src/smpi/colls/reduce/reduce-ompi.cpp

   1 /* Copyright (c) 2013-2021. 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) 2004-2009 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  *
  19  * Additional copyrights may follow
  20  */
  21
  22 #include "../coll_tuned_topo.hpp"
  23 #include "../colls_private.hpp"
  24
  25 namespace simgrid{
  26 namespace smpi{
  27
  28 int smpi_coll_tuned_ompi_reduce_generic(const void* sendbuf, void* recvbuf, int original_count,
  29                                     MPI_Datatype datatype, MPI_Op  op,
  30                                     int root, MPI_Comm comm,
  31                                     ompi_coll_tree_t* tree, int count_by_segment,
  32                                     int max_outstanding_reqs );
  33 /**
  34  * This is a generic implementation of the reduce protocol. It used the tree
  35  * provided as an argument and execute all operations using a segment of
  36  * count times a datatype.
  37  * For the last communication it will update the count in order to limit
  38  * the number of datatype to the original count (original_count)
  39  *
  40  * Note that for non-commutative operations we cannot save memory copy
  41  * for the first block: thus we must copy sendbuf to accumbuf on intermediate
  42  * to keep the optimized loop happy.
  43  */
  44 int smpi_coll_tuned_ompi_reduce_generic(const void* sendbuf, void* recvbuf, int original_count,
  45                                     MPI_Datatype datatype, MPI_Op  op,
  46                                     int root, MPI_Comm comm,
  47                                     ompi_coll_tree_t* tree, int count_by_segment,
  48                                     int max_outstanding_reqs )
  49 {
  50   unsigned char *inbuf[2] = {nullptr, nullptr}, *inbuf_free[2] = {nullptr, nullptr};
  51   unsigned char *accumbuf = nullptr, *accumbuf_free = nullptr;
  52   const unsigned char *local_op_buffer = nullptr, *sendtmpbuf = nullptr;
  53   ptrdiff_t extent, lower_bound, segment_increment;
  54   MPI_Request reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL};
  55   int num_segments, line, ret, segindex, i, rank;
  56   int recvcount, prevcount, inbi;
  57
  58   /**
  59    * Determine number of segments and number of elements
  60    * sent per operation
  61    */
  62   datatype->extent(&lower_bound, &extent);
  63   num_segments      = (original_count + count_by_segment - 1) / count_by_segment;
  64   segment_increment = count_by_segment * extent;
  65
  66   sendtmpbuf = static_cast<const unsigned char*>(sendbuf);
  67   if (sendbuf == MPI_IN_PLACE) {
  68     sendtmpbuf = static_cast<const unsigned char*>(recvbuf);
  69     }
  70
  71     XBT_DEBUG("coll:tuned:reduce_generic count %d, msg size %lu, segsize %lu, max_requests %d", original_count,
  72               (unsigned long)(num_segments * segment_increment), (unsigned long)segment_increment,
  73               max_outstanding_reqs);
  74
  75     rank = comm->rank();
  76
  77     /* non-leaf nodes - wait for children to send me data & forward up
  78        (if needed) */
  79     if( tree->tree_nextsize > 0 ) {
  80         ptrdiff_t true_extent, real_segment_size;
  81         true_extent=datatype->get_extent();
  82
  83         /* handle non existent recv buffer (i.e. its NULL) and
  84            protect the recv buffer on non-root nodes */
  85         accumbuf = static_cast<unsigned char*>(recvbuf);
  86         if (nullptr == accumbuf || root != rank) {
  87           /* Allocate temporary accumulator buffer. */
  88           accumbuf_free = smpi_get_tmp_sendbuffer(true_extent + (original_count - 1) * extent);
  89           if (accumbuf_free == nullptr) {
  90             line = __LINE__;
  91             ret  = -1;
  92             goto error_hndl;
  93           }
  94           accumbuf = accumbuf_free - lower_bound;
  95         }
  96
  97         /* If this is a non-commutative operation we must copy
  98            sendbuf to the accumbuf, in order to simplify the loops */
  99         if ((op != MPI_OP_NULL && not op->is_commutative())) {
 100           Datatype::copy(sendtmpbuf, original_count, datatype, accumbuf, original_count, datatype);
 101         }
 102         /* Allocate two buffers for incoming segments */
 103         real_segment_size = true_extent + (count_by_segment - 1) * extent;
 104         inbuf_free[0]     = smpi_get_tmp_recvbuffer(real_segment_size);
 105         if (inbuf_free[0] == nullptr) {
 106           line = __LINE__;
 107           ret  = -1;
 108           goto error_hndl;
 109         }
 110         inbuf[0] = inbuf_free[0] - lower_bound;
 111         /* if there is chance to overlap communication -
 112            allocate second buffer */
 113         if( (num_segments > 1) || (tree->tree_nextsize > 1) ) {
 114           inbuf_free[1] = smpi_get_tmp_recvbuffer(real_segment_size);
 115           if (inbuf_free[1] == nullptr) {
 116             line = __LINE__;
 117             ret  = -1;
 118             goto error_hndl;
 119             }
 120             inbuf[1] = inbuf_free[1] - lower_bound;
 121         }
 122
 123         /* reset input buffer index and receive count */
 124         inbi = 0;
 125         recvcount = 0;
 126         /* for each segment */
 127         for( segindex = 0; segindex <= num_segments; segindex++ ) {
 128             prevcount = recvcount;
 129             /* recvcount - number of elements in current segment */
 130             recvcount = count_by_segment;
 131             if( segindex == (num_segments-1) )
 132                 recvcount = original_count - count_by_segment * segindex;
 133
 134             /* for each child */
 135             for( i = 0; i < tree->tree_nextsize; i++ ) {
 136                 /**
 137                  * We try to overlap communication:
 138                  * either with next segment or with the next child
 139                  */
 140                 /* post irecv for current segindex on current child */
 141                 if( segindex < num_segments ) {
 142                     void* local_recvbuf = inbuf[inbi];
 143                     if( 0 == i ) {
 144                         /* for the first step (1st child per segment) and
 145                          * commutative operations we might be able to irecv
 146                          * directly into the accumulate buffer so that we can
 147                          * reduce(op) this with our sendbuf in one step as
 148                          * ompi_op_reduce only has two buffer pointers,
 149                          * this avoids an extra memory copy.
 150                          *
 151                          * BUT if the operation is non-commutative or
 152                          * we are root and are USING MPI_IN_PLACE this is wrong!
 153                          */
 154                         if(  (op==MPI_OP_NULL || op->is_commutative()) &&
 155                             !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
 156                             local_recvbuf = accumbuf + segindex * segment_increment;
 157                         }
 158                     }
 159
 160                     reqs[inbi]=Request::irecv(local_recvbuf, recvcount, datatype,
 161                                              tree->tree_next[i],
 162                                              COLL_TAG_REDUCE, comm
 163                                              );
 164                 }
 165                 /* wait for previous req to complete, if any.
 166                    if there are no requests reqs[inbi ^1] will be
 167                    MPI_REQUEST_NULL. */
 168                 /* wait on data from last child for previous segment */
 169                 Request::waitall( 1, &reqs[inbi ^ 1],
 170                                              MPI_STATUSES_IGNORE );
 171                 local_op_buffer = inbuf[inbi ^ 1];
 172                 if( i > 0 ) {
 173                     /* our first operation is to combine our own [sendbuf] data
 174                      * with the data we recvd from down stream (but only
 175                      * the operation is commutative and if we are not root and
 176                      * not using MPI_IN_PLACE)
 177                      */
 178                     if( 1 == i ) {
 179                         if( (op==MPI_OP_NULL || op->is_commutative())&&
 180                             !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
 181                             local_op_buffer = sendtmpbuf + segindex * segment_increment;
 182                         }
 183                     }
 184                     /* apply operation */
 185                     if(op!=MPI_OP_NULL) op->apply( local_op_buffer,
 186                                    accumbuf + segindex * segment_increment,
 187                                    &recvcount, datatype );
 188                 } else if ( segindex > 0 ) {
 189                     void* accumulator = accumbuf + (segindex-1) * segment_increment;
 190                     if( tree->tree_nextsize <= 1 ) {
 191                         if(  (op==MPI_OP_NULL || op->is_commutative()) &&
 192                             !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
 193                             local_op_buffer = sendtmpbuf + (segindex-1) * segment_increment;
 194                         }
 195                     }
 196                     if(op!=MPI_OP_NULL) op->apply( local_op_buffer, accumulator, &prevcount,
 197                                    datatype );
 198
 199                     /* all reduced on available data this step (i) complete,
 200                      * pass to the next process unless you are the root.
 201                      */
 202                     if (rank != tree->tree_root) {
 203                         /* send combined/accumulated data to parent */
 204                         Request::send( accumulator, prevcount,
 205                                                   datatype, tree->tree_prev,
 206                                                   COLL_TAG_REDUCE,
 207                                                   comm);
 208                     }
 209
 210                     /* we stop when segindex = number of segments
 211                        (i.e. we do num_segment+1 steps for pipelining */
 212                     if (segindex == num_segments) break;
 213                 }
 214
 215                 /* update input buffer index */
 216                 inbi = inbi ^ 1;
 217             } /* end of for each child */
 218         } /* end of for each segment */
 219
 220         /* clean up */
 221         smpi_free_tmp_buffer(inbuf_free[0]);
 222         smpi_free_tmp_buffer(inbuf_free[1]);
 223         smpi_free_tmp_buffer(accumbuf_free);
 224     }
 225
 226     /* leaf nodes
 227        Depending on the value of max_outstanding_reqs and
 228        the number of segments we have two options:
 229        - send all segments using blocking send to the parent, or
 230        - avoid overflooding the parent nodes by limiting the number of
 231        outstanding requests to max_oustanding_reqs.
 232        TODO/POSSIBLE IMPROVEMENT: If there is a way to determine the eager size
 233        for the current communication, synchronization should be used only
 234        when the message/segment size is smaller than the eager size.
 235     */
 236     else {
 237
 238         /* If the number of segments is less than a maximum number of outstanding
 239            requests or there is no limit on the maximum number of outstanding
 240            requests, we send data to the parent using blocking send */
 241         if ((0 == max_outstanding_reqs) ||
 242             (num_segments <= max_outstanding_reqs)) {
 243
 244             segindex = 0;
 245             while ( original_count > 0) {
 246                 if (original_count < count_by_segment) {
 247                     count_by_segment = original_count;
 248                 }
 249                 Request::send((char*)sendbuf +
 250                                          segindex * segment_increment,
 251                                          count_by_segment, datatype,
 252                                          tree->tree_prev,
 253                                          COLL_TAG_REDUCE,
 254                                          comm) ;
 255                 segindex++;
 256                 original_count -= count_by_segment;
 257             }
 258         }
 259
 260         /* Otherwise, introduce flow control:
 261            - post max_outstanding_reqs non-blocking synchronous send,
 262            - for remaining segments
 263            - wait for a ssend to complete, and post the next one.
 264            - wait for all outstanding sends to complete.
 265         */
 266         else {
 267
 268             int creq = 0;
 269             auto* sreq = new (std::nothrow) MPI_Request[max_outstanding_reqs];
 270             if (nullptr == sreq) {
 271               line = __LINE__;
 272               ret  = -1;
 273               goto error_hndl;
 274             }
 275
 276             /* post first group of requests */
 277             for (segindex = 0; segindex < max_outstanding_reqs; segindex++) {
 278                 sreq[segindex]=Request::isend((char*)sendbuf +
 279                                           segindex * segment_increment,
 280                                           count_by_segment, datatype,
 281                                           tree->tree_prev,
 282                                           COLL_TAG_REDUCE,
 283                                           comm);
 284                 original_count -= count_by_segment;
 285             }
 286
 287             creq = 0;
 288             while ( original_count > 0 ) {
 289                 /* wait on a posted request to complete */
 290                 Request::wait(&sreq[creq], MPI_STATUS_IGNORE);
 291                 sreq[creq] = MPI_REQUEST_NULL;
 292
 293                 if( original_count < count_by_segment ) {
 294                     count_by_segment = original_count;
 295                 }
 296                 sreq[creq]=Request::isend((char*)sendbuf +
 297                                           segindex * segment_increment,
 298                                           count_by_segment, datatype,
 299                                           tree->tree_prev,
 300                                           COLL_TAG_REDUCE,
 301                                           comm );
 302                 creq = (creq + 1) % max_outstanding_reqs;
 303                 segindex++;
 304                 original_count -= count_by_segment;
 305             }
 306
 307             /* Wait on the remaining request to complete */
 308             Request::waitall( max_outstanding_reqs, sreq,
 309                                          MPI_STATUSES_IGNORE );
 310
 311             /* free requests */
 312             delete[] sreq;
 313         }
 314     }
 315     ompi_coll_tuned_topo_destroy_tree(&tree);
 316     return MPI_SUCCESS;
 317
 318  error_hndl:  /* error handler */
 319     XBT_DEBUG("ERROR_HNDL: node %d file %s line %d error %d\n",
 320                    rank, __FILE__, line, ret );
 321     smpi_free_tmp_buffer(inbuf_free[0]);
 322     smpi_free_tmp_buffer(inbuf_free[1]);
 323     smpi_free_tmp_buffer(accumbuf);
 324     return ret;
 325 }
 326
 327 /* Attention: this version of the reduce operations does not
 328    work for:
 329    - non-commutative operations
 330    - segment sizes which are not multiplies of the extent of the datatype
 331      meaning that at least one datatype must fit in the segment !
 332 */
 333
 334
 335 int reduce__ompi_chain(const void *sendbuf, void *recvbuf, int count,
 336                        MPI_Datatype datatype,
 337                        MPI_Op  op, int root,
 338                        MPI_Comm  comm
 339                        )
 340 {
 341     uint32_t segsize=64*1024;
 342     int segcount = count;
 343     size_t typelng;
 344     int fanout = comm->size()/2;
 345
 346     XBT_DEBUG("coll:tuned:reduce_intra_chain rank %d fo %d ss %5u", comm->rank(), fanout, segsize);
 347
 348     /**
 349      * Determine number of segments and number of elements
 350      * sent per operation
 351      */
 352     typelng = datatype->size();
 353
 354     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
 355
 356     return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
 357                                            op, root, comm,
 358                                            ompi_coll_tuned_topo_build_chain(fanout, comm, root),
 359                                            segcount, 0 );
 360 }
 361
 362
 363 int reduce__ompi_pipeline(const void *sendbuf, void *recvbuf,
 364                           int count, MPI_Datatype datatype,
 365                           MPI_Op  op, int root,
 366                           MPI_Comm  comm  )
 367 {
 368
 369     uint32_t segsize;
 370     int segcount = count;
 371     size_t typelng;
 372 //    COLL_TUNED_UPDATE_PIPELINE( comm, tuned_module, root );
 373
 374     /**
 375      * Determine number of segments and number of elements
 376      * sent per operation
 377      */
 378     const double a2 =  0.0410 / 1024.0; /* [1/B] */
 379     const double b2 =  9.7128;
 380     const double a4 =  0.0033 / 1024.0; /* [1/B] */
 381     const double b4 =  1.6761;
 382     typelng= datatype->size();
 383     int communicator_size = comm->size();
 384     size_t message_size = typelng * count;
 385
 386     if (communicator_size > (a2 * message_size + b2)) {
 387         // Pipeline_1K
 388         segsize = 1024;
 389     }else if (communicator_size > (a4 * message_size + b4)) {
 390         // Pipeline_32K
 391         segsize = 32*1024;
 392     } else {
 393         // Pipeline_64K
 394         segsize = 64*1024;
 395     }
 396
 397     XBT_DEBUG("coll:tuned:reduce_intra_pipeline rank %d ss %5u", comm->rank(), segsize);
 398
 399     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
 400
 401     return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
 402                                            op, root, comm,
 403                                            ompi_coll_tuned_topo_build_chain( 1, comm, root),
 404                                            segcount, 0);
 405 }
 406
 407 int reduce__ompi_binary(const void *sendbuf, void *recvbuf,
 408                         int count, MPI_Datatype datatype,
 409                         MPI_Op  op, int root,
 410                         MPI_Comm  comm)
 411 {
 412     uint32_t segsize;
 413     int segcount = count;
 414     size_t typelng;
 415
 416
 417
 418     /**
 419      * Determine number of segments and number of elements
 420      * sent per operation
 421      */
 422     typelng=datatype->size();
 423
 424         // Binary_32K
 425     segsize = 32*1024;
 426
 427     XBT_DEBUG("coll:tuned:reduce_intra_binary rank %d ss %5u", comm->rank(), segsize);
 428
 429     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
 430
 431     return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
 432                                            op, root, comm,
 433                                            ompi_coll_tuned_topo_build_tree(2, comm, root),
 434                                            segcount, 0);
 435 }
 436
 437 int reduce__ompi_binomial(const void *sendbuf, void *recvbuf,
 438                           int count, MPI_Datatype datatype,
 439                           MPI_Op  op, int root,
 440                           MPI_Comm  comm)
 441 {
 442
 443     uint32_t segsize=0;
 444     int segcount = count;
 445     size_t typelng;
 446
 447     const double a1 =  0.6016 / 1024.0; /* [1/B] */
 448     const double b1 =  1.3496;
 449
 450 //    COLL_TUNED_UPDATE_IN_ORDER_BMTREE( comm, tuned_module, root );
 451
 452     /**
 453      * Determine number of segments and number of elements
 454      * sent per operation
 455      */
 456     typelng= datatype->size();
 457     int communicator_size = comm->size();
 458     size_t message_size = typelng * count;
 459     if (((communicator_size < 8) && (message_size < 20480)) ||
 460                (message_size < 2048) || (count <= 1)) {
 461         /* Binomial_0K */
 462         segsize = 0;
 463     } else if (communicator_size > (a1 * message_size + b1)) {
 464         // Binomial_1K
 465         segsize = 1024;
 466     }
 467
 468     XBT_DEBUG("coll:tuned:reduce_intra_binomial rank %d ss %5u", comm->rank(), segsize);
 469     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
 470
 471     return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
 472                                            op, root, comm,
 473                                            ompi_coll_tuned_topo_build_in_order_bmtree(comm, root),
 474                                            segcount, 0);
 475 }
 476
 477 /*
 478  * reduce_intra_in_order_binary
 479  *
 480  * Function:      Logarithmic reduce operation for non-commutative operations.
 481  * Accepts:       same as MPI_Reduce()
 482  * Returns:       MPI_SUCCESS or error code
 483  */
 484 int reduce__ompi_in_order_binary(const void *sendbuf, void *recvbuf,
 485                                  int count,
 486                                  MPI_Datatype datatype,
 487                                  MPI_Op  op, int root,
 488                                  MPI_Comm  comm)
 489 {
 490     uint32_t segsize=0;
 491     int ret;
 492     int rank, size, io_root;
 493     int segcount = count;
 494     size_t typelng;
 495
 496     rank = comm->rank();
 497     size = comm->size();
 498     XBT_DEBUG("coll:tuned:reduce_intra_in_order_binary rank %d ss %5u", rank, segsize);
 499
 500     /**
 501      * Determine number of segments and number of elements
 502      * sent per operation
 503      */
 504     typelng=datatype->size();
 505     COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
 506
 507     /* An in-order binary tree must use root (size-1) to preserve the order of
 508        operations.  Thus, if root is not rank (size - 1), then we must handle
 509        1. MPI_IN_PLACE option on real root, and
 510        2. we must allocate temporary recvbuf on rank (size - 1).
 511        Note that generic function must be careful not to switch order of
 512        operations for non-commutative ops.
 513     */
 514     io_root = size - 1;
 515     const void* use_this_sendbuf = sendbuf;
 516     void* use_this_recvbuf       = recvbuf;
 517     unsigned char* tmp_sendbuf   = nullptr;
 518     unsigned char* tmp_recvbuf   = nullptr;
 519     if (io_root != root) {
 520         ptrdiff_t text, ext;
 521
 522         ext=datatype->get_extent();
 523         text=datatype->get_extent();
 524
 525         if ((root == rank) && (MPI_IN_PLACE == sendbuf)) {
 526           tmp_sendbuf = smpi_get_tmp_sendbuffer(text + (count - 1) * ext);
 527           if (nullptr == tmp_sendbuf) {
 528             return MPI_ERR_INTERN;
 529           }
 530           Datatype::copy(recvbuf, count, datatype, tmp_sendbuf, count, datatype);
 531           use_this_sendbuf = tmp_sendbuf;
 532         } else if (io_root == rank) {
 533           tmp_recvbuf = smpi_get_tmp_recvbuffer(text + (count - 1) * ext);
 534           if (nullptr == tmp_recvbuf) {
 535             return MPI_ERR_INTERN;
 536           }
 537           use_this_recvbuf = tmp_recvbuf;
 538         }
 539     }
 540
 541     /* Use generic reduce with in-order binary tree topology and io_root */
 542     ret = smpi_coll_tuned_ompi_reduce_generic( use_this_sendbuf, use_this_recvbuf, count, datatype,
 543                                           op, io_root, comm,
 544                                           ompi_coll_tuned_topo_build_in_order_bintree(comm),
 545                                           segcount, 0 );
 546     if (MPI_SUCCESS != ret) { return ret; }
 547
 548     /* Clean up */
 549     if (io_root != root) {
 550         if (root == rank) {
 551             /* Receive result from rank io_root to recvbuf */
 552             Request::recv(recvbuf, count, datatype, io_root,
 553                                     COLL_TAG_REDUCE, comm,
 554                                     MPI_STATUS_IGNORE);
 555             if (MPI_IN_PLACE == sendbuf) {
 556               smpi_free_tmp_buffer(tmp_sendbuf);
 557             }
 558
 559         } else if (io_root == rank) {
 560             /* Send result from use_this_recvbuf to root */
 561             Request::send(use_this_recvbuf, count, datatype, root,
 562                                     COLL_TAG_REDUCE,
 563                                     comm);
 564             smpi_free_tmp_buffer(tmp_recvbuf);
 565         }
 566     }
 567
 568     return MPI_SUCCESS;
 569 }
 570
 571 /*
 572  * Linear functions are copied from the BASIC coll module
 573  * they do not segment the message and are simple implementations
 574  * but for some small number of nodes and/or small data sizes they
 575  * are just as fast as tuned/tree based segmenting operations
 576  * and as such may be selected by the decision functions
 577  * These are copied into this module due to the way we select modules
 578  * in V1. i.e. in V2 we will handle this differently and so will not
 579  * have to duplicate code.
 580  * GEF Oct05 after asking Jeff.
 581  */
 582
 583 /* copied function (with appropriate renaming) starts here */
 584
 585 /*
 586  *  reduce_lin_intra
 587  *
 588  *  Function:   - reduction using O(N) algorithm
 589  *  Accepts:    - same as MPI_Reduce()
 590  *  Returns:    - MPI_SUCCESS or error code
 591  */
 592
 593 int reduce__ompi_basic_linear(const void *sbuf, void *rbuf, int count,
 594                               MPI_Datatype dtype,
 595                               MPI_Op op,
 596                               int root,
 597                               MPI_Comm comm)
 598 {
 599     int i, rank, size;
 600     ptrdiff_t true_extent, lb, extent;
 601     unsigned char* free_buffer  = nullptr;
 602     unsigned char* pml_buffer   = nullptr;
 603     unsigned char* inplace_temp = nullptr;
 604     const unsigned char* inbuf;
 605
 606     /* Initialize */
 607
 608     rank = comm->rank();
 609     size = comm->size();
 610
 611     XBT_DEBUG("coll:tuned:reduce_intra_basic_linear rank %d", rank);
 612
 613     /* If not root, send data to the root. */
 614
 615     if (rank != root) {
 616         Request::send(sbuf, count, dtype, root,
 617                                 COLL_TAG_REDUCE,
 618                                 comm);
 619         return MPI_SUCCESS;
 620     }
 621
 622     /* see discussion in ompi_coll_basic_reduce_lin_intra about
 623        extent and true extent */
 624     /* for reducing buffer allocation lengths.... */
 625
 626     dtype->extent(&lb, &extent);
 627     true_extent = dtype->get_extent();
 628
 629     if (MPI_IN_PLACE == sbuf) {
 630         sbuf = rbuf;
 631         inplace_temp = smpi_get_tmp_recvbuffer(true_extent + (count - 1) * extent);
 632         if (nullptr == inplace_temp) {
 633           return -1;
 634         }
 635         rbuf = inplace_temp - lb;
 636     }
 637
 638     if (size > 1) {
 639       free_buffer = smpi_get_tmp_recvbuffer(true_extent + (count - 1) * extent);
 640       pml_buffer  = free_buffer - lb;
 641     }
 642
 643     /* Initialize the receive buffer. */
 644
 645     if (rank == (size - 1)) {
 646         Datatype::copy((char*)sbuf, count, dtype,(char*)rbuf, count, dtype);
 647     } else {
 648         Request::recv(rbuf, count, dtype, size - 1,
 649                                 COLL_TAG_REDUCE, comm,
 650                                 MPI_STATUS_IGNORE);
 651     }
 652
 653     /* Loop receiving and calling reduction function (C or Fortran). */
 654
 655     for (i = size - 2; i >= 0; --i) {
 656         if (rank == i) {
 657           inbuf = static_cast<const unsigned char*>(sbuf);
 658         } else {
 659             Request::recv(pml_buffer, count, dtype, i,
 660                                     COLL_TAG_REDUCE, comm,
 661                                     MPI_STATUS_IGNORE);
 662             inbuf = pml_buffer;
 663         }
 664
 665         /* Perform the reduction */
 666         if(op!=MPI_OP_NULL) op->apply( inbuf, rbuf, &count, dtype);
 667     }
 668
 669     if (nullptr != inplace_temp) {
 670       Datatype::copy(inplace_temp, count, dtype, (char*)sbuf, count, dtype);
 671       smpi_free_tmp_buffer(inplace_temp);
 672     }
 673     if (nullptr != free_buffer) {
 674       smpi_free_tmp_buffer(free_buffer);
 675     }
 676
 677     /* All done */
 678     return MPI_SUCCESS;
 679 }
 680
 681 /* copied function (with appropriate renaming) ends here */
 682
 683
 684 }
 685 }