src/gromacs/domdec/domdec.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2005,2006,2007,2008,2009 by the GROMACS development team.
   5  * Copyright (c) 2010,2011,2012,2013,2014 by the GROMACS development team.
   6  * Copyright (c) 2015,2016,2017,2018,2019 by the GROMACS development team.
   7  * Copyright (c) 2020,2021, by the GROMACS development team, led by
   8  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   9  * and including many others, as listed in the AUTHORS file in the
  10  * top-level source directory and at http://www.gromacs.org.
  11  *
  12  * GROMACS is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public License
  14  * as published by the Free Software Foundation; either version 2.1
  15  * of the License, or (at your option) any later version.
  16  *
  17  * GROMACS is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with GROMACS; if not, see
  24  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  25  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  26  *
  27  * If you want to redistribute modifications to GROMACS, please
  28  * consider that scientific software is very special. Version
  29  * control is crucial - bugs must be traceable. We will be happy to
  30  * consider code for inclusion in the official distribution, but
  31  * derived work must not be called official GROMACS. Details are found
  32  * in the README & COPYING files - if they are missing, get the
  33  * official version at http://www.gromacs.org.
  34  *
  35  * To help us fund GROMACS development, we humbly ask that you cite
  36  * the research papers on the package. Check out http://www.gromacs.org.
  37  */
  38
  39 #include "gmxpre.h"
  40
  41 #include "domdec.h"
  42
  43 #include "config.h"
  44
  45 #include <cassert>
  46 #include <cinttypes>
  47 #include <climits>
  48 #include <cmath>
  49 #include <cstdio>
  50 #include <cstdlib>
  51 #include <cstring>
  52
  53 #include <algorithm>
  54 #include <memory>
  55
  56 #include "gromacs/domdec/builder.h"
  57 #include "gromacs/domdec/collect.h"
  58 #include "gromacs/domdec/dlb.h"
  59 #include "gromacs/domdec/dlbtiming.h"
  60 #include "gromacs/domdec/domdec_network.h"
  61 #include "gromacs/domdec/ga2la.h"
  62 #include "gromacs/domdec/gpuhaloexchange.h"
  63 #include "gromacs/domdec/options.h"
  64 #include "gromacs/domdec/partition.h"
  65 #include "gromacs/gmxlib/network.h"
  66 #include "gromacs/gmxlib/nrnb.h"
  67 #include "gromacs/gpu_utils/device_stream_manager.h"
  68 #include "gromacs/gpu_utils/gpu_utils.h"
  69 #include "gromacs/hardware/hw_info.h"
  70 #include "gromacs/math/vec.h"
  71 #include "gromacs/math/vectypes.h"
  72 #include "gromacs/mdlib/calc_verletbuf.h"
  73 #include "gromacs/mdlib/constr.h"
  74 #include "gromacs/mdlib/constraintrange.h"
  75 #include "gromacs/mdlib/updategroups.h"
  76 #include "gromacs/mdlib/vsite.h"
  77 #include "gromacs/mdtypes/commrec.h"
  78 #include "gromacs/mdtypes/forceoutput.h"
  79 #include "gromacs/mdtypes/inputrec.h"
  80 #include "gromacs/mdtypes/mdrunoptions.h"
  81 #include "gromacs/mdtypes/state.h"
  82 #include "gromacs/pbcutil/ishift.h"
  83 #include "gromacs/pbcutil/pbc.h"
  84 #include "gromacs/pulling/pull.h"
  85 #include "gromacs/timing/wallcycle.h"
  86 #include "gromacs/topology/block.h"
  87 #include "gromacs/topology/idef.h"
  88 #include "gromacs/topology/ifunc.h"
  89 #include "gromacs/topology/mtop_lookup.h"
  90 #include "gromacs/topology/mtop_util.h"
  91 #include "gromacs/topology/topology.h"
  92 #include "gromacs/utility/basedefinitions.h"
  93 #include "gromacs/utility/basenetwork.h"
  94 #include "gromacs/utility/cstringutil.h"
  95 #include "gromacs/utility/exceptions.h"
  96 #include "gromacs/utility/fatalerror.h"
  97 #include "gromacs/utility/gmxmpi.h"
  98 #include "gromacs/utility/logger.h"
  99 #include "gromacs/utility/real.h"
 100 #include "gromacs/utility/smalloc.h"
 101 #include "gromacs/utility/strconvert.h"
 102 #include "gromacs/utility/stringstream.h"
 103 #include "gromacs/utility/stringutil.h"
 104 #include "gromacs/utility/textwriter.h"
 105
 106 #include "atomdistribution.h"
 107 #include "box.h"
 108 #include "cellsizes.h"
 109 #include "distribute.h"
 110 #include "domdec_constraints.h"
 111 #include "domdec_internal.h"
 112 #include "domdec_setup.h"
 113 #include "domdec_vsite.h"
 114 #include "redistribute.h"
 115 #include "utility.h"
 116
 117 // TODO remove this when moving domdec into gmx namespace
 118 using gmx::DdRankOrder;
 119 using gmx::DlbOption;
 120 using gmx::DomdecOptions;
 121
 122 static const char* edlbs_names[int(DlbState::nr)] = { "off", "off", "auto", "locked", "on", "on" };
 123
 124 /* The size per atom group of the cggl_flag buffer in gmx_domdec_comm_t */
 125 #define DD_CGIBS 2
 126
 127 /* The flags for the cggl_flag buffer in gmx_domdec_comm_t */
 128 #define DD_FLAG_NRCG 65535
 129 #define DD_FLAG_FW(d) (1 << (16 + (d)*2))
 130 #define DD_FLAG_BW(d) (1 << (16 + (d)*2 + 1))
 131
 132 /* The DD zone order */
 133 static const ivec dd_zo[DD_MAXZONE] = { { 0, 0, 0 }, { 1, 0, 0 }, { 1, 1, 0 }, { 0, 1, 0 },
 134                                         { 0, 1, 1 }, { 0, 0, 1 }, { 1, 0, 1 }, { 1, 1, 1 } };
 135
 136 /* The non-bonded zone-pair setup for domain decomposition
 137  * The first number is the i-zone, the second number the first j-zone seen by
 138  * this i-zone, the third number the last+1 j-zone seen by this i-zone.
 139  * As is, this is for 3D decomposition, where there are 4 i-zones.
 140  * With 2D decomposition use only the first 2 i-zones and a last+1 j-zone of 4.
 141  * With 1D decomposition use only the first i-zone and a last+1 j-zone of 2.
 142  */
 143 static const int ddNonbondedZonePairRanges[DD_MAXIZONE][3] = { { 0, 0, 8 },
 144                                                                { 1, 3, 6 },
 145                                                                { 2, 5, 6 },
 146                                                                { 3, 5, 7 } };
 147
 148 static void ddindex2xyz(const ivec nc, int ind, ivec xyz)
 149 {
 150     xyz[XX] = ind / (nc[YY] * nc[ZZ]);
 151     xyz[YY] = (ind / nc[ZZ]) % nc[YY];
 152     xyz[ZZ] = ind % nc[ZZ];
 153 }
 154
 155 static int ddcoord2ddnodeid(gmx_domdec_t* dd, ivec c)
 156 {
 157     int ddnodeid = -1;
 158
 159     const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
 160     const int                 ddindex   = dd_index(dd->numCells, c);
 161     if (cartSetup.bCartesianPP_PME)
 162     {
 163         ddnodeid = cartSetup.ddindex2ddnodeid[ddindex];
 164     }
 165     else if (cartSetup.bCartesianPP)
 166     {
 167 #if GMX_MPI
 168         MPI_Cart_rank(dd->mpi_comm_all, c, &ddnodeid);
 169 #endif
 170     }
 171     else
 172     {
 173         ddnodeid = ddindex;
 174     }
 175
 176     return ddnodeid;
 177 }
 178
 179 int ddglatnr(const gmx_domdec_t* dd, int i)
 180 {
 181     int atnr = 0;
 182
 183     if (dd == nullptr)
 184     {
 185         atnr = i + 1;
 186     }
 187     else
 188     {
 189         if (i >= dd->comm->atomRanges.numAtomsTotal())
 190         {
 191             gmx_fatal(FARGS,
 192                       "glatnr called with %d, which is larger than the local number of atoms (%d)",
 193                       i,
 194                       dd->comm->atomRanges.numAtomsTotal());
 195         }
 196         atnr = dd->globalAtomIndices[i] + 1;
 197     }
 198
 199     return atnr;
 200 }
 201
 202 gmx::ArrayRef<const gmx::RangePartitioning> getUpdateGroupingPerMoleculetype(const gmx_domdec_t& dd)
 203 {
 204     GMX_RELEASE_ASSERT(dd.comm, "Need a valid dd.comm");
 205     return dd.comm->systemInfo.updateGroupingPerMoleculetype;
 206 }
 207
 208 void dd_store_state(const gmx_domdec_t& dd, t_state* state)
 209 {
 210     if (state->ddp_count != dd.ddp_count)
 211     {
 212         gmx_incons("The MD state does not match the domain decomposition state");
 213     }
 214
 215     state->cg_gl.resize(dd.ncg_home);
 216     for (int i = 0; i < dd.ncg_home; i++)
 217     {
 218         state->cg_gl[i] = dd.globalAtomGroupIndices[i];
 219     }
 220
 221     state->ddp_count_cg_gl = dd.ddp_count;
 222 }
 223
 224 gmx_domdec_zones_t* domdec_zones(gmx_domdec_t* dd)
 225 {
 226     return &dd->comm->zones;
 227 }
 228
 229 int dd_numAtomsZones(const gmx_domdec_t& dd)
 230 {
 231     return dd.comm->atomRanges.end(DDAtomRanges::Type::Zones);
 232 }
 233
 234 int dd_numHomeAtoms(const gmx_domdec_t& dd)
 235 {
 236     return dd.comm->atomRanges.numHomeAtoms();
 237 }
 238
 239 int dd_natoms_mdatoms(const gmx_domdec_t& dd)
 240 {
 241     /* We currently set mdatoms entries for all atoms:
 242      * local + non-local + communicated for vsite + constraints
 243      */
 244
 245     return dd.comm->atomRanges.numAtomsTotal();
 246 }
 247
 248 int dd_natoms_vsite(const gmx_domdec_t& dd)
 249 {
 250     return dd.comm->atomRanges.end(DDAtomRanges::Type::Vsites);
 251 }
 252
 253 void dd_get_constraint_range(const gmx_domdec_t& dd, int* at_start, int* at_end)
 254 {
 255     *at_start = dd.comm->atomRanges.start(DDAtomRanges::Type::Constraints);
 256     *at_end   = dd.comm->atomRanges.end(DDAtomRanges::Type::Constraints);
 257 }
 258
 259 void dd_move_x(gmx_domdec_t* dd, const matrix box, gmx::ArrayRef<gmx::RVec> x, gmx_wallcycle* wcycle)
 260 {
 261     wallcycle_start(wcycle, WallCycleCounter::MoveX);
 262
 263     rvec shift = { 0, 0, 0 };
 264
 265     gmx_domdec_comm_t* comm = dd->comm;
 266
 267     int nzone   = 1;
 268     int nat_tot = comm->atomRanges.numHomeAtoms();
 269     for (int d = 0; d < dd->ndim; d++)
 270     {
 271         const bool bPBC   = (dd->ci[dd->dim[d]] == 0);
 272         const bool bScrew = (bPBC && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
 273         if (bPBC)
 274         {
 275             copy_rvec(box[dd->dim[d]], shift);
 276         }
 277         gmx_domdec_comm_dim_t* cd = &comm->cd[d];
 278         for (const gmx_domdec_ind_t& ind : cd->ind)
 279         {
 280             DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
 281             gmx::ArrayRef<gmx::RVec>& sendBuffer = sendBufferAccess.buffer;
 282             int                       n          = 0;
 283             if (!bPBC)
 284             {
 285                 for (int j : ind.index)
 286                 {
 287                     sendBuffer[n] = x[j];
 288                     n++;
 289                 }
 290             }
 291             else if (!bScrew)
 292             {
 293                 for (int j : ind.index)
 294                 {
 295                     /* We need to shift the coordinates */
 296                     for (int d = 0; d < DIM; d++)
 297                     {
 298                         sendBuffer[n][d] = x[j][d] + shift[d];
 299                     }
 300                     n++;
 301                 }
 302             }
 303             else
 304             {
 305                 for (int j : ind.index)
 306                 {
 307                     /* Shift x */
 308                     sendBuffer[n][XX] = x[j][XX] + shift[XX];
 309                     /* Rotate y and z.
 310                      * This operation requires a special shift force
 311                      * treatment, which is performed in calc_vir.
 312                      */
 313                     sendBuffer[n][YY] = box[YY][YY] - x[j][YY];
 314                     sendBuffer[n][ZZ] = box[ZZ][ZZ] - x[j][ZZ];
 315                     n++;
 316                 }
 317             }
 318
 319             DDBufferAccess<gmx::RVec> receiveBufferAccess(
 320                     comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
 321
 322             gmx::ArrayRef<gmx::RVec> receiveBuffer;
 323             if (cd->receiveInPlace)
 324             {
 325                 receiveBuffer = gmx::arrayRefFromArray(x.data() + nat_tot, ind.nrecv[nzone + 1]);
 326             }
 327             else
 328             {
 329                 receiveBuffer = receiveBufferAccess.buffer;
 330             }
 331             /* Send and receive the coordinates */
 332             ddSendrecv(dd, d, dddirBackward, sendBuffer, receiveBuffer);
 333
 334             if (!cd->receiveInPlace)
 335             {
 336                 int j = 0;
 337                 for (int zone = 0; zone < nzone; zone++)
 338                 {
 339                     for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
 340                     {
 341                         x[i] = receiveBuffer[j++];
 342                     }
 343                 }
 344             }
 345             nat_tot += ind.nrecv[nzone + 1];
 346         }
 347         nzone += nzone;
 348     }
 349
 350     wallcycle_stop(wcycle, WallCycleCounter::MoveX);
 351 }
 352
 353 void dd_move_f(gmx_domdec_t* dd, gmx::ForceWithShiftForces* forceWithShiftForces, gmx_wallcycle* wcycle)
 354 {
 355     wallcycle_start(wcycle, WallCycleCounter::MoveF);
 356
 357     gmx::ArrayRef<gmx::RVec> f      = forceWithShiftForces->force();
 358     gmx::ArrayRef<gmx::RVec> fshift = forceWithShiftForces->shiftForces();
 359
 360     gmx_domdec_comm_t& comm    = *dd->comm;
 361     int                nzone   = comm.zones.n / 2;
 362     int                nat_tot = comm.atomRanges.end(DDAtomRanges::Type::Zones);
 363     for (int d = dd->ndim - 1; d >= 0; d--)
 364     {
 365         /* Only forces in domains near the PBC boundaries need to
 366            consider PBC in the treatment of fshift */
 367         const bool shiftForcesNeedPbc =
 368                 (forceWithShiftForces->computeVirial() && dd->ci[dd->dim[d]] == 0);
 369         const bool applyScrewPbc =
 370                 (shiftForcesNeedPbc && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
 371         /* Determine which shift vector we need */
 372         ivec vis        = { 0, 0, 0 };
 373         vis[dd->dim[d]] = 1;
 374         const int is    = gmx::ivecToShiftIndex(vis);
 375
 376         /* Loop over the pulses */
 377         const gmx_domdec_comm_dim_t& cd = comm.cd[d];
 378         for (int p = cd.numPulses() - 1; p >= 0; p--)
 379         {
 380             const gmx_domdec_ind_t&   ind = cd.ind[p];
 381             DDBufferAccess<gmx::RVec> receiveBufferAccess(comm.rvecBuffer, ind.nsend[nzone + 1]);
 382             gmx::ArrayRef<gmx::RVec>& receiveBuffer = receiveBufferAccess.buffer;
 383
 384             nat_tot -= ind.nrecv[nzone + 1];
 385
 386             DDBufferAccess<gmx::RVec> sendBufferAccess(
 387                     comm.rvecBuffer2, cd.receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
 388
 389             gmx::ArrayRef<gmx::RVec> sendBuffer;
 390             if (cd.receiveInPlace)
 391             {
 392                 sendBuffer = gmx::arrayRefFromArray(f.data() + nat_tot, ind.nrecv[nzone + 1]);
 393             }
 394             else
 395             {
 396                 sendBuffer = sendBufferAccess.buffer;
 397                 int j      = 0;
 398                 for (int zone = 0; zone < nzone; zone++)
 399                 {
 400                     for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
 401                     {
 402                         sendBuffer[j++] = f[i];
 403                     }
 404                 }
 405             }
 406             /* Communicate the forces */
 407             ddSendrecv(dd, d, dddirForward, sendBuffer, receiveBuffer);
 408             /* Add the received forces */
 409             int n = 0;
 410             if (!shiftForcesNeedPbc)
 411             {
 412                 for (int j : ind.index)
 413                 {
 414                     for (int d = 0; d < DIM; d++)
 415                     {
 416                         f[j][d] += receiveBuffer[n][d];
 417                     }
 418                     n++;
 419                 }
 420             }
 421             else if (!applyScrewPbc)
 422             {
 423                 for (int j : ind.index)
 424                 {
 425                     for (int d = 0; d < DIM; d++)
 426                     {
 427                         f[j][d] += receiveBuffer[n][d];
 428                     }
 429                     /* Add this force to the shift force */
 430                     for (int d = 0; d < DIM; d++)
 431                     {
 432                         fshift[is][d] += receiveBuffer[n][d];
 433                     }
 434                     n++;
 435                 }
 436             }
 437             else
 438             {
 439                 for (int j : ind.index)
 440                 {
 441                     /* Rotate the force */
 442                     f[j][XX] += receiveBuffer[n][XX];
 443                     f[j][YY] -= receiveBuffer[n][YY];
 444                     f[j][ZZ] -= receiveBuffer[n][ZZ];
 445                     if (shiftForcesNeedPbc)
 446                     {
 447                         /* Add this force to the shift force */
 448                         for (int d = 0; d < DIM; d++)
 449                         {
 450                             fshift[is][d] += receiveBuffer[n][d];
 451                         }
 452                     }
 453                     n++;
 454                 }
 455             }
 456         }
 457         nzone /= 2;
 458     }
 459     wallcycle_stop(wcycle, WallCycleCounter::MoveF);
 460 }
 461
 462 /* Convenience function for extracting a real buffer from an rvec buffer
 463  *
 464  * To reduce the number of temporary communication buffers and avoid
 465  * cache polution, we reuse gmx::RVec buffers for storing reals.
 466  * This functions return a real buffer reference with the same number
 467  * of elements as the gmx::RVec buffer (so 1/3 of the size in bytes).
 468  */
 469 static gmx::ArrayRef<real> realArrayRefFromRvecArrayRef(gmx::ArrayRef<gmx::RVec> arrayRef)
 470 {
 471     return gmx::arrayRefFromArray(as_rvec_array(arrayRef.data())[0], arrayRef.size());
 472 }
 473
 474 void dd_atom_spread_real(gmx_domdec_t* dd, real v[])
 475 {
 476     gmx_domdec_comm_t* comm = dd->comm;
 477
 478     int nzone   = 1;
 479     int nat_tot = comm->atomRanges.numHomeAtoms();
 480     for (int d = 0; d < dd->ndim; d++)
 481     {
 482         gmx_domdec_comm_dim_t* cd = &comm->cd[d];
 483         for (const gmx_domdec_ind_t& ind : cd->ind)
 484         {
 485             /* Note: We provision for RVec instead of real, so a factor of 3
 486              * more than needed. The buffer actually already has this size
 487              * and we pass a plain pointer below, so this does not matter.
 488              */
 489             DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
 490             gmx::ArrayRef<real> sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
 491             int                 n          = 0;
 492             for (int j : ind.index)
 493             {
 494                 sendBuffer[n++] = v[j];
 495             }
 496
 497             DDBufferAccess<gmx::RVec> receiveBufferAccess(
 498                     comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
 499
 500             gmx::ArrayRef<real> receiveBuffer;
 501             if (cd->receiveInPlace)
 502             {
 503                 receiveBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
 504             }
 505             else
 506             {
 507                 receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
 508             }
 509             /* Send and receive the data */
 510             ddSendrecv(dd, d, dddirBackward, sendBuffer, receiveBuffer);
 511             if (!cd->receiveInPlace)
 512             {
 513                 int j = 0;
 514                 for (int zone = 0; zone < nzone; zone++)
 515                 {
 516                     for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
 517                     {
 518                         v[i] = receiveBuffer[j++];
 519                     }
 520                 }
 521             }
 522             nat_tot += ind.nrecv[nzone + 1];
 523         }
 524         nzone += nzone;
 525     }
 526 }
 527
 528 void dd_atom_sum_real(gmx_domdec_t* dd, real v[])
 529 {
 530     gmx_domdec_comm_t* comm = dd->comm;
 531
 532     int nzone   = comm->zones.n / 2;
 533     int nat_tot = comm->atomRanges.end(DDAtomRanges::Type::Zones);
 534     for (int d = dd->ndim - 1; d >= 0; d--)
 535     {
 536         gmx_domdec_comm_dim_t* cd = &comm->cd[d];
 537         for (int p = cd->numPulses() - 1; p >= 0; p--)
 538         {
 539             const gmx_domdec_ind_t& ind = cd->ind[p];
 540
 541             /* Note: We provision for RVec instead of real, so a factor of 3
 542              * more than needed. The buffer actually already has this size
 543              * and we typecast, so this works as intended.
 544              */
 545             DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
 546             gmx::ArrayRef<real> receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
 547             nat_tot -= ind.nrecv[nzone + 1];
 548
 549             DDBufferAccess<gmx::RVec> sendBufferAccess(
 550                     comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
 551
 552             gmx::ArrayRef<real> sendBuffer;
 553             if (cd->receiveInPlace)
 554             {
 555                 sendBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
 556             }
 557             else
 558             {
 559                 sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
 560                 int j      = 0;
 561                 for (int zone = 0; zone < nzone; zone++)
 562                 {
 563                     for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
 564                     {
 565                         sendBuffer[j++] = v[i];
 566                     }
 567                 }
 568             }
 569             /* Communicate the forces */
 570             ddSendrecv(dd, d, dddirForward, sendBuffer, receiveBuffer);
 571             /* Add the received forces */
 572             int n = 0;
 573             for (int j : ind.index)
 574             {
 575                 v[j] += receiveBuffer[n];
 576                 n++;
 577             }
 578         }
 579         nzone /= 2;
 580     }
 581 }
 582
 583 real dd_cutoff_multibody(const gmx_domdec_t* dd)
 584 {
 585     const gmx_domdec_comm_t& comm       = *dd->comm;
 586     const DDSystemInfo&      systemInfo = comm.systemInfo;
 587
 588     real r = -1;
 589     if (systemInfo.haveInterDomainMultiBodyBondeds)
 590     {
 591         if (comm.cutoff_mbody > 0)
 592         {
 593             r = comm.cutoff_mbody;
 594         }
 595         else
 596         {
 597             /* cutoff_mbody=0 means we do not have DLB */
 598             r = comm.cellsize_min[dd->dim[0]];
 599             for (int di = 1; di < dd->ndim; di++)
 600             {
 601                 r = std::min(r, comm.cellsize_min[dd->dim[di]]);
 602             }
 603             if (comm.systemInfo.filterBondedCommunication)
 604             {
 605                 r = std::max(r, comm.cutoff_mbody);
 606             }
 607             else
 608             {
 609                 r = std::min(r, systemInfo.cutoff);
 610             }
 611         }
 612     }
 613
 614     return r;
 615 }
 616
 617 real dd_cutoff_twobody(const gmx_domdec_t* dd)
 618 {
 619     const real r_mb = dd_cutoff_multibody(dd);
 620
 621     return std::max(dd->comm->systemInfo.cutoff, r_mb);
 622 }
 623
 624
 625 static void dd_cart_coord2pmecoord(const DDRankSetup&        ddRankSetup,
 626                                    const CartesianRankSetup& cartSetup,
 627                                    const ivec                coord,
 628                                    ivec                      coord_pme)
 629 {
 630     const int nc   = ddRankSetup.numPPCells[cartSetup.cartpmedim];
 631     const int ntot = cartSetup.ntot[cartSetup.cartpmedim];
 632     copy_ivec(coord, coord_pme);
 633     coord_pme[cartSetup.cartpmedim] =
 634             nc + (coord[cartSetup.cartpmedim] * (ntot - nc) + (ntot - nc) / 2) / nc;
 635 }
 636
 637 /* Returns the PME rank index in 0...npmenodes-1 for the PP cell with index ddCellIndex */
 638 static int ddindex2pmeindex(const DDRankSetup& ddRankSetup, const int ddCellIndex)
 639 {
 640     const int npp  = ddRankSetup.numPPRanks;
 641     const int npme = ddRankSetup.numRanksDoingPme;
 642
 643     /* Here we assign a PME node to communicate with this DD node
 644      * by assuming that the major index of both is x.
 645      * We add npme/2 to obtain an even distribution.
 646      */
 647     return (ddCellIndex * npme + npme / 2) / npp;
 648 }
 649
 650 static std::vector<int> dd_interleaved_pme_ranks(const DDRankSetup& ddRankSetup)
 651 {
 652     std::vector<int> pmeRanks(ddRankSetup.numRanksDoingPme);
 653
 654     int n = 0;
 655     for (int i = 0; i < ddRankSetup.numPPRanks; i++)
 656     {
 657         const int p0 = ddindex2pmeindex(ddRankSetup, i);
 658         const int p1 = ddindex2pmeindex(ddRankSetup, i + 1);
 659         if (i + 1 == ddRankSetup.numPPRanks || p1 > p0)
 660         {
 661             if (debug)
 662             {
 663                 fprintf(debug, "pme_rank[%d] = %d\n", n, i + 1 + n);
 664             }
 665             pmeRanks[n] = i + 1 + n;
 666             n++;
 667         }
 668     }
 669
 670     return pmeRanks;
 671 }
 672
 673 static int gmx_ddcoord2pmeindex(const gmx_domdec_t& dd, int x, int y, int z)
 674 {
 675     ivec coords;
 676
 677     coords[XX]     = x;
 678     coords[YY]     = y;
 679     coords[ZZ]     = z;
 680     const int slab = ddindex2pmeindex(dd.comm->ddRankSetup, dd_index(dd.numCells, coords));
 681
 682     return slab;
 683 }
 684
 685 static int ddcoord2simnodeid(const t_commrec* cr, int x, int y, int z)
 686 {
 687     const CartesianRankSetup& cartSetup = cr->dd->comm->cartesianRankSetup;
 688     ivec                      coords    = { x, y, z };
 689     int                       nodeid    = -1;
 690
 691     if (cartSetup.bCartesianPP_PME)
 692     {
 693 #if GMX_MPI
 694         MPI_Cart_rank(cr->mpi_comm_mysim, coords, &nodeid);
 695 #endif
 696     }
 697     else
 698     {
 699         int ddindex = dd_index(cr->dd->numCells, coords);
 700         if (cartSetup.bCartesianPP)
 701         {
 702             nodeid = cartSetup.ddindex2simnodeid[ddindex];
 703         }
 704         else
 705         {
 706             if (cr->dd->comm->ddRankSetup.usePmeOnlyRanks)
 707             {
 708                 nodeid = ddindex + gmx_ddcoord2pmeindex(*cr->dd, x, y, z);
 709             }
 710             else
 711             {
 712                 nodeid = ddindex;
 713             }
 714         }
 715     }
 716
 717     return nodeid;
 718 }
 719
 720 static int dd_simnode2pmenode(const DDRankSetup&        ddRankSetup,
 721                               const CartesianRankSetup& cartSetup,
 722                               gmx::ArrayRef<const int>  pmeRanks,
 723                               const t_commrec gmx_unused* cr,
 724                               const int                   sim_nodeid)
 725 {
 726     int pmenode = -1;
 727
 728     /* This assumes a uniform x domain decomposition grid cell size */
 729     if (cartSetup.bCartesianPP_PME)
 730     {
 731 #if GMX_MPI
 732         ivec coord, coord_pme;
 733         MPI_Cart_coords(cr->mpi_comm_mysim, sim_nodeid, DIM, coord);
 734         if (coord[cartSetup.cartpmedim] < ddRankSetup.numPPCells[cartSetup.cartpmedim])
 735         {
 736             /* This is a PP rank */
 737             dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
 738             MPI_Cart_rank(cr->mpi_comm_mysim, coord_pme, &pmenode);
 739         }
 740 #endif
 741     }
 742     else if (cartSetup.bCartesianPP)
 743     {
 744         if (sim_nodeid < ddRankSetup.numPPRanks)
 745         {
 746             pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
 747         }
 748     }
 749     else
 750     {
 751         /* This assumes DD cells with identical x coordinates
 752          * are numbered sequentially.
 753          */
 754         if (pmeRanks.empty())
 755         {
 756             if (sim_nodeid < ddRankSetup.numPPRanks)
 757             {
 758                 /* The DD index equals the nodeid */
 759                 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
 760             }
 761         }
 762         else
 763         {
 764             int i = 0;
 765             while (sim_nodeid > pmeRanks[i])
 766             {
 767                 i++;
 768             }
 769             if (sim_nodeid < pmeRanks[i])
 770             {
 771                 pmenode = pmeRanks[i];
 772             }
 773         }
 774     }
 775
 776     return pmenode;
 777 }
 778
 779 NumPmeDomains getNumPmeDomains(const gmx_domdec_t* dd)
 780 {
 781     if (dd != nullptr)
 782     {
 783         return { dd->comm->ddRankSetup.npmenodes_x, dd->comm->ddRankSetup.npmenodes_y };
 784     }
 785     else
 786     {
 787         return { 1, 1 };
 788     }
 789 }
 790
 791 std::vector<int> get_pme_ddranks(const t_commrec* cr, const int pmenodeid)
 792 {
 793     const DDRankSetup&        ddRankSetup = cr->dd->comm->ddRankSetup;
 794     const CartesianRankSetup& cartSetup   = cr->dd->comm->cartesianRankSetup;
 795     GMX_RELEASE_ASSERT(ddRankSetup.usePmeOnlyRanks,
 796                        "This function should only be called when PME-only ranks are in use");
 797     std::vector<int> ddranks;
 798     ddranks.reserve((ddRankSetup.numPPRanks + ddRankSetup.numRanksDoingPme - 1) / ddRankSetup.numRanksDoingPme);
 799
 800     for (int x = 0; x < ddRankSetup.numPPCells[XX]; x++)
 801     {
 802         for (int y = 0; y < ddRankSetup.numPPCells[YY]; y++)
 803         {
 804             for (int z = 0; z < ddRankSetup.numPPCells[ZZ]; z++)
 805             {
 806                 if (cartSetup.bCartesianPP_PME)
 807                 {
 808                     ivec coord = { x, y, z };
 809                     ivec coord_pme;
 810                     dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
 811                     if (cr->dd->ci[XX] == coord_pme[XX] && cr->dd->ci[YY] == coord_pme[YY]
 812                         && cr->dd->ci[ZZ] == coord_pme[ZZ])
 813                     {
 814                         ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
 815                     }
 816                 }
 817                 else
 818                 {
 819                     /* The slab corresponds to the nodeid in the PME group */
 820                     if (gmx_ddcoord2pmeindex(*cr->dd, x, y, z) == pmenodeid)
 821                     {
 822                         ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
 823                     }
 824                 }
 825             }
 826         }
 827     }
 828     return ddranks;
 829 }
 830
 831 static gmx_bool receive_vir_ener(const gmx_domdec_t* dd, gmx::ArrayRef<const int> pmeRanks, const t_commrec* cr)
 832 {
 833     bool bReceive = true;
 834
 835     const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
 836     if (ddRankSetup.usePmeOnlyRanks)
 837     {
 838         const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
 839         if (cartSetup.bCartesianPP_PME)
 840         {
 841 #if GMX_MPI
 842             int  pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
 843             ivec coords;
 844             MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, coords);
 845             coords[cartSetup.cartpmedim]++;
 846             if (coords[cartSetup.cartpmedim] < dd->numCells[cartSetup.cartpmedim])
 847             {
 848                 int rank = 0;
 849                 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &rank);
 850                 if (dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, rank) == pmenode)
 851                 {
 852                     /* This is not the last PP node for pmenode */
 853                     bReceive = FALSE;
 854                 }
 855             }
 856 #else
 857             GMX_RELEASE_ASSERT(
 858                     false,
 859                     "Without MPI we should not have Cartesian PP-PME with #PMEnodes < #DDnodes");
 860 #endif
 861         }
 862         else
 863         {
 864             int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
 865             if (cr->sim_nodeid + 1 < cr->nnodes
 866                 && dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid + 1) == pmenode)
 867             {
 868                 /* This is not the last PP node for pmenode */
 869                 bReceive = FALSE;
 870             }
 871         }
 872     }
 873
 874     return bReceive;
 875 }
 876
 877 static void set_slb_pme_dim_f(gmx_domdec_t* dd, int dim, real** dim_f)
 878 {
 879     gmx_domdec_comm_t* comm = dd->comm;
 880
 881     snew(*dim_f, dd->numCells[dim] + 1);
 882     (*dim_f)[0] = 0;
 883     for (int i = 1; i < dd->numCells[dim]; i++)
 884     {
 885         if (comm->slb_frac[dim])
 886         {
 887             (*dim_f)[i] = (*dim_f)[i - 1] + comm->slb_frac[dim][i - 1];
 888         }
 889         else
 890         {
 891             (*dim_f)[i] = static_cast<real>(i) / static_cast<real>(dd->numCells[dim]);
 892         }
 893     }
 894     (*dim_f)[dd->numCells[dim]] = 1;
 895 }
 896
 897 static void init_ddpme(gmx_domdec_t* dd, gmx_ddpme_t* ddpme, int dimind)
 898 {
 899     const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
 900
 901     if (dimind == 0 && dd->dim[0] == YY && ddRankSetup.npmenodes_x == 1)
 902     {
 903         ddpme->dim = YY;
 904     }
 905     else
 906     {
 907         ddpme->dim = dimind;
 908     }
 909     ddpme->dim_match = (ddpme->dim == dd->dim[dimind]);
 910
 911     ddpme->nslab = (ddpme->dim == 0 ? ddRankSetup.npmenodes_x : ddRankSetup.npmenodes_y);
 912
 913     if (ddpme->nslab <= 1)
 914     {
 915         return;
 916     }
 917
 918     const int nso = ddRankSetup.numRanksDoingPme / ddpme->nslab;
 919     /* Determine for each PME slab the PP location range for dimension dim */
 920     snew(ddpme->pp_min, ddpme->nslab);
 921     snew(ddpme->pp_max, ddpme->nslab);
 922     for (int slab = 0; slab < ddpme->nslab; slab++)
 923     {
 924         ddpme->pp_min[slab] = dd->numCells[dd->dim[dimind]] - 1;
 925         ddpme->pp_max[slab] = 0;
 926     }
 927     for (int i = 0; i < dd->nnodes; i++)
 928     {
 929         ivec xyz;
 930         ddindex2xyz(dd->numCells, i, xyz);
 931         /* For y only use our y/z slab.
 932          * This assumes that the PME x grid size matches the DD grid size.
 933          */
 934         if (dimind == 0 || xyz[XX] == dd->ci[XX])
 935         {
 936             const int pmeindex  = ddindex2pmeindex(ddRankSetup, i);
 937             const int slab      = (dimind == 0) ? (pmeindex / nso) : (pmeindex % ddpme->nslab);
 938             ddpme->pp_min[slab] = std::min(ddpme->pp_min[slab], xyz[dimind]);
 939             ddpme->pp_max[slab] = std::max(ddpme->pp_max[slab], xyz[dimind]);
 940         }
 941     }
 942
 943     set_slb_pme_dim_f(dd, ddpme->dim, &ddpme->slb_dim_f);
 944 }
 945
 946 int dd_pme_maxshift_x(const gmx_domdec_t& dd)
 947 {
 948     const DDRankSetup& ddRankSetup = dd.comm->ddRankSetup;
 949
 950     if (ddRankSetup.ddpme[0].dim == XX)
 951     {
 952         return ddRankSetup.ddpme[0].maxshift;
 953     }
 954     else
 955     {
 956         return 0;
 957     }
 958 }
 959
 960 int dd_pme_maxshift_y(const gmx_domdec_t& dd)
 961 {
 962     const DDRankSetup& ddRankSetup = dd.comm->ddRankSetup;
 963
 964     if (ddRankSetup.ddpme[0].dim == YY)
 965     {
 966         return ddRankSetup.ddpme[0].maxshift;
 967     }
 968     else if (ddRankSetup.npmedecompdim >= 2 && ddRankSetup.ddpme[1].dim == YY)
 969     {
 970         return ddRankSetup.ddpme[1].maxshift;
 971     }
 972     else
 973     {
 974         return 0;
 975     }
 976 }
 977
 978 bool ddHaveSplitConstraints(const gmx_domdec_t& dd)
 979 {
 980     return dd.comm->systemInfo.haveSplitConstraints;
 981 }
 982
 983 bool ddUsesUpdateGroups(const gmx_domdec_t& dd)
 984 {
 985     return dd.comm->systemInfo.useUpdateGroups;
 986 }
 987
 988 void dd_cycles_add(const gmx_domdec_t* dd, float cycles, int ddCycl)
 989 {
 990     /* Note that the cycles value can be incorrect, either 0 or some
 991      * extremely large value, when our thread migrated to another core
 992      * with an unsynchronized cycle counter. If this happens less often
 993      * that once per nstlist steps, this will not cause issues, since
 994      * we later subtract the maximum value from the sum over nstlist steps.
 995      * A zero count will slightly lower the total, but that's a small effect.
 996      * Note that the main purpose of the subtraction of the maximum value
 997      * is to avoid throwing off the load balancing when stalls occur due
 998      * e.g. system activity or network congestion.
 999      */
1000     dd->comm->cycl[ddCycl] += cycles;
1001     dd->comm->cycl_n[ddCycl]++;
1002     if (cycles > dd->comm->cycl_max[ddCycl])
1003     {
1004         dd->comm->cycl_max[ddCycl] = cycles;
1005     }
1006 }
1007
1008 #if GMX_MPI
1009 static void make_load_communicator(gmx_domdec_t* dd, int dim_ind, ivec loc)
1010 {
1011     MPI_Comm c_row = MPI_COMM_NULL;
1012     ivec     loc_c;
1013     bool     bPartOfGroup = false;
1014
1015     const int dim = dd->dim[dim_ind];
1016     copy_ivec(loc, loc_c);
1017     for (int i = 0; i < dd->numCells[dim]; i++)
1018     {
1019         loc_c[dim]     = i;
1020         const int rank = dd_index(dd->numCells, loc_c);
1021         if (rank == dd->rank)
1022         {
1023             /* This process is part of the group */
1024             bPartOfGroup = TRUE;
1025         }
1026     }
1027     MPI_Comm_split(dd->mpi_comm_all, bPartOfGroup ? 0 : MPI_UNDEFINED, dd->rank, &c_row);
1028     if (bPartOfGroup)
1029     {
1030         dd->comm->mpi_comm_load[dim_ind] = c_row;
1031         if (!isDlbDisabled(dd->comm))
1032         {
1033             DDCellsizesWithDlb& cellsizes = dd->comm->cellsizesWithDlb[dim_ind];
1034
1035             if (dd->ci[dim] == dd->master_ci[dim])
1036             {
1037                 /* This is the root process of this row */
1038                 cellsizes.rowMaster = std::make_unique<RowMaster>();
1039
1040                 RowMaster& rowMaster = *cellsizes.rowMaster;
1041                 rowMaster.cellFrac.resize(ddCellFractionBufferSize(dd, dim_ind));
1042                 rowMaster.oldCellFrac.resize(dd->numCells[dim] + 1);
1043                 rowMaster.isCellMin.resize(dd->numCells[dim]);
1044                 if (dim_ind > 0)
1045                 {
1046                     rowMaster.bounds.resize(dd->numCells[dim]);
1047                 }
1048                 rowMaster.buf_ncd.resize(dd->numCells[dim]);
1049             }
1050             else
1051             {
1052                 /* This is not a root process, we only need to receive cell_f */
1053                 cellsizes.fracRow.resize(ddCellFractionBufferSize(dd, dim_ind));
1054             }
1055         }
1056         if (dd->ci[dim] == dd->master_ci[dim])
1057         {
1058             snew(dd->comm->load[dim_ind].load, dd->numCells[dim] * DD_NLOAD_MAX);
1059         }
1060     }
1061 }
1062 #endif
1063
1064 void dd_setup_dlb_resource_sharing(const t_commrec* cr, int gpu_id)
1065 {
1066 #if GMX_MPI
1067     MPI_Comm mpi_comm_pp_physicalnode = MPI_COMM_NULL;
1068
1069     if (!thisRankHasDuty(cr, DUTY_PP) || gpu_id < 0)
1070     {
1071         /* Only ranks with short-ranged tasks (currently) use GPUs.
1072          * If we don't have GPUs assigned, there are no resources to share.
1073          */
1074         return;
1075     }
1076
1077     const int physicalnode_id_hash = gmx_physicalnode_id_hash();
1078
1079     gmx_domdec_t* dd = cr->dd;
1080
1081     if (debug)
1082     {
1083         fprintf(debug, "dd_setup_dd_dlb_gpu_sharing:\n");
1084         fprintf(debug, "DD PP rank %d physical node hash %d gpu_id %d\n", dd->rank, physicalnode_id_hash, gpu_id);
1085     }
1086     /* Split the PP communicator over the physical nodes */
1087     /* TODO: See if we should store this (before), as it's also used for
1088      * for the nodecomm summation.
1089      */
1090     // TODO PhysicalNodeCommunicator could be extended/used to handle
1091     // the need for per-node per-group communicators.
1092     MPI_Comm_split(dd->mpi_comm_all, physicalnode_id_hash, dd->rank, &mpi_comm_pp_physicalnode);
1093     MPI_Comm_split(mpi_comm_pp_physicalnode, gpu_id, dd->rank, &dd->comm->mpi_comm_gpu_shared);
1094     MPI_Comm_free(&mpi_comm_pp_physicalnode);
1095     MPI_Comm_size(dd->comm->mpi_comm_gpu_shared, &dd->comm->nrank_gpu_shared);
1096
1097     if (debug)
1098     {
1099         fprintf(debug, "nrank_gpu_shared %d\n", dd->comm->nrank_gpu_shared);
1100     }
1101
1102     /* Note that some ranks could share a GPU, while others don't */
1103
1104     if (dd->comm->nrank_gpu_shared == 1)
1105     {
1106         MPI_Comm_free(&dd->comm->mpi_comm_gpu_shared);
1107     }
1108 #else
1109     GMX_UNUSED_VALUE(cr);
1110     GMX_UNUSED_VALUE(gpu_id);
1111 #endif
1112 }
1113
1114 static void make_load_communicators(gmx_domdec_t gmx_unused* dd)
1115 {
1116 #if GMX_MPI
1117     ivec loc;
1118
1119     if (debug)
1120     {
1121         fprintf(debug, "Making load communicators\n");
1122     }
1123
1124     dd->comm->load = new domdec_load_t[std::max(dd->ndim, 1)];
1125     snew(dd->comm->mpi_comm_load, std::max(dd->ndim, 1));
1126
1127     if (dd->ndim == 0)
1128     {
1129         return;
1130     }
1131
1132     clear_ivec(loc);
1133     make_load_communicator(dd, 0, loc);
1134     if (dd->ndim > 1)
1135     {
1136         const int dim0 = dd->dim[0];
1137         for (int i = 0; i < dd->numCells[dim0]; i++)
1138         {
1139             loc[dim0] = i;
1140             make_load_communicator(dd, 1, loc);
1141         }
1142     }
1143     if (dd->ndim > 2)
1144     {
1145         const int dim0 = dd->dim[0];
1146         for (int i = 0; i < dd->numCells[dim0]; i++)
1147         {
1148             loc[dim0]      = i;
1149             const int dim1 = dd->dim[1];
1150             for (int j = 0; j < dd->numCells[dim1]; j++)
1151             {
1152                 loc[dim1] = j;
1153                 make_load_communicator(dd, 2, loc);
1154             }
1155         }
1156     }
1157
1158     if (debug)
1159     {
1160         fprintf(debug, "Finished making load communicators\n");
1161     }
1162 #endif
1163 }
1164
1165 /*! \brief Sets up the relation between neighboring domains and zones */
1166 static void setup_neighbor_relations(gmx_domdec_t* dd)
1167 {
1168     ivec tmp, s;
1169     GMX_ASSERT((dd->ndim >= 0) && (dd->ndim <= DIM), "Must have valid number of dimensions for DD");
1170
1171     for (int d = 0; d < dd->ndim; d++)
1172     {
1173         const int dim = dd->dim[d];
1174         copy_ivec(dd->ci, tmp);
1175         tmp[dim]           = (tmp[dim] + 1) % dd->numCells[dim];
1176         dd->neighbor[d][0] = ddcoord2ddnodeid(dd, tmp);
1177         copy_ivec(dd->ci, tmp);
1178         tmp[dim]           = (tmp[dim] - 1 + dd->numCells[dim]) % dd->numCells[dim];
1179         dd->neighbor[d][1] = ddcoord2ddnodeid(dd, tmp);
1180         if (debug)
1181         {
1182             fprintf(debug,
1183                     "DD rank %d neighbor ranks in dir %d are + %d - %d\n",
1184                     dd->rank,
1185                     dim,
1186                     dd->neighbor[d][0],
1187                     dd->neighbor[d][1]);
1188         }
1189     }
1190
1191     int nzone  = (1 << dd->ndim);
1192     int nizone = (1 << std::max(dd->ndim - 1, 0));
1193     assert(nizone >= 1 && nizone <= DD_MAXIZONE);
1194
1195     gmx_domdec_zones_t* zones = &dd->comm->zones;
1196
1197     for (int i = 0; i < nzone; i++)
1198     {
1199         int m = 0;
1200         clear_ivec(zones->shift[i]);
1201         for (int d = 0; d < dd->ndim; d++)
1202         {
1203             zones->shift[i][dd->dim[d]] = dd_zo[i][m++];
1204         }
1205     }
1206
1207     zones->n = nzone;
1208     for (int i = 0; i < nzone; i++)
1209     {
1210         for (int d = 0; d < DIM; d++)
1211         {
1212             s[d] = dd->ci[d] - zones->shift[i][d];
1213             if (s[d] < 0)
1214             {
1215                 s[d] += dd->numCells[d];
1216             }
1217             else if (s[d] >= dd->numCells[d])
1218             {
1219                 s[d] -= dd->numCells[d];
1220             }
1221         }
1222     }
1223     for (int iZoneIndex = 0; iZoneIndex < nizone; iZoneIndex++)
1224     {
1225         GMX_RELEASE_ASSERT(
1226                 ddNonbondedZonePairRanges[iZoneIndex][0] == iZoneIndex,
1227                 "The first element for each ddNonbondedZonePairRanges should match its index");
1228
1229         DDPairInteractionRanges iZone;
1230         iZone.iZoneIndex = iZoneIndex;
1231         /* dd_zp3 is for 3D decomposition, for fewer dimensions use only
1232          * j-zones up to nzone.
1233          */
1234         iZone.jZoneRange = gmx::Range<int>(std::min(ddNonbondedZonePairRanges[iZoneIndex][1], nzone),
1235                                            std::min(ddNonbondedZonePairRanges[iZoneIndex][2], nzone));
1236         for (int dim = 0; dim < DIM; dim++)
1237         {
1238             if (dd->numCells[dim] == 1)
1239             {
1240                 /* All shifts should be allowed */
1241                 iZone.shift0[dim] = -1;
1242                 iZone.shift1[dim] = 1;
1243             }
1244             else
1245             {
1246                 /* Determine the min/max j-zone shift wrt the i-zone */
1247                 iZone.shift0[dim] = 1;
1248                 iZone.shift1[dim] = -1;
1249                 for (int jZone : iZone.jZoneRange)
1250                 {
1251                     int shift_diff = zones->shift[jZone][dim] - zones->shift[iZoneIndex][dim];
1252                     if (shift_diff < iZone.shift0[dim])
1253                     {
1254                         iZone.shift0[dim] = shift_diff;
1255                     }
1256                     if (shift_diff > iZone.shift1[dim])
1257                     {
1258                         iZone.shift1[dim] = shift_diff;
1259                     }
1260                 }
1261             }
1262         }
1263
1264         zones->iZones.push_back(iZone);
1265     }
1266
1267     if (!isDlbDisabled(dd->comm))
1268     {
1269         dd->comm->cellsizesWithDlb.resize(dd->ndim);
1270     }
1271
1272     if (dd->comm->ddSettings.recordLoad)
1273     {
1274         make_load_communicators(dd);
1275     }
1276 }
1277
1278 static void make_pp_communicator(const gmx::MDLogger& mdlog,
1279                                  gmx_domdec_t*        dd,
1280                                  t_commrec gmx_unused* cr,
1281                                  bool gmx_unused reorder)
1282 {
1283 #if GMX_MPI
1284     gmx_domdec_comm_t*  comm      = dd->comm;
1285     CartesianRankSetup& cartSetup = comm->cartesianRankSetup;
1286
1287     if (cartSetup.bCartesianPP)
1288     {
1289         /* Set up cartesian communication for the particle-particle part */
1290         GMX_LOG(mdlog.info)
1291                 .appendTextFormatted("Will use a Cartesian communicator: %d x %d x %d",
1292                                      dd->numCells[XX],
1293                                      dd->numCells[YY],
1294                                      dd->numCells[ZZ]);
1295
1296         ivec periods;
1297         for (int i = 0; i < DIM; i++)
1298         {
1299             periods[i] = TRUE;
1300         }
1301         MPI_Comm comm_cart = MPI_COMM_NULL;
1302         MPI_Cart_create(cr->mpi_comm_mygroup, DIM, dd->numCells, periods, static_cast<int>(reorder), &comm_cart);
1303         /* We overwrite the old communicator with the new cartesian one */
1304         cr->mpi_comm_mygroup = comm_cart;
1305     }
1306
1307     dd->mpi_comm_all = cr->mpi_comm_mygroup;
1308     MPI_Comm_rank(dd->mpi_comm_all, &dd->rank);
1309
1310     if (cartSetup.bCartesianPP_PME)
1311     {
1312         /* Since we want to use the original cartesian setup for sim,
1313          * and not the one after split, we need to make an index.
1314          */
1315         cartSetup.ddindex2ddnodeid.resize(dd->nnodes);
1316         cartSetup.ddindex2ddnodeid[dd_index(dd->numCells, dd->ci)] = dd->rank;
1317         gmx_sumi(dd->nnodes, cartSetup.ddindex2ddnodeid.data(), cr);
1318         /* Get the rank of the DD master,
1319          * above we made sure that the master node is a PP node.
1320          */
1321         int rank = MASTER(cr) ? dd->rank : 0;
1322         MPI_Allreduce(&rank, &dd->masterrank, 1, MPI_INT, MPI_SUM, dd->mpi_comm_all);
1323     }
1324     else if (cartSetup.bCartesianPP)
1325     {
1326         if (!comm->ddRankSetup.usePmeOnlyRanks)
1327         {
1328             /* The PP communicator is also
1329              * the communicator for this simulation
1330              */
1331             cr->mpi_comm_mysim = cr->mpi_comm_mygroup;
1332         }
1333         cr->nodeid = dd->rank;
1334
1335         MPI_Cart_coords(dd->mpi_comm_all, dd->rank, DIM, dd->ci);
1336
1337         /* We need to make an index to go from the coordinates
1338          * to the nodeid of this simulation.
1339          */
1340         cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1341         std::vector<int> buf(dd->nnodes);
1342         if (thisRankHasDuty(cr, DUTY_PP))
1343         {
1344             buf[dd_index(dd->numCells, dd->ci)] = cr->sim_nodeid;
1345         }
1346         /* Communicate the ddindex to simulation nodeid index */
1347         MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM, cr->mpi_comm_mysim);
1348
1349         /* Determine the master coordinates and rank.
1350          * The DD master should be the same node as the master of this sim.
1351          */
1352         for (int i = 0; i < dd->nnodes; i++)
1353         {
1354             if (cartSetup.ddindex2simnodeid[i] == 0)
1355             {
1356                 ddindex2xyz(dd->numCells, i, dd->master_ci);
1357                 MPI_Cart_rank(dd->mpi_comm_all, dd->master_ci, &dd->masterrank);
1358             }
1359         }
1360         if (debug)
1361         {
1362             fprintf(debug, "The master rank is %d\n", dd->masterrank);
1363         }
1364     }
1365     else
1366     {
1367         /* No Cartesian communicators */
1368         /* We use the rank in dd->comm->all as DD index */
1369         ddindex2xyz(dd->numCells, dd->rank, dd->ci);
1370         /* The simulation master nodeid is 0, so the DD master rank is also 0 */
1371         dd->masterrank = 0;
1372         clear_ivec(dd->master_ci);
1373     }
1374 #endif
1375
1376     GMX_LOG(mdlog.info)
1377             .appendTextFormatted("Domain decomposition rank %d, coordinates %d %d %d\n",
1378                                  dd->rank,
1379                                  dd->ci[XX],
1380                                  dd->ci[YY],
1381                                  dd->ci[ZZ]);
1382     if (debug)
1383     {
1384         fprintf(debug,
1385                 "Domain decomposition rank %d, coordinates %d %d %d\n\n",
1386                 dd->rank,
1387                 dd->ci[XX],
1388                 dd->ci[YY],
1389                 dd->ci[ZZ]);
1390     }
1391 }
1392
1393 static void receive_ddindex2simnodeid(gmx_domdec_t* dd, t_commrec* cr)
1394 {
1395 #if GMX_MPI
1396     CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
1397
1398     if (!cartSetup.bCartesianPP_PME && cartSetup.bCartesianPP)
1399     {
1400         cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1401         std::vector<int> buf(dd->nnodes);
1402         if (thisRankHasDuty(cr, DUTY_PP))
1403         {
1404             buf[dd_index(dd->numCells, dd->ci)] = cr->sim_nodeid;
1405         }
1406         /* Communicate the ddindex to simulation nodeid index */
1407         MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM, cr->mpi_comm_mysim);
1408     }
1409 #else
1410     GMX_UNUSED_VALUE(dd);
1411     GMX_UNUSED_VALUE(cr);
1412 #endif
1413 }
1414
1415 static CartesianRankSetup split_communicator(const gmx::MDLogger& mdlog,
1416                                              t_commrec*           cr,
1417                                              const DdRankOrder    ddRankOrder,
1418                                              bool gmx_unused    reorder,
1419                                              const DDRankSetup& ddRankSetup,
1420                                              ivec               ddCellIndex,
1421                                              std::vector<int>*  pmeRanks)
1422 {
1423     CartesianRankSetup cartSetup;
1424
1425     cartSetup.bCartesianPP     = (ddRankOrder == DdRankOrder::cartesian);
1426     cartSetup.bCartesianPP_PME = false;
1427
1428     const ivec& numDDCells = ddRankSetup.numPPCells;
1429     /* Initially we set ntot to the number of PP cells */
1430     copy_ivec(numDDCells, cartSetup.ntot);
1431
1432     if (cartSetup.bCartesianPP)
1433     {
1434         const int numDDCellsTot = ddRankSetup.numPPRanks;
1435         bool      bDiv[DIM];
1436         for (int i = 1; i < DIM; i++)
1437         {
1438             bDiv[i] = ((ddRankSetup.numRanksDoingPme * numDDCells[i]) % numDDCellsTot == 0);
1439         }
1440         if (bDiv[YY] || bDiv[ZZ])
1441         {
1442             cartSetup.bCartesianPP_PME = TRUE;
1443             /* If we have 2D PME decomposition, which is always in x+y,
1444              * we stack the PME only nodes in z.
1445              * Otherwise we choose the direction that provides the thinnest slab
1446              * of PME only nodes as this will have the least effect
1447              * on the PP communication.
1448              * But for the PME communication the opposite might be better.
1449              */
1450             if (bDiv[ZZ] && (ddRankSetup.npmenodes_y > 1 || !bDiv[YY] || numDDCells[YY] > numDDCells[ZZ]))
1451             {
1452                 cartSetup.cartpmedim = ZZ;
1453             }
1454             else
1455             {
1456                 cartSetup.cartpmedim = YY;
1457             }
1458             cartSetup.ntot[cartSetup.cartpmedim] +=
1459                     (ddRankSetup.numRanksDoingPme * numDDCells[cartSetup.cartpmedim]) / numDDCellsTot;
1460         }
1461         else
1462         {
1463             GMX_LOG(mdlog.info)
1464                     .appendTextFormatted(
1465                             "Number of PME-only ranks (%d) is not a multiple of nx*ny (%d*%d) or "
1466                             "nx*nz (%d*%d)",
1467                             ddRankSetup.numRanksDoingPme,
1468                             numDDCells[XX],
1469                             numDDCells[YY],
1470                             numDDCells[XX],
1471                             numDDCells[ZZ]);
1472             GMX_LOG(mdlog.info)
1473                     .appendText("Will not use a Cartesian communicator for PP <-> PME\n");
1474         }
1475     }
1476
1477     if (cartSetup.bCartesianPP_PME)
1478     {
1479 #if GMX_MPI
1480         int  rank = 0;
1481         ivec periods;
1482
1483         GMX_LOG(mdlog.info)
1484                 .appendTextFormatted(
1485                         "Will use a Cartesian communicator for PP <-> PME: %d x %d x %d",
1486                         cartSetup.ntot[XX],
1487                         cartSetup.ntot[YY],
1488                         cartSetup.ntot[ZZ]);
1489
1490         for (int i = 0; i < DIM; i++)
1491         {
1492             periods[i] = TRUE;
1493         }
1494         MPI_Comm comm_cart = MPI_COMM_NULL;
1495         MPI_Cart_create(cr->mpi_comm_mysim, DIM, cartSetup.ntot, periods, static_cast<int>(reorder), &comm_cart);
1496         MPI_Comm_rank(comm_cart, &rank);
1497         if (MASTER(cr) && rank != 0)
1498         {
1499             gmx_fatal(FARGS, "MPI rank 0 was renumbered by MPI_Cart_create, we do not allow this");
1500         }
1501
1502         /* With this assigment we loose the link to the original communicator
1503          * which will usually be MPI_COMM_WORLD, unless have multisim.
1504          */
1505         cr->mpi_comm_mysim = comm_cart;
1506         cr->sim_nodeid     = rank;
1507
1508         MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, ddCellIndex);
1509
1510         GMX_LOG(mdlog.info)
1511                 .appendTextFormatted("Cartesian rank %d, coordinates %d %d %d\n",
1512                                      cr->sim_nodeid,
1513                                      ddCellIndex[XX],
1514                                      ddCellIndex[YY],
1515                                      ddCellIndex[ZZ]);
1516
1517         if (ddCellIndex[cartSetup.cartpmedim] < numDDCells[cartSetup.cartpmedim])
1518         {
1519             cr->duty = DUTY_PP;
1520         }
1521         if (!ddRankSetup.usePmeOnlyRanks
1522             || ddCellIndex[cartSetup.cartpmedim] >= numDDCells[cartSetup.cartpmedim])
1523         {
1524             cr->duty = DUTY_PME;
1525         }
1526
1527         /* Split the sim communicator into PP and PME only nodes */
1528         MPI_Comm_split(cr->mpi_comm_mysim,
1529                        getThisRankDuties(cr),
1530                        dd_index(cartSetup.ntot, ddCellIndex),
1531                        &cr->mpi_comm_mygroup);
1532 #else
1533         GMX_UNUSED_VALUE(ddCellIndex);
1534 #endif
1535     }
1536     else
1537     {
1538         switch (ddRankOrder)
1539         {
1540             case DdRankOrder::pp_pme:
1541                 GMX_LOG(mdlog.info).appendText("Order of the ranks: PP first, PME last");
1542                 break;
1543             case DdRankOrder::interleave:
1544                 /* Interleave the PP-only and PME-only ranks */
1545                 GMX_LOG(mdlog.info).appendText("Interleaving PP and PME ranks");
1546                 *pmeRanks = dd_interleaved_pme_ranks(ddRankSetup);
1547                 break;
1548             case DdRankOrder::cartesian: break;
1549             default: gmx_fatal(FARGS, "Invalid ddRankOrder=%d", static_cast<int>(ddRankOrder));
1550         }
1551
1552         if (dd_simnode2pmenode(ddRankSetup, cartSetup, *pmeRanks, cr, cr->sim_nodeid) == -1)
1553         {
1554             cr->duty = DUTY_PME;
1555         }
1556         else
1557         {
1558             cr->duty = DUTY_PP;
1559         }
1560 #if GMX_MPI
1561         /* Split the sim communicator into PP and PME only nodes */
1562         MPI_Comm_split(cr->mpi_comm_mysim, getThisRankDuties(cr), cr->nodeid, &cr->mpi_comm_mygroup);
1563         MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
1564 #endif
1565     }
1566
1567     GMX_LOG(mdlog.info)
1568             .appendTextFormatted("This rank does only %s work.\n",
1569                                  thisRankHasDuty(cr, DUTY_PP) ? "particle-particle" : "PME-mesh");
1570
1571     return cartSetup;
1572 }
1573
1574 /*! \brief Makes the PP communicator and the PME communicator, when needed
1575  *
1576  * Returns the Cartesian rank setup.
1577  * Sets \p cr->mpi_comm_mygroup
1578  * For PP ranks, sets the DD PP cell index in \p ddCellIndex.
1579  * With separate PME ranks in interleaved order, set the PME ranks in \p pmeRanks.
1580  */
1581 static CartesianRankSetup makeGroupCommunicators(const gmx::MDLogger& mdlog,
1582                                                  const DDSettings&    ddSettings,
1583                                                  const DdRankOrder    ddRankOrder,
1584                                                  const DDRankSetup&   ddRankSetup,
1585                                                  t_commrec*           cr,
1586                                                  ivec                 ddCellIndex,
1587                                                  std::vector<int>*    pmeRanks)
1588 {
1589     CartesianRankSetup cartSetup;
1590
1591     // As a default, both group and sim communicators are equal to the default communicator
1592     cr->mpi_comm_mygroup = cr->mpiDefaultCommunicator;
1593     cr->mpi_comm_mysim   = cr->mpiDefaultCommunicator;
1594     cr->nnodes           = cr->sizeOfDefaultCommunicator;
1595     cr->nodeid           = cr->rankInDefaultCommunicator;
1596     cr->sim_nodeid       = cr->rankInDefaultCommunicator;
1597
1598     if (ddRankSetup.usePmeOnlyRanks)
1599     {
1600         /* Split the communicator into a PP and PME part */
1601         cartSetup = split_communicator(
1602                 mdlog, cr, ddRankOrder, ddSettings.useCartesianReorder, ddRankSetup, ddCellIndex, pmeRanks);
1603     }
1604     else
1605     {
1606         /* All nodes do PP and PME */
1607         /* We do not require separate communicators */
1608         cartSetup.bCartesianPP     = false;
1609         cartSetup.bCartesianPP_PME = false;
1610     }
1611
1612     return cartSetup;
1613 }
1614
1615 /*! \brief For PP ranks, sets or makes the communicator
1616  *
1617  * For PME ranks get the rank id.
1618  * For PP only ranks, sets the PME-only rank.
1619  */
1620 static void setupGroupCommunication(const gmx::MDLogger&     mdlog,
1621                                     const DDSettings&        ddSettings,
1622                                     gmx::ArrayRef<const int> pmeRanks,
1623                                     t_commrec*               cr,
1624                                     const int                numAtomsInSystem,
1625                                     gmx_domdec_t*            dd)
1626 {
1627     const DDRankSetup&        ddRankSetup = dd->comm->ddRankSetup;
1628     const CartesianRankSetup& cartSetup   = dd->comm->cartesianRankSetup;
1629
1630     if (thisRankHasDuty(cr, DUTY_PP))
1631     {
1632         /* Copy or make a new PP communicator */
1633
1634         /* We (possibly) reordered the nodes in split_communicator,
1635          * so it is no longer required in make_pp_communicator.
1636          */
1637         const bool useCartesianReorder = (ddSettings.useCartesianReorder && !cartSetup.bCartesianPP_PME);
1638
1639         make_pp_communicator(mdlog, dd, cr, useCartesianReorder);
1640     }
1641     else
1642     {
1643         receive_ddindex2simnodeid(dd, cr);
1644     }
1645
1646     if (!thisRankHasDuty(cr, DUTY_PME))
1647     {
1648         /* Set up the commnuication to our PME node */
1649         dd->pme_nodeid = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
1650         dd->pme_receive_vir_ener = receive_vir_ener(dd, pmeRanks, cr);
1651         if (debug)
1652         {
1653             fprintf(debug,
1654                     "My pme_nodeid %d receive ener %s\n",
1655                     dd->pme_nodeid,
1656                     gmx::boolToString(dd->pme_receive_vir_ener));
1657         }
1658     }
1659     else
1660     {
1661         dd->pme_nodeid = -1;
1662     }
1663
1664     /* We can not use DDMASTER(dd), because dd->masterrank is set later */
1665     if (MASTER(cr))
1666     {
1667         dd->ma = std::make_unique<AtomDistribution>(dd->numCells, numAtomsInSystem, numAtomsInSystem);
1668     }
1669 }
1670
1671 static real* get_slb_frac(const gmx::MDLogger& mdlog, const char* dir, int nc, const char* size_string)
1672 {
1673     real* slb_frac = nullptr;
1674     if (nc > 1 && size_string != nullptr)
1675     {
1676         GMX_LOG(mdlog.info).appendTextFormatted("Using static load balancing for the %s direction", dir);
1677         snew(slb_frac, nc);
1678         real tot = 0;
1679         for (int i = 0; i < nc; i++)
1680         {
1681             double dbl = 0;
1682             int    n   = 0;
1683             sscanf(size_string, "%20lf%n", &dbl, &n);
1684             if (dbl == 0)
1685             {
1686                 gmx_fatal(FARGS,
1687                           "Incorrect or not enough DD cell size entries for direction %s: '%s'",
1688                           dir,
1689                           size_string);
1690             }
1691             slb_frac[i] = dbl;
1692             size_string += n;
1693             tot += slb_frac[i];
1694         }
1695         /* Normalize */
1696         std::string relativeCellSizes = "Relative cell sizes:";
1697         for (int i = 0; i < nc; i++)
1698         {
1699             slb_frac[i] /= tot;
1700             relativeCellSizes += gmx::formatString(" %5.3f", slb_frac[i]);
1701         }
1702         GMX_LOG(mdlog.info).appendText(relativeCellSizes);
1703     }
1704
1705     return slb_frac;
1706 }
1707
1708 static int multi_body_bondeds_count(const gmx_mtop_t& mtop)
1709 {
1710     int n = 0;
1711     for (const auto ilists : IListRange(mtop))
1712     {
1713         for (auto& ilist : extractILists(ilists.list(), IF_BOND))
1714         {
1715             if (NRAL(ilist.functionType) > 2)
1716             {
1717                 n += ilists.nmol() * (ilist.iatoms.size() / ilistStride(ilist));
1718             }
1719         }
1720     }
1721
1722     return n;
1723 }
1724
1725 static int dd_getenv(const gmx::MDLogger& mdlog, const char* env_var, int def)
1726 {
1727     int   nst = def;
1728     char* val = getenv(env_var);
1729     if (val)
1730     {
1731         if (sscanf(val, "%20d", &nst) <= 0)
1732         {
1733             nst = 1;
1734         }
1735         GMX_LOG(mdlog.info).appendTextFormatted("Found env.var. %s = %s, using value %d", env_var, val, nst);
1736     }
1737
1738     return nst;
1739 }
1740
1741 static void check_dd_restrictions(const gmx_domdec_t* dd, const t_inputrec& inputrec, const gmx::MDLogger& mdlog)
1742 {
1743     if (inputrec.pbcType == PbcType::Screw
1744         && (dd->numCells[XX] == 1 || dd->numCells[YY] > 1 || dd->numCells[ZZ] > 1))
1745     {
1746         gmx_fatal(FARGS,
1747                   "With pbc=%s can only do domain decomposition in the x-direction",
1748                   c_pbcTypeNames[inputrec.pbcType].c_str());
1749     }
1750
1751     if (inputrec.nstlist == 0)
1752     {
1753         gmx_fatal(FARGS, "Domain decomposition does not work with nstlist=0");
1754     }
1755
1756     if (inputrec.comm_mode == ComRemovalAlgorithm::Angular && inputrec.pbcType != PbcType::No)
1757     {
1758         GMX_LOG(mdlog.warning)
1759                 .appendText(
1760                         "comm-mode angular will give incorrect results when the comm group "
1761                         "partially crosses a periodic boundary");
1762     }
1763 }
1764
1765 static real average_cellsize_min(const gmx_ddbox_t& ddbox, const ivec numDomains)
1766 {
1767     real r = ddbox.box_size[XX];
1768     for (int d = 0; d < DIM; d++)
1769     {
1770         if (numDomains[d] > 1)
1771         {
1772             /* Check using the initial average cell size */
1773             r = std::min(r, ddbox.box_size[d] * ddbox.skew_fac[d] / numDomains[d]);
1774         }
1775     }
1776
1777     return r;
1778 }
1779
1780 /*! \brief Depending on the DLB initial value return the DLB switched off state or issue an error.
1781  */
1782 static DlbState forceDlbOffOrBail(DlbState             cmdlineDlbState,
1783                                   const std::string&   reasonStr,
1784                                   const gmx::MDLogger& mdlog)
1785 {
1786     std::string dlbNotSupportedErr = "Dynamic load balancing requested, but ";
1787     std::string dlbDisableNote     = "NOTE: disabling dynamic load balancing as ";
1788
1789     if (cmdlineDlbState == DlbState::onUser)
1790     {
1791         gmx_fatal(FARGS, "%s", (dlbNotSupportedErr + reasonStr).c_str());
1792     }
1793     else if (cmdlineDlbState == DlbState::offCanTurnOn)
1794     {
1795         GMX_LOG(mdlog.info).appendText(dlbDisableNote + reasonStr);
1796     }
1797     return DlbState::offForever;
1798 }
1799
1800 /*! \brief Return the dynamic load balancer's initial state based on initial conditions and user inputs.
1801  *
1802  * This function parses the parameters of "-dlb" command line option setting
1803  * corresponding state values. Then it checks the consistency of the determined
1804  * state with other run parameters and settings. As a result, the initial state
1805  * may be altered or an error may be thrown if incompatibility of options is detected.
1806  *
1807  * \param [in] mdlog       Logger.
1808  * \param [in] dlbOption   Enum value for the DLB option.
1809  * \param [in] bRecordLoad True if the load balancer is recording load information.
1810  * \param [in] mdrunOptions  Options for mdrun.
1811  * \param [in] inputrec    Pointer mdrun to input parameters.
1812  * \returns                DLB initial/startup state.
1813  */
1814 static DlbState determineInitialDlbState(const gmx::MDLogger&     mdlog,
1815                                          DlbOption                dlbOption,
1816                                          gmx_bool                 bRecordLoad,
1817                                          const gmx::MdrunOptions& mdrunOptions,
1818                                          const t_inputrec&        inputrec)
1819 {
1820     DlbState dlbState = DlbState::offCanTurnOn;
1821
1822     switch (dlbOption)
1823     {
1824         case DlbOption::turnOnWhenUseful: dlbState = DlbState::offCanTurnOn; break;
1825         case DlbOption::no: dlbState = DlbState::offUser; break;
1826         case DlbOption::yes: dlbState = DlbState::onUser; break;
1827         default: gmx_incons("Invalid dlbOption enum value");
1828     }
1829
1830     /* Reruns don't support DLB: bail or override auto mode */
1831     if (mdrunOptions.rerun)
1832     {
1833         std::string reasonStr = "it is not supported in reruns.";
1834         return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1835     }
1836
1837     /* Unsupported integrators */
1838     if (!EI_DYNAMICS(inputrec.eI))
1839     {
1840         auto reasonStr =
1841                 gmx::formatString("it is only supported with dynamics, not with integrator '%s'.",
1842                                   enumValueToString(inputrec.eI));
1843         return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1844     }
1845
1846     /* Without cycle counters we can't time work to balance on */
1847     if (!bRecordLoad)
1848     {
1849         std::string reasonStr =
1850                 "cycle counters unsupported or not enabled in the operating system kernel.";
1851         return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1852     }
1853
1854     if (mdrunOptions.reproducible)
1855     {
1856         std::string reasonStr = "you started a reproducible run.";
1857         switch (dlbState)
1858         {
1859             case DlbState::offUser: break;
1860             case DlbState::offForever:
1861                 GMX_RELEASE_ASSERT(false, "DlbState::offForever is not a valid initial state");
1862                 break;
1863             case DlbState::offCanTurnOn: return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1864             case DlbState::onCanTurnOff:
1865                 GMX_RELEASE_ASSERT(false, "DlbState::offCanTurnOff is not a valid initial state");
1866                 break;
1867             case DlbState::onUser:
1868                 return forceDlbOffOrBail(
1869                         dlbState,
1870                         reasonStr
1871                                 + " In load balanced runs binary reproducibility cannot be "
1872                                   "ensured.",
1873                         mdlog);
1874             default:
1875                 gmx_fatal(FARGS,
1876                           "Death horror: undefined case (%d) for load balancing choice",
1877                           static_cast<int>(dlbState));
1878         }
1879     }
1880
1881     return dlbState;
1882 }
1883
1884 static gmx_domdec_comm_t* init_dd_comm()
1885 {
1886     gmx_domdec_comm_t* comm = new gmx_domdec_comm_t;
1887
1888     comm->n_load_have    = 0;
1889     comm->n_load_collect = 0;
1890
1891     comm->haveTurnedOffDlb = false;
1892
1893     for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
1894     {
1895         comm->sum_nat[i] = 0;
1896     }
1897     comm->ndecomp   = 0;
1898     comm->nload     = 0;
1899     comm->load_step = 0;
1900     comm->load_sum  = 0;
1901     comm->load_max  = 0;
1902     clear_ivec(comm->load_lim);
1903     comm->load_mdf = 0;
1904     comm->load_pme = 0;
1905
1906     /* This should be replaced by a unique pointer */
1907     comm->balanceRegion = ddBalanceRegionAllocate();
1908
1909     return comm;
1910 }
1911
1912 /* Returns whether mtop contains constraints and/or vsites */
1913 static bool systemHasConstraintsOrVsites(const gmx_mtop_t& mtop)
1914 {
1915     for (const auto ilists : IListRange(mtop))
1916     {
1917         if (!extractILists(ilists.list(), IF_CONSTRAINT | IF_VSITE).empty())
1918         {
1919             return true;
1920         }
1921     }
1922     return false;
1923 }
1924
1925 static void setupUpdateGroups(const gmx::MDLogger& mdlog,
1926                               const gmx_mtop_t&    mtop,
1927                               const t_inputrec&    inputrec,
1928                               const real           cutoffMargin,
1929                               DDSystemInfo*        systemInfo)
1930 {
1931     /* When we have constraints and/or vsites, it is beneficial to use
1932      * update groups (when possible) to allow independent update of groups.
1933      */
1934     if (!systemHasConstraintsOrVsites(mtop))
1935     {
1936         /* No constraints or vsites, atoms can be updated independently */
1937         return;
1938     }
1939
1940     systemInfo->updateGroupingPerMoleculetype = gmx::makeUpdateGroups(mtop);
1941     systemInfo->useUpdateGroups               = (!systemInfo->updateGroupingPerMoleculetype.empty()
1942                                    && getenv("GMX_NO_UPDATEGROUPS") == nullptr);
1943
1944     if (systemInfo->useUpdateGroups)
1945     {
1946         int numUpdateGroups = 0;
1947         for (const auto& molblock : mtop.molblock)
1948         {
1949             numUpdateGroups += molblock.nmol
1950                                * systemInfo->updateGroupingPerMoleculetype[molblock.type].numBlocks();
1951         }
1952
1953         systemInfo->maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1954                 mtop, systemInfo->updateGroupingPerMoleculetype, maxReferenceTemperature(inputrec));
1955
1956         /* To use update groups, the large domain-to-domain cutoff distance
1957          * should be compatible with the box size.
1958          */
1959         systemInfo->useUpdateGroups = (atomToAtomIntoDomainToDomainCutoff(*systemInfo, 0) < cutoffMargin);
1960
1961         if (systemInfo->useUpdateGroups)
1962         {
1963             GMX_LOG(mdlog.info)
1964                     .appendTextFormatted(
1965                             "Using update groups, nr %d, average size %.1f atoms, max. radius %.3f "
1966                             "nm\n",
1967                             numUpdateGroups,
1968                             mtop.natoms / static_cast<double>(numUpdateGroups),
1969                             systemInfo->maxUpdateGroupRadius);
1970         }
1971         else
1972         {
1973             GMX_LOG(mdlog.info)
1974                     .appendTextFormatted(
1975                             "The combination of rlist and box size prohibits the use of update "
1976                             "groups\n");
1977             systemInfo->updateGroupingPerMoleculetype.clear();
1978         }
1979     }
1980 }
1981
1982 UnitCellInfo::UnitCellInfo(const t_inputrec& ir) :
1983     npbcdim(numPbcDimensions(ir.pbcType)),
1984     numBoundedDimensions(inputrec2nboundeddim(&ir)),
1985     ddBoxIsDynamic(numBoundedDimensions < DIM || inputrecDynamicBox(&ir)),
1986     haveScrewPBC(ir.pbcType == PbcType::Screw)
1987 {
1988 }
1989
1990 /* Returns whether molecules are always whole, i.e. not broken by PBC */
1991 static bool moleculesAreAlwaysWhole(const gmx_mtop_t&                           mtop,
1992                                     const bool                                  useUpdateGroups,
1993                                     gmx::ArrayRef<const gmx::RangePartitioning> updateGroupingPerMoleculetype)
1994 {
1995     if (useUpdateGroups)
1996     {
1997         GMX_RELEASE_ASSERT(updateGroupingPerMoleculetype.size() == mtop.moltype.size(),
1998                            "Need one grouping per moltype");
1999         for (size_t mol = 0; mol < mtop.moltype.size(); mol++)
2000         {
2001             if (updateGroupingPerMoleculetype[mol].numBlocks() > 1)
2002             {
2003                 return false;
2004             }
2005         }
2006     }
2007     else
2008     {
2009         for (const auto& moltype : mtop.moltype)
2010         {
2011             if (moltype.atoms.nr > 1)
2012             {
2013                 return false;
2014             }
2015         }
2016     }
2017
2018     return true;
2019 }
2020
2021 /*! \brief Generate the simulation system information */
2022 static DDSystemInfo getSystemInfo(const gmx::MDLogger&           mdlog,
2023                                   DDRole                         ddRole,
2024                                   MPI_Comm                       communicator,
2025                                   const DomdecOptions&           options,
2026                                   const gmx_mtop_t&              mtop,
2027                                   const t_inputrec&              ir,
2028                                   const matrix                   box,
2029                                   gmx::ArrayRef<const gmx::RVec> xGlobal)
2030 {
2031     const real tenPercentMargin = 1.1;
2032
2033     DDSystemInfo systemInfo;
2034
2035     /* We need to decide on update groups early, as this affects communication distances */
2036     systemInfo.useUpdateGroups = false;
2037     if (ir.cutoff_scheme == CutoffScheme::Verlet)
2038     {
2039         real cutoffMargin = std::sqrt(max_cutoff2(ir.pbcType, box)) - ir.rlist;
2040         setupUpdateGroups(mdlog, mtop, ir, cutoffMargin, &systemInfo);
2041     }
2042
2043     systemInfo.moleculesAreAlwaysWhole = moleculesAreAlwaysWhole(
2044             mtop, systemInfo.useUpdateGroups, systemInfo.updateGroupingPerMoleculetype);
2045     systemInfo.haveInterDomainBondeds =
2046             (!systemInfo.moleculesAreAlwaysWhole || mtop.bIntermolecularInteractions);
2047     systemInfo.haveInterDomainMultiBodyBondeds =
2048             (systemInfo.haveInterDomainBondeds && multi_body_bondeds_count(mtop) > 0);
2049
2050     if (systemInfo.useUpdateGroups)
2051     {
2052         systemInfo.haveSplitConstraints = false;
2053         systemInfo.haveSplitSettles     = false;
2054     }
2055     else
2056     {
2057         systemInfo.haveSplitConstraints = (gmx_mtop_ftype_count(mtop, F_CONSTR) > 0
2058                                            || gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0);
2059         systemInfo.haveSplitSettles     = (gmx_mtop_ftype_count(mtop, F_SETTLE) > 0);
2060     }
2061
2062     if (ir.rlist == 0)
2063     {
2064         /* Set the cut-off to some very large value,
2065          * so we don't need if statements everywhere in the code.
2066          * We use sqrt, since the cut-off is squared in some places.
2067          */
2068         systemInfo.cutoff = GMX_CUTOFF_INF;
2069     }
2070     else
2071     {
2072         systemInfo.cutoff = atomToAtomIntoDomainToDomainCutoff(systemInfo, ir.rlist);
2073     }
2074     systemInfo.minCutoffForMultiBody = 0;
2075
2076     /* Determine the minimum cell size limit, affected by many factors */
2077     systemInfo.cellsizeLimit             = 0;
2078     systemInfo.filterBondedCommunication = false;
2079
2080     /* We do not allow home atoms to move beyond the neighboring domain
2081      * between domain decomposition steps, which limits the cell size.
2082      * Get an estimate of cell size limit due to atom displacement.
2083      * In most cases this is a large overestimate, because it assumes
2084      * non-interaction atoms.
2085      * We set the chance to 1 in a trillion steps.
2086      */
2087     constexpr real c_chanceThatAtomMovesBeyondDomain = 1e-12;
2088     const real     limitForAtomDisplacement          = minCellSizeForAtomDisplacement(
2089             mtop, ir, systemInfo.updateGroupingPerMoleculetype, c_chanceThatAtomMovesBeyondDomain);
2090     GMX_LOG(mdlog.info).appendTextFormatted("Minimum cell size due to atom displacement: %.3f nm", limitForAtomDisplacement);
2091
2092     systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, limitForAtomDisplacement);
2093
2094     /* TODO: PME decomposition currently requires atoms not to be more than
2095      *       2/3 of comm->cutoff, which is >=rlist, outside of their domain.
2096      *       In nearly all cases, limitForAtomDisplacement will be smaller
2097      *       than 2/3*rlist, so the PME requirement is satisfied.
2098      *       But it would be better for both correctness and performance
2099      *       to use limitForAtomDisplacement instead of 2/3*comm->cutoff.
2100      *       Note that we would need to improve the pairlist buffer case.
2101      */
2102
2103     if (systemInfo.haveInterDomainBondeds)
2104     {
2105         if (options.minimumCommunicationRange > 0)
2106         {
2107             systemInfo.minCutoffForMultiBody =
2108                     atomToAtomIntoDomainToDomainCutoff(systemInfo, options.minimumCommunicationRange);
2109             if (options.useBondedCommunication)
2110             {
2111                 systemInfo.filterBondedCommunication =
2112                         (systemInfo.minCutoffForMultiBody > systemInfo.cutoff);
2113             }
2114             else
2115             {
2116                 systemInfo.cutoff = std::max(systemInfo.cutoff, systemInfo.minCutoffForMultiBody);
2117             }
2118         }
2119         else if (ir.bPeriodicMols)
2120         {
2121             /* Can not easily determine the required cut-off */
2122             GMX_LOG(mdlog.warning)
2123                     .appendText(
2124                             "NOTE: Periodic molecules are present in this system. Because of this, "
2125                             "the domain decomposition algorithm cannot easily determine the "
2126                             "minimum cell size that it requires for treating bonded interactions. "
2127                             "Instead, domain decomposition will assume that half the non-bonded "
2128                             "cut-off will be a suitable lower bound.");
2129             systemInfo.minCutoffForMultiBody = systemInfo.cutoff / 2;
2130         }
2131         else
2132         {
2133             real r_2b = 0;
2134             real r_mb = 0;
2135
2136             if (ddRole == DDRole::Master)
2137             {
2138                 const DDBondedChecking ddBondedChecking = options.checkBondedInteractions
2139                                                                   ? DDBondedChecking::All
2140                                                                   : DDBondedChecking::ExcludeZeroLimit;
2141
2142                 dd_bonded_cg_distance(mdlog, mtop, ir, xGlobal, box, ddBondedChecking, &r_2b, &r_mb);
2143             }
2144             gmx_bcast(sizeof(r_2b), &r_2b, communicator);
2145             gmx_bcast(sizeof(r_mb), &r_mb, communicator);
2146
2147             /* We use an initial margin of 10% for the minimum cell size,
2148              * except when we are just below the non-bonded cut-off.
2149              */
2150             if (options.useBondedCommunication)
2151             {
2152                 if (std::max(r_2b, r_mb) > systemInfo.cutoff)
2153                 {
2154                     const real r_bonded              = std::max(r_2b, r_mb);
2155                     systemInfo.minCutoffForMultiBody = tenPercentMargin * r_bonded;
2156                     /* This is the (only) place where we turn on the filtering */
2157                     systemInfo.filterBondedCommunication = true;
2158                 }
2159                 else
2160                 {
2161                     const real r_bonded = r_mb;
2162                     systemInfo.minCutoffForMultiBody =
2163                             std::min(tenPercentMargin * r_bonded, systemInfo.cutoff);
2164                 }
2165                 /* We determine cutoff_mbody later */
2166                 systemInfo.increaseMultiBodyCutoff = true;
2167             }
2168             else
2169             {
2170                 /* No special bonded communication,
2171                  * simply increase the DD cut-off.
2172                  */
2173                 systemInfo.minCutoffForMultiBody = tenPercentMargin * std::max(r_2b, r_mb);
2174                 systemInfo.cutoff = std::max(systemInfo.cutoff, systemInfo.minCutoffForMultiBody);
2175             }
2176         }
2177         GMX_LOG(mdlog.info)
2178                 .appendTextFormatted("Minimum cell size due to bonded interactions: %.3f nm",
2179                                      systemInfo.minCutoffForMultiBody);
2180
2181         systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, systemInfo.minCutoffForMultiBody);
2182     }
2183
2184     systemInfo.constraintCommunicationRange = 0;
2185     if (systemInfo.haveSplitConstraints && options.constraintCommunicationRange <= 0)
2186     {
2187         /* There is a cell size limit due to the constraints (P-LINCS) */
2188         systemInfo.constraintCommunicationRange = gmx::constr_r_max(mdlog, &mtop, &ir);
2189         GMX_LOG(mdlog.info)
2190                 .appendTextFormatted("Estimated maximum distance required for P-LINCS: %.3f nm",
2191                                      systemInfo.constraintCommunicationRange);
2192         if (systemInfo.constraintCommunicationRange > systemInfo.cellsizeLimit)
2193         {
2194             GMX_LOG(mdlog.info)
2195                     .appendText(
2196                             "This distance will limit the DD cell size, you can override this with "
2197                             "-rcon");
2198         }
2199     }
2200     else if (options.constraintCommunicationRange > 0)
2201     {
2202         /* Here we do not check for dd->splitConstraints.
2203          * because one can also set a cell size limit for virtual sites only
2204          * and at this point we don't know yet if there are intercg v-sites.
2205          */
2206         GMX_LOG(mdlog.info)
2207                 .appendTextFormatted("User supplied maximum distance required for P-LINCS: %.3f nm",
2208                                      options.constraintCommunicationRange);
2209         systemInfo.constraintCommunicationRange = options.constraintCommunicationRange;
2210     }
2211     systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, systemInfo.constraintCommunicationRange);
2212
2213     return systemInfo;
2214 }
2215
2216 /*! \brief Exit with a fatal error if the DDGridSetup cannot be
2217  * implemented. */
2218 static void checkDDGridSetup(const DDGridSetup&   ddGridSetup,
2219                              DDRole               ddRole,
2220                              MPI_Comm             communicator,
2221                              int                  numNodes,
2222                              const DomdecOptions& options,
2223                              const DDSettings&    ddSettings,
2224                              const DDSystemInfo&  systemInfo,
2225                              const real           cellsizeLimit,
2226                              const gmx_ddbox_t&   ddbox)
2227 {
2228     if (options.numCells[XX] <= 0 && (ddGridSetup.numDomains[XX] == 0))
2229     {
2230         const bool  bC = (systemInfo.haveSplitConstraints
2231                          && systemInfo.constraintCommunicationRange > systemInfo.minCutoffForMultiBody);
2232         std::string message =
2233                 gmx::formatString("Change the number of ranks or mdrun option %s%s%s",
2234                                   !bC ? "-rdd" : "-rcon",
2235                                   ddSettings.initialDlbState != DlbState::offUser ? " or -dds" : "",
2236                                   bC ? " or your LINCS settings" : "");
2237
2238         gmx_fatal_collective(FARGS,
2239                              communicator,
2240                              ddRole == DDRole::Master,
2241                              "There is no domain decomposition for %d ranks that is compatible "
2242                              "with the given box and a minimum cell size of %g nm\n"
2243                              "%s\n"
2244                              "Look in the log file for details on the domain decomposition",
2245                              numNodes - ddGridSetup.numPmeOnlyRanks,
2246                              cellsizeLimit,
2247                              message.c_str());
2248     }
2249
2250     const real acs = average_cellsize_min(ddbox, ddGridSetup.numDomains);
2251     if (acs < cellsizeLimit)
2252     {
2253         if (options.numCells[XX] <= 0)
2254         {
2255             GMX_RELEASE_ASSERT(
2256                     false,
2257                     "dd_choose_grid() should return a grid that satisfies the cell size limits");
2258         }
2259         else
2260         {
2261             gmx_fatal_collective(
2262                     FARGS,
2263                     communicator,
2264                     ddRole == DDRole::Master,
2265                     "The initial cell size (%f) is smaller than the cell size limit (%f), change "
2266                     "options -dd, -rdd or -rcon, see the log file for details",
2267                     acs,
2268                     cellsizeLimit);
2269         }
2270     }
2271
2272     const int numPPRanks =
2273             ddGridSetup.numDomains[XX] * ddGridSetup.numDomains[YY] * ddGridSetup.numDomains[ZZ];
2274     if (numNodes - numPPRanks != ddGridSetup.numPmeOnlyRanks)
2275     {
2276         gmx_fatal_collective(FARGS,
2277                              communicator,
2278                              ddRole == DDRole::Master,
2279                              "The size of the domain decomposition grid (%d) does not match the "
2280                              "number of PP ranks (%d). The total number of ranks is %d",
2281                              numPPRanks,
2282                              numNodes - ddGridSetup.numPmeOnlyRanks,
2283                              numNodes);
2284     }
2285     if (ddGridSetup.numPmeOnlyRanks > numPPRanks)
2286     {
2287         gmx_fatal_collective(FARGS,
2288                              communicator,
2289                              ddRole == DDRole::Master,
2290                              "The number of separate PME ranks (%d) is larger than the number of "
2291                              "PP ranks (%d), this is not supported.",
2292                              ddGridSetup.numPmeOnlyRanks,
2293                              numPPRanks);
2294     }
2295 }
2296
2297 /*! \brief Set the cell size and interaction limits, as well as the DD grid */
2298 static DDRankSetup getDDRankSetup(const gmx::MDLogger& mdlog,
2299                                   int                  numNodes,
2300                                   const DDGridSetup&   ddGridSetup,
2301                                   const t_inputrec&    ir)
2302 {
2303     GMX_LOG(mdlog.info)
2304             .appendTextFormatted("Domain decomposition grid %d x %d x %d, separate PME ranks %d",
2305                                  ddGridSetup.numDomains[XX],
2306                                  ddGridSetup.numDomains[YY],
2307                                  ddGridSetup.numDomains[ZZ],
2308                                  ddGridSetup.numPmeOnlyRanks);
2309
2310     DDRankSetup ddRankSetup;
2311
2312     ddRankSetup.numPPRanks = numNodes - ddGridSetup.numPmeOnlyRanks;
2313     copy_ivec(ddGridSetup.numDomains, ddRankSetup.numPPCells);
2314
2315     ddRankSetup.usePmeOnlyRanks = (ddGridSetup.numPmeOnlyRanks > 0);
2316     if (ddRankSetup.usePmeOnlyRanks)
2317     {
2318         ddRankSetup.numRanksDoingPme = ddGridSetup.numPmeOnlyRanks;
2319     }
2320     else
2321     {
2322         ddRankSetup.numRanksDoingPme =
2323                 ddGridSetup.numDomains[XX] * ddGridSetup.numDomains[YY] * ddGridSetup.numDomains[ZZ];
2324     }
2325
2326     if (EEL_PME(ir.coulombtype) || EVDW_PME(ir.vdwtype))
2327     {
2328         /* The following choices should match those
2329          * in comm_cost_est in domdec_setup.c.
2330          * Note that here the checks have to take into account
2331          * that the decomposition might occur in a different order than xyz
2332          * (for instance through the env.var. GMX_DD_ORDER_ZYX),
2333          * in which case they will not match those in comm_cost_est,
2334          * but since that is mainly for testing purposes that's fine.
2335          */
2336         if (ddGridSetup.numDDDimensions >= 2 && ddGridSetup.ddDimensions[0] == XX
2337             && ddGridSetup.ddDimensions[1] == YY
2338             && ddRankSetup.numRanksDoingPme > ddGridSetup.numDomains[XX]
2339             && ddRankSetup.numRanksDoingPme % ddGridSetup.numDomains[XX] == 0
2340             && getenv("GMX_PMEONEDD") == nullptr)
2341         {
2342             ddRankSetup.npmedecompdim = 2;
2343             ddRankSetup.npmenodes_x   = ddGridSetup.numDomains[XX];
2344             ddRankSetup.npmenodes_y   = ddRankSetup.numRanksDoingPme / ddRankSetup.npmenodes_x;
2345         }
2346         else
2347         {
2348             /* In case nc is 1 in both x and y we could still choose to
2349              * decompose pme in y instead of x, but we use x for simplicity.
2350              */
2351             ddRankSetup.npmedecompdim = 1;
2352             if (ddGridSetup.ddDimensions[0] == YY)
2353             {
2354                 ddRankSetup.npmenodes_x = 1;
2355                 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme;
2356             }
2357             else
2358             {
2359                 ddRankSetup.npmenodes_x = ddRankSetup.numRanksDoingPme;
2360                 ddRankSetup.npmenodes_y = 1;
2361             }
2362         }
2363         GMX_LOG(mdlog.info)
2364                 .appendTextFormatted("PME domain decomposition: %d x %d x %d",
2365                                      ddRankSetup.npmenodes_x,
2366                                      ddRankSetup.npmenodes_y,
2367                                      1);
2368     }
2369     else
2370     {
2371         ddRankSetup.npmedecompdim = 0;
2372         ddRankSetup.npmenodes_x   = 0;
2373         ddRankSetup.npmenodes_y   = 0;
2374     }
2375
2376     return ddRankSetup;
2377 }
2378
2379 /*! \brief Set the cell size and interaction limits */
2380 static void set_dd_limits(const gmx::MDLogger& mdlog,
2381                           DDRole               ddRole,
2382                           gmx_domdec_t*        dd,
2383                           const DomdecOptions& options,
2384                           const DDSettings&    ddSettings,
2385                           const DDSystemInfo&  systemInfo,
2386                           const DDGridSetup&   ddGridSetup,
2387                           const int            numPPRanks,
2388                           const gmx_mtop_t&    mtop,
2389                           const t_inputrec&    ir,
2390                           const gmx_ddbox_t&   ddbox)
2391 {
2392     gmx_domdec_comm_t* comm = dd->comm;
2393     comm->ddSettings        = ddSettings;
2394
2395     /* Initialize to GPU share count to 0, might change later */
2396     comm->nrank_gpu_shared = 0;
2397
2398     comm->dlbState = comm->ddSettings.initialDlbState;
2399     dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, TRUE);
2400     /* To consider turning DLB on after 2*nstlist steps we need to check
2401      * at partitioning count 3. Thus we need to increase the first count by 2.
2402      */
2403     comm->ddPartioningCountFirstDlbOff += 2;
2404
2405     comm->bPMELoadBalDLBLimits = FALSE;
2406
2407     /* Allocate the charge group/atom sorting struct */
2408     comm->sort = std::make_unique<gmx_domdec_sort_t>();
2409
2410     comm->systemInfo = systemInfo;
2411
2412     if (systemInfo.useUpdateGroups)
2413     {
2414         /* Note: We would like to use dd->nnodes for the atom count estimate,
2415          *       but that is not yet available here. But this anyhow only
2416          *       affect performance up to the second dd_partition_system call.
2417          */
2418         const int homeAtomCountEstimate = mtop.natoms / numPPRanks;
2419         comm->updateGroupsCog           = std::make_unique<gmx::UpdateGroupsCog>(
2420                 mtop, systemInfo.updateGroupingPerMoleculetype, maxReferenceTemperature(ir), homeAtomCountEstimate);
2421     }
2422
2423     /* Set the DD setup given by ddGridSetup */
2424     copy_ivec(ddGridSetup.numDomains, dd->numCells);
2425     dd->ndim = ddGridSetup.numDDDimensions;
2426     copy_ivec(ddGridSetup.ddDimensions, dd->dim);
2427
2428     dd->nnodes = dd->numCells[XX] * dd->numCells[YY] * dd->numCells[ZZ];
2429
2430     snew(comm->slb_frac, DIM);
2431     if (isDlbDisabled(comm))
2432     {
2433         comm->slb_frac[XX] = get_slb_frac(mdlog, "x", dd->numCells[XX], options.cellSizeX);
2434         comm->slb_frac[YY] = get_slb_frac(mdlog, "y", dd->numCells[YY], options.cellSizeY);
2435         comm->slb_frac[ZZ] = get_slb_frac(mdlog, "z", dd->numCells[ZZ], options.cellSizeZ);
2436     }
2437
2438     /* Set the multi-body cut-off and cellsize limit for DLB */
2439     comm->cutoff_mbody   = systemInfo.minCutoffForMultiBody;
2440     comm->cellsize_limit = systemInfo.cellsizeLimit;
2441     if (systemInfo.haveInterDomainBondeds && systemInfo.increaseMultiBodyCutoff)
2442     {
2443         if (systemInfo.filterBondedCommunication || !isDlbDisabled(comm))
2444         {
2445             /* Set the bonded communication distance to halfway
2446              * the minimum and the maximum,
2447              * since the extra communication cost is nearly zero.
2448              */
2449             real acs           = average_cellsize_min(ddbox, dd->numCells);
2450             comm->cutoff_mbody = 0.5 * (systemInfo.minCutoffForMultiBody + acs);
2451             if (!isDlbDisabled(comm))
2452             {
2453                 /* Check if this does not limit the scaling */
2454                 comm->cutoff_mbody = std::min(comm->cutoff_mbody, options.dlbScaling * acs);
2455             }
2456             if (!systemInfo.filterBondedCommunication)
2457             {
2458                 /* Without bBondComm do not go beyond the n.b. cut-off */
2459                 comm->cutoff_mbody = std::min(comm->cutoff_mbody, systemInfo.cutoff);
2460                 if (comm->cellsize_limit >= systemInfo.cutoff)
2461                 {
2462                     /* We don't loose a lot of efficieny
2463                      * when increasing it to the n.b. cut-off.
2464                      * It can even be slightly faster, because we need
2465                      * less checks for the communication setup.
2466                      */
2467                     comm->cutoff_mbody = systemInfo.cutoff;
2468                 }
2469             }
2470             /* Check if we did not end up below our original limit */
2471             comm->cutoff_mbody = std::max(comm->cutoff_mbody, systemInfo.minCutoffForMultiBody);
2472
2473             if (comm->cutoff_mbody > comm->cellsize_limit)
2474             {
2475                 comm->cellsize_limit = comm->cutoff_mbody;
2476             }
2477         }
2478         /* Without DLB and cutoff_mbody<cutoff, cutoff_mbody is dynamic */
2479     }
2480
2481     if (debug)
2482     {
2483         fprintf(debug,
2484                 "Bonded atom communication beyond the cut-off: %s\n"
2485                 "cellsize limit %f\n",
2486                 gmx::boolToString(systemInfo.filterBondedCommunication),
2487                 comm->cellsize_limit);
2488     }
2489
2490     if (ddRole == DDRole::Master)
2491     {
2492         check_dd_restrictions(dd, ir, mdlog);
2493     }
2494 }
2495
2496 void dd_init_bondeds(FILE*                           fplog,
2497                      gmx_domdec_t*                   dd,
2498                      const gmx_mtop_t&               mtop,
2499                      const gmx::VirtualSitesHandler* vsite,
2500                      const t_inputrec&               inputrec,
2501                      const DDBondedChecking          ddBondedChecking,
2502                      gmx::ArrayRef<cginfo_mb_t>      cginfo_mb)
2503 {
2504     dd_make_reverse_top(fplog, dd, mtop, vsite, inputrec, ddBondedChecking);
2505
2506     gmx_domdec_comm_t* comm = dd->comm;
2507
2508     if (comm->systemInfo.filterBondedCommunication)
2509     {
2510         /* Communicate atoms beyond the cut-off for bonded interactions */
2511         comm->bondedLinks = makeBondedLinks(mtop, cginfo_mb);
2512     }
2513     else
2514     {
2515         /* Only communicate atoms based on cut-off */
2516         comm->bondedLinks = nullptr;
2517     }
2518 }
2519
2520 static void writeSettings(gmx::TextWriter*   log,
2521                           gmx_domdec_t*      dd,
2522                           const gmx_mtop_t&  mtop,
2523                           const t_inputrec&  ir,
2524                           gmx_bool           bDynLoadBal,
2525                           real               dlb_scale,
2526                           const gmx_ddbox_t* ddbox)
2527 {
2528     gmx_domdec_comm_t* comm = dd->comm;
2529
2530     if (bDynLoadBal)
2531     {
2532         log->writeString("The maximum number of communication pulses is:");
2533         for (int d = 0; d < dd->ndim; d++)
2534         {
2535             log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), comm->cd[d].np_dlb);
2536         }
2537         log->ensureLineBreak();
2538         log->writeLineFormatted("The minimum size for domain decomposition cells is %.3f nm",
2539                                 comm->cellsize_limit);
2540         log->writeLineFormatted("The requested allowed shrink of DD cells (option -dds) is: %.2f", dlb_scale);
2541         log->writeString("The allowed shrink of domain decomposition cells is:");
2542         for (int d = 0; d < DIM; d++)
2543         {
2544             if (dd->numCells[d] > 1)
2545             {
2546                 const real shrink =
2547                         (d >= ddbox->npbcdim && dd->numCells[d] == 2)
2548                                 ? 0
2549                                 : comm->cellsize_min_dlb[d]
2550                                           / (ddbox->box_size[d] * ddbox->skew_fac[d] / dd->numCells[d]);
2551                 log->writeStringFormatted(" %c %.2f", dim2char(d), shrink);
2552             }
2553         }
2554         log->ensureLineBreak();
2555     }
2556     else
2557     {
2558         ivec np;
2559         set_dd_cell_sizes_slb(dd, ddbox, setcellsizeslbPULSE_ONLY, np);
2560         log->writeString("The initial number of communication pulses is:");
2561         for (int d = 0; d < dd->ndim; d++)
2562         {
2563             log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), np[dd->dim[d]]);
2564         }
2565         log->ensureLineBreak();
2566         log->writeString("The initial domain decomposition cell size is:");
2567         for (int d = 0; d < DIM; d++)
2568         {
2569             if (dd->numCells[d] > 1)
2570             {
2571                 log->writeStringFormatted(" %c %.2f nm", dim2char(d), dd->comm->cellsize_min[d]);
2572             }
2573         }
2574         log->ensureLineBreak();
2575         log->writeLine();
2576     }
2577
2578     const bool haveInterDomainVsites =
2579             (countInterUpdategroupVsites(mtop, comm->systemInfo.updateGroupingPerMoleculetype) != 0);
2580
2581     if (comm->systemInfo.haveInterDomainBondeds || haveInterDomainVsites
2582         || comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2583     {
2584         std::string decompUnits;
2585         if (comm->systemInfo.useUpdateGroups)
2586         {
2587             decompUnits = "atom groups";
2588         }
2589         else
2590         {
2591             decompUnits = "atoms";
2592         }
2593
2594         log->writeLineFormatted("The maximum allowed distance for %s involved in interactions is:",
2595                                 decompUnits.c_str());
2596         log->writeLineFormatted(
2597                 "%40s  %-7s %6.3f nm", "non-bonded interactions", "", comm->systemInfo.cutoff);
2598
2599         real limit = 0;
2600         if (bDynLoadBal)
2601         {
2602             limit = dd->comm->cellsize_limit;
2603         }
2604         else
2605         {
2606             if (dd->unitCellInfo.ddBoxIsDynamic)
2607             {
2608                 log->writeLine(
2609                         "(the following are initial values, they could change due to box "
2610                         "deformation)");
2611             }
2612             limit = dd->comm->cellsize_min[XX];
2613             for (int d = 1; d < DIM; d++)
2614             {
2615                 limit = std::min(limit, dd->comm->cellsize_min[d]);
2616             }
2617         }
2618
2619         if (comm->systemInfo.haveInterDomainBondeds)
2620         {
2621             log->writeLineFormatted("%40s  %-7s %6.3f nm",
2622                                     "two-body bonded interactions",
2623                                     "(-rdd)",
2624                                     std::max(comm->systemInfo.cutoff, comm->cutoff_mbody));
2625             log->writeLineFormatted("%40s  %-7s %6.3f nm",
2626                                     "multi-body bonded interactions",
2627                                     "(-rdd)",
2628                                     (comm->systemInfo.filterBondedCommunication || isDlbOn(dd->comm))
2629                                             ? comm->cutoff_mbody
2630                                             : std::min(comm->systemInfo.cutoff, limit));
2631         }
2632         if (haveInterDomainVsites)
2633         {
2634             log->writeLineFormatted("%40s  %-7s %6.3f nm", "virtual site constructions", "(-rcon)", limit);
2635         }
2636         if (comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2637         {
2638             std::string separation =
2639                     gmx::formatString("atoms separated by up to %d constraints", 1 + ir.nProjOrder);
2640             log->writeLineFormatted("%40s  %-7s %6.3f nm\n", separation.c_str(), "(-rcon)", limit);
2641         }
2642         log->ensureLineBreak();
2643     }
2644 }
2645
2646 static void logSettings(const gmx::MDLogger& mdlog,
2647                         gmx_domdec_t*        dd,
2648                         const gmx_mtop_t&    mtop,
2649                         const t_inputrec&    ir,
2650                         real                 dlb_scale,
2651                         const gmx_ddbox_t*   ddbox)
2652 {
2653     gmx::StringOutputStream stream;
2654     gmx::TextWriter         log(&stream);
2655     writeSettings(&log, dd, mtop, ir, isDlbOn(dd->comm), dlb_scale, ddbox);
2656     if (dd->comm->dlbState == DlbState::offCanTurnOn)
2657     {
2658         {
2659             log.ensureEmptyLine();
2660             log.writeLine(
2661                     "When dynamic load balancing gets turned on, these settings will change to:");
2662         }
2663         writeSettings(&log, dd, mtop, ir, true, dlb_scale, ddbox);
2664     }
2665     GMX_LOG(mdlog.info).asParagraph().appendText(stream.toString());
2666 }
2667
2668 static void set_cell_limits_dlb(const gmx::MDLogger& mdlog,
2669                                 gmx_domdec_t*        dd,
2670                                 real                 dlb_scale,
2671                                 const t_inputrec&    inputrec,
2672                                 const gmx_ddbox_t*   ddbox)
2673 {
2674     int npulse       = 0;
2675     int npulse_d_max = 0;
2676     int npulse_d     = 0;
2677
2678     gmx_domdec_comm_t* comm = dd->comm;
2679
2680     bool bNoCutOff = (inputrec.rvdw == 0 || inputrec.rcoulomb == 0);
2681
2682     /* Determine the maximum number of comm. pulses in one dimension */
2683
2684     comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2685
2686     /* Determine the maximum required number of grid pulses */
2687     if (comm->cellsize_limit >= comm->systemInfo.cutoff)
2688     {
2689         /* Only a single pulse is required */
2690         npulse = 1;
2691     }
2692     else if (!bNoCutOff && comm->cellsize_limit > 0)
2693     {
2694         /* We round down slightly here to avoid overhead due to the latency
2695          * of extra communication calls when the cut-off
2696          * would be only slightly longer than the cell size.
2697          * Later cellsize_limit is redetermined,
2698          * so we can not miss interactions due to this rounding.
2699          */
2700         npulse = static_cast<int>(0.96 + comm->systemInfo.cutoff / comm->cellsize_limit);
2701     }
2702     else
2703     {
2704         /* There is no cell size limit */
2705         npulse = std::max(dd->numCells[XX] - 1, std::max(dd->numCells[YY] - 1, dd->numCells[ZZ] - 1));
2706     }
2707
2708     if (!bNoCutOff && npulse > 1)
2709     {
2710         /* See if we can do with less pulses, based on dlb_scale */
2711         npulse_d_max = 0;
2712         for (int d = 0; d < dd->ndim; d++)
2713         {
2714             int dim  = dd->dim[d];
2715             npulse_d = static_cast<int>(
2716                     1
2717                     + dd->numCells[dim] * comm->systemInfo.cutoff
2718                               / (ddbox->box_size[dim] * ddbox->skew_fac[dim] * dlb_scale));
2719             npulse_d_max = std::max(npulse_d_max, npulse_d);
2720         }
2721         npulse = std::min(npulse, npulse_d_max);
2722     }
2723
2724     /* This env var can override npulse */
2725     const int ddPulseEnv = dd_getenv(mdlog, "GMX_DD_NPULSE", 0);
2726     if (ddPulseEnv > 0)
2727     {
2728         npulse = ddPulseEnv;
2729     }
2730
2731     comm->maxpulse       = 1;
2732     comm->bVacDLBNoLimit = (inputrec.pbcType == PbcType::No);
2733     for (int d = 0; d < dd->ndim; d++)
2734     {
2735         comm->cd[d].np_dlb = std::min(npulse, dd->numCells[dd->dim[d]] - 1);
2736         comm->maxpulse     = std::max(comm->maxpulse, comm->cd[d].np_dlb);
2737         if (comm->cd[d].np_dlb < dd->numCells[dd->dim[d]] - 1)
2738         {
2739             comm->bVacDLBNoLimit = FALSE;
2740         }
2741     }
2742
2743     /* cellsize_limit is set for LINCS in init_domain_decomposition */
2744     if (!comm->bVacDLBNoLimit)
2745     {
2746         comm->cellsize_limit = std::max(comm->cellsize_limit, comm->systemInfo.cutoff / comm->maxpulse);
2747     }
2748     comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2749     /* Set the minimum cell size for each DD dimension */
2750     for (int d = 0; d < dd->ndim; d++)
2751     {
2752         if (comm->bVacDLBNoLimit || comm->cd[d].np_dlb * comm->cellsize_limit >= comm->systemInfo.cutoff)
2753         {
2754             comm->cellsize_min_dlb[dd->dim[d]] = comm->cellsize_limit;
2755         }
2756         else
2757         {
2758             comm->cellsize_min_dlb[dd->dim[d]] = comm->systemInfo.cutoff / comm->cd[d].np_dlb;
2759         }
2760     }
2761     if (comm->cutoff_mbody <= 0)
2762     {
2763         comm->cutoff_mbody = std::min(comm->systemInfo.cutoff, comm->cellsize_limit);
2764     }
2765     if (isDlbOn(comm))
2766     {
2767         set_dlb_limits(dd);
2768     }
2769 }
2770
2771 bool dd_moleculesAreAlwaysWhole(const gmx_domdec_t& dd)
2772 {
2773     return dd.comm->systemInfo.moleculesAreAlwaysWhole;
2774 }
2775
2776 bool dd_bonded_molpbc(const gmx_domdec_t& dd, PbcType pbcType)
2777 {
2778     /* If each molecule is a single charge group
2779      * or we use domain decomposition for each periodic dimension,
2780      * we do not need to take pbc into account for the bonded interactions.
2781      */
2782     return (pbcType != PbcType::No && dd.comm->systemInfo.haveInterDomainBondeds
2783             && !(dd.numCells[XX] > 1 && dd.numCells[YY] > 1
2784                  && (dd.numCells[ZZ] > 1 || pbcType == PbcType::XY)));
2785 }
2786
2787 /*! \brief Sets grid size limits and PP-PME setup, prints settings to log */
2788 static void set_ddgrid_parameters(const gmx::MDLogger& mdlog,
2789                                   gmx_domdec_t*        dd,
2790                                   real                 dlb_scale,
2791                                   const gmx_mtop_t&    mtop,
2792                                   const t_inputrec&    inputrec,
2793                                   const gmx_ddbox_t*   ddbox)
2794 {
2795     gmx_domdec_comm_t* comm        = dd->comm;
2796     DDRankSetup&       ddRankSetup = comm->ddRankSetup;
2797
2798     if (EEL_PME(inputrec.coulombtype) || EVDW_PME(inputrec.vdwtype))
2799     {
2800         init_ddpme(dd, &ddRankSetup.ddpme[0], 0);
2801         if (ddRankSetup.npmedecompdim >= 2)
2802         {
2803             init_ddpme(dd, &ddRankSetup.ddpme[1], 1);
2804         }
2805     }
2806     else
2807     {
2808         ddRankSetup.numRanksDoingPme = 0;
2809         if (dd->pme_nodeid >= 0)
2810         {
2811             gmx_fatal_collective(FARGS,
2812                                  dd->mpi_comm_all,
2813                                  DDMASTER(dd),
2814                                  "Can not have separate PME ranks without PME electrostatics");
2815         }
2816     }
2817
2818     if (debug)
2819     {
2820         fprintf(debug, "The DD cut-off is %f\n", comm->systemInfo.cutoff);
2821     }
2822     if (!isDlbDisabled(comm))
2823     {
2824         set_cell_limits_dlb(mdlog, dd, dlb_scale, inputrec, ddbox);
2825     }
2826
2827     logSettings(mdlog, dd, mtop, inputrec, dlb_scale, ddbox);
2828
2829     const real vol_frac = (inputrec.pbcType == PbcType::No)
2830                                   ? (1 - 1 / static_cast<double>(dd->nnodes))
2831                                   : ((1 + comm_box_frac(dd->numCells, comm->systemInfo.cutoff, *ddbox))
2832                                      / static_cast<double>(dd->nnodes));
2833     if (debug)
2834     {
2835         fprintf(debug, "Volume fraction for all DD zones: %f\n", vol_frac);
2836     }
2837     int natoms_tot = mtop.natoms;
2838
2839     dd->ga2la = new gmx_ga2la_t(natoms_tot, static_cast<int>(vol_frac * natoms_tot));
2840 }
2841
2842 /*! \brief Get some important DD parameters which can be modified by env.vars */
2843 static DDSettings getDDSettings(const gmx::MDLogger&     mdlog,
2844                                 const DomdecOptions&     options,
2845                                 const gmx::MdrunOptions& mdrunOptions,
2846                                 const t_inputrec&        ir)
2847 {
2848     DDSettings ddSettings;
2849
2850     ddSettings.useSendRecv2        = (dd_getenv(mdlog, "GMX_DD_USE_SENDRECV2", 0) != 0);
2851     ddSettings.dlb_scale_lim       = dd_getenv(mdlog, "GMX_DLB_MAX_BOX_SCALING", 10);
2852     ddSettings.useDDOrderZYX       = bool(dd_getenv(mdlog, "GMX_DD_ORDER_ZYX", 0));
2853     ddSettings.useCartesianReorder = bool(dd_getenv(mdlog, "GMX_NO_CART_REORDER", 1));
2854     ddSettings.eFlop               = dd_getenv(mdlog, "GMX_DLB_BASED_ON_FLOPS", 0);
2855     const int recload              = dd_getenv(mdlog, "GMX_DD_RECORD_LOAD", 1);
2856     ddSettings.nstDDDump           = dd_getenv(mdlog, "GMX_DD_NST_DUMP", 0);
2857     ddSettings.nstDDDumpGrid       = dd_getenv(mdlog, "GMX_DD_NST_DUMP_GRID", 0);
2858     ddSettings.DD_debug            = dd_getenv(mdlog, "GMX_DD_DEBUG", 0);
2859
2860     if (ddSettings.useSendRecv2)
2861     {
2862         GMX_LOG(mdlog.info)
2863                 .appendText(
2864                         "Will use two sequential MPI_Sendrecv calls instead of two simultaneous "
2865                         "non-blocking MPI_Irecv and MPI_Isend pairs for constraint and vsite "
2866                         "communication");
2867     }
2868
2869     if (ddSettings.eFlop)
2870     {
2871         GMX_LOG(mdlog.info).appendText("Will load balance based on FLOP count");
2872         ddSettings.recordLoad = true;
2873     }
2874     else
2875     {
2876         ddSettings.recordLoad = (wallcycle_have_counter() && recload > 0);
2877     }
2878
2879     ddSettings.initialDlbState =
2880             determineInitialDlbState(mdlog, options.dlbOption, ddSettings.recordLoad, mdrunOptions, ir);
2881     GMX_LOG(mdlog.info)
2882             .appendTextFormatted("Dynamic load balancing: %s",
2883                                  edlbs_names[static_cast<int>(ddSettings.initialDlbState)]);
2884
2885     return ddSettings;
2886 }
2887
2888 gmx_domdec_t::gmx_domdec_t(const t_inputrec& ir) : unitCellInfo(ir) {}
2889
2890 gmx_domdec_t::~gmx_domdec_t() = default;
2891
2892 namespace gmx
2893 {
2894
2895 // TODO once the functionality stablizes, move this class and
2896 // supporting functionality into builder.cpp
2897 /*! \brief Impl class for DD builder */
2898 class DomainDecompositionBuilder::Impl
2899 {
2900 public:
2901     //! Constructor
2902     Impl(const MDLogger&      mdlog,
2903          t_commrec*           cr,
2904          const DomdecOptions& options,
2905          const MdrunOptions&  mdrunOptions,
2906          const gmx_mtop_t&    mtop,
2907          const t_inputrec&    ir,
2908          const matrix         box,
2909          ArrayRef<const RVec> xGlobal);
2910
2911     //! Build the resulting DD manager
2912     gmx_domdec_t* build(LocalAtomSetManager* atomSets);
2913
2914     //! Objects used in constructing and configuring DD
2915     //! {
2916     //! Logging object
2917     const MDLogger& mdlog_;
2918     //! Communication object
2919     t_commrec* cr_;
2920     //! User-supplied options configuring DD behavior
2921     const DomdecOptions options_;
2922     //! Global system topology
2923     const gmx_mtop_t& mtop_;
2924     //! User input values from the tpr file
2925     const t_inputrec& ir_;
2926     //! }
2927
2928     //! Internal objects used in constructing DD
2929     //! {
2930     //! Settings combined from the user input
2931     DDSettings ddSettings_;
2932     //! Information derived from the simulation system
2933     DDSystemInfo systemInfo_;
2934     //! Box structure
2935     gmx_ddbox_t ddbox_ = { 0 };
2936     //! Organization of the DD grids
2937     DDGridSetup ddGridSetup_;
2938     //! Organzation of the DD ranks
2939     DDRankSetup ddRankSetup_;
2940     //! Number of DD cells in each dimension
2941     ivec ddCellIndex_ = { 0, 0, 0 };
2942     //! IDs of PME-only ranks
2943     std::vector<int> pmeRanks_;
2944     //! Contains a valid Cartesian-communicator-based setup, or defaults.
2945     CartesianRankSetup cartSetup_;
2946     //! }
2947 };
2948
2949 DomainDecompositionBuilder::Impl::Impl(const MDLogger&      mdlog,
2950                                        t_commrec*           cr,
2951                                        const DomdecOptions& options,
2952                                        const MdrunOptions&  mdrunOptions,
2953                                        const gmx_mtop_t&    mtop,
2954                                        const t_inputrec&    ir,
2955                                        const matrix         box,
2956                                        ArrayRef<const RVec> xGlobal) :
2957     mdlog_(mdlog),
2958     cr_(cr),
2959     options_(options),
2960     mtop_(mtop),
2961     ir_(ir)
2962 {
2963     GMX_LOG(mdlog_.info).appendTextFormatted("\nInitializing Domain Decomposition on %d ranks", cr_->sizeOfDefaultCommunicator);
2964
2965     ddSettings_ = getDDSettings(mdlog_, options_, mdrunOptions, ir_);
2966
2967     if (ddSettings_.eFlop > 1)
2968     {
2969         /* Ensure that we have different random flop counts on different ranks */
2970         srand(1 + cr_->rankInDefaultCommunicator);
2971     }
2972
2973     systemInfo_ = getSystemInfo(mdlog_,
2974                                 MASTER(cr_) ? DDRole::Master : DDRole::Agent,
2975                                 cr->mpiDefaultCommunicator,
2976                                 options_,
2977                                 mtop_,
2978                                 ir_,
2979                                 box,
2980                                 xGlobal);
2981
2982     const int  numRanksRequested         = cr_->sizeOfDefaultCommunicator;
2983     const bool checkForLargePrimeFactors = (options_.numCells[0] <= 0);
2984     checkForValidRankCountRequests(
2985             numRanksRequested, EEL_PME(ir_.coulombtype), options_.numPmeRanks, checkForLargePrimeFactors);
2986
2987     // DD grid setup uses a more different cell size limit for
2988     // automated setup than the one in systemInfo_. The latter is used
2989     // in set_dd_limits() to configure DLB, for example.
2990     const real gridSetupCellsizeLimit =
2991             getDDGridSetupCellSizeLimit(mdlog_,
2992                                         !isDlbDisabled(ddSettings_.initialDlbState),
2993                                         options_.dlbScaling,
2994                                         ir_,
2995                                         systemInfo_.cellsizeLimit);
2996     ddGridSetup_ = getDDGridSetup(mdlog_,
2997                                   MASTER(cr_) ? DDRole::Master : DDRole::Agent,
2998                                   cr->mpiDefaultCommunicator,
2999                                   numRanksRequested,
3000                                   options_,
3001                                   ddSettings_,
3002                                   systemInfo_,
3003                                   gridSetupCellsizeLimit,
3004                                   mtop_,
3005                                   ir_,
3006                                   box,
3007                                   xGlobal,
3008                                   &ddbox_);
3009     checkDDGridSetup(ddGridSetup_,
3010                      MASTER(cr_) ? DDRole::Master : DDRole::Agent,
3011                      cr->mpiDefaultCommunicator,
3012                      cr->sizeOfDefaultCommunicator,
3013                      options_,
3014                      ddSettings_,
3015                      systemInfo_,
3016                      gridSetupCellsizeLimit,
3017                      ddbox_);
3018
3019     cr_->npmenodes = ddGridSetup_.numPmeOnlyRanks;
3020
3021     ddRankSetup_ = getDDRankSetup(mdlog_, cr_->sizeOfDefaultCommunicator, ddGridSetup_, ir_);
3022
3023     /* Generate the group communicator, also decides the duty of each rank */
3024     cartSetup_ = makeGroupCommunicators(
3025             mdlog_, ddSettings_, options_.rankOrder, ddRankSetup_, cr_, ddCellIndex_, &pmeRanks_);
3026 }
3027
3028 gmx_domdec_t* DomainDecompositionBuilder::Impl::build(LocalAtomSetManager* atomSets)
3029 {
3030     gmx_domdec_t* dd = new gmx_domdec_t(ir_);
3031
3032     copy_ivec(ddCellIndex_, dd->ci);
3033
3034     dd->comm = init_dd_comm();
3035
3036     dd->comm->ddRankSetup        = ddRankSetup_;
3037     dd->comm->cartesianRankSetup = cartSetup_;
3038
3039     set_dd_limits(mdlog_,
3040                   MASTER(cr_) ? DDRole::Master : DDRole::Agent,
3041                   dd,
3042                   options_,
3043                   ddSettings_,
3044                   systemInfo_,
3045                   ddGridSetup_,
3046                   ddRankSetup_.numPPRanks,
3047                   mtop_,
3048                   ir_,
3049                   ddbox_);
3050
3051     setupGroupCommunication(mdlog_, ddSettings_, pmeRanks_, cr_, mtop_.natoms, dd);
3052
3053     if (thisRankHasDuty(cr_, DUTY_PP))
3054     {
3055         set_ddgrid_parameters(mdlog_, dd, options_.dlbScaling, mtop_, ir_, &ddbox_);
3056
3057         setup_neighbor_relations(dd);
3058     }
3059
3060     /* Set overallocation to avoid frequent reallocation of arrays */
3061     set_over_alloc_dd(true);
3062
3063     dd->atomSets = atomSets;
3064
3065     return dd;
3066 }
3067
3068 DomainDecompositionBuilder::DomainDecompositionBuilder(const MDLogger&      mdlog,
3069                                                        t_commrec*           cr,
3070                                                        const DomdecOptions& options,
3071                                                        const MdrunOptions&  mdrunOptions,
3072                                                        const gmx_mtop_t&    mtop,
3073                                                        const t_inputrec&    ir,
3074                                                        const matrix         box,
3075                                                        ArrayRef<const RVec> xGlobal) :
3076     impl_(new Impl(mdlog, cr, options, mdrunOptions, mtop, ir, box, xGlobal))
3077 {
3078 }
3079
3080 gmx_domdec_t* DomainDecompositionBuilder::build(LocalAtomSetManager* atomSets)
3081 {
3082     return impl_->build(atomSets);
3083 }
3084
3085 DomainDecompositionBuilder::~DomainDecompositionBuilder() = default;
3086
3087 } // namespace gmx
3088
3089 static gmx_bool test_dd_cutoff(const t_commrec* cr, const matrix box, gmx::ArrayRef<const gmx::RVec> x, real cutoffRequested)
3090 {
3091     gmx_ddbox_t ddbox;
3092     int         LocallyLimited = 0;
3093
3094     const auto* dd = cr->dd;
3095
3096     set_ddbox(*dd, false, box, true, x, &ddbox);
3097
3098     LocallyLimited = 0;
3099
3100     for (int d = 0; d < dd->ndim; d++)
3101     {
3102         const int dim = dd->dim[d];
3103
3104         real inv_cell_size = DD_CELL_MARGIN * dd->numCells[dim] / ddbox.box_size[dim];
3105         if (dd->unitCellInfo.ddBoxIsDynamic)
3106         {
3107             inv_cell_size *= DD_PRES_SCALE_MARGIN;
3108         }
3109
3110         const int np = 1 + static_cast<int>(cutoffRequested * inv_cell_size * ddbox.skew_fac[dim]);
3111
3112         if (!isDlbDisabled(dd->comm) && (dim < ddbox.npbcdim) && (dd->comm->cd[d].np_dlb > 0))
3113         {
3114             if (np > dd->comm->cd[d].np_dlb)
3115             {
3116                 return FALSE;
3117             }
3118
3119             /* If a current local cell size is smaller than the requested
3120              * cut-off, we could still fix it, but this gets very complicated.
3121              * Without fixing here, we might actually need more checks.
3122              */
3123             real cellSizeAlongDim =
3124                     (dd->comm->cell_x1[dim] - dd->comm->cell_x0[dim]) * ddbox.skew_fac[dim];
3125             if (cellSizeAlongDim * dd->comm->cd[d].np_dlb < cutoffRequested)
3126             {
3127                 LocallyLimited = 1;
3128             }
3129         }
3130     }
3131
3132     if (!isDlbDisabled(dd->comm))
3133     {
3134         /* If DLB is not active yet, we don't need to check the grid jumps.
3135          * Actually we shouldn't, because then the grid jump data is not set.
3136          */
3137         if (isDlbOn(dd->comm) && gmx::check_grid_jump(0, dd, cutoffRequested, &ddbox, FALSE))
3138         {
3139             LocallyLimited = 1;
3140         }
3141
3142         gmx_sumi(1, &LocallyLimited, cr);
3143
3144         if (LocallyLimited > 0)
3145         {
3146             return FALSE;
3147         }
3148     }
3149
3150     return TRUE;
3151 }
3152
3153 bool change_dd_cutoff(t_commrec* cr, const matrix box, gmx::ArrayRef<const gmx::RVec> x, real cutoffRequested)
3154 {
3155     bool bCutoffAllowed = test_dd_cutoff(cr, box, x, cutoffRequested);
3156
3157     if (bCutoffAllowed)
3158     {
3159         cr->dd->comm->systemInfo.cutoff = cutoffRequested;
3160     }
3161
3162     return bCutoffAllowed;
3163 }
3164
3165 void constructGpuHaloExchange(const gmx::MDLogger&            mdlog,
3166                               const t_commrec&                cr,
3167                               const gmx::DeviceStreamManager& deviceStreamManager,
3168                               gmx_wallcycle*                  wcycle)
3169 {
3170     GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
3171                        "Local non-bonded stream should be valid when using"
3172                        "GPU halo exchange.");
3173     GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
3174                        "Non-local non-bonded stream should be valid when using "
3175                        "GPU halo exchange.");
3176
3177     if (cr.dd->gpuHaloExchange[0].empty())
3178     {
3179         GMX_LOG(mdlog.warning)
3180                 .asParagraph()
3181                 .appendTextFormatted(
3182                         "NOTE: Activating the 'GPU halo exchange' feature, enabled "
3183                         "by the "
3184                         "GMX_GPU_DD_COMMS environment variable.");
3185     }
3186
3187     for (int d = 0; d < cr.dd->ndim; d++)
3188     {
3189         for (int pulse = cr.dd->gpuHaloExchange[d].size(); pulse < cr.dd->comm->cd[d].numPulses(); pulse++)
3190         {
3191             cr.dd->gpuHaloExchange[d].push_back(std::make_unique<gmx::GpuHaloExchange>(
3192                     cr.dd,
3193                     d,
3194                     cr.mpi_comm_mygroup,
3195                     deviceStreamManager.context(),
3196                     deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal),
3197                     deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal),
3198                     pulse,
3199                     wcycle));
3200         }
3201     }
3202 }
3203
3204 void reinitGpuHaloExchange(const t_commrec&              cr,
3205                            const DeviceBuffer<gmx::RVec> d_coordinatesBuffer,
3206                            const DeviceBuffer<gmx::RVec> d_forcesBuffer)
3207 {
3208     for (int d = 0; d < cr.dd->ndim; d++)
3209     {
3210         for (int pulse = 0; pulse < cr.dd->comm->cd[d].numPulses(); pulse++)
3211         {
3212             cr.dd->gpuHaloExchange[d][pulse]->reinitHalo(d_coordinatesBuffer, d_forcesBuffer);
3213         }
3214     }
3215 }
3216
3217 void communicateGpuHaloCoordinates(const t_commrec&      cr,
3218                                    const matrix          box,
3219                                    GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
3220 {
3221     for (int d = 0; d < cr.dd->ndim; d++)
3222     {
3223         for (int pulse = 0; pulse < cr.dd->comm->cd[d].numPulses(); pulse++)
3224         {
3225             cr.dd->gpuHaloExchange[d][pulse]->communicateHaloCoordinates(box, coordinatesReadyOnDeviceEvent);
3226         }
3227     }
3228 }
3229
3230 void communicateGpuHaloForces(const t_commrec& cr, bool accumulateForces)
3231 {
3232     for (int d = cr.dd->ndim - 1; d >= 0; d--)
3233     {
3234         for (int pulse = cr.dd->comm->cd[d].numPulses() - 1; pulse >= 0; pulse--)
3235         {
3236             cr.dd->gpuHaloExchange[d][pulse]->communicateHaloForces(accumulateForces);
3237         }
3238     }
3239 }