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 gmx_device_info_t* gpuInfo,
 574                         PmeGpuProgramHandle      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->ePBC == epbcSCREW)
 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             GMX_LOG(mdlog.warning)
 745                     .asParagraph()
 746                     .appendTextFormatted(
 747                             "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
 748                             "      For optimal PME load balancing\n"
 749                             "      PME grid_x (%d) and grid_y (%d) should be divisible by "
 750                             "#PME_ranks_x "
 751                             "(%d)\n"
 752                             "      and PME grid_y (%d) and grid_z (%d) should be divisible by "
 753                             "#PME_ranks_y "
 754                             "(%d)",
 755                             gmx::roundToInt((imbal - 1) * 100), pme->nkx, pme->nky,
 756                             pme->nnodes_major, pme->nky, pme->nkz, pme->nnodes_minor);
 757         }
 758     }
 759
 760     /* For non-divisible grid we need pme_order iso pme_order-1 */
 761     /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
 762      * y is always copied through a buffer: we don't need padding in z,
 763      * but we do need the overlap in x because of the communication order.
 764      */
 765     init_overlap_comm(&pme->overlap[0], pme->pme_order, pme->mpi_comm_d[0], pme->nnodes_major,
 766                       pme->nodeid_major, pme->nkx,
 767                       (div_round_up(pme->nky, pme->nnodes_minor) + pme->pme_order)
 768                               * (pme->nkz + pme->pme_order - 1));
 769
 770     /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
 771      * We do this with an offset buffer of equal size, so we need to allocate
 772      * extra for the offset. That's what the (+1)*pme->nkz is for.
 773      */
 774     init_overlap_comm(&pme->overlap[1], pme->pme_order, pme->mpi_comm_d[1], pme->nnodes_minor,
 775                       pme->nodeid_minor, pme->nky,
 776                       (div_round_up(pme->nkx, pme->nnodes_major) + pme->pme_order + 1) * pme->nkz);
 777
 778     /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
 779      * Note that gmx_pme_check_restrictions checked for this already.
 780      */
 781     if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
 782     {
 783         gmx_incons(
 784                 "More than one communication pulse required for grid overlap communication along "
 785                 "the major dimension while using threads");
 786     }
 787
 788     snew(pme->bsp_mod[XX], pme->nkx);
 789     snew(pme->bsp_mod[YY], pme->nky);
 790     snew(pme->bsp_mod[ZZ], pme->nkz);
 791
 792     pme->gpu     = pmeGpu; /* Carrying over the single GPU structure */
 793     pme->runMode = runMode;
 794
 795     /* The required size of the interpolation grid, including overlap.
 796      * The allocated size (pmegrid_n?) might be slightly larger.
 797      */
 798     pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] - pme->overlap[0].s2g0[pme->nodeid_major];
 799     pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] - pme->overlap[1].s2g0[pme->nodeid_minor];
 800     pme->pmegrid_nz_base = pme->nkz;
 801     pme->pmegrid_nz      = pme->pmegrid_nz_base + pme->pme_order - 1;
 802     set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
 803     pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
 804     pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
 805     pme->pmegrid_start_iz = 0;
 806
 807     make_gridindex_to_localindex(pme->nkx, pme->pmegrid_start_ix,
 808                                  pme->pmegrid_nx - (pme->pme_order - 1), &pme->nnx, &pme->fshx);
 809     make_gridindex_to_localindex(pme->nky, pme->pmegrid_start_iy,
 810                                  pme->pmegrid_ny - (pme->pme_order - 1), &pme->nny, &pme->fshy);
 811     make_gridindex_to_localindex(pme->nkz, pme->pmegrid_start_iz, pme->pmegrid_nz_base, &pme->nnz,
 812                                  &pme->fshz);
 813
 814     pme->spline_work = make_pme_spline_work(pme->pme_order);
 815
 816     ndata[0] = pme->nkx;
 817     ndata[1] = pme->nky;
 818     ndata[2] = pme->nkz;
 819     /* It doesn't matter if we allocate too many grids here,
 820      * we only allocate and use the ones we need.
 821      */
 822     if (pme->doLJ)
 823     {
 824         pme->ngrids = ((ir->ljpme_combination_rule == eljpmeLB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
 825     }
 826     else
 827     {
 828         pme->ngrids = DO_Q;
 829     }
 830     snew(pme->fftgrid, pme->ngrids);
 831     snew(pme->cfftgrid, pme->ngrids);
 832     snew(pme->pfft_setup, pme->ngrids);
 833
 834     for (i = 0; i < pme->ngrids; ++i)
 835     {
 836         if ((i < DO_Q && pme->doCoulomb && (i == 0 || bFreeEnergy_q))
 837             || (i >= DO_Q && pme->doLJ
 838                 && (i == 2 || bFreeEnergy_lj || ir->ljpme_combination_rule == eljpmeLB)))
 839         {
 840             pmegrids_init(&pme->pmegrid[i], pme->pmegrid_nx, pme->pmegrid_ny, pme->pmegrid_nz,
 841                           pme->pmegrid_nz_base, pme->pme_order, pme->bUseThreads, pme->nthread,
 842                           pme->overlap[0].s2g1[pme->nodeid_major]
 843                                   - pme->overlap[0].s2g0[pme->nodeid_major + 1],
 844                           pme->overlap[1].s2g1[pme->nodeid_minor]
 845                                   - pme->overlap[1].s2g0[pme->nodeid_minor + 1]);
 846             /* This routine will allocate the grid data to fit the FFTs */
 847             const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed)
 848                                                         ? gmx::PinningPolicy::PinnedIfSupported
 849                                                         : gmx::PinningPolicy::CannotBePinned;
 850             gmx_parallel_3dfft_init(&pme->pfft_setup[i], ndata, &pme->fftgrid[i], &pme->cfftgrid[i],
 851                                     pme->mpi_comm_d, bReproducible, pme->nthread, allocateRealGridForGpu);
 852         }
 853     }
 854
 855     if (!pme->bP3M)
 856     {
 857         /* Use plain SPME B-spline interpolation */
 858         make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 859     }
 860     else
 861     {
 862         /* Use the P3M grid-optimized influence function */
 863         make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
 864     }
 865
 866     /* Use atc[0] for spreading */
 867     const int firstDimIndex   = (numPmeDomains.x > 1 ? 0 : 1);
 868     MPI_Comm  mpiCommFirstDim = (pme->nnodes > 1 ? pme->mpi_comm_d[firstDimIndex] : MPI_COMM_NULL);
 869     bool      doSpread        = true;
 870     pme->atc.emplace_back(mpiCommFirstDim, pme->nthread, pme->pme_order, firstDimIndex, doSpread);
 871     if (pme->ndecompdim >= 2)
 872     {
 873         const int secondDimIndex = 1;
 874         doSpread                 = false;
 875         pme->atc.emplace_back(pme->mpi_comm_d[1], pme->nthread, pme->pme_order, secondDimIndex, doSpread);
 876     }
 877
 878     if (pme_gpu_active(pme.get()))
 879     {
 880         if (!pme->gpu)
 881         {
 882             // Initial check of validity of the data
 883             std::string errorString;
 884             bool        canRunOnGpu = pme_gpu_check_restrictions(pme.get(), &errorString);
 885             if (!canRunOnGpu)
 886             {
 887                 GMX_THROW(gmx::NotImplementedError(errorString));
 888             }
 889         }
 890
 891         pme_gpu_reinit(pme.get(), gpuInfo, pmeGpuProgram);
 892     }
 893
 894     pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
 895
 896     // no exception was thrown during the init, so we hand over the PME structure handle
 897     return pme.release();
 898 }
 899
 900 void gmx_pme_reinit(struct gmx_pme_t** pmedata,
 901                     const t_commrec*   cr,
 902                     struct gmx_pme_t*  pme_src,
 903                     const t_inputrec*  ir,
 904                     const ivec         grid_size,
 905                     real               ewaldcoeff_q,
 906                     real               ewaldcoeff_lj)
 907 {
 908     // Create a copy of t_inputrec fields that are used in gmx_pme_init().
 909     // TODO: This would be better as just copying a sub-structure that contains
 910     // all the PME parameters and nothing else.
 911     t_inputrec irc;
 912     irc.ePBC                   = ir->ePBC;
 913     irc.coulombtype            = ir->coulombtype;
 914     irc.vdwtype                = ir->vdwtype;
 915     irc.efep                   = ir->efep;
 916     irc.pme_order              = ir->pme_order;
 917     irc.epsilon_r              = ir->epsilon_r;
 918     irc.ljpme_combination_rule = ir->ljpme_combination_rule;
 919     irc.nkx                    = grid_size[XX];
 920     irc.nky                    = grid_size[YY];
 921     irc.nkz                    = grid_size[ZZ];
 922
 923     try
 924     {
 925         // This is reinit. Any logging should have been done at first init.
 926         // Here we should avoid writing notes for settings the user did not
 927         // set directly.
 928         const gmx::MDLogger dummyLogger;
 929         GMX_ASSERT(pmedata, "Invalid PME pointer");
 930         NumPmeDomains numPmeDomains = { pme_src->nnodes_major, pme_src->nnodes_minor };
 931         *pmedata = gmx_pme_init(cr, numPmeDomains, &irc, pme_src->bFEP_q, pme_src->bFEP_lj, FALSE,
 932                                 ewaldcoeff_q, ewaldcoeff_lj, pme_src->nthread, pme_src->runMode,
 933                                 pme_src->gpu, nullptr, nullptr, dummyLogger);
 934         /* When running PME on the CPU not using domain decomposition,
 935          * the atom data is allocated once only in gmx_pme_(re)init().
 936          */
 937         if (!pme_src->gpu && pme_src->nnodes == 1)
 938         {
 939             gmx_pme_reinit_atoms(*pmedata, pme_src->atc[0].numAtoms(), nullptr);
 940         }
 941         // TODO this is mostly passing around current values
 942     }
 943     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 944
 945     /* We can easily reuse the allocated pme grids in pme_src */
 946     reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
 947     /* We would like to reuse the fft grids, but that's harder */
 948 }
 949
 950 void gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q, real* V)
 951 {
 952     pmegrids_t* grid;
 953
 954     if (pme->nnodes > 1)
 955     {
 956         gmx_incons("gmx_pme_calc_energy called in parallel");
 957     }
 958     if (pme->bFEP_q)
 959     {
 960         gmx_incons("gmx_pme_calc_energy with free energy");
 961     }
 962
 963     if (!pme->atc_energy)
 964     {
 965         pme->atc_energy = std::make_unique<PmeAtomComm>(MPI_COMM_NULL, 1, pme->pme_order, 0, true);
 966     }
 967     PmeAtomComm* atc = pme->atc_energy.get();
 968     atc->setNumAtoms(x.ssize());
 969     atc->x           = x;
 970     atc->coefficient = q;
 971
 972     /* We only use the A-charges grid */
 973     grid = &pme->pmegrid[PME_GRID_QA];
 974
 975     /* Only calculate the spline coefficients, don't actually spread */
 976     spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
 977
 978     *V = gather_energy_bsplines(pme, grid->grid.grid, atc);
 979 }
 980
 981 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
 982 static void calc_initial_lb_coeffs(gmx::ArrayRef<real> coefficient, const real* local_c6, const real* local_sigma)
 983 {
 984     for (gmx::index i = 0; i < coefficient.ssize(); ++i)
 985     {
 986         real sigma4    = local_sigma[i];
 987         sigma4         = sigma4 * sigma4;
 988         sigma4         = sigma4 * sigma4;
 989         coefficient[i] = local_c6[i] / sigma4;
 990     }
 991 }
 992
 993 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
 994 static void calc_next_lb_coeffs(gmx::ArrayRef<real> coefficient, const real* local_sigma)
 995 {
 996     for (gmx::index i = 0; i < coefficient.ssize(); ++i)
 997     {
 998         coefficient[i] *= local_sigma[i];
 999     }
1000 }
1001
1002 int gmx_pme_do(struct gmx_pme_t*              pme,
1003                gmx::ArrayRef<const gmx::RVec> coordinates,
1004                gmx::ArrayRef<gmx::RVec>       forces,
1005                real                           chargeA[],
1006                real                           chargeB[],
1007                real                           c6A[],
1008                real                           c6B[],
1009                real                           sigmaA[],
1010                real                           sigmaB[],
1011                const matrix                   box,
1012                const t_commrec*               cr,
1013                int                            maxshift_x,
1014                int                            maxshift_y,
1015                t_nrnb*                        nrnb,
1016                gmx_wallcycle*                 wcycle,
1017                matrix                         vir_q,
1018                matrix                         vir_lj,
1019                real*                          energy_q,
1020                real*                          energy_lj,
1021                real                           lambda_q,
1022                real                           lambda_lj,
1023                real*                          dvdlambda_q,
1024                real*                          dvdlambda_lj,
1025                int                            flags)
1026 {
1027     GMX_ASSERT(pme->runMode == PmeRunMode::CPU,
1028                "gmx_pme_do should not be called on the GPU PME run.");
1029
1030     int                  d, npme, grid_index, max_grid_index;
1031     PmeAtomComm&         atc         = pme->atc[0];
1032     pmegrids_t*          pmegrid     = nullptr;
1033     real*                grid        = nullptr;
1034     real*                coefficient = nullptr;
1035     PmeOutput            output[2]; // The second is used for the B state with FEP
1036     real                 scale, lambda;
1037     gmx_bool             bClearF;
1038     gmx_parallel_3dfft_t pfft_setup;
1039     real*                fftgrid;
1040     t_complex*           cfftgrid;
1041     int                  thread;
1042     gmx_bool             bFirst, bDoSplines;
1043     int                  fep_state;
1044     int                  fep_states_lj = pme->bFEP_lj ? 2 : 1;
1045     const gmx_bool       bCalcEnerVir  = (flags & GMX_PME_CALC_ENER_VIR) != 0;
1046     const gmx_bool       bBackFFT      = (flags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
1047     const gmx_bool       bCalcF        = (flags & GMX_PME_CALC_F) != 0;
1048
1049     /* We could be passing lambda!=0 while no q or LJ is actually perturbed */
1050     if (!pme->bFEP_q)
1051     {
1052         lambda_q = 0;
1053     }
1054     if (!pme->bFEP_lj)
1055     {
1056         lambda_lj = 0;
1057     }
1058
1059     assert(pme->nnodes > 0);
1060     assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1061
1062     if (pme->nnodes > 1)
1063     {
1064         atc.pd.resize(coordinates.ssize());
1065         for (int d = pme->ndecompdim - 1; d >= 0; d--)
1066         {
1067             PmeAtomComm& atc = pme->atc[d];
1068             atc.maxshift     = (atc.dimind == 0 ? maxshift_x : maxshift_y);
1069         }
1070     }
1071     else
1072     {
1073         GMX_ASSERT(coordinates.ssize() == atc.numAtoms(), "We expect atc.numAtoms() coordinates");
1074         GMX_ASSERT(forces.ssize() >= atc.numAtoms(),
1075                    "We need a force buffer with at least atc.numAtoms() elements");
1076
1077         atc.x = coordinates;
1078         atc.f = forces;
1079     }
1080
1081     matrix scaledBox;
1082     pme->boxScaler->scaleBox(box, scaledBox);
1083
1084     gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1085     bFirst = TRUE;
1086
1087     /* For simplicity, we construct the splines for all particles if
1088      * more than one PME calculations is needed. Some optimization
1089      * could be done by keeping track of which atoms have splines
1090      * constructed, and construct new splines on each pass for atoms
1091      * that don't yet have them.
1092      */
1093
1094     bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1095
1096     /* We need a maximum of four separate PME calculations:
1097      * grid_index=0: Coulomb PME with charges from state A
1098      * grid_index=1: Coulomb PME with charges from state B
1099      * grid_index=2: LJ PME with C6 from state A
1100      * grid_index=3: LJ PME with C6 from state B
1101      * For Lorentz-Berthelot combination rules, a separate loop is used to
1102      * calculate all the terms
1103      */
1104
1105     /* If we are doing LJ-PME with LB, we only do Q here */
1106     max_grid_index = (pme->ljpme_combination_rule == eljpmeLB) ? DO_Q : DO_Q_AND_LJ;
1107
1108     for (grid_index = 0; grid_index < max_grid_index; ++grid_index)
1109     {
1110         /* Check if we should do calculations at this grid_index
1111          * If grid_index is odd we should be doing FEP
1112          * If grid_index < 2 we should be doing electrostatic PME
1113          * If grid_index >= 2 we should be doing LJ-PME
1114          */
1115         if ((grid_index < DO_Q && (!pme->doCoulomb || (grid_index == 1 && !pme->bFEP_q)))
1116             || (grid_index >= DO_Q && (!pme->doLJ || (grid_index == 3 && !pme->bFEP_lj))))
1117         {
1118             continue;
1119         }
1120         /* Unpack structure */
1121         pmegrid    = &pme->pmegrid[grid_index];
1122         fftgrid    = pme->fftgrid[grid_index];
1123         cfftgrid   = pme->cfftgrid[grid_index];
1124         pfft_setup = pme->pfft_setup[grid_index];
1125         switch (grid_index)
1126         {
1127             case 0: coefficient = chargeA; break;
1128             case 1: coefficient = chargeB; break;
1129             case 2: coefficient = c6A; break;
1130             case 3: coefficient = c6B; break;
1131         }
1132
1133         grid = pmegrid->grid.grid;
1134
1135         if (debug)
1136         {
1137             fprintf(debug, "PME: number of ranks = %d, rank = %d\n", cr->nnodes, cr->nodeid);
1138             fprintf(debug, "Grid = %p\n", static_cast<void*>(grid));
1139             if (grid == nullptr)
1140             {
1141                 gmx_fatal(FARGS, "No grid!");
1142             }
1143         }
1144
1145         if (pme->nnodes == 1)
1146         {
1147             atc.coefficient = gmx::arrayRefFromArray(coefficient, coordinates.size());
1148         }
1149         else
1150         {
1151             wallcycle_start(wcycle, ewcPME_REDISTXF);
1152             do_redist_pos_coeffs(pme, cr, bFirst, coordinates, coefficient);
1153
1154             wallcycle_stop(wcycle, ewcPME_REDISTXF);
1155         }
1156
1157         if (debug)
1158         {
1159             fprintf(debug, "Rank= %6d, pme local particles=%6d\n", cr->nodeid, atc.numAtoms());
1160         }
1161
1162         if (flags & GMX_PME_SPREAD)
1163         {
1164             wallcycle_start(wcycle, ewcPME_SPREAD);
1165
1166             /* Spread the coefficients on a grid */
1167             spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1168
1169             if (bFirst)
1170             {
1171                 inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1172             }
1173             inc_nrnb(nrnb, eNR_SPREADBSP,
1174                      pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1175
1176             if (!pme->bUseThreads)
1177             {
1178                 wrap_periodic_pmegrid(pme, grid);
1179
1180                 /* sum contributions to local grid from other nodes */
1181                 if (pme->nnodes > 1)
1182                 {
1183                     gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1184                 }
1185
1186                 copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1187             }
1188
1189             wallcycle_stop(wcycle, ewcPME_SPREAD);
1190
1191             /* TODO If the OpenMP and single-threaded implementations
1192                converge, then spread_on_grid() and
1193                copy_pmegrid_to_fftgrid() will perhaps live in the same
1194                source file.
1195              */
1196         }
1197
1198         /* Here we start a large thread parallel region */
1199 #pragma omp parallel num_threads(pme->nthread) private(thread)
1200         {
1201             try
1202             {
1203                 thread = gmx_omp_get_thread_num();
1204                 if (flags & GMX_PME_SOLVE)
1205                 {
1206                     int loop_count;
1207
1208                     /* do 3d-fft */
1209                     if (thread == 0)
1210                     {
1211                         wallcycle_start(wcycle, ewcPME_FFT);
1212                     }
1213                     gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1214                     if (thread == 0)
1215                     {
1216                         wallcycle_stop(wcycle, ewcPME_FFT);
1217                     }
1218
1219                     /* solve in k-space for our local cells */
1220                     if (thread == 0)
1221                     {
1222                         wallcycle_start(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1223                     }
1224                     if (grid_index < DO_Q)
1225                     {
1226                         loop_count = solve_pme_yzx(
1227                                 pme, cfftgrid, scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1228                                 bCalcEnerVir, pme->nthread, thread);
1229                     }
1230                     else
1231                     {
1232                         loop_count = solve_pme_lj_yzx(
1233                                 pme, &cfftgrid, FALSE,
1234                                 scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1235                                 bCalcEnerVir, pme->nthread, thread);
1236                     }
1237
1238                     if (thread == 0)
1239                     {
1240                         wallcycle_stop(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1241                         inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1242                     }
1243                 }
1244
1245                 if (bBackFFT)
1246                 {
1247                     /* do 3d-invfft */
1248                     if (thread == 0)
1249                     {
1250                         wallcycle_start(wcycle, ewcPME_FFT);
1251                     }
1252                     gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1253                     if (thread == 0)
1254                     {
1255                         wallcycle_stop(wcycle, ewcPME_FFT);
1256
1257
1258                         if (pme->nodeid == 0)
1259                         {
1260                             real ntot = pme->nkx * pme->nky * pme->nkz;
1261                             npme      = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1262                             inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1263                         }
1264
1265                         /* Note: this wallcycle region is closed below
1266                            outside an OpenMP region, so take care if
1267                            refactoring code here. */
1268                         wallcycle_start(wcycle, ewcPME_GATHER);
1269                     }
1270
1271                     copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1272                 }
1273             }
1274             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1275         }
1276         /* End of thread parallel section.
1277          * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1278          */
1279
1280         if (bBackFFT)
1281         {
1282             /* distribute local grid to all nodes */
1283             if (pme->nnodes > 1)
1284             {
1285                 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1286             }
1287
1288             unwrap_periodic_pmegrid(pme, grid);
1289         }
1290
1291         if (bCalcF)
1292         {
1293             /* interpolate forces for our local atoms */
1294
1295
1296             /* If we are running without parallelization,
1297              * atc->f is the actual force array, not a buffer,
1298              * therefore we should not clear it.
1299              */
1300             lambda  = grid_index < DO_Q ? lambda_q : lambda_lj;
1301             bClearF = (bFirst && PAR(cr));
1302 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1303             for (thread = 0; thread < pme->nthread; thread++)
1304             {
1305                 try
1306                 {
1307                     gather_f_bsplines(pme, grid, bClearF, &atc, &atc.spline[thread],
1308                                       pme->bFEP ? (grid_index % 2 == 0 ? 1.0 - lambda : lambda) : 1.0);
1309                 }
1310                 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1311             }
1312
1313
1314             inc_nrnb(nrnb, eNR_GATHERFBSP,
1315                      pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1316             /* Note: this wallcycle region is opened above inside an OpenMP
1317                region, so take care if refactoring code here. */
1318             wallcycle_stop(wcycle, ewcPME_GATHER);
1319         }
1320
1321         if (bCalcEnerVir)
1322         {
1323             /* This should only be called on the master thread
1324              * and after the threads have synchronized.
1325              */
1326             if (grid_index < 2)
1327             {
1328                 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1329             }
1330             else
1331             {
1332                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1333             }
1334         }
1335         bFirst = FALSE;
1336     } /* of grid_index-loop */
1337
1338     /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1339      * seven terms. */
1340
1341     if (pme->doLJ && pme->ljpme_combination_rule == eljpmeLB)
1342     {
1343         /* Loop over A- and B-state if we are doing FEP */
1344         for (fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1345         {
1346             real *local_c6 = nullptr, *local_sigma = nullptr, *RedistC6 = nullptr, *RedistSigma = nullptr;
1347             gmx::ArrayRef<real> coefficientBuffer;
1348             if (pme->nnodes == 1)
1349             {
1350                 pme->lb_buf1.resize(atc.numAtoms());
1351                 coefficientBuffer = pme->lb_buf1;
1352                 switch (fep_state)
1353                 {
1354                     case 0:
1355                         local_c6    = c6A;
1356                         local_sigma = sigmaA;
1357                         break;
1358                     case 1:
1359                         local_c6    = c6B;
1360                         local_sigma = sigmaB;
1361                         break;
1362                     default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1363                 }
1364             }
1365             else
1366             {
1367                 coefficientBuffer = atc.coefficientBuffer;
1368                 switch (fep_state)
1369                 {
1370                     case 0:
1371                         RedistC6    = c6A;
1372                         RedistSigma = sigmaA;
1373                         break;
1374                     case 1:
1375                         RedistC6    = c6B;
1376                         RedistSigma = sigmaB;
1377                         break;
1378                     default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1379                 }
1380                 wallcycle_start(wcycle, ewcPME_REDISTXF);
1381
1382                 do_redist_pos_coeffs(pme, cr, bFirst, coordinates, RedistC6);
1383                 pme->lb_buf1.resize(atc.numAtoms());
1384                 pme->lb_buf2.resize(atc.numAtoms());
1385                 local_c6 = pme->lb_buf1.data();
1386                 for (int i = 0; i < atc.numAtoms(); ++i)
1387                 {
1388                     local_c6[i] = atc.coefficient[i];
1389                 }
1390
1391                 do_redist_pos_coeffs(pme, cr, FALSE, coordinates, RedistSigma);
1392                 local_sigma = pme->lb_buf2.data();
1393                 for (int i = 0; i < atc.numAtoms(); ++i)
1394                 {
1395                     local_sigma[i] = atc.coefficient[i];
1396                 }
1397
1398                 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1399             }
1400             atc.coefficient = coefficientBuffer;
1401             calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1402
1403             /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1404             for (grid_index = 2; grid_index < 9; ++grid_index)
1405             {
1406                 /* Unpack structure */
1407                 pmegrid    = &pme->pmegrid[grid_index];
1408                 fftgrid    = pme->fftgrid[grid_index];
1409                 pfft_setup = pme->pfft_setup[grid_index];
1410                 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1411                 grid = pmegrid->grid.grid;
1412
1413                 if (flags & GMX_PME_SPREAD)
1414                 {
1415                     wallcycle_start(wcycle, ewcPME_SPREAD);
1416                     /* Spread the c6 on a grid */
1417                     spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1418
1419                     if (bFirst)
1420                     {
1421                         inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1422                     }
1423
1424                     inc_nrnb(nrnb, eNR_SPREADBSP,
1425                              pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1426                     if (pme->nthread == 1)
1427                     {
1428                         wrap_periodic_pmegrid(pme, grid);
1429                         /* sum contributions to local grid from other nodes */
1430                         if (pme->nnodes > 1)
1431                         {
1432                             gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1433                         }
1434                         copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1435                     }
1436                     wallcycle_stop(wcycle, ewcPME_SPREAD);
1437                 }
1438                 /*Here we start a large thread parallel region*/
1439 #pragma omp parallel num_threads(pme->nthread) private(thread)
1440                 {
1441                     try
1442                     {
1443                         thread = gmx_omp_get_thread_num();
1444                         if (flags & GMX_PME_SOLVE)
1445                         {
1446                             /* do 3d-fft */
1447                             if (thread == 0)
1448                             {
1449                                 wallcycle_start(wcycle, ewcPME_FFT);
1450                             }
1451
1452                             gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1453                             if (thread == 0)
1454                             {
1455                                 wallcycle_stop(wcycle, ewcPME_FFT);
1456                             }
1457                         }
1458                     }
1459                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1460                 }
1461                 bFirst = FALSE;
1462             }
1463             if (flags & GMX_PME_SOLVE)
1464             {
1465                 /* solve in k-space for our local cells */
1466 #pragma omp parallel num_threads(pme->nthread) private(thread)
1467                 {
1468                     try
1469                     {
1470                         int loop_count;
1471                         thread = gmx_omp_get_thread_num();
1472                         if (thread == 0)
1473                         {
1474                             wallcycle_start(wcycle, ewcLJPME);
1475                         }
1476
1477                         loop_count = solve_pme_lj_yzx(
1478                                 pme, &pme->cfftgrid[2], TRUE,
1479                                 scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1480                                 bCalcEnerVir, pme->nthread, thread);
1481                         if (thread == 0)
1482                         {
1483                             wallcycle_stop(wcycle, ewcLJPME);
1484                             inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1485                         }
1486                     }
1487                     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1488                 }
1489             }
1490
1491             if (bCalcEnerVir)
1492             {
1493                 /* This should only be called on the master thread and
1494                  * after the threads have synchronized.
1495                  */
1496                 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[fep_state]);
1497             }
1498
1499             if (bBackFFT)
1500             {
1501                 bFirst = !pme->doCoulomb;
1502                 calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1503                 for (grid_index = 8; grid_index >= 2; --grid_index)
1504                 {
1505                     /* Unpack structure */
1506                     pmegrid    = &pme->pmegrid[grid_index];
1507                     fftgrid    = pme->fftgrid[grid_index];
1508                     pfft_setup = pme->pfft_setup[grid_index];
1509                     grid       = pmegrid->grid.grid;
1510                     calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1511 #pragma omp parallel num_threads(pme->nthread) private(thread)
1512                     {
1513                         try
1514                         {
1515                             thread = gmx_omp_get_thread_num();
1516                             /* do 3d-invfft */
1517                             if (thread == 0)
1518                             {
1519                                 wallcycle_start(wcycle, ewcPME_FFT);
1520                             }
1521
1522                             gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1523                             if (thread == 0)
1524                             {
1525                                 wallcycle_stop(wcycle, ewcPME_FFT);
1526
1527
1528                                 if (pme->nodeid == 0)
1529                                 {
1530                                     real ntot = pme->nkx * pme->nky * pme->nkz;
1531                                     npme = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1532                                     inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1533                                 }
1534                                 wallcycle_start(wcycle, ewcPME_GATHER);
1535                             }
1536
1537                             copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1538                         }
1539                         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1540                     } /*#pragma omp parallel*/
1541
1542                     /* distribute local grid to all nodes */
1543                     if (pme->nnodes > 1)
1544                     {
1545                         gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1546                     }
1547
1548                     unwrap_periodic_pmegrid(pme, grid);
1549
1550                     if (bCalcF)
1551                     {
1552                         /* interpolate forces for our local atoms */
1553                         bClearF = (bFirst && PAR(cr));
1554                         scale   = pme->bFEP ? (fep_state < 1 ? 1.0 - lambda_lj : lambda_lj) : 1.0;
1555                         scale *= lb_scale_factor[grid_index - 2];
1556
1557 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1558                         for (thread = 0; thread < pme->nthread; thread++)
1559                         {
1560                             try
1561                             {
1562                                 gather_f_bsplines(pme, grid, bClearF, &pme->atc[0],
1563                                                   &pme->atc[0].spline[thread], scale);
1564                             }
1565                             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1566                         }
1567
1568
1569                         inc_nrnb(nrnb, eNR_GATHERFBSP,
1570                                  pme->pme_order * pme->pme_order * pme->pme_order * pme->atc[0].numAtoms());
1571                     }
1572                     wallcycle_stop(wcycle, ewcPME_GATHER);
1573
1574                     bFirst = FALSE;
1575                 } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1576             }     /* if (bCalcF) */
1577         }         /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1578     }             /* if ((flags & GMX_PME_DO_LJ) && pme->ljpme_combination_rule == eljpmeLB) */
1579
1580     if (bCalcF && pme->nnodes > 1)
1581     {
1582         wallcycle_start(wcycle, ewcPME_REDISTXF);
1583         for (d = 0; d < pme->ndecompdim; d++)
1584         {
1585             gmx::ArrayRef<gmx::RVec> forcesRef;
1586             if (d == pme->ndecompdim - 1)
1587             {
1588                 const size_t numAtoms = coordinates.size();
1589                 GMX_ASSERT(forces.size() >= numAtoms, "Need at least numAtoms forces");
1590                 forcesRef = forces.subArray(0, numAtoms);
1591             }
1592             else
1593             {
1594                 forcesRef = pme->atc[d + 1].f;
1595             }
1596             if (DOMAINDECOMP(cr))
1597             {
1598                 dd_pmeredist_f(pme, &pme->atc[d], forcesRef, d == pme->ndecompdim - 1 && pme->bPPnode);
1599             }
1600         }
1601
1602         wallcycle_stop(wcycle, ewcPME_REDISTXF);
1603     }
1604
1605     if (bCalcEnerVir)
1606     {
1607         if (pme->doCoulomb)
1608         {
1609             if (!pme->bFEP_q)
1610             {
1611                 *energy_q = output[0].coulombEnergy_;
1612                 m_add(vir_q, output[0].coulombVirial_, vir_q);
1613             }
1614             else
1615             {
1616                 *energy_q = (1.0 - lambda_q) * output[0].coulombEnergy_ + lambda_q * output[1].coulombEnergy_;
1617                 *dvdlambda_q += output[1].coulombEnergy_ - output[0].coulombEnergy_;
1618                 for (int i = 0; i < DIM; i++)
1619                 {
1620                     for (int j = 0; j < DIM; j++)
1621                     {
1622                         vir_q[i][j] += (1.0 - lambda_q) * output[0].coulombVirial_[i][j]
1623                                        + lambda_q * output[1].coulombVirial_[i][j];
1624                     }
1625                 }
1626             }
1627             if (debug)
1628             {
1629                 fprintf(debug, "Electrostatic PME mesh energy: %g\n", *energy_q);
1630             }
1631         }
1632         else
1633         {
1634             *energy_q = 0;
1635         }
1636
1637         if (pme->doLJ)
1638         {
1639             if (!pme->bFEP_lj)
1640             {
1641                 *energy_lj = output[0].lennardJonesEnergy_;
1642                 m_add(vir_lj, output[0].lennardJonesVirial_, vir_lj);
1643             }
1644             else
1645             {
1646                 *energy_lj = (1.0 - lambda_lj) * output[0].lennardJonesEnergy_
1647                              + lambda_lj * output[1].lennardJonesEnergy_;
1648                 *dvdlambda_lj += output[1].lennardJonesEnergy_ - output[0].lennardJonesEnergy_;
1649                 for (int i = 0; i < DIM; i++)
1650                 {
1651                     for (int j = 0; j < DIM; j++)
1652                     {
1653                         vir_lj[i][j] += (1.0 - lambda_lj) * output[0].lennardJonesVirial_[i][j]
1654                                         + lambda_lj * output[1].lennardJonesVirial_[i][j];
1655                     }
1656                 }
1657             }
1658             if (debug)
1659             {
1660                 fprintf(debug, "Lennard-Jones PME mesh energy: %g\n", *energy_lj);
1661             }
1662         }
1663         else
1664         {
1665             *energy_lj = 0;
1666         }
1667     }
1668     return 0;
1669 }
1670
1671 void gmx_pme_destroy(gmx_pme_t* pme)
1672 {
1673     if (!pme)
1674     {
1675         return;
1676     }
1677
1678     delete pme->boxScaler;
1679
1680     sfree(pme->nnx);
1681     sfree(pme->nny);
1682     sfree(pme->nnz);
1683     sfree(pme->fshx);
1684     sfree(pme->fshy);
1685     sfree(pme->fshz);
1686
1687     for (int i = 0; i < pme->ngrids; ++i)
1688     {
1689         pmegrids_destroy(&pme->pmegrid[i]);
1690     }
1691     if (pme->pfft_setup)
1692     {
1693         for (int i = 0; i < pme->ngrids; ++i)
1694         {
1695             gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1696         }
1697     }
1698     sfree(pme->fftgrid);
1699     sfree(pme->cfftgrid);
1700     sfree(pme->pfft_setup);
1701
1702     for (int i = 0; i < DIM; i++)
1703     {
1704         sfree(pme->bsp_mod[i]);
1705     }
1706
1707     sfree(pme->bufv);
1708     sfree(pme->bufr);
1709
1710     if (pme->solve_work)
1711     {
1712         pme_free_all_work(&pme->solve_work, pme->nthread);
1713     }
1714
1715     sfree(pme->sum_qgrid_tmp);
1716     sfree(pme->sum_qgrid_dd_tmp);
1717
1718     destroy_pme_spline_work(pme->spline_work);
1719
1720     if (pme_gpu_active(pme) && pme->gpu)
1721     {
1722         pme_gpu_destroy(pme->gpu);
1723     }
1724
1725     delete pme;
1726 }
1727
1728 void gmx_pme_reinit_atoms(gmx_pme_t* pme, const int numAtoms, const real* charges)
1729 {
1730     if (pme_gpu_active(pme))
1731     {
1732         pme_gpu_reinit_atoms(pme->gpu, numAtoms, charges);
1733     }
1734     else
1735     {
1736         pme->atc[0].setNumAtoms(numAtoms);
1737         // TODO: set the charges here as well
1738     }
1739 }