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