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_output.h"
 105 #include "pme_pp_communication.h"
 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     /*! \brief whether GPU direct communications should send forces directly to remote GPU memory */
 138     bool sendForcesDirectToPpGpu = false;
 139 };
 140
 141 /*! \brief Initialize the PME-only side of the PME <-> PP communication */
 142 static std::unique_ptr<gmx_pme_pp> gmx_pme_pp_init(const t_commrec* cr)
 143 {
 144     auto pme_pp = std::make_unique<gmx_pme_pp>();
 145
 146 #if GMX_MPI
 147     int rank;
 148
 149     pme_pp->mpi_comm_mysim = cr->mpi_comm_mysim;
 150     MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
 151     auto ppRanks = get_pme_ddranks(cr, rank);
 152     pme_pp->ppRanks.reserve(ppRanks.size());
 153     for (const auto& ppRankId : ppRanks)
 154     {
 155         pme_pp->ppRanks.push_back({ ppRankId, 0 });
 156     }
 157     // The peer PP rank is the last one.
 158     pme_pp->peerRankId = pme_pp->ppRanks.back().rankId;
 159     pme_pp->req.resize(eCommType_NR * pme_pp->ppRanks.size());
 160     pme_pp->stat.resize(eCommType_NR * pme_pp->ppRanks.size());
 161 #else
 162     GMX_UNUSED_VALUE(cr);
 163 #endif
 164
 165     return pme_pp;
 166 }
 167
 168 static void reset_pmeonly_counters(gmx_wallcycle*            wcycle,
 169                                    gmx_walltime_accounting_t walltime_accounting,
 170                                    t_nrnb*                   nrnb,
 171                                    int64_t                   step,
 172                                    bool                      useGpuForPme)
 173 {
 174     /* Reset all the counters related to performance over the run */
 175     wallcycle_stop(wcycle, WallCycleCounter::Run);
 176     wallcycle_reset_all(wcycle);
 177     *nrnb = { 0 };
 178     wallcycle_start(wcycle, WallCycleCounter::Run);
 179     walltime_accounting_reset_time(walltime_accounting, step);
 180
 181     if (useGpuForPme)
 182     {
 183         resetGpuProfiler();
 184     }
 185 }
 186
 187 static gmx_pme_t* gmx_pmeonly_switch(std::vector<gmx_pme_t*>* pmedata,
 188                                      const ivec               grid_size,
 189                                      real                     ewaldcoeff_q,
 190                                      real                     ewaldcoeff_lj,
 191                                      const t_commrec*         cr,
 192                                      const t_inputrec*        ir)
 193 {
 194     GMX_ASSERT(pmedata, "Bad PME tuning list pointer");
 195     for (auto& pme : *pmedata)
 196     {
 197         GMX_ASSERT(pme, "Bad PME tuning list element pointer");
 198         if (gmx_pme_grid_matches(*pme, grid_size))
 199         {
 200             /* Here we have found an existing PME data structure that suits us.
 201              * However, in the GPU case, we have to reinitialize it - there's only one GPU structure.
 202              * This should not cause actual GPU reallocations, at least (the allocated buffers are never shrunk).
 203              * So, just some grid size updates in the GPU kernel parameters.
 204              * TODO: this should be something like gmx_pme_update_split_params()
 205              */
 206             gmx_pme_reinit(&pme, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 207             return pme;
 208         }
 209     }
 210
 211     const auto& pme          = pmedata->back();
 212     gmx_pme_t*  newStructure = nullptr;
 213     // Copy last structure with new grid params
 214     gmx_pme_reinit(&newStructure, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 215     pmedata->push_back(newStructure);
 216     return newStructure;
 217 }
 218
 219 /*! \brief Called by PME-only ranks to receive coefficients and coordinates
 220  *
 221  * \param[in] pme                     PME data structure.
 222  * \param[in,out] pme_pp              PME-PP communication structure.
 223  * \param[out] natoms                 Number of received atoms.
 224  * \param[out] box                    System box, if received.
 225  * \param[out] maxshift_x             Maximum shift in X direction, if received.
 226  * \param[out] maxshift_y             Maximum shift in Y direction, if received.
 227  * \param[out] lambda_q               Free-energy lambda for electrostatics, if received.
 228  * \param[out] lambda_lj              Free-energy lambda for Lennard-Jones, if received.
 229  * \param[out] computeEnergyAndVirial Set to true if this is an energy/virial calculation
 230  *                                    step, otherwise set to false.
 231  * \param[out] step                   MD integration step number.
 232  * \param[out] grid_size              PME grid size, if received.
 233  * \param[out] ewaldcoeff_q           Ewald cut-off parameter for electrostatics, if received.
 234  * \param[out] ewaldcoeff_lj          Ewald cut-off parameter for Lennard-Jones, if received.
 235  * \param[in]  useGpuForPme           Flag on whether PME is on GPU.
 236  * \param[in]  stateGpu               GPU state propagator object.
 237  * \param[in]  runMode                PME run mode.
 238  *
 239  * \retval pmerecvqxX                 All parameters were set, chargeA and chargeB can be NULL.
 240  * \retval pmerecvqxFINISH            No parameters were set.
 241  * \retval pmerecvqxSWITCHGRID        Only grid_size and *ewaldcoeff were set.
 242  * \retval pmerecvqxRESETCOUNTERS     *step was set.
 243  */
 244 static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t*            pme,
 245                                       gmx_pme_pp*                  pme_pp,
 246                                       int*                         natoms,
 247                                       matrix                       box,
 248                                       int*                         maxshift_x,
 249                                       int*                         maxshift_y,
 250                                       real*                        lambda_q,
 251                                       real*                        lambda_lj,
 252                                       gmx_bool*                    computeEnergyAndVirial,
 253                                       int64_t*                     step,
 254                                       ivec*                        grid_size,
 255                                       real*                        ewaldcoeff_q,
 256                                       real*                        ewaldcoeff_lj,
 257                                       bool                         useGpuForPme,
 258                                       gmx::StatePropagatorDataGpu* stateGpu,
 259                                       PmeRunMode gmx_unused        runMode)
 260 {
 261     int status = -1;
 262     int nat    = 0;
 263
 264 #if GMX_MPI
 265     int  messages       = 0;
 266     bool atomSetChanged = false;
 267
 268     do
 269     {
 270         gmx_pme_comm_n_box_t cnb;
 271         cnb.flags = 0;
 272
 273         /* Receive the send count, box and time step from the peer PP node */
 274         MPI_Recv(&cnb, sizeof(cnb), MPI_BYTE, pme_pp->peerRankId, eCommType_CNB, pme_pp->mpi_comm_mysim, MPI_STATUS_IGNORE);
 275
 276         *step = cnb.step;
 277
 278         if (debug)
 279         {
 280             fprintf(debug,
 281                     "PME only rank receiving:%s%s%s%s%s\n",
 282                     (cnb.flags & PP_PME_CHARGE) ? " charges" : "",
 283                     (cnb.flags & PP_PME_COORD) ? " coordinates" : "",
 284                     (cnb.flags & PP_PME_FINISH) ? " finish" : "",
 285                     (cnb.flags & PP_PME_SWITCHGRID) ? " switch grid" : "",
 286                     (cnb.flags & PP_PME_RESETCOUNTERS) ? " reset counters" : "");
 287         }
 288
 289         pme_pp->useGpuDirectComm = ((cnb.flags & PP_PME_GPUCOMMS) != 0);
 290         GMX_ASSERT(!pme_pp->useGpuDirectComm || (pme_pp->pmeForceSenderGpu != nullptr),
 291                    "The use of GPU direct communication for PME-PP is enabled, "
 292                    "but the PME GPU force reciever object does not exist");
 293         pme_pp->sendForcesDirectToPpGpu = ((cnb.flags & PP_PME_RECVFTOGPU) != 0);
 294
 295         if (cnb.flags & PP_PME_FINISH)
 296         {
 297             status = pmerecvqxFINISH;
 298         }
 299
 300         if (cnb.flags & PP_PME_SWITCHGRID)
 301         {
 302             /* Special case, receive the new parameters and return */
 303             copy_ivec(cnb.grid_size, *grid_size);
 304             *ewaldcoeff_q  = cnb.ewaldcoeff_q;
 305             *ewaldcoeff_lj = cnb.ewaldcoeff_lj;
 306
 307             status = pmerecvqxSWITCHGRID;
 308         }
 309
 310         if (cnb.flags & PP_PME_RESETCOUNTERS)
 311         {
 312             /* Special case, receive the step (set above) and return */
 313             status = pmerecvqxRESETCOUNTERS;
 314         }
 315
 316         if (cnb.flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
 317         {
 318             atomSetChanged = true;
 319
 320             /* Receive the send counts from the other PP nodes */
 321             for (auto& sender : pme_pp->ppRanks)
 322             {
 323                 if (sender.rankId == pme_pp->peerRankId)
 324                 {
 325                     sender.numAtoms = cnb.natoms;
 326                 }
 327                 else
 328                 {
 329                     MPI_Irecv(&sender.numAtoms,
 330                               sizeof(sender.numAtoms),
 331                               MPI_BYTE,
 332                               sender.rankId,
 333                               eCommType_CNB,
 334                               pme_pp->mpi_comm_mysim,
 335                               &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,
 403                                   sender.numAtoms * sizeof(real),
 404                                   MPI_BYTE,
 405                                   sender.rankId,
 406                                   q,
 407                                   pme_pp->mpi_comm_mysim,
 408                                   &pme_pp->req[messages++]);
 409                         nat += sender.numAtoms;
 410                         if (debug)
 411                         {
 412                             fprintf(debug,
 413                                     "Received from PP rank %d: %d %s\n",
 414                                     sender.rankId,
 415                                     sender.numAtoms,
 416                                     (q == eCommType_ChargeA || q == eCommType_ChargeB) ? "charges"
 417                                                                                        : "params");
 418                         }
 419                     }
 420                 }
 421             }
 422         }
 423
 424         if (cnb.flags & PP_PME_COORD)
 425         {
 426             if (atomSetChanged)
 427             {
 428                 gmx_pme_reinit_atoms(pme, nat, pme_pp->chargeA, pme_pp->chargeB);
 429                 if (useGpuForPme)
 430                 {
 431                     stateGpu->reinit(nat, nat);
 432                     pme_gpu_set_device_x(pme, stateGpu->getCoordinates());
 433                 }
 434                 if (pme_pp->useGpuDirectComm)
 435                 {
 436                     GMX_ASSERT(runMode == PmeRunMode::GPU,
 437                                "GPU Direct PME-PP communication has been enabled, "
 438                                "but PME run mode is not PmeRunMode::GPU\n");
 439
 440                     // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses.
 441                     pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(
 442                             stateGpu->getCoordinates());
 443                     pme_pp->pmeForceSenderGpu->setForceSendBuffer(pme_gpu_get_device_f(pme));
 444                 }
 445             }
 446
 447
 448             /* The box, FE flag and lambda are sent along with the coordinates
 449              *  */
 450             copy_mat(cnb.box, box);
 451             *lambda_q               = cnb.lambda_q;
 452             *lambda_lj              = cnb.lambda_lj;
 453             *computeEnergyAndVirial = ((cnb.flags & PP_PME_ENER_VIR) != 0U);
 454             *step                   = cnb.step;
 455
 456             /* Receive the coordinates in place */
 457             nat = 0;
 458             for (const auto& sender : pme_pp->ppRanks)
 459             {
 460                 if (sender.numAtoms > 0)
 461                 {
 462                     if (pme_pp->useGpuDirectComm)
 463                     {
 464                         if (GMX_THREAD_MPI)
 465                         {
 466                             pme_pp->pmeCoordinateReceiverGpu->receiveCoordinatesSynchronizerFromPpCudaDirect(
 467                                     sender.rankId);
 468                         }
 469                         else
 470                         {
 471                             pme_pp->pmeCoordinateReceiverGpu->launchReceiveCoordinatesFromPpCudaMpi(
 472                                     stateGpu->getCoordinates(), nat, sender.numAtoms * sizeof(rvec), sender.rankId);
 473                         }
 474                     }
 475                     else
 476                     {
 477                         MPI_Irecv(pme_pp->x[nat],
 478                                   sender.numAtoms * sizeof(rvec),
 479                                   MPI_BYTE,
 480                                   sender.rankId,
 481                                   eCommType_COORD,
 482                                   pme_pp->mpi_comm_mysim,
 483                                   &pme_pp->req[messages++]);
 484                     }
 485                     nat += sender.numAtoms;
 486                     if (debug)
 487                     {
 488                         fprintf(debug,
 489                                 "Received from PP rank %d: %d "
 490                                 "coordinates\n",
 491                                 sender.rankId,
 492                                 sender.numAtoms);
 493                     }
 494                 }
 495             }
 496
 497             if (pme_pp->useGpuDirectComm)
 498             {
 499                 pme_pp->pmeCoordinateReceiverGpu->synchronizeOnCoordinatesFromPpRanks();
 500             }
 501
 502             status = pmerecvqxX;
 503         }
 504
 505         /* Wait for the coordinates and/or charges to arrive */
 506         MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 507         messages = 0;
 508     } while (status == -1);
 509 #else
 510     GMX_UNUSED_VALUE(pme);
 511     GMX_UNUSED_VALUE(pme_pp);
 512     GMX_UNUSED_VALUE(box);
 513     GMX_UNUSED_VALUE(maxshift_x);
 514     GMX_UNUSED_VALUE(maxshift_y);
 515     GMX_UNUSED_VALUE(lambda_q);
 516     GMX_UNUSED_VALUE(lambda_lj);
 517     GMX_UNUSED_VALUE(computeEnergyAndVirial);
 518     GMX_UNUSED_VALUE(step);
 519     GMX_UNUSED_VALUE(grid_size);
 520     GMX_UNUSED_VALUE(ewaldcoeff_q);
 521     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 522     GMX_UNUSED_VALUE(useGpuForPme);
 523     GMX_UNUSED_VALUE(stateGpu);
 524
 525     status = pmerecvqxX;
 526 #endif
 527
 528     if (status == pmerecvqxX)
 529     {
 530         *natoms = nat;
 531     }
 532
 533     return status;
 534 }
 535
 536 /*! \brief Send the PME mesh force, virial and energy to the PP-only ranks. */
 537 static void gmx_pme_send_force_vir_ener(const gmx_pme_t& pme,
 538                                         gmx_pme_pp*      pme_pp,
 539                                         const PmeOutput& output,
 540                                         real             dvdlambda_q,
 541                                         real             dvdlambda_lj,
 542                                         float            cycles)
 543 {
 544 #if GMX_MPI
 545     gmx_pme_comm_vir_ene_t cve;
 546     int                    messages, ind_start, ind_end;
 547     cve.cycles = cycles;
 548
 549     /* Now the evaluated forces have to be transferred to the PP nodes */
 550     messages = 0;
 551     ind_end  = 0;
 552     for (const auto& receiver : pme_pp->ppRanks)
 553     {
 554         ind_start = ind_end;
 555         ind_end   = ind_start + receiver.numAtoms;
 556         if (pme_pp->useGpuDirectComm)
 557         {
 558             GMX_ASSERT((pme_pp->pmeForceSenderGpu != nullptr),
 559                        "The use of GPU direct communication for PME-PP is enabled, "
 560                        "but the PME GPU force reciever object does not exist");
 561
 562             if (GMX_THREAD_MPI)
 563             {
 564                 pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(
 565                         receiver.rankId, receiver.numAtoms, pme_pp->sendForcesDirectToPpGpu);
 566             }
 567             else
 568             {
 569                 pme_pp->pmeForceSenderGpu->sendFToPpCudaMpi(pme_gpu_get_device_f(&pme),
 570                                                             ind_start,
 571                                                             receiver.numAtoms * sizeof(rvec),
 572                                                             receiver.rankId,
 573                                                             &pme_pp->req[messages]);
 574
 575                 messages++;
 576             }
 577         }
 578         else
 579         {
 580             void* sendbuf = const_cast<void*>(static_cast<const void*>(output.forces_[ind_start]));
 581             // Send using MPI
 582             MPI_Isend(sendbuf,
 583                       receiver.numAtoms * sizeof(rvec),
 584                       MPI_BYTE,
 585                       receiver.rankId,
 586                       0,
 587                       pme_pp->mpi_comm_mysim,
 588                       &pme_pp->req[messages]);
 589             messages++;
 590         }
 591     }
 592
 593     /* send virial and energy to our last PP node */
 594     copy_mat(output.coulombVirial_, cve.vir_q);
 595     copy_mat(output.lennardJonesVirial_, cve.vir_lj);
 596     cve.energy_q     = output.coulombEnergy_;
 597     cve.energy_lj    = output.lennardJonesEnergy_;
 598     cve.dvdlambda_q  = dvdlambda_q;
 599     cve.dvdlambda_lj = dvdlambda_lj;
 600     /* check for the signals to send back to a PP node */
 601     cve.stop_cond = gmx_get_stop_condition();
 602
 603     cve.cycles = cycles;
 604
 605     if (debug)
 606     {
 607         fprintf(debug, "PME rank sending to PP rank %d: virial and energy\n", pme_pp->peerRankId);
 608     }
 609     MPI_Isend(&cve, sizeof(cve), MPI_BYTE, pme_pp->peerRankId, 1, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 610
 611     /* Wait for the forces to arrive */
 612     MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 613 #else
 614     GMX_RELEASE_ASSERT(false, "Invalid call to gmx_pme_send_force_vir_ener");
 615     GMX_UNUSED_VALUE(pme);
 616     GMX_UNUSED_VALUE(pme_pp);
 617     GMX_UNUSED_VALUE(output);
 618     GMX_UNUSED_VALUE(dvdlambda_q);
 619     GMX_UNUSED_VALUE(dvdlambda_lj);
 620     GMX_UNUSED_VALUE(cycles);
 621 #endif
 622 }
 623
 624 int gmx_pmeonly(struct gmx_pme_t*               pme,
 625                 const t_commrec*                cr,
 626                 t_nrnb*                         mynrnb,
 627                 gmx_wallcycle*                  wcycle,
 628                 gmx_walltime_accounting_t       walltime_accounting,
 629                 t_inputrec*                     ir,
 630                 PmeRunMode                      runMode,
 631                 bool                            useGpuPmePpCommunication,
 632                 const gmx::DeviceStreamManager* deviceStreamManager)
 633 {
 634     int     ret;
 635     int     natoms = 0;
 636     matrix  box;
 637     real    lambda_q   = 0;
 638     real    lambda_lj  = 0;
 639     int     maxshift_x = 0, maxshift_y = 0;
 640     real    dvdlambda_q, dvdlambda_lj;
 641     float   cycles;
 642     int     count;
 643     bool    computeEnergyAndVirial = false;
 644     int64_t step;
 645
 646     /* This data will only use with PME tuning, i.e. switching PME grids */
 647     std::vector<gmx_pme_t*> pmedata;
 648     pmedata.push_back(pme);
 649
 650     auto pme_pp = gmx_pme_pp_init(cr);
 651
 652     std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
 653     // TODO the variable below should be queried from the task assignment info
 654     const bool useGpuForPme = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
 655     if (useGpuForPme)
 656     {
 657         GMX_RELEASE_ASSERT(
 658                 deviceStreamManager != nullptr,
 659                 "Device stream manager can not be nullptr when using GPU in PME-only rank.");
 660         GMX_RELEASE_ASSERT(deviceStreamManager->streamIsValid(gmx::DeviceStreamType::Pme),
 661                            "Device stream can not be nullptr when using GPU in PME-only rank");
 662         changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
 663         changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
 664         if (useGpuPmePpCommunication)
 665         {
 666             pme_pp->pmeCoordinateReceiverGpu = std::make_unique<gmx::PmeCoordinateReceiverGpu>(
 667                     deviceStreamManager->stream(gmx::DeviceStreamType::Pme),
 668                     pme_pp->mpi_comm_mysim,
 669                     pme_pp->ppRanks);
 670             pme_pp->pmeForceSenderGpu =
 671                     std::make_unique<gmx::PmeForceSenderGpu>(pme_gpu_get_f_ready_synchronizer(pme),
 672                                                              pme_pp->mpi_comm_mysim,
 673                                                              deviceStreamManager->context(),
 674                                                              pme_pp->ppRanks);
 675         }
 676         // TODO: Special PME-only constructor is used here. There is no mechanism to prevent from using the other constructor here.
 677         //       This should be made safer.
 678         stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
 679                 &deviceStreamManager->stream(gmx::DeviceStreamType::Pme),
 680                 deviceStreamManager->context(),
 681                 GpuApiCallBehavior::Async,
 682                 pme_gpu_get_block_size(pme),
 683                 wcycle);
 684     }
 685
 686     clear_nrnb(mynrnb);
 687
 688     count = 0;
 689     do /****** this is a quasi-loop over time steps! */
 690     {
 691         /* The reason for having a loop here is PME grid tuning/switching */
 692         do
 693         {
 694             /* Domain decomposition */
 695             ivec newGridSize;
 696             real ewaldcoeff_q = 0, ewaldcoeff_lj = 0;
 697             ret = gmx_pme_recv_coeffs_coords(pme,
 698                                              pme_pp.get(),
 699                                              &natoms,
 700                                              box,
 701                                              &maxshift_x,
 702                                              &maxshift_y,
 703                                              &lambda_q,
 704                                              &lambda_lj,
 705                                              &computeEnergyAndVirial,
 706                                              &step,
 707                                              &newGridSize,
 708                                              &ewaldcoeff_q,
 709                                              &ewaldcoeff_lj,
 710                                              useGpuForPme,
 711                                              stateGpu.get(),
 712                                              runMode);
 713
 714             if (ret == pmerecvqxSWITCHGRID)
 715             {
 716                 /* Switch the PME grid to newGridSize */
 717                 pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
 718             }
 719
 720             if (ret == pmerecvqxRESETCOUNTERS)
 721             {
 722                 /* Reset the cycle and flop counters */
 723                 reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
 724             }
 725         } while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
 726
 727         if (ret == pmerecvqxFINISH)
 728         {
 729             /* We should stop: break out of the loop */
 730             break;
 731         }
 732
 733         if (count == 0)
 734         {
 735             wallcycle_start(wcycle, WallCycleCounter::Run);
 736             walltime_accounting_start_time(walltime_accounting);
 737         }
 738
 739         wallcycle_start(wcycle, WallCycleCounter::PmeMesh);
 740
 741         dvdlambda_q  = 0;
 742         dvdlambda_lj = 0;
 743
 744         // TODO Make a struct of array refs onto these per-atom fields
 745         // of pme_pp (maybe box, energy and virial, too; and likewise
 746         // from mdatoms for the other call to gmx_pme_do), so we have
 747         // fewer lines of code and less parameter passing.
 748         gmx::StepWorkload stepWork;
 749         stepWork.computeVirial = computeEnergyAndVirial;
 750         stepWork.computeEnergy = computeEnergyAndVirial;
 751         stepWork.computeForces = true;
 752         PmeOutput output       = { {}, false, 0, { { 0 } }, 0, 0, { { 0 } }, 0 };
 753         if (useGpuForPme)
 754         {
 755             stepWork.haveDynamicBox      = false;
 756             stepWork.useGpuPmeFReduction = pme_pp->useGpuDirectComm;
 757             // TODO this should be set properly by gmx_pme_recv_coeffs_coords,
 758             // or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
 759             pme_gpu_prepare_computation(pme, box, wcycle, stepWork);
 760             if (!pme_pp->useGpuDirectComm)
 761             {
 762                 stateGpu->copyCoordinatesToGpu(gmx::ArrayRef<gmx::RVec>(pme_pp->x), gmx::AtomLocality::All);
 763             }
 764             // On the separate PME rank we do not need a synchronizer as we schedule everything in a single stream
 765             // TODO: with pme on GPU the receive should make a list of synchronizers and pass it here #3157
 766             auto xReadyOnDevice = nullptr;
 767
 768             pme_gpu_launch_spread(pme, xReadyOnDevice, wcycle, lambda_q);
 769             pme_gpu_launch_complex_transforms(pme, wcycle, stepWork);
 770             pme_gpu_launch_gather(pme, wcycle, lambda_q);
 771             output = pme_gpu_wait_finish_task(pme, computeEnergyAndVirial, lambda_q, wcycle);
 772             pme_gpu_reinit_computation(pme, wcycle);
 773         }
 774         else
 775         {
 776             GMX_ASSERT(pme_pp->x.size() == static_cast<size_t>(natoms),
 777                        "The coordinate buffer should have size natoms");
 778
 779             gmx_pme_do(pme,
 780                        pme_pp->x,
 781                        pme_pp->f,
 782                        pme_pp->chargeA,
 783                        pme_pp->chargeB,
 784                        pme_pp->sqrt_c6A,
 785                        pme_pp->sqrt_c6B,
 786                        pme_pp->sigmaA,
 787                        pme_pp->sigmaB,
 788                        box,
 789                        cr,
 790                        maxshift_x,
 791                        maxshift_y,
 792                        mynrnb,
 793                        wcycle,
 794                        output.coulombVirial_,
 795                        output.lennardJonesVirial_,
 796                        &output.coulombEnergy_,
 797                        &output.lennardJonesEnergy_,
 798                        lambda_q,
 799                        lambda_lj,
 800                        &dvdlambda_q,
 801                        &dvdlambda_lj,
 802                        stepWork);
 803             output.forces_ = pme_pp->f;
 804         }
 805
 806         cycles = wallcycle_stop(wcycle, WallCycleCounter::PmeMesh);
 807         gmx_pme_send_force_vir_ener(*pme, pme_pp.get(), output, dvdlambda_q, dvdlambda_lj, cycles);
 808
 809         count++;
 810     } /***** end of quasi-loop, we stop with the break above */
 811     while (TRUE);
 812
 813     walltime_accounting_end_time(walltime_accounting);
 814
 815     return 0;
 816 }