1 /* Copyright (c) 2013-2019. 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.hpp"
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 ****************************************************************************/
33 Coll_alltoallv_ring_light_barrier::alltoallv(const void *send_buff, const int *send_counts, const int *send_disps,
34 MPI_Datatype send_type,
35 void *recv_buff, const int *recv_counts, const int *recv_disps,
36 MPI_Datatype recv_type,
39 MPI_Aint send_chunk, recv_chunk;
41 int i, src, dst, rank, num_procs, next_dst, next_src;
42 int tag = COLL_TAG_ALLTOALLV;
44 char send_sync = 'a', recv_sync = 'b';
45 char *send_ptr = (char *) send_buff;
46 char *recv_ptr = (char *) recv_buff;
49 num_procs = comm->size();
50 send_chunk = send_type->get_extent();
51 recv_chunk = recv_type->get_extent();
53 Request::sendrecv(send_ptr + send_disps[rank] * send_chunk, send_counts[rank], send_type, rank, tag,
54 recv_ptr + recv_disps[rank] * recv_chunk, recv_counts[rank], recv_type, rank, tag,
57 for (i = 1; i < num_procs; i++) {
58 src = (rank - i + num_procs) % num_procs;
59 dst = (rank + i) % num_procs;
61 Request::sendrecv(send_ptr + send_disps[dst] * send_chunk, send_counts[dst], send_type,
62 dst, tag, recv_ptr + recv_disps[src] * recv_chunk, recv_counts[src],
63 recv_type, src, tag, comm, &s);
65 if ((i + 1) < num_procs) {
66 next_src = (rank - (i + 1) + num_procs) % num_procs;
67 next_dst = (rank + (i + 1) + num_procs) % num_procs;
68 Request::sendrecv(&send_sync, 1, MPI_CHAR, next_src, tag,
69 &recv_sync, 1, MPI_CHAR, next_dst, tag, comm, &s);