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, 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 "config.h"
  64
  65 #include <cassert>
  66 #include <cmath>
  67 #include <cstdio>
  68 #include <cstdlib>
  69 #include <cstring>
  70
  71 #include <memory>
  72 #include <numeric>
  73 #include <vector>
  74
  75 #include "gromacs/domdec/domdec.h"
  76 #include "gromacs/ewald/pme.h"
  77 #include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
  78 #include "gromacs/ewald/pme_force_sender_gpu.h"
  79 #include "gromacs/fft/parallel_3dfft.h"
  80 #include "gromacs/fileio/pdbio.h"
  81 #include "gromacs/gmxlib/network.h"
  82 #include "gromacs/gmxlib/nrnb.h"
  83 #include "gromacs/gpu_utils/hostallocator.h"
  84 #include "gromacs/math/gmxcomplex.h"
  85 #include "gromacs/math/units.h"
  86 #include "gromacs/math/vec.h"
  87 #include "gromacs/mdtypes/commrec.h"
  88 #include "gromacs/mdtypes/forceoutput.h"
  89 #include "gromacs/mdtypes/inputrec.h"
  90 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  91 #include "gromacs/timing/cyclecounter.h"
  92 #include "gromacs/timing/wallcycle.h"
  93 #include "gromacs/utility/fatalerror.h"
  94 #include "gromacs/utility/futil.h"
  95 #include "gromacs/utility/gmxmpi.h"
  96 #include "gromacs/utility/gmxomp.h"
  97 #include "gromacs/utility/smalloc.h"
  98
  99 #include "pme_gpu_internal.h"
 100 #include "pme_internal.h"
 101 #include "pme_pp_communication.h"
 102
 103 /*! \brief environment variable to enable GPU P2P communication */
 104 static const bool c_enableGpuPmePpComms =
 105         (getenv("GMX_GPU_PME_PP_COMMS") != nullptr) && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA);
 106
 107 /*! \brief Master PP-PME communication data structure */
 108 struct gmx_pme_pp
 109 {
 110     MPI_Comm             mpi_comm_mysim; /**< MPI communicator for this simulation */
 111     std::vector<PpRanks> ppRanks;        /**< The PP partner ranks                 */
 112     int                  peerRankId;     /**< The peer PP rank id                  */
 113     //@{
 114     /**< Vectors of A- and B-state parameters used to transfer vectors to PME ranks  */
 115     gmx::PaddedHostVector<real> chargeA;
 116     std::vector<real>           chargeB;
 117     std::vector<real>           sqrt_c6A;
 118     std::vector<real>           sqrt_c6B;
 119     std::vector<real>           sigmaA;
 120     std::vector<real>           sigmaB;
 121     //@}
 122     gmx::HostVector<gmx::RVec> x; /**< Vector of atom coordinates to transfer to PME ranks */
 123     std::vector<gmx::RVec>     f; /**< Vector of atom forces received from PME ranks */
 124     //@{
 125     /**< Vectors of MPI objects used in non-blocking communication between multiple PP ranks per PME rank */
 126     std::vector<MPI_Request> req;
 127     std::vector<MPI_Status>  stat;
 128     //@}
 129
 130     /*! \brief object for receiving coordinates using communications operating on GPU memory space */
 131     std::unique_ptr<gmx::PmeCoordinateReceiverGpu> pmeCoordinateReceiverGpu;
 132     /*! \brief object for sending PME force using communications operating on GPU memory space */
 133     std::unique_ptr<gmx::PmeForceSenderGpu> pmeForceSenderGpu;
 134
 135     /*! \brief whether GPU direct communications are active for PME-PP transfers */
 136     bool useGpuDirectComm = false;
 137 };
 138
 139 /*! \brief Initialize the PME-only side of the PME <-> PP communication */
 140 static std::unique_ptr<gmx_pme_pp> gmx_pme_pp_init(const t_commrec* cr)
 141 {
 142     auto pme_pp = std::make_unique<gmx_pme_pp>();
 143
 144 #if GMX_MPI
 145     int rank;
 146
 147     pme_pp->mpi_comm_mysim = cr->mpi_comm_mysim;
 148     MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
 149     auto ppRanks = get_pme_ddranks(cr, rank);
 150     pme_pp->ppRanks.reserve(ppRanks.size());
 151     for (const auto& ppRankId : ppRanks)
 152     {
 153         pme_pp->ppRanks.push_back({ ppRankId, 0 });
 154     }
 155     // The peer PP rank is the last one.
 156     pme_pp->peerRankId = pme_pp->ppRanks.back().rankId;
 157     pme_pp->req.resize(eCommType_NR * pme_pp->ppRanks.size());
 158     pme_pp->stat.resize(eCommType_NR * pme_pp->ppRanks.size());
 159 #else
 160     GMX_UNUSED_VALUE(cr);
 161 #endif
 162
 163     return pme_pp;
 164 }
 165
 166 static void reset_pmeonly_counters(gmx_wallcycle_t           wcycle,
 167                                    gmx_walltime_accounting_t walltime_accounting,
 168                                    t_nrnb*                   nrnb,
 169                                    int64_t                   step,
 170                                    bool                      useGpuForPme)
 171 {
 172     /* Reset all the counters related to performance over the run */
 173     wallcycle_stop(wcycle, ewcRUN);
 174     wallcycle_reset_all(wcycle);
 175     *nrnb = { 0 };
 176     wallcycle_start(wcycle, ewcRUN);
 177     walltime_accounting_reset_time(walltime_accounting, step);
 178
 179     if (useGpuForPme)
 180     {
 181         resetGpuProfiler();
 182     }
 183 }
 184
 185 static gmx_pme_t* gmx_pmeonly_switch(std::vector<gmx_pme_t*>* pmedata,
 186                                      const ivec               grid_size,
 187                                      real                     ewaldcoeff_q,
 188                                      real                     ewaldcoeff_lj,
 189                                      const t_commrec*         cr,
 190                                      const t_inputrec*        ir)
 191 {
 192     GMX_ASSERT(pmedata, "Bad PME tuning list pointer");
 193     for (auto& pme : *pmedata)
 194     {
 195         GMX_ASSERT(pme, "Bad PME tuning list element pointer");
 196         if (pme->nkx == grid_size[XX] && pme->nky == grid_size[YY] && pme->nkz == grid_size[ZZ])
 197         {
 198             /* Here we have found an existing PME data structure that suits us.
 199              * However, in the GPU case, we have to reinitialize it - there's only one GPU structure.
 200              * This should not cause actual GPU reallocations, at least (the allocated buffers are never shrunk).
 201              * So, just some grid size updates in the GPU kernel parameters.
 202              * TODO: this should be something like gmx_pme_update_split_params()
 203              */
 204             gmx_pme_reinit(&pme, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 205             return pme;
 206         }
 207     }
 208
 209     const auto& pme          = pmedata->back();
 210     gmx_pme_t*  newStructure = nullptr;
 211     // Copy last structure with new grid params
 212     gmx_pme_reinit(&newStructure, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 213     pmedata->push_back(newStructure);
 214     return newStructure;
 215 }
 216
 217 /*! \brief Called by PME-only ranks to receive coefficients and coordinates
 218  *
 219  * \param[in] pme           PME data structure.
 220  * \param[in,out] pme_pp    PME-PP communication structure.
 221  * \param[out] natoms       Number of received atoms.
 222  * \param[out] box        System box, if received.
 223  * \param[out] maxshift_x        Maximum shift in X direction, if received.
 224  * \param[out] maxshift_y        Maximum shift in Y direction, if received.
 225  * \param[out] lambda_q         Free-energy lambda for electrostatics, if received.
 226  * \param[out] lambda_lj         Free-energy lambda for Lennard-Jones, if received.
 227  * \param[out] bEnerVir          Set to true if this is an energy/virial calculation step, otherwise
 228  * set to false. \param[out] step              MD integration step number. \param[out] grid_size PME
 229  * grid size, if received. \param[out] ewaldcoeff_q         Ewald cut-off parameter for
 230  * electrostatics, if received. \param[out] ewaldcoeff_lj         Ewald cut-off parameter for
 231  * Lennard-Jones, if received. \param[in] useGpuForPme      flag on whether PME is on GPU \param[in]
 232  * stateGpu          GPU state propagator object \param[in] runMode           PME run mode
 233  *
 234  * \retval pmerecvqxX             All parameters were set, chargeA and chargeB can be NULL.
 235  * \retval pmerecvqxFINISH        No parameters were set.
 236  * \retval pmerecvqxSWITCHGRID    Only grid_size and *ewaldcoeff were set.
 237  * \retval pmerecvqxRESETCOUNTERS *step was set.
 238  */
 239 static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t*            pme,
 240                                       gmx_pme_pp*                  pme_pp,
 241                                       int*                         natoms,
 242                                       matrix                       box,
 243                                       int*                         maxshift_x,
 244                                       int*                         maxshift_y,
 245                                       real*                        lambda_q,
 246                                       real*                        lambda_lj,
 247                                       gmx_bool*                    bEnerVir,
 248                                       int64_t*                     step,
 249                                       ivec*                        grid_size,
 250                                       real*                        ewaldcoeff_q,
 251                                       real*                        ewaldcoeff_lj,
 252                                       bool                         useGpuForPme,
 253                                       gmx::StatePropagatorDataGpu* stateGpu,
 254                                       PmeRunMode gmx_unused runMode)
 255 {
 256     int status = -1;
 257     int nat    = 0;
 258
 259 #if GMX_MPI
 260     unsigned int flags          = 0;
 261     int          messages       = 0;
 262     bool         atomSetChanged = false;
 263
 264     do
 265     {
 266         gmx_pme_comm_n_box_t cnb;
 267         cnb.flags = 0;
 268
 269         /* Receive the send count, box and time step from the peer PP node */
 270         MPI_Recv(&cnb, sizeof(cnb), MPI_BYTE, pme_pp->peerRankId, eCommType_CNB,
 271                  pme_pp->mpi_comm_mysim, MPI_STATUS_IGNORE);
 272
 273         /* We accumulate all received flags */
 274         flags |= cnb.flags;
 275
 276         *step = cnb.step;
 277
 278         if (debug)
 279         {
 280             fprintf(debug, "PME only rank receiving:%s%s%s%s%s\n",
 281                     (cnb.flags & PP_PME_CHARGE) ? " charges" : "",
 282                     (cnb.flags & PP_PME_COORD) ? " coordinates" : "",
 283                     (cnb.flags & PP_PME_FINISH) ? " finish" : "",
 284                     (cnb.flags & PP_PME_SWITCHGRID) ? " switch grid" : "",
 285                     (cnb.flags & PP_PME_RESETCOUNTERS) ? " reset counters" : "");
 286         }
 287
 288         pme_pp->useGpuDirectComm = ((cnb.flags & PP_PME_GPUCOMMS) != 0);
 289         GMX_ASSERT(!pme_pp->useGpuDirectComm || (pme_pp->pmeForceSenderGpu != nullptr),
 290                    "The use of GPU direct communication for PME-PP is enabled, "
 291                    "but the PME GPU force reciever object does not exist");
 292
 293         if (cnb.flags & PP_PME_FINISH)
 294         {
 295             status = pmerecvqxFINISH;
 296         }
 297
 298         if (cnb.flags & PP_PME_SWITCHGRID)
 299         {
 300             /* Special case, receive the new parameters and return */
 301             copy_ivec(cnb.grid_size, *grid_size);
 302             *ewaldcoeff_q  = cnb.ewaldcoeff_q;
 303             *ewaldcoeff_lj = cnb.ewaldcoeff_lj;
 304
 305             status = pmerecvqxSWITCHGRID;
 306         }
 307
 308         if (cnb.flags & PP_PME_RESETCOUNTERS)
 309         {
 310             /* Special case, receive the step (set above) and return */
 311             status = pmerecvqxRESETCOUNTERS;
 312         }
 313
 314         if (cnb.flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
 315         {
 316             atomSetChanged = true;
 317
 318             /* Receive the send counts from the other PP nodes */
 319             for (auto& sender : pme_pp->ppRanks)
 320             {
 321                 if (sender.rankId == pme_pp->peerRankId)
 322                 {
 323                     sender.numAtoms = cnb.natoms;
 324                 }
 325                 else
 326                 {
 327                     MPI_Irecv(&sender.numAtoms, sizeof(sender.numAtoms), MPI_BYTE, sender.rankId,
 328                               eCommType_CNB, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 329                 }
 330             }
 331             MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 332             messages = 0;
 333
 334             nat = 0;
 335             for (const auto& sender : pme_pp->ppRanks)
 336             {
 337                 nat += sender.numAtoms;
 338             }
 339
 340             if (cnb.flags & PP_PME_CHARGE)
 341             {
 342                 pme_pp->chargeA.resizeWithPadding(nat);
 343             }
 344             if (cnb.flags & PP_PME_CHARGEB)
 345             {
 346                 pme_pp->chargeB.resize(nat);
 347             }
 348             if (cnb.flags & PP_PME_SQRTC6)
 349             {
 350                 pme_pp->sqrt_c6A.resize(nat);
 351             }
 352             if (cnb.flags & PP_PME_SQRTC6B)
 353             {
 354                 pme_pp->sqrt_c6B.resize(nat);
 355             }
 356             if (cnb.flags & PP_PME_SIGMA)
 357             {
 358                 pme_pp->sigmaA.resize(nat);
 359             }
 360             if (cnb.flags & PP_PME_SIGMAB)
 361             {
 362                 pme_pp->sigmaB.resize(nat);
 363             }
 364             pme_pp->x.resize(nat);
 365             pme_pp->f.resize(nat);
 366
 367             /* maxshift is sent when the charges are sent */
 368             *maxshift_x = cnb.maxshift_x;
 369             *maxshift_y = cnb.maxshift_y;
 370
 371             /* Receive the charges in place */
 372             for (int q = 0; q < eCommType_NR; q++)
 373             {
 374                 real* bufferPtr;
 375
 376                 if (!(cnb.flags & (PP_PME_CHARGE << q)))
 377                 {
 378                     continue;
 379                 }
 380                 switch (q)
 381                 {
 382                     case eCommType_ChargeA: bufferPtr = pme_pp->chargeA.data(); break;
 383                     case eCommType_ChargeB: bufferPtr = pme_pp->chargeB.data(); break;
 384                     case eCommType_SQRTC6A: bufferPtr = pme_pp->sqrt_c6A.data(); break;
 385                     case eCommType_SQRTC6B: bufferPtr = pme_pp->sqrt_c6B.data(); break;
 386                     case eCommType_SigmaA: bufferPtr = pme_pp->sigmaA.data(); break;
 387                     case eCommType_SigmaB: bufferPtr = pme_pp->sigmaB.data(); break;
 388                     default: gmx_incons("Wrong eCommType");
 389                 }
 390                 nat = 0;
 391                 for (const auto& sender : pme_pp->ppRanks)
 392                 {
 393                     if (sender.numAtoms > 0)
 394                     {
 395                         MPI_Irecv(bufferPtr + nat, sender.numAtoms * sizeof(real), MPI_BYTE,
 396                                   sender.rankId, q, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 397                         nat += sender.numAtoms;
 398                         if (debug)
 399                         {
 400                             fprintf(debug, "Received from PP rank %d: %d %s\n", sender.rankId,
 401                                     sender.numAtoms,
 402                                     (q == eCommType_ChargeA || q == eCommType_ChargeB) ? "charges"
 403                                                                                        : "params");
 404                         }
 405                     }
 406                 }
 407             }
 408         }
 409
 410         if (cnb.flags & PP_PME_COORD)
 411         {
 412             if (atomSetChanged)
 413             {
 414                 gmx_pme_reinit_atoms(pme, nat, pme_pp->chargeA.data());
 415                 if (useGpuForPme)
 416                 {
 417                     stateGpu->reinit(nat, nat);
 418                     pme_gpu_set_device_x(pme, stateGpu->getCoordinates());
 419                 }
 420                 if (pme_pp->useGpuDirectComm)
 421                 {
 422                     GMX_ASSERT(runMode == PmeRunMode::GPU,
 423                                "GPU Direct PME-PP communication has been enabled, "
 424                                "but PME run mode is not PmeRunMode::GPU\n");
 425
 426                     // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses.
 427                     pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(
 428                             pme_gpu_get_device_x(pme));
 429                     pme_pp->pmeForceSenderGpu->sendForceBufferAddressToPpRanks(
 430                             reinterpret_cast<rvec*>(pme_gpu_get_device_f(pme)));
 431                 }
 432             }
 433
 434
 435             /* The box, FE flag and lambda are sent along with the coordinates
 436              *  */
 437             copy_mat(cnb.box, box);
 438             *lambda_q  = cnb.lambda_q;
 439             *lambda_lj = cnb.lambda_lj;
 440             *bEnerVir  = ((cnb.flags & PP_PME_ENER_VIR) != 0U);
 441             *step      = cnb.step;
 442
 443             /* Receive the coordinates in place */
 444             nat = 0;
 445             for (const auto& sender : pme_pp->ppRanks)
 446             {
 447                 if (sender.numAtoms > 0)
 448                 {
 449                     if (pme_pp->useGpuDirectComm)
 450                     {
 451                         pme_pp->pmeCoordinateReceiverGpu->launchReceiveCoordinatesFromPpCudaDirect(
 452                                 sender.rankId);
 453                     }
 454                     else
 455                     {
 456                         MPI_Irecv(pme_pp->x[nat], sender.numAtoms * sizeof(rvec), MPI_BYTE, sender.rankId,
 457                                   eCommType_COORD, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 458                     }
 459                     nat += sender.numAtoms;
 460                     if (debug)
 461                     {
 462                         fprintf(debug,
 463                                 "Received from PP rank %d: %d "
 464                                 "coordinates\n",
 465                                 sender.rankId, sender.numAtoms);
 466                     }
 467                 }
 468             }
 469
 470             if (pme_pp->useGpuDirectComm)
 471             {
 472                 pme_pp->pmeCoordinateReceiverGpu->enqueueWaitReceiveCoordinatesFromPpCudaDirect();
 473             }
 474
 475             status = pmerecvqxX;
 476         }
 477
 478         /* Wait for the coordinates and/or charges to arrive */
 479         MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 480         messages = 0;
 481     } while (status == -1);
 482 #else
 483     GMX_UNUSED_VALUE(pme);
 484     GMX_UNUSED_VALUE(pme_pp);
 485     GMX_UNUSED_VALUE(box);
 486     GMX_UNUSED_VALUE(maxshift_x);
 487     GMX_UNUSED_VALUE(maxshift_y);
 488     GMX_UNUSED_VALUE(lambda_q);
 489     GMX_UNUSED_VALUE(lambda_lj);
 490     GMX_UNUSED_VALUE(bEnerVir);
 491     GMX_UNUSED_VALUE(step);
 492     GMX_UNUSED_VALUE(grid_size);
 493     GMX_UNUSED_VALUE(ewaldcoeff_q);
 494     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 495     GMX_UNUSED_VALUE(useGpuForPme);
 496     GMX_UNUSED_VALUE(stateGpu);
 497
 498     status = pmerecvqxX;
 499 #endif
 500
 501     if (status == pmerecvqxX)
 502     {
 503         *natoms = nat;
 504     }
 505
 506     return status;
 507 }
 508
 509 #if GMX_MPI
 510 /*! \brief Send force data to PP ranks */
 511 static void sendFToPP(void* sendbuf, PpRanks receiver, gmx_pme_pp* pme_pp, int* messages)
 512 {
 513
 514     if (pme_pp->useGpuDirectComm)
 515     {
 516         GMX_ASSERT((pme_pp->pmeForceSenderGpu != nullptr),
 517                    "The use of GPU direct communication for PME-PP is enabled, "
 518                    "but the PME GPU force reciever object does not exist");
 519
 520         pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(receiver.rankId);
 521     }
 522     else
 523     {
 524         // Send using MPI
 525         MPI_Isend(sendbuf, receiver.numAtoms * sizeof(rvec), MPI_BYTE, receiver.rankId, 0,
 526                   pme_pp->mpi_comm_mysim, &pme_pp->req[*messages]);
 527         *messages = *messages + 1;
 528     }
 529 }
 530 #endif
 531
 532 /*! \brief Send the PME mesh force, virial and energy to the PP-only ranks. */
 533 static void gmx_pme_send_force_vir_ener(const gmx_pme_t& pme,
 534                                         gmx_pme_pp*      pme_pp,
 535                                         const PmeOutput& output,
 536                                         real             dvdlambda_q,
 537                                         real             dvdlambda_lj,
 538                                         float            cycles)
 539 {
 540 #if GMX_MPI
 541     gmx_pme_comm_vir_ene_t cve;
 542     int                    messages, ind_start, ind_end;
 543     cve.cycles = cycles;
 544
 545     /* Now the evaluated forces have to be transferred to the PP nodes */
 546     messages = 0;
 547     ind_end  = 0;
 548     for (const auto& receiver : pme_pp->ppRanks)
 549     {
 550         ind_start     = ind_end;
 551         ind_end       = ind_start + receiver.numAtoms;
 552         void* sendbuf = const_cast<void*>(static_cast<const void*>(output.forces_[ind_start]));
 553         if (pme_pp->useGpuDirectComm)
 554         {
 555             // Data will be transferred directly from GPU.
 556             rvec* d_f = reinterpret_cast<rvec*>(pme_gpu_get_device_f(&pme));
 557             sendbuf   = reinterpret_cast<void*>(&d_f[ind_start]);
 558         }
 559         sendFToPP(sendbuf, receiver, pme_pp, &messages);
 560     }
 561
 562     /* send virial and energy to our last PP node */
 563     copy_mat(output.coulombVirial_, cve.vir_q);
 564     copy_mat(output.lennardJonesVirial_, cve.vir_lj);
 565     cve.energy_q     = output.coulombEnergy_;
 566     cve.energy_lj    = output.lennardJonesEnergy_;
 567     cve.dvdlambda_q  = dvdlambda_q;
 568     cve.dvdlambda_lj = dvdlambda_lj;
 569     /* check for the signals to send back to a PP node */
 570     cve.stop_cond = gmx_get_stop_condition();
 571
 572     cve.cycles = cycles;
 573
 574     if (debug)
 575     {
 576         fprintf(debug, "PME rank sending to PP rank %d: virial and energy\n", pme_pp->peerRankId);
 577     }
 578     MPI_Isend(&cve, sizeof(cve), MPI_BYTE, pme_pp->peerRankId, 1, pme_pp->mpi_comm_mysim,
 579               &pme_pp->req[messages++]);
 580
 581     /* Wait for the forces to arrive */
 582     MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 583 #else
 584     gmx_call("MPI not enabled");
 585     GMX_UNUSED_VALUE(pme);
 586     GMX_UNUSED_VALUE(pme_pp);
 587     GMX_UNUSED_VALUE(output);
 588     GMX_UNUSED_VALUE(dvdlambda_q);
 589     GMX_UNUSED_VALUE(dvdlambda_lj);
 590     GMX_UNUSED_VALUE(cycles);
 591 #endif
 592 }
 593
 594 int gmx_pmeonly(struct gmx_pme_t*         pme,
 595                 const t_commrec*          cr,
 596                 t_nrnb*                   mynrnb,
 597                 gmx_wallcycle*            wcycle,
 598                 gmx_walltime_accounting_t walltime_accounting,
 599                 t_inputrec*               ir,
 600                 PmeRunMode                runMode)
 601 {
 602     int      ret;
 603     int      natoms = 0;
 604     matrix   box;
 605     real     lambda_q   = 0;
 606     real     lambda_lj  = 0;
 607     int      maxshift_x = 0, maxshift_y = 0;
 608     real     dvdlambda_q, dvdlambda_lj;
 609     float    cycles;
 610     int      count;
 611     gmx_bool bEnerVir = FALSE;
 612     int64_t  step;
 613
 614     /* This data will only use with PME tuning, i.e. switching PME grids */
 615     std::vector<gmx_pme_t*> pmedata;
 616     pmedata.push_back(pme);
 617
 618     auto pme_pp = gmx_pme_pp_init(cr);
 619     // TODO the variable below should be queried from the task assignment info
 620     const bool  useGpuForPme  = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
 621     const void* commandStream = useGpuForPme ? pme_gpu_get_device_stream(pme) : nullptr;
 622     const void* deviceContext = useGpuForPme ? pme_gpu_get_device_context(pme) : nullptr;
 623     const int   paddingSize   = pme_gpu_get_padding_size(pme);
 624     if (useGpuForPme)
 625     {
 626         changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
 627         changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
 628         if (c_enableGpuPmePpComms)
 629         {
 630             pme_pp->pmeCoordinateReceiverGpu = std::make_unique<gmx::PmeCoordinateReceiverGpu>(
 631                     pme_gpu_get_device_stream(pme), pme_pp->mpi_comm_mysim, pme_pp->ppRanks);
 632             pme_pp->pmeForceSenderGpu = std::make_unique<gmx::PmeForceSenderGpu>(
 633                     pme_gpu_get_device_stream(pme), pme_pp->mpi_comm_mysim, pme_pp->ppRanks);
 634         }
 635     }
 636
 637     std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
 638     if (useGpuForPme)
 639     {
 640         // TODO: Special PME-only constructor is used here. There is no mechanism to prevent from using the other constructor here.
 641         //       This should be made safer.
 642         stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
 643                 commandStream, deviceContext, GpuApiCallBehavior::Async, paddingSize, wcycle);
 644     }
 645
 646
 647     clear_nrnb(mynrnb);
 648
 649     count = 0;
 650     do /****** this is a quasi-loop over time steps! */
 651     {
 652         /* The reason for having a loop here is PME grid tuning/switching */
 653         do
 654         {
 655             /* Domain decomposition */
 656             ivec newGridSize;
 657             real ewaldcoeff_q = 0, ewaldcoeff_lj = 0;
 658             ret = gmx_pme_recv_coeffs_coords(pme, pme_pp.get(), &natoms, box, &maxshift_x,
 659                                              &maxshift_y, &lambda_q, &lambda_lj, &bEnerVir, &step,
 660                                              &newGridSize, &ewaldcoeff_q, &ewaldcoeff_lj,
 661                                              useGpuForPme, stateGpu.get(), runMode);
 662
 663             if (ret == pmerecvqxSWITCHGRID)
 664             {
 665                 /* Switch the PME grid to newGridSize */
 666                 pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
 667             }
 668
 669             if (ret == pmerecvqxRESETCOUNTERS)
 670             {
 671                 /* Reset the cycle and flop counters */
 672                 reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
 673             }
 674         } while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
 675
 676         if (ret == pmerecvqxFINISH)
 677         {
 678             /* We should stop: break out of the loop */
 679             break;
 680         }
 681
 682         if (count == 0)
 683         {
 684             wallcycle_start(wcycle, ewcRUN);
 685             walltime_accounting_start_time(walltime_accounting);
 686         }
 687
 688         wallcycle_start(wcycle, ewcPMEMESH);
 689
 690         dvdlambda_q  = 0;
 691         dvdlambda_lj = 0;
 692
 693         // TODO Make a struct of array refs onto these per-atom fields
 694         // of pme_pp (maybe box, energy and virial, too; and likewise
 695         // from mdatoms for the other call to gmx_pme_do), so we have
 696         // fewer lines of code and less parameter passing.
 697         const int pmeFlags = GMX_PME_DO_ALL_F | (bEnerVir ? GMX_PME_CALC_ENER_VIR : 0);
 698         PmeOutput output;
 699         if (useGpuForPme)
 700         {
 701             const bool boxChanged              = false;
 702             const bool useGpuPmeForceReduction = pme_pp->useGpuDirectComm;
 703             // TODO this should be set properly by gmx_pme_recv_coeffs_coords,
 704             // or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
 705             pme_gpu_prepare_computation(pme, boxChanged, box, wcycle, pmeFlags, useGpuPmeForceReduction);
 706             if (!pme_pp->useGpuDirectComm)
 707             {
 708                 stateGpu->copyCoordinatesToGpu(gmx::ArrayRef<gmx::RVec>(pme_pp->x), gmx::AtomLocality::All);
 709             }
 710             // On the separate PME rank we do not need a synchronizer as we schedule everything in a single stream
 711             // TODO: with pme on GPU the receive should make a list of synchronizers and pass it here #3157
 712             auto xReadyOnDevice = nullptr;
 713
 714             pme_gpu_launch_spread(pme, xReadyOnDevice, wcycle);
 715             pme_gpu_launch_complex_transforms(pme, wcycle);
 716             pme_gpu_launch_gather(pme, wcycle, PmeForceOutputHandling::Set);
 717             output = pme_gpu_wait_finish_task(pme, pmeFlags, wcycle);
 718             pme_gpu_reinit_computation(pme, wcycle);
 719         }
 720         else
 721         {
 722             GMX_ASSERT(pme_pp->x.size() == static_cast<size_t>(natoms),
 723                        "The coordinate buffer should have size natoms");
 724
 725             gmx_pme_do(pme, pme_pp->x, pme_pp->f, pme_pp->chargeA.data(), pme_pp->chargeB.data(),
 726                        pme_pp->sqrt_c6A.data(), pme_pp->sqrt_c6B.data(), pme_pp->sigmaA.data(),
 727                        pme_pp->sigmaB.data(), box, cr, maxshift_x, maxshift_y, mynrnb, wcycle,
 728                        output.coulombVirial_, output.lennardJonesVirial_, &output.coulombEnergy_,
 729                        &output.lennardJonesEnergy_, lambda_q, lambda_lj, &dvdlambda_q,
 730                        &dvdlambda_lj, pmeFlags);
 731             output.forces_ = pme_pp->f;
 732         }
 733
 734         cycles = wallcycle_stop(wcycle, ewcPMEMESH);
 735         gmx_pme_send_force_vir_ener(*pme, pme_pp.get(), output, dvdlambda_q, dvdlambda_lj, cycles);
 736
 737         count++;
 738     } /***** end of quasi-loop, we stop with the break above */
 739     while (TRUE);
 740
 741     walltime_accounting_end_time(walltime_accounting);
 742
 743     return 0;
 744 }