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/stringutil.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 /*! \brief Help build a descriptive message in \c error if there are
 131  * \c errorReasons why PME on GPU is not supported.
 132  *
 133  * \returns Whether the lack of errorReasons indicate there is support. */
 134 static bool addMessageIfNotSupported(const std::list<std::string>& errorReasons, std::string* error)
 135 {
 136     bool isSupported = errorReasons.empty();
 137     if (!isSupported && error)
 138     {
 139         std::string regressionTestMarker = "PME GPU does not support";
 140         // this prefix is tested for in the regression tests script gmxtest.pl
 141         *error = regressionTestMarker;
 142         if (errorReasons.size() == 1)
 143         {
 144             *error += " " + errorReasons.back();
 145         }
 146         else
 147         {
 148             *error += ": " + gmx::joinStrings(errorReasons, "; ");
 149         }
 150         *error += ".";
 151     }
 152     return isSupported;
 153 }
 154
 155 bool pme_gpu_supports_build(std::string* error)
 156 {
 157     std::list<std::string> errorReasons;
 158     if (GMX_DOUBLE)
 159     {
 160         errorReasons.emplace_back("a double-precision build");
 161     }
 162     if (!GMX_GPU)
 163     {
 164         errorReasons.emplace_back("a non-GPU build");
 165     }
 166     if (GMX_GPU_SYCL)
 167     {
 168         errorReasons.emplace_back("SYCL build"); // SYCL-TODO
 169     }
 170     return addMessageIfNotSupported(errorReasons, error);
 171 }
 172
 173 bool pme_gpu_supports_hardware(const gmx_hw_info_t gmx_unused& hwinfo, std::string* error)
 174 {
 175     std::list<std::string> errorReasons;
 176
 177     if (GMX_GPU_OPENCL)
 178     {
 179 #ifdef __APPLE__
 180         errorReasons.emplace_back("Apple OS X operating system");
 181 #endif
 182     }
 183     return addMessageIfNotSupported(errorReasons, error);
 184 }
 185
 186 bool pme_gpu_supports_input(const t_inputrec& ir, std::string* error)
 187 {
 188     std::list<std::string> errorReasons;
 189     if (!EEL_PME(ir.coulombtype))
 190     {
 191         errorReasons.emplace_back("systems that do not use PME for electrostatics");
 192     }
 193     if (ir.pme_order != 4)
 194     {
 195         errorReasons.emplace_back("interpolation orders other than 4");
 196     }
 197     if (EVDW_PME(ir.vdwtype))
 198     {
 199         errorReasons.emplace_back("Lennard-Jones PME");
 200     }
 201     if (!EI_DYNAMICS(ir.eI))
 202     {
 203         errorReasons.emplace_back(
 204                 "Cannot compute PME interactions on a GPU, because PME GPU requires a dynamical "
 205                 "integrator (md, sd, etc).");
 206     }
 207     return addMessageIfNotSupported(errorReasons, error);
 208 }
 209
 210 /*! \brief \libinternal
 211  * Finds out if PME with given inputs is possible to run on GPU.
 212  * This function is an internal final check, validating the whole PME structure on creation,
 213  * but it still duplicates the preliminary checks from the above (externally exposed) pme_gpu_supports_input() - just in case.
 214  *
 215  * \param[in]  pme          The PME structure.
 216  * \param[out] error        The error message if the input is not supported on GPU.
 217  * \returns                 True if this PME input is possible to run on GPU, false otherwise.
 218  */
 219 static bool pme_gpu_check_restrictions(const gmx_pme_t* pme, std::string* error)
 220 {
 221     std::list<std::string> errorReasons;
 222     if (pme->nnodes != 1)
 223     {
 224         errorReasons.emplace_back("PME decomposition");
 225     }
 226     if (pme->pme_order != 4)
 227     {
 228         errorReasons.emplace_back("interpolation orders other than 4");
 229     }
 230     if (pme->doLJ)
 231     {
 232         errorReasons.emplace_back("Lennard-Jones PME");
 233     }
 234     if (GMX_DOUBLE)
 235     {
 236         errorReasons.emplace_back("double precision");
 237     }
 238     if (!GMX_GPU)
 239     {
 240         errorReasons.emplace_back("non-GPU build of GROMACS");
 241     }
 242     if (GMX_GPU_SYCL)
 243     {
 244         errorReasons.emplace_back("SYCL build of GROMACS"); // SYCL-TODO
 245     }
 246     return addMessageIfNotSupported(errorReasons, error);
 247 }
 248
 249 PmeRunMode pme_run_mode(const gmx_pme_t* pme)
 250 {
 251     GMX_ASSERT(pme != nullptr, "Expecting valid PME data pointer");
 252     return pme->runMode;
 253 }
 254
 255 gmx::PinningPolicy pme_get_pinning_policy()
 256 {
 257     return gmx::PinningPolicy::PinnedIfSupported;
 258 }
 259
 260 /*! \brief Number of bytes in a cache line.
 261  *
 262  * Must also be a multiple of the SIMD and SIMD4 register size, to
 263  * preserve alignment.
 264  */
 265 const int gmxCacheLineSize = 64;
 266
 267 //! Set up coordinate communication
 268 static void setup_coordinate_communication(PmeAtomComm* atc)
 269 {
 270     int nslab, n, i;
 271     int fw, bw;
 272
 273     nslab = atc->nslab;
 274
 275     n = 0;
 276     for (i = 1; i <= nslab / 2; i++)
 277     {
 278         fw = (atc->nodeid + i) % nslab;
 279         bw = (atc->nodeid - i + nslab) % nslab;
 280         if (n < nslab - 1)
 281         {
 282             atc->slabCommSetup[n].node_dest = fw;
 283             atc->slabCommSetup[n].node_src  = bw;
 284             n++;
 285         }
 286         if (n < nslab - 1)
 287         {
 288             atc->slabCommSetup[n].node_dest = bw;
 289             atc->slabCommSetup[n].node_src  = fw;
 290             n++;
 291         }
 292     }
 293 }
 294
 295 /*! \brief Round \p n up to the next multiple of \p f */
 296 static int mult_up(int n, int f)
 297 {
 298     return ((n + f - 1) / f) * f;
 299 }
 300
 301 /*! \brief Return estimate of the load imbalance from the PME grid not being a good match for the number of PME ranks */
 302 static double estimate_pme_load_imbalance(struct gmx_pme_t* pme)
 303 {
 304     int    nma, nmi;
 305     double n1, n2, n3;
 306
 307     nma = pme->nnodes_major;
 308     nmi = pme->nnodes_minor;
 309
 310     n1 = mult_up(pme->nkx, nma) * mult_up(pme->nky, nmi) * pme->nkz;
 311     n2 = mult_up(pme->nkx, nma) * mult_up(pme->nkz, nmi) * pme->nky;
 312     n3 = mult_up(pme->nky, nma) * mult_up(pme->nkz, nmi) * pme->nkx;
 313
 314     /* pme_solve is roughly double the cost of an fft */
 315
 316     return (n1 + n2 + 3 * n3) / static_cast<double>(6 * pme->nkx * pme->nky * pme->nkz);
 317 }
 318
 319 #ifndef DOXYGEN
 320
 321 PmeAtomComm::PmeAtomComm(MPI_Comm   PmeMpiCommunicator,
 322                          const int  numThreads,
 323                          const int  pmeOrder,
 324                          const int  dimIndex,
 325                          const bool doSpread) :
 326     dimind(dimIndex), bSpread(doSpread), pme_order(pmeOrder), nthread(numThreads), spline(nthread)
 327 {
 328     if (PmeMpiCommunicator != MPI_COMM_NULL)
 329     {
 330         mpi_comm = PmeMpiCommunicator;
 331 #    if GMX_MPI
 332         MPI_Comm_size(mpi_comm, &nslab);
 333         MPI_Comm_rank(mpi_comm, &nodeid);
 334 #    endif
 335     }
 336     if (debug)
 337     {
 338         fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", dimind, nslab, nodeid);
 339     }
 340
 341     if (nslab > 1)
 342     {
 343         slabCommSetup.resize(nslab);
 344         setup_coordinate_communication(this);
 345
 346         count_thread.resize(nthread);
 347         for (auto& countThread : count_thread)
 348         {
 349             countThread.resize(nslab);
 350         }
 351     }
 352
 353     if (nthread > 1)
 354     {
 355         threadMap.resize(nthread);
 356
 357 #    pragma omp parallel for num_threads(nthread) schedule(static)
 358         for (int thread = 0; thread < nthread; thread++)
 359         {
 360             try
 361             {
 362                 /* Allocate buffer with padding to avoid cache polution */
 363                 threadMap[thread].nBuffer.resize(nthread + 2 * gmxCacheLineSize);
 364                 threadMap[thread].n = threadMap[thread].nBuffer.data() + gmxCacheLineSize;
 365             }
 366             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 367         }
 368     }
 369 }
 370
 371 #endif // !DOXYGEN
 372
 373 /*! \brief Initialize data structure for communication */
 374 static void init_overlap_comm(pme_overlap_t* ol, int norder, MPI_Comm comm, int nnodes, int nodeid, int ndata, int commplainsize)
 375 {
 376     gmx_bool bCont;
 377
 378     ol->mpi_comm = comm;
 379     ol->nnodes   = nnodes;
 380     ol->nodeid   = nodeid;
 381
 382     /* Linear translation of the PME grid won't affect reciprocal space
 383      * calculations, so to optimize we only interpolate "upwards",
 384      * which also means we only have to consider overlap in one direction.
 385      * I.e., particles on this node might also be spread to grid indices
 386      * that belong to higher nodes (modulo nnodes)
 387      */
 388
 389     ol->s2g0.resize(ol->nnodes + 1);
 390     ol->s2g1.resize(ol->nnodes);
 391     if (debug)
 392     {
 393         fprintf(debug, "PME slab boundaries:");
 394     }
 395     for (int i = 0; i < nnodes; i++)
 396     {
 397         /* s2g0 the local interpolation grid start.
 398          * s2g1 the local interpolation grid end.
 399          * Since in calc_pidx we divide particles, and not grid lines,
 400          * spatially uniform along dimension x or y, we need to round
 401          * s2g0 down and s2g1 up.
 402          */
 403         ol->s2g0[i] = (i * ndata + 0) / nnodes;
 404         ol->s2g1[i] = ((i + 1) * ndata + nnodes - 1) / nnodes + norder - 1;
 405
 406         if (debug)
 407         {
 408             fprintf(debug, "  %3d %3d", ol->s2g0[i], ol->s2g1[i]);
 409         }
 410     }
 411     ol->s2g0[nnodes] = ndata;
 412     if (debug)
 413     {
 414         fprintf(debug, "\n");
 415     }
 416
 417     /* Determine with how many nodes we need to communicate the grid overlap */
 418     int testRankCount = 0;
 419     do
 420     {
 421         testRankCount++;
 422         bCont = FALSE;
 423         for (int i = 0; i < nnodes; i++)
 424         {
 425             if ((i + testRankCount < nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount])
 426                 || (i + testRankCount >= nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount - nnodes] + ndata))
 427             {
 428                 bCont = TRUE;
 429             }
 430         }
 431     } while (bCont && testRankCount < nnodes);
 432
 433     ol->comm_data.resize(testRankCount - 1);
 434     ol->send_size = 0;
 435
 436     for (size_t b = 0; b < ol->comm_data.size(); b++)
 437     {
 438         pme_grid_comm_t* pgc = &ol->comm_data[b];
 439
 440         /* Send */
 441         pgc->send_id  = (ol->nodeid + (b + 1)) % ol->nnodes;
 442         int fft_start = ol->s2g0[pgc->send_id];
 443         int fft_end   = ol->s2g0[pgc->send_id + 1];
 444         if (pgc->send_id < nodeid)
 445         {
 446             fft_start += ndata;
 447             fft_end += ndata;
 448         }
 449         int send_index1  = ol->s2g1[nodeid];
 450         send_index1      = std::min(send_index1, fft_end);
 451         pgc->send_index0 = fft_start;
 452         pgc->send_nindex = std::max(0, send_index1 - pgc->send_index0);
 453         ol->send_size += pgc->send_nindex;
 454
 455         /* We always start receiving to the first index of our slab */
 456         pgc->recv_id    = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
 457         fft_start       = ol->s2g0[ol->nodeid];
 458         fft_end         = ol->s2g0[ol->nodeid + 1];
 459         int recv_index1 = ol->s2g1[pgc->recv_id];
 460         if (pgc->recv_id > nodeid)
 461         {
 462             recv_index1 -= ndata;
 463         }
 464         recv_index1      = std::min(recv_index1, fft_end);
 465         pgc->recv_index0 = fft_start;
 466         pgc->recv_nindex = std::max(0, recv_index1 - pgc->recv_index0);
 467     }
 468
 469 #if GMX_MPI
 470     /* Communicate the buffer sizes to receive */
 471     MPI_Status stat;
 472     for (size_t b = 0; b < ol->comm_data.size(); b++)
 473     {
 474         MPI_Sendrecv(&ol->send_size,
 475                      1,
 476                      MPI_INT,
 477                      ol->comm_data[b].send_id,
 478                      b,
 479                      &ol->comm_data[b].recv_size,
 480                      1,
 481                      MPI_INT,
 482                      ol->comm_data[b].recv_id,
 483                      b,
 484                      ol->mpi_comm,
 485                      &stat);
 486     }
 487 #endif
 488
 489     /* For non-divisible grid we need pme_order iso pme_order-1 */
 490     ol->sendbuf.resize(norder * commplainsize);
 491     ol->recvbuf.resize(norder * commplainsize);
 492 }
 493
 494 int minimalPmeGridSize(int pmeOrder)
 495 {
 496     /* The actual grid size limitations are:
 497      *   serial:        >= pme_order
 498      *   DD, no OpenMP: >= 2*(pme_order - 1)
 499      *   DD, OpenMP:    >= pme_order + 1
 500      * But we use the maximum for simplicity since in practice there is not
 501      * much performance difference between pme_order and 2*(pme_order -1).
 502      */
 503     int minimalSize = 2 * (pmeOrder - 1);
 504
 505     GMX_RELEASE_ASSERT(pmeOrder >= 3, "pmeOrder has to be >= 3");
 506     GMX_RELEASE_ASSERT(minimalSize >= pmeOrder + 1, "The grid size should be >= pmeOrder + 1");
 507
 508     return minimalSize;
 509 }
 510
 511 bool gmx_pme_check_restrictions(int pme_order, int nkx, int nky, int nkz, int numPmeDomainsAlongX, bool useThreads, bool errorsAreFatal)
 512 {
 513     if (pme_order > PME_ORDER_MAX)
 514     {
 515         if (!errorsAreFatal)
 516         {
 517             return false;
 518         }
 519
 520         std::string message = gmx::formatString(
 521                 "pme_order (%d) is larger than the maximum allowed value (%d). Modify and "
 522                 "recompile the code if you really need such a high order.",
 523                 pme_order,
 524                 PME_ORDER_MAX);
 525         GMX_THROW(gmx::InconsistentInputError(message));
 526     }
 527
 528     const int minGridSize = minimalPmeGridSize(pme_order);
 529     if (nkx < minGridSize || nky < minGridSize || nkz < minGridSize)
 530     {
 531         if (!errorsAreFatal)
 532         {
 533             return false;
 534         }
 535         std::string message =
 536                 gmx::formatString("The PME grid sizes need to be >= 2*(pme_order-1) (%d)", minGridSize);
 537         GMX_THROW(gmx::InconsistentInputError(message));
 538     }
 539
 540     /* Check for a limitation of the (current) sum_fftgrid_dd code.
 541      * We only allow multiple communication pulses in dim 1, not in dim 0.
 542      */
 543     if (useThreads
 544         && (nkx < numPmeDomainsAlongX * pme_order && nkx != numPmeDomainsAlongX * (pme_order - 1)))
 545     {
 546         if (!errorsAreFatal)
 547         {
 548             return false;
 549         }
 550         gmx_fatal(FARGS,
 551                   "The number of PME grid lines per rank along x is %g. But when using OpenMP "
 552                   "threads, the number of grid lines per rank along x should be >= pme_order (%d) "
 553                   "or = pmeorder-1. To resolve this issue, use fewer ranks along x (and possibly "
 554                   "more along y and/or z) by specifying -dd manually.",
 555                   nkx / static_cast<double>(numPmeDomainsAlongX),
 556                   pme_order);
 557     }
 558
 559     return true;
 560 }
 561
 562 /*! \brief Round \p enumerator */
 563 static int div_round_up(int enumerator, int denominator)
 564 {
 565     return (enumerator + denominator - 1) / denominator;
 566 }
 567
 568 gmx_pme_t* gmx_pme_init(const t_commrec*     cr,
 569                         const NumPmeDomains& numPmeDomains,
 570                         const t_inputrec*    ir,
 571                         gmx_bool             bFreeEnergy_q,
 572                         gmx_bool             bFreeEnergy_lj,
 573                         gmx_bool             bReproducible,
 574                         real                 ewaldcoeff_q,
 575                         real                 ewaldcoeff_lj,
 576                         int                  nthread,
 577                         PmeRunMode           runMode,
 578                         PmeGpu*              pmeGpu,
 579                         const DeviceContext* deviceContext,
 580                         const DeviceStream*  deviceStream,
 581                         const PmeGpuProgram* pmeGpuProgram,
 582                         const gmx::MDLogger& mdlog)
 583 {
 584     int  use_threads, sum_use_threads, i;
 585     ivec ndata;
 586
 587     if (debug)
 588     {
 589         fprintf(debug, "Creating PME data structures.\n");
 590     }
 591
 592     gmx::unique_cptr<gmx_pme_t, gmx_pme_destroy> pme(new gmx_pme_t());
 593
 594     pme->sum_qgrid_tmp    = nullptr;
 595     pme->sum_qgrid_dd_tmp = nullptr;
 596
 597     pme->buf_nalloc = 0;
 598
 599     pme->nnodes  = 1;
 600     pme->bPPnode = TRUE;
 601
 602     pme->nnodes_major = numPmeDomains.x;
 603     pme->nnodes_minor = numPmeDomains.y;
 604
 605     if (numPmeDomains.x * numPmeDomains.y > 1)
 606     {
 607         pme->mpi_comm = cr->mpi_comm_mygroup;
 608
 609 #if GMX_MPI
 610         MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
 611         MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
 612 #endif
 613         if (pme->nnodes != numPmeDomains.x * numPmeDomains.y)
 614         {
 615             gmx_incons("PME rank count mismatch");
 616         }
 617     }
 618     else
 619     {
 620         pme->mpi_comm = MPI_COMM_NULL;
 621     }
 622
 623     if (pme->nnodes == 1)
 624     {
 625         pme->mpi_comm_d[0] = MPI_COMM_NULL;
 626         pme->mpi_comm_d[1] = MPI_COMM_NULL;
 627         pme->ndecompdim    = 0;
 628         pme->nodeid_major  = 0;
 629         pme->nodeid_minor  = 0;
 630     }
 631     else
 632     {
 633         if (numPmeDomains.y == 1)
 634         {
 635             pme->mpi_comm_d[0] = pme->mpi_comm;
 636             pme->mpi_comm_d[1] = MPI_COMM_NULL;
 637             pme->ndecompdim    = 1;
 638             pme->nodeid_major  = pme->nodeid;
 639             pme->nodeid_minor  = 0;
 640         }
 641         else if (numPmeDomains.x == 1)
 642         {
 643             pme->mpi_comm_d[0] = MPI_COMM_NULL;
 644             pme->mpi_comm_d[1] = pme->mpi_comm;
 645             pme->ndecompdim    = 1;
 646             pme->nodeid_major  = 0;
 647             pme->nodeid_minor  = pme->nodeid;
 648         }
 649         else
 650         {
 651             if (pme->nnodes % numPmeDomains.x != 0)
 652             {
 653                 gmx_incons(
 654                         "For 2D PME decomposition, #PME ranks must be divisible by the number of "
 655                         "domains along x");
 656             }
 657             pme->ndecompdim = 2;
 658
 659 #if GMX_MPI
 660             MPI_Comm_split(pme->mpi_comm,
 661                            pme->nodeid % numPmeDomains.y,
 662                            pme->nodeid,
 663                            &pme->mpi_comm_d[0]); /* My communicator along major dimension */
 664             MPI_Comm_split(pme->mpi_comm,
 665                            pme->nodeid / numPmeDomains.y,
 666                            pme->nodeid,
 667                            &pme->mpi_comm_d[1]); /* My communicator along minor dimension */
 668
 669             MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
 670             MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
 671             MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
 672             MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
 673 #endif
 674         }
 675     }
 676     // cr is always initialized if there is a a PP rank, so we can safely assume
 677     // that when it is not, like in ewald tests, we not on a PP rank.
 678     pme->bPPnode = ((cr != nullptr && cr->duty != 0) && thisRankHasDuty(cr, DUTY_PP));
 679
 680     pme->nthread = nthread;
 681
 682     /* Check if any of the PME MPI ranks uses threads */
 683     use_threads = (pme->nthread > 1 ? 1 : 0);
 684 #if GMX_MPI
 685     if (pme->nnodes > 1)
 686     {
 687         MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT, MPI_SUM, pme->mpi_comm);
 688     }
 689     else
 690 #endif
 691     {
 692         sum_use_threads = use_threads;
 693     }
 694     pme->bUseThreads = (sum_use_threads > 0);
 695
 696     if (ir->pbcType == PbcType::Screw)
 697     {
 698         gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
 699     }
 700
 701     /* NOTE:
 702      * It is likely that the current gmx_pme_do() routine supports calculating
 703      * only Coulomb or LJ while gmx_pme_init() configures for both,
 704      * but that has never been tested.
 705      * It is likely that the current gmx_pme_do() routine supports calculating,
 706      * not calculating free-energy for Coulomb and/or LJ while gmx_pme_init()
 707      * configures with free-energy, but that has never been tested.
 708      */
 709     pme->doCoulomb = EEL_PME(ir->coulombtype);
 710     pme->doLJ      = EVDW_PME(ir->vdwtype);
 711     pme->bFEP_q    = ((ir->efep != FreeEnergyPerturbationType::No) && bFreeEnergy_q);
 712     pme->bFEP_lj   = ((ir->efep != FreeEnergyPerturbationType::No) && bFreeEnergy_lj);
 713     pme->bFEP      = (pme->bFEP_q || pme->bFEP_lj);
 714     pme->nkx       = ir->nkx;
 715     pme->nky       = ir->nky;
 716     pme->nkz       = ir->nkz;
 717     pme->bP3M = (ir->coulombtype == CoulombInteractionType::P3mAD || getenv("GMX_PME_P3M") != nullptr);
 718     pme->pme_order     = ir->pme_order;
 719     pme->ewaldcoeff_q  = ewaldcoeff_q;
 720     pme->ewaldcoeff_lj = ewaldcoeff_lj;
 721
 722     /* Always constant electrostatics coefficients */
 723     pme->epsilon_r = ir->epsilon_r;
 724
 725     /* Always constant LJ coefficients */
 726     pme->ljpme_combination_rule = ir->ljpme_combination_rule;
 727
 728     // The box requires scaling with nwalls = 2, we store that condition as well
 729     // as the scaling factor
 730     delete pme->boxScaler;
 731     pme->boxScaler = new EwaldBoxZScaler(inputrecPbcXY2Walls(ir), ir->wall_ewald_zfac);
 732
 733     /* If we violate restrictions, generate a fatal error here */
 734     gmx_pme_check_restrictions(
 735             pme->pme_order, pme->nkx, pme->nky, pme->nkz, pme->nnodes_major, pme->bUseThreads, true);
 736
 737     if (pme->nnodes > 1)
 738     {
 739         double imbal;
 740
 741 #if GMX_MPI
 742         MPI_Type_contiguous(DIM, GMX_MPI_REAL, &(pme->rvec_mpi));
 743         MPI_Type_commit(&(pme->rvec_mpi));
 744 #endif
 745
 746         /* Note that the coefficient spreading and force gathering, which usually
 747          * takes about the same amount of time as FFT+solve_pme,
 748          * is always fully load balanced
 749          * (unless the coefficient distribution is inhomogeneous).
 750          */
 751
 752         imbal = estimate_pme_load_imbalance(pme.get());
 753         if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
 754         {
 755             GMX_LOG(mdlog.warning)
 756                     .asParagraph()
 757                     .appendTextFormatted(
 758                             "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
 759                             "      For optimal PME load balancing\n"
 760                             "      PME grid_x (%d) and grid_y (%d) should be divisible by "
 761                             "#PME_ranks_x "
 762                             "(%d)\n"
 763                             "      and PME grid_y (%d) and grid_z (%d) should be divisible by "
 764                             "#PME_ranks_y "
 765                             "(%d)",
 766                             gmx::roundToInt((imbal - 1) * 100),
 767                             pme->nkx,
 768                             pme->nky,
 769                             pme->nnodes_major,
 770                             pme->nky,
 771                             pme->nkz,
 772                             pme->nnodes_minor);
 773         }
 774     }
 775
 776     /* For non-divisible grid we need pme_order iso pme_order-1 */
 777     /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
 778      * y is always copied through a buffer: we don't need padding in z,
 779      * but we do need the overlap in x because of the communication order.
 780      */
 781     init_overlap_comm(&pme->overlap[0],
 782                       pme->pme_order,
 783                       pme->mpi_comm_d[0],
 784                       pme->nnodes_major,
 785                       pme->nodeid_major,
 786                       pme->nkx,
 787                       (div_round_up(pme->nky, pme->nnodes_minor) + pme->pme_order)
 788                               * (pme->nkz + pme->pme_order - 1));
 789
 790     /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
 791      * We do this with an offset buffer of equal size, so we need to allocate
 792      * extra for the offset. That's what the (+1)*pme->nkz is for.
 793      */
 794     init_overlap_comm(&pme->overlap[1],
 795                       pme->pme_order,
 796                       pme->mpi_comm_d[1],
 797                       pme->nnodes_minor,
 798                       pme->nodeid_minor,
 799                       pme->nky,
 800                       (div_round_up(pme->nkx, pme->nnodes_major) + pme->pme_order + 1) * pme->nkz);
 801
 802     /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
 803      * Note that gmx_pme_check_restrictions checked for this already.
 804      */
 805     if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
 806     {
 807         gmx_incons(
 808                 "More than one communication pulse required for grid overlap communication along "
 809                 "the major dimension while using threads");
 810     }
 811
 812     snew(pme->bsp_mod[XX], pme->nkx);
 813     snew(pme->bsp_mod[YY], pme->nky);
 814     snew(pme->bsp_mod[ZZ], pme->nkz);
 815
 816     pme->gpu     = pmeGpu; /* Carrying over the single GPU structure */
 817     pme->runMode = runMode;
 818
 819     /* The required size of the interpolation grid, including overlap.
 820      * The allocated size (pmegrid_n?) might be slightly larger.
 821      */
 822     pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] - pme->overlap[0].s2g0[pme->nodeid_major];
 823     pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] - pme->overlap[1].s2g0[pme->nodeid_minor];
 824     pme->pmegrid_nz_base = pme->nkz;
 825     pme->pmegrid_nz      = pme->pmegrid_nz_base + pme->pme_order - 1;
 826     set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
 827     pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
 828     pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
 829     pme->pmegrid_start_iz = 0;
 830
 831     make_gridindex_to_localindex(
 832             pme->nkx, pme->pmegrid_start_ix, pme->pmegrid_nx - (pme->pme_order - 1), &pme->nnx, &pme->fshx);
 833     make_gridindex_to_localindex(
 834             pme->nky, pme->pmegrid_start_iy, pme->pmegrid_ny - (pme->pme_order - 1), &pme->nny, &pme->fshy);
 835     make_gridindex_to_localindex(
 836             pme->nkz, pme->pmegrid_start_iz, pme->pmegrid_nz_base, &pme->nnz, &pme->fshz);
 837
 838     pme->spline_work = make_pme_spline_work(pme->pme_order);
 839
 840     ndata[0] = pme->nkx;
 841     ndata[1] = pme->nky;
 842     ndata[2] = pme->nkz;
 843     /* It doesn't matter if we allocate too many grids here,
 844      * we only allocate and use the ones we need.
 845      */
 846     if (pme->doLJ)
 847     {
 848         pme->ngrids = ((ir->ljpme_combination_rule == LongRangeVdW::LB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
 849     }
 850     else
 851     {
 852         pme->ngrids = DO_Q;
 853     }
 854     snew(pme->fftgrid, pme->ngrids);
 855     snew(pme->cfftgrid, pme->ngrids);
 856     snew(pme->pfft_setup, pme->ngrids);
 857
 858     for (i = 0; i < pme->ngrids; ++i)
 859     {
 860         if ((i < DO_Q && pme->doCoulomb && (i == 0 || bFreeEnergy_q))
 861             || (i >= DO_Q && pme->doLJ
 862                 && (i == 2 || bFreeEnergy_lj || ir->ljpme_combination_rule == LongRangeVdW::LB)))
 863         {
 864             pmegrids_init(&pme->pmegrid[i],
 865                           pme->pmegrid_nx,
 866                           pme->pmegrid_ny,
 867                           pme->pmegrid_nz,
 868                           pme->pmegrid_nz_base,
 869                           pme->pme_order,
 870                           pme->bUseThreads,
 871                           pme->nthread,
 872                           pme->overlap[0].s2g1[pme->nodeid_major]
 873                                   - pme->overlap[0].s2g0[pme->nodeid_major + 1],
 874                           pme->overlap[1].s2g1[pme->nodeid_minor]
 875                                   - pme->overlap[1].s2g0[pme->nodeid_minor + 1]);
 876             /* This routine will allocate the grid data to fit the FFTs */
 877             const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed)
 878                                                         ? gmx::PinningPolicy::PinnedIfSupported
 879                                                         : gmx::PinningPolicy::CannotBePinned;
 880             gmx_parallel_3dfft_init(&pme->pfft_setup[i],
 881                                     ndata,
 882                                     &pme->fftgrid[i],
 883                                     &pme->cfftgrid[i],
 884                                     pme->mpi_comm_d,
 885                                     bReproducible,
 886                                     pme->nthread,
 887                                     allocateRealGridForGpu);
 888         }
 889     }
 890
 891     if (!pme->bP3M)
 892     {
 893         /* Use plain SPME B-spline interpolation */
 894         make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 895     }
 896     else
 897     {
 898         /* Use the P3M grid-optimized influence function */
 899         make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 900     }
 901
 902     /* Use atc[0] for spreading */
 903     const int firstDimIndex   = (numPmeDomains.x > 1 ? 0 : 1);
 904     MPI_Comm  mpiCommFirstDim = (pme->nnodes > 1 ? pme->mpi_comm_d[firstDimIndex] : MPI_COMM_NULL);
 905     bool      doSpread        = true;
 906     pme->atc.emplace_back(mpiCommFirstDim, pme->nthread, pme->pme_order, firstDimIndex, doSpread);
 907     if (pme->ndecompdim >= 2)
 908     {
 909         const int secondDimIndex = 1;
 910         doSpread                 = false;
 911         pme->atc.emplace_back(pme->mpi_comm_d[1], pme->nthread, pme->pme_order, secondDimIndex, doSpread);
 912     }
 913
 914     // Initial check of validity of the input for running on the GPU
 915     if (pme->runMode != PmeRunMode::CPU)
 916     {
 917         std::string errorString;
 918         bool        canRunOnGpu = pme_gpu_check_restrictions(pme.get(), &errorString);
 919         if (!canRunOnGpu)
 920         {
 921             GMX_THROW(gmx::NotImplementedError(errorString));
 922         }
 923         pme_gpu_reinit(pme.get(), deviceContext, deviceStream, pmeGpuProgram);
 924     }
 925     else
 926     {
 927         GMX_ASSERT(pme->gpu == nullptr, "Should not have PME GPU object when PME is on a CPU.");
 928     }
 929
 930
 931     pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
 932
 933     // no exception was thrown during the init, so we hand over the PME structure handle
 934     return pme.release();
 935 }
 936
 937 void gmx_pme_reinit(struct gmx_pme_t** pmedata,
 938                     const t_commrec*   cr,
 939                     struct gmx_pme_t*  pme_src,
 940                     const t_inputrec*  ir,
 941                     const ivec         grid_size,
 942                     real               ewaldcoeff_q,
 943                     real               ewaldcoeff_lj)
 944 {
 945     // Create a copy of t_inputrec fields that are used in gmx_pme_init().
 946     // TODO: This would be better as just copying a sub-structure that contains
 947     // all the PME parameters and nothing else.
 948     t_inputrec irc;
 949     irc.pbcType                = ir->pbcType;
 950     irc.coulombtype            = ir->coulombtype;
 951     irc.vdwtype                = ir->vdwtype;
 952     irc.efep                   = ir->efep;
 953     irc.pme_order              = ir->pme_order;
 954     irc.epsilon_r              = ir->epsilon_r;
 955     irc.ljpme_combination_rule = ir->ljpme_combination_rule;
 956     irc.nkx                    = grid_size[XX];
 957     irc.nky                    = grid_size[YY];
 958     irc.nkz                    = grid_size[ZZ];
 959
 960     try
 961     {
 962         // This is reinit. Any logging should have been done at first init.
 963         // Here we should avoid writing notes for settings the user did not
 964         // set directly.
 965         const gmx::MDLogger dummyLogger;
 966         GMX_ASSERT(pmedata, "Invalid PME pointer");
 967         NumPmeDomains numPmeDomains = { pme_src->nnodes_major, pme_src->nnodes_minor };
 968         *pmedata                    = gmx_pme_init(cr,
 969                                 numPmeDomains,
 970                                 &irc,
 971                                 pme_src->bFEP_q,
 972                                 pme_src->bFEP_lj,
 973                                 FALSE,
 974                                 ewaldcoeff_q,
 975                                 ewaldcoeff_lj,
 976                                 pme_src->nthread,
 977                                 pme_src->runMode,
 978                                 pme_src->gpu,
 979                                 nullptr,
 980                                 nullptr,
 981                                 nullptr,
 982                                 dummyLogger);
 983         /* When running PME on the CPU not using domain decomposition,
 984          * the atom data is allocated once only in gmx_pme_(re)init().
 985          */
 986         if (!pme_src->gpu && pme_src->nnodes == 1)
 987         {
 988             gmx_pme_reinit_atoms(*pmedata, pme_src->atc[0].numAtoms(), {}, {});
 989         }
 990         // TODO this is mostly passing around current values
 991     }
 992     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 993
 994     /* We can easily reuse the allocated pme grids in pme_src */
 995     reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
 996     /* We would like to reuse the fft grids, but that's harder */
 997 }
 998
 999 real gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q)
1000 {
1001     pmegrids_t* grid;
1002
1003     if (pme->nnodes > 1)
1004     {
1005         gmx_incons("gmx_pme_calc_energy called in parallel");
1006     }
1007     if (pme->bFEP_q)
1008     {
1009         gmx_incons("gmx_pme_calc_energy with free energy");
1010     }
1011
1012     if (!pme->atc_energy)
1013     {
1014         pme->atc_energy = std::make_unique<PmeAtomComm>(MPI_COMM_NULL, 1, pme->pme_order, 0, true);
1015     }
1016     PmeAtomComm* atc = pme->atc_energy.get();
1017     atc->setNumAtoms(x.ssize());
1018     atc->x           = x;
1019     atc->coefficient = q;
1020
1021     /* We only use the A-charges grid */
1022     grid = &pme->pmegrid[PME_GRID_QA];
1023
1024     /* Only calculate the spline coefficients, don't actually spread */
1025     spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
1026
1027     return gather_energy_bsplines(pme, grid->grid.grid, atc);
1028 }
1029
1030 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
1031 static void calc_initial_lb_coeffs(gmx::ArrayRef<real>       coefficient,
1032                                    gmx::ArrayRef<const real> local_c6,
1033                                    gmx::ArrayRef<const real> local_sigma)
1034 {
1035     for (gmx::index i = 0; i < coefficient.ssize(); ++i)
1036     {
1037         real sigma4    = local_sigma[i];
1038         sigma4         = sigma4 * sigma4;
1039         sigma4         = sigma4 * sigma4;
1040         coefficient[i] = local_c6[i] / sigma4;
1041     }
1042 }
1043
1044 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
1045 static void calc_next_lb_coeffs(gmx::ArrayRef<real> coefficient, gmx::ArrayRef<const real> local_sigma)
1046 {
1047     for (gmx::index i = 0; i < coefficient.ssize(); ++i)
1048     {
1049         coefficient[i] *= local_sigma[i];
1050     }
1051 }
1052
1053 int gmx_pme_do(struct gmx_pme_t*              pme,
1054                gmx::ArrayRef<const gmx::RVec> coordinates,
1055                gmx::ArrayRef<gmx::RVec>       forces,
1056                gmx::ArrayRef<const real>      chargeA,
1057                gmx::ArrayRef<const real>      chargeB,
1058                gmx::ArrayRef<const real>      c6A,
1059                gmx::ArrayRef<const real>      c6B,
1060                gmx::ArrayRef<const real>      sigmaA,
1061                gmx::ArrayRef<const real>      sigmaB,
1062                const matrix                   box,
1063                const t_commrec*               cr,
1064                int                            maxshift_x,
1065                int                            maxshift_y,
1066                t_nrnb*                        nrnb,
1067                gmx_wallcycle*                 wcycle,
1068                matrix                         vir_q,
1069                matrix                         vir_lj,
1070                real*                          energy_q,
1071                real*                          energy_lj,
1072                real                           lambda_q,
1073                real                           lambda_lj,
1074                real*                          dvdlambda_q,
1075                real*                          dvdlambda_lj,
1076                const gmx::StepWorkload&       stepWork)
1077 {
1078     GMX_ASSERT(pme->runMode == PmeRunMode::CPU,
1079                "gmx_pme_do should not be called on the GPU PME run.");
1080
1081     int                       d, npme, grid_index, max_grid_index;
1082     PmeAtomComm&              atc     = pme->atc[0];
1083     pmegrids_t*               pmegrid = nullptr;
1084     real*                     grid    = nullptr;
1085     gmx::ArrayRef<const real> coefficient;
1086     PmeOutput                 output[2]; // The second is used for the B state with FEP
1087     real                      scale, lambda;
1088     gmx_bool                  bClearF;
1089     gmx_parallel_3dfft_t      pfft_setup;
1090     real*                     fftgrid;
1091     t_complex*                cfftgrid;
1092     int                       thread;
1093     gmx_bool                  bFirst, bDoSplines;
1094     int                       fep_state;
1095     int                       fep_states_lj = pme->bFEP_lj ? 2 : 1;
1096     // There's no support for computing energy without virial, or vice versa
1097     const bool computeEnergyAndVirial = (stepWork.computeEnergy || stepWork.computeVirial);
1098
1099     /* We could be passing lambda!=0 while no q or LJ is actually perturbed */
1100     if (!pme->bFEP_q)
1101     {
1102         lambda_q = 0;
1103     }
1104     if (!pme->bFEP_lj)
1105     {
1106         lambda_lj = 0;
1107     }
1108
1109     assert(pme->nnodes > 0);
1110     assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1111
1112     if (pme->nnodes > 1)
1113     {
1114         atc.pd.resize(coordinates.ssize());
1115         for (int d = pme->ndecompdim - 1; d >= 0; d--)
1116         {
1117             PmeAtomComm& atc = pme->atc[d];
1118             atc.maxshift     = (atc.dimind == 0 ? maxshift_x : maxshift_y);
1119         }
1120     }
1121     else
1122     {
1123         GMX_ASSERT(coordinates.ssize() == atc.numAtoms(), "We expect atc.numAtoms() coordinates");
1124         GMX_ASSERT(forces.ssize() >= atc.numAtoms(),
1125                    "We need a force buffer with at least atc.numAtoms() elements");
1126
1127         atc.x = coordinates;
1128         atc.f = forces;
1129     }
1130
1131     matrix scaledBox;
1132     pme->boxScaler->scaleBox(box, scaledBox);
1133
1134     gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1135     bFirst = TRUE;
1136
1137     /* For simplicity, we construct the splines for all particles if
1138      * more than one PME calculations is needed. Some optimization
1139      * could be done by keeping track of which atoms have splines
1140      * constructed, and construct new splines on each pass for atoms
1141      * that don't yet have them.
1142      */
1143
1144     bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1145
1146     /* We need a maximum of four separate PME calculations:
1147      * grid_index=0: Coulomb PME with charges from state A
1148      * grid_index=1: Coulomb PME with charges from state B
1149      * grid_index=2: LJ PME with C6 from state A
1150      * grid_index=3: LJ PME with C6 from state B
1151      * For Lorentz-Berthelot combination rules, a separate loop is used to
1152      * calculate all the terms
1153      */
1154
1155     /* If we are doing LJ-PME with LB, we only do Q here */
1156     max_grid_index = (pme->ljpme_combination_rule == LongRangeVdW::LB) ? DO_Q : DO_Q_AND_LJ;
1157
1158     for (grid_index = 0; grid_index < max_grid_index; ++grid_index)
1159     {
1160         /* Check if we should do calculations at this grid_index
1161          * If grid_index is odd we should be doing FEP
1162          * If grid_index < 2 we should be doing electrostatic PME
1163          * If grid_index >= 2 we should be doing LJ-PME
1164          */
1165         if ((grid_index < DO_Q && (!pme->doCoulomb || (grid_index == 1 && !pme->bFEP_q)))
1166             || (grid_index >= DO_Q && (!pme->doLJ || (grid_index == 3 && !pme->bFEP_lj))))
1167         {
1168             continue;
1169         }
1170         /* Unpack structure */
1171         pmegrid    = &pme->pmegrid[grid_index];
1172         fftgrid    = pme->fftgrid[grid_index];
1173         cfftgrid   = pme->cfftgrid[grid_index];
1174         pfft_setup = pme->pfft_setup[grid_index];
1175         switch (grid_index)
1176         {
1177             case 0: coefficient = chargeA; break;
1178             case 1: coefficient = chargeB; break;
1179             case 2: coefficient = c6A; break;
1180             case 3: coefficient = c6B; break;
1181         }
1182
1183         grid = pmegrid->grid.grid;
1184
1185         if (debug)
1186         {
1187             fprintf(debug, "PME: number of ranks = %d, rank = %d\n", cr->nnodes, cr->nodeid);
1188             fprintf(debug, "Grid = %p\n", static_cast<void*>(grid));
1189             if (grid == nullptr)
1190             {
1191                 gmx_fatal(FARGS, "No grid!");
1192             }
1193         }
1194
1195         if (pme->nnodes == 1)
1196         {
1197             atc.coefficient = coefficient;
1198         }
1199         else
1200         {
1201             wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1202             do_redist_pos_coeffs(pme, cr, bFirst, coordinates, coefficient);
1203
1204             wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1205         }
1206
1207         if (debug)
1208         {
1209             fprintf(debug, "Rank= %6d, pme local particles=%6d\n", cr->nodeid, atc.numAtoms());
1210         }
1211
1212         wallcycle_start(wcycle, WallCycleCounter::PmeSpread);
1213
1214         /* Spread the coefficients on a grid */
1215         spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1216
1217         if (bFirst)
1218         {
1219             inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1220         }
1221         inc_nrnb(nrnb, eNR_SPREADBSP, pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1222
1223         if (!pme->bUseThreads)
1224         {
1225             wrap_periodic_pmegrid(pme, grid);
1226
1227             /* sum contributions to local grid from other nodes */
1228             if (pme->nnodes > 1)
1229             {
1230                 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1231             }
1232
1233             copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1234         }
1235
1236         wallcycle_stop(wcycle, WallCycleCounter::PmeSpread);
1237
1238         /* TODO If the OpenMP and single-threaded implementations
1239            converge, then spread_on_grid() and
1240            copy_pmegrid_to_fftgrid() will perhaps live in the same
1241            source file.
1242         */
1243
1244         /* Here we start a large thread parallel region */
1245 #pragma omp parallel num_threads(pme->nthread) private(thread)
1246         {
1247             try
1248             {
1249                 thread = gmx_omp_get_thread_num();
1250                 int loop_count;
1251
1252                 /* do 3d-fft */
1253                 if (thread == 0)
1254                 {
1255                     wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1256                 }
1257                 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1258                 if (thread == 0)
1259                 {
1260                     wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1261                 }
1262
1263                 /* solve in k-space for our local cells */
1264                 if (thread == 0)
1265                 {
1266                     wallcycle_start(
1267                             wcycle,
1268                             (grid_index < DO_Q ? WallCycleCounter::PmeSolve : WallCycleCounter::LJPme));
1269                 }
1270                 if (grid_index < DO_Q)
1271                 {
1272                     loop_count = solve_pme_yzx(pme,
1273                                                cfftgrid,
1274                                                scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1275                                                computeEnergyAndVirial,
1276                                                pme->nthread,
1277                                                thread);
1278                 }
1279                 else
1280                 {
1281                     loop_count =
1282                             solve_pme_lj_yzx(pme,
1283                                              &cfftgrid,
1284                                              FALSE,
1285                                              scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1286                                              computeEnergyAndVirial,
1287                                              pme->nthread,
1288                                              thread);
1289                 }
1290
1291                 if (thread == 0)
1292                 {
1293                     wallcycle_stop(
1294                             wcycle,
1295                             (grid_index < DO_Q ? WallCycleCounter::PmeSolve : WallCycleCounter::LJPme));
1296                     inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1297                 }
1298
1299                 /* do 3d-invfft */
1300                 if (thread == 0)
1301                 {
1302                     wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1303                 }
1304                 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1305                 if (thread == 0)
1306                 {
1307                     wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1308
1309
1310                     if (pme->nodeid == 0)
1311                     {
1312                         real ntot = pme->nkx * pme->nky * pme->nkz;
1313                         npme      = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1314                         inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1315                     }
1316
1317                     /* Note: this wallcycle region is closed below
1318                        outside an OpenMP region, so take care if
1319                        refactoring code here. */
1320                     wallcycle_start(wcycle, WallCycleCounter::PmeGather);
1321                 }
1322
1323                 copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1324             }
1325             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1326         }
1327         /* End of thread parallel section.
1328          * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1329          */
1330
1331         /* distribute local grid to all nodes */
1332         if (pme->nnodes > 1)
1333         {
1334             gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1335         }
1336
1337         unwrap_periodic_pmegrid(pme, grid);
1338
1339         if (stepWork.computeForces)
1340         {
1341             /* interpolate forces for our local atoms */
1342
1343
1344             /* If we are running without parallelization,
1345              * atc->f is the actual force array, not a buffer,
1346              * therefore we should not clear it.
1347              */
1348             lambda  = grid_index < DO_Q ? lambda_q : lambda_lj;
1349             bClearF = (bFirst && PAR(cr));
1350 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1351             for (thread = 0; thread < pme->nthread; thread++)
1352             {
1353                 try
1354                 {
1355                     gather_f_bsplines(pme,
1356                                       grid,
1357                                       bClearF,
1358                                       &atc,
1359                                       &atc.spline[thread],
1360                                       pme->bFEP ? (grid_index % 2 == 0 ? 1.0 - lambda : lambda) : 1.0);
1361                 }
1362                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1363             }
1364
1365
1366             inc_nrnb(nrnb, eNR_GATHERFBSP, pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1367             /* Note: this wallcycle region is opened above inside an OpenMP
1368                region, so take care if refactoring code here. */
1369             wallcycle_stop(wcycle, WallCycleCounter::PmeGather);
1370         }
1371
1372         if (computeEnergyAndVirial)
1373         {
1374             /* This should only be called on the master thread
1375              * and after the threads have synchronized.
1376              */
1377             if (grid_index < 2)
1378             {
1379                 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1380             }
1381             else
1382             {
1383                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1384             }
1385         }
1386         bFirst = FALSE;
1387     } /* of grid_index-loop */
1388
1389     /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1390      * seven terms. */
1391
1392     if (pme->doLJ && pme->ljpme_combination_rule == LongRangeVdW::LB)
1393     {
1394         /* Loop over A- and B-state if we are doing FEP */
1395         for (fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1396         {
1397             std::vector<real>         local_c6;
1398             std::vector<real>         local_sigma;
1399             gmx::ArrayRef<const real> RedistC6;
1400             gmx::ArrayRef<const real> RedistSigma;
1401             gmx::ArrayRef<real>       coefficientBuffer;
1402             if (pme->nnodes == 1)
1403             {
1404                 pme->lb_buf1.resize(atc.numAtoms());
1405                 coefficientBuffer = pme->lb_buf1;
1406                 switch (fep_state)
1407                 {
1408                     case 0:
1409                         local_c6.assign(c6A.begin(), c6A.end());
1410                         local_sigma.assign(sigmaA.begin(), sigmaA.end());
1411                         break;
1412                     case 1:
1413                         local_c6.assign(c6B.begin(), c6B.end());
1414                         local_sigma.assign(sigmaB.begin(), sigmaB.end());
1415                         break;
1416                     default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1417                 }
1418             }
1419             else
1420             {
1421                 coefficientBuffer = atc.coefficientBuffer;
1422                 switch (fep_state)
1423                 {
1424                     case 0:
1425                         RedistC6    = c6A;
1426                         RedistSigma = sigmaA;
1427                         break;
1428                     case 1:
1429                         RedistC6    = c6B;
1430                         RedistSigma = sigmaB;
1431                         break;
1432                     default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1433                 }
1434                 wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1435
1436                 do_redist_pos_coeffs(pme, cr, bFirst, coordinates, RedistC6);
1437                 pme->lb_buf1.resize(atc.numAtoms());
1438                 pme->lb_buf2.resize(atc.numAtoms());
1439                 local_c6.assign(pme->lb_buf1.begin(), pme->lb_buf1.end());
1440                 for (int i = 0; i < atc.numAtoms(); ++i)
1441                 {
1442                     local_c6[i] = atc.coefficient[i];
1443                 }
1444
1445                 do_redist_pos_coeffs(pme, cr, FALSE, coordinates, RedistSigma);
1446                 local_sigma.assign(pme->lb_buf2.begin(), pme->lb_buf2.end());
1447                 for (int i = 0; i < atc.numAtoms(); ++i)
1448                 {
1449                     local_sigma[i] = atc.coefficient[i];
1450                 }
1451
1452                 wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1453             }
1454             atc.coefficient = coefficientBuffer;
1455             calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1456
1457             /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1458             for (grid_index = 2; grid_index < 9; ++grid_index)
1459             {
1460                 /* Unpack structure */
1461                 pmegrid    = &pme->pmegrid[grid_index];
1462                 fftgrid    = pme->fftgrid[grid_index];
1463                 pfft_setup = pme->pfft_setup[grid_index];
1464                 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1465                 grid = pmegrid->grid.grid;
1466
1467                 wallcycle_start(wcycle, WallCycleCounter::PmeSpread);
1468                 /* Spread the c6 on a grid */
1469                 spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1470
1471                 if (bFirst)
1472                 {
1473                     inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1474                 }
1475
1476                 inc_nrnb(nrnb,
1477                          eNR_SPREADBSP,
1478                          pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1479                 if (pme->nthread == 1)
1480                 {
1481                     wrap_periodic_pmegrid(pme, grid);
1482                     /* sum contributions to local grid from other nodes */
1483                     if (pme->nnodes > 1)
1484                     {
1485                         gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1486                     }
1487                     copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1488                 }
1489                 wallcycle_stop(wcycle, WallCycleCounter::PmeSpread);
1490
1491                 /*Here we start a large thread parallel region*/
1492 #pragma omp parallel num_threads(pme->nthread) private(thread)
1493                 {
1494                     try
1495                     {
1496                         thread = gmx_omp_get_thread_num();
1497                         /* do 3d-fft */
1498                         if (thread == 0)
1499                         {
1500                             wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1501                         }
1502
1503                         gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1504                         if (thread == 0)
1505                         {
1506                             wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1507                         }
1508                     }
1509                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1510                 }
1511                 bFirst = FALSE;
1512             }
1513             /* solve in k-space for our local cells */
1514 #pragma omp parallel num_threads(pme->nthread) private(thread)
1515             {
1516                 try
1517                 {
1518                     int loop_count;
1519                     thread = gmx_omp_get_thread_num();
1520                     if (thread == 0)
1521                     {
1522                         wallcycle_start(wcycle, WallCycleCounter::LJPme);
1523                     }
1524
1525                     loop_count =
1526                             solve_pme_lj_yzx(pme,
1527                                              &pme->cfftgrid[2],
1528                                              TRUE,
1529                                              scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1530                                              computeEnergyAndVirial,
1531                                              pme->nthread,
1532                                              thread);
1533                     if (thread == 0)
1534                     {
1535                         wallcycle_stop(wcycle, WallCycleCounter::LJPme);
1536                         inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1537                     }
1538                 }
1539                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1540             }
1541
1542             if (computeEnergyAndVirial)
1543             {
1544                 /* This should only be called on the master thread and
1545                  * after the threads have synchronized.
1546                  */
1547                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[fep_state]);
1548             }
1549
1550             bFirst = !pme->doCoulomb;
1551             calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1552             for (grid_index = 8; grid_index >= 2; --grid_index)
1553             {
1554                 /* Unpack structure */
1555                 pmegrid    = &pme->pmegrid[grid_index];
1556                 fftgrid    = pme->fftgrid[grid_index];
1557                 pfft_setup = pme->pfft_setup[grid_index];
1558                 grid       = pmegrid->grid.grid;
1559                 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1560 #pragma omp parallel num_threads(pme->nthread) private(thread)
1561                 {
1562                     try
1563                     {
1564                         thread = gmx_omp_get_thread_num();
1565                         /* do 3d-invfft */
1566                         if (thread == 0)
1567                         {
1568                             wallcycle_start(wcycle, WallCycleCounter::PmeFft);
1569                         }
1570
1571                         gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1572                         if (thread == 0)
1573                         {
1574                             wallcycle_stop(wcycle, WallCycleCounter::PmeFft);
1575
1576
1577                             if (pme->nodeid == 0)
1578                             {
1579                                 real ntot = pme->nkx * pme->nky * pme->nkz;
1580                                 npme      = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1581                                 inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1582                             }
1583                             wallcycle_start(wcycle, WallCycleCounter::PmeGather);
1584                         }
1585
1586                         copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1587                     }
1588                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1589                 } /*#pragma omp parallel*/
1590
1591                 /* distribute local grid to all nodes */
1592                 if (pme->nnodes > 1)
1593                 {
1594                     gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1595                 }
1596
1597                 unwrap_periodic_pmegrid(pme, grid);
1598
1599                 if (stepWork.computeForces)
1600                 {
1601                     /* interpolate forces for our local atoms */
1602                     bClearF = (bFirst && PAR(cr));
1603                     scale   = pme->bFEP ? (fep_state < 1 ? 1.0 - lambda_lj : lambda_lj) : 1.0;
1604                     scale *= lb_scale_factor[grid_index - 2];
1605
1606 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1607                     for (thread = 0; thread < pme->nthread; thread++)
1608                     {
1609                         try
1610                         {
1611                             gather_f_bsplines(
1612                                     pme, grid, bClearF, &pme->atc[0], &pme->atc[0].spline[thread], scale);
1613                         }
1614                         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1615                     }
1616
1617
1618                     inc_nrnb(nrnb,
1619                              eNR_GATHERFBSP,
1620                              pme->pme_order * pme->pme_order * pme->pme_order * pme->atc[0].numAtoms());
1621                 }
1622                 wallcycle_stop(wcycle, WallCycleCounter::PmeGather);
1623
1624                 bFirst = FALSE;
1625             } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1626         }     /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1627     }         /* if (pme->doLJ && pme->ljpme_combination_rule == LongRangeVdW::LB) */
1628
1629     if (stepWork.computeForces && pme->nnodes > 1)
1630     {
1631         wallcycle_start(wcycle, WallCycleCounter::PmeRedistXF);
1632         for (d = 0; d < pme->ndecompdim; d++)
1633         {
1634             gmx::ArrayRef<gmx::RVec> forcesRef;
1635             if (d == pme->ndecompdim - 1)
1636             {
1637                 const size_t numAtoms = coordinates.size();
1638                 GMX_ASSERT(forces.size() >= numAtoms, "Need at least numAtoms forces");
1639                 forcesRef = forces.subArray(0, numAtoms);
1640             }
1641             else
1642             {
1643                 forcesRef = pme->atc[d + 1].f;
1644             }
1645             if (DOMAINDECOMP(cr))
1646             {
1647                 dd_pmeredist_f(pme, &pme->atc[d], forcesRef, d == pme->ndecompdim - 1 && pme->bPPnode);
1648             }
1649         }
1650
1651         wallcycle_stop(wcycle, WallCycleCounter::PmeRedistXF);
1652     }
1653
1654     if (computeEnergyAndVirial)
1655     {
1656         if (pme->doCoulomb)
1657         {
1658             if (!pme->bFEP_q)
1659             {
1660                 *energy_q = output[0].coulombEnergy_;
1661                 m_add(vir_q, output[0].coulombVirial_, vir_q);
1662             }
1663             else
1664             {
1665                 *energy_q = (1.0 - lambda_q) * output[0].coulombEnergy_ + lambda_q * output[1].coulombEnergy_;
1666                 *dvdlambda_q += output[1].coulombEnergy_ - output[0].coulombEnergy_;
1667                 for (int i = 0; i < DIM; i++)
1668                 {
1669                     for (int j = 0; j < DIM; j++)
1670                     {
1671                         vir_q[i][j] += (1.0 - lambda_q) * output[0].coulombVirial_[i][j]
1672                                        + lambda_q * output[1].coulombVirial_[i][j];
1673                     }
1674                 }
1675             }
1676             if (debug)
1677             {
1678                 fprintf(debug, "Electrostatic PME mesh energy: %g\n", *energy_q);
1679             }
1680         }
1681         else
1682         {
1683             *energy_q = 0;
1684         }
1685
1686         if (pme->doLJ)
1687         {
1688             if (!pme->bFEP_lj)
1689             {
1690                 *energy_lj = output[0].lennardJonesEnergy_;
1691                 m_add(vir_lj, output[0].lennardJonesVirial_, vir_lj);
1692             }
1693             else
1694             {
1695                 *energy_lj = (1.0 - lambda_lj) * output[0].lennardJonesEnergy_
1696                              + lambda_lj * output[1].lennardJonesEnergy_;
1697                 *dvdlambda_lj += output[1].lennardJonesEnergy_ - output[0].lennardJonesEnergy_;
1698                 for (int i = 0; i < DIM; i++)
1699                 {
1700                     for (int j = 0; j < DIM; j++)
1701                     {
1702                         vir_lj[i][j] += (1.0 - lambda_lj) * output[0].lennardJonesVirial_[i][j]
1703                                         + lambda_lj * output[1].lennardJonesVirial_[i][j];
1704                     }
1705                 }
1706             }
1707             if (debug)
1708             {
1709                 fprintf(debug, "Lennard-Jones PME mesh energy: %g\n", *energy_lj);
1710             }
1711         }
1712         else
1713         {
1714             *energy_lj = 0;
1715         }
1716     }
1717     return 0;
1718 }
1719
1720 void gmx_pme_destroy(gmx_pme_t* pme)
1721 {
1722     if (!pme)
1723     {
1724         return;
1725     }
1726
1727     delete pme->boxScaler;
1728
1729     sfree(pme->nnx);
1730     sfree(pme->nny);
1731     sfree(pme->nnz);
1732     sfree(pme->fshx);
1733     sfree(pme->fshy);
1734     sfree(pme->fshz);
1735
1736     for (int i = 0; i < pme->ngrids; ++i)
1737     {
1738         pmegrids_destroy(&pme->pmegrid[i]);
1739     }
1740     if (pme->pfft_setup)
1741     {
1742         for (int i = 0; i < pme->ngrids; ++i)
1743         {
1744             gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1745         }
1746     }
1747     sfree(pme->fftgrid);
1748     sfree(pme->cfftgrid);
1749     sfree(pme->pfft_setup);
1750
1751     for (int i = 0; i < DIM; i++)
1752     {
1753         sfree(pme->bsp_mod[i]);
1754     }
1755
1756     sfree(pme->bufv);
1757     sfree(pme->bufr);
1758
1759     if (pme->solve_work)
1760     {
1761         pme_free_all_work(&pme->solve_work, pme->nthread);
1762     }
1763
1764     sfree(pme->sum_qgrid_tmp);
1765     sfree(pme->sum_qgrid_dd_tmp);
1766
1767     destroy_pme_spline_work(pme->spline_work);
1768
1769     if (pme->gpu != nullptr)
1770     {
1771         pme_gpu_destroy(pme->gpu);
1772     }
1773
1774     delete pme;
1775 }
1776
1777 void gmx_pme_reinit_atoms(gmx_pme_t*                pme,
1778                           const int                 numAtoms,
1779                           gmx::ArrayRef<const real> chargesA,
1780                           gmx::ArrayRef<const real> chargesB)
1781 {
1782     if (pme->gpu != nullptr)
1783     {
1784         GMX_ASSERT(!(pme->bFEP_q && chargesB.empty()),
1785                    "B state charges must be specified if running Coulomb FEP on the GPU");
1786         pme_gpu_reinit_atoms(pme->gpu, numAtoms, chargesA.data(), pme->bFEP_q ? chargesB.data() : nullptr);
1787     }
1788     else
1789     {
1790         pme->atc[0].setNumAtoms(numAtoms);
1791         // TODO: set the charges here as well
1792     }
1793 }
1794
1795 bool gmx_pme_grid_matches(const gmx_pme_t& pme, const ivec grid_size)
1796 {
1797     return (pme.nkx == grid_size[XX] && pme.nky == grid_size[YY] && pme.nkz == grid_size[ZZ]);
1798 }
1799
1800 void gmx::SeparatePmeRanksPermitted::disablePmeRanks(const std::string& reason)
1801 {
1802     permitSeparatePmeRanks_ = false;
1803
1804     if (!reason.empty())
1805     {
1806         reasons_.push_back(reason);
1807     }
1808 }
1809
1810 bool gmx::SeparatePmeRanksPermitted::permitSeparatePmeRanks() const
1811 {
1812     return permitSeparatePmeRanks_;
1813 }
1814
1815 std::string gmx::SeparatePmeRanksPermitted::reasonsWhyDisabled() const
1816 {
1817     return joinStrings(reasons_, "; ");
1818 }