src/smpi/colls/reduce/reduce-mvapich-knomial.cpp

   1 /* Copyright (c) 2013-2014. The SimGrid Team.
   2  * All rights reserved.                                                     */
   3
   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. */
   6
   7 /*
   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) 2004-2012 The University of Tennessee and The University
  12  *                         of Tennessee Research Foundation.  All rights
  13  *                         reserved.
  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.
  18  * Copyright (c) 2008      Sun Microsystems, Inc.  All rights reserved.
  19  * Copyright (c) 2009      University of Houston. All rights reserved.
  20  *
  21  * Additional copyrights may follow
  22  */
  23
  24 /*
  25  *
  26  *  (C) 2001 by Argonne National Laboratory.
  27  *      See COPYRIGHT in top-level directory.
  28  */
  29 /* Copyright (c) 2001-2014, The Ohio State University. All rights
  30  * reserved.
  31  *
  32  * This file is part of the MVAPICH2 software package developed by the
  33  * team members of The Ohio State University's Network-Based Computing
  34  * Laboratory (NBCL), headed by Professor Dhabaleswar K. (DK) Panda.
  35  *
  36  * For detailed copyright and licensing information, please refer to the
  37  * copyright file COPYRIGHT in the top level MVAPICH2 directory.
  38  *
  39  */
  40
  41 #include "../colls_private.h"
  42 extern int mv2_reduce_intra_knomial_factor;
  43 extern int mv2_reduce_inter_knomial_factor;
  44
  45 #define SMPI_DEFAULT_KNOMIAL_FACTOR 4
  46
  47 //        int mv2_reduce_knomial_factor = 2;
  48
  49
  50
  51 static int MPIR_Reduce_knomial_trace(int root, int reduce_knomial_factor,
  52         MPI_Comm comm, int *dst, int *expected_send_count,
  53         int *expected_recv_count, int **src_array)
  54 {
  55     int mask=0x1, k, comm_size, src, rank, relative_rank, lroot=0;
  56     int orig_mask=0x1;
  57     int recv_iter=0, send_iter=0;
  58     int *knomial_reduce_src_array=NULL;
  59     comm_size =  comm->size();
  60     rank = comm->rank();
  61
  62     lroot = root;
  63     relative_rank = (rank - lroot + comm_size) % comm_size;
  64
  65     /* First compute to whom we need to send data */
  66     while (mask < comm_size) {
  67         if (relative_rank % (reduce_knomial_factor*mask)) {
  68             *dst = relative_rank/(reduce_knomial_factor*mask)*
  69                 (reduce_knomial_factor*mask)+root;
  70             if (*dst >= comm_size) {
  71                 *dst -= comm_size;
  72             }
  73             send_iter++;
  74             break;
  75         }
  76         mask *= reduce_knomial_factor;
  77     }
  78     mask /= reduce_knomial_factor;
  79
  80     /* Now compute how many children we have in the knomial-tree */
  81     orig_mask = mask;
  82     while (mask > 0) {
  83         for(k=1;k<reduce_knomial_factor;k++) {
  84             if (relative_rank + mask*k < comm_size) {
  85                 recv_iter++;
  86             }
  87         }
  88         mask /= reduce_knomial_factor;
  89     }
  90
  91     /* Finally, fill up the src array */
  92     if(recv_iter > 0) {
  93         knomial_reduce_src_array = static_cast<int*>(smpi_get_tmp_sendbuffer(sizeof(int)*recv_iter));
  94     }
  95
  96     mask = orig_mask;
  97     recv_iter=0;
  98     while (mask > 0) {
  99         for(k=1;k<reduce_knomial_factor;k++) {
 100             if (relative_rank + mask*k < comm_size) {
 101                 src = rank + mask*k;
 102                 if (src >= comm_size) {
 103                     src -= comm_size;
 104                 }
 105                 knomial_reduce_src_array[recv_iter++] = src;
 106             }
 107         }
 108         mask /= reduce_knomial_factor;
 109     }
 110
 111     *expected_recv_count = recv_iter;
 112     *expected_send_count = send_iter;
 113     *src_array = knomial_reduce_src_array;
 114     return 0;
 115 }
 116
 117 int smpi_coll_tuned_reduce_mvapich2_knomial (
 118         void *sendbuf,
 119         void *recvbuf,
 120         int count,
 121         MPI_Datatype datatype,
 122         MPI_Op op,
 123         int root,
 124         MPI_Comm comm)
 125 {
 126     int mpi_errno = MPI_SUCCESS;
 127     int rank, is_commutative;
 128     int src, k;
 129     MPI_Request send_request;
 130     int index=0;
 131     MPI_Aint true_lb, true_extent, extent;
 132     MPI_Status status;
 133     int recv_iter=0, dst=-1, expected_send_count, expected_recv_count;
 134     int *src_array=NULL;
 135     void **tmp_buf=NULL;
 136     MPI_Request *requests=NULL;
 137
 138
 139     if (count == 0) return MPI_SUCCESS;
 140
 141     rank = comm->rank();
 142
 143     /* Create a temporary buffer */
 144
 145     datatype->extent(&true_lb, &true_extent);
 146     extent = datatype->get_extent();
 147
 148     is_commutative =  (op==MPI_OP_NULL || op->is_commutative());
 149
 150     if (rank != root) {
 151         recvbuf=(void *)smpi_get_tmp_recvbuffer(count*(MAX(extent,true_extent)));
 152         recvbuf = (void *)((char*)recvbuf - true_lb);
 153     }
 154
 155     if ((rank != root) || (sendbuf != MPI_IN_PLACE)) {
 156         mpi_errno = Datatype::copy(sendbuf, count, datatype, recvbuf,
 157                 count, datatype);
 158     }
 159
 160
 161     if(mv2_reduce_intra_knomial_factor<0)
 162       {
 163         mv2_reduce_intra_knomial_factor = SMPI_DEFAULT_KNOMIAL_FACTOR;
 164       }
 165     if(mv2_reduce_inter_knomial_factor<0)
 166       {
 167         mv2_reduce_inter_knomial_factor = SMPI_DEFAULT_KNOMIAL_FACTOR;
 168       }
 169
 170
 171     MPIR_Reduce_knomial_trace(root, mv2_reduce_intra_knomial_factor, comm,
 172            &dst, &expected_send_count, &expected_recv_count, &src_array);
 173
 174     if(expected_recv_count > 0 ) {
 175         tmp_buf  = static_cast<void**>(xbt_malloc(sizeof(void *)*expected_recv_count));
 176         requests = static_cast<MPI_Request*>(xbt_malloc(sizeof(MPI_Request)*expected_recv_count));
 177         for(k=0; k < expected_recv_count; k++ ) {
 178             tmp_buf[k] = smpi_get_tmp_sendbuffer(count*(MAX(extent,true_extent)));
 179             tmp_buf[k] = (void *)((char*)tmp_buf[k] - true_lb);
 180         }
 181
 182         while(recv_iter  < expected_recv_count) {
 183             src = src_array[expected_recv_count - (recv_iter+1)];
 184
 185             requests[recv_iter]=Request::irecv (tmp_buf[recv_iter], count, datatype ,src,
 186                     COLL_TAG_REDUCE, comm);
 187             recv_iter++;
 188
 189         }
 190
 191         recv_iter=0;
 192         while(recv_iter < expected_recv_count) {
 193             index=Request::waitany(expected_recv_count, requests,
 194                     &status);
 195             recv_iter++;
 196
 197             if (is_commutative) {
 198               if(op!=MPI_OP_NULL) op->apply( tmp_buf[index], recvbuf, &count, datatype);
 199             }
 200         }
 201
 202         for(k=0; k < expected_recv_count; k++ ) {
 203             smpi_free_tmp_buffer(tmp_buf[k]);
 204         }
 205         xbt_free(tmp_buf);
 206         xbt_free(requests);
 207     }
 208
 209     if(src_array != NULL) {
 210         xbt_free(src_array);
 211     }
 212
 213     if(rank != root) {
 214         send_request=Request::isend(recvbuf,count, datatype, dst,
 215                 COLL_TAG_REDUCE,comm);
 216
 217         Request::waitall(1, &send_request, &status);
 218
 219         smpi_free_tmp_buffer((void *)((char*)recvbuf + true_lb));
 220     }
 221
 222     /* --END ERROR HANDLING-- */
 223
 224     return mpi_errno;
 225 }