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_ring_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 - i -> j -> j + i.
24 Light barriers are inserted between communications in different
29 ****************************************************************************/
31 smpi_coll_tuned_alltoall_ring_light_barrier(void *send_buff, int send_count,
32 MPI_Datatype send_type,
33 void *recv_buff, int recv_count,
34 MPI_Datatype recv_type,
37 MPI_Aint send_chunk, recv_chunk;
39 int i, src, dst, rank, num_procs, next_dst, next_src;
40 int tag = COLL_TAG_ALLTOALL;
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);
48 send_chunk = smpi_datatype_get_extent(send_type);
49 recv_chunk = smpi_datatype_get_extent(recv_type);
51 send_chunk *= send_count;
52 recv_chunk *= recv_count;
54 smpi_mpi_sendrecv(send_ptr + rank * send_chunk, send_count, send_type, rank, tag,
55 recv_ptr + rank * recv_chunk, recv_count, recv_type, rank, tag,
58 for (i = 1; i < num_procs; i++) {
59 src = (rank - i + num_procs) % num_procs;
60 dst = (rank + i) % num_procs;
62 smpi_mpi_sendrecv(send_ptr + dst * send_chunk, send_count, send_type,
63 dst, tag, recv_ptr + src * recv_chunk, recv_count,
64 recv_type, src, tag, comm, &s);
66 if ((i + 1) < num_procs) {
67 next_src = (rank - (i + 1) + num_procs) % num_procs;
68 next_dst = (rank + (i + 1) + num_procs) % num_procs;
69 smpi_mpi_sendrecv(&send_sync, 1, MPI_CHAR, next_src, tag,
70 &recv_sync, 1, MPI_CHAR, next_dst, tag, comm, &s);