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) 2013,2014,2015,2016,2017,2018, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 /*! \internal \file
  38  *
  39  * \brief This file contains function definitions necessary for
  40  * computing energies and forces for the PME long-ranged part (Coulomb
  41  * and LJ).
  42  *
  43  * \author Erik Lindahl <erik@kth.se>
  44  * \author Berk Hess <hess@kth.se>
  45  * \ingroup module_ewald
  46  */
  47 /* IMPORTANT FOR DEVELOPERS:
  48  *
  49  * Triclinic pme stuff isn't entirely trivial, and we've experienced
  50  * some bugs during development (many of them due to me). To avoid
  51  * this in the future, please check the following things if you make
  52  * changes in this file:
  53  *
  54  * 1. You should obtain identical (at least to the PME precision)
  55  *    energies, forces, and virial for
  56  *    a rectangular box and a triclinic one where the z (or y) axis is
  57  *    tilted a whole box side. For instance you could use these boxes:
  58  *
  59  *    rectangular       triclinic
  60  *     2  0  0           2  0  0
  61  *     0  2  0           0  2  0
  62  *     0  0  6           2  2  6
  63  *
  64  * 2. You should check the energy conservation in a triclinic box.
  65  *
  66  * It might seem an overkill, but better safe than sorry.
  67  * /Erik 001109
  68  */
  69
  70 #include "gmxpre.h"
  71
  72 #include "pme.h"
  73
  74 #include "config.h"
  75
  76 #include <assert.h>
  77 #include <stdio.h>
  78 #include <stdlib.h>
  79 #include <string.h>
  80
  81 #include <cmath>
  82
  83 #include <algorithm>
  84 #include <list>
  85
  86 #include "gromacs/ewald/ewald-utils.h"
  87 #include "gromacs/fft/parallel_3dfft.h"
  88 #include "gromacs/fileio/pdbio.h"
  89 #include "gromacs/gmxlib/network.h"
  90 #include "gromacs/gmxlib/nrnb.h"
  91 #include "gromacs/math/gmxcomplex.h"
  92 #include "gromacs/math/invertmatrix.h"
  93 #include "gromacs/math/units.h"
  94 #include "gromacs/math/vec.h"
  95 #include "gromacs/math/vectypes.h"
  96 #include "gromacs/mdtypes/commrec.h"
  97 #include "gromacs/mdtypes/forcerec.h"
  98 #include "gromacs/mdtypes/inputrec.h"
  99 #include "gromacs/mdtypes/md_enums.h"
 100 #include "gromacs/pbcutil/pbc.h"
 101 #include "gromacs/timing/cyclecounter.h"
 102 #include "gromacs/timing/wallcycle.h"
 103 #include "gromacs/timing/walltime_accounting.h"
 104 #include "gromacs/utility/basedefinitions.h"
 105 #include "gromacs/utility/exceptions.h"
 106 #include "gromacs/utility/fatalerror.h"
 107 #include "gromacs/utility/gmxmpi.h"
 108 #include "gromacs/utility/gmxomp.h"
 109 #include "gromacs/utility/logger.h"
 110 #include "gromacs/utility/real.h"
 111 #include "gromacs/utility/smalloc.h"
 112 #include "gromacs/utility/stringutil.h"
 113 #include "gromacs/utility/unique_cptr.h"
 114
 115 #include "calculate-spline-moduli.h"
 116 #include "pme-gather.h"
 117 #include "pme-gpu-internal.h"
 118 #include "pme-grid.h"
 119 #include "pme-internal.h"
 120 #include "pme-redistribute.h"
 121 #include "pme-solve.h"
 122 #include "pme-spline-work.h"
 123 #include "pme-spread.h"
 124
 125 bool pme_gpu_supports_input(const t_inputrec *ir, std::string *error)
 126 {
 127     std::list<std::string> errorReasons;
 128     if (!EEL_PME(ir->coulombtype))
 129     {
 130         errorReasons.push_back("systems that do not use PME for electrostatics");
 131     }
 132     if (ir->pme_order != 4)
 133     {
 134         errorReasons.push_back("interpolation orders other than 4");
 135     }
 136     if (ir->efep != efepNO)
 137     {
 138         errorReasons.push_back("free energy calculations (multiple grids)");
 139     }
 140     if (EVDW_PME(ir->vdwtype))
 141     {
 142         errorReasons.push_back("Lennard-Jones PME");
 143     }
 144 #if GMX_DOUBLE
 145     {
 146         errorReasons.push_back("double precision");
 147     }
 148 #endif
 149 #if GMX_GPU != GMX_GPU_CUDA
 150     {
 151         errorReasons.push_back("non-CUDA build of GROMACS");
 152     }
 153 #endif
 154     if (ir->cutoff_scheme == ecutsGROUP)
 155     {
 156         errorReasons.push_back("group cutoff scheme");
 157     }
 158     if (EI_TPI(ir->eI))
 159     {
 160         errorReasons.push_back("test particle insertion");
 161     }
 162
 163     bool inputSupported = errorReasons.empty();
 164     if (!inputSupported && error)
 165     {
 166         std::string regressionTestMarker = "PME GPU does not support";
 167         // this prefix is tested for in the regression tests script gmxtest.pl
 168         *error = regressionTestMarker + ": " + gmx::joinStrings(errorReasons, "; ") + ".";
 169     }
 170     return inputSupported;
 171 }
 172
 173 /*! \brief \libinternal
 174  * Finds out if PME with given inputs is possible to run on GPU.
 175  * This function is an internal final check, validating the whole PME structure on creation,
 176  * but it still duplicates the preliminary checks from the above (externally exposed) pme_gpu_supports_input() - just in case.
 177  *
 178  * \param[in]  pme          The PME structure.
 179  * \param[out] error        The error message if the input is not supported on GPU.
 180  * \returns                 True if this PME input is possible to run on GPU, false otherwise.
 181  */
 182 static bool pme_gpu_check_restrictions(const gmx_pme_t *pme, std::string *error)
 183 {
 184     std::list<std::string> errorReasons;
 185     if (pme->nnodes != 1)
 186     {
 187         errorReasons.push_back("PME decomposition");
 188     }
 189     if (pme->pme_order != 4)
 190     {
 191         errorReasons.push_back("interpolation orders other than 4");
 192     }
 193     if (pme->bFEP)
 194     {
 195         errorReasons.push_back("free energy calculations (multiple grids)");
 196     }
 197     if (pme->doLJ)
 198     {
 199         errorReasons.push_back("Lennard-Jones PME");
 200     }
 201 #if GMX_DOUBLE
 202     {
 203         errorReasons.push_back("double precision");
 204     }
 205 #endif
 206 #if GMX_GPU != GMX_GPU_CUDA
 207     {
 208         errorReasons.push_back("non-CUDA build of GROMACS");
 209     }
 210 #endif
 211
 212     bool inputSupported = errorReasons.empty();
 213     if (!inputSupported && error)
 214     {
 215         std::string regressionTestMarker = "PME GPU does not support";
 216         // this prefix is tested for in the regression tests script gmxtest.pl
 217         *error = regressionTestMarker + ": " + gmx::joinStrings(errorReasons, "; ") + ".";
 218     }
 219     return inputSupported;
 220 }
 221
 222 PmeRunMode pme_run_mode(const gmx_pme_t *pme)
 223 {
 224     GMX_ASSERT(pme != nullptr, "Expecting valid PME data pointer");
 225     return pme->runMode;
 226 }
 227
 228 /*! \brief Number of bytes in a cache line.
 229  *
 230  * Must also be a multiple of the SIMD and SIMD4 register size, to
 231  * preserve alignment.
 232  */
 233 const int gmxCacheLineSize = 64;
 234
 235 //! Set up coordinate communication
 236 static void setup_coordinate_communication(pme_atomcomm_t *atc)
 237 {
 238     int nslab, n, i;
 239     int fw, bw;
 240
 241     nslab = atc->nslab;
 242
 243     n = 0;
 244     for (i = 1; i <= nslab/2; i++)
 245     {
 246         fw = (atc->nodeid + i) % nslab;
 247         bw = (atc->nodeid - i + nslab) % nslab;
 248         if (n < nslab - 1)
 249         {
 250             atc->node_dest[n] = fw;
 251             atc->node_src[n]  = bw;
 252             n++;
 253         }
 254         if (n < nslab - 1)
 255         {
 256             atc->node_dest[n] = bw;
 257             atc->node_src[n]  = fw;
 258             n++;
 259         }
 260     }
 261 }
 262
 263 /*! \brief Round \p n up to the next multiple of \p f */
 264 static int mult_up(int n, int f)
 265 {
 266     return ((n + f - 1)/f)*f;
 267 }
 268
 269 /*! \brief Return estimate of the load imbalance from the PME grid not being a good match for the number of PME ranks */
 270 static double estimate_pme_load_imbalance(struct gmx_pme_t *pme)
 271 {
 272     int    nma, nmi;
 273     double n1, n2, n3;
 274
 275     nma = pme->nnodes_major;
 276     nmi = pme->nnodes_minor;
 277
 278     n1 = mult_up(pme->nkx, nma)*mult_up(pme->nky, nmi)*pme->nkz;
 279     n2 = mult_up(pme->nkx, nma)*mult_up(pme->nkz, nmi)*pme->nky;
 280     n3 = mult_up(pme->nky, nma)*mult_up(pme->nkz, nmi)*pme->nkx;
 281
 282     /* pme_solve is roughly double the cost of an fft */
 283
 284     return (n1 + n2 + 3*n3)/(double)(6*pme->nkx*pme->nky*pme->nkz);
 285 }
 286
 287 /*! \brief Initialize atom communication data structure */
 288 static void init_atomcomm(struct gmx_pme_t *pme, pme_atomcomm_t *atc,
 289                           int dimind, gmx_bool bSpread)
 290 {
 291     int thread;
 292
 293     atc->dimind    = dimind;
 294     atc->nslab     = 1;
 295     atc->nodeid    = 0;
 296     atc->pd_nalloc = 0;
 297 #if GMX_MPI
 298     if (pme->nnodes > 1)
 299     {
 300         atc->mpi_comm = pme->mpi_comm_d[dimind];
 301         MPI_Comm_size(atc->mpi_comm, &atc->nslab);
 302         MPI_Comm_rank(atc->mpi_comm, &atc->nodeid);
 303     }
 304     if (debug)
 305     {
 306         fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", atc->dimind, atc->nslab, atc->nodeid);
 307     }
 308 #endif
 309
 310     atc->bSpread   = bSpread;
 311     atc->pme_order = pme->pme_order;
 312
 313     if (atc->nslab > 1)
 314     {
 315         snew(atc->node_dest, atc->nslab);
 316         snew(atc->node_src, atc->nslab);
 317         setup_coordinate_communication(atc);
 318
 319         snew(atc->count_thread, pme->nthread);
 320         for (thread = 0; thread < pme->nthread; thread++)
 321         {
 322             snew(atc->count_thread[thread], atc->nslab);
 323         }
 324         atc->count = atc->count_thread[0];
 325         snew(atc->rcount, atc->nslab);
 326         snew(atc->buf_index, atc->nslab);
 327     }
 328
 329     atc->nthread = pme->nthread;
 330     if (atc->nthread > 1)
 331     {
 332         snew(atc->thread_plist, atc->nthread);
 333     }
 334     snew(atc->spline, atc->nthread);
 335     for (thread = 0; thread < atc->nthread; thread++)
 336     {
 337         if (atc->nthread > 1)
 338         {
 339             snew(atc->thread_plist[thread].n, atc->nthread+2*gmxCacheLineSize);
 340             atc->thread_plist[thread].n += gmxCacheLineSize;
 341         }
 342     }
 343 }
 344
 345 /*! \brief Destroy an atom communication data structure and its child structs */
 346 static void destroy_atomcomm(pme_atomcomm_t *atc)
 347 {
 348     sfree(atc->pd);
 349     if (atc->nslab > 1)
 350     {
 351         sfree(atc->node_dest);
 352         sfree(atc->node_src);
 353         for (int i = 0; i < atc->nthread; i++)
 354         {
 355             sfree(atc->count_thread[i]);
 356         }
 357         sfree(atc->count_thread);
 358         sfree(atc->rcount);
 359         sfree(atc->buf_index);
 360
 361         sfree(atc->x);
 362         sfree(atc->coefficient);
 363         sfree(atc->f);
 364     }
 365     sfree(atc->idx);
 366     sfree(atc->fractx);
 367
 368     sfree(atc->thread_idx);
 369     for (int i = 0; i < atc->nthread; i++)
 370     {
 371         if (atc->nthread > 1)
 372         {
 373             int *n_ptr = atc->thread_plist[i].n - gmxCacheLineSize;
 374             sfree(n_ptr);
 375             sfree(atc->thread_plist[i].i);
 376         }
 377         sfree(atc->spline[i].ind);
 378         for (int d = 0; d < ZZ; d++)
 379         {
 380             sfree(atc->spline[i].theta[d]);
 381             sfree(atc->spline[i].dtheta[d]);
 382         }
 383         sfree_aligned(atc->spline[i].ptr_dtheta_z);
 384         sfree_aligned(atc->spline[i].ptr_theta_z);
 385     }
 386     if (atc->nthread > 1)
 387     {
 388         sfree(atc->thread_plist);
 389     }
 390     sfree(atc->spline);
 391 }
 392
 393 /*! \brief Initialize data structure for communication */
 394 static void
 395 init_overlap_comm(pme_overlap_t *  ol,
 396                   int              norder,
 397 #if GMX_MPI
 398                   MPI_Comm         comm,
 399 #endif
 400                   int              nnodes,
 401                   int              nodeid,
 402                   int              ndata,
 403                   int              commplainsize)
 404 {
 405     gmx_bool         bCont;
 406 #if GMX_MPI
 407     MPI_Status       stat;
 408
 409     ol->mpi_comm = comm;
 410 #endif
 411
 412     ol->nnodes = nnodes;
 413     ol->nodeid = nodeid;
 414
 415     /* Linear translation of the PME grid won't affect reciprocal space
 416      * calculations, so to optimize we only interpolate "upwards",
 417      * which also means we only have to consider overlap in one direction.
 418      * I.e., particles on this node might also be spread to grid indices
 419      * that belong to higher nodes (modulo nnodes)
 420      */
 421
 422     ol->s2g0.resize(ol->nnodes + 1);
 423     ol->s2g1.resize(ol->nnodes);
 424     if (debug)
 425     {
 426         fprintf(debug, "PME slab boundaries:");
 427     }
 428     for (int i = 0; i < nnodes; i++)
 429     {
 430         /* s2g0 the local interpolation grid start.
 431          * s2g1 the local interpolation grid end.
 432          * Since in calc_pidx we divide particles, and not grid lines,
 433          * spatially uniform along dimension x or y, we need to round
 434          * s2g0 down and s2g1 up.
 435          */
 436         ol->s2g0[i] = (i * ndata + 0) / nnodes;
 437         ol->s2g1[i] = ((i + 1) * ndata + nnodes - 1) / nnodes + norder - 1;
 438
 439         if (debug)
 440         {
 441             fprintf(debug, "  %3d %3d", ol->s2g0[i], ol->s2g1[i]);
 442         }
 443     }
 444     ol->s2g0[nnodes] = ndata;
 445     if (debug)
 446     {
 447         fprintf(debug, "\n");
 448     }
 449
 450     /* Determine with how many nodes we need to communicate the grid overlap */
 451     int testRankCount = 0;
 452     do
 453     {
 454         testRankCount++;
 455         bCont = FALSE;
 456         for (int i = 0; i < nnodes; i++)
 457         {
 458             if ((i + testRankCount <  nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount]) ||
 459                 (i + testRankCount >= nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount - nnodes] + ndata))
 460             {
 461                 bCont = TRUE;
 462             }
 463         }
 464     }
 465     while (bCont && testRankCount < nnodes);
 466
 467     ol->comm_data.resize(testRankCount - 1);
 468     ol->send_size = 0;
 469
 470     for (size_t b = 0; b < ol->comm_data.size(); b++)
 471     {
 472         pme_grid_comm_t *pgc = &ol->comm_data[b];
 473
 474         /* Send */
 475         pgc->send_id = (ol->nodeid + (b + 1)) % ol->nnodes;
 476         int fft_start = ol->s2g0[pgc->send_id];
 477         int fft_end   = ol->s2g0[pgc->send_id + 1];
 478         if (pgc->send_id < nodeid)
 479         {
 480             fft_start += ndata;
 481             fft_end   += ndata;
 482         }
 483         int send_index1  = ol->s2g1[nodeid];
 484         send_index1      = std::min(send_index1, fft_end);
 485         pgc->send_index0 = fft_start;
 486         pgc->send_nindex = std::max(0, send_index1 - pgc->send_index0);
 487         ol->send_size   += pgc->send_nindex;
 488
 489         /* We always start receiving to the first index of our slab */
 490         pgc->recv_id     = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
 491         fft_start        = ol->s2g0[ol->nodeid];
 492         fft_end          = ol->s2g0[ol->nodeid + 1];
 493         int recv_index1  = ol->s2g1[pgc->recv_id];
 494         if (pgc->recv_id > nodeid)
 495         {
 496             recv_index1 -= ndata;
 497         }
 498         recv_index1      = std::min(recv_index1, fft_end);
 499         pgc->recv_index0 = fft_start;
 500         pgc->recv_nindex = std::max(0, recv_index1 - pgc->recv_index0);
 501     }
 502
 503 #if GMX_MPI
 504     /* Communicate the buffer sizes to receive */
 505     for (size_t b = 0; b < ol->comm_data.size(); b++)
 506     {
 507         MPI_Sendrecv(&ol->send_size, 1, MPI_INT, ol->comm_data[b].send_id, b,
 508                      &ol->comm_data[b].recv_size, 1, MPI_INT, ol->comm_data[b].recv_id, b,
 509                      ol->mpi_comm, &stat);
 510     }
 511 #endif
 512
 513     /* For non-divisible grid we need pme_order iso pme_order-1 */
 514     ol->sendbuf.resize(norder * commplainsize);
 515     ol->recvbuf.resize(norder * commplainsize);
 516 }
 517
 518 int minimalPmeGridSize(int pmeOrder)
 519 {
 520     /* The actual grid size limitations are:
 521      *   serial:        >= pme_order
 522      *   DD, no OpenMP: >= 2*(pme_order - 1)
 523      *   DD, OpenMP:    >= pme_order + 1
 524      * But we use the maximum for simplicity since in practice there is not
 525      * much performance difference between pme_order and 2*(pme_order -1).
 526      */
 527     int minimalSize = 2*(pmeOrder - 1);
 528
 529     GMX_RELEASE_ASSERT(pmeOrder >= 3, "pmeOrder has to be >= 3");
 530     GMX_RELEASE_ASSERT(minimalSize >= pmeOrder + 1, "The grid size should be >= pmeOrder + 1");
 531
 532     return minimalSize;
 533 }
 534
 535 bool gmx_pme_check_restrictions(int pme_order,
 536                                 int nkx, int nky, int nkz,
 537                                 int nnodes_major,
 538                                 bool useThreads,
 539                                 bool errorsAreFatal)
 540 {
 541     if (pme_order > PME_ORDER_MAX)
 542     {
 543         if (!errorsAreFatal)
 544         {
 545             return false;
 546         }
 547
 548         std::string message = gmx::formatString(
 549                     "pme_order (%d) is larger than the maximum allowed value (%d). Modify and recompile the code if you really need such a high order.",
 550                     pme_order, PME_ORDER_MAX);
 551         GMX_THROW(InconsistentInputError(message));
 552     }
 553
 554     const int minGridSize = minimalPmeGridSize(pme_order);
 555     if (nkx < minGridSize ||
 556         nky < minGridSize ||
 557         nkz < minGridSize)
 558     {
 559         if (!errorsAreFatal)
 560         {
 561             return false;
 562         }
 563         std::string message = gmx::formatString(
 564                     "The PME grid sizes need to be >= 2*(pme_order-1) (%d)",
 565                     minGridSize);
 566         GMX_THROW(InconsistentInputError(message));
 567     }
 568
 569     /* Check for a limitation of the (current) sum_fftgrid_dd code.
 570      * We only allow multiple communication pulses in dim 1, not in dim 0.
 571      */
 572     if (useThreads && (nkx < nnodes_major*pme_order &&
 573                        nkx != nnodes_major*(pme_order - 1)))
 574     {
 575         if (!errorsAreFatal)
 576         {
 577             return false;
 578         }
 579         gmx_fatal(FARGS, "The number of PME grid lines per rank along x is %g. But when using OpenMP threads, the number of grid lines per rank along x should be >= pme_order (%d) or = pmeorder-1. To resolve this issue, use fewer ranks along x (and possibly more along y and/or z) by specifying -dd manually.",
 580                   nkx/(double)nnodes_major, pme_order);
 581     }
 582
 583     return true;
 584 }
 585
 586 /*! \brief Round \p enumerator */
 587 static int div_round_up(int enumerator, int denominator)
 588 {
 589     return (enumerator + denominator - 1)/denominator;
 590 }
 591
 592 gmx_pme_t *gmx_pme_init(const t_commrec     *cr,
 593                         int                  nnodes_major,
 594                         int                  nnodes_minor,
 595                         const t_inputrec    *ir,
 596                         int                  homenr,
 597                         gmx_bool             bFreeEnergy_q,
 598                         gmx_bool             bFreeEnergy_lj,
 599                         gmx_bool             bReproducible,
 600                         real                 ewaldcoeff_q,
 601                         real                 ewaldcoeff_lj,
 602                         int                  nthread,
 603                         PmeRunMode           runMode,
 604                         PmeGpu              *pmeGpu,
 605                         gmx_device_info_t   *gpuInfo,
 606                         const gmx::MDLogger  & /*mdlog*/)
 607 {
 608     int               use_threads, sum_use_threads, i;
 609     ivec              ndata;
 610
 611     if (debug)
 612     {
 613         fprintf(debug, "Creating PME data structures.\n");
 614     }
 615
 616     unique_cptr<gmx_pme_t, gmx_pme_destroy> pme(new gmx_pme_t());
 617
 618     pme->sum_qgrid_tmp       = nullptr;
 619     pme->sum_qgrid_dd_tmp    = nullptr;
 620
 621     pme->buf_nalloc          = 0;
 622
 623     pme->nnodes              = 1;
 624     pme->bPPnode             = TRUE;
 625
 626     pme->nnodes_major        = nnodes_major;
 627     pme->nnodes_minor        = nnodes_minor;
 628
 629 #if GMX_MPI
 630     if (nnodes_major*nnodes_minor > 1)
 631     {
 632         pme->mpi_comm = cr->mpi_comm_mygroup;
 633
 634         MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
 635         MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
 636         if (pme->nnodes != nnodes_major*nnodes_minor)
 637         {
 638             gmx_incons("PME rank count mismatch");
 639         }
 640     }
 641     else
 642     {
 643         pme->mpi_comm = MPI_COMM_NULL;
 644     }
 645 #endif
 646
 647     if (pme->nnodes == 1)
 648     {
 649 #if GMX_MPI
 650         pme->mpi_comm_d[0] = MPI_COMM_NULL;
 651         pme->mpi_comm_d[1] = MPI_COMM_NULL;
 652 #endif
 653         pme->ndecompdim   = 0;
 654         pme->nodeid_major = 0;
 655         pme->nodeid_minor = 0;
 656 #if GMX_MPI
 657         pme->mpi_comm_d[0] = pme->mpi_comm_d[1] = MPI_COMM_NULL;
 658 #endif
 659     }
 660     else
 661     {
 662         if (nnodes_minor == 1)
 663         {
 664 #if GMX_MPI
 665             pme->mpi_comm_d[0] = pme->mpi_comm;
 666             pme->mpi_comm_d[1] = MPI_COMM_NULL;
 667 #endif
 668             pme->ndecompdim   = 1;
 669             pme->nodeid_major = pme->nodeid;
 670             pme->nodeid_minor = 0;
 671
 672         }
 673         else if (nnodes_major == 1)
 674         {
 675 #if GMX_MPI
 676             pme->mpi_comm_d[0] = MPI_COMM_NULL;
 677             pme->mpi_comm_d[1] = pme->mpi_comm;
 678 #endif
 679             pme->ndecompdim   = 1;
 680             pme->nodeid_major = 0;
 681             pme->nodeid_minor = pme->nodeid;
 682         }
 683         else
 684         {
 685             if (pme->nnodes % nnodes_major != 0)
 686             {
 687                 gmx_incons("For 2D PME decomposition, #PME ranks must be divisible by the number of ranks in the major dimension");
 688             }
 689             pme->ndecompdim = 2;
 690
 691 #if GMX_MPI
 692             MPI_Comm_split(pme->mpi_comm, pme->nodeid % nnodes_minor,
 693                            pme->nodeid, &pme->mpi_comm_d[0]);  /* My communicator along major dimension */
 694             MPI_Comm_split(pme->mpi_comm, pme->nodeid/nnodes_minor,
 695                            pme->nodeid, &pme->mpi_comm_d[1]);  /* My communicator along minor dimension */
 696
 697             MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
 698             MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
 699             MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
 700             MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
 701 #endif
 702         }
 703         pme->bPPnode = thisRankHasDuty(cr, DUTY_PP);
 704     }
 705
 706     pme->nthread = nthread;
 707
 708     /* Check if any of the PME MPI ranks uses threads */
 709     use_threads = (pme->nthread > 1 ? 1 : 0);
 710 #if GMX_MPI
 711     if (pme->nnodes > 1)
 712     {
 713         MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT,
 714                       MPI_SUM, pme->mpi_comm);
 715     }
 716     else
 717 #endif
 718     {
 719         sum_use_threads = use_threads;
 720     }
 721     pme->bUseThreads = (sum_use_threads > 0);
 722
 723     if (ir->ePBC == epbcSCREW)
 724     {
 725         gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
 726     }
 727
 728     /* NOTE:
 729      * It is likely that the current gmx_pme_do() routine supports calculating
 730      * only Coulomb or LJ while gmx_pme_init() configures for both,
 731      * but that has never been tested.
 732      * It is likely that the current gmx_pme_do() routine supports calculating,
 733      * not calculating free-energy for Coulomb and/or LJ while gmx_pme_init()
 734      * configures with free-energy, but that has never been tested.
 735      */
 736     pme->doCoulomb     = EEL_PME(ir->coulombtype);
 737     pme->doLJ          = EVDW_PME(ir->vdwtype);
 738     pme->bFEP_q        = ((ir->efep != efepNO) && bFreeEnergy_q);
 739     pme->bFEP_lj       = ((ir->efep != efepNO) && bFreeEnergy_lj);
 740     pme->bFEP          = (pme->bFEP_q || pme->bFEP_lj);
 741     pme->nkx           = ir->nkx;
 742     pme->nky           = ir->nky;
 743     pme->nkz           = ir->nkz;
 744     pme->bP3M          = (ir->coulombtype == eelP3M_AD || getenv("GMX_PME_P3M") != nullptr);
 745     pme->pme_order     = ir->pme_order;
 746     pme->ewaldcoeff_q  = ewaldcoeff_q;
 747     pme->ewaldcoeff_lj = ewaldcoeff_lj;
 748
 749     /* Always constant electrostatics coefficients */
 750     pme->epsilon_r     = ir->epsilon_r;
 751
 752     /* Always constant LJ coefficients */
 753     pme->ljpme_combination_rule = ir->ljpme_combination_rule;
 754
 755     // The box requires scaling with nwalls = 2, we store that condition as well
 756     // as the scaling factor
 757     delete pme->boxScaler;
 758     pme->boxScaler = new EwaldBoxZScaler(*ir);
 759
 760     /* If we violate restrictions, generate a fatal error here */
 761     gmx_pme_check_restrictions(pme->pme_order,
 762                                pme->nkx, pme->nky, pme->nkz,
 763                                pme->nnodes_major,
 764                                pme->bUseThreads,
 765                                true);
 766
 767     if (pme->nnodes > 1)
 768     {
 769         double imbal;
 770
 771 #if GMX_MPI
 772         MPI_Type_contiguous(DIM, GMX_MPI_REAL, &(pme->rvec_mpi));
 773         MPI_Type_commit(&(pme->rvec_mpi));
 774 #endif
 775
 776         /* Note that the coefficient spreading and force gathering, which usually
 777          * takes about the same amount of time as FFT+solve_pme,
 778          * is always fully load balanced
 779          * (unless the coefficient distribution is inhomogeneous).
 780          */
 781
 782         imbal = estimate_pme_load_imbalance(pme.get());
 783         if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
 784         {
 785             fprintf(stderr,
 786                     "\n"
 787                     "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
 788                     "      For optimal PME load balancing\n"
 789                     "      PME grid_x (%d) and grid_y (%d) should be divisible by #PME_ranks_x (%d)\n"
 790                     "      and PME grid_y (%d) and grid_z (%d) should be divisible by #PME_ranks_y (%d)\n"
 791                     "\n",
 792                     (int)((imbal-1)*100 + 0.5),
 793                     pme->nkx, pme->nky, pme->nnodes_major,
 794                     pme->nky, pme->nkz, pme->nnodes_minor);
 795         }
 796     }
 797
 798     /* For non-divisible grid we need pme_order iso pme_order-1 */
 799     /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
 800      * y is always copied through a buffer: we don't need padding in z,
 801      * but we do need the overlap in x because of the communication order.
 802      */
 803     init_overlap_comm(&pme->overlap[0], pme->pme_order,
 804 #if GMX_MPI
 805                       pme->mpi_comm_d[0],
 806 #endif
 807                       pme->nnodes_major, pme->nodeid_major,
 808                       pme->nkx,
 809                       (div_round_up(pme->nky, pme->nnodes_minor)+pme->pme_order)*(pme->nkz+pme->pme_order-1));
 810
 811     /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
 812      * We do this with an offset buffer of equal size, so we need to allocate
 813      * extra for the offset. That's what the (+1)*pme->nkz is for.
 814      */
 815     init_overlap_comm(&pme->overlap[1], pme->pme_order,
 816 #if GMX_MPI
 817                       pme->mpi_comm_d[1],
 818 #endif
 819                       pme->nnodes_minor, pme->nodeid_minor,
 820                       pme->nky,
 821                       (div_round_up(pme->nkx, pme->nnodes_major)+pme->pme_order+1)*pme->nkz);
 822
 823     /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
 824      * Note that gmx_pme_check_restrictions checked for this already.
 825      */
 826     if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
 827     {
 828         gmx_incons("More than one communication pulse required for grid overlap communication along the major dimension while using threads");
 829     }
 830
 831     snew(pme->bsp_mod[XX], pme->nkx);
 832     snew(pme->bsp_mod[YY], pme->nky);
 833     snew(pme->bsp_mod[ZZ], pme->nkz);
 834
 835     pme->gpu     = pmeGpu; /* Carrying over the single GPU structure */
 836     pme->runMode = runMode;
 837
 838     /* The required size of the interpolation grid, including overlap.
 839      * The allocated size (pmegrid_n?) might be slightly larger.
 840      */
 841     pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] -
 842         pme->overlap[0].s2g0[pme->nodeid_major];
 843     pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] -
 844         pme->overlap[1].s2g0[pme->nodeid_minor];
 845     pme->pmegrid_nz_base = pme->nkz;
 846     pme->pmegrid_nz      = pme->pmegrid_nz_base + pme->pme_order - 1;
 847     set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
 848     pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
 849     pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
 850     pme->pmegrid_start_iz = 0;
 851
 852     make_gridindex_to_localindex(pme->nkx,
 853                                  pme->pmegrid_start_ix,
 854                                  pme->pmegrid_nx - (pme->pme_order-1),
 855                                  &pme->nnx, &pme->fshx);
 856     make_gridindex_to_localindex(pme->nky,
 857                                  pme->pmegrid_start_iy,
 858                                  pme->pmegrid_ny - (pme->pme_order-1),
 859                                  &pme->nny, &pme->fshy);
 860     make_gridindex_to_localindex(pme->nkz,
 861                                  pme->pmegrid_start_iz,
 862                                  pme->pmegrid_nz_base,
 863                                  &pme->nnz, &pme->fshz);
 864
 865     pme->spline_work = make_pme_spline_work(pme->pme_order);
 866
 867     ndata[0]    = pme->nkx;
 868     ndata[1]    = pme->nky;
 869     ndata[2]    = pme->nkz;
 870     /* It doesn't matter if we allocate too many grids here,
 871      * we only allocate and use the ones we need.
 872      */
 873     if (pme->doLJ)
 874     {
 875         pme->ngrids = ((ir->ljpme_combination_rule == eljpmeLB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
 876     }
 877     else
 878     {
 879         pme->ngrids = DO_Q;
 880     }
 881     snew(pme->fftgrid, pme->ngrids);
 882     snew(pme->cfftgrid, pme->ngrids);
 883     snew(pme->pfft_setup, pme->ngrids);
 884
 885     for (i = 0; i < pme->ngrids; ++i)
 886     {
 887         if ((i <  DO_Q && pme->doCoulomb && (i == 0 ||
 888                                              bFreeEnergy_q)) ||
 889             (i >= DO_Q && pme->doLJ && (i == 2 ||
 890                                         bFreeEnergy_lj ||
 891                                         ir->ljpme_combination_rule == eljpmeLB)))
 892         {
 893             pmegrids_init(&pme->pmegrid[i],
 894                           pme->pmegrid_nx, pme->pmegrid_ny, pme->pmegrid_nz,
 895                           pme->pmegrid_nz_base,
 896                           pme->pme_order,
 897                           pme->bUseThreads,
 898                           pme->nthread,
 899                           pme->overlap[0].s2g1[pme->nodeid_major]-pme->overlap[0].s2g0[pme->nodeid_major+1],
 900                           pme->overlap[1].s2g1[pme->nodeid_minor]-pme->overlap[1].s2g0[pme->nodeid_minor+1]);
 901             /* This routine will allocate the grid data to fit the FFTs */
 902             const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed) ? gmx::PinningPolicy::CanBePinned : gmx::PinningPolicy::CannotBePinned;
 903             gmx_parallel_3dfft_init(&pme->pfft_setup[i], ndata,
 904                                     &pme->fftgrid[i], &pme->cfftgrid[i],
 905                                     pme->mpi_comm_d,
 906                                     bReproducible, pme->nthread, allocateRealGridForGpu);
 907
 908         }
 909     }
 910
 911     if (!pme->bP3M)
 912     {
 913         /* Use plain SPME B-spline interpolation */
 914         make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 915     }
 916     else
 917     {
 918         /* Use the P3M grid-optimized influence function */
 919         make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 920     }
 921
 922     /* Use atc[0] for spreading */
 923     init_atomcomm(pme.get(), &pme->atc[0], nnodes_major > 1 ? 0 : 1, TRUE);
 924     if (pme->ndecompdim >= 2)
 925     {
 926         init_atomcomm(pme.get(), &pme->atc[1], 1, FALSE);
 927     }
 928
 929     if (pme->nnodes == 1)
 930     {
 931         pme->atc[0].n = homenr;
 932         pme_realloc_atomcomm_things(&pme->atc[0]);
 933     }
 934
 935     pme->lb_buf1       = nullptr;
 936     pme->lb_buf2       = nullptr;
 937     pme->lb_buf_nalloc = 0;
 938
 939     if (pme_gpu_active(pme.get()))
 940     {
 941         if (!pme->gpu)
 942         {
 943             // Initial check of validity of the data
 944             std::string errorString;
 945             bool        canRunOnGpu = pme_gpu_check_restrictions(pme.get(), &errorString);
 946             if (!canRunOnGpu)
 947             {
 948                 GMX_THROW(gmx::NotImplementedError(errorString));
 949             }
 950         }
 951
 952         pme_gpu_reinit(pme.get(), gpuInfo);
 953     }
 954
 955     pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
 956
 957     // no exception was thrown during the init, so we hand over the PME structure handle
 958     return pme.release();
 959 }
 960
 961 void gmx_pme_reinit(struct gmx_pme_t **pmedata,
 962                     const t_commrec   *cr,
 963                     struct gmx_pme_t * pme_src,
 964                     const t_inputrec * ir,
 965                     const ivec         grid_size,
 966                     real               ewaldcoeff_q,
 967                     real               ewaldcoeff_lj)
 968 {
 969     int        homenr;
 970
 971     // Create a copy of t_inputrec fields that are used in gmx_pme_init().
 972     // TODO: This would be better as just copying a sub-structure that contains
 973     // all the PME parameters and nothing else.
 974     t_inputrec irc;
 975     irc.ePBC                   = ir->ePBC;
 976     irc.coulombtype            = ir->coulombtype;
 977     irc.vdwtype                = ir->vdwtype;
 978     irc.efep                   = ir->efep;
 979     irc.pme_order              = ir->pme_order;
 980     irc.epsilon_r              = ir->epsilon_r;
 981     irc.ljpme_combination_rule = ir->ljpme_combination_rule;
 982     irc.nkx                    = grid_size[XX];
 983     irc.nky                    = grid_size[YY];
 984     irc.nkz                    = grid_size[ZZ];
 985
 986     if (pme_src->nnodes == 1)
 987     {
 988         homenr = pme_src->atc[0].n;
 989     }
 990     else
 991     {
 992         homenr = -1;
 993     }
 994
 995     try
 996     {
 997         const gmx::MDLogger dummyLogger;
 998         // This is reinit which is currently only changing grid size/coefficients,
 999         // so we don't expect the actual logging.
1000         // TODO: when PME is an object, it should take reference to mdlog on construction and save it.
1001         GMX_ASSERT(pmedata, "Invalid PME pointer");
1002         *pmedata = gmx_pme_init(cr, pme_src->nnodes_major, pme_src->nnodes_minor,
1003                                 &irc, homenr, pme_src->bFEP_q, pme_src->bFEP_lj, FALSE, ewaldcoeff_q, ewaldcoeff_lj,
1004                                 pme_src->nthread, pme_src->runMode, pme_src->gpu, nullptr, dummyLogger);
1005         //TODO this is mostly passing around current values
1006     }
1007     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1008
1009     /* We can easily reuse the allocated pme grids in pme_src */
1010     reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
1011     /* We would like to reuse the fft grids, but that's harder */
1012 }
1013
1014 void gmx_pme_calc_energy(struct gmx_pme_t *pme, int n, rvec *x, real *q, real *V)
1015 {
1016     pme_atomcomm_t *atc;
1017     pmegrids_t     *grid;
1018
1019     if (pme->nnodes > 1)
1020     {
1021         gmx_incons("gmx_pme_calc_energy called in parallel");
1022     }
1023     if (pme->bFEP_q > 1)
1024     {
1025         gmx_incons("gmx_pme_calc_energy with free energy");
1026     }
1027
1028     atc            = &pme->atc_energy;
1029     atc->nthread   = 1;
1030     if (atc->spline == nullptr)
1031     {
1032         snew(atc->spline, atc->nthread);
1033     }
1034     atc->nslab     = 1;
1035     atc->bSpread   = TRUE;
1036     atc->pme_order = pme->pme_order;
1037     atc->n         = n;
1038     pme_realloc_atomcomm_things(atc);
1039     atc->x           = x;
1040     atc->coefficient = q;
1041
1042     /* We only use the A-charges grid */
1043     grid = &pme->pmegrid[PME_GRID_QA];
1044
1045     /* Only calculate the spline coefficients, don't actually spread */
1046     spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
1047
1048     *V = gather_energy_bsplines(pme, grid->grid.grid, atc);
1049 }
1050
1051 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
1052 static void
1053 calc_initial_lb_coeffs(struct gmx_pme_t *pme, real *local_c6, real *local_sigma)
1054 {
1055     int  i;
1056     for (i = 0; i < pme->atc[0].n; ++i)
1057     {
1058         real sigma4;
1059         sigma4                     = local_sigma[i];
1060         sigma4                     = sigma4*sigma4;
1061         sigma4                     = sigma4*sigma4;
1062         pme->atc[0].coefficient[i] = local_c6[i] / sigma4;
1063     }
1064 }
1065
1066 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
1067 static void
1068 calc_next_lb_coeffs(struct gmx_pme_t *pme, real *local_sigma)
1069 {
1070     int  i;
1071
1072     for (i = 0; i < pme->atc[0].n; ++i)
1073     {
1074         pme->atc[0].coefficient[i] *= local_sigma[i];
1075     }
1076 }
1077
1078 int gmx_pme_do(struct gmx_pme_t *pme,
1079                int start,       int homenr,
1080                rvec x[],        rvec f[],
1081                real chargeA[],  real chargeB[],
1082                real c6A[],      real c6B[],
1083                real sigmaA[],   real sigmaB[],
1084                matrix box,      const t_commrec *cr,
1085                int  maxshift_x, int maxshift_y,
1086                t_nrnb *nrnb,    gmx_wallcycle *wcycle,
1087                matrix vir_q,    matrix vir_lj,
1088                real *energy_q,  real *energy_lj,
1089                real lambda_q,   real lambda_lj,
1090                real *dvdlambda_q, real *dvdlambda_lj,
1091                int flags)
1092 {
1093     GMX_ASSERT(pme->runMode == PmeRunMode::CPU, "gmx_pme_do should not be called on the GPU PME run.");
1094
1095     int                  d, i, j, npme, grid_index, max_grid_index;
1096     int                  n_d;
1097     pme_atomcomm_t      *atc        = nullptr;
1098     pmegrids_t          *pmegrid    = nullptr;
1099     real                *grid       = nullptr;
1100     rvec                *f_d;
1101     real                *coefficient = nullptr;
1102     real                 energy_AB[4];
1103     matrix               vir_AB[4];
1104     real                 scale, lambda;
1105     gmx_bool             bClearF;
1106     gmx_parallel_3dfft_t pfft_setup;
1107     real              *  fftgrid;
1108     t_complex          * cfftgrid;
1109     int                  thread;
1110     gmx_bool             bFirst, bDoSplines;
1111     int                  fep_state;
1112     int                  fep_states_lj           = pme->bFEP_lj ? 2 : 1;
1113     const gmx_bool       bCalcEnerVir            = flags & GMX_PME_CALC_ENER_VIR;
1114     const gmx_bool       bBackFFT                = flags & (GMX_PME_CALC_F | GMX_PME_CALC_POT);
1115     const gmx_bool       bCalcF                  = flags & GMX_PME_CALC_F;
1116
1117     assert(pme->nnodes > 0);
1118     assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1119
1120     if (pme->nnodes > 1)
1121     {
1122         atc      = &pme->atc[0];
1123         atc->npd = homenr;
1124         if (atc->npd > atc->pd_nalloc)
1125         {
1126             atc->pd_nalloc = over_alloc_dd(atc->npd);
1127             srenew(atc->pd, atc->pd_nalloc);
1128         }
1129         for (d = pme->ndecompdim-1; d >= 0; d--)
1130         {
1131             atc           = &pme->atc[d];
1132             atc->maxshift = (atc->dimind == 0 ? maxshift_x : maxshift_y);
1133         }
1134     }
1135     else
1136     {
1137         atc = &pme->atc[0];
1138         /* This could be necessary for TPI */
1139         pme->atc[0].n = homenr;
1140         if (DOMAINDECOMP(cr))
1141         {
1142             pme_realloc_atomcomm_things(atc);
1143         }
1144         atc->x = x;
1145         atc->f = f;
1146     }
1147
1148     matrix scaledBox;
1149     pme->boxScaler->scaleBox(box, scaledBox);
1150
1151     gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1152     bFirst = TRUE;
1153
1154     /* For simplicity, we construct the splines for all particles if
1155      * more than one PME calculations is needed. Some optimization
1156      * could be done by keeping track of which atoms have splines
1157      * constructed, and construct new splines on each pass for atoms
1158      * that don't yet have them.
1159      */
1160
1161     bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1162
1163     /* We need a maximum of four separate PME calculations:
1164      * grid_index=0: Coulomb PME with charges from state A
1165      * grid_index=1: Coulomb PME with charges from state B
1166      * grid_index=2: LJ PME with C6 from state A
1167      * grid_index=3: LJ PME with C6 from state B
1168      * For Lorentz-Berthelot combination rules, a separate loop is used to
1169      * calculate all the terms
1170      */
1171
1172     /* If we are doing LJ-PME with LB, we only do Q here */
1173     max_grid_index = (pme->ljpme_combination_rule == eljpmeLB) ? DO_Q : DO_Q_AND_LJ;
1174
1175     for (grid_index = 0; grid_index < max_grid_index; ++grid_index)
1176     {
1177         /* Check if we should do calculations at this grid_index
1178          * If grid_index is odd we should be doing FEP
1179          * If grid_index < 2 we should be doing electrostatic PME
1180          * If grid_index >= 2 we should be doing LJ-PME
1181          */
1182         if ((grid_index <  DO_Q && (!pme->doCoulomb ||
1183                                     (grid_index == 1 && !pme->bFEP_q))) ||
1184             (grid_index >= DO_Q && (!pme->doLJ ||
1185                                     (grid_index == 3 && !pme->bFEP_lj))))
1186         {
1187             continue;
1188         }
1189         /* Unpack structure */
1190         pmegrid    = &pme->pmegrid[grid_index];
1191         fftgrid    = pme->fftgrid[grid_index];
1192         cfftgrid   = pme->cfftgrid[grid_index];
1193         pfft_setup = pme->pfft_setup[grid_index];
1194         switch (grid_index)
1195         {
1196             case 0: coefficient = chargeA + start; break;
1197             case 1: coefficient = chargeB + start; break;
1198             case 2: coefficient = c6A + start; break;
1199             case 3: coefficient = c6B + start; break;
1200         }
1201
1202         grid = pmegrid->grid.grid;
1203
1204         if (debug)
1205         {
1206             fprintf(debug, "PME: number of ranks = %d, rank = %d\n",
1207                     cr->nnodes, cr->nodeid);
1208             fprintf(debug, "Grid = %p\n", (void*)grid);
1209             if (grid == nullptr)
1210             {
1211                 gmx_fatal(FARGS, "No grid!");
1212             }
1213         }
1214
1215         if (pme->nnodes == 1)
1216         {
1217             atc->coefficient = coefficient;
1218         }
1219         else
1220         {
1221             wallcycle_start(wcycle, ewcPME_REDISTXF);
1222             do_redist_pos_coeffs(pme, cr, start, homenr, bFirst, x, coefficient);
1223
1224             wallcycle_stop(wcycle, ewcPME_REDISTXF);
1225         }
1226
1227         if (debug)
1228         {
1229             fprintf(debug, "Rank= %6d, pme local particles=%6d\n",
1230                     cr->nodeid, atc->n);
1231         }
1232
1233         if (flags & GMX_PME_SPREAD)
1234         {
1235             wallcycle_start(wcycle, ewcPME_SPREAD);
1236
1237             /* Spread the coefficients on a grid */
1238             spread_on_grid(pme, &pme->atc[0], pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1239
1240             if (bFirst)
1241             {
1242                 inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
1243             }
1244             inc_nrnb(nrnb, eNR_SPREADBSP,
1245                      pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
1246
1247             if (!pme->bUseThreads)
1248             {
1249                 wrap_periodic_pmegrid(pme, grid);
1250
1251                 /* sum contributions to local grid from other nodes */
1252 #if GMX_MPI
1253                 if (pme->nnodes > 1)
1254                 {
1255                     gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1256                 }
1257 #endif
1258
1259                 copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1260             }
1261
1262             wallcycle_stop(wcycle, ewcPME_SPREAD);
1263
1264             /* TODO If the OpenMP and single-threaded implementations
1265                converge, then spread_on_grid() and
1266                copy_pmegrid_to_fftgrid() will perhaps live in the same
1267                source file.
1268              */
1269         }
1270
1271         /* Here we start a large thread parallel region */
1272 #pragma omp parallel num_threads(pme->nthread) private(thread)
1273         {
1274             try
1275             {
1276                 thread = gmx_omp_get_thread_num();
1277                 if (flags & GMX_PME_SOLVE)
1278                 {
1279                     int loop_count;
1280
1281                     /* do 3d-fft */
1282                     if (thread == 0)
1283                     {
1284                         wallcycle_start(wcycle, ewcPME_FFT);
1285                     }
1286                     gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
1287                                                thread, wcycle);
1288                     if (thread == 0)
1289                     {
1290                         wallcycle_stop(wcycle, ewcPME_FFT);
1291                     }
1292
1293                     /* solve in k-space for our local cells */
1294                     if (thread == 0)
1295                     {
1296                         wallcycle_start(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1297                     }
1298                     if (grid_index < DO_Q)
1299                     {
1300                         loop_count =
1301                             solve_pme_yzx(pme, cfftgrid,
1302                                           scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1303                                           bCalcEnerVir,
1304                                           pme->nthread, thread);
1305                     }
1306                     else
1307                     {
1308                         loop_count =
1309                             solve_pme_lj_yzx(pme, &cfftgrid, FALSE,
1310                                              scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1311                                              bCalcEnerVir,
1312                                              pme->nthread, thread);
1313                     }
1314
1315                     if (thread == 0)
1316                     {
1317                         wallcycle_stop(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1318                         inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1319                     }
1320                 }
1321
1322                 if (bBackFFT)
1323                 {
1324                     /* do 3d-invfft */
1325                     if (thread == 0)
1326                     {
1327                         wallcycle_start(wcycle, ewcPME_FFT);
1328                     }
1329                     gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
1330                                                thread, wcycle);
1331                     if (thread == 0)
1332                     {
1333                         wallcycle_stop(wcycle, ewcPME_FFT);
1334
1335
1336                         if (pme->nodeid == 0)
1337                         {
1338                             real ntot = pme->nkx*pme->nky*pme->nkz;
1339                             npme  = static_cast<int>(ntot*std::log(ntot)/std::log(2.0));
1340                             inc_nrnb(nrnb, eNR_FFT, 2*npme);
1341                         }
1342
1343                         /* Note: this wallcycle region is closed below
1344                            outside an OpenMP region, so take care if
1345                            refactoring code here. */
1346                         wallcycle_start(wcycle, ewcPME_GATHER);
1347                     }
1348
1349                     copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1350                 }
1351             } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1352         }
1353         /* End of thread parallel section.
1354          * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1355          */
1356
1357         if (bBackFFT)
1358         {
1359             /* distribute local grid to all nodes */
1360 #if GMX_MPI
1361             if (pme->nnodes > 1)
1362             {
1363                 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1364             }
1365 #endif
1366
1367             unwrap_periodic_pmegrid(pme, grid);
1368         }
1369
1370         if (bCalcF)
1371         {
1372             /* interpolate forces for our local atoms */
1373
1374
1375             /* If we are running without parallelization,
1376              * atc->f is the actual force array, not a buffer,
1377              * therefore we should not clear it.
1378              */
1379             lambda  = grid_index < DO_Q ? lambda_q : lambda_lj;
1380             bClearF = (bFirst && PAR(cr));
1381 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1382             for (thread = 0; thread < pme->nthread; thread++)
1383             {
1384                 try
1385                 {
1386                     gather_f_bsplines(pme, grid, bClearF, atc,
1387                                       &atc->spline[thread],
1388                                       pme->bFEP ? (grid_index % 2 == 0 ? 1.0-lambda : lambda) : 1.0);
1389                 }
1390                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1391             }
1392
1393
1394             inc_nrnb(nrnb, eNR_GATHERFBSP,
1395                      pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
1396             /* Note: this wallcycle region is opened above inside an OpenMP
1397                region, so take care if refactoring code here. */
1398             wallcycle_stop(wcycle, ewcPME_GATHER);
1399         }
1400
1401         if (bCalcEnerVir)
1402         {
1403             /* This should only be called on the master thread
1404              * and after the threads have synchronized.
1405              */
1406             if (grid_index < 2)
1407             {
1408                 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &energy_AB[grid_index], vir_AB[grid_index]);
1409             }
1410             else
1411             {
1412                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &energy_AB[grid_index], vir_AB[grid_index]);
1413             }
1414         }
1415         bFirst = FALSE;
1416     } /* of grid_index-loop */
1417
1418     /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1419      * seven terms. */
1420
1421     if (pme->doLJ && pme->ljpme_combination_rule == eljpmeLB)
1422     {
1423         /* Loop over A- and B-state if we are doing FEP */
1424         for (fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1425         {
1426             real *local_c6 = nullptr, *local_sigma = nullptr, *RedistC6 = nullptr, *RedistSigma = nullptr;
1427             if (pme->nnodes == 1)
1428             {
1429                 if (pme->lb_buf1 == nullptr)
1430                 {
1431                     pme->lb_buf_nalloc = pme->atc[0].n;
1432                     snew(pme->lb_buf1, pme->lb_buf_nalloc);
1433                 }
1434                 pme->atc[0].coefficient = pme->lb_buf1;
1435                 switch (fep_state)
1436                 {
1437                     case 0:
1438                         local_c6      = c6A;
1439                         local_sigma   = sigmaA;
1440                         break;
1441                     case 1:
1442                         local_c6      = c6B;
1443                         local_sigma   = sigmaB;
1444                         break;
1445                     default:
1446                         gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1447                 }
1448             }
1449             else
1450             {
1451                 atc = &pme->atc[0];
1452                 switch (fep_state)
1453                 {
1454                     case 0:
1455                         RedistC6      = c6A;
1456                         RedistSigma   = sigmaA;
1457                         break;
1458                     case 1:
1459                         RedistC6      = c6B;
1460                         RedistSigma   = sigmaB;
1461                         break;
1462                     default:
1463                         gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1464                 }
1465                 wallcycle_start(wcycle, ewcPME_REDISTXF);
1466
1467                 do_redist_pos_coeffs(pme, cr, start, homenr, bFirst, x, RedistC6);
1468                 if (pme->lb_buf_nalloc < atc->n)
1469                 {
1470                     pme->lb_buf_nalloc = atc->nalloc;
1471                     srenew(pme->lb_buf1, pme->lb_buf_nalloc);
1472                     srenew(pme->lb_buf2, pme->lb_buf_nalloc);
1473                 }
1474                 local_c6 = pme->lb_buf1;
1475                 for (i = 0; i < atc->n; ++i)
1476                 {
1477                     local_c6[i] = atc->coefficient[i];
1478                 }
1479
1480                 do_redist_pos_coeffs(pme, cr, start, homenr, FALSE, x, RedistSigma);
1481                 local_sigma = pme->lb_buf2;
1482                 for (i = 0; i < atc->n; ++i)
1483                 {
1484                     local_sigma[i] = atc->coefficient[i];
1485                 }
1486
1487                 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1488             }
1489             calc_initial_lb_coeffs(pme, local_c6, local_sigma);
1490
1491             /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1492             for (grid_index = 2; grid_index < 9; ++grid_index)
1493             {
1494                 /* Unpack structure */
1495                 pmegrid    = &pme->pmegrid[grid_index];
1496                 fftgrid    = pme->fftgrid[grid_index];
1497                 pfft_setup = pme->pfft_setup[grid_index];
1498                 calc_next_lb_coeffs(pme, local_sigma);
1499                 grid = pmegrid->grid.grid;
1500
1501                 if (flags & GMX_PME_SPREAD)
1502                 {
1503                     wallcycle_start(wcycle, ewcPME_SPREAD);
1504                     /* Spread the c6 on a grid */
1505                     spread_on_grid(pme, &pme->atc[0], pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1506
1507                     if (bFirst)
1508                     {
1509                         inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
1510                     }
1511
1512                     inc_nrnb(nrnb, eNR_SPREADBSP,
1513                              pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
1514                     if (pme->nthread == 1)
1515                     {
1516                         wrap_periodic_pmegrid(pme, grid);
1517                         /* sum contributions to local grid from other nodes */
1518 #if GMX_MPI
1519                         if (pme->nnodes > 1)
1520                         {
1521                             gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1522                         }
1523 #endif
1524                         copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1525                     }
1526                     wallcycle_stop(wcycle, ewcPME_SPREAD);
1527                 }
1528                 /*Here we start a large thread parallel region*/
1529 #pragma omp parallel num_threads(pme->nthread) private(thread)
1530                 {
1531                     try
1532                     {
1533                         thread = gmx_omp_get_thread_num();
1534                         if (flags & GMX_PME_SOLVE)
1535                         {
1536                             /* do 3d-fft */
1537                             if (thread == 0)
1538                             {
1539                                 wallcycle_start(wcycle, ewcPME_FFT);
1540                             }
1541
1542                             gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
1543                                                        thread, wcycle);
1544                             if (thread == 0)
1545                             {
1546                                 wallcycle_stop(wcycle, ewcPME_FFT);
1547                             }
1548                         }
1549                     }
1550                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1551                 }
1552                 bFirst = FALSE;
1553             }
1554             if (flags & GMX_PME_SOLVE)
1555             {
1556                 /* solve in k-space for our local cells */
1557 #pragma omp parallel num_threads(pme->nthread) private(thread)
1558                 {
1559                     try
1560                     {
1561                         int loop_count;
1562                         thread = gmx_omp_get_thread_num();
1563                         if (thread == 0)
1564                         {
1565                             wallcycle_start(wcycle, ewcLJPME);
1566                         }
1567
1568                         loop_count =
1569                             solve_pme_lj_yzx(pme, &pme->cfftgrid[2], TRUE,
1570                                              scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1571                                              bCalcEnerVir,
1572                                              pme->nthread, thread);
1573                         if (thread == 0)
1574                         {
1575                             wallcycle_stop(wcycle, ewcLJPME);
1576                             inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1577                         }
1578                     }
1579                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1580                 }
1581             }
1582
1583             if (bCalcEnerVir)
1584             {
1585                 /* This should only be called on the master thread and
1586                  * after the threads have synchronized.
1587                  */
1588                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &energy_AB[2+fep_state], vir_AB[2+fep_state]);
1589             }
1590
1591             if (bBackFFT)
1592             {
1593                 bFirst = !pme->doCoulomb;
1594                 calc_initial_lb_coeffs(pme, local_c6, local_sigma);
1595                 for (grid_index = 8; grid_index >= 2; --grid_index)
1596                 {
1597                     /* Unpack structure */
1598                     pmegrid    = &pme->pmegrid[grid_index];
1599                     fftgrid    = pme->fftgrid[grid_index];
1600                     pfft_setup = pme->pfft_setup[grid_index];
1601                     grid       = pmegrid->grid.grid;
1602                     calc_next_lb_coeffs(pme, local_sigma);
1603 #pragma omp parallel num_threads(pme->nthread) private(thread)
1604                     {
1605                         try
1606                         {
1607                             thread = gmx_omp_get_thread_num();
1608                             /* do 3d-invfft */
1609                             if (thread == 0)
1610                             {
1611                                 wallcycle_start(wcycle, ewcPME_FFT);
1612                             }
1613
1614                             gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
1615                                                        thread, wcycle);
1616                             if (thread == 0)
1617                             {
1618                                 wallcycle_stop(wcycle, ewcPME_FFT);
1619
1620
1621                                 if (pme->nodeid == 0)
1622                                 {
1623                                     real ntot = pme->nkx*pme->nky*pme->nkz;
1624                                     npme  = static_cast<int>(ntot*std::log(ntot)/std::log(2.0));
1625                                     inc_nrnb(nrnb, eNR_FFT, 2*npme);
1626                                 }
1627                                 wallcycle_start(wcycle, ewcPME_GATHER);
1628                             }
1629
1630                             copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1631                         }
1632                         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1633                     } /*#pragma omp parallel*/
1634
1635                     /* distribute local grid to all nodes */
1636 #if GMX_MPI
1637                     if (pme->nnodes > 1)
1638                     {
1639                         gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1640                     }
1641 #endif
1642
1643                     unwrap_periodic_pmegrid(pme, grid);
1644
1645                     if (bCalcF)
1646                     {
1647                         /* interpolate forces for our local atoms */
1648                         bClearF = (bFirst && PAR(cr));
1649                         scale   = pme->bFEP ? (fep_state < 1 ? 1.0-lambda_lj : lambda_lj) : 1.0;
1650                         scale  *= lb_scale_factor[grid_index-2];
1651
1652 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1653                         for (thread = 0; thread < pme->nthread; thread++)
1654                         {
1655                             try
1656                             {
1657                                 gather_f_bsplines(pme, grid, bClearF, &pme->atc[0],
1658                                                   &pme->atc[0].spline[thread],
1659                                                   scale);
1660                             }
1661                             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1662                         }
1663
1664
1665                         inc_nrnb(nrnb, eNR_GATHERFBSP,
1666                                  pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
1667                     }
1668                     wallcycle_stop(wcycle, ewcPME_GATHER);
1669
1670                     bFirst = FALSE;
1671                 } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1672             }     /* if (bCalcF) */
1673         }         /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1674     }             /* if ((flags & GMX_PME_DO_LJ) && pme->ljpme_combination_rule == eljpmeLB) */
1675
1676     if (bCalcF && pme->nnodes > 1)
1677     {
1678         wallcycle_start(wcycle, ewcPME_REDISTXF);
1679         for (d = 0; d < pme->ndecompdim; d++)
1680         {
1681             atc = &pme->atc[d];
1682             if (d == pme->ndecompdim - 1)
1683             {
1684                 n_d = homenr;
1685                 f_d = f + start;
1686             }
1687             else
1688             {
1689                 n_d = pme->atc[d+1].n;
1690                 f_d = pme->atc[d+1].f;
1691             }
1692             if (DOMAINDECOMP(cr))
1693             {
1694                 dd_pmeredist_f(pme, atc, n_d, f_d,
1695                                d == pme->ndecompdim-1 && pme->bPPnode);
1696             }
1697         }
1698
1699         wallcycle_stop(wcycle, ewcPME_REDISTXF);
1700     }
1701
1702     if (bCalcEnerVir)
1703     {
1704         if (pme->doCoulomb)
1705         {
1706             if (!pme->bFEP_q)
1707             {
1708                 *energy_q = energy_AB[0];
1709                 m_add(vir_q, vir_AB[0], vir_q);
1710             }
1711             else
1712             {
1713                 *energy_q       = (1.0-lambda_q)*energy_AB[0] + lambda_q*energy_AB[1];
1714                 *dvdlambda_q   += energy_AB[1] - energy_AB[0];
1715                 for (i = 0; i < DIM; i++)
1716                 {
1717                     for (j = 0; j < DIM; j++)
1718                     {
1719                         vir_q[i][j] += (1.0-lambda_q)*vir_AB[0][i][j] +
1720                             lambda_q*vir_AB[1][i][j];
1721                     }
1722                 }
1723             }
1724             if (debug)
1725             {
1726                 fprintf(debug, "Electrostatic PME mesh energy: %g\n", *energy_q);
1727             }
1728         }
1729         else
1730         {
1731             *energy_q = 0;
1732         }
1733
1734         if (pme->doLJ)
1735         {
1736             if (!pme->bFEP_lj)
1737             {
1738                 *energy_lj = energy_AB[2];
1739                 m_add(vir_lj, vir_AB[2], vir_lj);
1740             }
1741             else
1742             {
1743                 *energy_lj     = (1.0-lambda_lj)*energy_AB[2] + lambda_lj*energy_AB[3];
1744                 *dvdlambda_lj += energy_AB[3] - energy_AB[2];
1745                 for (i = 0; i < DIM; i++)
1746                 {
1747                     for (j = 0; j < DIM; j++)
1748                     {
1749                         vir_lj[i][j] += (1.0-lambda_lj)*vir_AB[2][i][j] + lambda_lj*vir_AB[3][i][j];
1750                     }
1751                 }
1752             }
1753             if (debug)
1754             {
1755                 fprintf(debug, "Lennard-Jones PME mesh energy: %g\n", *energy_lj);
1756             }
1757         }
1758         else
1759         {
1760             *energy_lj = 0;
1761         }
1762     }
1763     return 0;
1764 }
1765
1766 void gmx_pme_destroy(gmx_pme_t *pme)
1767 {
1768     if (!pme)
1769     {
1770         return;
1771     }
1772
1773     delete pme->boxScaler;
1774
1775     sfree(pme->nnx);
1776     sfree(pme->nny);
1777     sfree(pme->nnz);
1778     sfree(pme->fshx);
1779     sfree(pme->fshy);
1780     sfree(pme->fshz);
1781
1782     for (int i = 0; i < pme->ngrids; ++i)
1783     {
1784         pmegrids_destroy(&pme->pmegrid[i]);
1785     }
1786     if (pme->pfft_setup)
1787     {
1788         for (int i = 0; i < pme->ngrids; ++i)
1789         {
1790             gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1791         }
1792     }
1793     sfree(pme->fftgrid);
1794     sfree(pme->cfftgrid);
1795     sfree(pme->pfft_setup);
1796
1797     for (int i = 0; i < std::max(1, pme->ndecompdim); i++) //pme->atc[0] is always allocated
1798     {
1799         destroy_atomcomm(&pme->atc[i]);
1800     }
1801
1802     for (int i = 0; i < DIM; i++)
1803     {
1804         sfree(pme->bsp_mod[i]);
1805     }
1806
1807     sfree(pme->lb_buf1);
1808     sfree(pme->lb_buf2);
1809
1810     sfree(pme->bufv);
1811     sfree(pme->bufr);
1812
1813     if (pme->solve_work)
1814     {
1815         pme_free_all_work(&pme->solve_work, pme->nthread);
1816     }
1817
1818     sfree(pme->sum_qgrid_tmp);
1819     sfree(pme->sum_qgrid_dd_tmp);
1820
1821     destroy_pme_spline_work(pme->spline_work);
1822
1823     if (pme_gpu_active(pme) && pme->gpu)
1824     {
1825         pme_gpu_destroy(pme->gpu);
1826     }
1827
1828     delete pme;
1829 }
1830
1831 void gmx_pme_reinit_atoms(const gmx_pme_t *pme, const int nAtoms, const real *charges)
1832 {
1833     if (pme_gpu_active(pme))
1834     {
1835         pme_gpu_reinit_atoms(pme->gpu, nAtoms, charges);
1836     }
1837     // TODO: handle the CPU case here; handle the whole t_mdatoms
1838 }