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