예제 #1
0
/* once */
int tMPI_Once(tMPI_Comm comm, void (*function)(void*), void *param, 
                int *was_first)
{
    int myrank;
    int ret=TMPI_SUCCESS;
    struct coll_sync *csync;
    struct coll_env *cev;
    int syncs;


    if (!comm)
    {
        return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_COMM);
    }
    myrank=tMPI_Comm_seek_rank(comm, tMPI_Get_current());

    /* we increase our counter, and determine which coll_env we get */
    csync=&(comm->csync[myrank]);
    csync->syncs++;
    cev=&(comm->cev[csync->syncs % N_COLL_ENV]);

    /* now do a compare-and-swap on the current_syncc */
    syncs=tMPI_Atomic_get( &(cev->coll.current_sync));
    if ((csync->syncs - syncs > 0) && /* check if sync was an earlier number. 
                                         If it is a later number, we can't 
                                         have been the first to arrive here. */
        tMPI_Atomic_cas(&(cev->coll.current_sync), syncs, csync->syncs))
    {
        /* we're the first! */
        function(param);
        if (was_first)
            *was_first=TRUE;
    }
    return ret;
}
예제 #2
0
int tMPI_Cart_get(tMPI_Comm comm, int maxdims, int *dims, int *periods,
                  int *coords)
{
    int i;
    int myrank=tMPI_Comm_seek_rank(comm, tMPI_Get_current());

#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Cart_get(%p, %d, %p, %p, %p)", comm, maxdims, 
                       dims, periods, coords);
#endif
    if (!comm)
    {
        return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_COMM);
    }
    if (!comm->cart || comm->cart->ndims==0)
        return TMPI_SUCCESS;

    tMPI_Cart_coords(comm, myrank, maxdims, coords);

    for(i=0;i<comm->cart->ndims;i++)
    {
        if (i>=maxdims)
        {
            return tMPI_Error(comm, TMPI_ERR_DIMS);
        }
        dims[i]=comm->cart->dims[i];
        periods[i]=comm->cart->periods[i];
    }

    return TMPI_SUCCESS;
}
예제 #3
0
int tMPI_Comm_dup(tMPI_Comm comm, tMPI_Comm *newcomm)
{
#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Comm_dup(%p, %p)", comm, newcomm);
#endif
    /* we just call Comm_split because it already contains all the
       neccesary synchronization constructs. */
    return tMPI_Comm_split(comm, 0, tMPI_Comm_seek_rank(comm,
                                                        tMPI_Get_current()), newcomm);
}
예제 #4
0
int tMPI_Cart_create(tMPI_Comm comm_old, int ndims, int *dims, int *periods,
                     int reorder, tMPI_Comm *comm_cart)
{
    int myrank = tMPI_Comm_seek_rank(comm_old, tMPI_Get_current());
    int key    = myrank;
    int color  = 0;
    int Ntot   = 1;
    int i;


#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Cart_create(%p, %d, %p, %p, %d, %p)", comm_old,
                     ndims, dims, periods, reorder, comm_cart);
#endif
    if (!comm_old)
    {
        return tMPI_Error(comm_old, TMPI_ERR_COMM);
    }
    /* calculate the total number of procs in cartesian comm */
    for (i = 0; i < ndims; i++)
    {
        Ntot *= dims[i];
    }
    /* refuse to create if there's not enough procs */
    if (comm_old->grp.N < Ntot)
    {
        *comm_cart = TMPI_COMM_NULL;
#if 1
        return tMPI_Error(comm_old, TMPI_ERR_CART_CREATE_NPROCS);
#endif
    }

    if (key >= Ntot)
    {
        key = TMPI_UNDEFINED;
    }

    if (reorder)
    {
        tMPI_Cart_map(comm_old, ndims, dims, periods, &key);
    }

    if (key == TMPI_UNDEFINED)
    {
        color = TMPI_UNDEFINED;
    }

    tMPI_Comm_split(comm_old, color, key, comm_cart);

    tMPI_Cart_init(comm_cart, ndims, dims, periods);

    return TMPI_SUCCESS;
}
예제 #5
0
int tMPI_Comm_create(tMPI_Comm comm, tMPI_Group group, tMPI_Comm *newcomm)
{
    int color = TMPI_UNDEFINED;
    int key   = tMPI_Comm_seek_rank(comm, tMPI_Get_current());

#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Comm_create(%p, %p, %p)", comm, group, newcomm);
#endif
    if (tMPI_In_group(group))
    {
        color = 1;
    }
    /* the MPI specs specifically say that this is equivalent */
    return tMPI_Comm_split(comm, color, key, newcomm);
}
예제 #6
0
int tMPI_Cart_map(tMPI_Comm comm, int ndims, int *dims, int *periods, 
                  int *newrank)
{
    /* this function doesn't actually do anything beyond returning the current 
       rank (or TMPI_UNDEFINED if it doesn't fit in the new topology */
    int myrank=tMPI_Comm_seek_rank(comm, tMPI_Get_current());
    int Ntot=1;
    int i;

#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Cart_map(%p, %d, %p, %p, %p)", comm, ndims, dims, 
                     periods, newrank);
#endif
    if (!comm)
    {
        return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_COMM);
    }
    if (!periods)
    {
        return tMPI_Error(comm, TMPI_ERR_DIMS);
    }
 
    /* calculate the total number of procs in cartesian comm */
    for(i=0;i<ndims;i++)
    {
        Ntot *= dims[i];
    }

    if (myrank >= Ntot)
    {
        *newrank=TMPI_UNDEFINED;
    }
    else
    {
        *newrank=myrank;
    }

    return TMPI_SUCCESS;
}
예제 #7
0
int tMPI_Scan(void* sendbuf, void* recvbuf, int count,
              tMPI_Datatype datatype, tMPI_Op op, tMPI_Comm comm)
{
    struct tmpi_thread *cur=tMPI_Get_current();
    int myrank=tMPI_Comm_seek_rank(comm, cur);
    int N=tMPI_Comm_N(comm);
    int prev=myrank - 1; /* my previous neighbor */
    int next=myrank + 1; /* my next neighbor */

#ifdef TMPI_PROFILE
    tMPI_Profile_count_start(cur);
#endif
#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Scan(%p, %p, %d, %p, %p, %p)",
                     sendbuf, recvbuf, count, datatype, op, comm);
#endif
    if (count==0)
        return TMPI_SUCCESS;
    if (!recvbuf)
    {
        return tMPI_Error(comm, TMPI_ERR_BUF);
    }
    if (sendbuf==TMPI_IN_PLACE) 
    {
        sendbuf=recvbuf;
    }

    /* we set our send and recv buffers */
    tMPI_Atomic_ptr_set(&(comm->reduce_sendbuf[myrank]),sendbuf);
    tMPI_Atomic_ptr_set(&(comm->reduce_recvbuf[myrank]),recvbuf);

    /* now wait for the previous rank to finish */
    if (myrank > 0)
    {
        void *a, *b;
        int ret;

#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
        tMPI_Profile_wait_start(cur);
#endif
        /* wait for the previous neighbor's data to be ready */
        tMPI_Event_wait( &(comm->csync[myrank].events[prev]) );
        tMPI_Event_process( &(comm->csync[myrank].events[prev]), 1);
#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
        tMPI_Profile_wait_stop(cur, TMPIWAIT_Reduce);
#endif
#ifdef TMPI_DEBUG
        printf("%d: scanning with %d \n", myrank, prev, iteration);
        fflush(stdout);
#endif
        /* now do the reduction */
        if (prev > 0)
        {
            a = (void*)tMPI_Atomic_ptr_get(&(comm->reduce_recvbuf[prev]));
        }
        else
        {
            a = (void*)tMPI_Atomic_ptr_get(&(comm->reduce_sendbuf[prev]));
        }
        b = sendbuf;

        if ((ret=tMPI_Reduce_run_op(recvbuf, a, b, datatype,
                                    count, op, comm)) != TMPI_SUCCESS)
        {
            return ret;
        }

        /* signal to my previous neighbor that I'm done with the data */
        tMPI_Event_signal( &(comm->csync[prev].events[prev]) );
    }
    else
    {
        if (sendbuf != recvbuf)
        {
            /* copy the data if this is rank 0, and not MPI_IN_PLACE */
            memcpy(recvbuf, sendbuf, count*datatype->size);
        }
    }

    if (myrank < N-1)
    {
        /* signal to my next neighbor that I have the data */
        tMPI_Event_signal( &(comm->csync[next].events[myrank]) );
        /* and wait for my next neighbor to finish */
        tMPI_Event_wait( &(comm->csync[myrank].events[myrank]) );
        tMPI_Event_process( &(comm->csync[myrank].events[myrank]), 1);
    }


#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
    tMPI_Profile_wait_start(cur);
#endif
    /*tMPI_Barrier_wait( &(comm->barrier));*/
#if defined(TMPI_PROFILE)
    /*tMPI_Profile_wait_stop(cur, TMPIWAIT_Reduce);*/
    tMPI_Profile_count_stop(cur, TMPIFN_Scan);
#endif
    return TMPI_SUCCESS;
}
예제 #8
0
/* this is the main comm creation function. All other functions that create
    comms use this*/
int tMPI_Comm_split(tMPI_Comm comm, int color, int key, tMPI_Comm *newcomm)
{
    int                 i, j;
    int                 N = tMPI_Comm_N(comm);
    volatile tMPI_Comm *newcomm_list;
    volatile int        colors[MAX_PREALLOC_THREADS]; /* array with the colors
                                                         of each thread */
    volatile int        keys[MAX_PREALLOC_THREADS];   /* same for keys (only one of
                                                         the threads actually suplies
                                                         these arrays to the comm
                                                         structure) */
    tmpi_bool          i_am_first = FALSE;
    int                myrank     = tMPI_Comm_seek_rank(comm, tMPI_Get_current());
    struct tmpi_split *spl;
    int                ret;

#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Comm_split(%p, %d, %d, %p)", comm, color, key,
                     newcomm);
#endif
    if (!comm)
    {
        *newcomm = NULL;
        return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_COMM);
    }

    ret = tMPI_Thread_mutex_lock(&(comm->comm_create_lock));
    if (ret != 0)
    {
        return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
    }
    /* first get the colors */
    if (!comm->new_comm)
    {
        /* i am apparently  first */
        comm->split    = (struct tmpi_split*)tMPI_Malloc(sizeof(struct tmpi_split));
        comm->new_comm = (tMPI_Comm*)tMPI_Malloc(N*sizeof(tMPI_Comm));
        if (N <= MAX_PREALLOC_THREADS)
        {
            comm->split->colors = colors;
            comm->split->keys   = keys;
        }
        else
        {
            comm->split->colors = (int*)tMPI_Malloc(N*sizeof(int));
            comm->split->keys   = (int*)tMPI_Malloc(N*sizeof(int));
        }
        comm->split->Ncol_init  = tMPI_Comm_N(comm);
        comm->split->can_finish = FALSE;
        i_am_first              = TRUE;
        /* the main communicator contains a list the size of grp.N */
    }
    newcomm_list        = comm->new_comm; /* we copy it to the local stacks because
                                             we can later erase comm->new_comm safely */
    spl                 = comm->split;    /* we do the same for spl */
    spl->colors[myrank] = color;
    spl->keys[myrank]   = key;
    spl->Ncol_init--;

    if (spl->Ncol_init == 0)
    {
        ret = tMPI_Thread_cond_signal(&(comm->comm_create_prep));
        if (ret != 0)
        {
            return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
        }
    }

    if (!i_am_first)
    {
        /* all other threads can just wait until the creator thread is
           finished */
        while (!spl->can_finish)
        {
            ret = tMPI_Thread_cond_wait(&(comm->comm_create_finish),
                                        &(comm->comm_create_lock) );
            if (ret != 0)
            {
                return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
            }
        }
    }
    else
    {
        int        Ncomms = 0;
        int        comm_color_[MAX_PREALLOC_THREADS];
        int        comm_N_[MAX_PREALLOC_THREADS];
        int       *comm_color = comm_color_; /* there can't be more comms than N*/
        int       *comm_N     = comm_N_;     /* the number of procs in a group */

        int       *comm_groups;              /* the groups */
        tMPI_Comm *comms;                    /* the communicators */

        /* wait for the colors to be done */
        /*if (N>1)*/
        while (spl->Ncol_init > 0)
        {
            ret = tMPI_Thread_cond_wait(&(comm->comm_create_prep),
                                        &(comm->comm_create_lock));
            if (ret != 0)
            {
                return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
            }
        }

        /* reset the state so that a new comm creating function can run */
        spl->Ncol_destroy = N;
        comm->new_comm    = 0;
        comm->split       = 0;

        comm_groups = (int*)tMPI_Malloc(N*N*sizeof(int));
        if (N > MAX_PREALLOC_THREADS)
        {
            comm_color = (int*)tMPI_Malloc(N*sizeof(int));
            comm_N     = (int*)tMPI_Malloc(N*sizeof(int));
        }

        /* count colors, allocate and split up communicators */
        tMPI_Split_colors(N, (int*)spl->colors,
                          (int*)spl->keys,
                          &Ncomms,
                          comm_N, comm_color, comm_groups);


        /* allocate a bunch of communicators */
        comms = (tMPI_Comm*)tMPI_Malloc(Ncomms*sizeof(tMPI_Comm));
        for (i = 0; i < Ncomms; i++)
        {
            ret = tMPI_Comm_alloc(&(comms[i]), comm, comm_N[i]);
            if (ret != TMPI_SUCCESS)
            {
                return ret;
            }
        }

        /* now distribute the comms */
        for (i = 0; i < Ncomms; i++)
        {
            comms[i]->grp.N = comm_N[i];
            for (j = 0; j < comm_N[i]; j++)
            {
                comms[i]->grp.peers[j] =
                    comm->grp.peers[comm_groups[i*comm->grp.N + j]];
            }
        }
        /* and put them into the newcomm_list */
        for (i = 0; i < N; i++)
        {
            newcomm_list[i] = TMPI_COMM_NULL;
            for (j = 0; j < Ncomms; j++)
            {
                if (spl->colors[i] == comm_color[j])
                {
                    newcomm_list[i] = comms[j];
                    break;
                }
            }
        }

#ifdef TMPI_DEBUG
        /* output */
        for (i = 0; i < Ncomms; i++)
        {
            printf("Group %d (color %d) has %d members: ",
                   i, comm_color[i], comm_N[i]);
            for (j = 0; j < comm_N[i]; j++)
            {
                printf(" %d ", comm_groups[comm->grp.N*i + j]);
            }

            printf(" rank: ");
            for (j = 0; j < comm_N[i]; j++)
            {
                printf(" %d ", spl->keys[comm_groups[N*i + j]]);
            }
            printf(" color: ");
            for (j = 0; j < comm_N[i]; j++)
            {
                printf(" %d ", spl->colors[comm_groups[N*i + j]]);
            }
            printf("\n");
        }
#endif
        if (N > MAX_PREALLOC_THREADS)
        {
            free((int*)spl->colors);
            free((int*)spl->keys);
            free(comm_color);
            free(comm_N);
        }
        free(comm_groups);
        free(comms);
        spl->can_finish = TRUE;

        /* tell the waiting threads that there's a comm ready */
        ret = tMPI_Thread_cond_broadcast(&(comm->comm_create_finish));
        if (ret != 0)
        {
            return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
        }
    }
    /* here the individual threads get their comm object */
    *newcomm = newcomm_list[myrank];

    /* free when we have assigned them all, so we can reuse the object*/
    spl->Ncol_destroy--;
    if (spl->Ncol_destroy == 0)
    {
        free((void*)newcomm_list);
        free(spl);
    }

    ret = tMPI_Thread_mutex_unlock(&(comm->comm_create_lock));
    if (ret != 0)
    {
        return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
    }

    return TMPI_SUCCESS;
}
예제 #9
0
void* tMPI_Once_wait(tMPI_Comm comm, void* (*function)(void*), void *param, 
                     int *was_first)
{
    int myrank;
    struct coll_sync *csync;
    struct coll_env *cev;
    int syncs;
    void *ret;


    if (!comm)
    {
        tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_COMM);
        return NULL;
    }
    myrank=tMPI_Comm_seek_rank(comm, tMPI_Get_current());

    /* we increase our counter, and determine which coll_env we get */
    csync=&(comm->csync[myrank]);
    csync->syncs++;
    cev=&(comm->cev[csync->syncs % N_COLL_ENV]);

    /* now do a compare-and-swap on the current_syncc */
    syncs=tMPI_Atomic_get( &(cev->coll.current_sync));
    tMPI_Atomic_memory_barrier_acq();
    if ((csync->syncs - syncs > 0) && /* check if sync was an earlier number. 
                                         If it is a later number, we can't 
                                         have been the first to arrive here. 
                                         Calculating the difference instead
                                         of comparing directly avoids ABA 
                                         problems. */
        tMPI_Atomic_cas(&(cev->coll.current_sync), syncs, csync->syncs))
    {
        /* we're the first! */
        ret=function(param);
        if (was_first)
            *was_first=TRUE;

        /* broadcast the output data */
        cev->coll.res=ret;

        tMPI_Atomic_memory_barrier_rel();
        /* signal that we're done */
        tMPI_Atomic_fetch_add(&(cev->coll.current_sync), 1);
        /* we need to keep being in sync */
        csync->syncs++;
    }
    else
    {
        /* we need to wait until the current_syncc gets increased again */
        csync->syncs++;
        do
        {
            /*tMPI_Atomic_memory_barrier();*/
            syncs=tMPI_Atomic_get( &(cev->coll.current_sync) );
        } while (csync->syncs - syncs > 0); /* difference again due to ABA 
                                               problems */
        tMPI_Atomic_memory_barrier_acq();
        ret=cev->coll.res;
    }
    return ret;
}
예제 #10
0
int tMPI_Gather(void* sendbuf, int sendcount, tMPI_Datatype sendtype,
                void* recvbuf, int recvcount, tMPI_Datatype recvtype,
                int root, tMPI_Comm comm)
{
    int                 synct;
    struct coll_env    *cev;
    int                 myrank;
    int                 ret = TMPI_SUCCESS;
    struct tmpi_thread *cur = tMPI_Get_current();

#ifdef TMPI_PROFILE
    tMPI_Profile_count_start(cur);
#endif
#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Gather(%p, %d, %p, %p, %d, %p, %d, %p)",
                     sendbuf, sendcount, sendtype,
                     recvbuf, recvcount, recvtype, root, comm);
#endif

    if (!comm)
    {
        return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_COMM);
    }
    myrank = tMPI_Comm_seek_rank(comm, cur);

    /* we increase our counter, and determine which coll_env we get */
    cev = tMPI_Get_cev(comm, myrank, &synct);

    if (myrank == root)
    {
        int i;
        int n_remaining = comm->grp.N-1;
        /* do root transfer */
        if (sendbuf != TMPI_IN_PLACE)
        {
            tMPI_Coll_root_xfer(comm, sendtype, recvtype,
                                sendtype->size*sendcount,
                                recvtype->size*recvcount,
                                sendbuf,
                                (char*)recvbuf+myrank*recvcount*recvtype->size,
                                &ret);
        }
        for (i = 0; i < comm->grp.N; i++)
        {
            cev->met[myrank].read_data[i] = FALSE;
        }
        cev->met[myrank].read_data[myrank] = TRUE;

        /* wait for data availability as long as there are xfers to be done */
        while (n_remaining > 0)
        {
#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
            tMPI_Profile_wait_start(cur);
#endif
            tMPI_Event_wait( &(cev->met[myrank]).recv_ev );
#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
            tMPI_Profile_wait_stop(cur, TMPIWAIT_Coll_recv);
#endif
            /* now check all of them */
            for (i = 0; i < comm->grp.N; i++)
            {
                if (!cev->met[myrank].read_data[i] &&
                    (tMPI_Atomic_get(&(cev->met[i].current_sync)) == synct))
                {
                    tMPI_Mult_recv(comm, cev, i, 0, TMPI_GATHER_TAG, recvtype,
                                   recvcount*recvtype->size,
                                   (char*)recvbuf+i*recvcount*recvtype->size,
                                   &ret);
                    tMPI_Event_process( &(cev->met[myrank]).recv_ev, 1);
                    if (ret != TMPI_SUCCESS)
                    {
                        return ret;
                    }
                    cev->met[myrank].read_data[i] = TRUE;
                    n_remaining--;
                }
            }
        }
    }
    else
    {
        if (!sendbuf) /* don't do pointer arithmetic on a NULL ptr */
        {
            return tMPI_Error(comm, TMPI_ERR_BUF);
        }

        /* first set up the data just to root. */
        ret = tMPI_Post_multi(cev, myrank, 0, TMPI_GATHER_TAG, sendtype,
                              sendcount*sendtype->size, sendbuf, 1, synct, root);
        if (ret != TMPI_SUCCESS)
        {
            return ret;
        }
        /* and wait until root is done copying */
        tMPI_Wait_for_others(cev, myrank);
    }
#ifdef TMPI_PROFILE
    tMPI_Profile_count_stop(cur, TMPIFN_Gather);
#endif
    return ret;
}
예제 #11
0
int tMPI_Alltoall(void* sendbuf, int sendcount, tMPI_Datatype sendtype,
                  void* recvbuf, int recvcount, tMPI_Datatype recvtype,
                  tMPI_Comm comm)
{
    int                 synct;
    struct coll_env    *cev;
    int                 myrank;
    int                 ret = TMPI_SUCCESS;
    int                 i;
    size_t              sendsize = sendtype->size*sendcount;
    size_t              recvsize = recvtype->size*recvcount;
    int                 n_remaining;
    struct tmpi_thread *cur = tMPI_Get_current();

#ifdef TMPI_PROFILE
    tMPI_Profile_count_start(cur);
#endif
#ifdef TMPI_TRACE
    tMPI_Trace_print("tMPI_Alltoall(%p, %d, %p, %p, %d, %p, %p)",
                     sendbuf, sendcount, sendtype,
                     recvbuf, recvcount, recvtype, comm);
#endif

    if (!comm)
    {
        return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_COMM);
    }
    if (!sendbuf || !recvbuf) /* don't do pointer arithmetic on a NULL ptr */
    {
        return tMPI_Error(comm, TMPI_ERR_BUF);
    }

    myrank = tMPI_Comm_seek_rank(comm, cur);

    /* we increase our counter, and determine which coll_env we get */
    cev = tMPI_Get_cev(comm, myrank, &synct);

    /* post our pointers */
    /* we set up multiple posts, so no Post_multi */
    cev->met[myrank].tag      = TMPI_ALLTOALL_TAG;
    cev->met[myrank].datatype = sendtype;
    tMPI_Atomic_set( &(cev->met[myrank].n_remaining), cev->N-1 );
    for (i = 0; i < comm->grp.N; i++)
    {
        cev->met[myrank].bufsize[i]   = sendsize;
        cev->met[myrank].buf[i]       = (char*)sendbuf+sendsize*i;
        cev->met[myrank].read_data[i] = FALSE;
    }
    tMPI_Atomic_memory_barrier_rel();
    tMPI_Atomic_set(&(cev->met[myrank].current_sync), synct);

    /* post availability */
    for (i = 0; i < cev->N; i++)
    {
        if (i != myrank)
        {
            tMPI_Event_signal( &(cev->met[i].recv_ev) );
        }
    }

    /* we don't do the copy buffer thing here because it's pointless:
       the processes have to synchronize anyway, because they all
       send and receive. */

    /* do root transfer */
    tMPI_Coll_root_xfer(comm, sendtype, recvtype,
                        sendsize, recvsize,
                        (char*)sendbuf+sendsize*myrank,
                        (char*)recvbuf+recvsize*myrank, &ret);
    cev->met[myrank].read_data[myrank] = TRUE;
    /* and poll data availability */
    n_remaining = cev->N-1;
    while (n_remaining > 0)
    {
#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
        tMPI_Profile_wait_start(cur);
#endif
        tMPI_Event_wait( &(cev->met[myrank]).recv_ev );
#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
        tMPI_Profile_wait_stop(cur, TMPIWAIT_Coll_recv);
#endif
        for (i = 0; i < cev->N; i++)
        {
            if ((!cev->met[myrank].read_data[i]) &&
                (tMPI_Atomic_get(&(cev->met[i].current_sync)) == synct))
            {
                tMPI_Event_process( &(cev->met[myrank]).recv_ev, 1);
                tMPI_Mult_recv(comm, cev, i, myrank, TMPI_ALLTOALL_TAG,
                               recvtype, recvsize, (char*)recvbuf+recvsize*i,
                               &ret);
                if (ret != TMPI_SUCCESS)
                {
                    return ret;
                }
                cev->met[myrank].read_data[i] = TRUE;
                n_remaining--;
            }
        }
    }


    /* and wait until everybody is done copying our data */
    tMPI_Wait_for_others(cev, myrank);

#ifdef TMPI_PROFILE
    tMPI_Profile_count_stop(cur, TMPIFN_Alltoall);
#endif
    return ret;
}