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