1 /* Copyright (c) 2013-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "colls_private.h"
8 /*****************************************************************************
10 * Function: alltoall_pair_light_barrier
15 send_buff: send input buffer
16 send_count: number of elements to send
17 send_type: data type of elements being sent
18 recv_buff: receive output buffer
19 recv_count: number of elements to received
20 recv_type: data type of elements being received
23 * Descrp: Function works in P - 1 steps. In step i, node j exchanges data
24 with node i ^ j. Light barriers are inserted between
25 communications in different phases.
29 ****************************************************************************/
31 smpi_coll_tuned_alltoallv_pair_light_barrier(void *send_buff, int *send_counts, int *send_disps,
32 MPI_Datatype send_type,
33 void *recv_buff, int *recv_counts, int *recv_disps,
34 MPI_Datatype recv_type,
37 MPI_Aint send_chunk, recv_chunk;
39 int i, src, dst, rank, num_procs, next_partner;
40 int tag = COLL_TAG_ALLTOALLV; /*, failure = 0; */
42 char send_sync = 'a', recv_sync = 'b';
43 char *send_ptr = (char *) send_buff;
44 char *recv_ptr = (char *) recv_buff;
46 rank = smpi_comm_rank(comm);
47 num_procs = smpi_comm_size(comm);
49 if((num_procs&(num_procs-1)))
50 THROWF(arg_error,0, "alltoallv pair algorithm can't be used with non power of two number of processes ! ");
52 send_chunk = smpi_datatype_get_extent(send_type);
53 recv_chunk = smpi_datatype_get_extent(recv_type);
55 smpi_mpi_sendrecv(send_ptr + send_disps[rank] * send_chunk, send_counts[rank], send_type, rank, tag,
56 recv_ptr + recv_disps[rank] * recv_chunk, recv_counts[rank], recv_type, rank, tag,
59 for (i = 1; i < num_procs; i++) {
62 smpi_mpi_sendrecv(send_ptr + send_disps[dst] * send_chunk, send_counts[dst], send_type,
63 dst, tag, recv_ptr + recv_disps[src] *recv_chunk, recv_counts[dst],
64 recv_type, src, tag, comm, &s);
66 if ((i + 1) < num_procs) {
67 next_partner = rank ^ (i + 1);
68 smpi_mpi_sendrecv(&send_sync, 1, MPI_CHAR, next_partner, tag,
69 &recv_sync, 1, MPI_CHAR, next_partner, tag, comm, &s);