root/ompi/mca/coll/libnbc/nbc_iallreduce.c

DEFINITIONS

1. NBC_Allreduce_args_compare
2. nbc_allreduce_init
3. ompi_coll_libnbc_iallreduce
4. nbc_allreduce_inter_init
5. ompi_coll_libnbc_iallreduce_inter
6. allred_sched_diss
7. allred_sched_ring
8. allred_sched_linear
9. allred_sched_redscat_allgather
10. ompi_coll_libnbc_allreduce_init
11. ompi_coll_libnbc_allreduce_inter_init

   1 /* -*- Mode: C; c-basic-offset:2 ; indent-tabs-mode:nil -*- */
   2 /*
   3  * Copyright (c) 2006      The Trustees of Indiana University and Indiana
   4  *                         University Research and Technology
   5  *                         Corporation.  All rights reserved.
   6  * Copyright (c) 2006      The Technical University of Chemnitz. All
   7  *                         rights reserved.
   8  * Copyright (c) 2013-2017 Los Alamos National Security, LLC. All rights
   9  *                         reserved.
  10  * Copyright (c) 2014-2018 Research Organization for Information Science
  11  *                         and Technology (RIST).  All rights reserved.
  12  * Copyright (c) 2017      IBM Corporation.  All rights reserved.
  13  * Copyright (c) 2018      FUJITSU LIMITED.  All rights reserved.
  14  * $COPYRIGHT$
  15  *
  16  * Additional copyrights may follow
  17  *
  18  * Author(s): Torsten Hoefler <htor@cs.indiana.edu>
  19  *
  20  */
  21 #include "nbc_internal.h"
  22 #include "ompi/communicator/communicator.h"
  23 #include "ompi/datatype/ompi_datatype.h"
  24 #include "ompi/op/op.h"
  25 #include "opal/util/bit_ops.h"
  26 
  27 #include <assert.h>
  28 
  29 static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, ptrdiff_t gap, const void *sendbuf,
  30                                     void *recvbuf, MPI_Op op, char inplace, NBC_Schedule *schedule, void *tmpbuf);
  31 static inline int allred_sched_ring(int rank, int p, int count, MPI_Datatype datatype, const void *sendbuf,
  32                                     void *recvbuf, MPI_Op op, int size, int ext, NBC_Schedule *schedule,
  33                                     void *tmpbuf);
  34 static inline int allred_sched_linear(int rank, int p, const void *sendbuf, void *recvbuf, int count,
  35                                       MPI_Datatype datatype, ptrdiff_t gap, MPI_Op op, int ext, int size,
  36                                       NBC_Schedule *schedule, void *tmpbuf);
  37 static inline int allred_sched_redscat_allgather(
  38     int rank, int comm_size, int count, MPI_Datatype datatype, ptrdiff_t gap,
  39     const void *sbuf, void *rbuf, MPI_Op op, char inplace,
  40     NBC_Schedule *schedule, void *tmpbuf, struct ompi_communicator_t *comm);
  41 
  42 #ifdef NBC_CACHE_SCHEDULE
  43 /* tree comparison function for schedule cache */
  44 int NBC_Allreduce_args_compare(NBC_Allreduce_args *a, NBC_Allreduce_args *b, void *param) {
  45   if ((a->sendbuf == b->sendbuf) &&
  46       (a->recvbuf == b->recvbuf) &&
  47       (a->count == b->count) &&
  48       (a->datatype == b->datatype) &&
  49       (a->op == b->op)) {
  50     return 0;
  51   }
  52 
  53   if( a->sendbuf < b->sendbuf ) {
  54     return -1;
  55   }
  56 
  57   return 1;
  58 }
  59 #endif
  60 
  61 static int nbc_allreduce_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
  62                               struct ompi_communicator_t *comm, ompi_request_t ** request,
  63                               struct mca_coll_base_module_2_3_0_t *module, bool persistent)
  64 {
  65   int rank, p, res;
  66   ptrdiff_t ext, lb;
  67   NBC_Schedule *schedule;
  68   size_t size;
  69 #ifdef NBC_CACHE_SCHEDULE
  70   NBC_Allreduce_args *args, *found, search;
  71 #endif
  72   enum { NBC_ARED_BINOMIAL, NBC_ARED_RING, NBC_ARED_REDSCAT_ALLGATHER } alg;
  73   char inplace;
  74   void *tmpbuf = NULL;
  75   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
  76   ptrdiff_t span, gap;
  77 
  78   NBC_IN_PLACE(sendbuf, recvbuf, inplace);
  79 
  80   rank = ompi_comm_rank (comm);
  81   p = ompi_comm_size (comm);
  82 
  83   res = ompi_datatype_get_extent(datatype, &lb, &ext);
  84   if (OMPI_SUCCESS != res) {
  85     NBC_Error ("MPI Error in ompi_datatype_type_extent() (%i)", res);
  86     return res;
  87   }
  88 
  89   res = ompi_datatype_type_size (datatype, &size);
  90   if (OMPI_SUCCESS != res) {
  91     NBC_Error ("MPI Error in ompi_datatype_type_size() (%i)", res);
  92     return res;
  93   }
  94 
  95   if (1 == p && (!persistent || inplace)) {
  96     if (!inplace) {
  97       /* for a single node - copy data to receivebuf */
  98       res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm);
  99       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 100         return res;
 101       }
 102     }
 103     return nbc_get_noop_request(persistent, request);
 104   }
 105 
 106   span = opal_datatype_span(&datatype->super, count, &gap);
 107   tmpbuf = malloc (span);
 108   if (OPAL_UNLIKELY(NULL == tmpbuf)) {
 109     return OMPI_ERR_OUT_OF_RESOURCE;
 110   }
 111 
 112   /* algorithm selection */
 113   int nprocs_pof2 = opal_next_poweroftwo(p) >> 1;
 114   if (libnbc_iallreduce_algorithm == 0) {
 115     if(p < 4 || size*count < 65536 || !ompi_op_is_commute(op) || inplace) {
 116       alg = NBC_ARED_BINOMIAL;
 117     } else if (count >= nprocs_pof2 && ompi_op_is_commute(op)) {
 118       alg = NBC_ARED_REDSCAT_ALLGATHER;
 119     } else {
 120       alg = NBC_ARED_RING;
 121     }
 122   } else {
 123     if (libnbc_iallreduce_algorithm == 1)
 124       alg = NBC_ARED_RING;
 125     else if (libnbc_iallreduce_algorithm == 2)
 126       alg = NBC_ARED_BINOMIAL;
 127     else if (libnbc_iallreduce_algorithm == 3 && count >= nprocs_pof2 && ompi_op_is_commute(op)) {
 128       alg = NBC_ARED_REDSCAT_ALLGATHER;
 129     } else
 130       alg = NBC_ARED_RING;
 131   }
 132 #ifdef NBC_CACHE_SCHEDULE
 133   /* search schedule in communicator specific tree */
 134   search.sendbuf = sendbuf;
 135   search.recvbuf = recvbuf;
 136   search.count = count;
 137   search.datatype = datatype;
 138   search.op = op;
 139   found = (NBC_Allreduce_args *) hb_tree_search ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLREDUCE], &search);
 140   if (NULL == found) {
 141 #endif
 142     schedule = OBJ_NEW(NBC_Schedule);
 143     if (NULL == schedule) {
 144       free(tmpbuf);
 145       return OMPI_ERR_OUT_OF_RESOURCE;
 146     }
 147 
 148     if (p == 1) {
 149       res = NBC_Sched_copy((void *)sendbuf, false, count, datatype,
 150                            recvbuf, false, count, datatype, schedule, false);
 151     } else {
 152       switch(alg) {
 153         case NBC_ARED_BINOMIAL:
 154           res = allred_sched_diss(rank, p, count, datatype, gap, sendbuf, recvbuf, op, inplace, schedule, tmpbuf);
 155           break;
 156         case NBC_ARED_REDSCAT_ALLGATHER:
 157           res = allred_sched_redscat_allgather(rank, p, count, datatype, gap, sendbuf, recvbuf, op, inplace, schedule, tmpbuf, comm);
 158           break;
 159         case NBC_ARED_RING:
 160           res = allred_sched_ring(rank, p, count, datatype, sendbuf, recvbuf, op, size, ext, schedule, tmpbuf);
 161           break;
 162       }
 163     }
 164 
 165     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 166       OBJ_RELEASE(schedule);
 167       free(tmpbuf);
 168       return res;
 169     }
 170 
 171     res = NBC_Sched_commit(schedule);
 172     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 173       OBJ_RELEASE(schedule);
 174       free(tmpbuf);
 175       return res;
 176     }
 177 
 178 #ifdef NBC_CACHE_SCHEDULE
 179     /* save schedule to tree */
 180     args = (NBC_Allreduce_args *) malloc (sizeof(args));
 181     if (NULL != args) {
 182       args->sendbuf = sendbuf;
 183       args->recvbuf = recvbuf;
 184       args->count = count;
 185       args->datatype = datatype;
 186       args->op = op;
 187       args->schedule = schedule;
 188       res = hb_tree_insert ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLREDUCE], args, args, 0);
 189       if (0 == res) {
 190         OBJ_RETAIN(schedule);
 191 
 192         /* increase number of elements for A2A */
 193         if (++libnbc_module->NBC_Dict_size[NBC_ALLREDUCE] > NBC_SCHED_DICT_UPPER) {
 194           NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLREDUCE],
 195                                    &libnbc_module->NBC_Dict_size[NBC_ALLREDUCE]);
 196         }
 197       } else {
 198         NBC_Error("error in dict_insert() (%i)", res);
 199         free (args);
 200       }
 201     }
 202   } else {
 203     /* found schedule */
 204     schedule = found->schedule;
 205     OBJ_RETAIN(schedule);
 206   }
 207 #endif
 208 
 209   res = NBC_Schedule_request (schedule, comm, libnbc_module, persistent, request, tmpbuf);
 210   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 211     OBJ_RELEASE(schedule);
 212     free(tmpbuf);
 213     return res;
 214   }
 215 
 216   return OMPI_SUCCESS;
 217 }
 218 
 219 int ompi_coll_libnbc_iallreduce(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
 220                                 struct ompi_communicator_t *comm, ompi_request_t ** request,
 221                                 struct mca_coll_base_module_2_3_0_t *module) {
 222     int res = nbc_allreduce_init(sendbuf, recvbuf, count, datatype, op,
 223                                  comm, request, module, false);
 224     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 225         return res;
 226     }
 227   
 228     res = NBC_Start(*(ompi_coll_libnbc_request_t **)request);
 229     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 230         NBC_Return_handle (*(ompi_coll_libnbc_request_t **)request);
 231         *request = &ompi_request_null.request;
 232         return res;
 233     }
 234 
 235     return OMPI_SUCCESS;
 236 }
 237 
 238 static int nbc_allreduce_inter_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
 239                                     struct ompi_communicator_t *comm, ompi_request_t ** request,
 240                                     struct mca_coll_base_module_2_3_0_t *module, bool persistent)
 241 {
 242   int rank, res, rsize;
 243   size_t size;
 244   MPI_Aint ext;
 245   NBC_Schedule *schedule;
 246   void *tmpbuf = NULL;
 247   ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
 248   ptrdiff_t span, gap;
 249 
 250   rank = ompi_comm_rank (comm);
 251   rsize = ompi_comm_remote_size (comm);
 252 
 253   res = ompi_datatype_type_extent(datatype, &ext);
 254   if (MPI_SUCCESS != res) {
 255     NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
 256     return res;
 257   }
 258 
 259   res = ompi_datatype_type_size(datatype, &size);
 260   if (MPI_SUCCESS != res) {
 261     NBC_Error("MPI Error in ompi_datatype_type_size() (%i)", res);
 262     return res;
 263   }
 264 
 265   span = opal_datatype_span(&datatype->super, count, &gap);
 266   tmpbuf = malloc (span);
 267   if (OPAL_UNLIKELY(NULL == tmpbuf)) {
 268     return OMPI_ERR_OUT_OF_RESOURCE;
 269   }
 270 
 271   schedule = OBJ_NEW(NBC_Schedule);
 272   if (OPAL_UNLIKELY(NULL == schedule)) {
 273     free(tmpbuf);
 274     return OMPI_ERR_OUT_OF_RESOURCE;
 275   }
 276 
 277   res = allred_sched_linear (rank, rsize, sendbuf, recvbuf, count, datatype, gap, op,
 278                              ext, size, schedule, tmpbuf);
 279   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 280     OBJ_RELEASE(schedule);
 281     free(tmpbuf);
 282     return res;
 283   }
 284 
 285   res = NBC_Sched_commit(schedule);
 286   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 287     OBJ_RELEASE(schedule);
 288     free(tmpbuf);
 289     return res;
 290   }
 291 
 292   res = NBC_Schedule_request(schedule, comm, libnbc_module, persistent, request, tmpbuf);
 293   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 294     OBJ_RELEASE(schedule);
 295     free(tmpbuf);
 296     return res;
 297   }
 298 
 299   return OMPI_SUCCESS;
 300 }
 301 
 302 int ompi_coll_libnbc_iallreduce_inter(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
 303                                       struct ompi_communicator_t *comm, ompi_request_t ** request,
 304                                       struct mca_coll_base_module_2_3_0_t *module) {
 305     int res = nbc_allreduce_inter_init(sendbuf, recvbuf, count, datatype, op,
 306                                        comm, request, module, false);
 307     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 308         return res;
 309     }
 310   
 311     res = NBC_Start(*(ompi_coll_libnbc_request_t **)request);
 312     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 313         NBC_Return_handle (*(ompi_coll_libnbc_request_t **)request);
 314         *request = &ompi_request_null.request;
 315         return res;
 316     }
 317 
 318     return OMPI_SUCCESS;
 319 }
 320 
 321 /* binomial allreduce (binomial tree up and binomial bcast down)
 322  * working principle:
 323  * - each node gets a virtual rank vrank
 324  * - the 'root' node get vrank 0
 325  * - node 0 gets the vrank of the 'root'
 326  * - all other ranks stay identical (they do not matter)
 327  *
 328  * Algorithm:
 329  * pairwise exchange
 330  * round r:
 331  *  grp = rank % 2^r
 332  *  if grp == 0: receive from rank + 2^(r-1) if it exists and reduce value
 333  *  if grp == 1: send to rank - 2^(r-1) and exit function
 334  *
 335  * do this for R=log_2(p) rounds
 336  * followed by a Bcast:
 337  * Algorithm:
 338  * - each node with vrank > 2^r and vrank < 2^r+1 receives from node
 339  *   vrank - 2^r (vrank=1 receives from 0, vrank 0 receives never)
 340  * - each node sends each round r to node vrank + 2^r
 341  * - a node stops to send if 2^r > commsize
 342  *
 343  */
 344 #define RANK2VRANK(rank, vrank, root) \
 345 { \
 346   vrank = rank; \
 347   if (rank == 0) vrank = root; \
 348   if (rank == root) vrank = 0; \
 349 }
 350 #define VRANK2RANK(rank, vrank, root) \
 351 { \
 352   rank = vrank; \
 353   if (vrank == 0) rank = root; \
 354   if (vrank == root) rank = 0; \
 355 }
 356 static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, ptrdiff_t gap, const void *sendbuf, void *recvbuf,
 357                                     MPI_Op op, char inplace, NBC_Schedule *schedule, void *tmpbuf) {
 358   int root, vrank, maxr, vpeer, peer, res;
 359   char *rbuf, *lbuf, *buf;
 360   int tmprbuf, tmplbuf;
 361 
 362   root = 0; /* this makes the code for ireduce and iallreduce nearly identical - could be changed to improve performance */
 363   RANK2VRANK(rank, vrank, root);
 364   maxr = (int)ceil((log((double)p)/LOG2));
 365   /* ensure the result ends up in recvbuf on vrank 0 */
 366   if (0 == (maxr%2)) {
 367     rbuf = (void *)(-gap);
 368     tmprbuf = true;
 369     lbuf = recvbuf;
 370     tmplbuf = false;
 371   } else {
 372     lbuf = (void *)(-gap);
 373     tmplbuf = true;
 374     rbuf = recvbuf;
 375     tmprbuf = false;
 376     if (inplace) {
 377         res = NBC_Sched_copy(rbuf, false, count, datatype,
 378                              ((char *)tmpbuf) - gap, false, count, datatype,
 379                              schedule, true);
 380         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 381           return res;
 382         }
 383     }
 384   }
 385 
 386   for (int r = 1, firstred = 1 ; r <= maxr ; ++r) {
 387     if ((vrank % (1 << r)) == 0) {
 388       /* we have to receive this round */
 389       vpeer = vrank + (1 << (r - 1));
 390       VRANK2RANK(peer, vpeer, root)
 391       if (peer < p) {
 392         /* we have to wait until we have the data */
 393         res = NBC_Sched_recv (rbuf, tmprbuf, count, datatype, peer, schedule, true);
 394         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 395           return res;
 396         }
 397 
 398         /* this cannot be done until tmpbuf is unused :-( so barrier after the op */
 399         if (firstred && !inplace) {
 400           /* perform the reduce with the senbuf */
 401           res = NBC_Sched_op (sendbuf, false, rbuf, tmprbuf, count, datatype, op, schedule, true);
 402           firstred = 0;
 403         } else {
 404           /* perform the reduce in my local buffer */
 405           res = NBC_Sched_op (lbuf, tmplbuf, rbuf, tmprbuf, count, datatype, op, schedule, true);
 406         }
 407         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 408           return res;
 409         }
 410         /* swap left and right buffers */
 411         buf = rbuf; rbuf = lbuf ; lbuf = buf;
 412         tmprbuf ^= 1; tmplbuf ^= 1;
 413       }
 414     } else {
 415       /* we have to send this round */
 416       vpeer = vrank - (1 << (r - 1));
 417       VRANK2RANK(peer, vpeer, root)
 418       if (firstred && !inplace) {
 419         /* we have to use the sendbuf in the first round .. */
 420         res = NBC_Sched_send (sendbuf, false, count, datatype, peer, schedule, false);
 421       } else {
 422         /* and the recvbuf in all remaining rounds */
 423         res = NBC_Sched_send (lbuf, tmplbuf, count, datatype, peer, schedule, false);
 424       }
 425 
 426       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 427         return res;
 428       }
 429 
 430       /* leave the game */
 431       break;
 432     }
 433   }
 434 
 435   /* this is the Bcast part - copied with minor changes from nbc_ibcast.c
 436    * changed: buffer -> recvbuf  */
 437   RANK2VRANK(rank, vrank, root);
 438 
 439   /* receive from the right hosts  */
 440   if (vrank != 0) {
 441     for (int r = 0; r < maxr ; ++r) {
 442       if ((vrank >= (1 << r)) && (vrank < (1 << (r + 1)))) {
 443         VRANK2RANK(peer, vrank - (1 << r), root);
 444         res = NBC_Sched_recv (recvbuf, false, count, datatype, peer, schedule, false);
 445         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 446           return res;
 447         }
 448       }
 449     }
 450 
 451     res = NBC_Sched_barrier (schedule);
 452     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 453       return res;
 454     }
 455   }
 456 
 457   if (0 == vrank) assert(lbuf == recvbuf);
 458   /* now send to the right hosts */
 459   for (int r = 0; r < maxr; ++r) {
 460     if (((vrank + (1 << r) < p) && (vrank < (1 << r))) || (vrank == 0)) {
 461       VRANK2RANK(peer, vrank + (1 << r), root);
 462       res = NBC_Sched_send (recvbuf, false, count, datatype, peer, schedule, false);
 463       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 464         return res;
 465       }
 466     }
 467   }
 468 
 469   /* end of the bcast */
 470   return OMPI_SUCCESS;
 471 }
 472 
 473 static inline int allred_sched_ring (int r, int p, int count, MPI_Datatype datatype, const void *sendbuf, void *recvbuf, MPI_Op op,
 474                                      int size, int ext, NBC_Schedule *schedule, void *tmpbuf) {
 475   int segsize, *segsizes, *segoffsets; /* segment sizes and offsets per segment (number of segments == number of nodes */
 476   int speer, rpeer; /* send and recvpeer */
 477   int res = OMPI_SUCCESS;
 478 
 479   if (count == 0) {
 480     return OMPI_SUCCESS;
 481   }
 482 
 483   segsizes = (int *) malloc (sizeof (int) * p);
 484   segoffsets = (int *) malloc (sizeof (int) * p);
 485   if (NULL == segsizes || NULL == segoffsets) {
 486     free (segsizes);
 487     free (segoffsets);
 488     return OMPI_ERR_OUT_OF_RESOURCE;
 489   }
 490 
 491   segsize = (count + p - 1) / p; /* size of the segments */
 492 
 493   segoffsets[0] = 0;
 494   for (int i = 0, mycount = count ; i < p ; ++i) {
 495     mycount -= segsize;
 496     segsizes[i] = segsize;
 497     if (mycount < 0) {
 498       segsizes[i] = segsize + mycount;
 499       mycount = 0;
 500     }
 501 
 502     if (i) {
 503       segoffsets[i] = segoffsets[i-1] + segsizes[i-1];
 504     }
 505   }
 506 
 507   /* reduce peers */
 508   speer = (r + 1) % p;
 509   rpeer = (r - 1 + p) % p;
 510 
 511   /*  + -> reduced this round
 512    *  / -> sum (reduced in a previous step)
 513    *
 514    *     *** round 0 ***
 515    *    0        1        2
 516    *
 517    *   00       10       20      0: [1] -> 1
 518    *   01       11       21      1: [2] -> 2
 519    *   02       12       22      2: [0] -> 0  --> send element (r+1)%p to node (r+1)%p
 520    *
 521    *      *** round 1 ***
 522    *    0        1        2
 523    *
 524    *   00+20    10       20     0: red(0), [0] -> 1
 525    *   01       11+01    21     1: red(1), [1] -> 2
 526    *   02       12       22+12  2: red(2), [2] -> 0 --> reduce and send element (r+0)%p to node (r+1)%p
 527    *
 528    *      *** round 2 ***
 529    *    0        1        2
 530    *
 531    *   00/20    all      20     0: red(2), [2] -> 1
 532    *   01       11/01    all    1: red(0), [0] -> 2
 533    *   all      12       22/12  2: red(1), [1] -> 0 --> reduce and send (r-1)%p to node (r+1)%p
 534    *
 535    *      *** round 3 ***
 536    *    0        1        2
 537    *
 538    *   00/20    all      all    0: [1] -> 1
 539    *   all      11/01    all    1: [2] -> 2
 540    *   all      all      22/12  2: [0] -> 0 --> send element (r-2)%p to node (r+1)%p
 541    *
 542    *      *** round 4 ***
 543    *    0        1        2
 544    *
 545    *   all      all      all    0: done
 546    *   all      all      all    1: done
 547    *   all      all      all    2: done
 548    *
 549    * -> 4
 550    *     *** round 0 ***
 551    *    0        1        2        3
 552    *
 553    *   00       10       20       30       0: [1] -> 1
 554    *   01       11       21       31       1: [2] -> 2
 555    *   02       12       22       32       2: [3] -> 3
 556    *   03       13       23       33       3: [0] -> 0 --> send element (r+1)%p to node (r+1)%p
 557    *
 558    *      *** round 1 ***
 559    *    0        1        2        3
 560    *
 561    *   00+30    10       20       30       0: red(0), [0] -> 1
 562    *   01       11+01    21       31       1: red(1), [1] -> 2
 563    *   02       12       22+12    32       2: red(2), [2] -> 3
 564    *   03       13       23       33+23    3: red(3), [3] -> 0 --> reduce and send element (r+0)%p to node (r+1)%p
 565    *
 566    *      *** round 2 ***
 567    *    0        1        2        3
 568    *
 569    *   00/30    10+00/30 20       30       0: red(3), [3] -> 1
 570    *   01       11/01    21+11/01 31       1: red(0), [0] -> 2
 571    *   02       12       22/12    32+22/12 2: red(1), [1] -> 3
 572    *   03+33/23 13       23       33/23    3: red(2), [2] -> 0 --> reduce and send (r-1)%p to node (r+1)%p
 573    *
 574    *      *** round 3 ***
 575    *    0        1        2        3
 576    *
 577    *   00/30    10/00/30 all      30       0: red(2), [2] -> 1
 578    *   01       11/01    21/11/01 all      1: red(3), [3] -> 2
 579    *   all      12       22/12    32/22/12 2: red(0), [0] -> 3
 580    *   03/33/23 all      23       33/23    3: red(1), [1] -> 0 --> reduce and send (r-2)%p to node (r+1)%p
 581    *
 582    *      *** round 4 ***
 583    *    0        1        2        3
 584    *
 585    *   00/30    10/00/30 all      all      0: [1] -> 1
 586    *   all      11/01    21/11/01 all      1: [2] -> 2
 587    *   all      all      22/12    32/22/12 2: [3] -> 3
 588    *   03/33/23 all      all      33/23    3: [0] -> 0 --> receive and send element (r+1)%p to node (r+1)%p
 589    *
 590    *      *** round 5 ***
 591    *    0        1        2        3
 592    *
 593    *   all      10/00/30 all      all      0: [0] -> 1
 594    *   all      all      21/11/01 all      1: [1] -> 2
 595    *   all      all      all      32/22/12 2: [3] -> 3
 596    *   03/33/23 all      all      all      3: [4] -> 4 --> receive and send element (r-0)%p to node (r+1)%p
 597    *
 598    *      *** round 6 ***
 599    *    0        1        2        3
 600    *
 601    *   all      all      all      all
 602    *   all      all      all      all
 603    *   all      all      all      all
 604    *   all      all      all      all     receive element (r-1)%p
 605    *
 606    *   2p-2 rounds ... every node does p-1 reductions and p-1 sends
 607    *
 608    */
 609   /* first p-1 rounds are reductions */
 610   for (int round = 0 ; round < p - 1 ; ++round) {
 611     int selement = (r+1-round + 2*p /*2*p avoids negative mod*/)%p; /* the element I am sending */
 612     int soffset = segoffsets[selement]*ext;
 613     int relement = (r-round + 2*p /*2*p avoids negative mod*/)%p; /* the element that I receive from my neighbor */
 614     int roffset = segoffsets[relement]*ext;
 615 
 616     /* first message come out of sendbuf */
 617     if (round == 0) {
 618       res = NBC_Sched_send ((char *) sendbuf + soffset, false, segsizes[selement], datatype, speer,
 619                             schedule, false);
 620     } else {
 621       res = NBC_Sched_send ((char *) recvbuf + soffset, false, segsizes[selement], datatype, speer,
 622                             schedule, false);
 623     }
 624 
 625     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 626       break;
 627     }
 628 
 629     res = NBC_Sched_recv ((char *) recvbuf + roffset, false, segsizes[relement], datatype, rpeer,
 630                           schedule, true);
 631     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 632       break;
 633     }
 634 
 635     res = NBC_Sched_op ((char *) sendbuf + roffset, false, (char *) recvbuf + roffset, false,
 636                          segsizes[relement], datatype, op, schedule, true);
 637     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 638       break;
 639     }
 640   }
 641 
 642   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 643     free (segsizes);
 644     free (segoffsets);
 645     return res;
 646   }
 647 
 648   for (int round = p - 1 ; round < 2 * p - 2 ; ++round) {
 649     int selement = (r+1-round + 2*p /*2*p avoids negative mod*/)%p; /* the element I am sending */
 650     int soffset = segoffsets[selement]*ext;
 651     int relement = (r-round + 2*p /*2*p avoids negative mod*/)%p; /* the element that I receive from my neighbor */
 652     int roffset = segoffsets[relement]*ext;
 653 
 654     res = NBC_Sched_send ((char *) recvbuf + soffset, false, segsizes[selement], datatype, speer,
 655                           schedule, false);
 656     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 657       break;
 658     }
 659 
 660     res = NBC_Sched_recv ((char *) recvbuf + roffset, false, segsizes[relement], datatype, rpeer,
 661                           schedule, true);
 662     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 663       break;
 664     }
 665   }
 666 
 667   free (segsizes);
 668   free (segoffsets);
 669 
 670   return res;
 671 }
 672 
 673 static inline int allred_sched_linear(int rank, int rsize, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
 674                                       ptrdiff_t gap, MPI_Op op, int ext, int size, NBC_Schedule *schedule, void *tmpbuf) {
 675   int res;
 676 
 677   if (0 == count) {
 678     return OMPI_SUCCESS;
 679   }
 680 
 681   /* send my data to the remote root */
 682   res = NBC_Sched_send (sendbuf, false, count, datatype, 0, schedule, false);
 683   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 684     return res;
 685   }
 686 
 687   /* recv my data to the remote root */
 688   if (0 != rank || 1 ==(rsize%2)) {
 689     res = NBC_Sched_recv (recvbuf, false, count, datatype, 0, schedule, false);
 690   } else {
 691     res = NBC_Sched_recv ((void *)(-gap), true, count, datatype, 0, schedule, false);
 692   }
 693   if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 694     return res;
 695   }
 696 
 697   if (0 == rank) {
 698     char *rbuf, *lbuf, *buf;
 699     int tmprbuf, tmplbuf;
 700 
 701     res = NBC_Sched_barrier (schedule);
 702     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 703       return res;
 704     }
 705 
 706     /* ensure the result ends up in recvbuf */
 707     if (0 == (rsize%2)) {
 708       lbuf = (void *)(-gap);
 709       tmplbuf = true;
 710       rbuf = recvbuf;
 711       tmprbuf = false;
 712     } else {
 713       rbuf = (void *)(-gap);
 714       tmprbuf = true;
 715       lbuf = recvbuf;
 716       tmplbuf = false;
 717     }
 718 
 719     /* get data from remote peers and reduce */
 720     for (int rpeer = 1 ; rpeer < rsize ; ++rpeer) {
 721       res = NBC_Sched_recv (rbuf, tmprbuf, count, datatype, rpeer, schedule, true);
 722       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 723         return res;
 724       }
 725 
 726       res = NBC_Sched_op (lbuf, tmplbuf, rbuf, tmprbuf, count, datatype, op, schedule, true);
 727       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 728         return res;
 729       }
 730       /* swap left and right buffers */
 731       buf = rbuf; rbuf = lbuf ; lbuf = buf;
 732       tmprbuf ^= 1; tmplbuf ^= 1;
 733     }
 734 
 735     /* exchange our result with the remote root (each root will broadcast to the other's peers) */
 736     res = NBC_Sched_recv ((void *)(-gap), true, count, datatype, 0, schedule, false);
 737     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 738       return res;
 739     }
 740 
 741     /* wait for data from remote root */
 742     res = NBC_Sched_send (recvbuf, false, count, datatype, 0, schedule, true);
 743     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 744       return res;
 745     }
 746 
 747     /* broadcast the result to all remote peers */
 748     for (int rpeer = 1 ; rpeer < rsize ; ++rpeer) {
 749       res = NBC_Sched_send ((void *)(-gap), true, count, datatype, rpeer, schedule, false);
 750       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
 751         return res;
 752       }
 753     }
 754   }
 755 
 756   return OMPI_SUCCESS;
 757 }
 758 
 759 /*
 760  * allred_sched_redscat_allgather:
 761  *
 762  * Description: an implementation of Rabenseifner's Allreduce algorithm [1, 2].
 763  *   [1] Rajeev Thakur, Rolf Rabenseifner and William Gropp.
 764  *       Optimization of Collective Communication Operations in MPICH //
 765  *       The Int. Journal of High Performance Computing Applications. Vol 19,
 766  *       Issue 1, pp. 49--66.
 767  *   [2] http://www.hlrs.de/mpi/myreduce.html.
 768  *
 769  * This algorithm is a combination of a reduce-scatter implemented with
 770  * recursive vector halving and recursive distance doubling, followed either
 771  * by an allgather implemented with recursive doubling.
 772  *
 773  * Step 1. If the number of processes is not a power of two, reduce it to
 774  * the nearest lower power of two (p' = 2^{\floor{\log_2 p}})
 775  * by removing r = p - p' extra processes as follows. In the first 2r processes
 776  * (ranks 0 to 2r - 1), all the even ranks send the second half of the input
 777  * vector to their right neighbor (rank + 1), and all the odd ranks send
 778  * the first half of the input vector to their left neighbor (rank - 1).
 779  * The even ranks compute the reduction on the first half of the vector and
 780  * the odd ranks compute the reduction on the second half. The odd ranks then
 781  * send the result to their left neighbors (the even ranks). As a result,
 782  * the even ranks among the first 2r processes now contain the reduction with
 783  * the input vector on their right neighbors (the odd ranks). These odd ranks
 784  * do not participate in the rest of the algorithm, which leaves behind
 785  * a power-of-two number of processes. The first r even-ranked processes and
 786  * the last p - 2r processes are now renumbered from 0 to p' - 1.
 787  *
 788  * Step 2. The remaining processes now perform a reduce-scatter by using
 789  * recursive vector halving and recursive distance doubling. The even-ranked
 790  * processes send the second half of their buffer to rank + 1 and the odd-ranked
 791  * processes send the first half of their buffer to rank - 1. All processes
 792  * then compute the reduction between the local buffer and the received buffer.
 793  * In the next log_2(p') - 1 steps, the buffers are recursively halved, and the
 794  * distance is doubled. At the end, each of the p' processes has 1 / p' of the
 795  * total reduction result.
 796  *
 797  * Step 3. An allgather is performed by using recursive vector doubling and
 798  * distance halving. All exchanges are executed in reverse order relative
 799  * to recursive doubling on previous step. If the number of processes is not
 800  * a power of two, the total result vector must be sent to the r processes
 801  * that were removed in the first step.
 802  *
 803  * Limitations:
 804  *   count >= 2^{\floor{\log_2 p}}
 805  *   commutative operations only
 806  *   intra-communicators only
 807  *
 808  * Memory requirements (per process):
 809  *   count * typesize + 4 * \log_2(p) * sizeof(int) = O(count)
 810  *
 811  * Schedule length (rounds): O(\log(p))
 812  */
 813 static inline int allred_sched_redscat_allgather(
 814     int rank, int comm_size, int count, MPI_Datatype datatype, ptrdiff_t gap,
 815     const void *sbuf, void *rbuf, MPI_Op op, char inplace,
 816     NBC_Schedule *schedule, void *tmpbuf, struct ompi_communicator_t *comm)
 817 {
 818     int res = OMPI_SUCCESS;
 819     int *rindex = NULL, *rcount = NULL, *sindex = NULL, *scount = NULL;
 820      /* Find nearest power-of-two less than or equal to comm_size */
 821     int nsteps = opal_hibit(comm_size, comm->c_cube_dim + 1);   /* ilog2(comm_size) */
 822     int nprocs_pof2 = 1 << nsteps;                              /* flp2(comm_size) */
 823      if (!inplace) {
 824         res = NBC_Sched_copy((char *)sbuf, false, count, datatype,
 825                              rbuf, false, count, datatype, schedule, true);
 826         if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 827     }
 828     char *tmp_buf = (char *)tmpbuf - gap;
 829     ptrdiff_t lb, extent;
 830     ompi_datatype_get_extent(datatype, &lb, &extent);
 831      /*
 832      * Step 1. Reduce the number of processes to the nearest lower power of two
 833      * p' = 2^{\floor{\log_2 p}} by removing r = p - p' processes.
 834      * 1. In the first 2r processes (ranks 0 to 2r - 1), all the even ranks send
 835      *    the second half of the input vector to their right neighbor (rank + 1)
 836      *    and all the odd ranks send the first half of the input vector to their
 837      *    left neighbor (rank - 1).
 838      * 2. All 2r processes compute the reduction on their half.
 839      * 3. The odd ranks then send the result to their left neighbors
 840      *    (the even ranks).
 841      *
 842      * The even ranks (0 to 2r - 1) now contain the reduction with the input
 843      * vector on their right neighbors (the odd ranks). The first r even
 844      * processes and the p - 2r last processes are renumbered from
 845      * 0 to 2^{\floor{\log_2 p}} - 1.
 846      */
 847     int vrank, step, wsize;
 848     int nprocs_rem = comm_size - nprocs_pof2;
 849      if (rank < 2 * nprocs_rem) {
 850         int count_lhalf = count / 2;
 851         int count_rhalf = count - count_lhalf;
 852          if (rank % 2 != 0) {
 853             /*
 854              * Odd process -- exchange with rank - 1
 855              * Send the left half of the input vector to the left neighbor,
 856              * Recv the right half of the input vector from the left neighbor
 857              */
 858             res = NBC_Sched_send(rbuf, false, count_lhalf, datatype, rank - 1,
 859                                  schedule, false);
 860             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 861              res = NBC_Sched_recv(tmp_buf + (ptrdiff_t)count_lhalf * extent,
 862                                  false, count_rhalf, datatype, rank - 1, schedule, true);
 863             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 864              res = NBC_Sched_op(tmp_buf + (ptrdiff_t)count_lhalf * extent,
 865                                false, (char *)rbuf + (ptrdiff_t)count_lhalf * extent,
 866                                false, count_rhalf, datatype, op, schedule, true);
 867             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 868              /* Send the right half to the left neighbor */
 869             res = NBC_Sched_send((char *)rbuf + (ptrdiff_t)count_lhalf * extent,
 870                                  false, count_rhalf, datatype, rank - 1, schedule, true);
 871             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 872              /* This process does not participate in recursive doubling phase */
 873             vrank = -1;
 874          } else {
 875             /*
 876              * Even process -- exchange with rank + 1
 877              * Send the right half of the input vector to the right neighbor,
 878              * Recv the left half of the input vector from the right neighbor
 879              */
 880             res = NBC_Sched_send((char *)rbuf + (ptrdiff_t)count_lhalf * extent,
 881                                  false, count_rhalf, datatype, rank + 1, schedule, false);
 882             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 883              res = NBC_Sched_recv(tmp_buf, false, count_lhalf, datatype, rank + 1,
 884                                  schedule, true);
 885             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 886              res = NBC_Sched_op(tmp_buf, false, rbuf, false, count_lhalf,
 887                                datatype, op, schedule, true);
 888             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 889              /* Recv the right half from the right neighbor */
 890             res = NBC_Sched_recv((char *)rbuf + (ptrdiff_t)count_lhalf * extent,
 891                                  false, count_rhalf, datatype, rank + 1, schedule, true);
 892             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 893              vrank = rank / 2;
 894         }
 895     } else { /* rank >= 2 * nprocs_rem */
 896         vrank = rank - nprocs_rem;
 897     }
 898      /*
 899      * Step 2. Reduce-scatter implemented with recursive vector halving and
 900      * recursive distance doubling. We have p' = 2^{\floor{\log_2 p}}
 901      * power-of-two number of processes with new ranks (vrank) and result in rbuf.
 902      *
 903      * The even-ranked processes send the right half of their buffer to rank + 1
 904      * and the odd-ranked processes send the left half of their buffer to
 905      * rank - 1. All processes then compute the reduction between the local
 906      * buffer and the received buffer. In the next \log_2(p') - 1 steps, the
 907      * buffers are recursively halved, and the distance is doubled. At the end,
 908      * each of the p' processes has 1 / p' of the total reduction result.
 909      */
 910     rindex = malloc(sizeof(*rindex) * nsteps);
 911     sindex = malloc(sizeof(*sindex) * nsteps);
 912     rcount = malloc(sizeof(*rcount) * nsteps);
 913     scount = malloc(sizeof(*scount) * nsteps);
 914     if (NULL == rindex || NULL == sindex || NULL == rcount || NULL == scount) {
 915         res = OMPI_ERR_OUT_OF_RESOURCE;
 916         goto cleanup_and_return;
 917     }
 918      if (vrank != -1) {
 919         step = 0;
 920         wsize = count;
 921         sindex[0] = rindex[0] = 0;
 922          for (int mask = 1; mask < nprocs_pof2; mask <<= 1) {
 923             /*
 924              * On each iteration: rindex[step] = sindex[step] -- begining of the
 925              * current window. Length of the current window is storded in wsize.
 926              */
 927             int vdest = vrank ^ mask;
 928             /* Translate vdest virtual rank to real rank */
 929             int dest = (vdest < nprocs_rem) ? vdest * 2 : vdest + nprocs_rem;
 930              if (rank < dest) {
 931                 /*
 932                  * Recv into the left half of the current window, send the right
 933                  * half of the window to the peer (perform reduce on the left
 934                  * half of the current window)
 935                  */
 936                 rcount[step] = wsize / 2;
 937                 scount[step] = wsize - rcount[step];
 938                 sindex[step] = rindex[step] + rcount[step];
 939             } else {
 940                 /*
 941                  * Recv into the right half of the current window, send the left
 942                  * half of the window to the peer (perform reduce on the right
 943                  * half of the current window)
 944                  */
 945                 scount[step] = wsize / 2;
 946                 rcount[step] = wsize - scount[step];
 947                 rindex[step] = sindex[step] + scount[step];
 948             }
 949              /* Send part of data from the rbuf, recv into the tmp_buf */
 950             res = NBC_Sched_send((char *)rbuf + (ptrdiff_t)sindex[step] * extent,
 951                                  false, scount[step], datatype, dest, schedule, false);
 952             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 953             res = NBC_Sched_recv((char *)tmp_buf + (ptrdiff_t)rindex[step] * extent,
 954                                  false, rcount[step], datatype, dest, schedule, true);
 955             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 956              /* Local reduce: rbuf[] = tmp_buf[] <op> rbuf[] */
 957             res = NBC_Sched_op((char *)tmp_buf + (ptrdiff_t)rindex[step] * extent,
 958                                false, (char *)rbuf + (ptrdiff_t)rindex[step] * extent,
 959                                false, rcount[step], datatype, op, schedule, true);
 960             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 961              /* Move the current window to the received message */
 962             if (step + 1 < nsteps) {
 963                 rindex[step + 1] = rindex[step];
 964                 sindex[step + 1] = rindex[step];
 965                 wsize = rcount[step];
 966                 step++;
 967             }
 968         }
 969         /*
 970          * Assertion: each process has 1 / p' of the total reduction result:
 971          * rcount[nsteps - 1] elements in the rbuf[rindex[nsteps - 1], ...].
 972          */
 973          /*
 974          * Step 3. Allgather by the recursive doubling algorithm.
 975          * Each process has 1 / p' of the total reduction result:
 976          * rcount[nsteps - 1] elements in the rbuf[rindex[nsteps - 1], ...].
 977          * All exchanges are executed in reverse order relative
 978          * to recursive doubling (previous step).
 979          */
 980         step = nsteps - 1;
 981          for (int mask = nprocs_pof2 >> 1; mask > 0; mask >>= 1) {
 982             int vdest = vrank ^ mask;
 983             /* Translate vdest virtual rank to real rank */
 984             int dest = (vdest < nprocs_rem) ? vdest * 2 : vdest + nprocs_rem;
 985              /*
 986              * Send rcount[step] elements from rbuf[rindex[step]...]
 987              * Recv scount[step] elements to rbuf[sindex[step]...]
 988              */
 989             res = NBC_Sched_send((char *)rbuf + (ptrdiff_t)rindex[step] * extent,
 990                                  false, rcount[step], datatype, dest, schedule, false);
 991             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 992             res = NBC_Sched_recv((char *)rbuf + (ptrdiff_t)sindex[step] * extent,
 993                                  false, scount[step], datatype, dest, schedule, true);
 994             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
 995             step--;
 996         }
 997     }
 998      /*
 999      * Step 4. Send total result to excluded odd ranks.
1000      */
1001     if (rank < 2 * nprocs_rem) {
1002         if (rank % 2 != 0) {
1003             /* Odd process -- recv result from rank - 1 */
1004             res = NBC_Sched_recv(rbuf, false, count, datatype, rank - 1, schedule, false);
1005             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
1006         } else {
1007             /* Even process -- send result to rank + 1 */
1008             res = NBC_Sched_send(rbuf, false, count, datatype, rank + 1, schedule, false);
1009             if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
1010         }
1011     }
1012    cleanup_and_return:
1013     if (NULL != rindex)
1014         free(rindex);
1015     if (NULL != sindex)
1016         free(sindex);
1017     if (NULL != rcount)
1018         free(rcount);
1019     if (NULL != scount)
1020         free(scount);
1021     return res;
1022 }
1023 
1024 int ompi_coll_libnbc_allreduce_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
1025                                     struct ompi_communicator_t *comm, MPI_Info info, ompi_request_t ** request,
1026                                     struct mca_coll_base_module_2_3_0_t *module) {
1027     int res = nbc_allreduce_init(sendbuf, recvbuf, count, datatype, op,
1028                                  comm, request, module, true);
1029     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
1030         return res;
1031     }
1032 
1033     return OMPI_SUCCESS;
1034 }
1035 
1036 int ompi_coll_libnbc_allreduce_inter_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
1037                                           struct ompi_communicator_t *comm, MPI_Info info, ompi_request_t ** request,
1038                                           struct mca_coll_base_module_2_3_0_t *module) {
1039     int res = nbc_allreduce_inter_init(sendbuf, recvbuf, count, datatype, op,
1040                                        comm, request, module, true);
1041     if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
1042         return res;
1043     }
1044 
1045     return OMPI_SUCCESS;
1046 }
1047