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

   1 /* Copyright (c) 2013-2020. 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 #include "../coll_tuned_topo.hpp"
  25 #include "../colls_private.hpp"
  26
  27 /*
  28  * Recursive-halving function is (*mostly*) copied from the BASIC coll module.
  29  * I have removed the part which handles "large" message sizes
  30  * (non-overlapping version of reduce_Scatter).
  31  */
  32
  33 /* copied function (with appropriate renaming) starts here */
  34
  35 /*
  36  *  reduce_scatter_ompi_basic_recursivehalving
  37  *
  38  *  Function:   - reduce scatter implementation using recursive-halving
  39  *                algorithm
  40  *  Accepts:    - same as MPI_Reduce_scatter()
  41  *  Returns:    - MPI_SUCCESS or error code
  42  *  Limitation: - Works only for commutative operations.
  43  */
  44 namespace simgrid {
  45 namespace smpi {
  46 int reduce_scatter__ompi_basic_recursivehalving(const void *sbuf,
  47                                                 void *rbuf,
  48                                                 const int *rcounts,
  49                                                 MPI_Datatype dtype,
  50                                                 MPI_Op op,
  51                                                 MPI_Comm comm
  52                                                 )
  53 {
  54     int i, rank, size, count, err = MPI_SUCCESS;
  55     int tmp_size = 1, remain = 0, tmp_rank;
  56     ptrdiff_t true_lb, true_extent, lb, extent, buf_size;
  57     unsigned char *result_buf = nullptr, *result_buf_free = nullptr;
  58
  59     /* Initialize */
  60     rank = comm->rank();
  61     size = comm->size();
  62
  63     XBT_DEBUG("coll:tuned:reduce_scatter_ompi_basic_recursivehalving, rank %d", rank);
  64     if ((op != MPI_OP_NULL && not op->is_commutative()))
  65       throw std::invalid_argument(
  66           "reduce_scatter ompi_basic_recursivehalving can only be used for commutative operations!");
  67
  68     /* Find displacements and the like */
  69     int* disps = new int[size];
  70
  71     disps[0] = 0;
  72     for (i = 0; i < (size - 1); ++i) {
  73         disps[i + 1] = disps[i] + rcounts[i];
  74     }
  75     count = disps[size - 1] + rcounts[size - 1];
  76
  77     /* short cut the trivial case */
  78     if (0 == count) {
  79       delete[] disps;
  80       return MPI_SUCCESS;
  81     }
  82
  83     /* get datatype information */
  84     dtype->extent(&lb, &extent);
  85     dtype->extent(&true_lb, &true_extent);
  86     buf_size = true_extent + (ptrdiff_t)(count - 1) * extent;
  87
  88     /* Handle MPI_IN_PLACE */
  89     if (MPI_IN_PLACE == sbuf) {
  90         sbuf = rbuf;
  91     }
  92
  93     /* Allocate temporary receive buffer. */
  94     unsigned char* recv_buf_free = smpi_get_tmp_recvbuffer(buf_size);
  95     unsigned char* recv_buf      = recv_buf_free - lb;
  96     if (nullptr == recv_buf_free) {
  97       err = MPI_ERR_OTHER;
  98       goto cleanup;
  99     }
 100
 101     /* allocate temporary buffer for results */
 102     result_buf_free = smpi_get_tmp_sendbuffer(buf_size);
 103     result_buf = result_buf_free - lb;
 104
 105     /* copy local buffer into the temporary results */
 106     err =Datatype::copy(sbuf, count, dtype, result_buf, count, dtype);
 107     if (MPI_SUCCESS != err) goto cleanup;
 108
 109     /* figure out power of two mapping: grow until larger than
 110        comm size, then go back one, to get the largest power of
 111        two less than comm size */
 112     while (tmp_size <= size) tmp_size <<= 1;
 113     tmp_size >>= 1;
 114     remain = size - tmp_size;
 115
 116     /* If comm size is not a power of two, have the first "remain"
 117        procs with an even rank send to rank + 1, leaving a power of
 118        two procs to do the rest of the algorithm */
 119     if (rank < 2 * remain) {
 120         if ((rank & 1) == 0) {
 121             Request::send(result_buf, count, dtype, rank + 1,
 122                                     COLL_TAG_REDUCE_SCATTER,
 123                                     comm);
 124             /* we don't participate from here on out */
 125             tmp_rank = -1;
 126         } else {
 127             Request::recv(recv_buf, count, dtype, rank - 1,
 128                                     COLL_TAG_REDUCE_SCATTER,
 129                                     comm, MPI_STATUS_IGNORE);
 130
 131             /* integrate their results into our temp results */
 132             if(op!=MPI_OP_NULL) op->apply( recv_buf, result_buf, &count, dtype);
 133
 134             /* adjust rank to be the bottom "remain" ranks */
 135             tmp_rank = rank / 2;
 136         }
 137     } else {
 138         /* just need to adjust rank to show that the bottom "even
 139            remain" ranks dropped out */
 140         tmp_rank = rank - remain;
 141     }
 142
 143     /* For ranks not kicked out by the above code, perform the
 144        recursive halving */
 145     if (tmp_rank >= 0) {
 146         int mask, send_index, recv_index, last_index;
 147
 148         /* recalculate disps and rcounts to account for the
 149            special "remainder" processes that are no longer doing
 150            anything */
 151         int* tmp_rcounts = new int[tmp_size];
 152         int* tmp_disps   = new int[tmp_size];
 153
 154         for (i = 0 ; i < tmp_size ; ++i) {
 155             if (i < remain) {
 156                 /* need to include old neighbor as well */
 157                 tmp_rcounts[i] = rcounts[i * 2 + 1] + rcounts[i * 2];
 158             } else {
 159                 tmp_rcounts[i] = rcounts[i + remain];
 160             }
 161         }
 162
 163         tmp_disps[0] = 0;
 164         for (i = 0; i < tmp_size - 1; ++i) {
 165             tmp_disps[i + 1] = tmp_disps[i] + tmp_rcounts[i];
 166         }
 167
 168         /* do the recursive halving communication.  Don't use the
 169            dimension information on the communicator because I
 170            think the information is invalidated by our "shrinking"
 171            of the communicator */
 172         mask = tmp_size >> 1;
 173         send_index = recv_index = 0;
 174         last_index = tmp_size;
 175         while (mask > 0) {
 176             int tmp_peer, peer, send_count, recv_count;
 177             MPI_Request request;
 178
 179             tmp_peer = tmp_rank ^ mask;
 180             peer = (tmp_peer < remain) ? tmp_peer * 2 + 1 : tmp_peer + remain;
 181
 182             /* figure out if we're sending, receiving, or both */
 183             send_count = recv_count = 0;
 184             if (tmp_rank < tmp_peer) {
 185                 send_index = recv_index + mask;
 186                 for (i = send_index ; i < last_index ; ++i) {
 187                     send_count += tmp_rcounts[i];
 188                 }
 189                 for (i = recv_index ; i < send_index ; ++i) {
 190                     recv_count += tmp_rcounts[i];
 191                 }
 192             } else {
 193                 recv_index = send_index + mask;
 194                 for (i = send_index ; i < recv_index ; ++i) {
 195                     send_count += tmp_rcounts[i];
 196                 }
 197                 for (i = recv_index ; i < last_index ; ++i) {
 198                     recv_count += tmp_rcounts[i];
 199                 }
 200             }
 201
 202             /* actual data transfer.  Send from result_buf,
 203                receive into recv_buf */
 204             if (send_count > 0 && recv_count != 0) {
 205                 request=Request::irecv(recv_buf + (ptrdiff_t)tmp_disps[recv_index] * extent,
 206                                          recv_count, dtype, peer,
 207                                          COLL_TAG_REDUCE_SCATTER,
 208                                          comm);
 209                 if (MPI_SUCCESS != err) {
 210                   delete[] tmp_rcounts;
 211                   delete[] tmp_disps;
 212                   goto cleanup;
 213                 }
 214             }
 215             if (recv_count > 0 && send_count != 0) {
 216                 Request::send(result_buf + (ptrdiff_t)tmp_disps[send_index] * extent,
 217                                         send_count, dtype, peer,
 218                                         COLL_TAG_REDUCE_SCATTER,
 219                                         comm);
 220                 if (MPI_SUCCESS != err) {
 221                   delete[] tmp_rcounts;
 222                   delete[] tmp_disps;
 223                   goto cleanup;
 224                 }
 225             }
 226             if (send_count > 0 && recv_count != 0) {
 227                 Request::wait(&request, MPI_STATUS_IGNORE);
 228             }
 229
 230             /* if we received something on this step, push it into
 231                the results buffer */
 232             if (recv_count > 0) {
 233                 if(op!=MPI_OP_NULL) op->apply(
 234                                recv_buf + (ptrdiff_t)tmp_disps[recv_index] * extent,
 235                                result_buf + (ptrdiff_t)tmp_disps[recv_index] * extent,
 236                                &recv_count, dtype);
 237             }
 238
 239             /* update for next iteration */
 240             send_index = recv_index;
 241             last_index = recv_index + mask;
 242             mask >>= 1;
 243         }
 244
 245         /* copy local results from results buffer into real receive buffer */
 246         if (0 != rcounts[rank]) {
 247             err = Datatype::copy(result_buf + disps[rank] * extent,
 248                                        rcounts[rank], dtype,
 249                                        rbuf, rcounts[rank], dtype);
 250             if (MPI_SUCCESS != err) {
 251               delete[] tmp_rcounts;
 252               delete[] tmp_disps;
 253               goto cleanup;
 254             }
 255         }
 256
 257         delete[] tmp_rcounts;
 258         delete[] tmp_disps;
 259     }
 260
 261     /* Now fix up the non-power of two case, by having the odd
 262        procs send the even procs the proper results */
 263     if (rank < (2 * remain)) {
 264         if ((rank & 1) == 0) {
 265             if (rcounts[rank]) {
 266                 Request::recv(rbuf, rcounts[rank], dtype, rank + 1,
 267                                         COLL_TAG_REDUCE_SCATTER,
 268                                         comm, MPI_STATUS_IGNORE);
 269             }
 270         } else {
 271             if (rcounts[rank - 1]) {
 272                 Request::send(result_buf + disps[rank - 1] * extent,
 273                                         rcounts[rank - 1], dtype, rank - 1,
 274                                         COLL_TAG_REDUCE_SCATTER,
 275                                         comm);
 276             }
 277         }
 278     }
 279
 280  cleanup:
 281     delete[] disps;
 282     if (nullptr != recv_buf_free)
 283       smpi_free_tmp_buffer(recv_buf_free);
 284     if (nullptr != result_buf_free)
 285       smpi_free_tmp_buffer(result_buf_free);
 286
 287     return err;
 288 }
 289
 290 /* copied function (with appropriate renaming) ends here */
 291
 292
 293 /*
 294  *   Coll_reduce_scatter_ompi_ring::reduce_scatter
 295  *
 296  *   Function:       Ring algorithm for reduce_scatter operation
 297  *   Accepts:        Same as MPI_Reduce_scatter()
 298  *   Returns:        MPI_SUCCESS or error code
 299  *
 300  *   Description:    Implements ring algorithm for reduce_scatter:
 301  *                   the block sizes defined in rcounts are exchanged and
 302  8                    updated until they reach proper destination.
 303  *                   Algorithm requires 2 * max(rcounts) extra buffering
 304  *
 305  *   Limitations:    The algorithm DOES NOT preserve order of operations so it
 306  *                   can be used only for commutative operations.
 307  *         Example on 5 nodes:
 308  *         Initial state
 309  *   #      0              1             2              3             4
 310  *        [00]           [10]   ->     [20]           [30]           [40]
 311  *        [01]           [11]          [21]  ->       [31]           [41]
 312  *        [02]           [12]          [22]           [32]  ->       [42]
 313  *    ->  [03]           [13]          [23]           [33]           [43] --> ..
 314  *        [04]  ->       [14]          [24]           [34]           [44]
 315  *
 316  *        COMPUTATION PHASE
 317  *         Step 0: rank r sends block (r-1) to rank (r+1) and
 318  *                 receives block (r+1) from rank (r-1) [with wraparound].
 319  *   #      0              1             2              3             4
 320  *        [00]           [10]        [10+20]   ->     [30]           [40]
 321  *        [01]           [11]          [21]          [21+31]  ->     [41]
 322  *    ->  [02]           [12]          [22]           [32]         [32+42] -->..
 323  *      [43+03] ->       [13]          [23]           [33]           [43]
 324  *        [04]         [04+14]  ->     [24]           [34]           [44]
 325  *
 326  *         Step 1:
 327  *   #      0              1             2              3             4
 328  *        [00]           [10]        [10+20]       [10+20+30] ->     [40]
 329  *    ->  [01]           [11]          [21]          [21+31]      [21+31+41] ->
 330  *     [32+42+02] ->     [12]          [22]           [32]         [32+42]
 331  *        [03]        [43+03+13] ->    [23]           [33]           [43]
 332  *        [04]         [04+14]      [04+14+24]  ->    [34]           [44]
 333  *
 334  *         Step 2:
 335  *   #      0              1             2              3             4
 336  *     -> [00]           [10]        [10+20]       [10+20+30]   [10+20+30+40] ->
 337  *   [21+31+41+01]->     [11]          [21]          [21+31]      [21+31+41]
 338  *     [32+42+02]   [32+42+02+12]->    [22]           [32]         [32+42]
 339  *        [03]        [43+03+13]   [43+03+13+23]->    [33]           [43]
 340  *        [04]         [04+14]      [04+14+24]    [04+14+24+34] ->   [44]
 341  *
 342  *         Step 3:
 343  *   #      0             1              2              3             4
 344  * [10+20+30+40+00]     [10]         [10+20]       [10+20+30]   [10+20+30+40]
 345  *  [21+31+41+01] [21+31+41+01+11]     [21]          [21+31]      [21+31+41]
 346  *    [32+42+02]   [32+42+02+12] [32+42+02+12+22]     [32]         [32+42]
 347  *       [03]        [43+03+13]    [43+03+13+23] [43+03+13+23+33]    [43]
 348  *       [04]         [04+14]       [04+14+24]    [04+14+24+34] [04+14+24+34+44]
 349  *    DONE :)
 350  *
 351  */
 352 int reduce_scatter__ompi_ring(const void *sbuf, void *rbuf, const int *rcounts,
 353                               MPI_Datatype dtype,
 354                               MPI_Op op,
 355                               MPI_Comm comm
 356                               )
 357 {
 358     int ret, line, rank, size, i, k, recv_from, send_to, total_count, max_block_count;
 359     int inbi;
 360     unsigned char *tmpsend = nullptr, *tmprecv = nullptr, *accumbuf = nullptr, *accumbuf_free = nullptr;
 361     unsigned char *inbuf_free[2] = {nullptr, nullptr}, *inbuf[2] = {nullptr, nullptr};
 362     ptrdiff_t true_lb, true_extent, lb, extent, max_real_segsize;
 363     MPI_Request reqs[2] = {nullptr, nullptr};
 364
 365     size = comm->size();
 366     rank = comm->rank();
 367
 368     XBT_DEBUG(  "coll:tuned:reduce_scatter_ompi_ring rank %d, size %d",
 369                  rank, size);
 370
 371     /* Determine the maximum number of elements per node,
 372        corresponding block size, and displacements array.
 373     */
 374     int* displs = new int[size];
 375
 376     displs[0] = 0;
 377     total_count = rcounts[0];
 378     max_block_count = rcounts[0];
 379     for (i = 1; i < size; i++) {
 380         displs[i] = total_count;
 381         total_count += rcounts[i];
 382         if (max_block_count < rcounts[i]) max_block_count = rcounts[i];
 383     }
 384
 385     /* Special case for size == 1 */
 386     if (1 == size) {
 387         if (MPI_IN_PLACE != sbuf) {
 388             ret = Datatype::copy((char*)sbuf, total_count, dtype, (char*)rbuf, total_count, dtype);
 389             if (ret < 0) { line = __LINE__; goto error_hndl; }
 390         }
 391         delete[] displs;
 392         return MPI_SUCCESS;
 393     }
 394
 395     /* Allocate and initialize temporary buffers, we need:
 396        - a temporary buffer to perform reduction (size total_count) since
 397        rbuf can be of rcounts[rank] size.
 398        - up to two temporary buffers used for communication/computation overlap.
 399     */
 400     dtype->extent(&lb, &extent);
 401     dtype->extent(&true_lb, &true_extent);
 402
 403     max_real_segsize = true_extent + (ptrdiff_t)(max_block_count - 1) * extent;
 404
 405     accumbuf_free = smpi_get_tmp_recvbuffer(true_extent + (ptrdiff_t)(total_count - 1) * extent);
 406     if (nullptr == accumbuf_free) {
 407       ret  = -1;
 408       line = __LINE__;
 409       goto error_hndl;
 410     }
 411     accumbuf = accumbuf_free - lb;
 412
 413     inbuf_free[0] = smpi_get_tmp_sendbuffer(max_real_segsize);
 414     if (nullptr == inbuf_free[0]) {
 415       ret  = -1;
 416       line = __LINE__;
 417       goto error_hndl;
 418     }
 419     inbuf[0] = inbuf_free[0] - lb;
 420     if (size > 2) {
 421       inbuf_free[1] = smpi_get_tmp_sendbuffer(max_real_segsize);
 422       if (nullptr == inbuf_free[1]) {
 423         ret  = -1;
 424         line = __LINE__;
 425         goto error_hndl;
 426       }
 427       inbuf[1] = inbuf_free[1] - lb;
 428     }
 429
 430     /* Handle MPI_IN_PLACE for size > 1 */
 431     if (MPI_IN_PLACE == sbuf) {
 432         sbuf = rbuf;
 433     }
 434
 435     ret = Datatype::copy((char*)sbuf, total_count, dtype, accumbuf, total_count, dtype);
 436     if (ret < 0) { line = __LINE__; goto error_hndl; }
 437
 438     /* Computation loop */
 439
 440     /*
 441        For each of the remote nodes:
 442        - post irecv for block (r-2) from (r-1) with wrap around
 443        - send block (r-1) to (r+1)
 444        - in loop for every step k = 2 .. n
 445        - post irecv for block (r - 1 + n - k) % n
 446        - wait on block (r + n - k) % n to arrive
 447        - compute on block (r + n - k ) % n
 448        - send block (r + n - k) % n
 449        - wait on block (r)
 450        - compute on block (r)
 451        - copy block (r) to rbuf
 452        Note that we must be careful when computing the beginning of buffers and
 453        for send operations and computation we must compute the exact block size.
 454     */
 455     send_to = (rank + 1) % size;
 456     recv_from = (rank + size - 1) % size;
 457
 458     inbi = 0;
 459     /* Initialize first receive from the neighbor on the left */
 460     reqs[inbi]=Request::irecv(inbuf[inbi], max_block_count, dtype, recv_from,
 461                              COLL_TAG_REDUCE_SCATTER, comm
 462                              );
 463     tmpsend = accumbuf + (ptrdiff_t)displs[recv_from] * extent;
 464     Request::send(tmpsend, rcounts[recv_from], dtype, send_to,
 465                             COLL_TAG_REDUCE_SCATTER,
 466                              comm);
 467
 468     for (k = 2; k < size; k++) {
 469         const int prevblock = (rank + size - k) % size;
 470
 471         inbi = inbi ^ 0x1;
 472
 473         /* Post irecv for the current block */
 474         reqs[inbi]=Request::irecv(inbuf[inbi], max_block_count, dtype, recv_from,
 475                                  COLL_TAG_REDUCE_SCATTER, comm
 476                                  );
 477
 478         /* Wait on previous block to arrive */
 479         Request::wait(&reqs[inbi ^ 0x1], MPI_STATUS_IGNORE);
 480
 481         /* Apply operation on previous block: result goes to rbuf
 482            rbuf[prevblock] = inbuf[inbi ^ 0x1] (op) rbuf[prevblock]
 483         */
 484         tmprecv = accumbuf + (ptrdiff_t)displs[prevblock] * extent;
 485         if (op != MPI_OP_NULL)
 486           op->apply(inbuf[inbi ^ 0x1], tmprecv, &rcounts[prevblock], dtype);
 487
 488         /* send previous block to send_to */
 489         Request::send(tmprecv, rcounts[prevblock], dtype, send_to,
 490                                 COLL_TAG_REDUCE_SCATTER,
 491                                  comm);
 492     }
 493
 494     /* Wait on the last block to arrive */
 495     Request::wait(&reqs[inbi], MPI_STATUS_IGNORE);
 496
 497     /* Apply operation on the last block (my block)
 498        rbuf[rank] = inbuf[inbi] (op) rbuf[rank] */
 499     tmprecv = accumbuf + (ptrdiff_t)displs[rank] * extent;
 500     if (op != MPI_OP_NULL)
 501       op->apply(inbuf[inbi], tmprecv, &rcounts[rank], dtype);
 502
 503     /* Copy result from tmprecv to rbuf */
 504     ret = Datatype::copy(tmprecv, rcounts[rank], dtype, (char*)rbuf, rcounts[rank], dtype);
 505     if (ret < 0) { line = __LINE__; goto error_hndl; }
 506
 507     delete[] displs;
 508     if (nullptr != accumbuf_free)
 509       smpi_free_tmp_buffer(accumbuf_free);
 510     if (nullptr != inbuf_free[0])
 511       smpi_free_tmp_buffer(inbuf_free[0]);
 512     if (nullptr != inbuf_free[1])
 513       smpi_free_tmp_buffer(inbuf_free[1]);
 514
 515     return MPI_SUCCESS;
 516
 517  error_hndl:
 518     XBT_DEBUG( "%s:%4d\tRank %d Error occurred %d\n",
 519                  __FILE__, line, rank, ret);
 520     delete[] displs;
 521     if (nullptr != accumbuf_free)
 522       smpi_free_tmp_buffer(accumbuf_free);
 523     if (nullptr != inbuf_free[0])
 524       smpi_free_tmp_buffer(inbuf_free[0]);
 525     if (nullptr != inbuf_free[1])
 526       smpi_free_tmp_buffer(inbuf_free[1]);
 527     return ret;
 528 }
 529 }
 530 }
 531