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/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/math/gmxcomplex.h"
  93 #include "gromacs/math/invertmatrix.h"
  94 #include "gromacs/math/units.h"
  95 #include "gromacs/math/vec.h"
  96 #include "gromacs/math/vectypes.h"
  97 #include "gromacs/mdtypes/commrec.h"
  98 #include "gromacs/mdtypes/forcerec.h"
  99 #include "gromacs/mdtypes/inputrec.h"
 100 #include "gromacs/mdtypes/md_enums.h"
 101 #include "gromacs/pbcutil/pbc.h"
 102 #include "gromacs/timing/cyclecounter.h"
 103 #include "gromacs/timing/wallcycle.h"
 104 #include "gromacs/timing/walltime_accounting.h"
 105 #include "gromacs/utility/basedefinitions.h"
 106 #include "gromacs/utility/exceptions.h"
 107 #include "gromacs/utility/fatalerror.h"
 108 #include "gromacs/utility/gmxmpi.h"
 109 #include "gromacs/utility/gmxomp.h"
 110 #include "gromacs/utility/logger.h"
 111 #include "gromacs/utility/real.h"
 112 #include "gromacs/utility/smalloc.h"
 113 #include "gromacs/utility/stringutil.h"
 114 #include "gromacs/utility/unique_cptr.h"
 115
 116 #include "calculate-spline-moduli.h"
 117 #include "pme-gather.h"
 118 #include "pme-gpu-internal.h"
 119 #include "pme-grid.h"
 120 #include "pme-internal.h"
 121 #include "pme-redistribute.h"
 122 #include "pme-solve.h"
 123 #include "pme-spline-work.h"
 124 #include "pme-spread.h"
 125
 126 /*! \brief Help build a descriptive message in \c error if there are
 127  * \c errorReasons why PME on GPU is not supported.
 128  *
 129  * \returns Whether the lack of errorReasons indicate there is support. */
 130 static bool
 131 addMessageIfNotSupported(const std::list<std::string> &errorReasons,
 132                          std::string                  *error)
 133 {
 134     bool foundErrorReasons = errorReasons.empty();
 135     if (!foundErrorReasons && error)
 136     {
 137         std::string regressionTestMarker = "PME GPU does not support";
 138         // this prefix is tested for in the regression tests script gmxtest.pl
 139         *error = regressionTestMarker + ": " + gmx::joinStrings(errorReasons, "; ") + ".";
 140     }
 141     return foundErrorReasons;
 142 }
 143
 144 bool pme_gpu_supports_build(std::string *error)
 145 {
 146     std::list<std::string> errorReasons;
 147     if (GMX_DOUBLE)
 148     {
 149         errorReasons.push_back("double precision");
 150     }
 151     if (GMX_GPU != GMX_GPU_CUDA)
 152     {
 153         errorReasons.push_back("non-CUDA build of GROMACS");
 154     }
 155     return addMessageIfNotSupported(errorReasons, error);
 156 }
 157
 158 bool pme_gpu_supports_input(const t_inputrec *ir, std::string *error)
 159 {
 160     std::list<std::string> errorReasons;
 161     if (!EEL_PME(ir->coulombtype))
 162     {
 163         errorReasons.push_back("systems that do not use PME for electrostatics");
 164     }
 165     if (ir->pme_order != 4)
 166     {
 167         errorReasons.push_back("interpolation orders other than 4");
 168     }
 169     if (ir->efep != efepNO)
 170     {
 171         errorReasons.push_back("free energy calculations (multiple grids)");
 172     }
 173     if (EVDW_PME(ir->vdwtype))
 174     {
 175         errorReasons.push_back("Lennard-Jones PME");
 176     }
 177     if (ir->cutoff_scheme == ecutsGROUP)
 178     {
 179         errorReasons.push_back("group cutoff scheme");
 180     }
 181     if (EI_TPI(ir->eI))
 182     {
 183         errorReasons.push_back("test particle insertion");
 184     }
 185     return addMessageIfNotSupported(errorReasons, error);
 186 }
 187
 188 /*! \brief \libinternal
 189  * Finds out if PME with given inputs is possible to run on GPU.
 190  * This function is an internal final check, validating the whole PME structure on creation,
 191  * but it still duplicates the preliminary checks from the above (externally exposed) pme_gpu_supports_input() - just in case.
 192  *
 193  * \param[in]  pme          The PME structure.
 194  * \param[out] error        The error message if the input is not supported on GPU.
 195  * \returns                 True if this PME input is possible to run on GPU, false otherwise.
 196  */
 197 static bool pme_gpu_check_restrictions(const gmx_pme_t *pme, std::string *error)
 198 {
 199     std::list<std::string> errorReasons;
 200     if (pme->nnodes != 1)
 201     {
 202         errorReasons.push_back("PME decomposition");
 203     }
 204     if (pme->pme_order != 4)
 205     {
 206         errorReasons.push_back("interpolation orders other than 4");
 207     }
 208     if (pme->bFEP)
 209     {
 210         errorReasons.push_back("free energy calculations (multiple grids)");
 211     }
 212     if (pme->doLJ)
 213     {
 214         errorReasons.push_back("Lennard-Jones PME");
 215     }
 216     if (GMX_DOUBLE)
 217     {
 218         errorReasons.push_back("double precision");
 219     }
 220     if (GMX_GPU != GMX_GPU_CUDA)
 221     {
 222         errorReasons.push_back("non-CUDA build of GROMACS");
 223     }
 224
 225     return addMessageIfNotSupported(errorReasons, error);
 226 }
 227
 228 PmeRunMode pme_run_mode(const gmx_pme_t *pme)
 229 {
 230     GMX_ASSERT(pme != nullptr, "Expecting valid PME data pointer");
 231     return pme->runMode;
 232 }
 233
 234 /*! \brief Number of bytes in a cache line.
 235  *
 236  * Must also be a multiple of the SIMD and SIMD4 register size, to
 237  * preserve alignment.
 238  */
 239 const int gmxCacheLineSize = 64;
 240
 241 //! Set up coordinate communication
 242 static void setup_coordinate_communication(pme_atomcomm_t *atc)
 243 {
 244     int nslab, n, i;
 245     int fw, bw;
 246
 247     nslab = atc->nslab;
 248
 249     n = 0;
 250     for (i = 1; i <= nslab/2; i++)
 251     {
 252         fw = (atc->nodeid + i) % nslab;
 253         bw = (atc->nodeid - i + nslab) % nslab;
 254         if (n < nslab - 1)
 255         {
 256             atc->node_dest[n] = fw;
 257             atc->node_src[n]  = bw;
 258             n++;
 259         }
 260         if (n < nslab - 1)
 261         {
 262             atc->node_dest[n] = bw;
 263             atc->node_src[n]  = fw;
 264             n++;
 265         }
 266     }
 267 }
 268
 269 /*! \brief Round \p n up to the next multiple of \p f */
 270 static int mult_up(int n, int f)
 271 {
 272     return ((n + f - 1)/f)*f;
 273 }
 274
 275 /*! \brief Return estimate of the load imbalance from the PME grid not being a good match for the number of PME ranks */
 276 static double estimate_pme_load_imbalance(struct gmx_pme_t *pme)
 277 {
 278     int    nma, nmi;
 279     double n1, n2, n3;
 280
 281     nma = pme->nnodes_major;
 282     nmi = pme->nnodes_minor;
 283
 284     n1 = mult_up(pme->nkx, nma)*mult_up(pme->nky, nmi)*pme->nkz;
 285     n2 = mult_up(pme->nkx, nma)*mult_up(pme->nkz, nmi)*pme->nky;
 286     n3 = mult_up(pme->nky, nma)*mult_up(pme->nkz, nmi)*pme->nkx;
 287
 288     /* pme_solve is roughly double the cost of an fft */
 289
 290     return (n1 + n2 + 3*n3)/(double)(6*pme->nkx*pme->nky*pme->nkz);
 291 }
 292
 293 /*! \brief Initialize atom communication data structure */
 294 static void init_atomcomm(struct gmx_pme_t *pme, pme_atomcomm_t *atc,
 295                           int dimind, gmx_bool bSpread)
 296 {
 297     int thread;
 298
 299     atc->dimind    = dimind;
 300     atc->nslab     = 1;
 301     atc->nodeid    = 0;
 302     atc->pd_nalloc = 0;
 303 #if GMX_MPI
 304     if (pme->nnodes > 1)
 305     {
 306         atc->mpi_comm = pme->mpi_comm_d[dimind];
 307         MPI_Comm_size(atc->mpi_comm, &atc->nslab);
 308         MPI_Comm_rank(atc->mpi_comm, &atc->nodeid);
 309     }
 310     if (debug)
 311     {
 312         fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", atc->dimind, atc->nslab, atc->nodeid);
 313     }
 314 #endif
 315
 316     atc->bSpread   = bSpread;
 317     atc->pme_order = pme->pme_order;
 318
 319     if (atc->nslab > 1)
 320     {
 321         snew(atc->node_dest, atc->nslab);
 322         snew(atc->node_src, atc->nslab);
 323         setup_coordinate_communication(atc);
 324
 325         snew(atc->count_thread, pme->nthread);
 326         for (thread = 0; thread < pme->nthread; thread++)
 327         {
 328             snew(atc->count_thread[thread], atc->nslab);
 329         }
 330         atc->count = atc->count_thread[0];
 331         snew(atc->rcount, atc->nslab);
 332         snew(atc->buf_index, atc->nslab);
 333     }
 334
 335     atc->nthread = pme->nthread;
 336     if (atc->nthread > 1)
 337     {
 338         snew(atc->thread_plist, atc->nthread);
 339     }
 340     snew(atc->spline, atc->nthread);
 341     for (thread = 0; thread < atc->nthread; thread++)
 342     {
 343         if (atc->nthread > 1)
 344         {
 345             snew(atc->thread_plist[thread].n, atc->nthread+2*gmxCacheLineSize);
 346             atc->thread_plist[thread].n += gmxCacheLineSize;
 347         }
 348     }
 349 }
 350
 351 /*! \brief Destroy an atom communication data structure and its child structs */
 352 static void destroy_atomcomm(pme_atomcomm_t *atc)
 353 {
 354     sfree(atc->pd);
 355     if (atc->nslab > 1)
 356     {
 357         sfree(atc->node_dest);
 358         sfree(atc->node_src);
 359         for (int i = 0; i < atc->nthread; i++)
 360         {
 361             sfree(atc->count_thread[i]);
 362         }
 363         sfree(atc->count_thread);
 364         sfree(atc->rcount);
 365         sfree(atc->buf_index);
 366
 367         sfree(atc->x);
 368         sfree(atc->coefficient);
 369         sfree(atc->f);
 370     }
 371     sfree(atc->idx);
 372     sfree(atc->fractx);
 373
 374     sfree(atc->thread_idx);
 375     for (int i = 0; i < atc->nthread; i++)
 376     {
 377         if (atc->nthread > 1)
 378         {
 379             int *n_ptr = atc->thread_plist[i].n - gmxCacheLineSize;
 380             sfree(n_ptr);
 381             sfree(atc->thread_plist[i].i);
 382         }
 383         sfree(atc->spline[i].ind);
 384         for (int d = 0; d < ZZ; d++)
 385         {
 386             sfree(atc->spline[i].theta[d]);
 387             sfree(atc->spline[i].dtheta[d]);
 388         }
 389         sfree_aligned(atc->spline[i].ptr_dtheta_z);
 390         sfree_aligned(atc->spline[i].ptr_theta_z);
 391     }
 392     if (atc->nthread > 1)
 393     {
 394         sfree(atc->thread_plist);
 395     }
 396     sfree(atc->spline);
 397 }
 398
 399 /*! \brief Initialize data structure for communication */
 400 static void
 401 init_overlap_comm(pme_overlap_t *  ol,
 402                   int              norder,
 403 #if GMX_MPI
 404                   MPI_Comm         comm,
 405 #endif
 406                   int              nnodes,
 407                   int              nodeid,
 408                   int              ndata,
 409                   int              commplainsize)
 410 {
 411     gmx_bool         bCont;
 412 #if GMX_MPI
 413     MPI_Status       stat;
 414
 415     ol->mpi_comm = comm;
 416 #endif
 417
 418     ol->nnodes = nnodes;
 419     ol->nodeid = nodeid;
 420
 421     /* Linear translation of the PME grid won't affect reciprocal space
 422      * calculations, so to optimize we only interpolate "upwards",
 423      * which also means we only have to consider overlap in one direction.
 424      * I.e., particles on this node might also be spread to grid indices
 425      * that belong to higher nodes (modulo nnodes)
 426      */
 427
 428     ol->s2g0.resize(ol->nnodes + 1);
 429     ol->s2g1.resize(ol->nnodes);
 430     if (debug)
 431     {
 432         fprintf(debug, "PME slab boundaries:");
 433     }
 434     for (int i = 0; i < nnodes; i++)
 435     {
 436         /* s2g0 the local interpolation grid start.
 437          * s2g1 the local interpolation grid end.
 438          * Since in calc_pidx we divide particles, and not grid lines,
 439          * spatially uniform along dimension x or y, we need to round
 440          * s2g0 down and s2g1 up.
 441          */
 442         ol->s2g0[i] = (i * ndata + 0) / nnodes;
 443         ol->s2g1[i] = ((i + 1) * ndata + nnodes - 1) / nnodes + norder - 1;
 444
 445         if (debug)
 446         {
 447             fprintf(debug, "  %3d %3d", ol->s2g0[i], ol->s2g1[i]);
 448         }
 449     }
 450     ol->s2g0[nnodes] = ndata;
 451     if (debug)
 452     {
 453         fprintf(debug, "\n");
 454     }
 455
 456     /* Determine with how many nodes we need to communicate the grid overlap */
 457     int testRankCount = 0;
 458     do
 459     {
 460         testRankCount++;
 461         bCont = FALSE;
 462         for (int i = 0; i < nnodes; i++)
 463         {
 464             if ((i + testRankCount <  nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount]) ||
 465                 (i + testRankCount >= nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount - nnodes] + ndata))
 466             {
 467                 bCont = TRUE;
 468             }
 469         }
 470     }
 471     while (bCont && testRankCount < nnodes);
 472
 473     ol->comm_data.resize(testRankCount - 1);
 474     ol->send_size = 0;
 475
 476     for (size_t b = 0; b < ol->comm_data.size(); b++)
 477     {
 478         pme_grid_comm_t *pgc = &ol->comm_data[b];
 479
 480         /* Send */
 481         pgc->send_id = (ol->nodeid + (b + 1)) % ol->nnodes;
 482         int fft_start = ol->s2g0[pgc->send_id];
 483         int fft_end   = ol->s2g0[pgc->send_id + 1];
 484         if (pgc->send_id < nodeid)
 485         {
 486             fft_start += ndata;
 487             fft_end   += ndata;
 488         }
 489         int send_index1  = ol->s2g1[nodeid];
 490         send_index1      = std::min(send_index1, fft_end);
 491         pgc->send_index0 = fft_start;
 492         pgc->send_nindex = std::max(0, send_index1 - pgc->send_index0);
 493         ol->send_size   += pgc->send_nindex;
 494
 495         /* We always start receiving to the first index of our slab */
 496         pgc->recv_id     = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
 497         fft_start        = ol->s2g0[ol->nodeid];
 498         fft_end          = ol->s2g0[ol->nodeid + 1];
 499         int recv_index1  = ol->s2g1[pgc->recv_id];
 500         if (pgc->recv_id > nodeid)
 501         {
 502             recv_index1 -= ndata;
 503         }
 504         recv_index1      = std::min(recv_index1, fft_end);
 505         pgc->recv_index0 = fft_start;
 506         pgc->recv_nindex = std::max(0, recv_index1 - pgc->recv_index0);
 507     }
 508
 509 #if GMX_MPI
 510     /* Communicate the buffer sizes to receive */
 511     for (size_t b = 0; b < ol->comm_data.size(); b++)
 512     {
 513         MPI_Sendrecv(&ol->send_size, 1, MPI_INT, ol->comm_data[b].send_id, b,
 514                      &ol->comm_data[b].recv_size, 1, MPI_INT, ol->comm_data[b].recv_id, b,
 515                      ol->mpi_comm, &stat);
 516     }
 517 #endif
 518
 519     /* For non-divisible grid we need pme_order iso pme_order-1 */
 520     ol->sendbuf.resize(norder * commplainsize);
 521     ol->recvbuf.resize(norder * commplainsize);
 522 }
 523
 524 int minimalPmeGridSize(int pmeOrder)
 525 {
 526     /* The actual grid size limitations are:
 527      *   serial:        >= pme_order
 528      *   DD, no OpenMP: >= 2*(pme_order - 1)
 529      *   DD, OpenMP:    >= pme_order + 1
 530      * But we use the maximum for simplicity since in practice there is not
 531      * much performance difference between pme_order and 2*(pme_order -1).
 532      */
 533     int minimalSize = 2*(pmeOrder - 1);
 534
 535     GMX_RELEASE_ASSERT(pmeOrder >= 3, "pmeOrder has to be >= 3");
 536     GMX_RELEASE_ASSERT(minimalSize >= pmeOrder + 1, "The grid size should be >= pmeOrder + 1");
 537
 538     return minimalSize;
 539 }
 540
 541 bool gmx_pme_check_restrictions(int pme_order,
 542                                 int nkx, int nky, int nkz,
 543                                 int numPmeDomainsAlongX,
 544                                 bool useThreads,
 545                                 bool errorsAreFatal)
 546 {
 547     if (pme_order > PME_ORDER_MAX)
 548     {
 549         if (!errorsAreFatal)
 550         {
 551             return false;
 552         }
 553
 554         std::string message = gmx::formatString(
 555                     "pme_order (%d) is larger than the maximum allowed value (%d). Modify and recompile the code if you really need such a high order.",
 556                     pme_order, PME_ORDER_MAX);
 557         GMX_THROW(InconsistentInputError(message));
 558     }
 559
 560     const int minGridSize = minimalPmeGridSize(pme_order);
 561     if (nkx < minGridSize ||
 562         nky < minGridSize ||
 563         nkz < minGridSize)
 564     {
 565         if (!errorsAreFatal)
 566         {
 567             return false;
 568         }
 569         std::string message = gmx::formatString(
 570                     "The PME grid sizes need to be >= 2*(pme_order-1) (%d)",
 571                     minGridSize);
 572         GMX_THROW(InconsistentInputError(message));
 573     }
 574
 575     /* Check for a limitation of the (current) sum_fftgrid_dd code.
 576      * We only allow multiple communication pulses in dim 1, not in dim 0.
 577      */
 578     if (useThreads && (nkx < numPmeDomainsAlongX*pme_order &&
 579                        nkx != numPmeDomainsAlongX*(pme_order - 1)))
 580     {
 581         if (!errorsAreFatal)
 582         {
 583             return false;
 584         }
 585         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.",
 586                   nkx/static_cast<double>(numPmeDomainsAlongX), pme_order);
 587     }
 588
 589     return true;
 590 }
 591
 592 /*! \brief Round \p enumerator */
 593 static int div_round_up(int enumerator, int denominator)
 594 {
 595     return (enumerator + denominator - 1)/denominator;
 596 }
 597
 598 gmx_pme_t *gmx_pme_init(const t_commrec     *cr,
 599                         const NumPmeDomains &numPmeDomains,
 600                         const t_inputrec    *ir,
 601                         int                  homenr,
 602                         gmx_bool             bFreeEnergy_q,
 603                         gmx_bool             bFreeEnergy_lj,
 604                         gmx_bool             bReproducible,
 605                         real                 ewaldcoeff_q,
 606                         real                 ewaldcoeff_lj,
 607                         int                  nthread,
 608                         PmeRunMode           runMode,
 609                         PmeGpu              *pmeGpu,
 610                         gmx_device_info_t   *gpuInfo,
 611                         PmeGpuProgramHandle  pmeGpuProgram,
 612                         const gmx::MDLogger  & /*mdlog*/)
 613 {
 614     int               use_threads, sum_use_threads, i;
 615     ivec              ndata;
 616
 617     if (debug)
 618     {
 619         fprintf(debug, "Creating PME data structures.\n");
 620     }
 621
 622     unique_cptr<gmx_pme_t, gmx_pme_destroy> pme(new gmx_pme_t());
 623
 624     pme->sum_qgrid_tmp       = nullptr;
 625     pme->sum_qgrid_dd_tmp    = nullptr;
 626
 627     pme->buf_nalloc          = 0;
 628
 629     pme->nnodes              = 1;
 630     pme->bPPnode             = TRUE;
 631
 632     pme->nnodes_major        = numPmeDomains.x;
 633     pme->nnodes_minor        = numPmeDomains.y;
 634
 635 #if GMX_MPI
 636     if (numPmeDomains.x*numPmeDomains.y > 1)
 637     {
 638         pme->mpi_comm = cr->mpi_comm_mygroup;
 639
 640         MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
 641         MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
 642         if (pme->nnodes != numPmeDomains.x*numPmeDomains.y)
 643         {
 644             gmx_incons("PME rank count mismatch");
 645         }
 646     }
 647     else
 648     {
 649         pme->mpi_comm = MPI_COMM_NULL;
 650     }
 651 #endif
 652
 653     if (pme->nnodes == 1)
 654     {
 655 #if GMX_MPI
 656         pme->mpi_comm_d[0] = MPI_COMM_NULL;
 657         pme->mpi_comm_d[1] = MPI_COMM_NULL;
 658 #endif
 659         pme->ndecompdim   = 0;
 660         pme->nodeid_major = 0;
 661         pme->nodeid_minor = 0;
 662 #if GMX_MPI
 663         pme->mpi_comm_d[0] = pme->mpi_comm_d[1] = MPI_COMM_NULL;
 664 #endif
 665     }
 666     else
 667     {
 668         if (numPmeDomains.y == 1)
 669         {
 670 #if GMX_MPI
 671             pme->mpi_comm_d[0] = pme->mpi_comm;
 672             pme->mpi_comm_d[1] = MPI_COMM_NULL;
 673 #endif
 674             pme->ndecompdim   = 1;
 675             pme->nodeid_major = pme->nodeid;
 676             pme->nodeid_minor = 0;
 677
 678         }
 679         else if (numPmeDomains.x == 1)
 680         {
 681 #if GMX_MPI
 682             pme->mpi_comm_d[0] = MPI_COMM_NULL;
 683             pme->mpi_comm_d[1] = pme->mpi_comm;
 684 #endif
 685             pme->ndecompdim   = 1;
 686             pme->nodeid_major = 0;
 687             pme->nodeid_minor = pme->nodeid;
 688         }
 689         else
 690         {
 691             if (pme->nnodes % numPmeDomains.x != 0)
 692             {
 693                 gmx_incons("For 2D PME decomposition, #PME ranks must be divisible by the number of domains along x");
 694             }
 695             pme->ndecompdim = 2;
 696
 697 #if GMX_MPI
 698             MPI_Comm_split(pme->mpi_comm, pme->nodeid % numPmeDomains.y,
 699                            pme->nodeid, &pme->mpi_comm_d[0]);  /* My communicator along major dimension */
 700             MPI_Comm_split(pme->mpi_comm, pme->nodeid/numPmeDomains.y,
 701                            pme->nodeid, &pme->mpi_comm_d[1]);  /* My communicator along minor dimension */
 702
 703             MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
 704             MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
 705             MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
 706             MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
 707 #endif
 708         }
 709         pme->bPPnode = thisRankHasDuty(cr, DUTY_PP);
 710     }
 711
 712     pme->nthread = nthread;
 713
 714     /* Check if any of the PME MPI ranks uses threads */
 715     use_threads = (pme->nthread > 1 ? 1 : 0);
 716 #if GMX_MPI
 717     if (pme->nnodes > 1)
 718     {
 719         MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT,
 720                       MPI_SUM, pme->mpi_comm);
 721     }
 722     else
 723 #endif
 724     {
 725         sum_use_threads = use_threads;
 726     }
 727     pme->bUseThreads = (sum_use_threads > 0);
 728
 729     if (ir->ePBC == epbcSCREW)
 730     {
 731         gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
 732     }
 733
 734     /* NOTE:
 735      * It is likely that the current gmx_pme_do() routine supports calculating
 736      * only Coulomb or LJ while gmx_pme_init() configures for both,
 737      * but that has never been tested.
 738      * It is likely that the current gmx_pme_do() routine supports calculating,
 739      * not calculating free-energy for Coulomb and/or LJ while gmx_pme_init()
 740      * configures with free-energy, but that has never been tested.
 741      */
 742     pme->doCoulomb     = EEL_PME(ir->coulombtype);
 743     pme->doLJ          = EVDW_PME(ir->vdwtype);
 744     pme->bFEP_q        = ((ir->efep != efepNO) && bFreeEnergy_q);
 745     pme->bFEP_lj       = ((ir->efep != efepNO) && bFreeEnergy_lj);
 746     pme->bFEP          = (pme->bFEP_q || pme->bFEP_lj);
 747     pme->nkx           = ir->nkx;
 748     pme->nky           = ir->nky;
 749     pme->nkz           = ir->nkz;
 750     pme->bP3M          = (ir->coulombtype == eelP3M_AD || getenv("GMX_PME_P3M") != nullptr);
 751     pme->pme_order     = ir->pme_order;
 752     pme->ewaldcoeff_q  = ewaldcoeff_q;
 753     pme->ewaldcoeff_lj = ewaldcoeff_lj;
 754
 755     /* Always constant electrostatics coefficients */
 756     pme->epsilon_r     = ir->epsilon_r;
 757
 758     /* Always constant LJ coefficients */
 759     pme->ljpme_combination_rule = ir->ljpme_combination_rule;
 760
 761     // The box requires scaling with nwalls = 2, we store that condition as well
 762     // as the scaling factor
 763     delete pme->boxScaler;
 764     pme->boxScaler = new EwaldBoxZScaler(*ir);
 765
 766     /* If we violate restrictions, generate a fatal error here */
 767     gmx_pme_check_restrictions(pme->pme_order,
 768                                pme->nkx, pme->nky, pme->nkz,
 769                                pme->nnodes_major,
 770                                pme->bUseThreads,
 771                                true);
 772
 773     if (pme->nnodes > 1)
 774     {
 775         double imbal;
 776
 777 #if GMX_MPI
 778         MPI_Type_contiguous(DIM, GMX_MPI_REAL, &(pme->rvec_mpi));
 779         MPI_Type_commit(&(pme->rvec_mpi));
 780 #endif
 781
 782         /* Note that the coefficient spreading and force gathering, which usually
 783          * takes about the same amount of time as FFT+solve_pme,
 784          * is always fully load balanced
 785          * (unless the coefficient distribution is inhomogeneous).
 786          */
 787
 788         imbal = estimate_pme_load_imbalance(pme.get());
 789         if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
 790         {
 791             fprintf(stderr,
 792                     "\n"
 793                     "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
 794                     "      For optimal PME load balancing\n"
 795                     "      PME grid_x (%d) and grid_y (%d) should be divisible by #PME_ranks_x (%d)\n"
 796                     "      and PME grid_y (%d) and grid_z (%d) should be divisible by #PME_ranks_y (%d)\n"
 797                     "\n",
 798                     (int)((imbal-1)*100 + 0.5),
 799                     pme->nkx, pme->nky, pme->nnodes_major,
 800                     pme->nky, pme->nkz, pme->nnodes_minor);
 801         }
 802     }
 803
 804     /* For non-divisible grid we need pme_order iso pme_order-1 */
 805     /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
 806      * y is always copied through a buffer: we don't need padding in z,
 807      * but we do need the overlap in x because of the communication order.
 808      */
 809     init_overlap_comm(&pme->overlap[0], pme->pme_order,
 810 #if GMX_MPI
 811                       pme->mpi_comm_d[0],
 812 #endif
 813                       pme->nnodes_major, pme->nodeid_major,
 814                       pme->nkx,
 815                       (div_round_up(pme->nky, pme->nnodes_minor)+pme->pme_order)*(pme->nkz+pme->pme_order-1));
 816
 817     /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
 818      * We do this with an offset buffer of equal size, so we need to allocate
 819      * extra for the offset. That's what the (+1)*pme->nkz is for.
 820      */
 821     init_overlap_comm(&pme->overlap[1], pme->pme_order,
 822 #if GMX_MPI
 823                       pme->mpi_comm_d[1],
 824 #endif
 825                       pme->nnodes_minor, pme->nodeid_minor,
 826                       pme->nky,
 827                       (div_round_up(pme->nkx, pme->nnodes_major)+pme->pme_order+1)*pme->nkz);
 828
 829     /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
 830      * Note that gmx_pme_check_restrictions checked for this already.
 831      */
 832     if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
 833     {
 834         gmx_incons("More than one communication pulse required for grid overlap communication along the major dimension while using threads");
 835     }
 836
 837     snew(pme->bsp_mod[XX], pme->nkx);
 838     snew(pme->bsp_mod[YY], pme->nky);
 839     snew(pme->bsp_mod[ZZ], pme->nkz);
 840
 841     pme->gpu     = pmeGpu; /* Carrying over the single GPU structure */
 842     pme->runMode = runMode;
 843
 844     /* The required size of the interpolation grid, including overlap.
 845      * The allocated size (pmegrid_n?) might be slightly larger.
 846      */
 847     pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] -
 848         pme->overlap[0].s2g0[pme->nodeid_major];
 849     pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] -
 850         pme->overlap[1].s2g0[pme->nodeid_minor];
 851     pme->pmegrid_nz_base = pme->nkz;
 852     pme->pmegrid_nz      = pme->pmegrid_nz_base + pme->pme_order - 1;
 853     set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
 854     pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
 855     pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
 856     pme->pmegrid_start_iz = 0;
 857
 858     make_gridindex_to_localindex(pme->nkx,
 859                                  pme->pmegrid_start_ix,
 860                                  pme->pmegrid_nx - (pme->pme_order-1),
 861                                  &pme->nnx, &pme->fshx);
 862     make_gridindex_to_localindex(pme->nky,
 863                                  pme->pmegrid_start_iy,
 864                                  pme->pmegrid_ny - (pme->pme_order-1),
 865                                  &pme->nny, &pme->fshy);
 866     make_gridindex_to_localindex(pme->nkz,
 867                                  pme->pmegrid_start_iz,
 868                                  pme->pmegrid_nz_base,
 869                                  &pme->nnz, &pme->fshz);
 870
 871     pme->spline_work = make_pme_spline_work(pme->pme_order);
 872
 873     ndata[0]    = pme->nkx;
 874     ndata[1]    = pme->nky;
 875     ndata[2]    = pme->nkz;
 876     /* It doesn't matter if we allocate too many grids here,
 877      * we only allocate and use the ones we need.
 878      */
 879     if (pme->doLJ)
 880     {
 881         pme->ngrids = ((ir->ljpme_combination_rule == eljpmeLB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
 882     }
 883     else
 884     {
 885         pme->ngrids = DO_Q;
 886     }
 887     snew(pme->fftgrid, pme->ngrids);
 888     snew(pme->cfftgrid, pme->ngrids);
 889     snew(pme->pfft_setup, pme->ngrids);
 890
 891     for (i = 0; i < pme->ngrids; ++i)
 892     {
 893         if ((i <  DO_Q && pme->doCoulomb && (i == 0 ||
 894                                              bFreeEnergy_q)) ||
 895             (i >= DO_Q && pme->doLJ && (i == 2 ||
 896                                         bFreeEnergy_lj ||
 897                                         ir->ljpme_combination_rule == eljpmeLB)))
 898         {
 899             pmegrids_init(&pme->pmegrid[i],
 900                           pme->pmegrid_nx, pme->pmegrid_ny, pme->pmegrid_nz,
 901                           pme->pmegrid_nz_base,
 902                           pme->pme_order,
 903                           pme->bUseThreads,
 904                           pme->nthread,
 905                           pme->overlap[0].s2g1[pme->nodeid_major]-pme->overlap[0].s2g0[pme->nodeid_major+1],
 906                           pme->overlap[1].s2g1[pme->nodeid_minor]-pme->overlap[1].s2g0[pme->nodeid_minor+1]);
 907             /* This routine will allocate the grid data to fit the FFTs */
 908             const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed) ? gmx::PinningPolicy::CanBePinned : gmx::PinningPolicy::CannotBePinned;
 909             gmx_parallel_3dfft_init(&pme->pfft_setup[i], ndata,
 910                                     &pme->fftgrid[i], &pme->cfftgrid[i],
 911                                     pme->mpi_comm_d,
 912                                     bReproducible, pme->nthread, allocateRealGridForGpu);
 913
 914         }
 915     }
 916
 917     if (!pme->bP3M)
 918     {
 919         /* Use plain SPME B-spline interpolation */
 920         make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 921     }
 922     else
 923     {
 924         /* Use the P3M grid-optimized influence function */
 925         make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 926     }
 927
 928     /* Use atc[0] for spreading */
 929     init_atomcomm(pme.get(), &pme->atc[0], numPmeDomains.x > 1 ? 0 : 1, TRUE);
 930     if (pme->ndecompdim >= 2)
 931     {
 932         init_atomcomm(pme.get(), &pme->atc[1], 1, FALSE);
 933     }
 934
 935     if (pme->nnodes == 1)
 936     {
 937         pme->atc[0].n = homenr;
 938         pme_realloc_atomcomm_things(&pme->atc[0]);
 939     }
 940
 941     pme->lb_buf1       = nullptr;
 942     pme->lb_buf2       = nullptr;
 943     pme->lb_buf_nalloc = 0;
 944
 945     if (pme_gpu_active(pme.get()))
 946     {
 947         if (!pme->gpu)
 948         {
 949             // Initial check of validity of the data
 950             std::string errorString;
 951             bool        canRunOnGpu = pme_gpu_check_restrictions(pme.get(), &errorString);
 952             if (!canRunOnGpu)
 953             {
 954                 GMX_THROW(gmx::NotImplementedError(errorString));
 955             }
 956         }
 957
 958         pme_gpu_reinit(pme.get(), gpuInfo, pmeGpuProgram);
 959     }
 960
 961     pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
 962
 963     // no exception was thrown during the init, so we hand over the PME structure handle
 964     return pme.release();
 965 }
 966
 967 void gmx_pme_reinit(struct gmx_pme_t **pmedata,
 968                     const t_commrec   *cr,
 969                     struct gmx_pme_t * pme_src,
 970                     const t_inputrec * ir,
 971                     const ivec         grid_size,
 972                     real               ewaldcoeff_q,
 973                     real               ewaldcoeff_lj)
 974 {
 975     int        homenr;
 976
 977     // Create a copy of t_inputrec fields that are used in gmx_pme_init().
 978     // TODO: This would be better as just copying a sub-structure that contains
 979     // all the PME parameters and nothing else.
 980     t_inputrec irc;
 981     irc.ePBC                   = ir->ePBC;
 982     irc.coulombtype            = ir->coulombtype;
 983     irc.vdwtype                = ir->vdwtype;
 984     irc.efep                   = ir->efep;
 985     irc.pme_order              = ir->pme_order;
 986     irc.epsilon_r              = ir->epsilon_r;
 987     irc.ljpme_combination_rule = ir->ljpme_combination_rule;
 988     irc.nkx                    = grid_size[XX];
 989     irc.nky                    = grid_size[YY];
 990     irc.nkz                    = grid_size[ZZ];
 991
 992     if (pme_src->nnodes == 1)
 993     {
 994         homenr = pme_src->atc[0].n;
 995     }
 996     else
 997     {
 998         homenr = -1;
 999     }
1000
1001     try
1002     {
1003         const gmx::MDLogger dummyLogger;
1004         // This is reinit which is currently only changing grid size/coefficients,
1005         // so we don't expect the actual logging.
1006         // TODO: when PME is an object, it should take reference to mdlog on construction and save it.
1007         GMX_ASSERT(pmedata, "Invalid PME pointer");
1008         NumPmeDomains numPmeDomains = { pme_src->nnodes_major, pme_src->nnodes_minor };
1009         *pmedata = gmx_pme_init(cr, numPmeDomains,
1010                                 &irc, homenr, pme_src->bFEP_q, pme_src->bFEP_lj, FALSE, ewaldcoeff_q, ewaldcoeff_lj,
1011                                 pme_src->nthread, pme_src->runMode, pme_src->gpu, nullptr, nullptr, dummyLogger);
1012         //TODO this is mostly passing around current values
1013     }
1014     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1015
1016     /* We can easily reuse the allocated pme grids in pme_src */
1017     reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
1018     /* We would like to reuse the fft grids, but that's harder */
1019 }
1020
1021 void gmx_pme_calc_energy(struct gmx_pme_t *pme, int n, rvec *x, real *q, real *V)
1022 {
1023     pme_atomcomm_t *atc;
1024     pmegrids_t     *grid;
1025
1026     if (pme->nnodes > 1)
1027     {
1028         gmx_incons("gmx_pme_calc_energy called in parallel");
1029     }
1030     if (pme->bFEP_q > 1)
1031     {
1032         gmx_incons("gmx_pme_calc_energy with free energy");
1033     }
1034
1035     atc            = &pme->atc_energy;
1036     atc->nthread   = 1;
1037     if (atc->spline == nullptr)
1038     {
1039         snew(atc->spline, atc->nthread);
1040     }
1041     atc->nslab     = 1;
1042     atc->bSpread   = TRUE;
1043     atc->pme_order = pme->pme_order;
1044     atc->n         = n;
1045     pme_realloc_atomcomm_things(atc);
1046     atc->x           = x;
1047     atc->coefficient = q;
1048
1049     /* We only use the A-charges grid */
1050     grid = &pme->pmegrid[PME_GRID_QA];
1051
1052     /* Only calculate the spline coefficients, don't actually spread */
1053     spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
1054
1055     *V = gather_energy_bsplines(pme, grid->grid.grid, atc);
1056 }
1057
1058 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
1059 static void
1060 calc_initial_lb_coeffs(struct gmx_pme_t *pme, real *local_c6, real *local_sigma)
1061 {
1062     int  i;
1063     for (i = 0; i < pme->atc[0].n; ++i)
1064     {
1065         real sigma4;
1066         sigma4                     = local_sigma[i];
1067         sigma4                     = sigma4*sigma4;
1068         sigma4                     = sigma4*sigma4;
1069         pme->atc[0].coefficient[i] = local_c6[i] / sigma4;
1070     }
1071 }
1072
1073 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
1074 static void
1075 calc_next_lb_coeffs(struct gmx_pme_t *pme, real *local_sigma)
1076 {
1077     int  i;
1078
1079     for (i = 0; i < pme->atc[0].n; ++i)
1080     {
1081         pme->atc[0].coefficient[i] *= local_sigma[i];
1082     }
1083 }
1084
1085 int gmx_pme_do(struct gmx_pme_t *pme,
1086                int start,       int homenr,
1087                rvec x[],        rvec f[],
1088                real chargeA[],  real chargeB[],
1089                real c6A[],      real c6B[],
1090                real sigmaA[],   real sigmaB[],
1091                matrix box,      const t_commrec *cr,
1092                int  maxshift_x, int maxshift_y,
1093                t_nrnb *nrnb,    gmx_wallcycle *wcycle,
1094                matrix vir_q,    matrix vir_lj,
1095                real *energy_q,  real *energy_lj,
1096                real lambda_q,   real lambda_lj,
1097                real *dvdlambda_q, real *dvdlambda_lj,
1098                int flags)
1099 {
1100     GMX_ASSERT(pme->runMode == PmeRunMode::CPU, "gmx_pme_do should not be called on the GPU PME run.");
1101
1102     int                  d, i, j, npme, grid_index, max_grid_index;
1103     int                  n_d;
1104     pme_atomcomm_t      *atc        = nullptr;
1105     pmegrids_t          *pmegrid    = nullptr;
1106     real                *grid       = nullptr;
1107     rvec                *f_d;
1108     real                *coefficient = nullptr;
1109     real                 energy_AB[4];
1110     matrix               vir_AB[4];
1111     real                 scale, lambda;
1112     gmx_bool             bClearF;
1113     gmx_parallel_3dfft_t pfft_setup;
1114     real              *  fftgrid;
1115     t_complex          * cfftgrid;
1116     int                  thread;
1117     gmx_bool             bFirst, bDoSplines;
1118     int                  fep_state;
1119     int                  fep_states_lj           = pme->bFEP_lj ? 2 : 1;
1120     const gmx_bool       bCalcEnerVir            = flags & GMX_PME_CALC_ENER_VIR;
1121     const gmx_bool       bBackFFT                = flags & (GMX_PME_CALC_F | GMX_PME_CALC_POT);
1122     const gmx_bool       bCalcF                  = flags & GMX_PME_CALC_F;
1123
1124     assert(pme->nnodes > 0);
1125     assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1126
1127     if (pme->nnodes > 1)
1128     {
1129         atc      = &pme->atc[0];
1130         atc->npd = homenr;
1131         if (atc->npd > atc->pd_nalloc)
1132         {
1133             atc->pd_nalloc = over_alloc_dd(atc->npd);
1134             srenew(atc->pd, atc->pd_nalloc);
1135         }
1136         for (d = pme->ndecompdim-1; d >= 0; d--)
1137         {
1138             atc           = &pme->atc[d];
1139             atc->maxshift = (atc->dimind == 0 ? maxshift_x : maxshift_y);
1140         }
1141     }
1142     else
1143     {
1144         atc = &pme->atc[0];
1145         /* This could be necessary for TPI */
1146         pme->atc[0].n = homenr;
1147         if (DOMAINDECOMP(cr))
1148         {
1149             pme_realloc_atomcomm_things(atc);
1150         }
1151         atc->x = x;
1152         atc->f = f;
1153     }
1154
1155     matrix scaledBox;
1156     pme->boxScaler->scaleBox(box, scaledBox);
1157
1158     gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1159     bFirst = TRUE;
1160
1161     /* For simplicity, we construct the splines for all particles if
1162      * more than one PME calculations is needed. Some optimization
1163      * could be done by keeping track of which atoms have splines
1164      * constructed, and construct new splines on each pass for atoms
1165      * that don't yet have them.
1166      */
1167
1168     bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1169
1170     /* We need a maximum of four separate PME calculations:
1171      * grid_index=0: Coulomb PME with charges from state A
1172      * grid_index=1: Coulomb PME with charges from state B
1173      * grid_index=2: LJ PME with C6 from state A
1174      * grid_index=3: LJ PME with C6 from state B
1175      * For Lorentz-Berthelot combination rules, a separate loop is used to
1176      * calculate all the terms
1177      */
1178
1179     /* If we are doing LJ-PME with LB, we only do Q here */
1180     max_grid_index = (pme->ljpme_combination_rule == eljpmeLB) ? DO_Q : DO_Q_AND_LJ;
1181
1182     for (grid_index = 0; grid_index < max_grid_index; ++grid_index)
1183     {
1184         /* Check if we should do calculations at this grid_index
1185          * If grid_index is odd we should be doing FEP
1186          * If grid_index < 2 we should be doing electrostatic PME
1187          * If grid_index >= 2 we should be doing LJ-PME
1188          */
1189         if ((grid_index <  DO_Q && (!pme->doCoulomb ||
1190                                     (grid_index == 1 && !pme->bFEP_q))) ||
1191             (grid_index >= DO_Q && (!pme->doLJ ||
1192                                     (grid_index == 3 && !pme->bFEP_lj))))
1193         {
1194             continue;
1195         }
1196         /* Unpack structure */
1197         pmegrid    = &pme->pmegrid[grid_index];
1198         fftgrid    = pme->fftgrid[grid_index];
1199         cfftgrid   = pme->cfftgrid[grid_index];
1200         pfft_setup = pme->pfft_setup[grid_index];
1201         switch (grid_index)
1202         {
1203             case 0: coefficient = chargeA + start; break;
1204             case 1: coefficient = chargeB + start; break;
1205             case 2: coefficient = c6A + start; break;
1206             case 3: coefficient = c6B + start; break;
1207         }
1208
1209         grid = pmegrid->grid.grid;
1210
1211         if (debug)
1212         {
1213             fprintf(debug, "PME: number of ranks = %d, rank = %d\n",
1214                     cr->nnodes, cr->nodeid);
1215             fprintf(debug, "Grid = %p\n", (void*)grid);
1216             if (grid == nullptr)
1217             {
1218                 gmx_fatal(FARGS, "No grid!");
1219             }
1220         }
1221
1222         if (pme->nnodes == 1)
1223         {
1224             atc->coefficient = coefficient;
1225         }
1226         else
1227         {
1228             wallcycle_start(wcycle, ewcPME_REDISTXF);
1229             do_redist_pos_coeffs(pme, cr, start, homenr, bFirst, x, coefficient);
1230
1231             wallcycle_stop(wcycle, ewcPME_REDISTXF);
1232         }
1233
1234         if (debug)
1235         {
1236             fprintf(debug, "Rank= %6d, pme local particles=%6d\n",
1237                     cr->nodeid, atc->n);
1238         }
1239
1240         if (flags & GMX_PME_SPREAD)
1241         {
1242             wallcycle_start(wcycle, ewcPME_SPREAD);
1243
1244             /* Spread the coefficients on a grid */
1245             spread_on_grid(pme, &pme->atc[0], pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1246
1247             if (bFirst)
1248             {
1249                 inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
1250             }
1251             inc_nrnb(nrnb, eNR_SPREADBSP,
1252                      pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
1253
1254             if (!pme->bUseThreads)
1255             {
1256                 wrap_periodic_pmegrid(pme, grid);
1257
1258                 /* sum contributions to local grid from other nodes */
1259 #if GMX_MPI
1260                 if (pme->nnodes > 1)
1261                 {
1262                     gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1263                 }
1264 #endif
1265
1266                 copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1267             }
1268
1269             wallcycle_stop(wcycle, ewcPME_SPREAD);
1270
1271             /* TODO If the OpenMP and single-threaded implementations
1272                converge, then spread_on_grid() and
1273                copy_pmegrid_to_fftgrid() will perhaps live in the same
1274                source file.
1275              */
1276         }
1277
1278         /* Here we start a large thread parallel region */
1279 #pragma omp parallel num_threads(pme->nthread) private(thread)
1280         {
1281             try
1282             {
1283                 thread = gmx_omp_get_thread_num();
1284                 if (flags & GMX_PME_SOLVE)
1285                 {
1286                     int loop_count;
1287
1288                     /* do 3d-fft */
1289                     if (thread == 0)
1290                     {
1291                         wallcycle_start(wcycle, ewcPME_FFT);
1292                     }
1293                     gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
1294                                                thread, wcycle);
1295                     if (thread == 0)
1296                     {
1297                         wallcycle_stop(wcycle, ewcPME_FFT);
1298                     }
1299
1300                     /* solve in k-space for our local cells */
1301                     if (thread == 0)
1302                     {
1303                         wallcycle_start(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1304                     }
1305                     if (grid_index < DO_Q)
1306                     {
1307                         loop_count =
1308                             solve_pme_yzx(pme, cfftgrid,
1309                                           scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1310                                           bCalcEnerVir,
1311                                           pme->nthread, thread);
1312                     }
1313                     else
1314                     {
1315                         loop_count =
1316                             solve_pme_lj_yzx(pme, &cfftgrid, FALSE,
1317                                              scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1318                                              bCalcEnerVir,
1319                                              pme->nthread, thread);
1320                     }
1321
1322                     if (thread == 0)
1323                     {
1324                         wallcycle_stop(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1325                         inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1326                     }
1327                 }
1328
1329                 if (bBackFFT)
1330                 {
1331                     /* do 3d-invfft */
1332                     if (thread == 0)
1333                     {
1334                         wallcycle_start(wcycle, ewcPME_FFT);
1335                     }
1336                     gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
1337                                                thread, wcycle);
1338                     if (thread == 0)
1339                     {
1340                         wallcycle_stop(wcycle, ewcPME_FFT);
1341
1342
1343                         if (pme->nodeid == 0)
1344                         {
1345                             real ntot = pme->nkx*pme->nky*pme->nkz;
1346                             npme  = static_cast<int>(ntot*std::log(ntot)/std::log(2.0));
1347                             inc_nrnb(nrnb, eNR_FFT, 2*npme);
1348                         }
1349
1350                         /* Note: this wallcycle region is closed below
1351                            outside an OpenMP region, so take care if
1352                            refactoring code here. */
1353                         wallcycle_start(wcycle, ewcPME_GATHER);
1354                     }
1355
1356                     copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1357                 }
1358             } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1359         }
1360         /* End of thread parallel section.
1361          * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1362          */
1363
1364         if (bBackFFT)
1365         {
1366             /* distribute local grid to all nodes */
1367 #if GMX_MPI
1368             if (pme->nnodes > 1)
1369             {
1370                 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1371             }
1372 #endif
1373
1374             unwrap_periodic_pmegrid(pme, grid);
1375         }
1376
1377         if (bCalcF)
1378         {
1379             /* interpolate forces for our local atoms */
1380
1381
1382             /* If we are running without parallelization,
1383              * atc->f is the actual force array, not a buffer,
1384              * therefore we should not clear it.
1385              */
1386             lambda  = grid_index < DO_Q ? lambda_q : lambda_lj;
1387             bClearF = (bFirst && PAR(cr));
1388 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1389             for (thread = 0; thread < pme->nthread; thread++)
1390             {
1391                 try
1392                 {
1393                     gather_f_bsplines(pme, grid, bClearF, atc,
1394                                       &atc->spline[thread],
1395                                       pme->bFEP ? (grid_index % 2 == 0 ? 1.0-lambda : lambda) : 1.0);
1396                 }
1397                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1398             }
1399
1400
1401             inc_nrnb(nrnb, eNR_GATHERFBSP,
1402                      pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
1403             /* Note: this wallcycle region is opened above inside an OpenMP
1404                region, so take care if refactoring code here. */
1405             wallcycle_stop(wcycle, ewcPME_GATHER);
1406         }
1407
1408         if (bCalcEnerVir)
1409         {
1410             /* This should only be called on the master thread
1411              * and after the threads have synchronized.
1412              */
1413             if (grid_index < 2)
1414             {
1415                 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &energy_AB[grid_index], vir_AB[grid_index]);
1416             }
1417             else
1418             {
1419                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &energy_AB[grid_index], vir_AB[grid_index]);
1420             }
1421         }
1422         bFirst = FALSE;
1423     } /* of grid_index-loop */
1424
1425     /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1426      * seven terms. */
1427
1428     if (pme->doLJ && pme->ljpme_combination_rule == eljpmeLB)
1429     {
1430         /* Loop over A- and B-state if we are doing FEP */
1431         for (fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1432         {
1433             real *local_c6 = nullptr, *local_sigma = nullptr, *RedistC6 = nullptr, *RedistSigma = nullptr;
1434             if (pme->nnodes == 1)
1435             {
1436                 if (pme->lb_buf1 == nullptr)
1437                 {
1438                     pme->lb_buf_nalloc = pme->atc[0].n;
1439                     snew(pme->lb_buf1, pme->lb_buf_nalloc);
1440                 }
1441                 pme->atc[0].coefficient = pme->lb_buf1;
1442                 switch (fep_state)
1443                 {
1444                     case 0:
1445                         local_c6      = c6A;
1446                         local_sigma   = sigmaA;
1447                         break;
1448                     case 1:
1449                         local_c6      = c6B;
1450                         local_sigma   = sigmaB;
1451                         break;
1452                     default:
1453                         gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1454                 }
1455             }
1456             else
1457             {
1458                 atc = &pme->atc[0];
1459                 switch (fep_state)
1460                 {
1461                     case 0:
1462                         RedistC6      = c6A;
1463                         RedistSigma   = sigmaA;
1464                         break;
1465                     case 1:
1466                         RedistC6      = c6B;
1467                         RedistSigma   = sigmaB;
1468                         break;
1469                     default:
1470                         gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1471                 }
1472                 wallcycle_start(wcycle, ewcPME_REDISTXF);
1473
1474                 do_redist_pos_coeffs(pme, cr, start, homenr, bFirst, x, RedistC6);
1475                 if (pme->lb_buf_nalloc < atc->n)
1476                 {
1477                     pme->lb_buf_nalloc = atc->nalloc;
1478                     srenew(pme->lb_buf1, pme->lb_buf_nalloc);
1479                     srenew(pme->lb_buf2, pme->lb_buf_nalloc);
1480                 }
1481                 local_c6 = pme->lb_buf1;
1482                 for (i = 0; i < atc->n; ++i)
1483                 {
1484                     local_c6[i] = atc->coefficient[i];
1485                 }
1486
1487                 do_redist_pos_coeffs(pme, cr, start, homenr, FALSE, x, RedistSigma);
1488                 local_sigma = pme->lb_buf2;
1489                 for (i = 0; i < atc->n; ++i)
1490                 {
1491                     local_sigma[i] = atc->coefficient[i];
1492                 }
1493
1494                 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1495             }
1496             calc_initial_lb_coeffs(pme, local_c6, local_sigma);
1497
1498             /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1499             for (grid_index = 2; grid_index < 9; ++grid_index)
1500             {
1501                 /* Unpack structure */
1502                 pmegrid    = &pme->pmegrid[grid_index];
1503                 fftgrid    = pme->fftgrid[grid_index];
1504                 pfft_setup = pme->pfft_setup[grid_index];
1505                 calc_next_lb_coeffs(pme, local_sigma);
1506                 grid = pmegrid->grid.grid;
1507
1508                 if (flags & GMX_PME_SPREAD)
1509                 {
1510                     wallcycle_start(wcycle, ewcPME_SPREAD);
1511                     /* Spread the c6 on a grid */
1512                     spread_on_grid(pme, &pme->atc[0], pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1513
1514                     if (bFirst)
1515                     {
1516                         inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
1517                     }
1518
1519                     inc_nrnb(nrnb, eNR_SPREADBSP,
1520                              pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
1521                     if (pme->nthread == 1)
1522                     {
1523                         wrap_periodic_pmegrid(pme, grid);
1524                         /* sum contributions to local grid from other nodes */
1525 #if GMX_MPI
1526                         if (pme->nnodes > 1)
1527                         {
1528                             gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1529                         }
1530 #endif
1531                         copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1532                     }
1533                     wallcycle_stop(wcycle, ewcPME_SPREAD);
1534                 }
1535                 /*Here we start a large thread parallel region*/
1536 #pragma omp parallel num_threads(pme->nthread) private(thread)
1537                 {
1538                     try
1539                     {
1540                         thread = gmx_omp_get_thread_num();
1541                         if (flags & GMX_PME_SOLVE)
1542                         {
1543                             /* do 3d-fft */
1544                             if (thread == 0)
1545                             {
1546                                 wallcycle_start(wcycle, ewcPME_FFT);
1547                             }
1548
1549                             gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
1550                                                        thread, wcycle);
1551                             if (thread == 0)
1552                             {
1553                                 wallcycle_stop(wcycle, ewcPME_FFT);
1554                             }
1555                         }
1556                     }
1557                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1558                 }
1559                 bFirst = FALSE;
1560             }
1561             if (flags & GMX_PME_SOLVE)
1562             {
1563                 /* solve in k-space for our local cells */
1564 #pragma omp parallel num_threads(pme->nthread) private(thread)
1565                 {
1566                     try
1567                     {
1568                         int loop_count;
1569                         thread = gmx_omp_get_thread_num();
1570                         if (thread == 0)
1571                         {
1572                             wallcycle_start(wcycle, ewcLJPME);
1573                         }
1574
1575                         loop_count =
1576                             solve_pme_lj_yzx(pme, &pme->cfftgrid[2], TRUE,
1577                                              scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1578                                              bCalcEnerVir,
1579                                              pme->nthread, thread);
1580                         if (thread == 0)
1581                         {
1582                             wallcycle_stop(wcycle, ewcLJPME);
1583                             inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1584                         }
1585                     }
1586                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1587                 }
1588             }
1589
1590             if (bCalcEnerVir)
1591             {
1592                 /* This should only be called on the master thread and
1593                  * after the threads have synchronized.
1594                  */
1595                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &energy_AB[2+fep_state], vir_AB[2+fep_state]);
1596             }
1597
1598             if (bBackFFT)
1599             {
1600                 bFirst = !pme->doCoulomb;
1601                 calc_initial_lb_coeffs(pme, local_c6, local_sigma);
1602                 for (grid_index = 8; grid_index >= 2; --grid_index)
1603                 {
1604                     /* Unpack structure */
1605                     pmegrid    = &pme->pmegrid[grid_index];
1606                     fftgrid    = pme->fftgrid[grid_index];
1607                     pfft_setup = pme->pfft_setup[grid_index];
1608                     grid       = pmegrid->grid.grid;
1609                     calc_next_lb_coeffs(pme, local_sigma);
1610 #pragma omp parallel num_threads(pme->nthread) private(thread)
1611                     {
1612                         try
1613                         {
1614                             thread = gmx_omp_get_thread_num();
1615                             /* do 3d-invfft */
1616                             if (thread == 0)
1617                             {
1618                                 wallcycle_start(wcycle, ewcPME_FFT);
1619                             }
1620
1621                             gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
1622                                                        thread, wcycle);
1623                             if (thread == 0)
1624                             {
1625                                 wallcycle_stop(wcycle, ewcPME_FFT);
1626
1627
1628                                 if (pme->nodeid == 0)
1629                                 {
1630                                     real ntot = pme->nkx*pme->nky*pme->nkz;
1631                                     npme  = static_cast<int>(ntot*std::log(ntot)/std::log(2.0));
1632                                     inc_nrnb(nrnb, eNR_FFT, 2*npme);
1633                                 }
1634                                 wallcycle_start(wcycle, ewcPME_GATHER);
1635                             }
1636
1637                             copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1638                         }
1639                         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1640                     } /*#pragma omp parallel*/
1641
1642                     /* distribute local grid to all nodes */
1643 #if GMX_MPI
1644                     if (pme->nnodes > 1)
1645                     {
1646                         gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1647                     }
1648 #endif
1649
1650                     unwrap_periodic_pmegrid(pme, grid);
1651
1652                     if (bCalcF)
1653                     {
1654                         /* interpolate forces for our local atoms */
1655                         bClearF = (bFirst && PAR(cr));
1656                         scale   = pme->bFEP ? (fep_state < 1 ? 1.0-lambda_lj : lambda_lj) : 1.0;
1657                         scale  *= lb_scale_factor[grid_index-2];
1658
1659 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1660                         for (thread = 0; thread < pme->nthread; thread++)
1661                         {
1662                             try
1663                             {
1664                                 gather_f_bsplines(pme, grid, bClearF, &pme->atc[0],
1665                                                   &pme->atc[0].spline[thread],
1666                                                   scale);
1667                             }
1668                             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1669                         }
1670
1671
1672                         inc_nrnb(nrnb, eNR_GATHERFBSP,
1673                                  pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
1674                     }
1675                     wallcycle_stop(wcycle, ewcPME_GATHER);
1676
1677                     bFirst = FALSE;
1678                 } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1679             }     /* if (bCalcF) */
1680         }         /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1681     }             /* if ((flags & GMX_PME_DO_LJ) && pme->ljpme_combination_rule == eljpmeLB) */
1682
1683     if (bCalcF && pme->nnodes > 1)
1684     {
1685         wallcycle_start(wcycle, ewcPME_REDISTXF);
1686         for (d = 0; d < pme->ndecompdim; d++)
1687         {
1688             atc = &pme->atc[d];
1689             if (d == pme->ndecompdim - 1)
1690             {
1691                 n_d = homenr;
1692                 f_d = f + start;
1693             }
1694             else
1695             {
1696                 n_d = pme->atc[d+1].n;
1697                 f_d = pme->atc[d+1].f;
1698             }
1699             if (DOMAINDECOMP(cr))
1700             {
1701                 dd_pmeredist_f(pme, atc, n_d, f_d,
1702                                d == pme->ndecompdim-1 && pme->bPPnode);
1703             }
1704         }
1705
1706         wallcycle_stop(wcycle, ewcPME_REDISTXF);
1707     }
1708
1709     if (bCalcEnerVir)
1710     {
1711         if (pme->doCoulomb)
1712         {
1713             if (!pme->bFEP_q)
1714             {
1715                 *energy_q = energy_AB[0];
1716                 m_add(vir_q, vir_AB[0], vir_q);
1717             }
1718             else
1719             {
1720                 *energy_q       = (1.0-lambda_q)*energy_AB[0] + lambda_q*energy_AB[1];
1721                 *dvdlambda_q   += energy_AB[1] - energy_AB[0];
1722                 for (i = 0; i < DIM; i++)
1723                 {
1724                     for (j = 0; j < DIM; j++)
1725                     {
1726                         vir_q[i][j] += (1.0-lambda_q)*vir_AB[0][i][j] +
1727                             lambda_q*vir_AB[1][i][j];
1728                     }
1729                 }
1730             }
1731             if (debug)
1732             {
1733                 fprintf(debug, "Electrostatic PME mesh energy: %g\n", *energy_q);
1734             }
1735         }
1736         else
1737         {
1738             *energy_q = 0;
1739         }
1740
1741         if (pme->doLJ)
1742         {
1743             if (!pme->bFEP_lj)
1744             {
1745                 *energy_lj = energy_AB[2];
1746                 m_add(vir_lj, vir_AB[2], vir_lj);
1747             }
1748             else
1749             {
1750                 *energy_lj     = (1.0-lambda_lj)*energy_AB[2] + lambda_lj*energy_AB[3];
1751                 *dvdlambda_lj += energy_AB[3] - energy_AB[2];
1752                 for (i = 0; i < DIM; i++)
1753                 {
1754                     for (j = 0; j < DIM; j++)
1755                     {
1756                         vir_lj[i][j] += (1.0-lambda_lj)*vir_AB[2][i][j] + lambda_lj*vir_AB[3][i][j];
1757                     }
1758                 }
1759             }
1760             if (debug)
1761             {
1762                 fprintf(debug, "Lennard-Jones PME mesh energy: %g\n", *energy_lj);
1763             }
1764         }
1765         else
1766         {
1767             *energy_lj = 0;
1768         }
1769     }
1770     return 0;
1771 }
1772
1773 void gmx_pme_destroy(gmx_pme_t *pme)
1774 {
1775     if (!pme)
1776     {
1777         return;
1778     }
1779
1780     delete pme->boxScaler;
1781
1782     sfree(pme->nnx);
1783     sfree(pme->nny);
1784     sfree(pme->nnz);
1785     sfree(pme->fshx);
1786     sfree(pme->fshy);
1787     sfree(pme->fshz);
1788
1789     for (int i = 0; i < pme->ngrids; ++i)
1790     {
1791         pmegrids_destroy(&pme->pmegrid[i]);
1792     }
1793     if (pme->pfft_setup)
1794     {
1795         for (int i = 0; i < pme->ngrids; ++i)
1796         {
1797             gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1798         }
1799     }
1800     sfree(pme->fftgrid);
1801     sfree(pme->cfftgrid);
1802     sfree(pme->pfft_setup);
1803
1804     for (int i = 0; i < std::max(1, pme->ndecompdim); i++) //pme->atc[0] is always allocated
1805     {
1806         destroy_atomcomm(&pme->atc[i]);
1807     }
1808
1809     for (int i = 0; i < DIM; i++)
1810     {
1811         sfree(pme->bsp_mod[i]);
1812     }
1813
1814     sfree(pme->lb_buf1);
1815     sfree(pme->lb_buf2);
1816
1817     sfree(pme->bufv);
1818     sfree(pme->bufr);
1819
1820     if (pme->solve_work)
1821     {
1822         pme_free_all_work(&pme->solve_work, pme->nthread);
1823     }
1824
1825     sfree(pme->sum_qgrid_tmp);
1826     sfree(pme->sum_qgrid_dd_tmp);
1827
1828     destroy_pme_spline_work(pme->spline_work);
1829
1830     if (pme_gpu_active(pme) && pme->gpu)
1831     {
1832         pme_gpu_destroy(pme->gpu);
1833     }
1834
1835     delete pme;
1836 }
1837
1838 void gmx_pme_reinit_atoms(const gmx_pme_t *pme, const int nAtoms, const real *charges)
1839 {
1840     if (pme_gpu_active(pme))
1841     {
1842         pme_gpu_reinit_atoms(pme->gpu, nAtoms, charges);
1843     }
1844     // TODO: handle the CPU case here; handle the whole t_mdatoms
1845 }