src/gromacs/gmxlib/network.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016,2017, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "network.h"
  40
  41 #include "config.h"
  42
  43 #include <cctype>
  44 #include <cstdarg>
  45 #include <cstdlib>
  46 #include <cstring>
  47
  48 #include "gromacs/commandline/filenm.h"
  49 #include "gromacs/mdtypes/commrec.h"
  50 #include "gromacs/utility/basenetwork.h"
  51 #include "gromacs/utility/cstringutil.h"
  52 #include "gromacs/utility/fatalerror.h"
  53 #include "gromacs/utility/futil.h"
  54 #include "gromacs/utility/gmxmpi.h"
  55 #include "gromacs/utility/smalloc.h"
  56
  57 /* The source code in this file should be thread-safe.
  58       Please keep it that way. */
  59
  60 void gmx_fill_commrec_from_mpi(t_commrec gmx_unused *cr)
  61 {
  62 #if !GMX_MPI
  63     gmx_call("gmx_fill_commrec_from_mpi");
  64 #else
  65     if (!gmx_mpi_initialized())
  66     {
  67         gmx_comm("MPI has not been initialized properly");
  68     }
  69
  70     cr->nnodes           = gmx_node_num();
  71     cr->nodeid           = gmx_node_rank();
  72     if (PAR(cr) || MULTISIM(cr))
  73     {
  74         MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(), cr->nodeid, &cr->mpi_comm_physicalnode);
  75     }
  76     cr->sim_nodeid       = cr->nodeid;
  77     cr->mpi_comm_mysim   = MPI_COMM_WORLD;
  78     cr->mpi_comm_mygroup = MPI_COMM_WORLD;
  79
  80 #endif
  81 }
  82
  83 t_commrec *init_commrec()
  84 {
  85     t_commrec    *cr;
  86
  87     snew(cr, 1);
  88
  89 #if GMX_LIB_MPI
  90     gmx_fill_commrec_from_mpi(cr);
  91 #else
  92     cr->mpi_comm_mysim   = nullptr;
  93     cr->mpi_comm_mygroup = nullptr;
  94     cr->nnodes           = 1;
  95     cr->sim_nodeid       = 0;
  96     cr->nodeid           = cr->sim_nodeid;
  97 #endif
  98
  99     // TODO cr->duty should not be initialized here
 100     cr->duty = (DUTY_PP | DUTY_PME);
 101
 102 #if GMX_MPI && !MPI_IN_PLACE_EXISTS
 103     /* initialize the MPI_IN_PLACE replacement buffers */
 104     snew(cr->mpb, 1);
 105     cr->mpb->ibuf        = NULL;
 106     cr->mpb->libuf       = NULL;
 107     cr->mpb->fbuf        = NULL;
 108     cr->mpb->dbuf        = NULL;
 109     cr->mpb->ibuf_alloc  = 0;
 110     cr->mpb->libuf_alloc = 0;
 111     cr->mpb->fbuf_alloc  = 0;
 112     cr->mpb->dbuf_alloc  = 0;
 113 #endif
 114
 115     return cr;
 116 }
 117
 118 static void done_mpi_in_place_buf(mpi_in_place_buf_t *buf)
 119 {
 120     if (nullptr != buf)
 121     {
 122         sfree(buf->ibuf);
 123         sfree(buf->libuf);
 124         sfree(buf->fbuf);
 125         sfree(buf->dbuf);
 126         sfree(buf);
 127     }
 128 }
 129
 130 void done_commrec(t_commrec *cr)
 131 {
 132 #if GMX_MPI
 133     if (PAR(cr) || MULTISIM(cr))
 134     {
 135         MPI_Comm_free(&cr->mpi_comm_physicalnode);
 136     }
 137 #endif
 138     if (nullptr != cr->dd)
 139     {
 140         // TODO: implement
 141         // done_domdec(cr->dd);
 142     }
 143     if (nullptr != cr->ms)
 144     {
 145         done_mpi_in_place_buf(cr->ms->mpb);
 146         sfree(cr->ms);
 147     }
 148     done_mpi_in_place_buf(cr->mpb);
 149     sfree(cr);
 150 }
 151
 152 t_commrec *reinitialize_commrec_for_this_thread(const t_commrec gmx_unused *cro)
 153 {
 154 #if GMX_THREAD_MPI
 155     t_commrec *cr;
 156
 157     /* make a thread-specific commrec */
 158     snew(cr, 1);
 159     /* now copy the whole thing, so settings like the number of PME nodes
 160        get propagated. */
 161     *cr = *cro;
 162
 163     /* and we start setting our own thread-specific values for things */
 164     gmx_fill_commrec_from_mpi(cr);
 165
 166     // TODO cr->duty should not be initialized here
 167     cr->duty             = (DUTY_PP | DUTY_PME);
 168
 169     return cr;
 170 #else
 171     return nullptr;
 172 #endif
 173 }
 174
 175 void gmx_setup_nodecomm(FILE gmx_unused *fplog, t_commrec *cr)
 176 {
 177     gmx_nodecomm_t *nc;
 178
 179     /* Many MPI implementations do not optimize MPI_Allreduce
 180      * (and probably also other global communication calls)
 181      * for multi-core nodes connected by a network.
 182      * We can optimize such communication by using one MPI call
 183      * within each node and one between the nodes.
 184      * For MVAPICH2 and Intel MPI this reduces the time for
 185      * the global_stat communication by 25%
 186      * for 2x2-core 3 GHz Woodcrest connected by mixed DDR/SDR Infiniband.
 187      * B. Hess, November 2007
 188      */
 189
 190     nc = &cr->nc;
 191
 192     nc->bUse = FALSE;
 193 #if !GMX_THREAD_MPI
 194 #if GMX_MPI
 195     int n, rank;
 196
 197     MPI_Comm_size(cr->mpi_comm_mygroup, &n);
 198     MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
 199
 200     int nodehash = gmx_physicalnode_id_hash();
 201
 202     if (debug)
 203     {
 204         fprintf(debug, "In gmx_setup_nodecomm: splitting communicator of size %d\n", n);
 205     }
 206
 207
 208     /* The intra-node communicator, split on node number */
 209     MPI_Comm_split(cr->mpi_comm_mygroup, nodehash, rank, &nc->comm_intra);
 210     MPI_Comm_rank(nc->comm_intra, &nc->rank_intra);
 211     if (debug)
 212     {
 213         fprintf(debug, "In gmx_setup_nodecomm: node ID %d rank within node %d\n",
 214                 rank, nc->rank_intra);
 215     }
 216     /* The inter-node communicator, split on rank_intra.
 217      * We actually only need the one for rank=0,
 218      * but it is easier to create them all.
 219      */
 220     MPI_Comm_split(cr->mpi_comm_mygroup, nc->rank_intra, rank, &nc->comm_inter);
 221     /* Check if this really created two step communication */
 222     int ng, ni;
 223
 224     MPI_Comm_size(nc->comm_inter, &ng);
 225     MPI_Comm_size(nc->comm_intra, &ni);
 226     if (debug)
 227     {
 228         fprintf(debug, "In gmx_setup_nodecomm: groups %d, my group size %d\n",
 229                 ng, ni);
 230     }
 231
 232     if (getenv("GMX_NO_NODECOMM") == nullptr &&
 233         ((ng > 1 && ng < n) || (ni > 1 && ni < n)))
 234     {
 235         nc->bUse = TRUE;
 236         if (fplog)
 237         {
 238             fprintf(fplog, "Using two step summing over %d groups of on average %.1f ranks\n\n",
 239                     ng, (real)n/(real)ng);
 240         }
 241         if (nc->rank_intra > 0)
 242         {
 243             MPI_Comm_free(&nc->comm_inter);
 244         }
 245     }
 246     else
 247     {
 248         /* One group or all processes in a separate group, use normal summing */
 249         MPI_Comm_free(&nc->comm_inter);
 250         MPI_Comm_free(&nc->comm_intra);
 251         if (debug)
 252         {
 253             fprintf(debug, "In gmx_setup_nodecomm: not unsing separate inter- and intra-node communicators.\n");
 254         }
 255     }
 256 #endif
 257 #else
 258     /* tMPI runs only on a single node so just use the nodeid */
 259     nc->rank_intra = cr->nodeid;
 260 #endif
 261 }
 262
 263 void gmx_init_intranode_counters(t_commrec *cr)
 264 {
 265     /* counters for PP+PME and PP-only processes on my physical node */
 266     int nrank_intranode, rank_intranode;
 267     int nrank_pp_intranode, rank_pp_intranode;
 268     /* thread-MPI is not initialized when not running in parallel */
 269 #if GMX_MPI && !GMX_THREAD_MPI
 270     int nrank_world, rank_world;
 271     int i, myhash, *hash, *hash_s, *hash_pp, *hash_pp_s;
 272
 273     MPI_Comm_size(MPI_COMM_WORLD, &nrank_world);
 274     MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
 275
 276     /* Get a (hopefully unique) hash that identifies our physical node */
 277     myhash = gmx_physicalnode_id_hash();
 278
 279     /* We can't rely on MPI_IN_PLACE, so we need send and receive buffers */
 280     snew(hash,   nrank_world);
 281     snew(hash_s, nrank_world);
 282     snew(hash_pp,   nrank_world);
 283     snew(hash_pp_s, nrank_world);
 284
 285     hash_s[rank_world]    = myhash;
 286     hash_pp_s[rank_world] = (cr->duty & DUTY_PP) ? myhash : -1;
 287
 288     MPI_Allreduce(hash_s,    hash,    nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
 289     MPI_Allreduce(hash_pp_s, hash_pp, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
 290
 291     nrank_intranode    = 0;
 292     rank_intranode     = 0;
 293     nrank_pp_intranode = 0;
 294     rank_pp_intranode  = 0;
 295     for (i = 0; i < nrank_world; i++)
 296     {
 297         if (hash[i] == myhash)
 298         {
 299             nrank_intranode++;
 300             if (i < rank_world)
 301             {
 302                 rank_intranode++;
 303             }
 304         }
 305         if (hash_pp[i] == myhash)
 306         {
 307             nrank_pp_intranode++;
 308             if ((cr->duty & DUTY_PP) && i < rank_world)
 309             {
 310                 rank_pp_intranode++;
 311             }
 312         }
 313     }
 314     sfree(hash);
 315     sfree(hash_s);
 316     sfree(hash_pp);
 317     sfree(hash_pp_s);
 318 #else
 319     /* Serial or thread-MPI code: we run within a single physical node */
 320     nrank_intranode    = cr->nnodes;
 321     rank_intranode     = cr->sim_nodeid;
 322     nrank_pp_intranode = cr->nnodes - cr->npmenodes;
 323     rank_pp_intranode  = cr->nodeid;
 324 #endif
 325
 326     if (debug)
 327     {
 328         char sbuf[STRLEN];
 329         if ((cr->duty & DUTY_PP) && (cr->duty & DUTY_PME))
 330         {
 331             sprintf(sbuf, "PP+PME");
 332         }
 333         else
 334         {
 335             sprintf(sbuf, "%s", (cr->duty & DUTY_PP) ? "PP" : "PME");
 336         }
 337         fprintf(debug, "On %3s rank %d: nrank_intranode=%d, rank_intranode=%d, "
 338                 "nrank_pp_intranode=%d, rank_pp_intranode=%d\n",
 339                 sbuf, cr->sim_nodeid,
 340                 nrank_intranode, rank_intranode,
 341                 nrank_pp_intranode, rank_pp_intranode);
 342     }
 343
 344     cr->nrank_intranode    = nrank_intranode;
 345     cr->rank_intranode     = rank_intranode;
 346     cr->nrank_pp_intranode = nrank_pp_intranode;
 347     cr->rank_pp_intranode  = rank_pp_intranode;
 348 }
 349
 350
 351 void gmx_barrier(const t_commrec gmx_unused *cr)
 352 {
 353 #if !GMX_MPI
 354     gmx_call("gmx_barrier");
 355 #else
 356     MPI_Barrier(cr->mpi_comm_mygroup);
 357 #endif
 358 }
 359
 360 void gmx_barrier_physical_node(const t_commrec gmx_unused *cr)
 361 {
 362 #if !GMX_MPI
 363     gmx_call("gmx_barrier_physical_node");
 364 #else
 365     MPI_Barrier(cr->mpi_comm_physicalnode);
 366 #endif
 367 }
 368
 369 void gmx_bcast(int gmx_unused nbytes, void gmx_unused *b, const t_commrec gmx_unused *cr)
 370 {
 371 #if !GMX_MPI
 372     gmx_call("gmx_bast");
 373 #else
 374     MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mygroup);
 375 #endif
 376 }
 377
 378 void gmx_bcast_sim(int gmx_unused nbytes, void gmx_unused *b, const t_commrec gmx_unused *cr)
 379 {
 380 #if !GMX_MPI
 381     gmx_call("gmx_bast");
 382 #else
 383     MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mysim);
 384 #endif
 385 }
 386
 387 void gmx_sumd(int gmx_unused nr, double gmx_unused r[], const t_commrec gmx_unused *cr)
 388 {
 389 #if !GMX_MPI
 390     gmx_call("gmx_sumd");
 391 #else
 392 #if MPI_IN_PLACE_EXISTS
 393     if (cr->nc.bUse)
 394     {
 395         if (cr->nc.rank_intra == 0)
 396         {
 397             /* Use two step summing. */
 398             MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, 0,
 399                        cr->nc.comm_intra);
 400             /* Sum the roots of the internal (intra) buffers. */
 401             MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
 402                           cr->nc.comm_inter);
 403         }
 404         else
 405         {
 406             /* This is here because of the silly MPI specification
 407                 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
 408             MPI_Reduce(r, nullptr, nr, MPI_DOUBLE, MPI_SUM, 0, cr->nc.comm_intra);
 409         }
 410         MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
 411     }
 412     else
 413     {
 414         MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
 415                       cr->mpi_comm_mygroup);
 416     }
 417 #else
 418     int i;
 419
 420     if (nr > cr->mpb->dbuf_alloc)
 421     {
 422         cr->mpb->dbuf_alloc = nr;
 423         srenew(cr->mpb->dbuf, cr->mpb->dbuf_alloc);
 424     }
 425     if (cr->nc.bUse)
 426     {
 427         /* Use two step summing */
 428         MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM, cr->nc.comm_intra);
 429         if (cr->nc.rank_intra == 0)
 430         {
 431             /* Sum with the buffers reversed */
 432             MPI_Allreduce(cr->mpb->dbuf, r, nr, MPI_DOUBLE, MPI_SUM,
 433                           cr->nc.comm_inter);
 434         }
 435         MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
 436     }
 437     else
 438     {
 439         MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM,
 440                       cr->mpi_comm_mygroup);
 441         for (i = 0; i < nr; i++)
 442         {
 443             r[i] = cr->mpb->dbuf[i];
 444         }
 445     }
 446 #endif
 447 #endif
 448 }
 449
 450 void gmx_sumf(int gmx_unused nr, float gmx_unused r[], const t_commrec gmx_unused *cr)
 451 {
 452 #if !GMX_MPI
 453     gmx_call("gmx_sumf");
 454 #else
 455 #if MPI_IN_PLACE_EXISTS
 456     if (cr->nc.bUse)
 457     {
 458         /* Use two step summing.  */
 459         if (cr->nc.rank_intra == 0)
 460         {
 461             MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, 0,
 462                        cr->nc.comm_intra);
 463             /* Sum the roots of the internal (intra) buffers */
 464             MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM,
 465                           cr->nc.comm_inter);
 466         }
 467         else
 468         {
 469             /* This is here because of the silly MPI specification
 470                 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
 471             MPI_Reduce(r, nullptr, nr, MPI_FLOAT, MPI_SUM, 0, cr->nc.comm_intra);
 472         }
 473         MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
 474     }
 475     else
 476     {
 477         MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, cr->mpi_comm_mygroup);
 478     }
 479 #else
 480     int i;
 481
 482     if (nr > cr->mpb->fbuf_alloc)
 483     {
 484         cr->mpb->fbuf_alloc = nr;
 485         srenew(cr->mpb->fbuf, cr->mpb->fbuf_alloc);
 486     }
 487     if (cr->nc.bUse)
 488     {
 489         /* Use two step summing */
 490         MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM, cr->nc.comm_intra);
 491         if (cr->nc.rank_intra == 0)
 492         {
 493             /* Sum with the buffers reversed */
 494             MPI_Allreduce(cr->mpb->fbuf, r, nr, MPI_FLOAT, MPI_SUM,
 495                           cr->nc.comm_inter);
 496         }
 497         MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
 498     }
 499     else
 500     {
 501         MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM,
 502                       cr->mpi_comm_mygroup);
 503         for (i = 0; i < nr; i++)
 504         {
 505             r[i] = cr->mpb->fbuf[i];
 506         }
 507     }
 508 #endif
 509 #endif
 510 }
 511
 512 void gmx_sumi(int gmx_unused nr, int gmx_unused r[], const t_commrec gmx_unused *cr)
 513 {
 514 #if !GMX_MPI
 515     gmx_call("gmx_sumi");
 516 #else
 517 #if MPI_IN_PLACE_EXISTS
 518     if (cr->nc.bUse)
 519     {
 520         /* Use two step summing */
 521         if (cr->nc.rank_intra == 0)
 522         {
 523             MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
 524             /* Sum with the buffers reversed */
 525             MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
 526         }
 527         else
 528         {
 529             /* This is here because of the silly MPI specification
 530                 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
 531             MPI_Reduce(r, nullptr, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
 532         }
 533         MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
 534     }
 535     else
 536     {
 537         MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
 538     }
 539 #else
 540     int i;
 541
 542     if (nr > cr->mpb->ibuf_alloc)
 543     {
 544         cr->mpb->ibuf_alloc = nr;
 545         srenew(cr->mpb->ibuf, cr->mpb->ibuf_alloc);
 546     }
 547     if (cr->nc.bUse)
 548     {
 549         /* Use two step summing */
 550         MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->nc.comm_intra);
 551         if (cr->nc.rank_intra == 0)
 552         {
 553             /* Sum with the buffers reversed */
 554             MPI_Allreduce(cr->mpb->ibuf, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
 555         }
 556         MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
 557     }
 558     else
 559     {
 560         MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
 561         for (i = 0; i < nr; i++)
 562         {
 563             r[i] = cr->mpb->ibuf[i];
 564         }
 565     }
 566 #endif
 567 #endif
 568 }
 569
 570 void gmx_sumli(int gmx_unused nr, gmx_int64_t gmx_unused r[], const t_commrec gmx_unused *cr)
 571 {
 572 #if !GMX_MPI
 573     gmx_call("gmx_sumli");
 574 #else
 575 #if MPI_IN_PLACE_EXISTS
 576     if (cr->nc.bUse)
 577     {
 578         /* Use two step summing */
 579         if (cr->nc.rank_intra == 0)
 580         {
 581             MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, 0,
 582                        cr->nc.comm_intra);
 583             /* Sum with the buffers reversed */
 584             MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM,
 585                           cr->nc.comm_inter);
 586         }
 587         else
 588         {
 589             /* This is here because of the silly MPI specification
 590                 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
 591             MPI_Reduce(r, nullptr, nr, MPI_INT64_T, MPI_SUM, 0, cr->nc.comm_intra);
 592         }
 593         MPI_Bcast(r, nr, MPI_INT64_T, 0, cr->nc.comm_intra);
 594     }
 595     else
 596     {
 597         MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, cr->mpi_comm_mygroup);
 598     }
 599 #else
 600     int i;
 601
 602     if (nr > cr->mpb->libuf_alloc)
 603     {
 604         cr->mpb->libuf_alloc = nr;
 605         srenew(cr->mpb->libuf, cr->mpb->libuf_alloc);
 606     }
 607     if (cr->nc.bUse)
 608     {
 609         /* Use two step summing */
 610         MPI_Allreduce(r, cr->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
 611                       cr->nc.comm_intra);
 612         if (cr->nc.rank_intra == 0)
 613         {
 614             /* Sum with the buffers reversed */
 615             MPI_Allreduce(cr->mpb->libuf, r, nr, MPI_INT64_T, MPI_SUM,
 616                           cr->nc.comm_inter);
 617         }
 618         MPI_Bcast(r, nr, MPI_INT64_T, 0, cr->nc.comm_intra);
 619     }
 620     else
 621     {
 622         MPI_Allreduce(r, cr->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
 623                       cr->mpi_comm_mygroup);
 624         for (i = 0; i < nr; i++)
 625         {
 626             r[i] = cr->mpb->libuf[i];
 627         }
 628     }
 629 #endif
 630 #endif
 631 }
 632
 633
 634
 635 #if GMX_MPI
 636 static void gmx_sumd_comm(int nr, double r[], MPI_Comm mpi_comm)
 637 {
 638 #if MPI_IN_PLACE_EXISTS
 639     MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
 640 #else
 641     /* this function is only used in code that is not performance critical,
 642        (during setup, when comm_rec is not the appropriate communication
 643        structure), so this isn't as bad as it looks. */
 644     double *buf;
 645     int     i;
 646
 647     snew(buf, nr);
 648     MPI_Allreduce(r, buf, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
 649     for (i = 0; i < nr; i++)
 650     {
 651         r[i] = buf[i];
 652     }
 653     sfree(buf);
 654 #endif
 655 }
 656 #endif
 657
 658 #if GMX_MPI
 659 static void gmx_sumf_comm(int nr, float r[], MPI_Comm mpi_comm)
 660 {
 661 #if MPI_IN_PLACE_EXISTS
 662     MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
 663 #else
 664     /* this function is only used in code that is not performance critical,
 665        (during setup, when comm_rec is not the appropriate communication
 666        structure), so this isn't as bad as it looks. */
 667     float *buf;
 668     int    i;
 669
 670     snew(buf, nr);
 671     MPI_Allreduce(r, buf, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
 672     for (i = 0; i < nr; i++)
 673     {
 674         r[i] = buf[i];
 675     }
 676     sfree(buf);
 677 #endif
 678 }
 679 #endif
 680
 681 void gmx_sumd_sim(int gmx_unused nr, double gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
 682 {
 683 #if !GMX_MPI
 684     gmx_call("gmx_sumd_sim");
 685 #else
 686     gmx_sumd_comm(nr, r, ms->mpi_comm_masters);
 687 #endif
 688 }
 689
 690 void gmx_sumf_sim(int gmx_unused nr, float gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
 691 {
 692 #if !GMX_MPI
 693     gmx_call("gmx_sumf_sim");
 694 #else
 695     gmx_sumf_comm(nr, r, ms->mpi_comm_masters);
 696 #endif
 697 }
 698
 699 void gmx_sumi_sim(int gmx_unused nr, int gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
 700 {
 701 #if !GMX_MPI
 702     gmx_call("gmx_sumi_sim");
 703 #else
 704 #if MPI_IN_PLACE_EXISTS
 705     MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
 706 #else
 707     /* this is thread-unsafe, but it will do for now: */
 708     int i;
 709
 710     if (nr > ms->mpb->ibuf_alloc)
 711     {
 712         ms->mpb->ibuf_alloc = nr;
 713         srenew(ms->mpb->ibuf, ms->mpb->ibuf_alloc);
 714     }
 715     MPI_Allreduce(r, ms->mpb->ibuf, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
 716     for (i = 0; i < nr; i++)
 717     {
 718         r[i] = ms->mpb->ibuf[i];
 719     }
 720 #endif
 721 #endif
 722 }
 723
 724 void gmx_sumli_sim(int gmx_unused nr, gmx_int64_t gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
 725 {
 726 #if !GMX_MPI
 727     gmx_call("gmx_sumli_sim");
 728 #else
 729 #if MPI_IN_PLACE_EXISTS
 730     MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM,
 731                   ms->mpi_comm_masters);
 732 #else
 733     /* this is thread-unsafe, but it will do for now: */
 734     int i;
 735
 736     if (nr > ms->mpb->libuf_alloc)
 737     {
 738         ms->mpb->libuf_alloc = nr;
 739         srenew(ms->mpb->libuf, ms->mpb->libuf_alloc);
 740     }
 741     MPI_Allreduce(r, ms->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
 742                   ms->mpi_comm_masters);
 743     for (i = 0; i < nr; i++)
 744     {
 745         r[i] = ms->mpb->libuf[i];
 746     }
 747 #endif
 748 #endif
 749 }
 750
 751 const char *opt2fn_master(const char *opt, int nfile, const t_filenm fnm[],
 752                           t_commrec *cr)
 753 {
 754     return SIMMASTER(cr) ? opt2fn(opt, nfile, fnm) : nullptr;
 755 }
 756
 757 void gmx_fatal_collective(int f_errno, const char *file, int line,
 758                           MPI_Comm comm, gmx_bool bMaster,
 759                           const char *fmt, ...)
 760 {
 761     va_list  ap;
 762     gmx_bool bFinalize;
 763 #if GMX_MPI
 764     int      result;
 765     /* Check if we are calling on all processes in MPI_COMM_WORLD */
 766     MPI_Comm_compare(comm, MPI_COMM_WORLD, &result);
 767     /* Any result except MPI_UNEQUAL allows us to call MPI_Finalize */
 768     bFinalize = (result != MPI_UNEQUAL);
 769 #else
 770     GMX_UNUSED_VALUE(comm);
 771     bFinalize = TRUE;
 772 #endif
 773
 774     va_start(ap, fmt);
 775     gmx_fatal_mpi_va(f_errno, file, line, bMaster, bFinalize, fmt, ap);
 776     va_end(ap);
 777 }