
    {
      return 0;
    }

    /**
     * @brief Compute something over all processes
     * for each component of an array and return the result
     * in every process.
     *
     * The template parameter BinaryFunction is the type of
     * the binary function to use for the computation
     *
     * @param in The array to compute on.
     * @param out The array to store the results in.
     * @param len The number of components in the array
     */
    template<typename BinaryFunction, typename Type>
    void allreduce(Type* in, Type* out, int len) const
    {
      std::copy(in, in+len, out);
      return;
    }

  };
}

#endif