src/gromacs/ewald/pme.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016,2017 The GROMACS development team.
   7  * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
   8  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   9  * and including many others, as listed in the AUTHORS file in the
  10  * top-level source directory and at http://www.gromacs.org.
  11  *
  12  * GROMACS is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public License
  14  * as published by the Free Software Foundation; either version 2.1
  15  * of the License, or (at your option) any later version.
  16  *
  17  * GROMACS is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with GROMACS; if not, see
  24  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  25  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  26  *
  27  * If you want to redistribute modifications to GROMACS, please
  28  * consider that scientific software is very special. Version
  29  * control is crucial - bugs must be traceable. We will be happy to
  30  * consider code for inclusion in the official distribution, but
  31  * derived work must not be called official GROMACS. Details are found
  32  * in the README & COPYING files - if they are missing, get the
  33  * official version at http://www.gromacs.org.
  34  *
  35  * To help us fund GROMACS development, we humbly ask that you cite
  36  * the research papers on the package. Check out http://www.gromacs.org.
  37  */
  38 /*! \internal \file
  39  *
  40  * \brief This file contains function definitions necessary for
  41  * computing energies and forces for the PME long-ranged part (Coulomb
  42  * and LJ).
  43  *
  44  * \author Erik Lindahl <erik@kth.se>
  45  * \author Berk Hess <hess@kth.se>
  46  * \ingroup module_ewald
  47  */
  48 /* IMPORTANT FOR DEVELOPERS:
  49  *
  50  * Triclinic pme stuff isn't entirely trivial, and we've experienced
  51  * some bugs during development (many of them due to me). To avoid
  52  * this in the future, please check the following things if you make
  53  * changes in this file:
  54  *
  55  * 1. You should obtain identical (at least to the PME precision)
  56  *    energies, forces, and virial for
  57  *    a rectangular box and a triclinic one where the z (or y) axis is
  58  *    tilted a whole box side. For instance you could use these boxes:
  59  *
  60  *    rectangular       triclinic
  61  *     2  0  0           2  0  0
  62  *     0  2  0           0  2  0
  63  *     0  0  6           2  2  6
  64  *
  65  * 2. You should check the energy conservation in a triclinic box.
  66  *
  67  * It might seem an overkill, but better safe than sorry.
  68  * /Erik 001109
  69  */
  70
  71 #include "gmxpre.h"
  72
  73 #include "pme.h"
  74
  75 #include "config.h"
  76
  77 #include <cassert>
  78 #include <cmath>
  79 #include <cstdio>
  80 #include <cstdlib>
  81 #include <cstring>
  82
  83 #include <algorithm>
  84 #include <list>
  85
  86 #include "gromacs/domdec/domdec.h"
  87 #include "gromacs/ewald/ewald-utils.h"
  88 #include "gromacs/fft/parallel_3dfft.h"
  89 #include "gromacs/fileio/pdbio.h"
  90 #include "gromacs/gmxlib/network.h"
  91 #include "gromacs/gmxlib/nrnb.h"
  92 #include "gromacs/hardware/hw_info.h"
  93 #include "gromacs/math/gmxcomplex.h"
  94 #include "gromacs/math/invertmatrix.h"
  95 #include "gromacs/math/units.h"
  96 #include "gromacs/math/vec.h"
  97 #include "gromacs/math/vectypes.h"
  98 #include "gromacs/mdtypes/commrec.h"
  99 #include "gromacs/mdtypes/forcerec.h"
 100 #include "gromacs/mdtypes/inputrec.h"
 101 #include "gromacs/mdtypes/md_enums.h"
 102 #include "gromacs/pbcutil/pbc.h"
 103 #include "gromacs/timing/cyclecounter.h"
 104 #include "gromacs/timing/wallcycle.h"
 105 #include "gromacs/timing/walltime_accounting.h"
 106 #include "gromacs/topology/topology.h"
 107 #include "gromacs/utility/basedefinitions.h"
 108 #include "gromacs/utility/cstringutil.h"
 109 #include "gromacs/utility/exceptions.h"
 110 #include "gromacs/utility/fatalerror.h"
 111 #include "gromacs/utility/gmxmpi.h"
 112 #include "gromacs/utility/gmxomp.h"
 113 #include "gromacs/utility/logger.h"
 114 #include "gromacs/utility/real.h"
 115 #include "gromacs/utility/smalloc.h"
 116 #include "gromacs/utility/stringutil.h"
 117 #include "gromacs/utility/unique_cptr.h"
 118
 119 #include "calculate-spline-moduli.h"
 120 #include "pme-gather.h"
 121 #include "pme-gpu-internal.h"
 122 #include "pme-grid.h"
 123 #include "pme-internal.h"
 124 #include "pme-redistribute.h"
 125 #include "pme-solve.h"
 126 #include "pme-spline-work.h"
 127 #include "pme-spread.h"
 128
 129 /*! \brief Help build a descriptive message in \c error if there are
 130  * \c errorReasons why PME on GPU is not supported.
 131  *
 132  * \returns Whether the lack of errorReasons indicate there is support. */
 133 static bool
 134 addMessageIfNotSupported(const std::list<std::string> &errorReasons,
 135                          std::string                  *error)
 136 {
 137     bool isSupported = errorReasons.empty();
 138     if (!isSupported && error)
 139     {
 140         std::string regressionTestMarker = "PME GPU does not support";
 141         // this prefix is tested for in the regression tests script gmxtest.pl
 142         *error = regressionTestMarker;
 143         if (errorReasons.size() == 1)
 144         {
 145             *error += " " + errorReasons.back();
 146         }
 147         else
 148         {
 149             *error += ": " + gmx::joinStrings(errorReasons, "; ");
 150         }
 151         *error += ".";
 152     }
 153     return isSupported;
 154 }
 155
 156 bool pme_gpu_supports_build(std::string *error)
 157 {
 158     std::list<std::string> errorReasons;
 159     if (GMX_DOUBLE)
 160     {
 161         errorReasons.emplace_back("a double-precision build");
 162     }
 163     if (GMX_GPU == GMX_GPU_NONE)
 164     {
 165         errorReasons.emplace_back("a non-GPU build");
 166     }
 167     return addMessageIfNotSupported(errorReasons, error);
 168 }
 169
 170 bool pme_gpu_supports_hardware(const gmx_hw_info_t &hwinfo,
 171                                std::string         *error)
 172 {
 173     std::list<std::string> errorReasons;
 174     if (GMX_GPU == GMX_GPU_OPENCL)
 175     {
 176         if (!areAllGpuDevicesFromAmd(hwinfo.gpu_info))
 177         {
 178             errorReasons.emplace_back("non-AMD devices");
 179         }
 180 #ifdef __APPLE__
 181         errorReasons.emplace_back("Apple OS X operating system");
 182 #endif
 183     }
 184     return addMessageIfNotSupported(errorReasons, error);
 185 }
 186
 187 bool pme_gpu_supports_input(const t_inputrec &ir, const gmx_mtop_t &mtop, std::string *error)
 188 {
 189     std::list<std::string> errorReasons;
 190     if (!EEL_PME(ir.coulombtype))
 191     {
 192         errorReasons.emplace_back("systems that do not use PME for electrostatics");
 193     }
 194     if (ir.pme_order != 4)
 195     {
 196         errorReasons.emplace_back("interpolation orders other than 4");
 197     }
 198     if (ir.efep != efepNO)
 199     {
 200         if (gmx_mtop_has_perturbed_charges(mtop))
 201         {
 202             errorReasons.emplace_back("free energy calculations with perturbed charges (multiple grids)");
 203         }
 204     }
 205     if (EVDW_PME(ir.vdwtype))
 206     {
 207         errorReasons.emplace_back("Lennard-Jones PME");
 208     }
 209     if (ir.cutoff_scheme == ecutsGROUP)
 210     {
 211         errorReasons.emplace_back("group cutoff scheme");
 212     }
 213     if (!EI_DYNAMICS(ir.eI))
 214     {
 215         errorReasons.emplace_back("not a dynamical integrator");
 216     }
 217     return addMessageIfNotSupported(errorReasons, error);
 218 }
 219
 220 /*! \brief \libinternal
 221  * Finds out if PME with given inputs is possible to run on GPU.
 222  * This function is an internal final check, validating the whole PME structure on creation,
 223  * but it still duplicates the preliminary checks from the above (externally exposed) pme_gpu_supports_input() - just in case.
 224  *
 225  * \param[in]  pme          The PME structure.
 226  * \param[out] error        The error message if the input is not supported on GPU.
 227  * \returns                 True if this PME input is possible to run on GPU, false otherwise.
 228  */
 229 static bool pme_gpu_check_restrictions(const gmx_pme_t *pme, std::string *error)
 230 {
 231     std::list<std::string> errorReasons;
 232     if (pme->nnodes != 1)
 233     {
 234         errorReasons.emplace_back("PME decomposition");
 235     }
 236     if (pme->pme_order != 4)
 237     {
 238         errorReasons.emplace_back("interpolation orders other than 4");
 239     }
 240     if (pme->bFEP)
 241     {
 242         errorReasons.emplace_back("free energy calculations (multiple grids)");
 243     }
 244     if (pme->doLJ)
 245     {
 246         errorReasons.emplace_back("Lennard-Jones PME");
 247     }
 248     if (GMX_DOUBLE)
 249     {
 250         errorReasons.emplace_back("double precision");
 251     }
 252     if (GMX_GPU == GMX_GPU_NONE)
 253     {
 254         errorReasons.emplace_back("non-GPU build of GROMACS");
 255     }
 256
 257     return addMessageIfNotSupported(errorReasons, error);
 258 }
 259
 260 PmeRunMode pme_run_mode(const gmx_pme_t *pme)
 261 {
 262     GMX_ASSERT(pme != nullptr, "Expecting valid PME data pointer");
 263     return pme->runMode;
 264 }
 265
 266 gmx::PinningPolicy pme_get_pinning_policy()
 267 {
 268     return gmx::PinningPolicy::PinnedIfSupported;
 269 }
 270
 271 /*! \brief Number of bytes in a cache line.
 272  *
 273  * Must also be a multiple of the SIMD and SIMD4 register size, to
 274  * preserve alignment.
 275  */
 276 const int gmxCacheLineSize = 64;
 277
 278 //! Set up coordinate communication
 279 static void setup_coordinate_communication(pme_atomcomm_t *atc)
 280 {
 281     int nslab, n, i;
 282     int fw, bw;
 283
 284     nslab = atc->nslab;
 285
 286     n = 0;
 287     for (i = 1; i <= nslab/2; i++)
 288     {
 289         fw = (atc->nodeid + i) % nslab;
 290         bw = (atc->nodeid - i + nslab) % nslab;
 291         if (n < nslab - 1)
 292         {
 293             atc->node_dest[n] = fw;
 294             atc->node_src[n]  = bw;
 295             n++;
 296         }
 297         if (n < nslab - 1)
 298         {
 299             atc->node_dest[n] = bw;
 300             atc->node_src[n]  = fw;
 301             n++;
 302         }
 303     }
 304 }
 305
 306 /*! \brief Round \p n up to the next multiple of \p f */
 307 static int mult_up(int n, int f)
 308 {
 309     return ((n + f - 1)/f)*f;
 310 }
 311
 312 /*! \brief Return estimate of the load imbalance from the PME grid not being a good match for the number of PME ranks */
 313 static double estimate_pme_load_imbalance(struct gmx_pme_t *pme)
 314 {
 315     int    nma, nmi;
 316     double n1, n2, n3;
 317
 318     nma = pme->nnodes_major;
 319     nmi = pme->nnodes_minor;
 320
 321     n1 = mult_up(pme->nkx, nma)*mult_up(pme->nky, nmi)*pme->nkz;
 322     n2 = mult_up(pme->nkx, nma)*mult_up(pme->nkz, nmi)*pme->nky;
 323     n3 = mult_up(pme->nky, nma)*mult_up(pme->nkz, nmi)*pme->nkx;
 324
 325     /* pme_solve is roughly double the cost of an fft */
 326
 327     return (n1 + n2 + 3*n3)/static_cast<double>(6*pme->nkx*pme->nky*pme->nkz);
 328 }
 329
 330 /*! \brief Initialize atom communication data structure */
 331 static void init_atomcomm(struct gmx_pme_t *pme, pme_atomcomm_t *atc,
 332                           int dimind, gmx_bool bSpread)
 333 {
 334     int thread;
 335
 336     atc->dimind    = dimind;
 337     atc->nslab     = 1;
 338     atc->nodeid    = 0;
 339     atc->pd_nalloc = 0;
 340 #if GMX_MPI
 341     if (pme->nnodes > 1)
 342     {
 343         atc->mpi_comm = pme->mpi_comm_d[dimind];
 344         MPI_Comm_size(atc->mpi_comm, &atc->nslab);
 345         MPI_Comm_rank(atc->mpi_comm, &atc->nodeid);
 346     }
 347     if (debug)
 348     {
 349         fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", atc->dimind, atc->nslab, atc->nodeid);
 350     }
 351 #endif
 352
 353     atc->bSpread   = bSpread;
 354     atc->pme_order = pme->pme_order;
 355
 356     if (atc->nslab > 1)
 357     {
 358         snew(atc->node_dest, atc->nslab);
 359         snew(atc->node_src, atc->nslab);
 360         setup_coordinate_communication(atc);
 361
 362         snew(atc->count_thread, pme->nthread);
 363         for (thread = 0; thread < pme->nthread; thread++)
 364         {
 365             snew(atc->count_thread[thread], atc->nslab);
 366         }
 367         atc->count = atc->count_thread[0];
 368         snew(atc->rcount, atc->nslab);
 369         snew(atc->buf_index, atc->nslab);
 370     }
 371
 372     atc->nthread = pme->nthread;
 373     if (atc->nthread > 1)
 374     {
 375         snew(atc->thread_plist, atc->nthread);
 376     }
 377     snew(atc->spline, atc->nthread);
 378     for (thread = 0; thread < atc->nthread; thread++)
 379     {
 380         if (atc->nthread > 1)
 381         {
 382             snew(atc->thread_plist[thread].n, atc->nthread+2*gmxCacheLineSize);
 383             atc->thread_plist[thread].n += gmxCacheLineSize;
 384         }
 385     }
 386 }
 387
 388 /*! \brief Destroy an atom communication data structure and its child structs */
 389 static void destroy_atomcomm(pme_atomcomm_t *atc)
 390 {
 391     sfree(atc->pd);
 392     if (atc->nslab > 1)
 393     {
 394         sfree(atc->node_dest);
 395         sfree(atc->node_src);
 396         for (int i = 0; i < atc->nthread; i++)
 397         {
 398             sfree(atc->count_thread[i]);
 399         }
 400         sfree(atc->count_thread);
 401         sfree(atc->rcount);
 402         sfree(atc->buf_index);
 403
 404         sfree(atc->x);
 405         sfree(atc->coefficient);
 406         sfree(atc->f);
 407     }
 408     sfree(atc->idx);
 409     sfree(atc->fractx);
 410
 411     sfree(atc->thread_idx);
 412     for (int i = 0; i < atc->nthread; i++)
 413     {
 414         if (atc->nthread > 1)
 415         {
 416             int *n_ptr = atc->thread_plist[i].n - gmxCacheLineSize;
 417             sfree(n_ptr);
 418             sfree(atc->thread_plist[i].i);
 419         }
 420         sfree(atc->spline[i].ind);
 421         for (int d = 0; d < ZZ; d++)
 422         {
 423             sfree(atc->spline[i].theta[d]);
 424             sfree(atc->spline[i].dtheta[d]);
 425         }
 426         sfree_aligned(atc->spline[i].ptr_dtheta_z);
 427         sfree_aligned(atc->spline[i].ptr_theta_z);
 428     }
 429     if (atc->nthread > 1)
 430     {
 431         sfree(atc->thread_plist);
 432     }
 433     sfree(atc->spline);
 434 }
 435
 436 /*! \brief Initialize data structure for communication */
 437 static void
 438 init_overlap_comm(pme_overlap_t *  ol,
 439                   int              norder,
 440 #if GMX_MPI
 441                   MPI_Comm         comm,
 442 #endif
 443                   int              nnodes,
 444                   int              nodeid,
 445                   int              ndata,
 446                   int              commplainsize)
 447 {
 448     gmx_bool         bCont;
 449 #if GMX_MPI
 450     MPI_Status       stat;
 451
 452     ol->mpi_comm = comm;
 453 #endif
 454
 455     ol->nnodes = nnodes;
 456     ol->nodeid = nodeid;
 457
 458     /* Linear translation of the PME grid won't affect reciprocal space
 459      * calculations, so to optimize we only interpolate "upwards",
 460      * which also means we only have to consider overlap in one direction.
 461      * I.e., particles on this node might also be spread to grid indices
 462      * that belong to higher nodes (modulo nnodes)
 463      */
 464
 465     ol->s2g0.resize(ol->nnodes + 1);
 466     ol->s2g1.resize(ol->nnodes);
 467     if (debug)
 468     {
 469         fprintf(debug, "PME slab boundaries:");
 470     }
 471     for (int i = 0; i < nnodes; i++)
 472     {
 473         /* s2g0 the local interpolation grid start.
 474          * s2g1 the local interpolation grid end.
 475          * Since in calc_pidx we divide particles, and not grid lines,
 476          * spatially uniform along dimension x or y, we need to round
 477          * s2g0 down and s2g1 up.
 478          */
 479         ol->s2g0[i] = (i * ndata + 0) / nnodes;
 480         ol->s2g1[i] = ((i + 1) * ndata + nnodes - 1) / nnodes + norder - 1;
 481
 482         if (debug)
 483         {
 484             fprintf(debug, "  %3d %3d", ol->s2g0[i], ol->s2g1[i]);
 485         }
 486     }
 487     ol->s2g0[nnodes] = ndata;
 488     if (debug)
 489     {
 490         fprintf(debug, "\n");
 491     }
 492
 493     /* Determine with how many nodes we need to communicate the grid overlap */
 494     int testRankCount = 0;
 495     do
 496     {
 497         testRankCount++;
 498         bCont = FALSE;
 499         for (int i = 0; i < nnodes; i++)
 500         {
 501             if ((i + testRankCount <  nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount]) ||
 502                 (i + testRankCount >= nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount - nnodes] + ndata))
 503             {
 504                 bCont = TRUE;
 505             }
 506         }
 507     }
 508     while (bCont && testRankCount < nnodes);
 509
 510     ol->comm_data.resize(testRankCount - 1);
 511     ol->send_size = 0;
 512
 513     for (size_t b = 0; b < ol->comm_data.size(); b++)
 514     {
 515         pme_grid_comm_t *pgc = &ol->comm_data[b];
 516
 517         /* Send */
 518         pgc->send_id = (ol->nodeid + (b + 1)) % ol->nnodes;
 519         int fft_start = ol->s2g0[pgc->send_id];
 520         int fft_end   = ol->s2g0[pgc->send_id + 1];
 521         if (pgc->send_id < nodeid)
 522         {
 523             fft_start += ndata;
 524             fft_end   += ndata;
 525         }
 526         int send_index1  = ol->s2g1[nodeid];
 527         send_index1      = std::min(send_index1, fft_end);
 528         pgc->send_index0 = fft_start;
 529         pgc->send_nindex = std::max(0, send_index1 - pgc->send_index0);
 530         ol->send_size   += pgc->send_nindex;
 531
 532         /* We always start receiving to the first index of our slab */
 533         pgc->recv_id     = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
 534         fft_start        = ol->s2g0[ol->nodeid];
 535         fft_end          = ol->s2g0[ol->nodeid + 1];
 536         int recv_index1  = ol->s2g1[pgc->recv_id];
 537         if (pgc->recv_id > nodeid)
 538         {
 539             recv_index1 -= ndata;
 540         }
 541         recv_index1      = std::min(recv_index1, fft_end);
 542         pgc->recv_index0 = fft_start;
 543         pgc->recv_nindex = std::max(0, recv_index1 - pgc->recv_index0);
 544     }
 545
 546 #if GMX_MPI
 547     /* Communicate the buffer sizes to receive */
 548     for (size_t b = 0; b < ol->comm_data.size(); b++)
 549     {
 550         MPI_Sendrecv(&ol->send_size, 1, MPI_INT, ol->comm_data[b].send_id, b,
 551                      &ol->comm_data[b].recv_size, 1, MPI_INT, ol->comm_data[b].recv_id, b,
 552                      ol->mpi_comm, &stat);
 553     }
 554 #endif
 555
 556     /* For non-divisible grid we need pme_order iso pme_order-1 */
 557     ol->sendbuf.resize(norder * commplainsize);
 558     ol->recvbuf.resize(norder * commplainsize);
 559 }
 560
 561 int minimalPmeGridSize(int pmeOrder)
 562 {
 563     /* The actual grid size limitations are:
 564      *   serial:        >= pme_order
 565      *   DD, no OpenMP: >= 2*(pme_order - 1)
 566      *   DD, OpenMP:    >= pme_order + 1
 567      * But we use the maximum for simplicity since in practice there is not
 568      * much performance difference between pme_order and 2*(pme_order -1).
 569      */
 570     int minimalSize = 2*(pmeOrder - 1);
 571
 572     GMX_RELEASE_ASSERT(pmeOrder >= 3, "pmeOrder has to be >= 3");
 573     GMX_RELEASE_ASSERT(minimalSize >= pmeOrder + 1, "The grid size should be >= pmeOrder + 1");
 574
 575     return minimalSize;
 576 }
 577
 578 bool gmx_pme_check_restrictions(int pme_order,
 579                                 int nkx, int nky, int nkz,
 580                                 int numPmeDomainsAlongX,
 581                                 bool useThreads,
 582                                 bool errorsAreFatal)
 583 {
 584     if (pme_order > PME_ORDER_MAX)
 585     {
 586         if (!errorsAreFatal)
 587         {
 588             return false;
 589         }
 590
 591         std::string message = gmx::formatString(
 592                     "pme_order (%d) is larger than the maximum allowed value (%d). Modify and recompile the code if you really need such a high order.",
 593                     pme_order, PME_ORDER_MAX);
 594         GMX_THROW(gmx::InconsistentInputError(message));
 595     }
 596
 597     const int minGridSize = minimalPmeGridSize(pme_order);
 598     if (nkx < minGridSize ||
 599         nky < minGridSize ||
 600         nkz < minGridSize)
 601     {
 602         if (!errorsAreFatal)
 603         {
 604             return false;
 605         }
 606         std::string message = gmx::formatString(
 607                     "The PME grid sizes need to be >= 2*(pme_order-1) (%d)",
 608                     minGridSize);
 609         GMX_THROW(gmx::InconsistentInputError(message));
 610     }
 611
 612     /* Check for a limitation of the (current) sum_fftgrid_dd code.
 613      * We only allow multiple communication pulses in dim 1, not in dim 0.
 614      */
 615     if (useThreads && (nkx < numPmeDomainsAlongX*pme_order &&
 616                        nkx != numPmeDomainsAlongX*(pme_order - 1)))
 617     {
 618         if (!errorsAreFatal)
 619         {
 620             return false;
 621         }
 622         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.",
 623                   nkx/static_cast<double>(numPmeDomainsAlongX), pme_order);
 624     }
 625
 626     return true;
 627 }
 628
 629 /*! \brief Round \p enumerator */
 630 static int div_round_up(int enumerator, int denominator)
 631 {
 632     return (enumerator + denominator - 1)/denominator;
 633 }
 634
 635 gmx_pme_t *gmx_pme_init(const t_commrec         *cr,
 636                         const NumPmeDomains     &numPmeDomains,
 637                         const t_inputrec        *ir,
 638                         int                      homenr,
 639                         gmx_bool                 bFreeEnergy_q,
 640                         gmx_bool                 bFreeEnergy_lj,
 641                         gmx_bool                 bReproducible,
 642                         real                     ewaldcoeff_q,
 643                         real                     ewaldcoeff_lj,
 644                         int                      nthread,
 645                         PmeRunMode               runMode,
 646                         PmeGpu                  *pmeGpu,
 647                         const gmx_device_info_t *gpuInfo,
 648                         PmeGpuProgramHandle      pmeGpuProgram,
 649                         const gmx::MDLogger      & /*mdlog*/)
 650 {
 651     int               use_threads, sum_use_threads, i;
 652     ivec              ndata;
 653
 654     if (debug)
 655     {
 656         fprintf(debug, "Creating PME data structures.\n");
 657     }
 658
 659     gmx::unique_cptr<gmx_pme_t, gmx_pme_destroy> pme(new gmx_pme_t());
 660
 661     pme->sum_qgrid_tmp       = nullptr;
 662     pme->sum_qgrid_dd_tmp    = nullptr;
 663
 664     pme->buf_nalloc          = 0;
 665
 666     pme->nnodes              = 1;
 667     pme->bPPnode             = TRUE;
 668
 669     pme->nnodes_major        = numPmeDomains.x;
 670     pme->nnodes_minor        = numPmeDomains.y;
 671
 672 #if GMX_MPI
 673     if (numPmeDomains.x*numPmeDomains.y > 1)
 674     {
 675         pme->mpi_comm = cr->mpi_comm_mygroup;
 676
 677         MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
 678         MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
 679         if (pme->nnodes != numPmeDomains.x*numPmeDomains.y)
 680         {
 681             gmx_incons("PME rank count mismatch");
 682         }
 683     }
 684     else
 685     {
 686         pme->mpi_comm = MPI_COMM_NULL;
 687     }
 688 #endif
 689
 690     if (pme->nnodes == 1)
 691     {
 692 #if GMX_MPI
 693         pme->mpi_comm_d[0] = MPI_COMM_NULL;
 694         pme->mpi_comm_d[1] = MPI_COMM_NULL;
 695 #endif
 696         pme->ndecompdim   = 0;
 697         pme->nodeid_major = 0;
 698         pme->nodeid_minor = 0;
 699 #if GMX_MPI
 700         pme->mpi_comm_d[0] = pme->mpi_comm_d[1] = MPI_COMM_NULL;
 701 #endif
 702     }
 703     else
 704     {
 705         if (numPmeDomains.y == 1)
 706         {
 707 #if GMX_MPI
 708             pme->mpi_comm_d[0] = pme->mpi_comm;
 709             pme->mpi_comm_d[1] = MPI_COMM_NULL;
 710 #endif
 711             pme->ndecompdim   = 1;
 712             pme->nodeid_major = pme->nodeid;
 713             pme->nodeid_minor = 0;
 714
 715         }
 716         else if (numPmeDomains.x == 1)
 717         {
 718 #if GMX_MPI
 719             pme->mpi_comm_d[0] = MPI_COMM_NULL;
 720             pme->mpi_comm_d[1] = pme->mpi_comm;
 721 #endif
 722             pme->ndecompdim   = 1;
 723             pme->nodeid_major = 0;
 724             pme->nodeid_minor = pme->nodeid;
 725         }
 726         else
 727         {
 728             if (pme->nnodes % numPmeDomains.x != 0)
 729             {
 730                 gmx_incons("For 2D PME decomposition, #PME ranks must be divisible by the number of domains along x");
 731             }
 732             pme->ndecompdim = 2;
 733
 734 #if GMX_MPI
 735             MPI_Comm_split(pme->mpi_comm, pme->nodeid % numPmeDomains.y,
 736                            pme->nodeid, &pme->mpi_comm_d[0]);  /* My communicator along major dimension */
 737             MPI_Comm_split(pme->mpi_comm, pme->nodeid/numPmeDomains.y,
 738                            pme->nodeid, &pme->mpi_comm_d[1]);  /* My communicator along minor dimension */
 739
 740             MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
 741             MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
 742             MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
 743             MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
 744 #endif
 745         }
 746         pme->bPPnode = thisRankHasDuty(cr, DUTY_PP);
 747     }
 748
 749     pme->nthread = nthread;
 750
 751     /* Check if any of the PME MPI ranks uses threads */
 752     use_threads = (pme->nthread > 1 ? 1 : 0);
 753 #if GMX_MPI
 754     if (pme->nnodes > 1)
 755     {
 756         MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT,
 757                       MPI_SUM, pme->mpi_comm);
 758     }
 759     else
 760 #endif
 761     {
 762         sum_use_threads = use_threads;
 763     }
 764     pme->bUseThreads = (sum_use_threads > 0);
 765
 766     if (ir->ePBC == epbcSCREW)
 767     {
 768         gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
 769     }
 770
 771     /* NOTE:
 772      * It is likely that the current gmx_pme_do() routine supports calculating
 773      * only Coulomb or LJ while gmx_pme_init() configures for both,
 774      * but that has never been tested.
 775      * It is likely that the current gmx_pme_do() routine supports calculating,
 776      * not calculating free-energy for Coulomb and/or LJ while gmx_pme_init()
 777      * configures with free-energy, but that has never been tested.
 778      */
 779     pme->doCoulomb     = EEL_PME(ir->coulombtype);
 780     pme->doLJ          = EVDW_PME(ir->vdwtype);
 781     pme->bFEP_q        = ((ir->efep != efepNO) && bFreeEnergy_q);
 782     pme->bFEP_lj       = ((ir->efep != efepNO) && bFreeEnergy_lj);
 783     pme->bFEP          = (pme->bFEP_q || pme->bFEP_lj);
 784     pme->nkx           = ir->nkx;
 785     pme->nky           = ir->nky;
 786     pme->nkz           = ir->nkz;
 787     pme->bP3M          = (ir->coulombtype == eelP3M_AD || getenv("GMX_PME_P3M") != nullptr);
 788     pme->pme_order     = ir->pme_order;
 789     pme->ewaldcoeff_q  = ewaldcoeff_q;
 790     pme->ewaldcoeff_lj = ewaldcoeff_lj;
 791
 792     /* Always constant electrostatics coefficients */
 793     pme->epsilon_r     = ir->epsilon_r;
 794
 795     /* Always constant LJ coefficients */
 796     pme->ljpme_combination_rule = ir->ljpme_combination_rule;
 797
 798     // The box requires scaling with nwalls = 2, we store that condition as well
 799     // as the scaling factor
 800     delete pme->boxScaler;
 801     pme->boxScaler = new EwaldBoxZScaler(*ir);
 802
 803     /* If we violate restrictions, generate a fatal error here */
 804     gmx_pme_check_restrictions(pme->pme_order,
 805                                pme->nkx, pme->nky, pme->nkz,
 806                                pme->nnodes_major,
 807                                pme->bUseThreads,
 808                                true);
 809
 810     if (pme->nnodes > 1)
 811     {
 812         double imbal;
 813
 814 #if GMX_MPI
 815         MPI_Type_contiguous(DIM, GMX_MPI_REAL, &(pme->rvec_mpi));
 816         MPI_Type_commit(&(pme->rvec_mpi));
 817 #endif
 818
 819         /* Note that the coefficient spreading and force gathering, which usually
 820          * takes about the same amount of time as FFT+solve_pme,
 821          * is always fully load balanced
 822          * (unless the coefficient distribution is inhomogeneous).
 823          */
 824
 825         imbal = estimate_pme_load_imbalance(pme.get());
 826         if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
 827         {
 828             fprintf(stderr,
 829                     "\n"
 830                     "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
 831                     "      For optimal PME load balancing\n"
 832                     "      PME grid_x (%d) and grid_y (%d) should be divisible by #PME_ranks_x (%d)\n"
 833                     "      and PME grid_y (%d) and grid_z (%d) should be divisible by #PME_ranks_y (%d)\n"
 834                     "\n",
 835                     gmx::roundToInt((imbal-1)*100),
 836                     pme->nkx, pme->nky, pme->nnodes_major,
 837                     pme->nky, pme->nkz, pme->nnodes_minor);
 838         }
 839     }
 840
 841     /* For non-divisible grid we need pme_order iso pme_order-1 */
 842     /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
 843      * y is always copied through a buffer: we don't need padding in z,
 844      * but we do need the overlap in x because of the communication order.
 845      */
 846     init_overlap_comm(&pme->overlap[0], pme->pme_order,
 847 #if GMX_MPI
 848                       pme->mpi_comm_d[0],
 849 #endif
 850                       pme->nnodes_major, pme->nodeid_major,
 851                       pme->nkx,
 852                       (div_round_up(pme->nky, pme->nnodes_minor)+pme->pme_order)*(pme->nkz+pme->pme_order-1));
 853
 854     /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
 855      * We do this with an offset buffer of equal size, so we need to allocate
 856      * extra for the offset. That's what the (+1)*pme->nkz is for.
 857      */
 858     init_overlap_comm(&pme->overlap[1], pme->pme_order,
 859 #if GMX_MPI
 860                       pme->mpi_comm_d[1],
 861 #endif
 862                       pme->nnodes_minor, pme->nodeid_minor,
 863                       pme->nky,
 864                       (div_round_up(pme->nkx, pme->nnodes_major)+pme->pme_order+1)*pme->nkz);
 865
 866     /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
 867      * Note that gmx_pme_check_restrictions checked for this already.
 868      */
 869     if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
 870     {
 871         gmx_incons("More than one communication pulse required for grid overlap communication along the major dimension while using threads");
 872     }
 873
 874     snew(pme->bsp_mod[XX], pme->nkx);
 875     snew(pme->bsp_mod[YY], pme->nky);
 876     snew(pme->bsp_mod[ZZ], pme->nkz);
 877
 878     pme->gpu     = pmeGpu; /* Carrying over the single GPU structure */
 879     pme->runMode = runMode;
 880
 881     /* The required size of the interpolation grid, including overlap.
 882      * The allocated size (pmegrid_n?) might be slightly larger.
 883      */
 884     pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] -
 885         pme->overlap[0].s2g0[pme->nodeid_major];
 886     pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] -
 887         pme->overlap[1].s2g0[pme->nodeid_minor];
 888     pme->pmegrid_nz_base = pme->nkz;
 889     pme->pmegrid_nz      = pme->pmegrid_nz_base + pme->pme_order - 1;
 890     set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
 891     pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
 892     pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
 893     pme->pmegrid_start_iz = 0;
 894
 895     make_gridindex_to_localindex(pme->nkx,
 896                                  pme->pmegrid_start_ix,
 897                                  pme->pmegrid_nx - (pme->pme_order-1),
 898                                  &pme->nnx, &pme->fshx);
 899     make_gridindex_to_localindex(pme->nky,
 900                                  pme->pmegrid_start_iy,
 901                                  pme->pmegrid_ny - (pme->pme_order-1),
 902                                  &pme->nny, &pme->fshy);
 903     make_gridindex_to_localindex(pme->nkz,
 904                                  pme->pmegrid_start_iz,
 905                                  pme->pmegrid_nz_base,
 906                                  &pme->nnz, &pme->fshz);
 907
 908     pme->spline_work = make_pme_spline_work(pme->pme_order);
 909
 910     ndata[0]    = pme->nkx;
 911     ndata[1]    = pme->nky;
 912     ndata[2]    = pme->nkz;
 913     /* It doesn't matter if we allocate too many grids here,
 914      * we only allocate and use the ones we need.
 915      */
 916     if (pme->doLJ)
 917     {
 918         pme->ngrids = ((ir->ljpme_combination_rule == eljpmeLB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
 919     }
 920     else
 921     {
 922         pme->ngrids = DO_Q;
 923     }
 924     snew(pme->fftgrid, pme->ngrids);
 925     snew(pme->cfftgrid, pme->ngrids);
 926     snew(pme->pfft_setup, pme->ngrids);
 927
 928     for (i = 0; i < pme->ngrids; ++i)
 929     {
 930         if ((i <  DO_Q && pme->doCoulomb && (i == 0 ||
 931                                              bFreeEnergy_q)) ||
 932             (i >= DO_Q && pme->doLJ && (i == 2 ||
 933                                         bFreeEnergy_lj ||
 934                                         ir->ljpme_combination_rule == eljpmeLB)))
 935         {
 936             pmegrids_init(&pme->pmegrid[i],
 937                           pme->pmegrid_nx, pme->pmegrid_ny, pme->pmegrid_nz,
 938                           pme->pmegrid_nz_base,
 939                           pme->pme_order,
 940                           pme->bUseThreads,
 941                           pme->nthread,
 942                           pme->overlap[0].s2g1[pme->nodeid_major]-pme->overlap[0].s2g0[pme->nodeid_major+1],
 943                           pme->overlap[1].s2g1[pme->nodeid_minor]-pme->overlap[1].s2g0[pme->nodeid_minor+1]);
 944             /* This routine will allocate the grid data to fit the FFTs */
 945             const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed) ? gmx::PinningPolicy::PinnedIfSupported : gmx::PinningPolicy::CannotBePinned;
 946             gmx_parallel_3dfft_init(&pme->pfft_setup[i], ndata,
 947                                     &pme->fftgrid[i], &pme->cfftgrid[i],
 948                                     pme->mpi_comm_d,
 949                                     bReproducible, pme->nthread, allocateRealGridForGpu);
 950
 951         }
 952     }
 953
 954     if (!pme->bP3M)
 955     {
 956         /* Use plain SPME B-spline interpolation */
 957         make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 958     }
 959     else
 960     {
 961         /* Use the P3M grid-optimized influence function */
 962         make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 963     }
 964
 965     /* Use atc[0] for spreading */
 966     init_atomcomm(pme.get(), &pme->atc[0], numPmeDomains.x > 1 ? 0 : 1, TRUE);
 967     if (pme->ndecompdim >= 2)
 968     {
 969         init_atomcomm(pme.get(), &pme->atc[1], 1, FALSE);
 970     }
 971
 972     if (pme->nnodes == 1)
 973     {
 974         pme->atc[0].n = homenr;
 975         pme_realloc_atomcomm_things(&pme->atc[0]);
 976     }
 977
 978     pme->lb_buf1       = nullptr;
 979     pme->lb_buf2       = nullptr;
 980     pme->lb_buf_nalloc = 0;
 981
 982     if (pme_gpu_active(pme.get()))
 983     {
 984         if (!pme->gpu)
 985         {
 986             // Initial check of validity of the data
 987             std::string errorString;
 988             bool        canRunOnGpu = pme_gpu_check_restrictions(pme.get(), &errorString);
 989             if (!canRunOnGpu)
 990             {
 991                 GMX_THROW(gmx::NotImplementedError(errorString));
 992             }
 993         }
 994
 995         pme_gpu_reinit(pme.get(), gpuInfo, pmeGpuProgram);
 996     }
 997
 998     pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
 999
1000     // no exception was thrown during the init, so we hand over the PME structure handle
1001     return pme.release();
1002 }
1003
1004 void gmx_pme_reinit(struct gmx_pme_t **pmedata,
1005                     const t_commrec   *cr,
1006                     struct gmx_pme_t * pme_src,
1007                     const t_inputrec * ir,
1008                     const ivec         grid_size,
1009                     real               ewaldcoeff_q,
1010                     real               ewaldcoeff_lj)
1011 {
1012     int        homenr;
1013
1014     // Create a copy of t_inputrec fields that are used in gmx_pme_init().
1015     // TODO: This would be better as just copying a sub-structure that contains
1016     // all the PME parameters and nothing else.
1017     t_inputrec irc;
1018     irc.ePBC                   = ir->ePBC;
1019     irc.coulombtype            = ir->coulombtype;
1020     irc.vdwtype                = ir->vdwtype;
1021     irc.efep                   = ir->efep;
1022     irc.pme_order              = ir->pme_order;
1023     irc.epsilon_r              = ir->epsilon_r;
1024     irc.ljpme_combination_rule = ir->ljpme_combination_rule;
1025     irc.nkx                    = grid_size[XX];
1026     irc.nky                    = grid_size[YY];
1027     irc.nkz                    = grid_size[ZZ];
1028
1029     if (pme_src->nnodes == 1)
1030     {
1031         homenr = pme_src->atc[0].n;
1032     }
1033     else
1034     {
1035         homenr = -1;
1036     }
1037
1038     try
1039     {
1040         const gmx::MDLogger dummyLogger;
1041         // This is reinit which is currently only changing grid size/coefficients,
1042         // so we don't expect the actual logging.
1043         // TODO: when PME is an object, it should take reference to mdlog on construction and save it.
1044         GMX_ASSERT(pmedata, "Invalid PME pointer");
1045         NumPmeDomains numPmeDomains = { pme_src->nnodes_major, pme_src->nnodes_minor };
1046         *pmedata = gmx_pme_init(cr, numPmeDomains,
1047                                 &irc, homenr, pme_src->bFEP_q, pme_src->bFEP_lj, FALSE, ewaldcoeff_q, ewaldcoeff_lj,
1048                                 pme_src->nthread, pme_src->runMode, pme_src->gpu, nullptr, nullptr, dummyLogger);
1049         //TODO this is mostly passing around current values
1050     }
1051     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1052
1053     /* We can easily reuse the allocated pme grids in pme_src */
1054     reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
1055     /* We would like to reuse the fft grids, but that's harder */
1056 }
1057
1058 void gmx_pme_calc_energy(struct gmx_pme_t *pme, int n, rvec *x, real *q, real *V)
1059 {
1060     pme_atomcomm_t *atc;
1061     pmegrids_t     *grid;
1062
1063     if (pme->nnodes > 1)
1064     {
1065         gmx_incons("gmx_pme_calc_energy called in parallel");
1066     }
1067     if (pme->bFEP_q)
1068     {
1069         gmx_incons("gmx_pme_calc_energy with free energy");
1070     }
1071
1072     atc            = &pme->atc_energy;
1073     atc->nthread   = 1;
1074     if (atc->spline == nullptr)
1075     {
1076         snew(atc->spline, atc->nthread);
1077     }
1078     atc->nslab     = 1;
1079     atc->bSpread   = TRUE;
1080     atc->pme_order = pme->pme_order;
1081     atc->n         = n;
1082     pme_realloc_atomcomm_things(atc);
1083     atc->x           = x;
1084     atc->coefficient = q;
1085
1086     /* We only use the A-charges grid */
1087     grid = &pme->pmegrid[PME_GRID_QA];
1088
1089     /* Only calculate the spline coefficients, don't actually spread */
1090     spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
1091
1092     *V = gather_energy_bsplines(pme, grid->grid.grid, atc);
1093 }
1094
1095 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
1096 static void
1097 calc_initial_lb_coeffs(struct gmx_pme_t *pme, const real *local_c6, const real *local_sigma)
1098 {
1099     int  i;
1100     for (i = 0; i < pme->atc[0].n; ++i)
1101     {
1102         real sigma4;
1103         sigma4                     = local_sigma[i];
1104         sigma4                     = sigma4*sigma4;
1105         sigma4                     = sigma4*sigma4;
1106         pme->atc[0].coefficient[i] = local_c6[i] / sigma4;
1107     }
1108 }
1109
1110 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
1111 static void
1112 calc_next_lb_coeffs(struct gmx_pme_t *pme, const real *local_sigma)
1113 {
1114     int  i;
1115
1116     for (i = 0; i < pme->atc[0].n; ++i)
1117     {
1118         pme->atc[0].coefficient[i] *= local_sigma[i];
1119     }
1120 }
1121
1122 int gmx_pme_do(struct gmx_pme_t *pme,
1123                int start,       int homenr,
1124                rvec x[],        rvec f[],
1125                real chargeA[],  real chargeB[],
1126                real c6A[],      real c6B[],
1127                real sigmaA[],   real sigmaB[],
1128                matrix box,      const t_commrec *cr,
1129                int  maxshift_x, int maxshift_y,
1130                t_nrnb *nrnb,    gmx_wallcycle *wcycle,
1131                matrix vir_q,    matrix vir_lj,
1132                real *energy_q,  real *energy_lj,
1133                real lambda_q,   real lambda_lj,
1134                real *dvdlambda_q, real *dvdlambda_lj,
1135                int flags)
1136 {
1137     GMX_ASSERT(pme->runMode == PmeRunMode::CPU, "gmx_pme_do should not be called on the GPU PME run.");
1138
1139     int                  d, i, j, npme, grid_index, max_grid_index;
1140     int                  n_d;
1141     pme_atomcomm_t      *atc        = nullptr;
1142     pmegrids_t          *pmegrid    = nullptr;
1143     real                *grid       = nullptr;
1144     rvec                *f_d;
1145     real                *coefficient = nullptr;
1146     PmeOutput            output[2]; // The second is used for the B state with FEP
1147     real                 scale, lambda;
1148     gmx_bool             bClearF;
1149     gmx_parallel_3dfft_t pfft_setup;
1150     real              *  fftgrid;
1151     t_complex          * cfftgrid;
1152     int                  thread;
1153     gmx_bool             bFirst, bDoSplines;
1154     int                  fep_state;
1155     int                  fep_states_lj           = pme->bFEP_lj ? 2 : 1;
1156     const gmx_bool       bCalcEnerVir            = (flags & GMX_PME_CALC_ENER_VIR) != 0;
1157     const gmx_bool       bBackFFT                = (flags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
1158     const gmx_bool       bCalcF                  = (flags & GMX_PME_CALC_F) != 0;
1159
1160     /* We could be passing lambda!=0 while no q or LJ is actually perturbed */
1161     if (!pme->bFEP_q)
1162     {
1163         lambda_q = 0;
1164     }
1165     if (!pme->bFEP_lj)
1166     {
1167         lambda_lj = 0;
1168     }
1169
1170     assert(pme->nnodes > 0);
1171     assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1172
1173     if (pme->nnodes > 1)
1174     {
1175         atc      = &pme->atc[0];
1176         atc->npd = homenr;
1177         if (atc->npd > atc->pd_nalloc)
1178         {
1179             atc->pd_nalloc = over_alloc_dd(atc->npd);
1180             srenew(atc->pd, atc->pd_nalloc);
1181         }
1182         for (d = pme->ndecompdim-1; d >= 0; d--)
1183         {
1184             atc           = &pme->atc[d];
1185             atc->maxshift = (atc->dimind == 0 ? maxshift_x : maxshift_y);
1186         }
1187     }
1188     else
1189     {
1190         atc = &pme->atc[0];
1191         /* This could be necessary for TPI */
1192         pme->atc[0].n = homenr;
1193         if (DOMAINDECOMP(cr))
1194         {
1195             pme_realloc_atomcomm_things(atc);
1196         }
1197         atc->x = x;
1198         atc->f = f;
1199     }
1200
1201     matrix scaledBox;
1202     pme->boxScaler->scaleBox(box, scaledBox);
1203
1204     gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1205     bFirst = TRUE;
1206
1207     /* For simplicity, we construct the splines for all particles if
1208      * more than one PME calculations is needed. Some optimization
1209      * could be done by keeping track of which atoms have splines
1210      * constructed, and construct new splines on each pass for atoms
1211      * that don't yet have them.
1212      */
1213
1214     bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1215
1216     /* We need a maximum of four separate PME calculations:
1217      * grid_index=0: Coulomb PME with charges from state A
1218      * grid_index=1: Coulomb PME with charges from state B
1219      * grid_index=2: LJ PME with C6 from state A
1220      * grid_index=3: LJ PME with C6 from state B
1221      * For Lorentz-Berthelot combination rules, a separate loop is used to
1222      * calculate all the terms
1223      */
1224
1225     /* If we are doing LJ-PME with LB, we only do Q here */
1226     max_grid_index = (pme->ljpme_combination_rule == eljpmeLB) ? DO_Q : DO_Q_AND_LJ;
1227
1228     for (grid_index = 0; grid_index < max_grid_index; ++grid_index)
1229     {
1230         /* Check if we should do calculations at this grid_index
1231          * If grid_index is odd we should be doing FEP
1232          * If grid_index < 2 we should be doing electrostatic PME
1233          * If grid_index >= 2 we should be doing LJ-PME
1234          */
1235         if ((grid_index <  DO_Q && (!pme->doCoulomb ||
1236                                     (grid_index == 1 && !pme->bFEP_q))) ||
1237             (grid_index >= DO_Q && (!pme->doLJ ||
1238                                     (grid_index == 3 && !pme->bFEP_lj))))
1239         {
1240             continue;
1241         }
1242         /* Unpack structure */
1243         pmegrid    = &pme->pmegrid[grid_index];
1244         fftgrid    = pme->fftgrid[grid_index];
1245         cfftgrid   = pme->cfftgrid[grid_index];
1246         pfft_setup = pme->pfft_setup[grid_index];
1247         switch (grid_index)
1248         {
1249             case 0: coefficient = chargeA + start; break;
1250             case 1: coefficient = chargeB + start; break;
1251             case 2: coefficient = c6A + start; break;
1252             case 3: coefficient = c6B + start; break;
1253         }
1254
1255         grid = pmegrid->grid.grid;
1256
1257         if (debug)
1258         {
1259             fprintf(debug, "PME: number of ranks = %d, rank = %d\n",
1260                     cr->nnodes, cr->nodeid);
1261             fprintf(debug, "Grid = %p\n", static_cast<void*>(grid));
1262             if (grid == nullptr)
1263             {
1264                 gmx_fatal(FARGS, "No grid!");
1265             }
1266         }
1267
1268         if (pme->nnodes == 1)
1269         {
1270             atc->coefficient = coefficient;
1271         }
1272         else
1273         {
1274             wallcycle_start(wcycle, ewcPME_REDISTXF);
1275             do_redist_pos_coeffs(pme, cr, start, homenr, bFirst, x, coefficient);
1276
1277             wallcycle_stop(wcycle, ewcPME_REDISTXF);
1278         }
1279
1280         if (debug)
1281         {
1282             fprintf(debug, "Rank= %6d, pme local particles=%6d\n",
1283                     cr->nodeid, atc->n);
1284         }
1285
1286         if (flags & GMX_PME_SPREAD)
1287         {
1288             wallcycle_start(wcycle, ewcPME_SPREAD);
1289
1290             /* Spread the coefficients on a grid */
1291             spread_on_grid(pme, &pme->atc[0], pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1292
1293             if (bFirst)
1294             {
1295                 inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
1296             }
1297             inc_nrnb(nrnb, eNR_SPREADBSP,
1298                      pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
1299
1300             if (!pme->bUseThreads)
1301             {
1302                 wrap_periodic_pmegrid(pme, grid);
1303
1304                 /* sum contributions to local grid from other nodes */
1305 #if GMX_MPI
1306                 if (pme->nnodes > 1)
1307                 {
1308                     gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1309                 }
1310 #endif
1311
1312                 copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1313             }
1314
1315             wallcycle_stop(wcycle, ewcPME_SPREAD);
1316
1317             /* TODO If the OpenMP and single-threaded implementations
1318                converge, then spread_on_grid() and
1319                copy_pmegrid_to_fftgrid() will perhaps live in the same
1320                source file.
1321              */
1322         }
1323
1324         /* Here we start a large thread parallel region */
1325 #pragma omp parallel num_threads(pme->nthread) private(thread)
1326         {
1327             try
1328             {
1329                 thread = gmx_omp_get_thread_num();
1330                 if (flags & GMX_PME_SOLVE)
1331                 {
1332                     int loop_count;
1333
1334                     /* do 3d-fft */
1335                     if (thread == 0)
1336                     {
1337                         wallcycle_start(wcycle, ewcPME_FFT);
1338                     }
1339                     gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
1340                                                thread, wcycle);
1341                     if (thread == 0)
1342                     {
1343                         wallcycle_stop(wcycle, ewcPME_FFT);
1344                     }
1345
1346                     /* solve in k-space for our local cells */
1347                     if (thread == 0)
1348                     {
1349                         wallcycle_start(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1350                     }
1351                     if (grid_index < DO_Q)
1352                     {
1353                         loop_count =
1354                             solve_pme_yzx(pme, cfftgrid,
1355                                           scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1356                                           bCalcEnerVir,
1357                                           pme->nthread, thread);
1358                     }
1359                     else
1360                     {
1361                         loop_count =
1362                             solve_pme_lj_yzx(pme, &cfftgrid, FALSE,
1363                                              scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1364                                              bCalcEnerVir,
1365                                              pme->nthread, thread);
1366                     }
1367
1368                     if (thread == 0)
1369                     {
1370                         wallcycle_stop(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1371                         inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1372                     }
1373                 }
1374
1375                 if (bBackFFT)
1376                 {
1377                     /* do 3d-invfft */
1378                     if (thread == 0)
1379                     {
1380                         wallcycle_start(wcycle, ewcPME_FFT);
1381                     }
1382                     gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
1383                                                thread, wcycle);
1384                     if (thread == 0)
1385                     {
1386                         wallcycle_stop(wcycle, ewcPME_FFT);
1387
1388
1389                         if (pme->nodeid == 0)
1390                         {
1391                             real ntot = pme->nkx*pme->nky*pme->nkz;
1392                             npme  = static_cast<int>(ntot*std::log(ntot)/std::log(2.0));
1393                             inc_nrnb(nrnb, eNR_FFT, 2*npme);
1394                         }
1395
1396                         /* Note: this wallcycle region is closed below
1397                            outside an OpenMP region, so take care if
1398                            refactoring code here. */
1399                         wallcycle_start(wcycle, ewcPME_GATHER);
1400                     }
1401
1402                     copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1403                 }
1404             } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1405         }
1406         /* End of thread parallel section.
1407          * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1408          */
1409
1410         if (bBackFFT)
1411         {
1412             /* distribute local grid to all nodes */
1413 #if GMX_MPI
1414             if (pme->nnodes > 1)
1415             {
1416                 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1417             }
1418 #endif
1419
1420             unwrap_periodic_pmegrid(pme, grid);
1421         }
1422
1423         if (bCalcF)
1424         {
1425             /* interpolate forces for our local atoms */
1426
1427
1428             /* If we are running without parallelization,
1429              * atc->f is the actual force array, not a buffer,
1430              * therefore we should not clear it.
1431              */
1432             lambda  = grid_index < DO_Q ? lambda_q : lambda_lj;
1433             bClearF = (bFirst && PAR(cr));
1434 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1435             for (thread = 0; thread < pme->nthread; thread++)
1436             {
1437                 try
1438                 {
1439                     gather_f_bsplines(pme, grid, bClearF, atc,
1440                                       &atc->spline[thread],
1441                                       pme->bFEP ? (grid_index % 2 == 0 ? 1.0-lambda : lambda) : 1.0);
1442                 }
1443                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1444             }
1445
1446
1447             inc_nrnb(nrnb, eNR_GATHERFBSP,
1448                      pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
1449             /* Note: this wallcycle region is opened above inside an OpenMP
1450                region, so take care if refactoring code here. */
1451             wallcycle_stop(wcycle, ewcPME_GATHER);
1452         }
1453
1454         if (bCalcEnerVir)
1455         {
1456             /* This should only be called on the master thread
1457              * and after the threads have synchronized.
1458              */
1459             if (grid_index < 2)
1460             {
1461                 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1462             }
1463             else
1464             {
1465                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1466             }
1467         }
1468         bFirst = FALSE;
1469     } /* of grid_index-loop */
1470
1471     /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1472      * seven terms. */
1473
1474     if (pme->doLJ && pme->ljpme_combination_rule == eljpmeLB)
1475     {
1476         /* Loop over A- and B-state if we are doing FEP */
1477         for (fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1478         {
1479             real *local_c6 = nullptr, *local_sigma = nullptr, *RedistC6 = nullptr, *RedistSigma = nullptr;
1480             if (pme->nnodes == 1)
1481             {
1482                 if (pme->lb_buf1 == nullptr)
1483                 {
1484                     pme->lb_buf_nalloc = pme->atc[0].n;
1485                     snew(pme->lb_buf1, pme->lb_buf_nalloc);
1486                 }
1487                 pme->atc[0].coefficient = pme->lb_buf1;
1488                 switch (fep_state)
1489                 {
1490                     case 0:
1491                         local_c6      = c6A;
1492                         local_sigma   = sigmaA;
1493                         break;
1494                     case 1:
1495                         local_c6      = c6B;
1496                         local_sigma   = sigmaB;
1497                         break;
1498                     default:
1499                         gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1500                 }
1501             }
1502             else
1503             {
1504                 atc = &pme->atc[0];
1505                 switch (fep_state)
1506                 {
1507                     case 0:
1508                         RedistC6      = c6A;
1509                         RedistSigma   = sigmaA;
1510                         break;
1511                     case 1:
1512                         RedistC6      = c6B;
1513                         RedistSigma   = sigmaB;
1514                         break;
1515                     default:
1516                         gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1517                 }
1518                 wallcycle_start(wcycle, ewcPME_REDISTXF);
1519
1520                 do_redist_pos_coeffs(pme, cr, start, homenr, bFirst, x, RedistC6);
1521                 if (pme->lb_buf_nalloc < atc->n)
1522                 {
1523                     pme->lb_buf_nalloc = atc->nalloc;
1524                     srenew(pme->lb_buf1, pme->lb_buf_nalloc);
1525                     srenew(pme->lb_buf2, pme->lb_buf_nalloc);
1526                 }
1527                 local_c6 = pme->lb_buf1;
1528                 for (i = 0; i < atc->n; ++i)
1529                 {
1530                     local_c6[i] = atc->coefficient[i];
1531                 }
1532
1533                 do_redist_pos_coeffs(pme, cr, start, homenr, FALSE, x, RedistSigma);
1534                 local_sigma = pme->lb_buf2;
1535                 for (i = 0; i < atc->n; ++i)
1536                 {
1537                     local_sigma[i] = atc->coefficient[i];
1538                 }
1539
1540                 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1541             }
1542             calc_initial_lb_coeffs(pme, local_c6, local_sigma);
1543
1544             /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1545             for (grid_index = 2; grid_index < 9; ++grid_index)
1546             {
1547                 /* Unpack structure */
1548                 pmegrid    = &pme->pmegrid[grid_index];
1549                 fftgrid    = pme->fftgrid[grid_index];
1550                 pfft_setup = pme->pfft_setup[grid_index];
1551                 calc_next_lb_coeffs(pme, local_sigma);
1552                 grid = pmegrid->grid.grid;
1553
1554                 if (flags & GMX_PME_SPREAD)
1555                 {
1556                     wallcycle_start(wcycle, ewcPME_SPREAD);
1557                     /* Spread the c6 on a grid */
1558                     spread_on_grid(pme, &pme->atc[0], pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1559
1560                     if (bFirst)
1561                     {
1562                         inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
1563                     }
1564
1565                     inc_nrnb(nrnb, eNR_SPREADBSP,
1566                              pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
1567                     if (pme->nthread == 1)
1568                     {
1569                         wrap_periodic_pmegrid(pme, grid);
1570                         /* sum contributions to local grid from other nodes */
1571 #if GMX_MPI
1572                         if (pme->nnodes > 1)
1573                         {
1574                             gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1575                         }
1576 #endif
1577                         copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1578                     }
1579                     wallcycle_stop(wcycle, ewcPME_SPREAD);
1580                 }
1581                 /*Here we start a large thread parallel region*/
1582 #pragma omp parallel num_threads(pme->nthread) private(thread)
1583                 {
1584                     try
1585                     {
1586                         thread = gmx_omp_get_thread_num();
1587                         if (flags & GMX_PME_SOLVE)
1588                         {
1589                             /* do 3d-fft */
1590                             if (thread == 0)
1591                             {
1592                                 wallcycle_start(wcycle, ewcPME_FFT);
1593                             }
1594
1595                             gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
1596                                                        thread, wcycle);
1597                             if (thread == 0)
1598                             {
1599                                 wallcycle_stop(wcycle, ewcPME_FFT);
1600                             }
1601                         }
1602                     }
1603                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1604                 }
1605                 bFirst = FALSE;
1606             }
1607             if (flags & GMX_PME_SOLVE)
1608             {
1609                 /* solve in k-space for our local cells */
1610 #pragma omp parallel num_threads(pme->nthread) private(thread)
1611                 {
1612                     try
1613                     {
1614                         int loop_count;
1615                         thread = gmx_omp_get_thread_num();
1616                         if (thread == 0)
1617                         {
1618                             wallcycle_start(wcycle, ewcLJPME);
1619                         }
1620
1621                         loop_count =
1622                             solve_pme_lj_yzx(pme, &pme->cfftgrid[2], TRUE,
1623                                              scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1624                                              bCalcEnerVir,
1625                                              pme->nthread, thread);
1626                         if (thread == 0)
1627                         {
1628                             wallcycle_stop(wcycle, ewcLJPME);
1629                             inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1630                         }
1631                     }
1632                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1633                 }
1634             }
1635
1636             if (bCalcEnerVir)
1637             {
1638                 /* This should only be called on the master thread and
1639                  * after the threads have synchronized.
1640                  */
1641                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[fep_state]);
1642             }
1643
1644             if (bBackFFT)
1645             {
1646                 bFirst = !pme->doCoulomb;
1647                 calc_initial_lb_coeffs(pme, local_c6, local_sigma);
1648                 for (grid_index = 8; grid_index >= 2; --grid_index)
1649                 {
1650                     /* Unpack structure */
1651                     pmegrid    = &pme->pmegrid[grid_index];
1652                     fftgrid    = pme->fftgrid[grid_index];
1653                     pfft_setup = pme->pfft_setup[grid_index];
1654                     grid       = pmegrid->grid.grid;
1655                     calc_next_lb_coeffs(pme, local_sigma);
1656 #pragma omp parallel num_threads(pme->nthread) private(thread)
1657                     {
1658                         try
1659                         {
1660                             thread = gmx_omp_get_thread_num();
1661                             /* do 3d-invfft */
1662                             if (thread == 0)
1663                             {
1664                                 wallcycle_start(wcycle, ewcPME_FFT);
1665                             }
1666
1667                             gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
1668                                                        thread, wcycle);
1669                             if (thread == 0)
1670                             {
1671                                 wallcycle_stop(wcycle, ewcPME_FFT);
1672
1673
1674                                 if (pme->nodeid == 0)
1675                                 {
1676                                     real ntot = pme->nkx*pme->nky*pme->nkz;
1677                                     npme  = static_cast<int>(ntot*std::log(ntot)/std::log(2.0));
1678                                     inc_nrnb(nrnb, eNR_FFT, 2*npme);
1679                                 }
1680                                 wallcycle_start(wcycle, ewcPME_GATHER);
1681                             }
1682
1683                             copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1684                         }
1685                         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1686                     } /*#pragma omp parallel*/
1687
1688                     /* distribute local grid to all nodes */
1689 #if GMX_MPI
1690                     if (pme->nnodes > 1)
1691                     {
1692                         gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1693                     }
1694 #endif
1695
1696                     unwrap_periodic_pmegrid(pme, grid);
1697
1698                     if (bCalcF)
1699                     {
1700                         /* interpolate forces for our local atoms */
1701                         bClearF = (bFirst && PAR(cr));
1702                         scale   = pme->bFEP ? (fep_state < 1 ? 1.0-lambda_lj : lambda_lj) : 1.0;
1703                         scale  *= lb_scale_factor[grid_index-2];
1704
1705 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1706                         for (thread = 0; thread < pme->nthread; thread++)
1707                         {
1708                             try
1709                             {
1710                                 gather_f_bsplines(pme, grid, bClearF, &pme->atc[0],
1711                                                   &pme->atc[0].spline[thread],
1712                                                   scale);
1713                             }
1714                             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1715                         }
1716
1717
1718                         inc_nrnb(nrnb, eNR_GATHERFBSP,
1719                                  pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
1720                     }
1721                     wallcycle_stop(wcycle, ewcPME_GATHER);
1722
1723                     bFirst = FALSE;
1724                 } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1725             }     /* if (bCalcF) */
1726         }         /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1727     }             /* if ((flags & GMX_PME_DO_LJ) && pme->ljpme_combination_rule == eljpmeLB) */
1728
1729     if (bCalcF && pme->nnodes > 1)
1730     {
1731         wallcycle_start(wcycle, ewcPME_REDISTXF);
1732         for (d = 0; d < pme->ndecompdim; d++)
1733         {
1734             atc = &pme->atc[d];
1735             if (d == pme->ndecompdim - 1)
1736             {
1737                 n_d = homenr;
1738                 f_d = f + start;
1739             }
1740             else
1741             {
1742                 n_d = pme->atc[d+1].n;
1743                 f_d = pme->atc[d+1].f;
1744             }
1745             if (DOMAINDECOMP(cr))
1746             {
1747                 dd_pmeredist_f(pme, atc, n_d, f_d,
1748                                d == pme->ndecompdim-1 && pme->bPPnode);
1749             }
1750         }
1751
1752         wallcycle_stop(wcycle, ewcPME_REDISTXF);
1753     }
1754
1755     if (bCalcEnerVir)
1756     {
1757         if (pme->doCoulomb)
1758         {
1759             if (!pme->bFEP_q)
1760             {
1761                 *energy_q = output[0].coulombEnergy_;
1762                 m_add(vir_q, output[0].coulombVirial_, vir_q);
1763             }
1764             else
1765             {
1766                 *energy_q       = (1.0-lambda_q)*output[0].coulombEnergy_ + lambda_q*output[1].coulombEnergy_;
1767                 *dvdlambda_q   += output[1].coulombEnergy_ - output[0].coulombEnergy_;
1768                 for (i = 0; i < DIM; i++)
1769                 {
1770                     for (j = 0; j < DIM; j++)
1771                     {
1772                         vir_q[i][j] += (1.0-lambda_q)*output[0].coulombVirial_[i][j] +
1773                             lambda_q*output[1].coulombVirial_[i][j];
1774                     }
1775                 }
1776             }
1777             if (debug)
1778             {
1779                 fprintf(debug, "Electrostatic PME mesh energy: %g\n", *energy_q);
1780             }
1781         }
1782         else
1783         {
1784             *energy_q = 0;
1785         }
1786
1787         if (pme->doLJ)
1788         {
1789             if (!pme->bFEP_lj)
1790             {
1791                 *energy_lj = output[0].lennardJonesEnergy_;
1792                 m_add(vir_lj, output[0].lennardJonesVirial_, vir_lj);
1793             }
1794             else
1795             {
1796                 *energy_lj     = (1.0-lambda_lj)*output[0].lennardJonesEnergy_ + lambda_lj*output[1].lennardJonesEnergy_;
1797                 *dvdlambda_lj += output[1].lennardJonesEnergy_ - output[0].lennardJonesEnergy_;
1798                 for (i = 0; i < DIM; i++)
1799                 {
1800                     for (j = 0; j < DIM; j++)
1801                     {
1802                         vir_lj[i][j] += (1.0-lambda_lj)*output[0].lennardJonesVirial_[i][j] + lambda_lj*output[1].lennardJonesVirial_[i][j];
1803                     }
1804                 }
1805             }
1806             if (debug)
1807             {
1808                 fprintf(debug, "Lennard-Jones PME mesh energy: %g\n", *energy_lj);
1809             }
1810         }
1811         else
1812         {
1813             *energy_lj = 0;
1814         }
1815     }
1816     return 0;
1817 }
1818
1819 void gmx_pme_destroy(gmx_pme_t *pme)
1820 {
1821     if (!pme)
1822     {
1823         return;
1824     }
1825
1826     delete pme->boxScaler;
1827
1828     sfree(pme->nnx);
1829     sfree(pme->nny);
1830     sfree(pme->nnz);
1831     sfree(pme->fshx);
1832     sfree(pme->fshy);
1833     sfree(pme->fshz);
1834
1835     for (int i = 0; i < pme->ngrids; ++i)
1836     {
1837         pmegrids_destroy(&pme->pmegrid[i]);
1838     }
1839     if (pme->pfft_setup)
1840     {
1841         for (int i = 0; i < pme->ngrids; ++i)
1842         {
1843             gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1844         }
1845     }
1846     sfree(pme->fftgrid);
1847     sfree(pme->cfftgrid);
1848     sfree(pme->pfft_setup);
1849
1850     for (int i = 0; i < std::max(1, pme->ndecompdim); i++) //pme->atc[0] is always allocated
1851     {
1852         destroy_atomcomm(&pme->atc[i]);
1853     }
1854
1855     for (int i = 0; i < DIM; i++)
1856     {
1857         sfree(pme->bsp_mod[i]);
1858     }
1859
1860     sfree(pme->lb_buf1);
1861     sfree(pme->lb_buf2);
1862
1863     sfree(pme->bufv);
1864     sfree(pme->bufr);
1865
1866     if (pme->solve_work)
1867     {
1868         pme_free_all_work(&pme->solve_work, pme->nthread);
1869     }
1870
1871     sfree(pme->sum_qgrid_tmp);
1872     sfree(pme->sum_qgrid_dd_tmp);
1873
1874     destroy_pme_spline_work(pme->spline_work);
1875
1876     if (pme_gpu_active(pme) && pme->gpu)
1877     {
1878         pme_gpu_destroy(pme->gpu);
1879     }
1880
1881     delete pme;
1882 }
1883
1884 void gmx_pme_reinit_atoms(const gmx_pme_t *pme, const int nAtoms, const real *charges)
1885 {
1886     if (pme_gpu_active(pme))
1887     {
1888         pme_gpu_reinit_atoms(pme->gpu, nAtoms, charges);
1889     }
1890     // TODO: handle the CPU case here; handle the whole t_mdatoms
1891 }