src/gromacs/ewald/pme.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016,2017 The GROMACS development team.
   7  * Copyright (c) 2018,2019,2020,2021, 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 /*! \internal \file
  39  *
  40  * \brief This file contains function definitions necessary for
  41  * computing energies and forces for the PME long-ranged part (Coulomb
  42  * and LJ).
  43  *
  44  * \author Erik Lindahl <erik@kth.se>
  45  * \author Berk Hess <hess@kth.se>
  46  * \ingroup module_ewald
  47  */
  48 /* IMPORTANT FOR DEVELOPERS:
  49  *
  50  * Triclinic pme stuff isn't entirely trivial, and we've experienced
  51  * some bugs during development (many of them due to me). To avoid
  52  * this in the future, please check the following things if you make
  53  * changes in this file:
  54  *
  55  * 1. You should obtain identical (at least to the PME precision)
  56  *    energies, forces, and virial for
  57  *    a rectangular box and a triclinic one where the z (or y) axis is
  58  *    tilted a whole box side. For instance you could use these boxes:
  59  *
  60  *    rectangular       triclinic
  61  *     2  0  0           2  0  0
  62  *     0  2  0           0  2  0
  63  *     0  0  6           2  2  6
  64  *
  65  * 2. You should check the energy conservation in a triclinic box.
  66  *
  67  * It might seem an overkill, but better safe than sorry.
  68  * /Erik 001109
  69  */
  70
  71 #include "gmxpre.h"
  72
  73 #include "pme.h"
  74
  75 #include "config.h"
  76
  77 #include <cassert>
  78 #include <cmath>
  79 #include <cstdio>
  80 #include <cstdlib>
  81 #include <cstring>
  82
  83 #include <algorithm>
  84 #include <list>
  85
  86 #include "gromacs/domdec/domdec.h"
  87 #include "gromacs/ewald/ewald_utils.h"
  88 #include "gromacs/fft/parallel_3dfft.h"
  89 #include "gromacs/fileio/pdbio.h"
  90 #include "gromacs/gmxlib/network.h"
  91 #include "gromacs/gmxlib/nrnb.h"
  92 #include "gromacs/hardware/hw_info.h"
  93 #include "gromacs/math/gmxcomplex.h"
  94 #include "gromacs/math/invertmatrix.h"
  95 #include "gromacs/math/units.h"
  96 #include "gromacs/math/vec.h"
  97 #include "gromacs/math/vectypes.h"
  98 #include "gromacs/mdtypes/commrec.h"
  99 #include "gromacs/mdtypes/forcerec.h"
 100 #include "gromacs/mdtypes/inputrec.h"
 101 #include "gromacs/mdtypes/md_enums.h"
 102 #include "gromacs/mdtypes/simulation_workload.h"
 103 #include "gromacs/pbcutil/pbc.h"
 104 #include "gromacs/timing/cyclecounter.h"
 105 #include "gromacs/timing/wallcycle.h"
 106 #include "gromacs/timing/walltime_accounting.h"
 107 #include "gromacs/topology/topology.h"
 108 #include "gromacs/utility/basedefinitions.h"
 109 #include "gromacs/utility/cstringutil.h"
 110 #include "gromacs/utility/exceptions.h"
 111 #include "gromacs/utility/fatalerror.h"
 112 #include "gromacs/utility/gmxmpi.h"
 113 #include "gromacs/utility/gmxomp.h"
 114 #include "gromacs/utility/logger.h"
 115 #include "gromacs/utility/real.h"
 116 #include "gromacs/utility/smalloc.h"
 117 #include "gromacs/utility/message_string_collector.h"
 118 #include "gromacs/utility/unique_cptr.h"
 119
 120 #include "calculate_spline_moduli.h"
 121 #include "pme_gather.h"
 122 #include "pme_gpu_internal.h"
 123 #include "pme_grid.h"
 124 #include "pme_internal.h"
 125 #include "pme_redistribute.h"
 126 #include "pme_solve.h"
 127 #include "pme_spline_work.h"
 128 #include "pme_spread.h"
 129
 130 bool pme_gpu_supports_build(std::string* error)
 131 {
 132     gmx::MessageStringCollector errorReasons;
 133     // Before changing the prefix string, make sure that it is not searched for in regression tests.
 134     errorReasons.startContext("PME GPU does not support:");
 135     errorReasons.appendIf(GMX_DOUBLE, "Double-precision build of GROMACS.");
 136     errorReasons.appendIf(!GMX_GPU, "Non-GPU build of GROMACS.");
 137     errorReasons.appendIf(GMX_GPU_SYCL, "SYCL build."); // SYCL-TODO
 138     errorReasons.finishContext();
 139     if (error != nullptr)
 140     {
 141         *error = errorReasons.toString();
 142     }
 143     return errorReasons.isEmpty();
 144 }
 145
 146 bool pme_gpu_supports_hardware(const gmx_hw_info_t gmx_unused& hwinfo, std::string* error)
 147 {
 148     gmx::MessageStringCollector errorReasons;
 149     // Before changing the prefix string, make sure that it is not searched for in regression tests.
 150     errorReasons.startContext("PME GPU does not support:");
 151 #ifdef __APPLE__
 152     errorReasons.appendIf(GMX_GPU_OPENCL, "Apple OS X operating system");
 153 #endif
 154     errorReasons.finishContext();
 155     if (error != nullptr)
 156     {
 157         *error = errorReasons.toString();
 158     }
 159     return errorReasons.isEmpty();
 160 }
 161
 162 bool pme_gpu_supports_input(const t_inputrec& ir, std::string* error)
 163 {
 164     gmx::MessageStringCollector errorReasons;
 165     // Before changing the prefix string, make sure that it is not searched for in regression tests.
 166     errorReasons.startContext("PME GPU does not support:");
 167     errorReasons.appendIf(!EEL_PME(ir.coulombtype),
 168                           "Systems that do not use PME for electrostatics.");
 169     errorReasons.appendIf((ir.pme_order != 4), "Interpolation orders other than 4.");
 170     errorReasons.appendIf(EVDW_PME(ir.vdwtype), "Lennard-Jones PME.");
 171     errorReasons.appendIf(!EI_DYNAMICS(ir.eI), "Non-dynamical integrator (use md, sd, etc).");
 172     errorReasons.finishContext();
 173     if (error != nullptr)
 174     {
 175         *error = errorReasons.toString();
 176     }
 177     return errorReasons.isEmpty();
 178 }
 179
 180 /*! \brief \libinternal
 181  * Finds out if PME with given inputs is possible to run on GPU.
 182  * This function is an internal final check, validating the whole PME structure on creation,
 183  * but it still duplicates the preliminary checks from the above (externally exposed) pme_gpu_supports_input() - just in case.
 184  *
 185  * \param[in]  pme          The PME structure.
 186  * \param[out] error        The error message if the input is not supported on GPU.
 187  * \returns                 True if this PME input is possible to run on GPU, false otherwise.
 188  */
 189 static bool pme_gpu_check_restrictions(const gmx_pme_t* pme, std::string* error)
 190 {
 191     gmx::MessageStringCollector errorReasons;
 192     // Before changing the prefix string, make sure that it is not searched for in regression tests.
 193     errorReasons.startContext("PME GPU does not support:");
 194     errorReasons.appendIf((pme->nnodes != 1), "PME decomposition.");
 195     errorReasons.appendIf((pme->pme_order != 4), "interpolation orders other than 4.");
 196     errorReasons.appendIf(pme->doLJ, "Lennard-Jones PME.");
 197     errorReasons.appendIf(GMX_DOUBLE, "Double precision build of GROMACS.");
 198     errorReasons.appendIf(!GMX_GPU, "Non-GPU build of GROMACS.");
 199     errorReasons.appendIf(GMX_GPU_SYCL, "SYCL build of GROMACS."); // SYCL-TODO
 200     errorReasons.finishContext();
 201     if (error != nullptr)
 202     {
 203         *error = errorReasons.toString();
 204     }
 205     return errorReasons.isEmpty();
 206 }
 207
 208 PmeRunMode pme_run_mode(const gmx_pme_t* pme)
 209 {
 210     GMX_ASSERT(pme != nullptr, "Expecting valid PME data pointer");
 211     return pme->runMode;
 212 }
 213
 214 gmx::PinningPolicy pme_get_pinning_policy()
 215 {
 216     return gmx::PinningPolicy::PinnedIfSupported;
 217 }
 218
 219 /*! \brief Number of bytes in a cache line.
 220  *
 221  * Must also be a multiple of the SIMD and SIMD4 register size, to
 222  * preserve alignment.
 223  */
 224 const int gmxCacheLineSize = 64;
 225
 226 //! Set up coordinate communication
 227 static void setup_coordinate_communication(PmeAtomComm* atc)
 228 {
 229     int nslab, n, i;
 230     int fw, bw;
 231
 232     nslab = atc->nslab;
 233
 234     n = 0;
 235     for (i = 1; i <= nslab / 2; i++)
 236     {
 237         fw = (atc->nodeid + i) % nslab;
 238         bw = (atc->nodeid - i + nslab) % nslab;
 239         if (n < nslab - 1)
 240         {
 241             atc->slabCommSetup[n].node_dest = fw;
 242             atc->slabCommSetup[n].node_src  = bw;
 243             n++;
 244         }
 245         if (n < nslab - 1)
 246         {
 247             atc->slabCommSetup[n].node_dest = bw;
 248             atc->slabCommSetup[n].node_src  = fw;
 249             n++;
 250         }
 251     }
 252 }
 253
 254 /*! \brief Round \p n up to the next multiple of \p f */
 255 static int mult_up(int n, int f)
 256 {
 257     return ((n + f - 1) / f) * f;
 258 }
 259
 260 /*! \brief Return estimate of the load imbalance from the PME grid not being a good match for the number of PME ranks */
 261 static double estimate_pme_load_imbalance(struct gmx_pme_t* pme)
 262 {
 263     int    nma, nmi;
 264     double n1, n2, n3;
 265
 266     nma = pme->nnodes_major;
 267     nmi = pme->nnodes_minor;
 268
 269     n1 = mult_up(pme->nkx, nma) * mult_up(pme->nky, nmi) * pme->nkz;
 270     n2 = mult_up(pme->nkx, nma) * mult_up(pme->nkz, nmi) * pme->nky;
 271     n3 = mult_up(pme->nky, nma) * mult_up(pme->nkz, nmi) * pme->nkx;
 272
 273     /* pme_solve is roughly double the cost of an fft */
 274
 275     return (n1 + n2 + 3 * n3) / static_cast<double>(6 * pme->nkx * pme->nky * pme->nkz);
 276 }
 277
 278 #ifndef DOXYGEN
 279
 280 PmeAtomComm::PmeAtomComm(MPI_Comm   PmeMpiCommunicator,
 281                          const int  numThreads,
 282                          const int  pmeOrder,
 283                          const int  dimIndex,
 284                          const bool doSpread) :
 285     dimind(dimIndex), bSpread(doSpread), pme_order(pmeOrder), nthread(numThreads), spline(nthread)
 286 {
 287     if (PmeMpiCommunicator != MPI_COMM_NULL)
 288     {
 289         mpi_comm = PmeMpiCommunicator;
 290 #    if GMX_MPI
 291         MPI_Comm_size(mpi_comm, &nslab);
 292         MPI_Comm_rank(mpi_comm, &nodeid);
 293 #    endif
 294     }
 295     if (debug)
 296     {
 297         fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", dimind, nslab, nodeid);
 298     }
 299
 300     if (nslab > 1)
 301     {
 302         slabCommSetup.resize(nslab);
 303         setup_coordinate_communication(this);
 304
 305         count_thread.resize(nthread);
 306         for (auto& countThread : count_thread)
 307         {
 308             countThread.resize(nslab);
 309         }
 310     }
 311
 312     if (nthread > 1)
 313     {
 314         threadMap.resize(nthread);
 315
 316 #    pragma omp parallel for num_threads(nthread) schedule(static)
 317         for (int thread = 0; thread < nthread; thread++)
 318         {
 319             try
 320             {
 321                 /* Allocate buffer with padding to avoid cache polution */
 322                 threadMap[thread].nBuffer.resize(nthread + 2 * gmxCacheLineSize);
 323                 threadMap[thread].n = threadMap[thread].nBuffer.data() + gmxCacheLineSize;
 324             }
 325             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 326         }
 327     }
 328 }
 329
 330 #endif // !DOXYGEN
 331
 332 /*! \brief Initialize data structure for communication */
 333 static void init_overlap_comm(pme_overlap_t* ol, int norder, MPI_Comm comm, int nnodes, int nodeid, int ndata, int commplainsize)
 334 {
 335     gmx_bool bCont;
 336
 337     ol->mpi_comm = comm;
 338     ol->nnodes   = nnodes;
 339     ol->nodeid   = nodeid;
 340
 341     /* Linear translation of the PME grid won't affect reciprocal space
 342      * calculations, so to optimize we only interpolate "upwards",
 343      * which also means we only have to consider overlap in one direction.
 344      * I.e., particles on this node might also be spread to grid indices
 345      * that belong to higher nodes (modulo nnodes)
 346      */
 347
 348     ol->s2g0.resize(ol->nnodes + 1);
 349     ol->s2g1.resize(ol->nnodes);
 350     if (debug)
 351     {
 352         fprintf(debug, "PME slab boundaries:");
 353     }
 354     for (int i = 0; i < nnodes; i++)
 355     {
 356         /* s2g0 the local interpolation grid start.
 357          * s2g1 the local interpolation grid end.
 358          * Since in calc_pidx we divide particles, and not grid lines,
 359          * spatially uniform along dimension x or y, we need to round
 360          * s2g0 down and s2g1 up.
 361          */
 362         ol->s2g0[i] = (i * ndata + 0) / nnodes;
 363         ol->s2g1[i] = ((i + 1) * ndata + nnodes - 1) / nnodes + norder - 1;
 364
 365         if (debug)
 366         {
 367             fprintf(debug, "  %3d %3d", ol->s2g0[i], ol->s2g1[i]);
 368         }
 369     }
 370     ol->s2g0[nnodes] = ndata;
 371     if (debug)
 372     {
 373         fprintf(debug, "\n");
 374     }
 375
 376     /* Determine with how many nodes we need to communicate the grid overlap */
 377     int testRankCount = 0;
 378     do
 379     {
 380         testRankCount++;
 381         bCont = FALSE;
 382         for (int i = 0; i < nnodes; i++)
 383         {
 384             if ((i + testRankCount < nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount])
 385                 || (i + testRankCount >= nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount - nnodes] + ndata))
 386             {
 387                 bCont = TRUE;
 388             }
 389         }
 390     } while (bCont && testRankCount < nnodes);
 391
 392     ol->comm_data.resize(testRankCount - 1);
 393     ol->send_size = 0;
 394
 395     for (size_t b = 0; b < ol->comm_data.size(); b++)
 396     {
 397         pme_grid_comm_t* pgc = &ol->comm_data[b];
 398
 399         /* Send */
 400         pgc->send_id  = (ol->nodeid + (b + 1)) % ol->nnodes;
 401         int fft_start = ol->s2g0[pgc->send_id];
 402         int fft_end   = ol->s2g0[pgc->send_id + 1];
 403         if (pgc->send_id < nodeid)
 404         {
 405             fft_start += ndata;
 406             fft_end += ndata;
 407         }
 408         int send_index1  = ol->s2g1[nodeid];
 409         send_index1      = std::min(send_index1, fft_end);
 410         pgc->send_index0 = fft_start;
 411         pgc->send_nindex = std::max(0, send_index1 - pgc->send_index0);
 412         ol->send_size += pgc->send_nindex;
 413
 414         /* We always start receiving to the first index of our slab */
 415         pgc->recv_id    = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
 416         fft_start       = ol->s2g0[ol->nodeid];
 417         fft_end         = ol->s2g0[ol->nodeid + 1];
 418         int recv_index1 = ol->s2g1[pgc->recv_id];
 419         if (pgc->recv_id > nodeid)
 420         {
 421             recv_index1 -= ndata;
 422         }
 423         recv_index1      = std::min(recv_index1, fft_end);
 424         pgc->recv_index0 = fft_start;
 425         pgc->recv_nindex = std::max(0, recv_index1 - pgc->recv_index0);
 426     }
 427
 428 #if GMX_MPI
 429     /* Communicate the buffer sizes to receive */
 430     MPI_Status stat;
 431     for (size_t b = 0; b < ol->comm_data.size(); b++)
 432     {
 433         MPI_Sendrecv(&ol->send_size,
 434                      1,
 435                      MPI_INT,
 436                      ol->comm_data[b].send_id,
 437                      b,
 438                      &ol->comm_data[b].recv_size,
 439                      1,
 440                      MPI_INT,
 441                      ol->comm_data[b].recv_id,
 442                      b,
 443                      ol->mpi_comm,
 444                      &stat);
 445     }
 446 #endif
 447
 448     /* For non-divisible grid we need pme_order iso pme_order-1 */
 449     ol->sendbuf.resize(norder * commplainsize);
 450     ol->recvbuf.resize(norder * commplainsize);
 451 }
 452
 453 int minimalPmeGridSize(int pmeOrder)
 454 {
 455     /* The actual grid size limitations are:
 456      *   serial:        >= pme_order
 457      *   DD, no OpenMP: >= 2*(pme_order - 1)
 458      *   DD, OpenMP:    >= pme_order + 1
 459      * But we use the maximum for simplicity since in practice there is not
 460      * much performance difference between pme_order and 2*(pme_order -1).
 461      */
 462     int minimalSize = 2 * (pmeOrder - 1);
 463
 464     GMX_RELEASE_ASSERT(pmeOrder >= 3, "pmeOrder has to be >= 3");
 465     GMX_RELEASE_ASSERT(minimalSize >= pmeOrder + 1, "The grid size should be >= pmeOrder + 1");
 466
 467     return minimalSize;
 468 }
 469
 470 bool gmx_pme_check_restrictions(int pme_order, int nkx, int nky, int nkz, int numPmeDomainsAlongX, bool useThreads, bool errorsAreFatal)
 471 {
 472     if (pme_order > PME_ORDER_MAX)
 473     {
 474         if (!errorsAreFatal)
 475         {
 476             return false;
 477         }
 478
 479         std::string message = gmx::formatString(
 480                 "pme_order (%d) is larger than the maximum allowed value (%d). Modify and "
 481                 "recompile the code if you really need such a high order.",
 482                 pme_order,
 483                 PME_ORDER_MAX);
 484         GMX_THROW(gmx::InconsistentInputError(message));
 485     }
 486
 487     const int minGridSize = minimalPmeGridSize(pme_order);
 488     if (nkx < minGridSize || nky < minGridSize || nkz < minGridSize)
 489     {
 490         if (!errorsAreFatal)
 491         {
 492             return false;
 493         }
 494         std::string message =
 495                 gmx::formatString("The PME grid sizes need to be >= 2*(pme_order-1) (%d)", minGridSize);
 496         GMX_THROW(gmx::InconsistentInputError(message));
 497     }
 498
 499     /* Check for a limitation of the (current) sum_fftgrid_dd code.
 500      * We only allow multiple communication pulses in dim 1, not in dim 0.
 501      */
 502     if (useThreads
 503         && (nkx < numPmeDomainsAlongX * pme_order && nkx != numPmeDomainsAlongX * (pme_order - 1)))
 504     {
 505         if (!errorsAreFatal)
 506         {
 507             return false;
 508         }
 509         gmx_fatal(FARGS,
 510                   "The number of PME grid lines per rank along x is %g. But when using OpenMP "
 511                   "threads, the number of grid lines per rank along x should be >= pme_order (%d) "
 512                   "or = pmeorder-1. To resolve this issue, use fewer ranks along x (and possibly "
 513                   "more along y and/or z) by specifying -dd manually.",
 514                   nkx / static_cast<double>(numPmeDomainsAlongX),
 515                   pme_order);
 516     }
 517
 518     return true;
 519 }
 520
 521 /*! \brief Round \p enumerator */
 522 static int div_round_up(int enumerator, int denominator)
 523 {
 524     return (enumerator + denominator - 1) / denominator;
 525 }
 526
 527 gmx_pme_t* gmx_pme_init(const t_commrec*     cr,
 528                         const NumPmeDomains& numPmeDomains,
 529                         const t_inputrec*    ir,
 530                         gmx_bool             bFreeEnergy_q,
 531                         gmx_bool             bFreeEnergy_lj,
 532                         gmx_bool             bReproducible,
 533                         real                 ewaldcoeff_q,
 534                         real                 ewaldcoeff_lj,
 535                         int                  nthread,
 536                         PmeRunMode           runMode,
 537                         PmeGpu*              pmeGpu,
 538                         const DeviceContext* deviceContext,
 539                         const DeviceStream*  deviceStream,
 540                         const PmeGpuProgram* pmeGpuProgram,
 541                         const gmx::MDLogger& mdlog)
 542 {
 543     int  use_threads, sum_use_threads, i;
 544     ivec ndata;
 545
 546     if (debug)
 547     {
 548         fprintf(debug, "Creating PME data structures.\n");
 549     }
 550
 551     gmx::unique_cptr<gmx_pme_t, gmx_pme_destroy> pme(new gmx_pme_t());
 552
 553     pme->buf_nalloc = 0;
 554
 555     pme->nnodes  = 1;
 556     pme->bPPnode = TRUE;
 557
 558     pme->nnodes_major = numPmeDomains.x;
 559     pme->nnodes_minor = numPmeDomains.y;
 560
 561     if (numPmeDomains.x * numPmeDomains.y > 1)
 562     {
 563         pme->mpi_comm = cr->mpi_comm_mygroup;
 564
 565 #if GMX_MPI
 566         MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
 567         MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
 568 #endif
 569         if (pme->nnodes != numPmeDomains.x * numPmeDomains.y)
 570         {
 571             gmx_incons("PME rank count mismatch");
 572         }
 573     }
 574     else
 575     {
 576         pme->mpi_comm = MPI_COMM_NULL;
 577     }
 578
 579     if (pme->nnodes == 1)
 580     {
 581         pme->mpi_comm_d[0] = MPI_COMM_NULL;
 582         pme->mpi_comm_d[1] = MPI_COMM_NULL;
 583         pme->ndecompdim    = 0;
 584         pme->nodeid_major  = 0;
 585         pme->nodeid_minor  = 0;
 586     }
 587     else
 588     {
 589         if (numPmeDomains.y == 1)
 590         {
 591             pme->mpi_comm_d[0] = pme->mpi_comm;
 592             pme->mpi_comm_d[1] = MPI_COMM_NULL;
 593             pme->ndecompdim    = 1;
 594             pme->nodeid_major  = pme->nodeid;
 595             pme->nodeid_minor  = 0;
 596         }
 597         else if (numPmeDomains.x == 1)
 598         {
 599             pme->mpi_comm_d[0] = MPI_COMM_NULL;
 600             pme->mpi_comm_d[1] = pme->mpi_comm;
 601             pme->ndecompdim    = 1;
 602             pme->nodeid_major  = 0;
 603             pme->nodeid_minor  = pme->nodeid;
 604         }
 605         else
 606         {
 607             if (pme->nnodes % numPmeDomains.x != 0)
 608             {
 609                 gmx_incons(
 610                         "For 2D PME decomposition, #PME ranks must be divisible by the number of "
 611                         "domains along x");
 612             }
 613             pme->ndecompdim = 2;
 614
 615 #if GMX_MPI
 616             MPI_Comm_split(pme->mpi_comm,
 617                            pme->nodeid % numPmeDomains.y,
 618                            pme->nodeid,
 619                            &pme->mpi_comm_d[0]); /* My communicator along major dimension */
 620             MPI_Comm_split(pme->mpi_comm,
 621                            pme->nodeid / numPmeDomains.y,
 622                            pme->nodeid,
 623                            &pme->mpi_comm_d[1]); /* My communicator along minor dimension */
 624
 625             MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
 626             MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
 627             MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
 628             MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
 629 #endif
 630         }
 631     }
 632     // cr is always initialized if there is a a PP rank, so we can safely assume
 633     // that when it is not, like in ewald tests, we not on a PP rank.
 634     pme->bPPnode = ((cr != nullptr && cr->duty != 0) && thisRankHasDuty(cr, DUTY_PP));
 635
 636     pme->nthread = nthread;
 637
 638     /* Check if any of the PME MPI ranks uses threads */
 639     use_threads = (pme->nthread > 1 ? 1 : 0);
 640 #if GMX_MPI
 641     if (pme->nnodes > 1)
 642     {
 643         MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT, MPI_SUM, pme->mpi_comm);
 644     }
 645     else
 646 #endif
 647     {
 648         sum_use_threads = use_threads;
 649     }
 650     pme->bUseThreads = (sum_use_threads > 0);
 651
 652     if (ir->pbcType == PbcType::Screw)
 653     {
 654         gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
 655     }
 656
 657     /* NOTE:
 658      * It is likely that the current gmx_pme_do() routine supports calculating
 659      * only Coulomb or LJ while gmx_pme_init() configures for both,
 660      * but that has never been tested.
 661      * It is likely that the current gmx_pme_do() routine supports calculating,
 662      * not calculating free-energy for Coulomb and/or LJ while gmx_pme_init()
 663      * configures with free-energy, but that has never been tested.
 664      */
 665     pme->doCoulomb = EEL_PME(ir->coulombtype);
 666     pme->doLJ      = EVDW_PME(ir->vdwtype);
 667     pme->bFEP_q    = ((ir->efep != FreeEnergyPerturbationType::No) && bFreeEnergy_q);
 668     pme->bFEP_lj   = ((ir->efep != FreeEnergyPerturbationType::No) && bFreeEnergy_lj);
 669     pme->bFEP      = (pme->bFEP_q || pme->bFEP_lj);
 670     pme->nkx       = ir->nkx;
 671     pme->nky       = ir->nky;
 672     pme->nkz       = ir->nkz;
 673     pme->bP3M = (ir->coulombtype == CoulombInteractionType::P3mAD || getenv("GMX_PME_P3M") != nullptr);
 674     pme->pme_order     = ir->pme_order;
 675     pme->ewaldcoeff_q  = ewaldcoeff_q;
 676     pme->ewaldcoeff_lj = ewaldcoeff_lj;
 677
 678     /* Always constant electrostatics coefficients */
 679     pme->epsilon_r = ir->epsilon_r;
 680
 681     /* Always constant LJ coefficients */
 682     pme->ljpme_combination_rule = ir->ljpme_combination_rule;
 683
 684     // The box requires scaling with nwalls = 2, we store that condition as well
 685     // as the scaling factor
 686     pme->boxScaler = std::make_unique<EwaldBoxZScaler>(
 687             EwaldBoxZScaler(inputrecPbcXY2Walls(ir), ir->wall_ewald_zfac));
 688
 689     /* If we violate restrictions, generate a fatal error here */
 690     gmx_pme_check_restrictions(
 691             pme->pme_order, pme->nkx, pme->nky, pme->nkz, pme->nnodes_major, pme->bUseThreads, true);
 692
 693     if (pme->nnodes > 1)
 694     {
 695         double imbal;
 696
 697 #if GMX_MPI
 698         MPI_Type_contiguous(DIM, GMX_MPI_REAL, &(pme->rvec_mpi));
 699         MPI_Type_commit(&(pme->rvec_mpi));
 700 #endif
 701
 702         /* Note that the coefficient spreading and force gathering, which usually
 703          * takes about the same amount of time as FFT+solve_pme,
 704          * is always fully load balanced
 705          * (unless the coefficient distribution is inhomogeneous).
 706          */
 707
 708         imbal = estimate_pme_load_imbalance(pme.get());
 709         if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
 710         {
 711             GMX_LOG(mdlog.warning)
 712                     .asParagraph()
 713                     .appendTextFormatted(
 714                             "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
 715                             "      For optimal PME load balancing\n"
 716                             "      PME grid_x (%d) and grid_y (%d) should be divisible by "
 717                             "#PME_ranks_x "
 718                             "(%d)\n"
 719                             "      and PME grid_y (%d) and grid_z (%d) should be divisible by "
 720                             "#PME_ranks_y "
 721                             "(%d)",
 722                             gmx::roundToInt((imbal - 1) * 100),
 723                             pme->nkx,
 724                             pme->nky,
 725                             pme->nnodes_major,
 726                             pme->nky,
 727                             pme->nkz,
 728                             pme->nnodes_minor);
 729         }
 730     }
 731
 732     /* For non-divisible grid we need pme_order iso pme_order-1 */
 733     /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
 734      * y is always copied through a buffer: we don't need padding in z,
 735      * but we do need the overlap in x because of the communication order.
 736      */
 737     init_overlap_comm(&pme->overlap[0],
 738                       pme->pme_order,
 739                       pme->mpi_comm_d[0],
 740                       pme->nnodes_major,
 741                       pme->nodeid_major,
 742                       pme->nkx,
 743                       (div_round_up(pme->nky, pme->nnodes_minor) + pme->pme_order)
 744                               * (pme->nkz + pme->pme_order - 1));
 745
 746     /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
 747      * We do this with an offset buffer of equal size, so we need to allocate
 748      * extra for the offset. That's what the (+1)*pme->nkz is for.
 749      */
 750     init_overlap_comm(&pme->overlap[1],
 751                       pme->pme_order,
 752                       pme->mpi_comm_d[1],
 753                       pme->nnodes_minor,
 754                       pme->nodeid_minor,
 755                       pme->nky,
 756                       (div_round_up(pme->nkx, pme->nnodes_major) + pme->pme_order + 1) * pme->nkz);
 757
 758     /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
 759      * Note that gmx_pme_check_restrictions checked for this already.
 760      */
 761     if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
 762     {
 763         gmx_incons(
 764                 "More than one communication pulse required for grid overlap communication along "
 765                 "the major dimension while using threads");
 766     }
 767
 768     snew(pme->bsp_mod[XX], pme->nkx);
 769     snew(pme->bsp_mod[YY], pme->nky);
 770     snew(pme->bsp_mod[ZZ], pme->nkz);
 771
 772     pme->gpu     = pmeGpu; /* Carrying over the single GPU structure */
 773     pme->runMode = runMode;
 774
 775     /* The required size of the interpolation grid, including overlap.
 776      * The allocated size (pmegrid_n?) might be slightly larger.
 777      */
 778     pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] - pme->overlap[0].s2g0[pme->nodeid_major];
 779     pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] - pme->overlap[1].s2g0[pme->nodeid_minor];
 780     pme->pmegrid_nz_base = pme->nkz;
 781     pme->pmegrid_nz      = pme->pmegrid_nz_base + pme->pme_order - 1;
 782     set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
 783     pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
 784     pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
 785     pme->pmegrid_start_iz = 0;
 786
 787     make_gridindex_to_localindex(
 788             pme->nkx, pme->pmegrid_start_ix, pme->pmegrid_nx - (pme->pme_order - 1), &pme->nnx, &pme->fshx);
 789     make_gridindex_to_localindex(
 790             pme->nky, pme->pmegrid_start_iy, pme->pmegrid_ny - (pme->pme_order - 1), &pme->nny, &pme->fshy);
 791     make_gridindex_to_localindex(
 792             pme->nkz, pme->pmegrid_start_iz, pme->pmegrid_nz_base, &pme->nnz, &pme->fshz);
 793
 794     pme->spline_work = make_pme_spline_work(pme->pme_order);
 795
 796     ndata[0] = pme->nkx;
 797     ndata[1] = pme->nky;
 798     ndata[2] = pme->nkz;
 799     /* It doesn't matter if we allocate too many grids here,
 800      * we only allocate and use the ones we need.
 801      */
 802     if (pme->doLJ)
 803     {
 804         pme->ngrids = ((ir->ljpme_combination_rule == LongRangeVdW::LB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
 805     }
 806     else
 807     {
 808         pme->ngrids = DO_Q;
 809     }
 810     snew(pme->fftgrid, pme->ngrids);
 811     snew(pme->cfftgrid, pme->ngrids);
 812     snew(pme->pfft_setup, pme->ngrids);
 813
 814     for (i = 0; i < pme->ngrids; ++i)
 815     {
 816         if ((i < DO_Q && pme->doCoulomb && (i == 0 || bFreeEnergy_q))
 817             || (i >= DO_Q && pme->doLJ
 818                 && (i == 2 || bFreeEnergy_lj || ir->ljpme_combination_rule == LongRangeVdW::LB)))
 819         {
 820             pmegrids_init(&pme->pmegrid[i],
 821                           pme->pmegrid_nx,
 822                           pme->pmegrid_ny,
 823                           pme->pmegrid_nz,
 824                           pme->pmegrid_nz_base,
 825                           pme->pme_order,
 826                           pme->bUseThreads,
 827                           pme->nthread,
 828                           pme->overlap[0].s2g1[pme->nodeid_major]
 829                                   - pme->overlap[0].s2g0[pme->nodeid_major + 1],
 830                           pme->overlap[1].s2g1[pme->nodeid_minor]
 831                                   - pme->overlap[1].s2g0[pme->nodeid_minor + 1]);
 832             /* This routine will allocate the grid data to fit the FFTs */
 833             const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed)
 834                                                         ? gmx::PinningPolicy::PinnedIfSupported
 835                                                         : gmx::PinningPolicy::CannotBePinned;
 836             gmx_parallel_3dfft_init(&pme->pfft_setup[i],
 837                                     ndata,
 838                                     &pme->fftgrid[i],
 839                                     &pme->cfftgrid[i],
 840                                     pme->mpi_comm_d,
 841                                     bReproducible,
 842                                     pme->nthread,
 843                                     allocateRealGridForGpu);
 844         }
 845     }
 846
 847     if (!pme->bP3M)
 848     {
 849         /* Use plain SPME B-spline interpolation */
 850         make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 851     }
 852     else
 853     {
 854         /* Use the P3M grid-optimized influence function */
 855         make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 856     }
 857
 858     /* Use atc[0] for spreading */
 859     const int firstDimIndex   = (numPmeDomains.x > 1 ? 0 : 1);
 860     MPI_Comm  mpiCommFirstDim = (pme->nnodes > 1 ? pme->mpi_comm_d[firstDimIndex] : MPI_COMM_NULL);
 861     bool      doSpread        = true;
 862     pme->atc.emplace_back(mpiCommFirstDim, pme->nthread, pme->pme_order, firstDimIndex, doSpread);
 863     if (pme->ndecompdim >= 2)
 864     {
 865         const int secondDimIndex = 1;
 866         doSpread                 = false;
 867         pme->atc.emplace_back(pme->mpi_comm_d[1], pme->nthread, pme->pme_order, secondDimIndex, doSpread);
 868     }
 869
 870     // Initial check of validity of the input for running on the GPU
 871     if (pme->runMode != PmeRunMode::CPU)
 872     {
 873         std::string errorString;
 874         bool        canRunOnGpu = pme_gpu_check_restrictions(pme.get(), &errorString);
 875         if (!canRunOnGpu)
 876         {
 877             GMX_THROW(gmx::NotImplementedError(errorString));
 878         }
 879         pme_gpu_reinit(pme.get(), deviceContext, deviceStream, pmeGpuProgram);
 880     }
 881     else
 882     {
 883         GMX_ASSERT(pme->gpu == nullptr, "Should not have PME GPU object when PME is on a CPU.");
 884     }
 885
 886
 887     pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
 888
 889     // no exception was thrown during the init, so we hand over the PME structure handle
 890     return pme.release();
 891 }
 892
 893 void gmx_pme_reinit(struct gmx_pme_t** pmedata,
 894                     const t_commrec*   cr,
 895                     struct gmx_pme_t*  pme_src,
 896                     const t_inputrec*  ir,
 897                     const ivec         grid_size,
 898                     real               ewaldcoeff_q,
 899                     real               ewaldcoeff_lj)
 900 {
 901     // Create a copy of t_inputrec fields that are used in gmx_pme_init().
 902     // TODO: This would be better as just copying a sub-structure that contains
 903     // all the PME parameters and nothing else.
 904     t_inputrec irc;
 905     irc.pbcType                = ir->pbcType;
 906     irc.coulombtype            = ir->coulombtype;
 907     irc.vdwtype                = ir->vdwtype;
 908     irc.efep                   = ir->efep;
 909     irc.pme_order              = ir->pme_order;
 910     irc.epsilon_r              = ir->epsilon_r;
 911     irc.ljpme_combination_rule = ir->ljpme_combination_rule;
 912     irc.nkx                    = grid_size[XX];
 913     irc.nky                    = grid_size[YY];
 914     irc.nkz                    = grid_size[ZZ];
 915
 916     try
 917     {
 918         // This is reinit. Any logging should have been done at first init.
 919         // Here we should avoid writing notes for settings the user did not
 920         // set directly.
 921         const gmx::MDLogger dummyLogger;
 922         GMX_ASSERT(pmedata, "Invalid PME pointer");
 923         NumPmeDomains numPmeDomains = { pme_src->nnodes_major, pme_src->nnodes_minor };
 924         *pmedata                    = gmx_pme_init(cr,
 925                                 numPmeDomains,
 926                                 &irc,
 927                                 pme_src->bFEP_q,
 928                                 pme_src->bFEP_lj,
 929                                 FALSE,
 930                                 ewaldcoeff_q,
 931                                 ewaldcoeff_lj,
 932                                 pme_src->nthread,
 933                                 pme_src->runMode,
 934                                 pme_src->gpu,
 935                                 nullptr,
 936                                 nullptr,
 937                                 nullptr,
 938                                 dummyLogger);
 939         /* When running PME on the CPU not using domain decomposition,
 940          * the atom data is allocated once only in gmx_pme_(re)init().
 941          */
 942         if (!pme_src->gpu && pme_src->nnodes == 1)
 943         {
 944             gmx_pme_reinit_atoms(*pmedata, pme_src->atc[0].numAtoms(), {}, {});
 945         }
 946         // TODO this is mostly passing around current values
 947     }
 948     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 949
 950     /* We can easily reuse the allocated pme grids in pme_src */
 951     reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
 952     /* We would like to reuse the fft grids, but that's harder */
 953 }
 954
 955 real gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q)
 956 {
 957     pmegrids_t* grid;
 958
 959     if (pme->nnodes > 1)
 960     {
 961         gmx_incons("gmx_pme_calc_energy called in parallel");
 962     }
 963     if (pme->bFEP_q)
 964     {
 965         gmx_incons("gmx_pme_calc_energy with free energy");
 966     }
 967
 968     if (!pme->atc_energy)
 969     {
 970         pme->atc_energy = std::make_unique<PmeAtomComm>(MPI_COMM_NULL, 1, pme->pme_order, 0, true);
 971     }
 972     PmeAtomComm* atc = pme->atc_energy.get();
 973     atc->setNumAtoms(x.ssize());
 974     atc->x           = x;
 975     atc->coefficient = q;
 976
 977     /* We only use the A-charges grid */
 978     grid = &pme->pmegrid[PME_GRID_QA];
 979
 980     /* Only calculate the spline coefficients, don't actually spread */
 981     spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
 982
 983     return gather_energy_bsplines(pme, grid->grid.grid, atc);
 984 }
 985
 986 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
 987 static void calc_initial_lb_coeffs(gmx::ArrayRef<real>       coefficient,
 988                                    gmx::ArrayRef<const real> local_c6,
 989                                    gmx::ArrayRef<const real> local_sigma)
 990 {
 991     for (gmx::index i = 0; i < coefficient.ssize(); ++i)
 992     {
 993         real sigma4    = local_sigma[i];
 994         sigma4         = sigma4 * sigma4;
 995         sigma4         = sigma4 * sigma4;
 996         coefficient[i] = local_c6[i] / sigma4;
 997     }
 998 }
 999
1000 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
1001 static void calc_next_lb_coeffs(gmx::ArrayRef<real> coefficient, gmx::ArrayRef<const real> local_sigma)
1002 {
1003     for (gmx::index i = 0; i < coefficient.ssize(); ++i)
1004     {
1005         coefficient[i] *= local_sigma[i];
1006     }
1007 }
1008
1009 int gmx_pme_do(struct gmx_pme_t*              pme,
1010                gmx::ArrayRef<const gmx::RVec> coordinates,
1011                gmx::ArrayRef<gmx::RVec>       forces,
1012                gmx::ArrayRef<const real>      chargeA,
1013                gmx::ArrayRef<const real>      chargeB,
1014                gmx::ArrayRef<const real>      c6A,
1015                gmx::ArrayRef<const real>      c6B,
1016                gmx::ArrayRef<const real>      sigmaA,
1017                gmx::ArrayRef<const real>      sigmaB,
1018                const matrix                   box,
1019                const t_commrec*               cr,
1020                int                            maxshift_x,
1021                int                            maxshift_y,
1022                t_nrnb*                        nrnb,
1023                gmx_wallcycle*                 wcycle,
1024                matrix                         vir_q,
1025                matrix                         vir_lj,
1026                real*                          energy_q,
1027                real*                          energy_lj,
1028                real                           lambda_q,
1029                real                           lambda_lj,
1030                real*                          dvdlambda_q,
1031                real*                          dvdlambda_lj,
1032                const gmx::StepWorkload&       stepWork)
1033 {
1034     GMX_ASSERT(pme->runMode == PmeRunMode::CPU,
1035                "gmx_pme_do should not be called on the GPU PME run.");
1036
1037     PmeAtomComm&              atc     = pme->atc[0];
1038     pmegrids_t*               pmegrid = nullptr;
1039     real*                     grid    = nullptr;
1040     gmx::ArrayRef<const real> coefficient;
1041     std::array<PmeOutput, 2>  output; // The second is used for the B state with FEP
1042     gmx_parallel_3dfft_t      pfft_setup;
1043     real*                     fftgrid;
1044     t_complex*                cfftgrid;
1045     int                       thread;
1046     const int                 fep_states_lj = pme->bFEP_lj ? 2 : 1;
1047     // There's no support for computing energy without virial, or vice versa
1048     const bool computeEnergyAndVirial = (stepWork.computeEnergy || stepWork.computeVirial);
1049
1050     /* We could be passing lambda!=0 while no q or LJ is actually perturbed */
1051     if (!pme->bFEP_q)
1052     {
1053         lambda_q = 0;
1054     }
1055     if (!pme->bFEP_lj)
1056     {
1057         lambda_lj = 0;
1058     }
1059
1060     assert(pme->nnodes > 0);
1061     assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1062
1063     if (pme->nnodes > 1)
1064     {
1065         atc.pd.resize(coordinates.ssize());
1066         for (int d = pme->ndecompdim - 1; d >= 0; d--)
1067         {
1068             PmeAtomComm& atc = pme->atc[d];
1069             atc.maxshift     = (atc.dimind == 0 ? maxshift_x : maxshift_y);
1070         }
1071     }
1072     else
1073     {
1074         GMX_ASSERT(coordinates.ssize() == atc.numAtoms(), "We expect atc.numAtoms() coordinates");
1075         GMX_ASSERT(forces.ssize() >= atc.numAtoms(),
1076                    "We need a force buffer with at least atc.numAtoms() elements");
1077
1078         atc.x = coordinates;
1079         atc.f = forces;
1080     }
1081
1082     matrix scaledBox;
1083     pme->boxScaler->scaleBox(box, scaledBox);
1084
1085     gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1086     bool bFirst = true;
1087
1088     /* For simplicity, we construct the splines for all particles if
1089      * more than one PME calculations is needed. Some optimization
1090      * could be done by keeping track of which atoms have splines
1091      * constructed, and construct new splines on each pass for atoms
1092      * that don't yet have them.
1093      */
1094
1095     bool bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1096
1097     /* We need a maximum of four separate PME calculations:
1098      * grid_index=0: Coulomb PME with charges from state A
1099      * grid_index=1: Coulomb PME with charges from state B
1100      * grid_index=2: LJ PME with C6 from state A
1101      * grid_index=3: LJ PME with C6 from state B
1102      * For Lorentz-Berthelot combination rules, a separate loop is used to
1103      * calculate all the terms
1104      */
1105
1106     /* If we are doing LJ-PME with LB, we only do Q here */
1107     const int max_grid_index = (pme->ljpme_combination_rule == LongRangeVdW::LB) ? DO_Q : DO_Q_AND_LJ;
1108     bool      bClearF;
1109     for (int grid_index = 0; grid_index < max_grid_index; ++grid_index)
1110     {
1111         /* Check if we should do calculations at this grid_index
1112          * If grid_index is odd we should be doing FEP
1113          * If grid_index < 2 we should be doing electrostatic PME
1114          * If grid_index >= 2 we should be doing LJ-PME
1115          */
1116         if ((grid_index < DO_Q && (!pme->doCoulomb || (grid_index == 1 && !pme->bFEP_q)))
1117             || (grid_index >= DO_Q && (!pme->doLJ || (grid_index == 3 && !pme->bFEP_lj))))
1118         {
1119             continue;
1120         }
1121         /* Unpack structure */
1122         pmegrid    = &pme->pmegrid[grid_index];
1123         fftgrid    = pme->fftgrid[grid_index];
1124         cfftgrid   = pme->cfftgrid[grid_index];
1125         pfft_setup = pme->pfft_setup[grid_index];
1126         switch (grid_index)
1127         {
1128             case 0: coefficient = chargeA; break;
1129             case 1: coefficient = chargeB; break;
1130             case 2: coefficient = c6A; break;
1131             case 3: coefficient = c6B; break;
1132         }
1133
1134         grid = pmegrid->grid.grid;
1135
1136         if (pme->nnodes == 1)
1137         {
1138             atc.coefficient = coefficient;
1139         }
1140         else
1141         {
1142             wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1143             do_redist_pos_coeffs(pme, cr, bFirst, coordinates, coefficient);
1144
1145             wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1146         }
1147
1148         wallcycle_start(wcycle, WallCycleCounter::PmeSpread);
1149
1150         /* Spread the coefficients on a grid */
1151         spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1152
1153         if (bFirst)
1154         {
1155             inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1156         }
1157         inc_nrnb(nrnb, eNR_SPREADBSP, pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1158
1159         if (!pme->bUseThreads)
1160         {
1161             wrap_periodic_pmegrid(pme, grid);
1162
1163             /* sum contributions to local grid from other nodes */
1164             if (pme->nnodes > 1)
1165             {
1166                 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1167             }
1168
1169             copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1170         }
1171
1172         wallcycle_stop(wcycle, WallCycleCounter::PmeSpread);
1173
1174         /* TODO If the OpenMP and single-threaded implementations
1175            converge, then spread_on_grid() and
1176            copy_pmegrid_to_fftgrid() will perhaps live in the same
1177            source file.
1178         */
1179
1180         /* Here we start a large thread parallel region */
1181 #pragma omp parallel num_threads(pme->nthread) private(thread)
1182         {
1183             try
1184             {
1185                 thread = gmx_omp_get_thread_num();
1186                 int loop_count;
1187
1188                 /* do 3d-fft */
1189                 if (thread == 0)
1190                 {
1191                     wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1192                 }
1193                 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1194                 if (thread == 0)
1195                 {
1196                     wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1197                 }
1198
1199                 /* solve in k-space for our local cells */
1200                 if (thread == 0)
1201                 {
1202                     wallcycle_start(
1203                             wcycle,
1204                             (grid_index < DO_Q ? WallCycleCounter::PmeSolve : WallCycleCounter::LJPme));
1205                 }
1206                 if (grid_index < DO_Q)
1207                 {
1208                     loop_count = solve_pme_yzx(pme,
1209                                                cfftgrid,
1210                                                scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1211                                                computeEnergyAndVirial,
1212                                                pme->nthread,
1213                                                thread);
1214                 }
1215                 else
1216                 {
1217                     loop_count =
1218                             solve_pme_lj_yzx(pme,
1219                                              &cfftgrid,
1220                                              FALSE,
1221                                              scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1222                                              computeEnergyAndVirial,
1223                                              pme->nthread,
1224                                              thread);
1225                 }
1226
1227                 if (thread == 0)
1228                 {
1229                     wallcycle_stop(
1230                             wcycle,
1231                             (grid_index < DO_Q ? WallCycleCounter::PmeSolve : WallCycleCounter::LJPme));
1232                     inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1233                 }
1234
1235                 /* do 3d-invfft */
1236                 if (thread == 0)
1237                 {
1238                     wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1239                 }
1240                 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1241                 if (thread == 0)
1242                 {
1243                     wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1244
1245
1246                     if (pme->nodeid == 0)
1247                     {
1248                         real      ntot = pme->nkx * pme->nky * pme->nkz;
1249                         const int npme = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1250                         inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1251                     }
1252
1253                     /* Note: this wallcycle region is closed below
1254                        outside an OpenMP region, so take care if
1255                        refactoring code here. */
1256                     wallcycle_start(wcycle, WallCycleCounter::PmeGather);
1257                 }
1258
1259                 copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1260             }
1261             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1262         }
1263         /* End of thread parallel section.
1264          * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1265          */
1266
1267         /* distribute local grid to all nodes */
1268         if (pme->nnodes > 1)
1269         {
1270             gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1271         }
1272
1273         unwrap_periodic_pmegrid(pme, grid);
1274
1275         if (stepWork.computeForces)
1276         {
1277             /* interpolate forces for our local atoms */
1278
1279
1280             /* If we are running without parallelization,
1281              * atc->f is the actual force array, not a buffer,
1282              * therefore we should not clear it.
1283              */
1284             real lambda = grid_index < DO_Q ? lambda_q : lambda_lj;
1285             bClearF     = (bFirst && PAR(cr));
1286 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1287             for (thread = 0; thread < pme->nthread; thread++)
1288             {
1289                 try
1290                 {
1291                     gather_f_bsplines(pme,
1292                                       grid,
1293                                       bClearF,
1294                                       &atc,
1295                                       &atc.spline[thread],
1296                                       pme->bFEP ? (grid_index % 2 == 0 ? 1.0 - lambda : lambda) : 1.0);
1297                 }
1298                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1299             }
1300
1301
1302             inc_nrnb(nrnb, eNR_GATHERFBSP, pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1303             /* Note: this wallcycle region is opened above inside an OpenMP
1304                region, so take care if refactoring code here. */
1305             wallcycle_stop(wcycle, WallCycleCounter::PmeGather);
1306         }
1307
1308         if (computeEnergyAndVirial)
1309         {
1310             /* This should only be called on the master thread
1311              * and after the threads have synchronized.
1312              */
1313             if (grid_index < 2)
1314             {
1315                 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1316             }
1317             else
1318             {
1319                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1320             }
1321         }
1322         bFirst = FALSE;
1323     } /* of grid_index-loop */
1324
1325     /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1326      * seven terms. */
1327
1328     if (pme->doLJ && pme->ljpme_combination_rule == LongRangeVdW::LB)
1329     {
1330         /* Loop over A- and B-state if we are doing FEP */
1331         for (int fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1332         {
1333             std::vector<real>         local_c6;
1334             std::vector<real>         local_sigma;
1335             gmx::ArrayRef<const real> RedistC6;
1336             gmx::ArrayRef<const real> RedistSigma;
1337             gmx::ArrayRef<real>       coefficientBuffer;
1338             if (pme->nnodes == 1)
1339             {
1340                 pme->lb_buf1.resize(atc.numAtoms());
1341                 coefficientBuffer = pme->lb_buf1;
1342                 switch (fep_state)
1343                 {
1344                     case 0:
1345                         local_c6.assign(c6A.begin(), c6A.end());
1346                         local_sigma.assign(sigmaA.begin(), sigmaA.end());
1347                         break;
1348                     case 1:
1349                         local_c6.assign(c6B.begin(), c6B.end());
1350                         local_sigma.assign(sigmaB.begin(), sigmaB.end());
1351                         break;
1352                     default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1353                 }
1354             }
1355             else
1356             {
1357                 coefficientBuffer = atc.coefficientBuffer;
1358                 switch (fep_state)
1359                 {
1360                     case 0:
1361                         RedistC6    = c6A;
1362                         RedistSigma = sigmaA;
1363                         break;
1364                     case 1:
1365                         RedistC6    = c6B;
1366                         RedistSigma = sigmaB;
1367                         break;
1368                     default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1369                 }
1370                 wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1371
1372                 do_redist_pos_coeffs(pme, cr, bFirst, coordinates, RedistC6);
1373                 pme->lb_buf1.resize(atc.numAtoms());
1374                 pme->lb_buf2.resize(atc.numAtoms());
1375                 local_c6.assign(pme->lb_buf1.begin(), pme->lb_buf1.end());
1376                 for (int i = 0; i < atc.numAtoms(); ++i)
1377                 {
1378                     local_c6[i] = atc.coefficient[i];
1379                 }
1380
1381                 do_redist_pos_coeffs(pme, cr, FALSE, coordinates, RedistSigma);
1382                 local_sigma.assign(pme->lb_buf2.begin(), pme->lb_buf2.end());
1383                 for (int i = 0; i < atc.numAtoms(); ++i)
1384                 {
1385                     local_sigma[i] = atc.coefficient[i];
1386                 }
1387
1388                 wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1389             }
1390             atc.coefficient = coefficientBuffer;
1391             calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1392
1393             /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1394             for (int grid_index = 2; grid_index < 9; ++grid_index)
1395             {
1396                 /* Unpack structure */
1397                 pmegrid    = &pme->pmegrid[grid_index];
1398                 fftgrid    = pme->fftgrid[grid_index];
1399                 pfft_setup = pme->pfft_setup[grid_index];
1400                 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1401                 grid = pmegrid->grid.grid;
1402
1403                 wallcycle_start(wcycle, WallCycleCounter::PmeSpread);
1404                 /* Spread the c6 on a grid */
1405                 spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1406
1407                 if (bFirst)
1408                 {
1409                     inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1410                 }
1411
1412                 inc_nrnb(nrnb,
1413                          eNR_SPREADBSP,
1414                          pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1415                 if (pme->nthread == 1)
1416                 {
1417                     wrap_periodic_pmegrid(pme, grid);
1418                     /* sum contributions to local grid from other nodes */
1419                     if (pme->nnodes > 1)
1420                     {
1421                         gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1422                     }
1423                     copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1424                 }
1425                 wallcycle_stop(wcycle, WallCycleCounter::PmeSpread);
1426
1427                 /*Here we start a large thread parallel region*/
1428 #pragma omp parallel num_threads(pme->nthread) private(thread)
1429                 {
1430                     try
1431                     {
1432                         thread = gmx_omp_get_thread_num();
1433                         /* do 3d-fft */
1434                         if (thread == 0)
1435                         {
1436                             wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1437                         }
1438
1439                         gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1440                         if (thread == 0)
1441                         {
1442                             wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1443                         }
1444                     }
1445                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1446                 }
1447                 bFirst = FALSE;
1448             }
1449             /* solve in k-space for our local cells */
1450 #pragma omp parallel num_threads(pme->nthread) private(thread)
1451             {
1452                 try
1453                 {
1454                     int loop_count;
1455                     thread = gmx_omp_get_thread_num();
1456                     if (thread == 0)
1457                     {
1458                         wallcycle_start(wcycle, WallCycleCounter::LJPme);
1459                     }
1460
1461                     loop_count =
1462                             solve_pme_lj_yzx(pme,
1463                                              &pme->cfftgrid[2],
1464                                              TRUE,
1465                                              scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1466                                              computeEnergyAndVirial,
1467                                              pme->nthread,
1468                                              thread);
1469                     if (thread == 0)
1470                     {
1471                         wallcycle_stop(wcycle, WallCycleCounter::LJPme);
1472                         inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1473                     }
1474                 }
1475                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1476             }
1477
1478             if (computeEnergyAndVirial)
1479             {
1480                 /* This should only be called on the master thread and
1481                  * after the threads have synchronized.
1482                  */
1483                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[fep_state]);
1484             }
1485
1486             bFirst = !pme->doCoulomb;
1487             calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1488             for (int grid_index = 8; grid_index >= 2; --grid_index)
1489             {
1490                 /* Unpack structure */
1491                 pmegrid    = &pme->pmegrid[grid_index];
1492                 fftgrid    = pme->fftgrid[grid_index];
1493                 pfft_setup = pme->pfft_setup[grid_index];
1494                 grid       = pmegrid->grid.grid;
1495                 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1496 #pragma omp parallel num_threads(pme->nthread) private(thread)
1497                 {
1498                     try
1499                     {
1500                         thread = gmx_omp_get_thread_num();
1501                         /* do 3d-invfft */
1502                         if (thread == 0)
1503                         {
1504                             wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1505                         }
1506
1507                         gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1508                         if (thread == 0)
1509                         {
1510                             wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1511
1512
1513                             if (pme->nodeid == 0)
1514                             {
1515                                 real      ntot = pme->nkx * pme->nky * pme->nkz;
1516                                 const int npme = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1517                                 inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1518                             }
1519                             wallcycle_start(wcycle, WallCycleCounter::PmeGather);
1520                         }
1521
1522                         copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1523                     }
1524                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1525                 } /*#pragma omp parallel*/
1526
1527                 /* distribute local grid to all nodes */
1528                 if (pme->nnodes > 1)
1529                 {
1530                     gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1531                 }
1532
1533                 unwrap_periodic_pmegrid(pme, grid);
1534
1535                 if (stepWork.computeForces)
1536                 {
1537                     /* interpolate forces for our local atoms */
1538                     bClearF    = (bFirst && PAR(cr));
1539                     real scale = pme->bFEP ? (fep_state < 1 ? 1.0 - lambda_lj : lambda_lj) : 1.0;
1540                     scale *= lb_scale_factor[grid_index - 2];
1541
1542 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1543                     for (thread = 0; thread < pme->nthread; thread++)
1544                     {
1545                         try
1546                         {
1547                             gather_f_bsplines(
1548                                     pme, grid, bClearF, &pme->atc[0], &pme->atc[0].spline[thread], scale);
1549                         }
1550                         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1551                     }
1552
1553
1554                     inc_nrnb(nrnb,
1555                              eNR_GATHERFBSP,
1556                              pme->pme_order * pme->pme_order * pme->pme_order * pme->atc[0].numAtoms());
1557                 }
1558                 wallcycle_stop(wcycle, WallCycleCounter::PmeGather);
1559
1560                 bFirst = FALSE;
1561             } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1562         }     /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1563     }         /* if (pme->doLJ && pme->ljpme_combination_rule == LongRangeVdW::LB) */
1564
1565     if (stepWork.computeForces && pme->nnodes > 1)
1566     {
1567         wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1568         for (int d = 0; d < pme->ndecompdim; d++)
1569         {
1570             gmx::ArrayRef<gmx::RVec> forcesRef;
1571             if (d == pme->ndecompdim - 1)
1572             {
1573                 const size_t numAtoms = coordinates.size();
1574                 GMX_ASSERT(forces.size() >= numAtoms, "Need at least numAtoms forces");
1575                 forcesRef = forces.subArray(0, numAtoms);
1576             }
1577             else
1578             {
1579                 forcesRef = pme->atc[d + 1].f;
1580             }
1581             if (DOMAINDECOMP(cr))
1582             {
1583                 dd_pmeredist_f(pme, &pme->atc[d], forcesRef, d == pme->ndecompdim - 1 && pme->bPPnode);
1584             }
1585         }
1586
1587         wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1588     }
1589
1590     if (computeEnergyAndVirial)
1591     {
1592         if (pme->doCoulomb)
1593         {
1594             if (!pme->bFEP_q)
1595             {
1596                 *energy_q = output[0].coulombEnergy_;
1597                 m_add(vir_q, output[0].coulombVirial_, vir_q);
1598             }
1599             else
1600             {
1601                 *energy_q = (1.0 - lambda_q) * output[0].coulombEnergy_ + lambda_q * output[1].coulombEnergy_;
1602                 *dvdlambda_q += output[1].coulombEnergy_ - output[0].coulombEnergy_;
1603                 for (int i = 0; i < DIM; i++)
1604                 {
1605                     for (int j = 0; j < DIM; j++)
1606                     {
1607                         vir_q[i][j] += (1.0 - lambda_q) * output[0].coulombVirial_[i][j]
1608                                        + lambda_q * output[1].coulombVirial_[i][j];
1609                     }
1610                 }
1611             }
1612         }
1613         else
1614         {
1615             *energy_q = 0;
1616         }
1617
1618         if (pme->doLJ)
1619         {
1620             if (!pme->bFEP_lj)
1621             {
1622                 *energy_lj = output[0].lennardJonesEnergy_;
1623                 m_add(vir_lj, output[0].lennardJonesVirial_, vir_lj);
1624             }
1625             else
1626             {
1627                 *energy_lj = (1.0 - lambda_lj) * output[0].lennardJonesEnergy_
1628                              + lambda_lj * output[1].lennardJonesEnergy_;
1629                 *dvdlambda_lj += output[1].lennardJonesEnergy_ - output[0].lennardJonesEnergy_;
1630                 for (int i = 0; i < DIM; i++)
1631                 {
1632                     for (int j = 0; j < DIM; j++)
1633                     {
1634                         vir_lj[i][j] += (1.0 - lambda_lj) * output[0].lennardJonesVirial_[i][j]
1635                                         + lambda_lj * output[1].lennardJonesVirial_[i][j];
1636                     }
1637                 }
1638             }
1639         }
1640         else
1641         {
1642             *energy_lj = 0;
1643         }
1644     }
1645     return 0;
1646 }
1647
1648 void gmx_pme_destroy(gmx_pme_t* pme)
1649 {
1650     if (!pme)
1651     {
1652         return;
1653     }
1654
1655     sfree(pme->nnx);
1656     sfree(pme->nny);
1657     sfree(pme->nnz);
1658     sfree(pme->fshx);
1659     sfree(pme->fshy);
1660     sfree(pme->fshz);
1661
1662     for (int i = 0; i < pme->ngrids; ++i)
1663     {
1664         pmegrids_destroy(&pme->pmegrid[i]);
1665     }
1666     if (pme->pfft_setup)
1667     {
1668         for (int i = 0; i < pme->ngrids; ++i)
1669         {
1670             gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1671         }
1672     }
1673     sfree(pme->fftgrid);
1674     sfree(pme->cfftgrid);
1675     sfree(pme->pfft_setup);
1676
1677     for (int i = 0; i < DIM; i++)
1678     {
1679         sfree(pme->bsp_mod[i]);
1680     }
1681
1682     sfree(pme->bufv);
1683     sfree(pme->bufr);
1684
1685     if (pme->solve_work)
1686     {
1687         pme_free_all_work(&pme->solve_work, pme->nthread);
1688     }
1689
1690     destroy_pme_spline_work(pme->spline_work);
1691
1692     if (pme->gpu != nullptr)
1693     {
1694         pme_gpu_destroy(pme->gpu);
1695     }
1696
1697     delete pme;
1698 }
1699
1700 void gmx_pme_reinit_atoms(gmx_pme_t*                pme,
1701                           const int                 numAtoms,
1702                           gmx::ArrayRef<const real> chargesA,
1703                           gmx::ArrayRef<const real> chargesB)
1704 {
1705     if (pme->gpu != nullptr)
1706     {
1707         GMX_ASSERT(!(pme->bFEP_q && chargesB.empty()),
1708                    "B state charges must be specified if running Coulomb FEP on the GPU");
1709         pme_gpu_reinit_atoms(pme->gpu, numAtoms, chargesA.data(), pme->bFEP_q ? chargesB.data() : nullptr);
1710     }
1711     else
1712     {
1713         pme->atc[0].setNumAtoms(numAtoms);
1714         // TODO: set the charges here as well
1715     }
1716 }
1717
1718 bool gmx_pme_grid_matches(const gmx_pme_t& pme, const ivec grid_size)
1719 {
1720     return (pme.nkx == grid_size[XX] && pme.nky == grid_size[YY] && pme.nkz == grid_size[ZZ]);
1721 }
1722
1723 void gmx::SeparatePmeRanksPermitted::disablePmeRanks(const std::string& reason)
1724 {
1725     permitSeparatePmeRanks_ = false;
1726
1727     if (!reason.empty())
1728     {
1729         reasons_.push_back(reason);
1730     }
1731 }
1732
1733 bool gmx::SeparatePmeRanksPermitted::permitSeparatePmeRanks() const
1734 {
1735     return permitSeparatePmeRanks_;
1736 }
1737
1738 std::string gmx::SeparatePmeRanksPermitted::reasonsWhyDisabled() const
1739 {
1740     return joinStrings(reasons_, "; ");
1741 }