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"
14 Coll_allgather_rdb::allgather(void *sbuf, int send_count,
15 MPI_Datatype send_type, void *rbuf,
16 int recv_count, MPI_Datatype recv_type,
21 MPI_Aint send_chunk, recv_chunk;
23 // local int variables
24 unsigned int i, j, k, dst, send_offset, recv_offset, tree_root;
25 int dst_tree_root, rank_tree_root, last_recv_count = 0, num_procs_completed;
27 int tag = COLL_TAG_ALLGATHER;
28 unsigned int mask = 1;
30 int curr_count = recv_count;
32 // local string variables
33 char *send_ptr = (char *) sbuf;
34 char *recv_ptr = (char *) rbuf;
36 // get size of the communicator, followed by rank
37 unsigned int num_procs = comm->size();
38 unsigned int rank = comm->rank();
40 // get size of single element's type for send buffer and recv buffer
41 send_chunk = send_type->get_extent();
42 recv_chunk = recv_type->get_extent();
44 // multiply size of each element by number of elements to send or recv
45 send_chunk *= send_count;
46 recv_chunk *= recv_count;
48 // perform a local copy
49 Request::sendrecv(send_ptr, send_count, send_type, rank, tag,
50 recv_ptr + rank * recv_chunk, recv_count, recv_type, rank, tag,
54 while (mask < num_procs) {
56 dst_tree_root = dst >> i;
58 rank_tree_root = rank >> i;
60 send_offset = rank_tree_root * send_chunk;
61 recv_offset = dst_tree_root * recv_chunk;
63 if (dst < num_procs) {
64 Request::sendrecv(recv_ptr + send_offset, curr_count, send_type, dst,
65 tag, recv_ptr + recv_offset, mask * recv_count,
66 recv_type, dst, tag, comm, &status);
67 last_recv_count = smpi_mpi_get_count(&status, recv_type);
68 curr_count += last_recv_count;
71 if (dst_tree_root + mask > num_procs) {
72 num_procs_completed = num_procs - rank_tree_root - mask;
73 /* num_procs_completed is the number of processes in this
74 subtree that have all the data. Send data to others
75 in a tree fashion. First find root of current tree
76 that is being divided into two. k is the number of
77 least-significant bits in this process's rank that
78 must be zeroed out to find the rank of the root */
88 offset = recv_chunk * (rank_tree_root + mask);
92 dst = rank ^ tmp_mask;
94 tree_root = rank >> k;
97 /* send only if this proc has data and destination
98 doesn't have data. at any step, multiple processes
99 can send if they have the data */
101 && (rank < tree_root + num_procs_completed)
102 && (dst >= tree_root + num_procs_completed)) {
103 Request::send(recv_ptr + offset, last_recv_count, recv_type, dst,
106 /* last_recv_cnt was set in the previous
107 receive. that's the amount of data to be
110 /* recv only if this proc. doesn't have data and sender
112 else if ((dst < rank)
113 && (dst < tree_root + num_procs_completed)
114 && (rank >= tree_root + num_procs_completed)) {
115 Request::recv(recv_ptr + offset,
116 recv_count * num_procs_completed,
117 recv_type, dst, tag, comm, &status);
118 // num_procs_completed is also equal to the no. of processes
119 // whose data we don't have
120 last_recv_count = smpi_mpi_get_count(&status, recv_type);
121 curr_count += last_recv_count;