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. */
8 * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
9 * University Research and Technology
10 * Corporation. All rights reserved.
11 * Copyright (c) 2004-2009 The University of Tennessee and The University
12 * of Tennessee Research Foundation. All rights
14 * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
15 * University of Stuttgart. All rights reserved.
16 * Copyright (c) 2004-2005 The Regents of the University of California.
17 * All rights reserved.
19 * Additional copyrights may follow
21 /* -*- Mode: C; c-basic-offset:4 ; -*- */
22 /* Copyright (c) 2001-2014, The Ohio State University. All rights
25 * This file is part of the MVAPICH2 software package developed by the
26 * team members of The Ohio State University's Network-Based Computing
27 * Laboratory (NBCL), headed by Professor Dhabaleswar K. (DK) Panda.
29 * For detailed copyright and licensing information, please refer to the
30 * copyright file COPYRIGHT in the top level MVAPICH2 directory.
34 * (C) 2001 by Argonne National Laboratory.
35 * See COPYRIGHT in top-level directory.
37 #include "colls_private.h"
41 int smpi_coll_tuned_allgather_mvapich2_smp(void *sendbuf,int sendcnt, MPI_Datatype sendtype,
42 void *recvbuf, int recvcnt,MPI_Datatype recvtype,
46 int local_rank, local_size;
47 int leader_comm_size = 0;
48 int mpi_errno = MPI_SUCCESS;
49 MPI_Aint recvtype_extent = 0; /* Datatype extent */
50 MPI_Comm shmem_comm, leader_comm;
52 if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL){
53 smpi_comm_init_smp(comm);
56 if(!smpi_comm_is_uniform(comm) || !smpi_comm_is_blocked(comm))
57 THROWF(arg_error,0, "allgather MVAPICH2 smp algorithm can't be used with irregular deployment. Please insure that processes deployed on the same node are contiguous and that each node has the same number of processes");
63 rank = smpi_comm_rank(comm);
64 size = smpi_comm_size(comm);
66 /* extract the rank,size information for the intra-node
68 recvtype_extent=smpi_datatype_get_extent(recvtype);
70 shmem_comm = smpi_comm_get_intra_comm(comm);
71 local_rank = smpi_comm_rank(shmem_comm);
72 local_size = smpi_comm_size(shmem_comm);
74 if (local_rank == 0) {
75 /* Node leader. Extract the rank, size information for the leader
77 leader_comm = smpi_comm_get_leaders_comm(comm);
78 if(leader_comm==MPI_COMM_NULL){
79 leader_comm = MPI_COMM_WORLD;
81 leader_comm_size = smpi_comm_size(leader_comm);
84 /*If there is just one node, after gather itself,
85 * root has all the data and it can do bcast*/
87 mpi_errno = mpi_coll_gather_fun(sendbuf, sendcnt,sendtype,
88 (void*)((char*)recvbuf + (rank * recvcnt * recvtype_extent)),
92 /*Since in allgather all the processes could have
93 * its own data in place*/
94 if(sendbuf == MPI_IN_PLACE) {
95 mpi_errno = mpi_coll_gather_fun((void*)((char*)recvbuf + (rank * recvcnt * recvtype_extent)),
97 recvbuf, recvcnt, recvtype,
100 mpi_errno = mpi_coll_gather_fun(sendbuf, sendcnt,sendtype,
101 recvbuf, recvcnt, recvtype,
105 /* Exchange the data between the node leaders*/
106 if (local_rank == 0 && (leader_comm_size > 1)) {
107 /*When data in each socket is different*/
108 if (smpi_comm_is_uniform(comm) != 1) {
111 int *recvcnts = NULL;
112 int *node_sizes = NULL;
115 node_sizes = smpi_comm_get_non_uniform_map(comm);
117 displs = xbt_malloc(sizeof (int) * leader_comm_size);
118 recvcnts = xbt_malloc(sizeof (int) * leader_comm_size);
119 if (!displs || !recvcnts) {
120 return MPI_ERR_OTHER;
122 recvcnts[0] = node_sizes[0] * recvcnt;
125 for (i = 1; i < leader_comm_size; i++) {
126 displs[i] = displs[i - 1] + node_sizes[i - 1] * recvcnt;
127 recvcnts[i] = node_sizes[i] * recvcnt;
131 void* sendbuf=((char*)recvbuf)+smpi_datatype_get_extent(recvtype)*displs[smpi_comm_rank(leader_comm)];
133 mpi_errno = mpi_coll_allgatherv_fun(sendbuf,
134 (recvcnt*local_size),
142 void* sendtmpbuf=((char*)recvbuf)+smpi_datatype_get_extent(recvtype)*(recvcnt*local_size)*smpi_comm_rank(leader_comm);
146 mpi_errno = smpi_coll_tuned_allgather_mpich(sendtmpbuf,
147 (recvcnt*local_size),
149 recvbuf, (recvcnt*local_size), recvtype,
155 /*Bcast the entire data from node leaders to all other cores*/
156 mpi_errno = mpi_coll_bcast_fun (recvbuf, recvcnt * size, recvtype, 0, shmem_comm);