src/gromacs/ewald/pme_only.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 by the GROMACS development team.
   7  * Copyright (c) 2018,2019,2020,2021, 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 /* IMPORTANT FOR DEVELOPERS:
  39  *
  40  * Triclinic pme stuff isn't entirely trivial, and we've experienced
  41  * some bugs during development (many of them due to me). To avoid
  42  * this in the future, please check the following things if you make
  43  * changes in this file:
  44  *
  45  * 1. You should obtain identical (at least to the PME precision)
  46  *    energies, forces, and virial for
  47  *    a rectangular box and a triclinic one where the z (or y) axis is
  48  *    tilted a whole box side. For instance you could use these boxes:
  49  *
  50  *    rectangular       triclinic
  51  *     2  0  0           2  0  0
  52  *     0  2  0           0  2  0
  53  *     0  0  6           2  2  6
  54  *
  55  * 2. You should check the energy conservation in a triclinic box.
  56  *
  57  * It might seem an overkill, but better safe than sorry.
  58  * /Erik 001109
  59  */
  60
  61 #include "gmxpre.h"
  62
  63 #include "pme_only.h"
  64
  65 #include "config.h"
  66
  67 #include <cassert>
  68 #include <cmath>
  69 #include <cstdio>
  70 #include <cstdlib>
  71 #include <cstring>
  72
  73 #include <memory>
  74 #include <numeric>
  75 #include <vector>
  76
  77 #include "gromacs/domdec/domdec.h"
  78 #include "gromacs/ewald/pme.h"
  79 #include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
  80 #include "gromacs/ewald/pme_force_sender_gpu.h"
  81 #include "gromacs/fft/parallel_3dfft.h"
  82 #include "gromacs/fileio/pdbio.h"
  83 #include "gromacs/gmxlib/network.h"
  84 #include "gromacs/gmxlib/nrnb.h"
  85 #include "gromacs/gpu_utils/device_stream_manager.h"
  86 #include "gromacs/gpu_utils/hostallocator.h"
  87 #include "gromacs/math/gmxcomplex.h"
  88 #include "gromacs/math/units.h"
  89 #include "gromacs/math/vec.h"
  90 #include "gromacs/mdtypes/commrec.h"
  91 #include "gromacs/mdtypes/forceoutput.h"
  92 #include "gromacs/mdtypes/inputrec.h"
  93 #include "gromacs/mdtypes/simulation_workload.h"
  94 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  95 #include "gromacs/timing/cyclecounter.h"
  96 #include "gromacs/timing/wallcycle.h"
  97 #include "gromacs/utility/fatalerror.h"
  98 #include "gromacs/utility/futil.h"
  99 #include "gromacs/utility/gmxmpi.h"
 100 #include "gromacs/utility/gmxomp.h"
 101 #include "gromacs/utility/smalloc.h"
 102
 103 #include "pme_gpu_internal.h"
 104 #include "pme_internal.h"
 105 #include "pme_output.h"
 106 #include "pme_pp_communication.h"
 107
 108 /*! \brief Master PP-PME communication data structure */
 109 struct gmx_pme_pp
 110 {
 111     MPI_Comm             mpi_comm_mysim; /**< MPI communicator for this simulation */
 112     std::vector<PpRanks> ppRanks;        /**< The PP partner ranks                 */
 113     int                  peerRankId;     /**< The peer PP rank id                  */
 114     //@{
 115     /**< Vectors of A- and B-state parameters used to transfer vectors to PME ranks  */
 116     gmx::PaddedHostVector<real> chargeA;
 117     std::vector<real>           chargeB;
 118     std::vector<real>           sqrt_c6A;
 119     std::vector<real>           sqrt_c6B;
 120     std::vector<real>           sigmaA;
 121     std::vector<real>           sigmaB;
 122     //@}
 123     gmx::HostVector<gmx::RVec> x; /**< Vector of atom coordinates to transfer to PME ranks */
 124     std::vector<gmx::RVec>     f; /**< Vector of atom forces received from PME ranks */
 125     //@{
 126     /**< Vectors of MPI objects used in non-blocking communication between multiple PP ranks per PME rank */
 127     std::vector<MPI_Request> req;
 128     std::vector<MPI_Status>  stat;
 129     //@}
 130
 131     /*! \brief object for receiving coordinates using communications operating on GPU memory space */
 132     std::unique_ptr<gmx::PmeCoordinateReceiverGpu> pmeCoordinateReceiverGpu;
 133     /*! \brief object for sending PME force using communications operating on GPU memory space */
 134     std::unique_ptr<gmx::PmeForceSenderGpu> pmeForceSenderGpu;
 135
 136     /*! \brief whether GPU direct communications are active for PME-PP transfers */
 137     bool useGpuDirectComm = false;
 138     /*! \brief whether GPU direct communications should send forces directly to remote GPU memory */
 139     bool sendForcesDirectToPpGpu = false;
 140 };
 141
 142 /*! \brief Initialize the PME-only side of the PME <-> PP communication */
 143 static std::unique_ptr<gmx_pme_pp> gmx_pme_pp_init(const t_commrec* cr)
 144 {
 145     auto pme_pp = std::make_unique<gmx_pme_pp>();
 146
 147 #if GMX_MPI
 148     int rank;
 149
 150     pme_pp->mpi_comm_mysim = cr->mpi_comm_mysim;
 151     MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
 152     auto ppRanks = get_pme_ddranks(cr, rank);
 153     pme_pp->ppRanks.reserve(ppRanks.size());
 154     for (const auto& ppRankId : ppRanks)
 155     {
 156         pme_pp->ppRanks.push_back({ ppRankId, 0 });
 157     }
 158     // The peer PP rank is the last one.
 159     pme_pp->peerRankId = pme_pp->ppRanks.back().rankId;
 160     pme_pp->req.resize(eCommType_NR * pme_pp->ppRanks.size());
 161     pme_pp->stat.resize(eCommType_NR * pme_pp->ppRanks.size());
 162 #else
 163     GMX_UNUSED_VALUE(cr);
 164 #endif
 165
 166     return pme_pp;
 167 }
 168
 169 static void reset_pmeonly_counters(gmx_wallcycle*            wcycle,
 170                                    gmx_walltime_accounting_t walltime_accounting,
 171                                    t_nrnb*                   nrnb,
 172                                    int64_t                   step,
 173                                    bool                      useGpuForPme)
 174 {
 175     /* Reset all the counters related to performance over the run */
 176     wallcycle_stop(wcycle, WallCycleCounter::Run);
 177     wallcycle_reset_all(wcycle);
 178     *nrnb = { 0 };
 179     wallcycle_start(wcycle, WallCycleCounter::Run);
 180     walltime_accounting_reset_time(walltime_accounting, step);
 181
 182     if (useGpuForPme)
 183     {
 184         resetGpuProfiler();
 185     }
 186 }
 187
 188 static gmx_pme_t* gmx_pmeonly_switch(std::vector<gmx_pme_t*>* pmedata,
 189                                      const ivec               grid_size,
 190                                      real                     ewaldcoeff_q,
 191                                      real                     ewaldcoeff_lj,
 192                                      const t_commrec*         cr,
 193                                      const t_inputrec*        ir)
 194 {
 195     GMX_ASSERT(pmedata, "Bad PME tuning list pointer");
 196     for (auto& pme : *pmedata)
 197     {
 198         GMX_ASSERT(pme, "Bad PME tuning list element pointer");
 199         if (gmx_pme_grid_matches(*pme, grid_size))
 200         {
 201             /* Here we have found an existing PME data structure that suits us.
 202              * However, in the GPU case, we have to reinitialize it - there's only one GPU structure.
 203              * This should not cause actual GPU reallocations, at least (the allocated buffers are never shrunk).
 204              * So, just some grid size updates in the GPU kernel parameters.
 205              * TODO: this should be something like gmx_pme_update_split_params()
 206              */
 207             gmx_pme_reinit(&pme, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 208             return pme;
 209         }
 210     }
 211
 212     const auto& pme          = pmedata->back();
 213     gmx_pme_t*  newStructure = nullptr;
 214     // Copy last structure with new grid params
 215     gmx_pme_reinit(&newStructure, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 216     pmedata->push_back(newStructure);
 217     return newStructure;
 218 }
 219
 220 /*! \brief Called by PME-only ranks to receive coefficients and coordinates
 221  *
 222  * Note that with GPU direct communication the transfer is only initiated, it is the responsibility
 223  * of the caller to synchronize prior to launching spread.
 224  *
 225  * \param[in] pme                     PME data structure.
 226  * \param[in,out] pme_pp              PME-PP communication structure.
 227  * \param[out] natoms                 Number of received atoms.
 228  * \param[out] box                    System box, if received.
 229  * \param[out] maxshift_x             Maximum shift in X direction, if received.
 230  * \param[out] maxshift_y             Maximum shift in Y direction, if received.
 231  * \param[out] lambda_q               Free-energy lambda for electrostatics, if received.
 232  * \param[out] lambda_lj              Free-energy lambda for Lennard-Jones, if received.
 233  * \param[out] computeEnergyAndVirial Set to true if this is an energy/virial calculation
 234  *                                    step, otherwise set to false.
 235  * \param[out] step                   MD integration step number.
 236  * \param[out] grid_size              PME grid size, if received.
 237  * \param[out] ewaldcoeff_q           Ewald cut-off parameter for electrostatics, if received.
 238  * \param[out] ewaldcoeff_lj          Ewald cut-off parameter for Lennard-Jones, if received.
 239  * \param[in]  useGpuForPme           Flag on whether PME is on GPU.
 240  * \param[in]  stateGpu               GPU state propagator object.
 241  * \param[in]  runMode                PME run mode.
 242  *
 243  * \retval pmerecvqxX                 All parameters were set, chargeA and chargeB can be NULL.
 244  * \retval pmerecvqxFINISH            No parameters were set.
 245  * \retval pmerecvqxSWITCHGRID        Only grid_size and *ewaldcoeff were set.
 246  * \retval pmerecvqxRESETCOUNTERS     *step was set.
 247  */
 248 static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t*            pme,
 249                                       gmx_pme_pp*                  pme_pp,
 250                                       int*                         natoms,
 251                                       matrix                       box,
 252                                       int*                         maxshift_x,
 253                                       int*                         maxshift_y,
 254                                       real*                        lambda_q,
 255                                       real*                        lambda_lj,
 256                                       gmx_bool*                    computeEnergyAndVirial,
 257                                       int64_t*                     step,
 258                                       ivec*                        grid_size,
 259                                       real*                        ewaldcoeff_q,
 260                                       real*                        ewaldcoeff_lj,
 261                                       bool                         useGpuForPme,
 262                                       gmx::StatePropagatorDataGpu* stateGpu,
 263                                       PmeRunMode gmx_unused        runMode)
 264 {
 265     int status = -1;
 266     int nat    = 0;
 267
 268 #if GMX_MPI
 269     int  messages       = 0;
 270     bool atomSetChanged = false;
 271
 272     do
 273     {
 274         gmx_pme_comm_n_box_t cnb;
 275         cnb.flags = 0;
 276
 277         /* Receive the send count, box and time step from the peer PP node */
 278         MPI_Recv(&cnb, sizeof(cnb), MPI_BYTE, pme_pp->peerRankId, eCommType_CNB, pme_pp->mpi_comm_mysim, MPI_STATUS_IGNORE);
 279
 280         *step = cnb.step;
 281
 282         if (debug)
 283         {
 284             fprintf(debug,
 285                     "PME only rank receiving:%s%s%s%s%s\n",
 286                     (cnb.flags & PP_PME_CHARGE) ? " charges" : "",
 287                     (cnb.flags & PP_PME_COORD) ? " coordinates" : "",
 288                     (cnb.flags & PP_PME_FINISH) ? " finish" : "",
 289                     (cnb.flags & PP_PME_SWITCHGRID) ? " switch grid" : "",
 290                     (cnb.flags & PP_PME_RESETCOUNTERS) ? " reset counters" : "");
 291         }
 292
 293         pme_pp->useGpuDirectComm = ((cnb.flags & PP_PME_GPUCOMMS) != 0);
 294         GMX_ASSERT(!pme_pp->useGpuDirectComm || (pme_pp->pmeForceSenderGpu != nullptr),
 295                    "The use of GPU direct communication for PME-PP is enabled, "
 296                    "but the PME GPU force reciever object does not exist");
 297         pme_pp->sendForcesDirectToPpGpu = ((cnb.flags & PP_PME_RECVFTOGPU) != 0);
 298
 299         if (cnb.flags & PP_PME_FINISH)
 300         {
 301             status = pmerecvqxFINISH;
 302         }
 303
 304         if (cnb.flags & PP_PME_SWITCHGRID)
 305         {
 306             /* Special case, receive the new parameters and return */
 307             copy_ivec(cnb.grid_size, *grid_size);
 308             *ewaldcoeff_q  = cnb.ewaldcoeff_q;
 309             *ewaldcoeff_lj = cnb.ewaldcoeff_lj;
 310
 311             status = pmerecvqxSWITCHGRID;
 312         }
 313
 314         if (cnb.flags & PP_PME_RESETCOUNTERS)
 315         {
 316             /* Special case, receive the step (set above) and return */
 317             status = pmerecvqxRESETCOUNTERS;
 318         }
 319
 320         if (cnb.flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
 321         {
 322             atomSetChanged = true;
 323
 324             /* Receive the send counts from the other PP nodes */
 325             for (auto& sender : pme_pp->ppRanks)
 326             {
 327                 if (sender.rankId == pme_pp->peerRankId)
 328                 {
 329                     sender.numAtoms = cnb.natoms;
 330                 }
 331                 else
 332                 {
 333                     MPI_Irecv(&sender.numAtoms,
 334                               sizeof(sender.numAtoms),
 335                               MPI_BYTE,
 336                               sender.rankId,
 337                               eCommType_CNB,
 338                               pme_pp->mpi_comm_mysim,
 339                               &pme_pp->req[messages++]);
 340                 }
 341             }
 342             MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 343             messages = 0;
 344
 345             nat = 0;
 346             for (const auto& sender : pme_pp->ppRanks)
 347             {
 348                 nat += sender.numAtoms;
 349             }
 350
 351             if (cnb.flags & PP_PME_CHARGE)
 352             {
 353                 pme_pp->chargeA.resizeWithPadding(nat);
 354             }
 355             if (cnb.flags & PP_PME_CHARGEB)
 356             {
 357                 pme_pp->chargeB.resize(nat);
 358             }
 359             if (cnb.flags & PP_PME_SQRTC6)
 360             {
 361                 pme_pp->sqrt_c6A.resize(nat);
 362             }
 363             if (cnb.flags & PP_PME_SQRTC6B)
 364             {
 365                 pme_pp->sqrt_c6B.resize(nat);
 366             }
 367             if (cnb.flags & PP_PME_SIGMA)
 368             {
 369                 pme_pp->sigmaA.resize(nat);
 370             }
 371             if (cnb.flags & PP_PME_SIGMAB)
 372             {
 373                 pme_pp->sigmaB.resize(nat);
 374             }
 375             pme_pp->x.resize(nat);
 376             pme_pp->f.resize(nat);
 377
 378             /* maxshift is sent when the charges are sent */
 379             *maxshift_x = cnb.maxshift_x;
 380             *maxshift_y = cnb.maxshift_y;
 381
 382             /* Receive the charges in place */
 383             for (int q = 0; q < eCommType_NR; q++)
 384             {
 385                 real* bufferPtr;
 386
 387                 if (!(cnb.flags & (PP_PME_CHARGE << q)))
 388                 {
 389                     continue;
 390                 }
 391                 switch (q)
 392                 {
 393                     case eCommType_ChargeA: bufferPtr = pme_pp->chargeA.data(); break;
 394                     case eCommType_ChargeB: bufferPtr = pme_pp->chargeB.data(); break;
 395                     case eCommType_SQRTC6A: bufferPtr = pme_pp->sqrt_c6A.data(); break;
 396                     case eCommType_SQRTC6B: bufferPtr = pme_pp->sqrt_c6B.data(); break;
 397                     case eCommType_SigmaA: bufferPtr = pme_pp->sigmaA.data(); break;
 398                     case eCommType_SigmaB: bufferPtr = pme_pp->sigmaB.data(); break;
 399                     default: gmx_incons("Wrong eCommType");
 400                 }
 401                 nat = 0;
 402                 for (const auto& sender : pme_pp->ppRanks)
 403                 {
 404                     if (sender.numAtoms > 0)
 405                     {
 406                         MPI_Irecv(bufferPtr + nat,
 407                                   sender.numAtoms * sizeof(real),
 408                                   MPI_BYTE,
 409                                   sender.rankId,
 410                                   q,
 411                                   pme_pp->mpi_comm_mysim,
 412                                   &pme_pp->req[messages++]);
 413                         nat += sender.numAtoms;
 414                         if (debug)
 415                         {
 416                             fprintf(debug,
 417                                     "Received from PP rank %d: %d %s\n",
 418                                     sender.rankId,
 419                                     sender.numAtoms,
 420                                     (q == eCommType_ChargeA || q == eCommType_ChargeB) ? "charges"
 421                                                                                        : "params");
 422                         }
 423                     }
 424                 }
 425             }
 426         }
 427
 428         if (cnb.flags & PP_PME_COORD)
 429         {
 430             if (atomSetChanged)
 431             {
 432                 gmx_pme_reinit_atoms(pme, nat, pme_pp->chargeA, pme_pp->chargeB);
 433                 if (useGpuForPme)
 434                 {
 435                     stateGpu->reinit(nat, nat);
 436                     pme_gpu_set_device_x(pme, stateGpu->getCoordinates());
 437                 }
 438                 if (pme_pp->useGpuDirectComm)
 439                 {
 440                     GMX_ASSERT(runMode == PmeRunMode::GPU,
 441                                "GPU Direct PME-PP communication has been enabled, "
 442                                "but PME run mode is not PmeRunMode::GPU\n");
 443
 444                     // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses and re-set atom ranges associated with transfers.
 445                     pme_pp->pmeCoordinateReceiverGpu->reinitCoordinateReceiver(stateGpu->getCoordinates());
 446                     pme_pp->pmeForceSenderGpu->setForceSendBuffer(pme_gpu_get_device_f(pme));
 447                 }
 448             }
 449
 450
 451             /* The box, FE flag and lambda are sent along with the coordinates
 452              *  */
 453             copy_mat(cnb.box, box);
 454             *lambda_q               = cnb.lambda_q;
 455             *lambda_lj              = cnb.lambda_lj;
 456             *computeEnergyAndVirial = ((cnb.flags & PP_PME_ENER_VIR) != 0U);
 457             *step                   = cnb.step;
 458
 459             /* Receive the coordinates in place */
 460             nat = 0;
 461             for (const auto& sender : pme_pp->ppRanks)
 462             {
 463                 if (sender.numAtoms > 0)
 464                 {
 465                     if (pme_pp->useGpuDirectComm)
 466                     {
 467                         if (GMX_THREAD_MPI)
 468                         {
 469                             pme_pp->pmeCoordinateReceiverGpu->receiveCoordinatesSynchronizerFromPpCudaDirect(
 470                                     sender.rankId);
 471                         }
 472                         else
 473                         {
 474                             pme_pp->pmeCoordinateReceiverGpu->launchReceiveCoordinatesFromPpCudaMpi(
 475                                     stateGpu->getCoordinates(), nat, sender.numAtoms * sizeof(rvec), sender.rankId);
 476                         }
 477                     }
 478                     else
 479                     {
 480                         MPI_Irecv(pme_pp->x[nat],
 481                                   sender.numAtoms * sizeof(rvec),
 482                                   MPI_BYTE,
 483                                   sender.rankId,
 484                                   eCommType_COORD,
 485                                   pme_pp->mpi_comm_mysim,
 486                                   &pme_pp->req[messages++]);
 487                     }
 488                     nat += sender.numAtoms;
 489                     if (debug)
 490                     {
 491                         fprintf(debug,
 492                                 "Received from PP rank %d: %d "
 493                                 "coordinates\n",
 494                                 sender.rankId,
 495                                 sender.numAtoms);
 496                     }
 497                 }
 498             }
 499
 500             status = pmerecvqxX;
 501         }
 502
 503         /* Wait for the coordinates and/or charges to arrive */
 504         MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 505         messages = 0;
 506     } while (status == -1);
 507 #else
 508     GMX_UNUSED_VALUE(pme);
 509     GMX_UNUSED_VALUE(pme_pp);
 510     GMX_UNUSED_VALUE(box);
 511     GMX_UNUSED_VALUE(maxshift_x);
 512     GMX_UNUSED_VALUE(maxshift_y);
 513     GMX_UNUSED_VALUE(lambda_q);
 514     GMX_UNUSED_VALUE(lambda_lj);
 515     GMX_UNUSED_VALUE(computeEnergyAndVirial);
 516     GMX_UNUSED_VALUE(step);
 517     GMX_UNUSED_VALUE(grid_size);
 518     GMX_UNUSED_VALUE(ewaldcoeff_q);
 519     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 520     GMX_UNUSED_VALUE(useGpuForPme);
 521     GMX_UNUSED_VALUE(stateGpu);
 522
 523     status = pmerecvqxX;
 524 #endif
 525
 526     if (status == pmerecvqxX)
 527     {
 528         *natoms = nat;
 529     }
 530
 531     return status;
 532 }
 533
 534 /*! \brief Send the PME mesh force, virial and energy to the PP-only ranks. */
 535 static void gmx_pme_send_force_vir_ener(const gmx_pme_t& pme,
 536                                         gmx_pme_pp*      pme_pp,
 537                                         const PmeOutput& output,
 538                                         real             dvdlambda_q,
 539                                         real             dvdlambda_lj,
 540                                         float            cycles)
 541 {
 542 #if GMX_MPI
 543     gmx_pme_comm_vir_ene_t cve;
 544     int                    messages, ind_start, ind_end;
 545     cve.cycles = cycles;
 546
 547     if (pme_pp->useGpuDirectComm)
 548     {
 549         GMX_ASSERT((pme_pp->pmeForceSenderGpu != nullptr),
 550                    "The use of GPU direct communication for PME-PP is enabled, "
 551                    "but the PME GPU force reciever object does not exist");
 552     }
 553
 554     messages = 0;
 555     ind_end  = 0;
 556
 557     /* Now the evaluated forces have to be transferred to the PP ranks */
 558     if (pme_pp->useGpuDirectComm && GMX_THREAD_MPI)
 559     {
 560         int numPpRanks = static_cast<int>(pme_pp->ppRanks.size());
 561 #    pragma omp parallel for num_threads(std::min(numPpRanks, pme.nthread)) schedule(static)
 562         for (int i = 0; i < numPpRanks; i++)
 563         {
 564             auto& receiver = pme_pp->ppRanks[i];
 565             pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(
 566                     receiver.rankId, receiver.numAtoms, pme_pp->sendForcesDirectToPpGpu);
 567         }
 568     }
 569     else
 570     {
 571         for (const auto& receiver : pme_pp->ppRanks)
 572         {
 573             ind_start = ind_end;
 574             ind_end   = ind_start + receiver.numAtoms;
 575             if (pme_pp->useGpuDirectComm)
 576             {
 577                 pme_pp->pmeForceSenderGpu->sendFToPpCudaMpi(pme_gpu_get_device_f(&pme),
 578                                                             ind_start,
 579                                                             receiver.numAtoms * sizeof(rvec),
 580                                                             receiver.rankId,
 581                                                             &pme_pp->req[messages]);
 582             }
 583             else
 584             {
 585                 void* sendbuf = const_cast<void*>(static_cast<const void*>(output.forces_[ind_start]));
 586                 // Send using MPI
 587                 MPI_Isend(sendbuf,
 588                           receiver.numAtoms * sizeof(rvec),
 589                           MPI_BYTE,
 590                           receiver.rankId,
 591                           0,
 592                           pme_pp->mpi_comm_mysim,
 593                           &pme_pp->req[messages]);
 594             }
 595             messages++;
 596         }
 597     }
 598
 599     /* send virial and energy to our last PP node */
 600     copy_mat(output.coulombVirial_, cve.vir_q);
 601     copy_mat(output.lennardJonesVirial_, cve.vir_lj);
 602     cve.energy_q     = output.coulombEnergy_;
 603     cve.energy_lj    = output.lennardJonesEnergy_;
 604     cve.dvdlambda_q  = dvdlambda_q;
 605     cve.dvdlambda_lj = dvdlambda_lj;
 606     /* check for the signals to send back to a PP node */
 607     cve.stop_cond = gmx_get_stop_condition();
 608
 609     cve.cycles = cycles;
 610
 611     if (debug)
 612     {
 613         fprintf(debug, "PME rank sending to PP rank %d: virial and energy\n", pme_pp->peerRankId);
 614     }
 615     MPI_Isend(&cve, sizeof(cve), MPI_BYTE, pme_pp->peerRankId, 1, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 616
 617     /* Wait for the forces to arrive */
 618     MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 619 #else
 620     GMX_RELEASE_ASSERT(false, "Invalid call to gmx_pme_send_force_vir_ener");
 621     GMX_UNUSED_VALUE(pme);
 622     GMX_UNUSED_VALUE(pme_pp);
 623     GMX_UNUSED_VALUE(output);
 624     GMX_UNUSED_VALUE(dvdlambda_q);
 625     GMX_UNUSED_VALUE(dvdlambda_lj);
 626     GMX_UNUSED_VALUE(cycles);
 627 #endif
 628 }
 629
 630 int gmx_pmeonly(struct gmx_pme_t*               pme,
 631                 const t_commrec*                cr,
 632                 t_nrnb*                         mynrnb,
 633                 gmx_wallcycle*                  wcycle,
 634                 gmx_walltime_accounting_t       walltime_accounting,
 635                 t_inputrec*                     ir,
 636                 PmeRunMode                      runMode,
 637                 bool                            useGpuPmePpCommunication,
 638                 const gmx::DeviceStreamManager* deviceStreamManager)
 639 {
 640     int     ret;
 641     int     natoms = 0;
 642     matrix  box;
 643     real    lambda_q   = 0;
 644     real    lambda_lj  = 0;
 645     int     maxshift_x = 0, maxshift_y = 0;
 646     real    dvdlambda_q, dvdlambda_lj;
 647     float   cycles;
 648     int     count;
 649     bool    computeEnergyAndVirial = false;
 650     int64_t step;
 651
 652     /* This data will only use with PME tuning, i.e. switching PME grids */
 653     std::vector<gmx_pme_t*> pmedata;
 654     pmedata.push_back(pme);
 655
 656     auto pme_pp = gmx_pme_pp_init(cr);
 657
 658     std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
 659     // TODO the variable below should be queried from the task assignment info
 660     const bool useGpuForPme = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
 661     if (useGpuForPme)
 662     {
 663         GMX_RELEASE_ASSERT(
 664                 deviceStreamManager != nullptr,
 665                 "Device stream manager can not be nullptr when using GPU in PME-only rank.");
 666         GMX_RELEASE_ASSERT(deviceStreamManager->streamIsValid(gmx::DeviceStreamType::Pme),
 667                            "Device stream can not be nullptr when using GPU in PME-only rank");
 668         changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
 669         changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
 670         if (useGpuPmePpCommunication)
 671         {
 672             pme_pp->pmeCoordinateReceiverGpu = std::make_unique<gmx::PmeCoordinateReceiverGpu>(
 673                     pme_pp->mpi_comm_mysim, deviceStreamManager->context(), pme_pp->ppRanks);
 674             pme_pp->pmeForceSenderGpu =
 675                     std::make_unique<gmx::PmeForceSenderGpu>(pme_gpu_get_f_ready_synchronizer(pme),
 676                                                              pme_pp->mpi_comm_mysim,
 677                                                              deviceStreamManager->context(),
 678                                                              pme_pp->ppRanks);
 679         }
 680         // TODO: Special PME-only constructor is used here. There is no mechanism to prevent from using the other constructor here.
 681         //       This should be made safer.
 682         stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
 683                 &deviceStreamManager->stream(gmx::DeviceStreamType::Pme),
 684                 deviceStreamManager->context(),
 685                 GpuApiCallBehavior::Async,
 686                 pme_gpu_get_block_size(pme),
 687                 wcycle);
 688     }
 689
 690     clear_nrnb(mynrnb);
 691
 692     count = 0;
 693     do /****** this is a quasi-loop over time steps! */
 694     {
 695         /* The reason for having a loop here is PME grid tuning/switching */
 696         do
 697         {
 698             /* Domain decomposition */
 699             ivec newGridSize;
 700             real ewaldcoeff_q = 0, ewaldcoeff_lj = 0;
 701             ret = gmx_pme_recv_coeffs_coords(pme,
 702                                              pme_pp.get(),
 703                                              &natoms,
 704                                              box,
 705                                              &maxshift_x,
 706                                              &maxshift_y,
 707                                              &lambda_q,
 708                                              &lambda_lj,
 709                                              &computeEnergyAndVirial,
 710                                              &step,
 711                                              &newGridSize,
 712                                              &ewaldcoeff_q,
 713                                              &ewaldcoeff_lj,
 714                                              useGpuForPme,
 715                                              stateGpu.get(),
 716                                              runMode);
 717
 718             if (ret == pmerecvqxSWITCHGRID)
 719             {
 720                 /* Switch the PME grid to newGridSize */
 721                 pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
 722             }
 723
 724             if (ret == pmerecvqxRESETCOUNTERS)
 725             {
 726                 /* Reset the cycle and flop counters */
 727                 reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
 728             }
 729         } while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
 730
 731         if (ret == pmerecvqxFINISH)
 732         {
 733             /* We should stop: break out of the loop */
 734             break;
 735         }
 736
 737         if (count == 0)
 738         {
 739             wallcycle_start(wcycle, WallCycleCounter::Run);
 740             walltime_accounting_start_time(walltime_accounting);
 741         }
 742
 743         wallcycle_start(wcycle, WallCycleCounter::PmeMesh);
 744
 745         dvdlambda_q  = 0;
 746         dvdlambda_lj = 0;
 747
 748         // TODO Make a struct of array refs onto these per-atom fields
 749         // of pme_pp (maybe box, energy and virial, too; and likewise
 750         // from mdatoms for the other call to gmx_pme_do), so we have
 751         // fewer lines of code and less parameter passing.
 752         gmx::StepWorkload stepWork;
 753         stepWork.computeVirial = computeEnergyAndVirial;
 754         stepWork.computeEnergy = computeEnergyAndVirial;
 755         stepWork.computeForces = true;
 756         PmeOutput output       = { {}, false, 0, { { 0 } }, 0, 0, { { 0 } }, 0 };
 757         if (useGpuForPme)
 758         {
 759             stepWork.haveDynamicBox      = false;
 760             stepWork.useGpuPmeFReduction = pme_pp->useGpuDirectComm;
 761             // TODO this should be set properly by gmx_pme_recv_coeffs_coords,
 762             // or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
 763             pme_gpu_prepare_computation(pme, box, wcycle, stepWork);
 764             if (!pme_pp->useGpuDirectComm)
 765             {
 766                 stateGpu->copyCoordinatesToGpu(gmx::ArrayRef<gmx::RVec>(pme_pp->x),
 767                                                gmx::AtomLocality::Local);
 768             }
 769             // On the separate PME rank we do not need a synchronizer as we schedule everything in a single stream
 770             // TODO: with pme on GPU the receive should make a list of synchronizers and pass it here #3157
 771             auto xReadyOnDevice = nullptr;
 772
 773             pme_gpu_launch_spread(pme,
 774                                   xReadyOnDevice,
 775                                   wcycle,
 776                                   lambda_q,
 777                                   pme_pp->useGpuDirectComm,
 778                                   pme_pp->pmeCoordinateReceiverGpu.get());
 779             pme_gpu_launch_complex_transforms(pme, wcycle, stepWork);
 780             pme_gpu_launch_gather(pme, wcycle, lambda_q);
 781             output = pme_gpu_wait_finish_task(pme, computeEnergyAndVirial, lambda_q, wcycle);
 782             pme_gpu_reinit_computation(pme, wcycle);
 783         }
 784         else
 785         {
 786             GMX_ASSERT(pme_pp->x.size() == static_cast<size_t>(natoms),
 787                        "The coordinate buffer should have size natoms");
 788
 789             gmx_pme_do(pme,
 790                        pme_pp->x,
 791                        pme_pp->f,
 792                        pme_pp->chargeA,
 793                        pme_pp->chargeB,
 794                        pme_pp->sqrt_c6A,
 795                        pme_pp->sqrt_c6B,
 796                        pme_pp->sigmaA,
 797                        pme_pp->sigmaB,
 798                        box,
 799                        cr,
 800                        maxshift_x,
 801                        maxshift_y,
 802                        mynrnb,
 803                        wcycle,
 804                        output.coulombVirial_,
 805                        output.lennardJonesVirial_,
 806                        &output.coulombEnergy_,
 807                        &output.lennardJonesEnergy_,
 808                        lambda_q,
 809                        lambda_lj,
 810                        &dvdlambda_q,
 811                        &dvdlambda_lj,
 812                        stepWork);
 813             output.forces_ = pme_pp->f;
 814         }
 815
 816         cycles = wallcycle_stop(wcycle, WallCycleCounter::PmeMesh);
 817         gmx_pme_send_force_vir_ener(*pme, pme_pp.get(), output, dvdlambda_q, dvdlambda_lj, cycles);
 818
 819         count++;
 820     } /***** end of quasi-loop, we stop with the break above */
 821     while (TRUE);
 822
 823     walltime_accounting_end_time(walltime_accounting);
 824
 825     return 0;
 826 }