gather.c

mpi并行计算的c++代码 可用vc或gcc编译通过 可以用来搭建并行计算试验环境

/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
 *
 *  (C) 2001 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */

#include "mpiimpl.h"

/* -- Begin Profiling Symbol Block for routine MPI_Gather */
#if defined(HAVE_PRAGMA_WEAK)
#pragma weak MPI_Gather = PMPI_Gather
#elif defined(HAVE_PRAGMA_HP_SEC_DEF)
#pragma _HP_SECONDARY_DEF PMPI_Gather  MPI_Gather
#elif defined(HAVE_PRAGMA_CRI_DUP)
#pragma _CRI duplicate MPI_Gather as PMPI_Gather
#endif
/* -- End Profiling Symbol Block */

/* Define MPICH_MPI_FROM_PMPI if weak symbols are not supported to build
   the MPI routines */
#ifndef MPICH_MPI_FROM_PMPI
#define MPI_Gather PMPI_Gather
/* This is the default implementation of gather. The algorithm is:
   
   Algorithm: MPI_Gather

   We use a binomial tree algorithm for both short and
   long messages. At nodes other than leaf nodes we need to allocate
   a temporary buffer to store the incoming message. If the root is
   not rank 0, we receive data in a temporary buffer on the root and
   then reorder it into the right order. In the heterogeneous case
   we first pack the buffers by using MPI_Pack and then do the gather. 

   Cost = lgp.alpha + n.((p-1)/p).beta
   where n is the total size of the data gathered at the root.

   Possible improvements: 

   End Algorithm: MPI_Gather
*/

/* not declared static because it is called in intercomm. allgather */
/* begin:nested */
int MPIR_Gather ( 
	void *sendbuf, 
	int sendcnt, 
	MPI_Datatype sendtype, 
	void *recvbuf, 
	int recvcnt, 
	MPI_Datatype recvtype, 
	int root, 
	MPID_Comm *comm_ptr )
{
    static const char FCNAME[] = "MPIR_Gather";
    int        comm_size, rank;
    int        mpi_errno = MPI_SUCCESS;
    int curr_cnt=0, relative_rank, nbytes, recv_size, is_homogeneous;
    int mask, sendtype_size, recvtype_size, src, dst, position;
#ifdef MPID_HAS_HETERO
    int tmp_buf_size;
#endif
    void *tmp_buf=NULL;
    MPI_Status status;
    MPI_Aint   extent=0;            /* Datatype extent */
    MPI_Comm comm;
    
    comm = comm_ptr->handle;
    comm_size = comm_ptr->local_size;
    rank = comm_ptr->rank;

    if ( ((rank == root) && (recvcnt == 0)) ||
         ((rank != root) && (sendcnt == 0)) )
        return MPI_SUCCESS;

    is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
    if (comm_ptr->is_hetero)
        is_homogeneous = 0;
#endif

    /* check if multiple threads are calling this collective function */
    MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
    
/* Use binomial tree algorithm. */
    
    relative_rank = (rank >= root) ? rank - root : rank - root + comm_size;
    
    if (rank == root) 
        MPID_Datatype_get_extent_macro(recvtype, extent);
    
    if (is_homogeneous)
    {
        /* communicator is homogeneous. no need to pack buffer. */

        if (rank == root)
	{
            MPID_Datatype_get_size_macro(recvtype, recvtype_size);
            nbytes = recvtype_size * recvcnt;
        }
        else
	{
            MPID_Datatype_get_size_macro(sendtype, sendtype_size);
            nbytes = sendtype_size * sendcnt;
        }

        if (rank == root)
	{
            if (root != 0)
	    {
                /* allocate temporary buffer to receive data because it
                   will not be in the right order. We will need to
                   reorder it into the recv_buf. */
                tmp_buf = MPIU_Malloc(nbytes*comm_size);
		/* --BEGIN ERROR HANDLING-- */
                if (!tmp_buf)
		{
                    mpi_errno = MPIR_Err_create_code( MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem", 0 );
                    return mpi_errno;
                }
		/* --END ERROR HANDLING-- */

                if (sendbuf != MPI_IN_PLACE)
		{
                    /* copy root's sendbuf into tmpbuf just so that it is
                       easier to unpack everything later into the recv_buf */
                    mpi_errno = MPIR_Localcopy(sendbuf, sendcnt, sendtype,
                                               tmp_buf, nbytes, MPI_BYTE);
		    /* --BEGIN ERROR HANDLING-- */
                    if (mpi_errno)
		    {
			mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
			return mpi_errno;
		    }
		    /* --END ERROR HANDLING-- */
                }
                curr_cnt = nbytes;
            }
            else
	    {
                /* root is 0. no tmp_buf needed at root. */
                /* copy root's sendbuf into recvbuf */
                if (sendbuf != MPI_IN_PLACE)
		{
                    mpi_errno = MPIR_Localcopy(sendbuf, sendcnt, sendtype,
                                               recvbuf, recvcnt, recvtype);
		    /* --BEGIN ERROR HANDLING-- */
                    if (mpi_errno)
		    {
			mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
			return mpi_errno;
		    }
		    /* --END ERROR HANDLING-- */
                }
                curr_cnt = recvcnt;
            }          
        }
        else if (!(relative_rank % 2))
	{
            /* allocate temporary buffer for nodes other than leaf
               nodes. max size needed is (nbytes*comm_size)/2. */
            tmp_buf = MPIU_Malloc((nbytes*comm_size)/2);
	    /* --BEGIN ERROR HANDLING-- */
            if (!tmp_buf)
	    {
                mpi_errno = MPIR_Err_create_code( MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem", 0 );
                return mpi_errno;
            }
	    /* --END ERROR HANDLING-- */
            /* copy from sendbuf into tmp_buf */
            mpi_errno = MPIR_Localcopy(sendbuf, sendcnt, sendtype,
                                       tmp_buf, nbytes, MPI_BYTE);
	    /* --BEGIN ERROR HANDLING-- */
            if (mpi_errno)
	    {
		mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
		return mpi_errno;
	    }
	    /* --END ERROR HANDLING-- */
            curr_cnt = nbytes;
        }
        
        mask = 0x1;
        while (mask < comm_size)
	{
            if ((mask & relative_rank) == 0)
	    {
                src = relative_rank | mask;
                if (src < comm_size)
		{
                    src = (src + root) % comm_size;
                    if ((rank == root) && (root == 0))
		    {
                        /* root is 0. Receive directly into recvbuf */
                        mpi_errno = MPIC_Recv(((char *)recvbuf + 
                                               src*recvcnt*extent), 
                                              recvcnt*mask, recvtype, src,
                                              MPIR_GATHER_TAG, comm, 
                                              &status);
			/* --BEGIN ERROR HANDLING-- */
                        if (mpi_errno)
			{
			    mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
			    return mpi_errno;
			}
			/* --END ERROR HANDLING-- */
                    }
                    else
		    {
                        /* intermediate nodes or nonzero root. store in
                           tmp_buf */
                        mpi_errno = MPIC_Recv(((char *)tmp_buf + curr_cnt), 
                                              mask*nbytes, MPI_BYTE, src,
                                              MPIR_GATHER_TAG, comm, 
                                              &status);
			/* --BEGIN ERROR HANDLING-- */
                        if (mpi_errno)
			{
			    mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
			    return mpi_errno;
			}
			/* --END ERROR HANDLING-- */
                        /* the recv size is larger than what may be sent in
                           some cases. query amount of data actually received */
                        NMPI_Get_count(&status, MPI_BYTE, &recv_size);
                        curr_cnt += recv_size;
                    }
                }
            }
            else
	    {
                dst = relative_rank ^ mask;
                dst = (dst + root) % comm_size;
                if (relative_rank % 2)
		{
                    /* leaf nodes send directly from sendbuf */
                    mpi_errno = MPIC_Send(sendbuf, sendcnt, sendtype, dst,
                                          MPIR_GATHER_TAG, comm); 
		    /* --BEGIN ERROR HANDLING-- */
                    if (mpi_errno)
		    {
			mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
			return mpi_errno;
		    }
		    /* --END ERROR HANDLING-- */
                }
                else
		{
                    mpi_errno = MPIC_Send(tmp_buf, curr_cnt, MPI_BYTE, dst,
                                          MPIR_GATHER_TAG, comm); 
		    /* --BEGIN ERROR HANDLING-- */
                    if (mpi_errno)
		    {
			mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
			return mpi_errno;
		    }
		    /* --END ERROR HANDLING-- */
                }
                break;
            }
            mask <<= 1;
        }
        
        if ((rank == root) && (root != 0))
	{
            /* reorder and copy from tmp_buf into recvbuf */
            position = 0;

            if (sendbuf != MPI_IN_PLACE)
	    {
                MPIR_Localcopy(tmp_buf, nbytes*(comm_size-rank), MPI_BYTE, 
                               ((char *) recvbuf + extent*recvcnt*rank),
                               recvcnt*(comm_size-rank), recvtype);
            }
            else
	    {
                MPIR_Localcopy((char *) tmp_buf + nbytes,
                               nbytes*(comm_size-rank-1), MPI_BYTE,  
                               ((char *) recvbuf + extent*recvcnt*(rank+1)),
                               recvcnt*(comm_size-rank-1), recvtype);
            }
            MPIR_Localcopy((char *) tmp_buf + nbytes*(comm_size-rank),
                           nbytes*rank, MPI_BYTE, recvbuf, 
                           recvcnt*rank, recvtype); 

            MPIU_Free(tmp_buf);
        }
        else if (relative_rank && !(relative_rank % 2))
            MPIU_Free(tmp_buf);
    }
    
#ifdef MPID_HAS_HETERO
    else
    { /* communicator is heterogeneous. pack data into tmp_buf. */
        if (rank == root)
            NMPI_Pack_size(recvcnt*comm_size, recvtype, comm,
                           &tmp_buf_size); 
        else
            NMPI_Pack_size(sendcnt*(comm_size/2), sendtype, comm,
                           &tmp_buf_size);

        tmp_buf = MPIU_Malloc(tmp_buf_size);
	/* --BEGIN ERROR HANDLING-- */
        if (!tmp_buf)
	{ 
            mpi_errno = MPIR_Err_create_code( MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem", 0 );
            return mpi_errno;
        }
	/* --END ERROR HANDLING-- */

        position = 0;
        if (sendbuf != MPI_IN_PLACE)
	{
            NMPI_Pack(sendbuf, sendcnt, sendtype, tmp_buf,
                      tmp_buf_size, &position, comm);
            nbytes = position;
        }
        else
	{
            /* do a dummy pack just to calculate nbytes */
            NMPI_Pack(recvbuf, 1, recvtype, tmp_buf,
                      tmp_buf_size, &position, comm);
            nbytes = position*recvcnt;
        }
        
        curr_cnt = nbytes;
        
        mask = 0x1;
        while (mask < comm_size)
	{
            if ((mask & relative_rank) == 0)
	    {
                src = relative_rank | mask;
                if (src < comm_size)
		{
                    src = (src + root) % comm_size;
                    mpi_errno = MPIC_Recv(((char *)tmp_buf + curr_cnt), 
                                          mask*nbytes, MPI_BYTE, src,
                                          MPIR_GATHER_TAG, comm, 
                                          &status);
		    /* --BEGIN ERROR HANDLING-- */
                    if (mpi_errno)
		    {
			mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
			return mpi_errno;
		    }
		    /* --END ERROR HANDLING-- */
                    /* the recv size is larger than what may be sent in
                       some cases. query amount of data actually received */
                    NMPI_Get_count(&status, MPI_BYTE, &recv_size);
                    curr_cnt += recv_size;
                }
            }
            else
	    {
                dst = relative_rank ^ mask;
                dst = (dst + root) % comm_size;
                mpi_errno = MPIC_Send(tmp_buf, curr_cnt, MPI_BYTE, dst,
                                      MPIR_GATHER_TAG, comm); 
		/* --BEGIN ERROR HANDLING-- */
                if (mpi_errno)
		{
		    mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
		    return mpi_errno;
		}
		/* --END ERROR HANDLING-- */
                break;
            }
            mask <<= 1;
        }
        
        if (rank == root)
	{
            /* reorder and copy from tmp_buf into recvbuf */
            if (sendbuf != MPI_IN_PLACE)
	    {
                position = 0;
                NMPI_Unpack(tmp_buf, tmp_buf_size, &position,