src/gromacs/ewald/pme.cpp

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