allreduce.c

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

#include "mpiimpl.h"
#include "collutil.h"

/* -- Begin Profiling Symbol Block for routine MPI_Allreduce */
#if defined(HAVE_PRAGMA_WEAK)
#pragma weak MPI_Allreduce = PMPI_Allreduce
#elif defined(HAVE_PRAGMA_HP_SEC_DEF)
#pragma _HP_SECONDARY_DEF PMPI_Allreduce  MPI_Allreduce
#elif defined(HAVE_PRAGMA_CRI_DUP)
#pragma _CRI duplicate MPI_Allreduce as PMPI_Allreduce
#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
#undef MPI_Allreduce
#define MPI_Allreduce PMPI_Allreduce

/* The order of entries in this table must match the definitions in 
   mpi.h.in */
MPI_User_function *MPIR_Op_table[] = { MPIR_MAXF, MPIR_MINF, MPIR_SUM,
                                       MPIR_PROD, MPIR_LAND,
                                       MPIR_BAND, MPIR_LOR, MPIR_BOR,
                                       MPIR_LXOR, MPIR_BXOR,
                                       MPIR_MINLOC, MPIR_MAXLOC, };

MPIR_Op_check_dtype_fn *MPIR_Op_check_dtype_table[] = {
    MPIR_MAXF_check_dtype, MPIR_MINF_check_dtype,
    MPIR_SUM_check_dtype,
    MPIR_PROD_check_dtype, MPIR_LAND_check_dtype,
    MPIR_BAND_check_dtype, MPIR_LOR_check_dtype, MPIR_BOR_check_dtype,
    MPIR_LXOR_check_dtype, MPIR_BXOR_check_dtype,
    MPIR_MINLOC_check_dtype, MPIR_MAXLOC_check_dtype, }; 


/* This is the default implementation of allreduce. The algorithm is:
   
   Algorithm: MPI_Allreduce

   For the heterogeneous case, we call MPI_Reduce followed by MPI_Bcast
   in order to meet the requirement that all processes must have the
   same result. For the homogeneous case, we use the following algorithms.


   For long messages and for builtin ops and if count >= pof2 (where
   pof2 is the nearest power-of-two less than or equal to the number
   of processes), we use Rabenseifner's algorithm (see 
   http://www.hlrs.de/mpi/myreduce.html).
   This algorithm implements the allreduce in two steps: first a
   reduce-scatter, followed by an allgather. A recursive-halving
   algorithm (beginning with processes that are distance 1 apart) is
   used for the reduce-scatter, and a recursive doubling 
   algorithm is used for the allgather. The non-power-of-two case is
   handled by dropping to the nearest lower power-of-two: the first
   few even-numbered processes send their data to their right neighbors
   (rank+1), and the reduce-scatter and allgather happen among the remaining
   power-of-two processes. At the end, the first few even-numbered
   processes get the result from their right neighbors.

   For the power-of-two case, the cost for the reduce-scatter is 
   lgp.alpha + n.((p-1)/p).beta + n.((p-1)/p).gamma. The cost for the
   allgather lgp.alpha + n.((p-1)/p).beta. Therefore, the
   total cost is:
   Cost = 2.lgp.alpha + 2.n.((p-1)/p).beta + n.((p-1)/p).gamma

   For the non-power-of-two case, 
   Cost = (2.floor(lgp)+2).alpha + (2.((p-1)/p) + 2).n.beta + n.(1+(p-1)/p).gamma

   
   For short messages, for user-defined ops, and for count < pof2 
   we use a recursive doubling algorithm (similar to the one in
   MPI_Allgather). We use this algorithm in the case of user-defined ops
   because in this case derived datatypes are allowed, and the user
   could pass basic datatypes on one process and derived on another as
   long as the type maps are the same. Breaking up derived datatypes
   to do the reduce-scatter is tricky. 

   Cost = lgp.alpha + n.lgp.beta + n.lgp.gamma

   Possible improvements: 

   End Algorithm: MPI_Allreduce
*/

#undef FUNCNAME
#define FUNCNAME allreduce_intra_or_coll_fn
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
static inline int allreduce_intra_or_coll_fn(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
                                             MPID_Comm *comm_ptr, int *errflag)
{
    int mpi_errno = MPI_SUCCESS;

    if (comm_ptr->coll_fns != NULL && comm_ptr->coll_fns->Allreduce != NULL) {
	mpi_errno = comm_ptr->coll_fns->Allreduce(sendbuf, recvbuf, count, datatype, op, comm_ptr, errflag);
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    } else {
        mpi_errno = MPIR_Allreduce_intra(sendbuf, recvbuf, count, datatype, op, comm_ptr, errflag);
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    }
        
 fn_exit:
    return mpi_errno;
 fn_fail:
    goto fn_exit;
}


/* not declared static because a machine-specific function may call this one 
   in some cases */
#undef FUNCNAME
#define FUNCNAME MPIR_Allreduce_intra
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
int MPIR_Allreduce_intra ( 
    void *sendbuf, 
    void *recvbuf, 
    int count, 
    MPI_Datatype datatype, 
    MPI_Op op, 
    MPID_Comm *comm_ptr,
    int *errflag )
{
    int is_homogeneous;
#ifdef MPID_HAS_HETERO
    int rc;
#endif
    int        comm_size, rank, type_size;
    int mpi_errno = MPI_SUCCESS;
    int mpi_errno_ret = MPI_SUCCESS;
    int mask, dst, is_commutative, pof2, newrank, rem, newdst, i,
        send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps; 
    MPI_Aint true_extent, true_lb, extent;
    void *tmp_buf;
    MPI_Comm comm;
    MPIU_CHKLMEM_DECL(3);
    
    /* check if multiple threads are calling this collective function */
    MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );

    if (count == 0) goto fn_exit;
    comm = comm_ptr->handle;

    is_commutative = MPIR_Op_is_commutative(op);

#if defined(USE_SMP_COLLECTIVES)
    /* is the op commutative? We do SMP optimizations only if it is. */
    if (MPIR_Comm_is_node_aware(comm_ptr) && is_commutative) {
        /* on each node, do a reduce to the local root */ 
        if (comm_ptr->node_comm != NULL) {
            /* take care of the MPI_IN_PLACE case. For reduce, 
               MPI_IN_PLACE is specified only on the root; 
               for allreduce it is specified on all processes. */

            if ((sendbuf == MPI_IN_PLACE) && (comm_ptr->node_comm->rank != 0)) {
                /* IN_PLACE and not root of reduce. Data supplied to this
                   allreduce is in recvbuf. Pass that as the sendbuf to reduce. */
			
                mpi_errno = MPIR_Reduce_impl(recvbuf, NULL, count, datatype, op, 0, comm_ptr->node_comm, errflag);
                if (mpi_errno) {
                    /* for communication errors, just record the error but continue */
                    *errflag = TRUE;
                    MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                    MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                }
            } else {
                mpi_errno = MPIR_Reduce_impl(sendbuf, recvbuf, count, datatype, op, 0, comm_ptr->node_comm, errflag);
                if (mpi_errno) {
                    /* for communication errors, just record the error but continue */
                    *errflag = TRUE;
                    MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                    MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                }
            }
        } else {
            /* only one process on the node. copy sendbuf to recvbuf */
            if (sendbuf != MPI_IN_PLACE) {
                mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf, count, datatype);
                if (mpi_errno) MPIU_ERR_POP(mpi_errno);
            }
        }

        /* now do an IN_PLACE allreduce among the local roots of all nodes */
        if (comm_ptr->node_roots_comm != NULL) {
            mpi_errno = allreduce_intra_or_coll_fn(MPI_IN_PLACE, recvbuf, count, datatype, op, comm_ptr->node_roots_comm,
                                                   errflag);
            if (mpi_errno) {
                /* for communication errors, just record the error but continue */
                *errflag = TRUE;
                MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
            }
        }

        /* now broadcast the result among local processes */
        if (comm_ptr->node_comm != NULL) {
            mpi_errno = MPIR_Bcast_impl(recvbuf, count, datatype, 0, comm_ptr->node_comm, errflag);
            if (mpi_errno) {
                /* for communication errors, just record the error but continue */
                *errflag = TRUE;
                MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
            }
        }
        goto fn_exit;
    }
#endif
            
    
    is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
    if (comm_ptr->is_hetero)
        is_homogeneous = 0;
#endif
    
#ifdef MPID_HAS_HETERO
    if (!is_homogeneous) {
        /* heterogeneous. To get the same result on all processes, we
           do a reduce to 0 and then broadcast. */
        mpi_errno = MPIR_Reduce_impl ( sendbuf, recvbuf, count, datatype,
                                       op, 0, comm_ptr, errflag );
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
            *errflag = TRUE;
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }

        mpi_errno = MPIR_Bcast_impl( recvbuf, count, datatype, 0, comm_ptr, errflag );
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
            *errflag = TRUE;
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }
    }
    else 
#endif /* MPID_HAS_HETERO */
    {
        /* homogeneous */

        comm_size = comm_ptr->local_size;
        rank = comm_ptr->rank;

        is_commutative = MPIR_Op_is_commutative(op);

        /* need to allocate temporary buffer to store incoming data*/
        MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
        MPID_Datatype_get_extent_macro(datatype, extent);

        MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
        MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "temporary buffer");
	
        /* adjust for potential negative lower bound in datatype */
        tmp_buf = (void *)((char*)tmp_buf - true_lb);
        
        /* copy local data into recvbuf */
        if (sendbuf != MPI_IN_PLACE) {
            mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
                                       count, datatype);
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }

        MPID_Datatype_get_size_macro(datatype, type_size);

        /* find nearest power-of-two less than or equal to comm_size */
        pof2 = 1;
        while (pof2 <= comm_size) pof2 <<= 1;
        pof2 >>=1;

        rem = comm_size - pof2;

        /* In the non-power-of-two case, all even-numbered
           processes of rank < 2*rem send their data to
           (rank+1). These even-numbered processes no longer
           participate in the algorithm until the very end. The
           remaining processes form a nice power-of-two. */
        
        if (rank < 2*rem) {
            if (rank % 2 == 0) { /* even */
                mpi_errno = MPIC_Send_ft(recvbuf, count, 
                                         datatype, rank+1,
                                         MPIR_ALLREDUCE_TAG, comm, errflag);
                if (mpi_errno) {
                    /* for communication errors, just record the error but continue */
                    *errflag = TRUE;
                    MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                    MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                }
                
                /* temporarily set the rank to -1 so that this
                   process does not pariticipate in recursive
                   doubling */
                newrank = -1; 
            }
            else { /* odd */
                mpi_errno = MPIC_Recv_ft(tmp_buf, count, 
                                         datatype, rank-1,
                                         MPIR_ALLREDUCE_TAG, comm,
                                         MPI_STATUS_IGNORE, errflag);
                if (mpi_errno) {
                    /* for communication errors, just record the error but continue */
                    *errflag = TRUE;
                    MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                    MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                }

                /* do the reduction on received data. since the
                   ordering is right, it doesn't matter whether
                   the operation is commutative or not. */
                mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
                if (mpi_errno) MPIU_ERR_POP(mpi_errno);

                /* change the rank */
                newrank = rank / 2;
            }
        }
        else  /* rank >= 2*rem */
            newrank = rank - rem;
        
        /* If op is user-defined or count is less than pof2, use
           recursive doubling algorithm. Otherwise do a reduce-scatter
           followed by allgather. (If op is user-defined,
           derived datatypes are allowed and the user could pass basic
           datatypes on one process and derived on another as long as
           the type maps are the same. Breaking up derived
           datatypes to do the reduce-scatter is tricky, therefore
           using recursive doubling in that case.) */

        if (newrank != -1) {
            if ((count*type_size <= MPIR_PARAM_ALLREDUCE_SHORT_MSG_SIZE) ||
                (HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) ||  
                (count < pof2)) { /* use recursive doubling */
                mask = 0x1;
                while (mask < pof2) {
                    newdst = newrank ^ mask;
                    /* find real rank of dest */
                    dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;

                    /* Send the most current data, which is in recvbuf. Recv
                       into tmp_buf */ 
                    mpi_errno = MPIC_Sendrecv_ft(recvbuf, count, datatype, 
                                                 dst, MPIR_ALLREDUCE_TAG, tmp_buf,
                                                 count, datatype, dst,
                                                 MPIR_ALLREDUCE_TAG, comm,
                                                 MPI_STATUS_IGNORE, errflag);
                    if (mpi_errno) {
                        /* for communication errors, just record the error but continue */
                        *errflag = TRUE;
                        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                        MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                    }
                    
                    /* tmp_buf contains data received in this step.
                       recvbuf contains data accumulated so far */
                    
                    if (is_commutative  || (dst < rank)) {
                        /* op is commutative OR the order is already right */
                        mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
                        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                    }
                    else {
                        /* op is noncommutative and the order is not right */
                        mpi_errno = MPIR_Reduce_local_impl(recvbuf, tmp_buf, count, datatype, op);
                        if (mpi_errno) MPIU_ERR_POP(mpi_errno);

                        /* copy result back into recvbuf */
                        mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
                                                   recvbuf, count, datatype);
                        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
                    }
                    mask <<= 1;
                }
            }

            else {

                /* do a reduce-scatter followed by allgather */

                /* for the reduce-scatter, calculate the count that
                   each process receives and the displacement within
                   the buffer */

		MPIU_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts");
		MPIU_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements");

                for (i=0; i<(pof2-1); i++) 
                    cnts[i] = count/pof2;
                cnts[pof2-1] = count - (count/pof2)*(pof2-1);

                disps[0] = 0;
                for (i=1; i<pof2; i++)
                    disps[i] = disps[i-1] + cnts[i-1];

                mask = 0x1;
                send_idx = recv_idx = 0;
                last_idx = pof2;
                while (mask < pof2) {
                    newdst = newrank ^ mask;
                    /* find real rank of dest */
                    dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;

                    send_cnt = recv_cnt = 0;
                    if (newrank < newdst) {
                        send_idx = recv_idx + pof2/(mask*2);
                        for (i=send_idx; i<last_idx; i++)
                            send_cnt += cnts[i];
                        for (i=recv_idx; i<send_idx; i++)
                            recv_cnt += cnts[i];
                    }
                    else {
                        recv_idx = send_idx + pof2/(mask*2);
                        for (i=send_idx; i<recv_idx; i++)
                            send_cnt += cnts[i];
                        for (i=recv_idx; i<last_idx; i++)
                            recv_cnt += cnts[i];
                    }

/*                    printf("Rank %d, send_idx %d, recv_idx %d, send_cnt %d, recv_cnt %d, last_idx %d\n", newrank, send_idx, recv_idx,
                           send_cnt, recv_cnt, last_idx);
                           */
                    /* Send data from recvbuf. Recv into tmp_buf */ 
                    mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
                                                 disps[send_idx]*extent,
                                                 send_cnt, datatype,  
                                                 dst, MPIR_ALLREDUCE_TAG, 
                                                 (char *) tmp_buf +
                                                 disps[recv_idx]*extent,
                                                 recv_cnt, datatype, dst,
                                                 MPIR_ALLREDUCE_TAG, comm,
                                                 MPI_STATUS_IGNORE, errflag);
                    if (mpi_errno) {
                        /* for communication errors, just record the error but continue */
                        *errflag = TRUE;
                        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                        MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                    }
                    
                    /* tmp_buf contains data received in this step.
                       recvbuf contains data accumulated so far */
                    
                    /* This algorithm is used only for predefined ops
                       and predefined ops are always commutative. */
                    mpi_errno = MPIR_Reduce_local_impl(((char *) tmp_buf + disps[recv_idx]*extent),
                                                       ((char *) recvbuf + disps[recv_idx]*extent),
                                                       recv_cnt, datatype, op);
                    if (mpi_errno) MPIU_ERR_POP(mpi_errno);

                    /* update send_idx for next iteration */
                    send_idx = recv_idx;
                    mask <<= 1;

                    /* update last_idx, but not in last iteration
                       because the value is needed in the allgather
                       step below. */
                    if (mask < pof2)
                        last_idx = recv_idx + pof2/mask;
                }

                /* now do the allgather */

                mask >>= 1;
                while (mask > 0) {
                    newdst = newrank ^ mask;
                    /* find real rank of dest */
                    dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;

                    send_cnt = recv_cnt = 0;
                    if (newrank < newdst) {
                        /* update last_idx except on first iteration */
                        if (mask != pof2/2)
                            last_idx = last_idx + pof2/(mask*2);

                        recv_idx = send_idx + pof2/(mask*2);
                        for (i=send_idx; i<recv_idx; i++)
                            send_cnt += cnts[i];
                        for (i=recv_idx; i<last_idx; i++)
                            recv_cnt += cnts[i];
                    }
                    else {
                        recv_idx = send_idx - pof2/(mask*2);
                        for (i=send_idx; i<last_idx; i++)
                            send_cnt += cnts[i];
                        for (i=recv_idx; i<send_idx; i++)
                            recv_cnt += cnts[i];
                    }

                    mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
                                                 disps[send_idx]*extent,
                                                 send_cnt, datatype,  
                                                 dst, MPIR_ALLREDUCE_TAG, 
                                                 (char *) recvbuf +
                                                 disps[recv_idx]*extent,
                                                 recv_cnt, datatype, dst,
                                                 MPIR_ALLREDUCE_TAG, comm,
                                                 MPI_STATUS_IGNORE, errflag);
                    if (mpi_errno) {
                        /* for communication errors, just record the error but continue */
                        *errflag = TRUE;
                        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                        MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
                    }

                    if (newrank > newdst) send_idx = recv_idx;

                    mask >>= 1;
                }
            }
        }

        /* In the non-power-of-two case, all odd-numbered
           processes of rank < 2*rem send the result to
           (rank-1), the ranks who didn't participate above. */
        if (rank < 2*rem) {
            if (rank % 2)  /* odd */
                mpi_errno = MPIC_Send_ft(recvbuf, count, 
                                         datatype, rank-1,
                                         MPIR_ALLREDUCE_TAG, comm, errflag);
            else  /* even */
                mpi_errno = MPIC_Recv_ft(recvbuf, count,
                                         datatype, rank+1,
                                         MPIR_ALLREDUCE_TAG, comm,
                                         MPI_STATUS_IGNORE, errflag);
            if (mpi_errno) {
                /* for communication errors, just record the error but continue */
                *errflag = TRUE;
                MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
                MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
            }
        }
    }

  fn_exit:
    /* check if multiple threads are calling this collective function */
    MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );

    MPIU_CHKLMEM_FREEALL();
    if (mpi_errno_ret)
        mpi_errno = mpi_errno_ret;
    return (mpi_errno);

  fn_fail:
    goto fn_exit;
}


/* not declared static because a machine-specific function may call this one 
   in some cases */
#undef FUNCNAME
#define FUNCNAME MPIR_Allreduce_inter
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
int MPIR_Allreduce_inter ( 
    void *sendbuf, 
    void *recvbuf, 
    int count, 
    MPI_Datatype datatype, 
    MPI_Op op, 
    MPID_Comm *comm_ptr,
    int *errflag )
{
/* Intercommunicator Allreduce.
   We first do an intercommunicator reduce to rank 0 on left group,
   then an intercommunicator reduce to rank 0 on right group, followed
   by local intracommunicator broadcasts in each group.

   We don't do local reduces first and then intercommunicator
   broadcasts because it would require allocation of a temporary buffer. 
*/
    int rank, mpi_errno, root;
    int mpi_errno_ret = MPI_SUCCESS;
    MPID_Comm *newcomm_ptr = NULL;
    
    rank = comm_ptr->rank;

    /* first do a reduce from right group to rank 0 in left group,
       then from left group to rank 0 in right group*/
    if (comm_ptr->is_low_group) {
        /* reduce from right group to rank 0*/
        root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
        mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
				      root, comm_ptr, errflag);
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
            *errflag = TRUE;
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }

        /* reduce to rank 0 of right group */
        root = 0;
        mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
				      root, comm_ptr, errflag);
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
            *errflag = TRUE;
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }
    }
    else {
        /* reduce to rank 0 of left group */
        root = 0;
        mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
				      root, comm_ptr, errflag);
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
            *errflag = TRUE;
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }

        /* reduce from right group to rank 0 */
        root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
        mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
				      root, comm_ptr, errflag);
        if (mpi_errno) {
            /* for communication errors, just record the error but continue */
            *errflag = TRUE;
            MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
            MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
        }
    }

    /* Get the local intracommunicator */
    if (!comm_ptr->local_comm)
	MPIR_Setup_intercomm_localcomm( comm_ptr );

    newcomm_ptr = comm_ptr->local_comm;

    mpi_errno = MPIR_Bcast_impl(recvbuf, count, datatype, 0, newcomm_ptr, errflag);
    if (mpi_errno) {
        /* for communication errors, just record the error but continue */
        *errflag = TRUE;
        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
        MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
    }

  fn_exit:
    if (mpi_errno_ret)
        mpi_errno = mpi_errno_ret;
    else if (*errflag)
        MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");

    return mpi_errno;

  fn_fail:
    goto fn_exit;
}

/* MPIR_Allreduce performs an allreduce using point-to-point messages.
   This is intended to be used by device-specific implementations of
   allreduce.  In all other cases MPIR_Allreduce_impl should be
   used. */
#undef FUNCNAME
#define FUNCNAME MPIR_Allreduce
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
int MPIR_Allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPID_Comm *comm_ptr,
                   int *errflag)
{
    int mpi_errno = MPI_SUCCESS;

    if (comm_ptr->comm_kind == MPID_INTRACOMM) {
        /* intracommunicator */
        mpi_errno = MPIR_Allreduce_intra(sendbuf, recvbuf, count, datatype, op, comm_ptr, errflag);
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    }
    else {
        /* intercommunicator */
        mpi_errno = MPIR_Allreduce_inter(sendbuf, recvbuf, count, datatype, op, comm_ptr, errflag);
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    }

fn_exit:
    return mpi_errno;
fn_fail:

    goto fn_exit;
}

/* MPIR_Allreduce_impl should be called by any internal component that
   would otherwise call MPI_Allreduce.  This differs from
   MPIR_Allreduce in that this will call the coll_fns version if it
   exists.  */
#undef FUNCNAME
#define FUNCNAME MPIR_Allreduce_impl
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
int MPIR_Allreduce_impl(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPID_Comm *comm_ptr,
                        int *errflag)
{
    int mpi_errno = MPI_SUCCESS;

    if (comm_ptr->coll_fns != NULL && comm_ptr->coll_fns->Allreduce != NULL)
    {
	mpi_errno = comm_ptr->coll_fns->Allreduce(sendbuf, recvbuf, count, datatype, op, comm_ptr, errflag);
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
    }
    else
    {
        if (comm_ptr->comm_kind == MPID_INTRACOMM) {
            /* intracommunicator */
            mpi_errno = MPIR_Allreduce_intra(sendbuf, recvbuf, count, datatype, op, comm_ptr, errflag);
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
	}
        else {
            /* intercommunicator */
            mpi_errno = MPIR_Allreduce_inter(sendbuf, recvbuf, count, datatype, op, comm_ptr, errflag);
            if (mpi_errno) MPIU_ERR_POP(mpi_errno);
        }
    }

fn_exit:
    return mpi_errno;
fn_fail:
    goto fn_exit;
}


#endif

#undef FUNCNAME
#define FUNCNAME MPI_Allreduce
#undef FCNAME

/*@
MPI_Allreduce - Combines values from all processes and distributes the result
                back to all processes

Input Parameters:
+ sendbuf - starting address of send buffer (choice) 
. count - number of elements in send buffer (integer) 
. datatype - data type of elements of send buffer (handle) 
. op - operation (handle) 
- comm - communicator (handle) 

Output Parameter:
. recvbuf - starting address of receive buffer (choice) 

.N ThreadSafe

.N Fortran

.N collops

.N Errors
.N MPI_ERR_BUFFER
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_OP
.N MPI_ERR_COMM
@*/
int MPI_Allreduce ( void *sendbuf, void *recvbuf, int count, 
		    MPI_Datatype datatype, MPI_Op op, MPI_Comm comm )
{
    static const char FCNAME[] = "MPI_Allreduce";
    int mpi_errno = MPI_SUCCESS;
    MPID_Comm *comm_ptr = NULL;
    int errflag = FALSE;
    MPID_MPI_STATE_DECL(MPID_STATE_MPI_ALLREDUCE);

    MPIR_ERRTEST_INITIALIZED_ORDIE();
    
    MPIU_THREAD_CS_ENTER(ALLFUNC,);
    MPID_MPI_COLL_FUNC_ENTER(MPID_STATE_MPI_ALLREDUCE);

    /* Validate parameters, especially handles needing to be converted */
#   ifdef HAVE_ERROR_CHECKING
    {
        MPID_BEGIN_ERROR_CHECKS;
        {
	    MPIR_ERRTEST_COMM(comm, mpi_errno);
            if (mpi_errno != MPI_SUCCESS) goto fn_fail;
	}
        MPID_END_ERROR_CHECKS;
    }
#   endif /* HAVE_ERROR_CHECKING */

    /* Convert MPI object handles to object pointers */
    MPID_Comm_get_ptr( comm, comm_ptr );

    /* Validate parameters and objects (post conversion) */
#   ifdef HAVE_ERROR_CHECKING
    {
        MPID_BEGIN_ERROR_CHECKS;
        {
            MPID_Datatype *datatype_ptr = NULL;
            MPID_Op *op_ptr = NULL;

            MPID_Comm_valid_ptr( comm_ptr, mpi_errno );
            if (mpi_errno != MPI_SUCCESS) goto fn_fail;
	    MPIR_ERRTEST_COUNT(count, mpi_errno);
	    MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
	    MPIR_ERRTEST_OP(op, mpi_errno);
	    
            if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
                MPID_Datatype_get_ptr(datatype, datatype_ptr);
                MPID_Datatype_valid_ptr( datatype_ptr, mpi_errno );
                MPID_Datatype_committed_ptr( datatype_ptr, mpi_errno );
            }

	    if (comm_ptr->comm_kind == MPID_INTERCOMM)
                MPIR_ERRTEST_SENDBUF_INPLACE(sendbuf, count, mpi_errno);
            
            if (sendbuf != MPI_IN_PLACE) 
                MPIR_ERRTEST_USERBUFFER(sendbuf,count,datatype,mpi_errno);

            MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, count, mpi_errno);
	    MPIR_ERRTEST_USERBUFFER(recvbuf,count,datatype,mpi_errno);

	    if (mpi_errno != MPI_SUCCESS) goto fn_fail;

            if (HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) {
                MPID_Op_get_ptr(op, op_ptr);
                MPID_Op_valid_ptr( op_ptr, mpi_errno );
            }
            if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
                mpi_errno = 
                    ( * MPIR_Op_check_dtype_table[op%16 - 1] )(datatype); 
            }
	    if (count != 0) {
		MPIR_ERRTEST_ALIAS_COLL(sendbuf, recvbuf, mpi_errno);
	    }
	    if (mpi_errno != MPI_SUCCESS) goto fn_fail;
	}
        MPID_END_ERROR_CHECKS;
    }
#   endif /* HAVE_ERROR_CHECKING */

    /* ... body of routine ...  */

    mpi_errno = MPIR_Allreduce_impl(sendbuf, recvbuf, count, datatype, op, comm_ptr, &errflag);
    if (mpi_errno) goto fn_fail;

    /* ... end of body of routine ... */
    
  fn_exit:
    MPID_MPI_COLL_FUNC_EXIT(MPID_STATE_MPI_ALLREDUCE);
    MPIU_THREAD_CS_EXIT(ALLFUNC,);
    return mpi_errno;

  fn_fail:
    /* --BEGIN ERROR HANDLING-- */
#   ifdef HAVE_ERROR_CHECKING
    {
	mpi_errno = MPIR_Err_create_code(
	    mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_allreduce",
	    "**mpi_allreduce %p %p %d %D %O %C", sendbuf, recvbuf, count, datatype, op, comm);
    }
#   endif
    mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}