1 /* Copyright (c) 2013-2017. 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"
9 * Linear functions are copied from the basic coll module. For
10 * some small number of nodes and/or small data sizes they are just as
11 * fast as tuned/tree based segmenting operations and as such may be
12 * selected by the decision functions. These are copied into this module
13 * due to the way we select modules in V1. i.e. in V2 we will handle this
14 * differently and so will not have to duplicate code.
15 * GEF Oct05 after asking Jeff.
20 Coll_alltoallv_ompi_basic_linear::alltoallv(void *sbuf, int *scounts, int *sdisps,
22 void *rbuf, int *rcounts, int *rdisps,
33 MPI_Request *ireqs= static_cast<MPI_Request*>(xbt_malloc(sizeof(MPI_Request) * size * 2));
35 "coll:tuned:alltoallv_intra_basic_linear rank %d", rank);
37 sext=sdtype->get_extent();
38 rext=rdtype->get_extent();
40 /* Simple optimization - handle send to self first */
41 psnd = ((char *) sbuf) + (sdisps[rank] * sext);
42 prcv = ((char *) rbuf) + (rdisps[rank] * rext);
43 if (0 != scounts[rank]) {
44 Datatype::copy(psnd, scounts[rank], sdtype,
45 prcv, rcounts[rank], rdtype);
48 /* If only one process, we're done. */
53 /* Now, initiate all send/recv to/from others. */
57 /* Post all receives first */
58 for (i = 0; i < size; ++i) {
59 if (i == rank || 0 == rcounts[i]) {
63 prcv = ((char *) rbuf) + (rdisps[i] * rext);
65 *preq = Request::irecv_init(prcv, rcounts[i], rdtype,
66 i, COLL_TAG_ALLTOALLV, comm
73 /* Now post all sends */
74 for (i = 0; i < size; ++i) {
75 if (i == rank || 0 == scounts[i]) {
79 psnd = ((char *) sbuf) + (sdisps[i] * sext);
80 *preq=Request::isend_init(psnd, scounts[i], sdtype,
81 i, COLL_TAG_ALLTOALLV, comm
87 /* Start your engines. This will never return an error. */
88 Request::startall(nreqs, ireqs);
90 /* Wait for them all. If there's an error, note that we don't care
91 * what the error was -- just that there *was* an error. The PML
92 * will finish all requests, even if one or more of them fail.
93 * i.e., by the end of this call, all the requests are free-able.
94 * So free them anyway -- even if there was an error, and return the
95 * error after we free everything. */
96 Request::waitall(nreqs, ireqs,
99 /* Free the requests. */
100 for (i = 0; i < nreqs; ++i) {
101 if(ireqs[i]!=MPI_REQUEST_NULL)
102 Request::unref(&ireqs[i]);