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 "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_output.h"
 105 #include "pme_pp_communication.h"
 106
 107 /*! \brief environment variable to enable GPU P2P communication */
 108 static const bool c_enableGpuPmePpComms =
 109         (getenv("GMX_GPU_PME_PP_COMMS") != nullptr) && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA);
 110
 111 /*! \brief Master PP-PME communication data structure */
 112 struct gmx_pme_pp
 113 {
 114     MPI_Comm             mpi_comm_mysim; /**< MPI communicator for this simulation */
 115     std::vector<PpRanks> ppRanks;        /**< The PP partner ranks                 */
 116     int                  peerRankId;     /**< The peer PP rank id                  */
 117     //@{
 118     /**< Vectors of A- and B-state parameters used to transfer vectors to PME ranks  */
 119     gmx::PaddedHostVector<real> chargeA;
 120     std::vector<real>           chargeB;
 121     std::vector<real>           sqrt_c6A;
 122     std::vector<real>           sqrt_c6B;
 123     std::vector<real>           sigmaA;
 124     std::vector<real>           sigmaB;
 125     //@}
 126     gmx::HostVector<gmx::RVec> x; /**< Vector of atom coordinates to transfer to PME ranks */
 127     std::vector<gmx::RVec>     f; /**< Vector of atom forces received from PME ranks */
 128     //@{
 129     /**< Vectors of MPI objects used in non-blocking communication between multiple PP ranks per PME rank */
 130     std::vector<MPI_Request> req;
 131     std::vector<MPI_Status>  stat;
 132     //@}
 133
 134     /*! \brief object for receiving coordinates using communications operating on GPU memory space */
 135     std::unique_ptr<gmx::PmeCoordinateReceiverGpu> pmeCoordinateReceiverGpu;
 136     /*! \brief object for sending PME force using communications operating on GPU memory space */
 137     std::unique_ptr<gmx::PmeForceSenderGpu> pmeForceSenderGpu;
 138
 139     /*! \brief whether GPU direct communications are active for PME-PP transfers */
 140     bool useGpuDirectComm = false;
 141 };
 142
 143 /*! \brief Initialize the PME-only side of the PME <-> PP communication */
 144 static std::unique_ptr<gmx_pme_pp> gmx_pme_pp_init(const t_commrec* cr)
 145 {
 146     auto pme_pp = std::make_unique<gmx_pme_pp>();
 147
 148 #if GMX_MPI
 149     int rank;
 150
 151     pme_pp->mpi_comm_mysim = cr->mpi_comm_mysim;
 152     MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
 153     auto ppRanks = get_pme_ddranks(cr, rank);
 154     pme_pp->ppRanks.reserve(ppRanks.size());
 155     for (const auto& ppRankId : ppRanks)
 156     {
 157         pme_pp->ppRanks.push_back({ ppRankId, 0 });
 158     }
 159     // The peer PP rank is the last one.
 160     pme_pp->peerRankId = pme_pp->ppRanks.back().rankId;
 161     pme_pp->req.resize(eCommType_NR * pme_pp->ppRanks.size());
 162     pme_pp->stat.resize(eCommType_NR * pme_pp->ppRanks.size());
 163 #else
 164     GMX_UNUSED_VALUE(cr);
 165 #endif
 166
 167     return pme_pp;
 168 }
 169
 170 static void reset_pmeonly_counters(gmx_wallcycle_t           wcycle,
 171                                    gmx_walltime_accounting_t walltime_accounting,
 172                                    t_nrnb*                   nrnb,
 173                                    int64_t                   step,
 174                                    bool                      useGpuForPme)
 175 {
 176     /* Reset all the counters related to performance over the run */
 177     wallcycle_stop(wcycle, ewcRUN);
 178     wallcycle_reset_all(wcycle);
 179     *nrnb = { 0 };
 180     wallcycle_start(wcycle, ewcRUN);
 181     walltime_accounting_reset_time(walltime_accounting, step);
 182
 183     if (useGpuForPme)
 184     {
 185         resetGpuProfiler();
 186     }
 187 }
 188
 189 static gmx_pme_t* gmx_pmeonly_switch(std::vector<gmx_pme_t*>* pmedata,
 190                                      const ivec               grid_size,
 191                                      real                     ewaldcoeff_q,
 192                                      real                     ewaldcoeff_lj,
 193                                      const t_commrec*         cr,
 194                                      const t_inputrec*        ir)
 195 {
 196     GMX_ASSERT(pmedata, "Bad PME tuning list pointer");
 197     for (auto& pme : *pmedata)
 198     {
 199         GMX_ASSERT(pme, "Bad PME tuning list element pointer");
 200         if (gmx_pme_grid_matches(*pme, grid_size))
 201         {
 202             /* Here we have found an existing PME data structure that suits us.
 203              * However, in the GPU case, we have to reinitialize it - there's only one GPU structure.
 204              * This should not cause actual GPU reallocations, at least (the allocated buffers are never shrunk).
 205              * So, just some grid size updates in the GPU kernel parameters.
 206              * TODO: this should be something like gmx_pme_update_split_params()
 207              */
 208             gmx_pme_reinit(&pme, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 209             return pme;
 210         }
 211     }
 212
 213     const auto& pme          = pmedata->back();
 214     gmx_pme_t*  newStructure = nullptr;
 215     // Copy last structure with new grid params
 216     gmx_pme_reinit(&newStructure, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 217     pmedata->push_back(newStructure);
 218     return newStructure;
 219 }
 220
 221 /*! \brief Called by PME-only ranks to receive coefficients and coordinates
 222  *
 223  * \param[in] pme                     PME data structure.
 224  * \param[in,out] pme_pp              PME-PP communication structure.
 225  * \param[out] natoms                 Number of received atoms.
 226  * \param[out] box                    System box, if received.
 227  * \param[out] maxshift_x             Maximum shift in X direction, if received.
 228  * \param[out] maxshift_y             Maximum shift in Y direction, if received.
 229  * \param[out] lambda_q               Free-energy lambda for electrostatics, if received.
 230  * \param[out] lambda_lj              Free-energy lambda for Lennard-Jones, if received.
 231  * \param[out] computeEnergyAndVirial Set to true if this is an energy/virial calculation
 232  *                                    step, otherwise set to false.
 233  * \param[out] step                   MD integration step number.
 234  * \param[out] grid_size              PME grid size, if received.
 235  * \param[out] ewaldcoeff_q           Ewald cut-off parameter for electrostatics, if received.
 236  * \param[out] ewaldcoeff_lj          Ewald cut-off parameter for Lennard-Jones, if received.
 237  * \param[in]  useGpuForPme           Flag on whether PME is on GPU.
 238  * \param[in]  stateGpu               GPU state propagator object.
 239  * \param[in]  runMode                PME run mode.
 240  *
 241  * \retval pmerecvqxX                 All parameters were set, chargeA and chargeB can be NULL.
 242  * \retval pmerecvqxFINISH            No parameters were set.
 243  * \retval pmerecvqxSWITCHGRID        Only grid_size and *ewaldcoeff were set.
 244  * \retval pmerecvqxRESETCOUNTERS     *step was set.
 245  */
 246 static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t*            pme,
 247                                       gmx_pme_pp*                  pme_pp,
 248                                       int*                         natoms,
 249                                       matrix                       box,
 250                                       int*                         maxshift_x,
 251                                       int*                         maxshift_y,
 252                                       real*                        lambda_q,
 253                                       real*                        lambda_lj,
 254                                       gmx_bool*                    computeEnergyAndVirial,
 255                                       int64_t*                     step,
 256                                       ivec*                        grid_size,
 257                                       real*                        ewaldcoeff_q,
 258                                       real*                        ewaldcoeff_lj,
 259                                       bool                         useGpuForPme,
 260                                       gmx::StatePropagatorDataGpu* stateGpu,
 261                                       PmeRunMode gmx_unused runMode)
 262 {
 263     int status = -1;
 264     int nat    = 0;
 265
 266 #if GMX_MPI
 267     unsigned int flags          = 0;
 268     int          messages       = 0;
 269     bool         atomSetChanged = false;
 270
 271     do
 272     {
 273         gmx_pme_comm_n_box_t cnb;
 274         cnb.flags = 0;
 275
 276         /* Receive the send count, box and time step from the peer PP node */
 277         MPI_Recv(&cnb, sizeof(cnb), MPI_BYTE, pme_pp->peerRankId, eCommType_CNB,
 278                  pme_pp->mpi_comm_mysim, MPI_STATUS_IGNORE);
 279
 280         /* We accumulate all received flags */
 281         flags |= cnb.flags;
 282
 283         *step = cnb.step;
 284
 285         if (debug)
 286         {
 287             fprintf(debug, "PME only rank receiving:%s%s%s%s%s\n",
 288                     (cnb.flags & PP_PME_CHARGE) ? " charges" : "",
 289                     (cnb.flags & PP_PME_COORD) ? " coordinates" : "",
 290                     (cnb.flags & PP_PME_FINISH) ? " finish" : "",
 291                     (cnb.flags & PP_PME_SWITCHGRID) ? " switch grid" : "",
 292                     (cnb.flags & PP_PME_RESETCOUNTERS) ? " reset counters" : "");
 293         }
 294
 295         pme_pp->useGpuDirectComm = ((cnb.flags & PP_PME_GPUCOMMS) != 0);
 296         GMX_ASSERT(!pme_pp->useGpuDirectComm || (pme_pp->pmeForceSenderGpu != nullptr),
 297                    "The use of GPU direct communication for PME-PP is enabled, "
 298                    "but the PME GPU force reciever object does not exist");
 299
 300         if (cnb.flags & PP_PME_FINISH)
 301         {
 302             status = pmerecvqxFINISH;
 303         }
 304
 305         if (cnb.flags & PP_PME_SWITCHGRID)
 306         {
 307             /* Special case, receive the new parameters and return */
 308             copy_ivec(cnb.grid_size, *grid_size);
 309             *ewaldcoeff_q  = cnb.ewaldcoeff_q;
 310             *ewaldcoeff_lj = cnb.ewaldcoeff_lj;
 311
 312             status = pmerecvqxSWITCHGRID;
 313         }
 314
 315         if (cnb.flags & PP_PME_RESETCOUNTERS)
 316         {
 317             /* Special case, receive the step (set above) and return */
 318             status = pmerecvqxRESETCOUNTERS;
 319         }
 320
 321         if (cnb.flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
 322         {
 323             atomSetChanged = true;
 324
 325             /* Receive the send counts from the other PP nodes */
 326             for (auto& sender : pme_pp->ppRanks)
 327             {
 328                 if (sender.rankId == pme_pp->peerRankId)
 329                 {
 330                     sender.numAtoms = cnb.natoms;
 331                 }
 332                 else
 333                 {
 334                     MPI_Irecv(&sender.numAtoms, sizeof(sender.numAtoms), MPI_BYTE, sender.rankId,
 335                               eCommType_CNB, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 336                 }
 337             }
 338             MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 339             messages = 0;
 340
 341             nat = 0;
 342             for (const auto& sender : pme_pp->ppRanks)
 343             {
 344                 nat += sender.numAtoms;
 345             }
 346
 347             if (cnb.flags & PP_PME_CHARGE)
 348             {
 349                 pme_pp->chargeA.resizeWithPadding(nat);
 350             }
 351             if (cnb.flags & PP_PME_CHARGEB)
 352             {
 353                 pme_pp->chargeB.resize(nat);
 354             }
 355             if (cnb.flags & PP_PME_SQRTC6)
 356             {
 357                 pme_pp->sqrt_c6A.resize(nat);
 358             }
 359             if (cnb.flags & PP_PME_SQRTC6B)
 360             {
 361                 pme_pp->sqrt_c6B.resize(nat);
 362             }
 363             if (cnb.flags & PP_PME_SIGMA)
 364             {
 365                 pme_pp->sigmaA.resize(nat);
 366             }
 367             if (cnb.flags & PP_PME_SIGMAB)
 368             {
 369                 pme_pp->sigmaB.resize(nat);
 370             }
 371             pme_pp->x.resize(nat);
 372             pme_pp->f.resize(nat);
 373
 374             /* maxshift is sent when the charges are sent */
 375             *maxshift_x = cnb.maxshift_x;
 376             *maxshift_y = cnb.maxshift_y;
 377
 378             /* Receive the charges in place */
 379             for (int q = 0; q < eCommType_NR; q++)
 380             {
 381                 real* bufferPtr;
 382
 383                 if (!(cnb.flags & (PP_PME_CHARGE << q)))
 384                 {
 385                     continue;
 386                 }
 387                 switch (q)
 388                 {
 389                     case eCommType_ChargeA: bufferPtr = pme_pp->chargeA.data(); break;
 390                     case eCommType_ChargeB: bufferPtr = pme_pp->chargeB.data(); break;
 391                     case eCommType_SQRTC6A: bufferPtr = pme_pp->sqrt_c6A.data(); break;
 392                     case eCommType_SQRTC6B: bufferPtr = pme_pp->sqrt_c6B.data(); break;
 393                     case eCommType_SigmaA: bufferPtr = pme_pp->sigmaA.data(); break;
 394                     case eCommType_SigmaB: bufferPtr = pme_pp->sigmaB.data(); break;
 395                     default: gmx_incons("Wrong eCommType");
 396                 }
 397                 nat = 0;
 398                 for (const auto& sender : pme_pp->ppRanks)
 399                 {
 400                     if (sender.numAtoms > 0)
 401                     {
 402                         MPI_Irecv(bufferPtr + nat, sender.numAtoms * sizeof(real), MPI_BYTE,
 403                                   sender.rankId, q, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 404                         nat += sender.numAtoms;
 405                         if (debug)
 406                         {
 407                             fprintf(debug, "Received from PP rank %d: %d %s\n", sender.rankId,
 408                                     sender.numAtoms,
 409                                     (q == eCommType_ChargeA || q == eCommType_ChargeB) ? "charges"
 410                                                                                        : "params");
 411                         }
 412                     }
 413                 }
 414             }
 415         }
 416
 417         if (cnb.flags & PP_PME_COORD)
 418         {
 419             if (atomSetChanged)
 420             {
 421                 gmx_pme_reinit_atoms(pme, nat, pme_pp->chargeA.data(), pme_pp->chargeB.data());
 422                 if (useGpuForPme)
 423                 {
 424                     stateGpu->reinit(nat, nat);
 425                     pme_gpu_set_device_x(pme, stateGpu->getCoordinates());
 426                 }
 427                 if (pme_pp->useGpuDirectComm)
 428                 {
 429                     GMX_ASSERT(runMode == PmeRunMode::GPU,
 430                                "GPU Direct PME-PP communication has been enabled, "
 431                                "but PME run mode is not PmeRunMode::GPU\n");
 432
 433                     // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses.
 434                     pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(
 435                             stateGpu->getCoordinates());
 436                     pme_pp->pmeForceSenderGpu->sendForceBufferAddressToPpRanks(
 437                             reinterpret_cast<rvec*>(pme_gpu_get_device_f(pme)));
 438                 }
 439             }
 440
 441
 442             /* The box, FE flag and lambda are sent along with the coordinates
 443              *  */
 444             copy_mat(cnb.box, box);
 445             *lambda_q               = cnb.lambda_q;
 446             *lambda_lj              = cnb.lambda_lj;
 447             *computeEnergyAndVirial = ((cnb.flags & PP_PME_ENER_VIR) != 0U);
 448             *step                   = cnb.step;
 449
 450             /* Receive the coordinates in place */
 451             nat = 0;
 452             for (const auto& sender : pme_pp->ppRanks)
 453             {
 454                 if (sender.numAtoms > 0)
 455                 {
 456                     if (pme_pp->useGpuDirectComm)
 457                     {
 458                         pme_pp->pmeCoordinateReceiverGpu->launchReceiveCoordinatesFromPpCudaDirect(
 459                                 sender.rankId);
 460                     }
 461                     else
 462                     {
 463                         MPI_Irecv(pme_pp->x[nat], sender.numAtoms * sizeof(rvec), MPI_BYTE, sender.rankId,
 464                                   eCommType_COORD, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 465                     }
 466                     nat += sender.numAtoms;
 467                     if (debug)
 468                     {
 469                         fprintf(debug,
 470                                 "Received from PP rank %d: %d "
 471                                 "coordinates\n",
 472                                 sender.rankId, sender.numAtoms);
 473                     }
 474                 }
 475             }
 476
 477             if (pme_pp->useGpuDirectComm)
 478             {
 479                 pme_pp->pmeCoordinateReceiverGpu->enqueueWaitReceiveCoordinatesFromPpCudaDirect();
 480             }
 481
 482             status = pmerecvqxX;
 483         }
 484
 485         /* Wait for the coordinates and/or charges to arrive */
 486         MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 487         messages = 0;
 488     } while (status == -1);
 489 #else
 490     GMX_UNUSED_VALUE(pme);
 491     GMX_UNUSED_VALUE(pme_pp);
 492     GMX_UNUSED_VALUE(box);
 493     GMX_UNUSED_VALUE(maxshift_x);
 494     GMX_UNUSED_VALUE(maxshift_y);
 495     GMX_UNUSED_VALUE(lambda_q);
 496     GMX_UNUSED_VALUE(lambda_lj);
 497     GMX_UNUSED_VALUE(computeEnergyAndVirial);
 498     GMX_UNUSED_VALUE(step);
 499     GMX_UNUSED_VALUE(grid_size);
 500     GMX_UNUSED_VALUE(ewaldcoeff_q);
 501     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 502     GMX_UNUSED_VALUE(useGpuForPme);
 503     GMX_UNUSED_VALUE(stateGpu);
 504
 505     status = pmerecvqxX;
 506 #endif
 507
 508     if (status == pmerecvqxX)
 509     {
 510         *natoms = nat;
 511     }
 512
 513     return status;
 514 }
 515
 516 #if GMX_MPI
 517 /*! \brief Send force data to PP ranks */
 518 static void sendFToPP(void* sendbuf, PpRanks receiver, gmx_pme_pp* pme_pp, int* messages)
 519 {
 520
 521     if (pme_pp->useGpuDirectComm)
 522     {
 523         GMX_ASSERT((pme_pp->pmeForceSenderGpu != nullptr),
 524                    "The use of GPU direct communication for PME-PP is enabled, "
 525                    "but the PME GPU force reciever object does not exist");
 526
 527         pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(receiver.rankId);
 528     }
 529     else
 530     {
 531         // Send using MPI
 532         MPI_Isend(sendbuf, receiver.numAtoms * sizeof(rvec), MPI_BYTE, receiver.rankId, 0,
 533                   pme_pp->mpi_comm_mysim, &pme_pp->req[*messages]);
 534         *messages = *messages + 1;
 535     }
 536 }
 537 #endif
 538
 539 /*! \brief Send the PME mesh force, virial and energy to the PP-only ranks. */
 540 static void gmx_pme_send_force_vir_ener(const gmx_pme_t& pme,
 541                                         gmx_pme_pp*      pme_pp,
 542                                         const PmeOutput& output,
 543                                         real             dvdlambda_q,
 544                                         real             dvdlambda_lj,
 545                                         float            cycles)
 546 {
 547 #if GMX_MPI
 548     gmx_pme_comm_vir_ene_t cve;
 549     int                    messages, ind_start, ind_end;
 550     cve.cycles = cycles;
 551
 552     /* Now the evaluated forces have to be transferred to the PP nodes */
 553     messages = 0;
 554     ind_end  = 0;
 555     for (const auto& receiver : pme_pp->ppRanks)
 556     {
 557         ind_start     = ind_end;
 558         ind_end       = ind_start + receiver.numAtoms;
 559         void* sendbuf = const_cast<void*>(static_cast<const void*>(output.forces_[ind_start]));
 560         if (pme_pp->useGpuDirectComm)
 561         {
 562             // Data will be transferred directly from GPU.
 563             rvec* d_f = reinterpret_cast<rvec*>(pme_gpu_get_device_f(&pme));
 564             sendbuf   = reinterpret_cast<void*>(&d_f[ind_start]);
 565         }
 566         sendFToPP(sendbuf, receiver, pme_pp, &messages);
 567     }
 568
 569     /* send virial and energy to our last PP node */
 570     copy_mat(output.coulombVirial_, cve.vir_q);
 571     copy_mat(output.lennardJonesVirial_, cve.vir_lj);
 572     cve.energy_q     = output.coulombEnergy_;
 573     cve.energy_lj    = output.lennardJonesEnergy_;
 574     cve.dvdlambda_q  = dvdlambda_q;
 575     cve.dvdlambda_lj = dvdlambda_lj;
 576     /* check for the signals to send back to a PP node */
 577     cve.stop_cond = gmx_get_stop_condition();
 578
 579     cve.cycles = cycles;
 580
 581     if (debug)
 582     {
 583         fprintf(debug, "PME rank sending to PP rank %d: virial and energy\n", pme_pp->peerRankId);
 584     }
 585     MPI_Isend(&cve, sizeof(cve), MPI_BYTE, pme_pp->peerRankId, 1, pme_pp->mpi_comm_mysim,
 586               &pme_pp->req[messages++]);
 587
 588     /* Wait for the forces to arrive */
 589     MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 590 #else
 591     GMX_RELEASE_ASSERT(false, "Invalid call to gmx_pme_send_force_vir_ener");
 592     GMX_UNUSED_VALUE(pme);
 593     GMX_UNUSED_VALUE(pme_pp);
 594     GMX_UNUSED_VALUE(output);
 595     GMX_UNUSED_VALUE(dvdlambda_q);
 596     GMX_UNUSED_VALUE(dvdlambda_lj);
 597     GMX_UNUSED_VALUE(cycles);
 598 #endif
 599 }
 600
 601 int gmx_pmeonly(struct gmx_pme_t*               pme,
 602                 const t_commrec*                cr,
 603                 t_nrnb*                         mynrnb,
 604                 gmx_wallcycle*                  wcycle,
 605                 gmx_walltime_accounting_t       walltime_accounting,
 606                 t_inputrec*                     ir,
 607                 PmeRunMode                      runMode,
 608                 const gmx::DeviceStreamManager* deviceStreamManager)
 609 {
 610     int     ret;
 611     int     natoms = 0;
 612     matrix  box;
 613     real    lambda_q   = 0;
 614     real    lambda_lj  = 0;
 615     int     maxshift_x = 0, maxshift_y = 0;
 616     real    dvdlambda_q, dvdlambda_lj;
 617     float   cycles;
 618     int     count;
 619     bool    computeEnergyAndVirial = false;
 620     int64_t step;
 621
 622     /* This data will only use with PME tuning, i.e. switching PME grids */
 623     std::vector<gmx_pme_t*> pmedata;
 624     pmedata.push_back(pme);
 625
 626     auto pme_pp = gmx_pme_pp_init(cr);
 627
 628     std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
 629     // TODO the variable below should be queried from the task assignment info
 630     const bool useGpuForPme = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
 631     if (useGpuForPme)
 632     {
 633         GMX_RELEASE_ASSERT(
 634                 deviceStreamManager != nullptr,
 635                 "Device stream manager can not be nullptr when using GPU in PME-only rank.");
 636         GMX_RELEASE_ASSERT(deviceStreamManager->streamIsValid(gmx::DeviceStreamType::Pme),
 637                            "Device stream can not be nullptr when using GPU in PME-only rank");
 638         changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
 639         changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
 640         if (c_enableGpuPmePpComms)
 641         {
 642             pme_pp->pmeCoordinateReceiverGpu = std::make_unique<gmx::PmeCoordinateReceiverGpu>(
 643                     deviceStreamManager->stream(gmx::DeviceStreamType::Pme), pme_pp->mpi_comm_mysim,
 644                     pme_pp->ppRanks);
 645             pme_pp->pmeForceSenderGpu = std::make_unique<gmx::PmeForceSenderGpu>(
 646                     deviceStreamManager->stream(gmx::DeviceStreamType::Pme), pme_pp->mpi_comm_mysim,
 647                     pme_pp->ppRanks);
 648         }
 649         // TODO: Special PME-only constructor is used here. There is no mechanism to prevent from using the other constructor here.
 650         //       This should be made safer.
 651         stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
 652                 &deviceStreamManager->stream(gmx::DeviceStreamType::Pme), deviceStreamManager->context(),
 653                 GpuApiCallBehavior::Async, pme_gpu_get_block_size(pme), wcycle);
 654     }
 655
 656     clear_nrnb(mynrnb);
 657
 658     count = 0;
 659     do /****** this is a quasi-loop over time steps! */
 660     {
 661         /* The reason for having a loop here is PME grid tuning/switching */
 662         do
 663         {
 664             /* Domain decomposition */
 665             ivec newGridSize;
 666             real ewaldcoeff_q = 0, ewaldcoeff_lj = 0;
 667             ret = gmx_pme_recv_coeffs_coords(pme, pme_pp.get(), &natoms, box, &maxshift_x, &maxshift_y,
 668                                              &lambda_q, &lambda_lj, &computeEnergyAndVirial, &step,
 669                                              &newGridSize, &ewaldcoeff_q, &ewaldcoeff_lj,
 670                                              useGpuForPme, stateGpu.get(), runMode);
 671
 672             if (ret == pmerecvqxSWITCHGRID)
 673             {
 674                 /* Switch the PME grid to newGridSize */
 675                 pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
 676             }
 677
 678             if (ret == pmerecvqxRESETCOUNTERS)
 679             {
 680                 /* Reset the cycle and flop counters */
 681                 reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
 682             }
 683         } while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
 684
 685         if (ret == pmerecvqxFINISH)
 686         {
 687             /* We should stop: break out of the loop */
 688             break;
 689         }
 690
 691         if (count == 0)
 692         {
 693             wallcycle_start(wcycle, ewcRUN);
 694             walltime_accounting_start_time(walltime_accounting);
 695         }
 696
 697         wallcycle_start(wcycle, ewcPMEMESH);
 698
 699         dvdlambda_q  = 0;
 700         dvdlambda_lj = 0;
 701
 702         // TODO Make a struct of array refs onto these per-atom fields
 703         // of pme_pp (maybe box, energy and virial, too; and likewise
 704         // from mdatoms for the other call to gmx_pme_do), so we have
 705         // fewer lines of code and less parameter passing.
 706         gmx::StepWorkload stepWork;
 707         stepWork.computeVirial = computeEnergyAndVirial;
 708         stepWork.computeEnergy = computeEnergyAndVirial;
 709         stepWork.computeForces = true;
 710         PmeOutput output       = { {}, false, 0, { { 0 } }, 0, 0, { { 0 } }, 0 };
 711         if (useGpuForPme)
 712         {
 713             stepWork.haveDynamicBox      = false;
 714             stepWork.useGpuPmeFReduction = pme_pp->useGpuDirectComm;
 715             // TODO this should be set properly by gmx_pme_recv_coeffs_coords,
 716             // or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
 717             pme_gpu_prepare_computation(pme, box, wcycle, stepWork);
 718             if (!pme_pp->useGpuDirectComm)
 719             {
 720                 stateGpu->copyCoordinatesToGpu(gmx::ArrayRef<gmx::RVec>(pme_pp->x), gmx::AtomLocality::All);
 721             }
 722             // On the separate PME rank we do not need a synchronizer as we schedule everything in a single stream
 723             // TODO: with pme on GPU the receive should make a list of synchronizers and pass it here #3157
 724             auto xReadyOnDevice = nullptr;
 725
 726             pme_gpu_launch_spread(pme, xReadyOnDevice, wcycle, lambda_q);
 727             pme_gpu_launch_complex_transforms(pme, wcycle, stepWork);
 728             pme_gpu_launch_gather(pme, wcycle, lambda_q);
 729             output = pme_gpu_wait_finish_task(pme, computeEnergyAndVirial, lambda_q, wcycle);
 730             pme_gpu_reinit_computation(pme, wcycle);
 731         }
 732         else
 733         {
 734             GMX_ASSERT(pme_pp->x.size() == static_cast<size_t>(natoms),
 735                        "The coordinate buffer should have size natoms");
 736
 737             gmx_pme_do(pme, pme_pp->x, pme_pp->f, pme_pp->chargeA.data(), pme_pp->chargeB.data(),
 738                        pme_pp->sqrt_c6A.data(), pme_pp->sqrt_c6B.data(), pme_pp->sigmaA.data(),
 739                        pme_pp->sigmaB.data(), box, cr, maxshift_x, maxshift_y, mynrnb, wcycle,
 740                        output.coulombVirial_, output.lennardJonesVirial_, &output.coulombEnergy_,
 741                        &output.lennardJonesEnergy_, lambda_q, lambda_lj, &dvdlambda_q,
 742                        &dvdlambda_lj, stepWork);
 743             output.forces_ = pme_pp->f;
 744         }
 745
 746         cycles = wallcycle_stop(wcycle, ewcPMEMESH);
 747         gmx_pme_send_force_vir_ener(*pme, pme_pp.get(), output, dvdlambda_q, dvdlambda_lj, cycles);
 748
 749         count++;
 750     } /***** end of quasi-loop, we stop with the break above */
 751     while (TRUE);
 752
 753     walltime_accounting_end_time(walltime_accounting);
 754
 755     return 0;
 756 }