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 //NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
 131 bool g_allowPmeWithSyclForTesting = false;
 132
 133 bool pme_gpu_supports_build(std::string* error)
 134 {
 135     gmx::MessageStringCollector errorReasons;
 136     // Before changing the prefix string, make sure that it is not searched for in regression tests.
 137     errorReasons.startContext("PME GPU does not support:");
 138     errorReasons.appendIf(GMX_DOUBLE, "Double-precision build of GROMACS.");
 139     errorReasons.appendIf(!GMX_GPU, "Non-GPU build of GROMACS.");
 140     errorReasons.appendIf(GMX_GPU_SYCL && !g_allowPmeWithSyclForTesting, "SYCL build."); // SYCL-TODO
 141     errorReasons.finishContext();
 142     if (error != nullptr)
 143     {
 144         *error = errorReasons.toString();
 145     }
 146     return errorReasons.isEmpty();
 147 }
 148
 149 bool pme_gpu_supports_hardware(const gmx_hw_info_t gmx_unused& hwinfo, std::string* error)
 150 {
 151     gmx::MessageStringCollector errorReasons;
 152     // Before changing the prefix string, make sure that it is not searched for in regression tests.
 153     errorReasons.startContext("PME GPU does not support:");
 154 #ifdef __APPLE__
 155     errorReasons.appendIf(GMX_GPU_OPENCL, "Apple OS X operating system");
 156 #endif
 157     errorReasons.finishContext();
 158     if (error != nullptr)
 159     {
 160         *error = errorReasons.toString();
 161     }
 162     return errorReasons.isEmpty();
 163 }
 164
 165 bool pme_gpu_supports_input(const t_inputrec& ir, std::string* error)
 166 {
 167     gmx::MessageStringCollector errorReasons;
 168     // Before changing the prefix string, make sure that it is not searched for in regression tests.
 169     errorReasons.startContext("PME GPU does not support:");
 170     errorReasons.appendIf(!EEL_PME(ir.coulombtype),
 171                           "Systems that do not use PME for electrostatics.");
 172     errorReasons.appendIf((ir.pme_order != 4), "Interpolation orders other than 4.");
 173     errorReasons.appendIf(EVDW_PME(ir.vdwtype), "Lennard-Jones PME.");
 174     errorReasons.appendIf(!EI_DYNAMICS(ir.eI), "Non-dynamical integrator (use md, sd, etc).");
 175     errorReasons.finishContext();
 176     if (error != nullptr)
 177     {
 178         *error = errorReasons.toString();
 179     }
 180     return errorReasons.isEmpty();
 181 }
 182
 183 /*! \brief \libinternal
 184  * Finds out if PME with given inputs is possible to run on GPU.
 185  * This function is an internal final check, validating the whole PME structure on creation,
 186  * but it still duplicates the preliminary checks from the above (externally exposed) pme_gpu_supports_input() - just in case.
 187  *
 188  * \param[in]  pme          The PME structure.
 189  * \param[out] error        The error message if the input is not supported on GPU.
 190  * \returns                 True if this PME input is possible to run on GPU, false otherwise.
 191  */
 192 static bool pme_gpu_check_restrictions(const gmx_pme_t* pme, std::string* error)
 193 {
 194     gmx::MessageStringCollector errorReasons;
 195     // Before changing the prefix string, make sure that it is not searched for in regression tests.
 196     errorReasons.startContext("PME GPU does not support:");
 197     errorReasons.appendIf((pme->nnodes != 1), "PME decomposition.");
 198     errorReasons.appendIf((pme->pme_order != 4), "interpolation orders other than 4.");
 199     errorReasons.appendIf(pme->doLJ, "Lennard-Jones PME.");
 200     errorReasons.appendIf(GMX_DOUBLE, "Double precision build of GROMACS.");
 201     errorReasons.appendIf(!GMX_GPU, "Non-GPU build of GROMACS.");
 202     errorReasons.appendIf(GMX_GPU_SYCL && !g_allowPmeWithSyclForTesting, "SYCL build of GROMACS."); // SYCL-TODO
 203     errorReasons.finishContext();
 204     if (error != nullptr)
 205     {
 206         *error = errorReasons.toString();
 207     }
 208     return errorReasons.isEmpty();
 209 }
 210
 211 PmeRunMode pme_run_mode(const gmx_pme_t* pme)
 212 {
 213     GMX_ASSERT(pme != nullptr, "Expecting valid PME data pointer");
 214     return pme->runMode;
 215 }
 216
 217 gmx::PinningPolicy pme_get_pinning_policy()
 218 {
 219     return gmx::PinningPolicy::PinnedIfSupported;
 220 }
 221
 222 /*! \brief Number of bytes in a cache line.
 223  *
 224  * Must also be a multiple of the SIMD and SIMD4 register size, to
 225  * preserve alignment.
 226  */
 227 const int gmxCacheLineSize = 64;
 228
 229 //! Set up coordinate communication
 230 static void setup_coordinate_communication(PmeAtomComm* atc)
 231 {
 232     int nslab, n, i;
 233     int fw, bw;
 234
 235     nslab = atc->nslab;
 236
 237     n = 0;
 238     for (i = 1; i <= nslab / 2; i++)
 239     {
 240         fw = (atc->nodeid + i) % nslab;
 241         bw = (atc->nodeid - i + nslab) % nslab;
 242         if (n < nslab - 1)
 243         {
 244             atc->slabCommSetup[n].node_dest = fw;
 245             atc->slabCommSetup[n].node_src  = bw;
 246             n++;
 247         }
 248         if (n < nslab - 1)
 249         {
 250             atc->slabCommSetup[n].node_dest = bw;
 251             atc->slabCommSetup[n].node_src  = fw;
 252             n++;
 253         }
 254     }
 255 }
 256
 257 /*! \brief Round \p n up to the next multiple of \p f */
 258 static int mult_up(int n, int f)
 259 {
 260     return ((n + f - 1) / f) * f;
 261 }
 262
 263 /*! \brief Return estimate of the load imbalance from the PME grid not being a good match for the number of PME ranks */
 264 static double estimate_pme_load_imbalance(struct gmx_pme_t* pme)
 265 {
 266     int    nma, nmi;
 267     double n1, n2, n3;
 268
 269     nma = pme->nnodes_major;
 270     nmi = pme->nnodes_minor;
 271
 272     n1 = mult_up(pme->nkx, nma) * mult_up(pme->nky, nmi) * pme->nkz;
 273     n2 = mult_up(pme->nkx, nma) * mult_up(pme->nkz, nmi) * pme->nky;
 274     n3 = mult_up(pme->nky, nma) * mult_up(pme->nkz, nmi) * pme->nkx;
 275
 276     /* pme_solve is roughly double the cost of an fft */
 277
 278     return (n1 + n2 + 3 * n3) / static_cast<double>(6 * pme->nkx * pme->nky * pme->nkz);
 279 }
 280
 281 #ifndef DOXYGEN
 282
 283 PmeAtomComm::PmeAtomComm(MPI_Comm   PmeMpiCommunicator,
 284                          const int  numThreads,
 285                          const int  pmeOrder,
 286                          const int  dimIndex,
 287                          const bool doSpread) :
 288     dimind(dimIndex), bSpread(doSpread), pme_order(pmeOrder), nthread(numThreads), spline(nthread)
 289 {
 290     if (PmeMpiCommunicator != MPI_COMM_NULL)
 291     {
 292         mpi_comm = PmeMpiCommunicator;
 293 #    if GMX_MPI
 294         MPI_Comm_size(mpi_comm, &nslab);
 295         MPI_Comm_rank(mpi_comm, &nodeid);
 296 #    endif
 297     }
 298     if (debug)
 299     {
 300         fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", dimind, nslab, nodeid);
 301     }
 302
 303     if (nslab > 1)
 304     {
 305         slabCommSetup.resize(nslab);
 306         setup_coordinate_communication(this);
 307
 308         count_thread.resize(nthread);
 309         for (auto& countThread : count_thread)
 310         {
 311             countThread.resize(nslab);
 312         }
 313     }
 314
 315     if (nthread > 1)
 316     {
 317         threadMap.resize(nthread);
 318
 319 #    pragma omp parallel for num_threads(nthread) schedule(static)
 320         for (int thread = 0; thread < nthread; thread++)
 321         {
 322             try
 323             {
 324                 /* Allocate buffer with padding to avoid cache polution */
 325                 threadMap[thread].nBuffer.resize(nthread + 2 * gmxCacheLineSize);
 326                 threadMap[thread].n = threadMap[thread].nBuffer.data() + gmxCacheLineSize;
 327             }
 328             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 329         }
 330     }
 331 }
 332
 333 #endif // !DOXYGEN
 334
 335 /*! \brief Initialize data structure for communication */
 336 static void init_overlap_comm(pme_overlap_t* ol, int norder, MPI_Comm comm, int nnodes, int nodeid, int ndata, int commplainsize)
 337 {
 338     gmx_bool bCont;
 339
 340     ol->mpi_comm = comm;
 341     ol->nnodes   = nnodes;
 342     ol->nodeid   = nodeid;
 343
 344     /* Linear translation of the PME grid won't affect reciprocal space
 345      * calculations, so to optimize we only interpolate "upwards",
 346      * which also means we only have to consider overlap in one direction.
 347      * I.e., particles on this node might also be spread to grid indices
 348      * that belong to higher nodes (modulo nnodes)
 349      */
 350
 351     ol->s2g0.resize(ol->nnodes + 1);
 352     ol->s2g1.resize(ol->nnodes);
 353     if (debug)
 354     {
 355         fprintf(debug, "PME slab boundaries:");
 356     }
 357     for (int i = 0; i < nnodes; i++)
 358     {
 359         /* s2g0 the local interpolation grid start.
 360          * s2g1 the local interpolation grid end.
 361          * Since in calc_pidx we divide particles, and not grid lines,
 362          * spatially uniform along dimension x or y, we need to round
 363          * s2g0 down and s2g1 up.
 364          */
 365         ol->s2g0[i] = (i * ndata + 0) / nnodes;
 366         ol->s2g1[i] = ((i + 1) * ndata + nnodes - 1) / nnodes + norder - 1;
 367
 368         if (debug)
 369         {
 370             fprintf(debug, "  %3d %3d", ol->s2g0[i], ol->s2g1[i]);
 371         }
 372     }
 373     ol->s2g0[nnodes] = ndata;
 374     if (debug)
 375     {
 376         fprintf(debug, "\n");
 377     }
 378
 379     /* Determine with how many nodes we need to communicate the grid overlap */
 380     int testRankCount = 0;
 381     do
 382     {
 383         testRankCount++;
 384         bCont = FALSE;
 385         for (int i = 0; i < nnodes; i++)
 386         {
 387             if ((i + testRankCount < nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount])
 388                 || (i + testRankCount >= nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount - nnodes] + ndata))
 389             {
 390                 bCont = TRUE;
 391             }
 392         }
 393     } while (bCont && testRankCount < nnodes);
 394
 395     ol->comm_data.resize(testRankCount - 1);
 396     ol->send_size = 0;
 397
 398     for (size_t b = 0; b < ol->comm_data.size(); b++)
 399     {
 400         pme_grid_comm_t* pgc = &ol->comm_data[b];
 401
 402         /* Send */
 403         pgc->send_id  = (ol->nodeid + (b + 1)) % ol->nnodes;
 404         int fft_start = ol->s2g0[pgc->send_id];
 405         int fft_end   = ol->s2g0[pgc->send_id + 1];
 406         if (pgc->send_id < nodeid)
 407         {
 408             fft_start += ndata;
 409             fft_end += ndata;
 410         }
 411         int send_index1  = ol->s2g1[nodeid];
 412         send_index1      = std::min(send_index1, fft_end);
 413         pgc->send_index0 = fft_start;
 414         pgc->send_nindex = std::max(0, send_index1 - pgc->send_index0);
 415         ol->send_size += pgc->send_nindex;
 416
 417         /* We always start receiving to the first index of our slab */
 418         pgc->recv_id    = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
 419         fft_start       = ol->s2g0[ol->nodeid];
 420         fft_end         = ol->s2g0[ol->nodeid + 1];
 421         int recv_index1 = ol->s2g1[pgc->recv_id];
 422         if (pgc->recv_id > nodeid)
 423         {
 424             recv_index1 -= ndata;
 425         }
 426         recv_index1      = std::min(recv_index1, fft_end);
 427         pgc->recv_index0 = fft_start;
 428         pgc->recv_nindex = std::max(0, recv_index1 - pgc->recv_index0);
 429     }
 430
 431 #if GMX_MPI
 432     /* Communicate the buffer sizes to receive */
 433     MPI_Status stat;
 434     for (size_t b = 0; b < ol->comm_data.size(); b++)
 435     {
 436         MPI_Sendrecv(&ol->send_size,
 437                      1,
 438                      MPI_INT,
 439                      ol->comm_data[b].send_id,
 440                      b,
 441                      &ol->comm_data[b].recv_size,
 442                      1,
 443                      MPI_INT,
 444                      ol->comm_data[b].recv_id,
 445                      b,
 446                      ol->mpi_comm,
 447                      &stat);
 448     }
 449 #endif
 450
 451     /* For non-divisible grid we need pme_order iso pme_order-1 */
 452     ol->sendbuf.resize(norder * commplainsize);
 453     ol->recvbuf.resize(norder * commplainsize);
 454 }
 455
 456 int minimalPmeGridSize(int pmeOrder)
 457 {
 458     /* The actual grid size limitations are:
 459      *   serial:        >= pme_order
 460      *   DD, no OpenMP: >= 2*(pme_order - 1)
 461      *   DD, OpenMP:    >= pme_order + 1
 462      * But we use the maximum for simplicity since in practice there is not
 463      * much performance difference between pme_order and 2*(pme_order -1).
 464      */
 465     int minimalSize = 2 * (pmeOrder - 1);
 466
 467     GMX_RELEASE_ASSERT(pmeOrder >= 3, "pmeOrder has to be >= 3");
 468     GMX_RELEASE_ASSERT(minimalSize >= pmeOrder + 1, "The grid size should be >= pmeOrder + 1");
 469
 470     return minimalSize;
 471 }
 472
 473 bool gmx_pme_check_restrictions(int pme_order, int nkx, int nky, int nkz, int numPmeDomainsAlongX, bool useThreads, bool errorsAreFatal)
 474 {
 475     if (pme_order > PME_ORDER_MAX)
 476     {
 477         if (!errorsAreFatal)
 478         {
 479             return false;
 480         }
 481
 482         std::string message = gmx::formatString(
 483                 "pme_order (%d) is larger than the maximum allowed value (%d). Modify and "
 484                 "recompile the code if you really need such a high order.",
 485                 pme_order,
 486                 PME_ORDER_MAX);
 487         GMX_THROW(gmx::InconsistentInputError(message));
 488     }
 489
 490     const int minGridSize = minimalPmeGridSize(pme_order);
 491     if (nkx < minGridSize || nky < minGridSize || nkz < minGridSize)
 492     {
 493         if (!errorsAreFatal)
 494         {
 495             return false;
 496         }
 497         std::string message =
 498                 gmx::formatString("The PME grid sizes need to be >= 2*(pme_order-1) (%d)", minGridSize);
 499         GMX_THROW(gmx::InconsistentInputError(message));
 500     }
 501
 502     /* Check for a limitation of the (current) sum_fftgrid_dd code.
 503      * We only allow multiple communication pulses in dim 1, not in dim 0.
 504      */
 505     if (useThreads
 506         && (nkx < numPmeDomainsAlongX * pme_order && nkx != numPmeDomainsAlongX * (pme_order - 1)))
 507     {
 508         if (!errorsAreFatal)
 509         {
 510             return false;
 511         }
 512         gmx_fatal(FARGS,
 513                   "The number of PME grid lines per rank along x is %g. But when using OpenMP "
 514                   "threads, the number of grid lines per rank along x should be >= pme_order (%d) "
 515                   "or = pmeorder-1. To resolve this issue, use fewer ranks along x (and possibly "
 516                   "more along y and/or z) by specifying -dd manually.",
 517                   nkx / static_cast<double>(numPmeDomainsAlongX),
 518                   pme_order);
 519     }
 520
 521     return true;
 522 }
 523
 524 /*! \brief Round \p enumerator */
 525 static int div_round_up(int enumerator, int denominator)
 526 {
 527     return (enumerator + denominator - 1) / denominator;
 528 }
 529
 530 gmx_pme_t* gmx_pme_init(const t_commrec*     cr,
 531                         const NumPmeDomains& numPmeDomains,
 532                         const t_inputrec*    ir,
 533                         gmx_bool             bFreeEnergy_q,
 534                         gmx_bool             bFreeEnergy_lj,
 535                         gmx_bool             bReproducible,
 536                         real                 ewaldcoeff_q,
 537                         real                 ewaldcoeff_lj,
 538                         int                  nthread,
 539                         PmeRunMode           runMode,
 540                         PmeGpu*              pmeGpu,
 541                         const DeviceContext* deviceContext,
 542                         const DeviceStream*  deviceStream,
 543                         const PmeGpuProgram* pmeGpuProgram,
 544                         const gmx::MDLogger& mdlog)
 545 {
 546     int  use_threads, sum_use_threads, i;
 547     ivec ndata;
 548
 549     if (debug)
 550     {
 551         fprintf(debug, "Creating PME data structures.\n");
 552     }
 553
 554     gmx::unique_cptr<gmx_pme_t, gmx_pme_destroy> pme(new gmx_pme_t());
 555
 556     pme->buf_nalloc = 0;
 557
 558     pme->nnodes  = 1;
 559     pme->bPPnode = TRUE;
 560
 561     pme->nnodes_major = numPmeDomains.x;
 562     pme->nnodes_minor = numPmeDomains.y;
 563
 564     if (numPmeDomains.x * numPmeDomains.y > 1)
 565     {
 566         pme->mpi_comm = cr->mpi_comm_mygroup;
 567
 568 #if GMX_MPI
 569         MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
 570         MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
 571 #endif
 572         if (pme->nnodes != numPmeDomains.x * numPmeDomains.y)
 573         {
 574             gmx_incons("PME rank count mismatch");
 575         }
 576     }
 577     else
 578     {
 579         pme->mpi_comm = MPI_COMM_NULL;
 580     }
 581
 582     if (pme->nnodes == 1)
 583     {
 584         pme->mpi_comm_d[0] = MPI_COMM_NULL;
 585         pme->mpi_comm_d[1] = MPI_COMM_NULL;
 586         pme->ndecompdim    = 0;
 587         pme->nodeid_major  = 0;
 588         pme->nodeid_minor  = 0;
 589     }
 590     else
 591     {
 592         if (numPmeDomains.y == 1)
 593         {
 594             pme->mpi_comm_d[0] = pme->mpi_comm;
 595             pme->mpi_comm_d[1] = MPI_COMM_NULL;
 596             pme->ndecompdim    = 1;
 597             pme->nodeid_major  = pme->nodeid;
 598             pme->nodeid_minor  = 0;
 599         }
 600         else if (numPmeDomains.x == 1)
 601         {
 602             pme->mpi_comm_d[0] = MPI_COMM_NULL;
 603             pme->mpi_comm_d[1] = pme->mpi_comm;
 604             pme->ndecompdim    = 1;
 605             pme->nodeid_major  = 0;
 606             pme->nodeid_minor  = pme->nodeid;
 607         }
 608         else
 609         {
 610             if (pme->nnodes % numPmeDomains.x != 0)
 611             {
 612                 gmx_incons(
 613                         "For 2D PME decomposition, #PME ranks must be divisible by the number of "
 614                         "domains along x");
 615             }
 616             pme->ndecompdim = 2;
 617
 618 #if GMX_MPI
 619             MPI_Comm_split(pme->mpi_comm,
 620                            pme->nodeid % numPmeDomains.y,
 621                            pme->nodeid,
 622                            &pme->mpi_comm_d[0]); /* My communicator along major dimension */
 623             MPI_Comm_split(pme->mpi_comm,
 624                            pme->nodeid / numPmeDomains.y,
 625                            pme->nodeid,
 626                            &pme->mpi_comm_d[1]); /* My communicator along minor dimension */
 627
 628             MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
 629             MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
 630             MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
 631             MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
 632 #endif
 633         }
 634     }
 635     // cr is always initialized if there is a a PP rank, so we can safely assume
 636     // that when it is not, like in ewald tests, we not on a PP rank.
 637     pme->bPPnode = ((cr != nullptr && cr->duty != 0) && thisRankHasDuty(cr, DUTY_PP));
 638
 639     pme->nthread = nthread;
 640
 641     /* Check if any of the PME MPI ranks uses threads */
 642     use_threads = (pme->nthread > 1 ? 1 : 0);
 643 #if GMX_MPI
 644     if (pme->nnodes > 1)
 645     {
 646         MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT, MPI_SUM, pme->mpi_comm);
 647     }
 648     else
 649 #endif
 650     {
 651         sum_use_threads = use_threads;
 652     }
 653     pme->bUseThreads = (sum_use_threads > 0);
 654
 655     if (ir->pbcType == PbcType::Screw)
 656     {
 657         gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
 658     }
 659
 660     /* NOTE:
 661      * It is likely that the current gmx_pme_do() routine supports calculating
 662      * only Coulomb or LJ while gmx_pme_init() configures for both,
 663      * but that has never been tested.
 664      * It is likely that the current gmx_pme_do() routine supports calculating,
 665      * not calculating free-energy for Coulomb and/or LJ while gmx_pme_init()
 666      * configures with free-energy, but that has never been tested.
 667      */
 668     pme->doCoulomb = EEL_PME(ir->coulombtype);
 669     pme->doLJ      = EVDW_PME(ir->vdwtype);
 670     pme->bFEP_q    = ((ir->efep != FreeEnergyPerturbationType::No) && bFreeEnergy_q);
 671     pme->bFEP_lj   = ((ir->efep != FreeEnergyPerturbationType::No) && bFreeEnergy_lj);
 672     pme->bFEP      = (pme->bFEP_q || pme->bFEP_lj);
 673     pme->nkx       = ir->nkx;
 674     pme->nky       = ir->nky;
 675     pme->nkz       = ir->nkz;
 676     pme->bP3M = (ir->coulombtype == CoulombInteractionType::P3mAD || getenv("GMX_PME_P3M") != nullptr);
 677     pme->pme_order     = ir->pme_order;
 678     pme->ewaldcoeff_q  = ewaldcoeff_q;
 679     pme->ewaldcoeff_lj = ewaldcoeff_lj;
 680
 681     /* Always constant electrostatics coefficients */
 682     pme->epsilon_r = ir->epsilon_r;
 683
 684     /* Always constant LJ coefficients */
 685     pme->ljpme_combination_rule = ir->ljpme_combination_rule;
 686
 687     // The box requires scaling with nwalls = 2, we store that condition as well
 688     // as the scaling factor
 689     pme->boxScaler = std::make_unique<EwaldBoxZScaler>(
 690             EwaldBoxZScaler(inputrecPbcXY2Walls(ir), ir->wall_ewald_zfac));
 691
 692     /* If we violate restrictions, generate a fatal error here */
 693     gmx_pme_check_restrictions(
 694             pme->pme_order, pme->nkx, pme->nky, pme->nkz, pme->nnodes_major, pme->bUseThreads, true);
 695
 696     if (pme->nnodes > 1)
 697     {
 698         double imbal;
 699
 700 #if GMX_MPI
 701         MPI_Type_contiguous(DIM, GMX_MPI_REAL, &(pme->rvec_mpi));
 702         MPI_Type_commit(&(pme->rvec_mpi));
 703 #endif
 704
 705         /* Note that the coefficient spreading and force gathering, which usually
 706          * takes about the same amount of time as FFT+solve_pme,
 707          * is always fully load balanced
 708          * (unless the coefficient distribution is inhomogeneous).
 709          */
 710
 711         imbal = estimate_pme_load_imbalance(pme.get());
 712         if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
 713         {
 714             GMX_LOG(mdlog.warning)
 715                     .asParagraph()
 716                     .appendTextFormatted(
 717                             "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
 718                             "      For optimal PME load balancing\n"
 719                             "      PME grid_x (%d) and grid_y (%d) should be divisible by "
 720                             "#PME_ranks_x "
 721                             "(%d)\n"
 722                             "      and PME grid_y (%d) and grid_z (%d) should be divisible by "
 723                             "#PME_ranks_y "
 724                             "(%d)",
 725                             gmx::roundToInt((imbal - 1) * 100),
 726                             pme->nkx,
 727                             pme->nky,
 728                             pme->nnodes_major,
 729                             pme->nky,
 730                             pme->nkz,
 731                             pme->nnodes_minor);
 732         }
 733     }
 734
 735     /* For non-divisible grid we need pme_order iso pme_order-1 */
 736     /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
 737      * y is always copied through a buffer: we don't need padding in z,
 738      * but we do need the overlap in x because of the communication order.
 739      */
 740     init_overlap_comm(&pme->overlap[0],
 741                       pme->pme_order,
 742                       pme->mpi_comm_d[0],
 743                       pme->nnodes_major,
 744                       pme->nodeid_major,
 745                       pme->nkx,
 746                       (div_round_up(pme->nky, pme->nnodes_minor) + pme->pme_order)
 747                               * (pme->nkz + pme->pme_order - 1));
 748
 749     /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
 750      * We do this with an offset buffer of equal size, so we need to allocate
 751      * extra for the offset. That's what the (+1)*pme->nkz is for.
 752      */
 753     init_overlap_comm(&pme->overlap[1],
 754                       pme->pme_order,
 755                       pme->mpi_comm_d[1],
 756                       pme->nnodes_minor,
 757                       pme->nodeid_minor,
 758                       pme->nky,
 759                       (div_round_up(pme->nkx, pme->nnodes_major) + pme->pme_order + 1) * pme->nkz);
 760
 761     /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
 762      * Note that gmx_pme_check_restrictions checked for this already.
 763      */
 764     if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
 765     {
 766         gmx_incons(
 767                 "More than one communication pulse required for grid overlap communication along "
 768                 "the major dimension while using threads");
 769     }
 770
 771     snew(pme->bsp_mod[XX], pme->nkx);
 772     snew(pme->bsp_mod[YY], pme->nky);
 773     snew(pme->bsp_mod[ZZ], pme->nkz);
 774
 775     pme->gpu     = pmeGpu; /* Carrying over the single GPU structure */
 776     pme->runMode = runMode;
 777
 778     /* The required size of the interpolation grid, including overlap.
 779      * The allocated size (pmegrid_n?) might be slightly larger.
 780      */
 781     pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] - pme->overlap[0].s2g0[pme->nodeid_major];
 782     pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] - pme->overlap[1].s2g0[pme->nodeid_minor];
 783     pme->pmegrid_nz_base = pme->nkz;
 784     pme->pmegrid_nz      = pme->pmegrid_nz_base + pme->pme_order - 1;
 785     set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
 786     pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
 787     pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
 788     pme->pmegrid_start_iz = 0;
 789
 790     make_gridindex_to_localindex(
 791             pme->nkx, pme->pmegrid_start_ix, pme->pmegrid_nx - (pme->pme_order - 1), &pme->nnx, &pme->fshx);
 792     make_gridindex_to_localindex(
 793             pme->nky, pme->pmegrid_start_iy, pme->pmegrid_ny - (pme->pme_order - 1), &pme->nny, &pme->fshy);
 794     make_gridindex_to_localindex(
 795             pme->nkz, pme->pmegrid_start_iz, pme->pmegrid_nz_base, &pme->nnz, &pme->fshz);
 796
 797     pme->spline_work = make_pme_spline_work(pme->pme_order);
 798
 799     ndata[0] = pme->nkx;
 800     ndata[1] = pme->nky;
 801     ndata[2] = pme->nkz;
 802     /* It doesn't matter if we allocate too many grids here,
 803      * we only allocate and use the ones we need.
 804      */
 805     if (pme->doLJ)
 806     {
 807         pme->ngrids = ((ir->ljpme_combination_rule == LongRangeVdW::LB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
 808     }
 809     else
 810     {
 811         pme->ngrids = DO_Q;
 812     }
 813     snew(pme->fftgrid, pme->ngrids);
 814     snew(pme->cfftgrid, pme->ngrids);
 815     snew(pme->pfft_setup, pme->ngrids);
 816
 817     for (i = 0; i < pme->ngrids; ++i)
 818     {
 819         if ((i < DO_Q && pme->doCoulomb && (i == 0 || bFreeEnergy_q))
 820             || (i >= DO_Q && pme->doLJ
 821                 && (i == 2 || bFreeEnergy_lj || ir->ljpme_combination_rule == LongRangeVdW::LB)))
 822         {
 823             pmegrids_init(&pme->pmegrid[i],
 824                           pme->pmegrid_nx,
 825                           pme->pmegrid_ny,
 826                           pme->pmegrid_nz,
 827                           pme->pmegrid_nz_base,
 828                           pme->pme_order,
 829                           pme->bUseThreads,
 830                           pme->nthread,
 831                           pme->overlap[0].s2g1[pme->nodeid_major]
 832                                   - pme->overlap[0].s2g0[pme->nodeid_major + 1],
 833                           pme->overlap[1].s2g1[pme->nodeid_minor]
 834                                   - pme->overlap[1].s2g0[pme->nodeid_minor + 1]);
 835             /* This routine will allocate the grid data to fit the FFTs */
 836             const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed)
 837                                                         ? gmx::PinningPolicy::PinnedIfSupported
 838                                                         : gmx::PinningPolicy::CannotBePinned;
 839             gmx_parallel_3dfft_init(&pme->pfft_setup[i],
 840                                     ndata,
 841                                     &pme->fftgrid[i],
 842                                     &pme->cfftgrid[i],
 843                                     pme->mpi_comm_d,
 844                                     bReproducible,
 845                                     pme->nthread,
 846                                     allocateRealGridForGpu);
 847         }
 848     }
 849
 850     if (!pme->bP3M)
 851     {
 852         /* Use plain SPME B-spline interpolation */
 853         make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 854     }
 855     else
 856     {
 857         /* Use the P3M grid-optimized influence function */
 858         make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 859     }
 860
 861     /* Use atc[0] for spreading */
 862     const int firstDimIndex   = (numPmeDomains.x > 1 ? 0 : 1);
 863     MPI_Comm  mpiCommFirstDim = (pme->nnodes > 1 ? pme->mpi_comm_d[firstDimIndex] : MPI_COMM_NULL);
 864     bool      doSpread        = true;
 865     pme->atc.emplace_back(mpiCommFirstDim, pme->nthread, pme->pme_order, firstDimIndex, doSpread);
 866     if (pme->ndecompdim >= 2)
 867     {
 868         const int secondDimIndex = 1;
 869         doSpread                 = false;
 870         pme->atc.emplace_back(pme->mpi_comm_d[1], pme->nthread, pme->pme_order, secondDimIndex, doSpread);
 871     }
 872
 873     // Initial check of validity of the input for running on the GPU
 874     if (pme->runMode != PmeRunMode::CPU)
 875     {
 876         std::string errorString;
 877         bool        canRunOnGpu = pme_gpu_check_restrictions(pme.get(), &errorString);
 878         if (!canRunOnGpu)
 879         {
 880             GMX_THROW(gmx::NotImplementedError(errorString));
 881         }
 882         pme_gpu_reinit(pme.get(), deviceContext, deviceStream, pmeGpuProgram);
 883     }
 884     else
 885     {
 886         GMX_ASSERT(pme->gpu == nullptr, "Should not have PME GPU object when PME is on a CPU.");
 887     }
 888
 889
 890     pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
 891
 892     // no exception was thrown during the init, so we hand over the PME structure handle
 893     return pme.release();
 894 }
 895
 896 void gmx_pme_reinit(struct gmx_pme_t** pmedata,
 897                     const t_commrec*   cr,
 898                     struct gmx_pme_t*  pme_src,
 899                     const t_inputrec*  ir,
 900                     const ivec         grid_size,
 901                     real               ewaldcoeff_q,
 902                     real               ewaldcoeff_lj)
 903 {
 904     // Create a copy of t_inputrec fields that are used in gmx_pme_init().
 905     // TODO: This would be better as just copying a sub-structure that contains
 906     // all the PME parameters and nothing else.
 907     t_inputrec irc;
 908     irc.pbcType                = ir->pbcType;
 909     irc.coulombtype            = ir->coulombtype;
 910     irc.vdwtype                = ir->vdwtype;
 911     irc.efep                   = ir->efep;
 912     irc.pme_order              = ir->pme_order;
 913     irc.epsilon_r              = ir->epsilon_r;
 914     irc.ljpme_combination_rule = ir->ljpme_combination_rule;
 915     irc.nkx                    = grid_size[XX];
 916     irc.nky                    = grid_size[YY];
 917     irc.nkz                    = grid_size[ZZ];
 918
 919     try
 920     {
 921         // This is reinit. Any logging should have been done at first init.
 922         // Here we should avoid writing notes for settings the user did not
 923         // set directly.
 924         const gmx::MDLogger dummyLogger;
 925         GMX_ASSERT(pmedata, "Invalid PME pointer");
 926         NumPmeDomains numPmeDomains = { pme_src->nnodes_major, pme_src->nnodes_minor };
 927         *pmedata                    = gmx_pme_init(cr,
 928                                 numPmeDomains,
 929                                 &irc,
 930                                 pme_src->bFEP_q,
 931                                 pme_src->bFEP_lj,
 932                                 FALSE,
 933                                 ewaldcoeff_q,
 934                                 ewaldcoeff_lj,
 935                                 pme_src->nthread,
 936                                 pme_src->runMode,
 937                                 pme_src->gpu,
 938                                 nullptr,
 939                                 nullptr,
 940                                 nullptr,
 941                                 dummyLogger);
 942         /* When running PME on the CPU not using domain decomposition,
 943          * the atom data is allocated once only in gmx_pme_(re)init().
 944          */
 945         if (!pme_src->gpu && pme_src->nnodes == 1)
 946         {
 947             gmx_pme_reinit_atoms(*pmedata, pme_src->atc[0].numAtoms(), {}, {});
 948         }
 949         // TODO this is mostly passing around current values
 950     }
 951     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 952
 953     /* We can easily reuse the allocated pme grids in pme_src */
 954     reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
 955     /* We would like to reuse the fft grids, but that's harder */
 956 }
 957
 958 real gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q)
 959 {
 960     pmegrids_t* grid;
 961
 962     if (pme->nnodes > 1)
 963     {
 964         gmx_incons("gmx_pme_calc_energy called in parallel");
 965     }
 966     if (pme->bFEP_q)
 967     {
 968         gmx_incons("gmx_pme_calc_energy with free energy");
 969     }
 970
 971     if (!pme->atc_energy)
 972     {
 973         pme->atc_energy = std::make_unique<PmeAtomComm>(MPI_COMM_NULL, 1, pme->pme_order, 0, true);
 974     }
 975     PmeAtomComm* atc = pme->atc_energy.get();
 976     atc->setNumAtoms(x.ssize());
 977     atc->x           = x;
 978     atc->coefficient = q;
 979
 980     /* We only use the A-charges grid */
 981     grid = &pme->pmegrid[PME_GRID_QA];
 982
 983     /* Only calculate the spline coefficients, don't actually spread */
 984     spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
 985
 986     return gather_energy_bsplines(pme, grid->grid.grid, atc);
 987 }
 988
 989 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
 990 static void calc_initial_lb_coeffs(gmx::ArrayRef<real>       coefficient,
 991                                    gmx::ArrayRef<const real> local_c6,
 992                                    gmx::ArrayRef<const real> local_sigma)
 993 {
 994     for (gmx::index i = 0; i < coefficient.ssize(); ++i)
 995     {
 996         real sigma4    = local_sigma[i];
 997         sigma4         = sigma4 * sigma4;
 998         sigma4         = sigma4 * sigma4;
 999         coefficient[i] = local_c6[i] / sigma4;
1000     }
1001 }
1002
1003 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
1004 static void calc_next_lb_coeffs(gmx::ArrayRef<real> coefficient, gmx::ArrayRef<const real> local_sigma)
1005 {
1006     for (gmx::index i = 0; i < coefficient.ssize(); ++i)
1007     {
1008         coefficient[i] *= local_sigma[i];
1009     }
1010 }
1011
1012 int gmx_pme_do(struct gmx_pme_t*              pme,
1013                gmx::ArrayRef<const gmx::RVec> coordinates,
1014                gmx::ArrayRef<gmx::RVec>       forces,
1015                gmx::ArrayRef<const real>      chargeA,
1016                gmx::ArrayRef<const real>      chargeB,
1017                gmx::ArrayRef<const real>      c6A,
1018                gmx::ArrayRef<const real>      c6B,
1019                gmx::ArrayRef<const real>      sigmaA,
1020                gmx::ArrayRef<const real>      sigmaB,
1021                const matrix                   box,
1022                const t_commrec*               cr,
1023                int                            maxshift_x,
1024                int                            maxshift_y,
1025                t_nrnb*                        nrnb,
1026                gmx_wallcycle*                 wcycle,
1027                matrix                         vir_q,
1028                matrix                         vir_lj,
1029                real*                          energy_q,
1030                real*                          energy_lj,
1031                real                           lambda_q,
1032                real                           lambda_lj,
1033                real*                          dvdlambda_q,
1034                real*                          dvdlambda_lj,
1035                const gmx::StepWorkload&       stepWork)
1036 {
1037     GMX_ASSERT(pme->runMode == PmeRunMode::CPU,
1038                "gmx_pme_do should not be called on the GPU PME run.");
1039
1040     PmeAtomComm&              atc     = pme->atc[0];
1041     pmegrids_t*               pmegrid = nullptr;
1042     real*                     grid    = nullptr;
1043     gmx::ArrayRef<const real> coefficient;
1044     std::array<PmeOutput, 2>  output; // The second is used for the B state with FEP
1045     gmx_parallel_3dfft_t      pfft_setup;
1046     real*                     fftgrid;
1047     t_complex*                cfftgrid;
1048     int                       thread;
1049     const int                 fep_states_lj = pme->bFEP_lj ? 2 : 1;
1050     // There's no support for computing energy without virial, or vice versa
1051     const bool computeEnergyAndVirial = (stepWork.computeEnergy || stepWork.computeVirial);
1052
1053     /* We could be passing lambda!=0 while no q or LJ is actually perturbed */
1054     if (!pme->bFEP_q)
1055     {
1056         lambda_q = 0;
1057     }
1058     if (!pme->bFEP_lj)
1059     {
1060         lambda_lj = 0;
1061     }
1062
1063     assert(pme->nnodes > 0);
1064     assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1065
1066     if (pme->nnodes > 1)
1067     {
1068         atc.pd.resize(coordinates.ssize());
1069         for (int d = pme->ndecompdim - 1; d >= 0; d--)
1070         {
1071             PmeAtomComm& atc = pme->atc[d];
1072             atc.maxshift     = (atc.dimind == 0 ? maxshift_x : maxshift_y);
1073         }
1074     }
1075     else
1076     {
1077         GMX_ASSERT(coordinates.ssize() == atc.numAtoms(), "We expect atc.numAtoms() coordinates");
1078         GMX_ASSERT(forces.ssize() >= atc.numAtoms(),
1079                    "We need a force buffer with at least atc.numAtoms() elements");
1080
1081         atc.x = coordinates;
1082         atc.f = forces;
1083     }
1084
1085     matrix scaledBox;
1086     pme->boxScaler->scaleBox(box, scaledBox);
1087
1088     gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1089     bool bFirst = true;
1090
1091     /* For simplicity, we construct the splines for all particles if
1092      * more than one PME calculations is needed. Some optimization
1093      * could be done by keeping track of which atoms have splines
1094      * constructed, and construct new splines on each pass for atoms
1095      * that don't yet have them.
1096      */
1097
1098     bool bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1099
1100     /* We need a maximum of four separate PME calculations:
1101      * grid_index=0: Coulomb PME with charges from state A
1102      * grid_index=1: Coulomb PME with charges from state B
1103      * grid_index=2: LJ PME with C6 from state A
1104      * grid_index=3: LJ PME with C6 from state B
1105      * For Lorentz-Berthelot combination rules, a separate loop is used to
1106      * calculate all the terms
1107      */
1108
1109     /* If we are doing LJ-PME with LB, we only do Q here */
1110     const int max_grid_index = (pme->ljpme_combination_rule == LongRangeVdW::LB) ? DO_Q : DO_Q_AND_LJ;
1111     bool      bClearF;
1112     for (int grid_index = 0; grid_index < max_grid_index; ++grid_index)
1113     {
1114         /* Check if we should do calculations at this grid_index
1115          * If grid_index is odd we should be doing FEP
1116          * If grid_index < 2 we should be doing electrostatic PME
1117          * If grid_index >= 2 we should be doing LJ-PME
1118          */
1119         if ((grid_index < DO_Q && (!pme->doCoulomb || (grid_index == 1 && !pme->bFEP_q)))
1120             || (grid_index >= DO_Q && (!pme->doLJ || (grid_index == 3 && !pme->bFEP_lj))))
1121         {
1122             continue;
1123         }
1124         /* Unpack structure */
1125         pmegrid    = &pme->pmegrid[grid_index];
1126         fftgrid    = pme->fftgrid[grid_index];
1127         cfftgrid   = pme->cfftgrid[grid_index];
1128         pfft_setup = pme->pfft_setup[grid_index];
1129         switch (grid_index)
1130         {
1131             case 0: coefficient = chargeA; break;
1132             case 1: coefficient = chargeB; break;
1133             case 2: coefficient = c6A; break;
1134             case 3: coefficient = c6B; break;
1135         }
1136
1137         grid = pmegrid->grid.grid;
1138
1139         if (pme->nnodes == 1)
1140         {
1141             atc.coefficient = coefficient;
1142         }
1143         else
1144         {
1145             wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1146             do_redist_pos_coeffs(pme, cr, bFirst, coordinates, coefficient);
1147
1148             wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1149         }
1150
1151         wallcycle_start(wcycle, WallCycleCounter::PmeSpread);
1152
1153         /* Spread the coefficients on a grid */
1154         spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1155
1156         if (bFirst)
1157         {
1158             inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1159         }
1160         inc_nrnb(nrnb, eNR_SPREADBSP, pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1161
1162         if (!pme->bUseThreads)
1163         {
1164             wrap_periodic_pmegrid(pme, grid);
1165
1166             /* sum contributions to local grid from other nodes */
1167             if (pme->nnodes > 1)
1168             {
1169                 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1170             }
1171
1172             copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1173         }
1174
1175         wallcycle_stop(wcycle, WallCycleCounter::PmeSpread);
1176
1177         /* TODO If the OpenMP and single-threaded implementations
1178            converge, then spread_on_grid() and
1179            copy_pmegrid_to_fftgrid() will perhaps live in the same
1180            source file.
1181         */
1182
1183         /* Here we start a large thread parallel region */
1184 #pragma omp parallel num_threads(pme->nthread) private(thread)
1185         {
1186             try
1187             {
1188                 thread = gmx_omp_get_thread_num();
1189                 int loop_count;
1190
1191                 /* do 3d-fft */
1192                 if (thread == 0)
1193                 {
1194                     wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1195                 }
1196                 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1197                 if (thread == 0)
1198                 {
1199                     wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1200                 }
1201
1202                 /* solve in k-space for our local cells */
1203                 if (thread == 0)
1204                 {
1205                     wallcycle_start(
1206                             wcycle,
1207                             (grid_index < DO_Q ? WallCycleCounter::PmeSolve : WallCycleCounter::LJPme));
1208                 }
1209                 if (grid_index < DO_Q)
1210                 {
1211                     loop_count = solve_pme_yzx(pme,
1212                                                cfftgrid,
1213                                                scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1214                                                computeEnergyAndVirial,
1215                                                pme->nthread,
1216                                                thread);
1217                 }
1218                 else
1219                 {
1220                     loop_count =
1221                             solve_pme_lj_yzx(pme,
1222                                              &cfftgrid,
1223                                              FALSE,
1224                                              scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1225                                              computeEnergyAndVirial,
1226                                              pme->nthread,
1227                                              thread);
1228                 }
1229
1230                 if (thread == 0)
1231                 {
1232                     wallcycle_stop(
1233                             wcycle,
1234                             (grid_index < DO_Q ? WallCycleCounter::PmeSolve : WallCycleCounter::LJPme));
1235                     inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1236                 }
1237
1238                 /* do 3d-invfft */
1239                 if (thread == 0)
1240                 {
1241                     wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1242                 }
1243                 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1244                 if (thread == 0)
1245                 {
1246                     wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1247
1248
1249                     if (pme->nodeid == 0)
1250                     {
1251                         real      ntot = pme->nkx * pme->nky * pme->nkz;
1252                         const int npme = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1253                         inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1254                     }
1255
1256                     /* Note: this wallcycle region is closed below
1257                        outside an OpenMP region, so take care if
1258                        refactoring code here. */
1259                     wallcycle_start(wcycle, WallCycleCounter::PmeGather);
1260                 }
1261
1262                 copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1263             }
1264             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1265         }
1266         /* End of thread parallel section.
1267          * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1268          */
1269
1270         /* distribute local grid to all nodes */
1271         if (pme->nnodes > 1)
1272         {
1273             gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1274         }
1275
1276         unwrap_periodic_pmegrid(pme, grid);
1277
1278         if (stepWork.computeForces)
1279         {
1280             /* interpolate forces for our local atoms */
1281
1282
1283             /* If we are running without parallelization,
1284              * atc->f is the actual force array, not a buffer,
1285              * therefore we should not clear it.
1286              */
1287             real lambda = grid_index < DO_Q ? lambda_q : lambda_lj;
1288             bClearF     = (bFirst && PAR(cr));
1289 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1290             for (thread = 0; thread < pme->nthread; thread++)
1291             {
1292                 try
1293                 {
1294                     gather_f_bsplines(pme,
1295                                       grid,
1296                                       bClearF,
1297                                       &atc,
1298                                       &atc.spline[thread],
1299                                       pme->bFEP ? (grid_index % 2 == 0 ? 1.0 - lambda : lambda) : 1.0);
1300                 }
1301                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1302             }
1303
1304
1305             inc_nrnb(nrnb, eNR_GATHERFBSP, pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1306             /* Note: this wallcycle region is opened above inside an OpenMP
1307                region, so take care if refactoring code here. */
1308             wallcycle_stop(wcycle, WallCycleCounter::PmeGather);
1309         }
1310
1311         if (computeEnergyAndVirial)
1312         {
1313             /* This should only be called on the master thread
1314              * and after the threads have synchronized.
1315              */
1316             if (grid_index < 2)
1317             {
1318                 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1319             }
1320             else
1321             {
1322                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1323             }
1324         }
1325         bFirst = FALSE;
1326     } /* of grid_index-loop */
1327
1328     /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1329      * seven terms. */
1330
1331     if (pme->doLJ && pme->ljpme_combination_rule == LongRangeVdW::LB)
1332     {
1333         /* Loop over A- and B-state if we are doing FEP */
1334         for (int fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1335         {
1336             std::vector<real>         local_c6;
1337             std::vector<real>         local_sigma;
1338             gmx::ArrayRef<const real> RedistC6;
1339             gmx::ArrayRef<const real> RedistSigma;
1340             gmx::ArrayRef<real>       coefficientBuffer;
1341             if (pme->nnodes == 1)
1342             {
1343                 pme->lb_buf1.resize(atc.numAtoms());
1344                 coefficientBuffer = pme->lb_buf1;
1345                 switch (fep_state)
1346                 {
1347                     case 0:
1348                         local_c6.assign(c6A.begin(), c6A.end());
1349                         local_sigma.assign(sigmaA.begin(), sigmaA.end());
1350                         break;
1351                     case 1:
1352                         local_c6.assign(c6B.begin(), c6B.end());
1353                         local_sigma.assign(sigmaB.begin(), sigmaB.end());
1354                         break;
1355                     default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1356                 }
1357             }
1358             else
1359             {
1360                 coefficientBuffer = atc.coefficientBuffer;
1361                 switch (fep_state)
1362                 {
1363                     case 0:
1364                         RedistC6    = c6A;
1365                         RedistSigma = sigmaA;
1366                         break;
1367                     case 1:
1368                         RedistC6    = c6B;
1369                         RedistSigma = sigmaB;
1370                         break;
1371                     default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1372                 }
1373                 wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1374
1375                 do_redist_pos_coeffs(pme, cr, bFirst, coordinates, RedistC6);
1376                 pme->lb_buf1.resize(atc.numAtoms());
1377                 pme->lb_buf2.resize(atc.numAtoms());
1378                 local_c6.assign(pme->lb_buf1.begin(), pme->lb_buf1.end());
1379                 for (int i = 0; i < atc.numAtoms(); ++i)
1380                 {
1381                     local_c6[i] = atc.coefficient[i];
1382                 }
1383
1384                 do_redist_pos_coeffs(pme, cr, FALSE, coordinates, RedistSigma);
1385                 local_sigma.assign(pme->lb_buf2.begin(), pme->lb_buf2.end());
1386                 for (int i = 0; i < atc.numAtoms(); ++i)
1387                 {
1388                     local_sigma[i] = atc.coefficient[i];
1389                 }
1390
1391                 wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1392             }
1393             atc.coefficient = coefficientBuffer;
1394             calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1395
1396             /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1397             for (int grid_index = 2; grid_index < 9; ++grid_index)
1398             {
1399                 /* Unpack structure */
1400                 pmegrid    = &pme->pmegrid[grid_index];
1401                 fftgrid    = pme->fftgrid[grid_index];
1402                 pfft_setup = pme->pfft_setup[grid_index];
1403                 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1404                 grid = pmegrid->grid.grid;
1405
1406                 wallcycle_start(wcycle, WallCycleCounter::PmeSpread);
1407                 /* Spread the c6 on a grid */
1408                 spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1409
1410                 if (bFirst)
1411                 {
1412                     inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1413                 }
1414
1415                 inc_nrnb(nrnb,
1416                          eNR_SPREADBSP,
1417                          pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1418                 if (pme->nthread == 1)
1419                 {
1420                     wrap_periodic_pmegrid(pme, grid);
1421                     /* sum contributions to local grid from other nodes */
1422                     if (pme->nnodes > 1)
1423                     {
1424                         gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1425                     }
1426                     copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1427                 }
1428                 wallcycle_stop(wcycle, WallCycleCounter::PmeSpread);
1429
1430                 /*Here we start a large thread parallel region*/
1431 #pragma omp parallel num_threads(pme->nthread) private(thread)
1432                 {
1433                     try
1434                     {
1435                         thread = gmx_omp_get_thread_num();
1436                         /* do 3d-fft */
1437                         if (thread == 0)
1438                         {
1439                             wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1440                         }
1441
1442                         gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1443                         if (thread == 0)
1444                         {
1445                             wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1446                         }
1447                     }
1448                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1449                 }
1450                 bFirst = FALSE;
1451             }
1452             /* solve in k-space for our local cells */
1453 #pragma omp parallel num_threads(pme->nthread) private(thread)
1454             {
1455                 try
1456                 {
1457                     int loop_count;
1458                     thread = gmx_omp_get_thread_num();
1459                     if (thread == 0)
1460                     {
1461                         wallcycle_start(wcycle, WallCycleCounter::LJPme);
1462                     }
1463
1464                     loop_count =
1465                             solve_pme_lj_yzx(pme,
1466                                              &pme->cfftgrid[2],
1467                                              TRUE,
1468                                              scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1469                                              computeEnergyAndVirial,
1470                                              pme->nthread,
1471                                              thread);
1472                     if (thread == 0)
1473                     {
1474                         wallcycle_stop(wcycle, WallCycleCounter::LJPme);
1475                         inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1476                     }
1477                 }
1478                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1479             }
1480
1481             if (computeEnergyAndVirial)
1482             {
1483                 /* This should only be called on the master thread and
1484                  * after the threads have synchronized.
1485                  */
1486                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[fep_state]);
1487             }
1488
1489             bFirst = !pme->doCoulomb;
1490             calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1491             for (int grid_index = 8; grid_index >= 2; --grid_index)
1492             {
1493                 /* Unpack structure */
1494                 pmegrid    = &pme->pmegrid[grid_index];
1495                 fftgrid    = pme->fftgrid[grid_index];
1496                 pfft_setup = pme->pfft_setup[grid_index];
1497                 grid       = pmegrid->grid.grid;
1498                 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1499 #pragma omp parallel num_threads(pme->nthread) private(thread)
1500                 {
1501                     try
1502                     {
1503                         thread = gmx_omp_get_thread_num();
1504                         /* do 3d-invfft */
1505                         if (thread == 0)
1506                         {
1507                             wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1508                         }
1509
1510                         gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1511                         if (thread == 0)
1512                         {
1513                             wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1514
1515
1516                             if (pme->nodeid == 0)
1517                             {
1518                                 real      ntot = pme->nkx * pme->nky * pme->nkz;
1519                                 const int npme = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1520                                 inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1521                             }
1522                             wallcycle_start(wcycle, WallCycleCounter::PmeGather);
1523                         }
1524
1525                         copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1526                     }
1527                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1528                 } /*#pragma omp parallel*/
1529
1530                 /* distribute local grid to all nodes */
1531                 if (pme->nnodes > 1)
1532                 {
1533                     gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1534                 }
1535
1536                 unwrap_periodic_pmegrid(pme, grid);
1537
1538                 if (stepWork.computeForces)
1539                 {
1540                     /* interpolate forces for our local atoms */
1541                     bClearF    = (bFirst && PAR(cr));
1542                     real scale = pme->bFEP ? (fep_state < 1 ? 1.0 - lambda_lj : lambda_lj) : 1.0;
1543                     scale *= lb_scale_factor[grid_index - 2];
1544
1545 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1546                     for (thread = 0; thread < pme->nthread; thread++)
1547                     {
1548                         try
1549                         {
1550                             gather_f_bsplines(
1551                                     pme, grid, bClearF, &pme->atc[0], &pme->atc[0].spline[thread], scale);
1552                         }
1553                         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1554                     }
1555
1556
1557                     inc_nrnb(nrnb,
1558                              eNR_GATHERFBSP,
1559                              pme->pme_order * pme->pme_order * pme->pme_order * pme->atc[0].numAtoms());
1560                 }
1561                 wallcycle_stop(wcycle, WallCycleCounter::PmeGather);
1562
1563                 bFirst = FALSE;
1564             } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1565         }     /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1566     }         /* if (pme->doLJ && pme->ljpme_combination_rule == LongRangeVdW::LB) */
1567
1568     if (stepWork.computeForces && pme->nnodes > 1)
1569     {
1570         wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1571         for (int d = 0; d < pme->ndecompdim; d++)
1572         {
1573             gmx::ArrayRef<gmx::RVec> forcesRef;
1574             if (d == pme->ndecompdim - 1)
1575             {
1576                 const size_t numAtoms = coordinates.size();
1577                 GMX_ASSERT(forces.size() >= numAtoms, "Need at least numAtoms forces");
1578                 forcesRef = forces.subArray(0, numAtoms);
1579             }
1580             else
1581             {
1582                 forcesRef = pme->atc[d + 1].f;
1583             }
1584             if (haveDDAtomOrdering(*cr))
1585             {
1586                 dd_pmeredist_f(pme, &pme->atc[d], forcesRef, d == pme->ndecompdim - 1 && pme->bPPnode);
1587             }
1588         }
1589
1590         wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1591     }
1592
1593     if (computeEnergyAndVirial)
1594     {
1595         if (pme->doCoulomb)
1596         {
1597             if (!pme->bFEP_q)
1598             {
1599                 *energy_q = output[0].coulombEnergy_;
1600                 m_add(vir_q, output[0].coulombVirial_, vir_q);
1601             }
1602             else
1603             {
1604                 *energy_q = (1.0 - lambda_q) * output[0].coulombEnergy_ + lambda_q * output[1].coulombEnergy_;
1605                 *dvdlambda_q += output[1].coulombEnergy_ - output[0].coulombEnergy_;
1606                 for (int i = 0; i < DIM; i++)
1607                 {
1608                     for (int j = 0; j < DIM; j++)
1609                     {
1610                         vir_q[i][j] += (1.0 - lambda_q) * output[0].coulombVirial_[i][j]
1611                                        + lambda_q * output[1].coulombVirial_[i][j];
1612                     }
1613                 }
1614             }
1615         }
1616         else
1617         {
1618             *energy_q = 0;
1619         }
1620
1621         if (pme->doLJ)
1622         {
1623             if (!pme->bFEP_lj)
1624             {
1625                 *energy_lj = output[0].lennardJonesEnergy_;
1626                 m_add(vir_lj, output[0].lennardJonesVirial_, vir_lj);
1627             }
1628             else
1629             {
1630                 *energy_lj = (1.0 - lambda_lj) * output[0].lennardJonesEnergy_
1631                              + lambda_lj * output[1].lennardJonesEnergy_;
1632                 *dvdlambda_lj += output[1].lennardJonesEnergy_ - output[0].lennardJonesEnergy_;
1633                 for (int i = 0; i < DIM; i++)
1634                 {
1635                     for (int j = 0; j < DIM; j++)
1636                     {
1637                         vir_lj[i][j] += (1.0 - lambda_lj) * output[0].lennardJonesVirial_[i][j]
1638                                         + lambda_lj * output[1].lennardJonesVirial_[i][j];
1639                     }
1640                 }
1641             }
1642         }
1643         else
1644         {
1645             *energy_lj = 0;
1646         }
1647     }
1648     return 0;
1649 }
1650
1651 void gmx_pme_destroy(gmx_pme_t* pme)
1652 {
1653     if (!pme)
1654     {
1655         return;
1656     }
1657
1658     sfree(pme->nnx);
1659     sfree(pme->nny);
1660     sfree(pme->nnz);
1661     sfree(pme->fshx);
1662     sfree(pme->fshy);
1663     sfree(pme->fshz);
1664
1665     for (int i = 0; i < pme->ngrids; ++i)
1666     {
1667         pmegrids_destroy(&pme->pmegrid[i]);
1668     }
1669     if (pme->pfft_setup)
1670     {
1671         for (int i = 0; i < pme->ngrids; ++i)
1672         {
1673             gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1674         }
1675     }
1676     sfree(pme->fftgrid);
1677     sfree(pme->cfftgrid);
1678     sfree(pme->pfft_setup);
1679
1680     for (int i = 0; i < DIM; i++)
1681     {
1682         sfree(pme->bsp_mod[i]);
1683     }
1684
1685     sfree(pme->bufv);
1686     sfree(pme->bufr);
1687
1688     if (pme->solve_work)
1689     {
1690         pme_free_all_work(&pme->solve_work, pme->nthread);
1691     }
1692
1693     destroy_pme_spline_work(pme->spline_work);
1694
1695     if (pme->gpu != nullptr)
1696     {
1697         pme_gpu_destroy(pme->gpu);
1698     }
1699
1700     delete pme;
1701 }
1702
1703 void gmx_pme_reinit_atoms(gmx_pme_t*                pme,
1704                           const int                 numAtoms,
1705                           gmx::ArrayRef<const real> chargesA,
1706                           gmx::ArrayRef<const real> chargesB)
1707 {
1708     if (pme->gpu != nullptr)
1709     {
1710         GMX_ASSERT(!(pme->bFEP_q && chargesB.empty()),
1711                    "B state charges must be specified if running Coulomb FEP on the GPU");
1712         pme_gpu_reinit_atoms(pme->gpu, numAtoms, chargesA.data(), pme->bFEP_q ? chargesB.data() : nullptr);
1713     }
1714     else
1715     {
1716         pme->atc[0].setNumAtoms(numAtoms);
1717         // TODO: set the charges here as well
1718     }
1719 }
1720
1721 bool gmx_pme_grid_matches(const gmx_pme_t& pme, const ivec grid_size)
1722 {
1723     return (pme.nkx == grid_size[XX] && pme.nky == grid_size[YY] && pme.nkz == grid_size[ZZ]);
1724 }
1725
1726 void gmx::SeparatePmeRanksPermitted::disablePmeRanks(const std::string& reason)
1727 {
1728     permitSeparatePmeRanks_ = false;
1729
1730     if (!reason.empty())
1731     {
1732         reasons_.push_back(reason);
1733     }
1734 }
1735
1736 bool gmx::SeparatePmeRanksPermitted::permitSeparatePmeRanks() const
1737 {
1738     return permitSeparatePmeRanks_;
1739 }
1740
1741 std::string gmx::SeparatePmeRanksPermitted::reasonsWhyDisabled() const
1742 {
1743     return joinStrings(reasons_, "; ");
1744 }