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/hostallocator.h"
  86 #include "gromacs/math/gmxcomplex.h"
  87 #include "gromacs/math/units.h"
  88 #include "gromacs/math/vec.h"
  89 #include "gromacs/mdtypes/commrec.h"
  90 #include "gromacs/mdtypes/forceoutput.h"
  91 #include "gromacs/mdtypes/inputrec.h"
  92 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  93 #include "gromacs/timing/cyclecounter.h"
  94 #include "gromacs/timing/wallcycle.h"
  95 #include "gromacs/utility/fatalerror.h"
  96 #include "gromacs/utility/futil.h"
  97 #include "gromacs/utility/gmxmpi.h"
  98 #include "gromacs/utility/gmxomp.h"
  99 #include "gromacs/utility/smalloc.h"
 100
 101 #include "pme_gpu_internal.h"
 102 #include "pme_output.h"
 103 #include "pme_pp_communication.h"
 104
 105 /*! \brief environment variable to enable GPU P2P communication */
 106 static const bool c_enableGpuPmePpComms =
 107         (getenv("GMX_GPU_PME_PP_COMMS") != nullptr) && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA);
 108
 109 /*! \brief Master PP-PME communication data structure */
 110 struct gmx_pme_pp
 111 {
 112     MPI_Comm             mpi_comm_mysim; /**< MPI communicator for this simulation */
 113     std::vector<PpRanks> ppRanks;        /**< The PP partner ranks                 */
 114     int                  peerRankId;     /**< The peer PP rank id                  */
 115     //@{
 116     /**< Vectors of A- and B-state parameters used to transfer vectors to PME ranks  */
 117     gmx::PaddedHostVector<real> chargeA;
 118     std::vector<real>           chargeB;
 119     std::vector<real>           sqrt_c6A;
 120     std::vector<real>           sqrt_c6B;
 121     std::vector<real>           sigmaA;
 122     std::vector<real>           sigmaB;
 123     //@}
 124     gmx::HostVector<gmx::RVec> x; /**< Vector of atom coordinates to transfer to PME ranks */
 125     std::vector<gmx::RVec>     f; /**< Vector of atom forces received from PME ranks */
 126     //@{
 127     /**< Vectors of MPI objects used in non-blocking communication between multiple PP ranks per PME rank */
 128     std::vector<MPI_Request> req;
 129     std::vector<MPI_Status>  stat;
 130     //@}
 131
 132     /*! \brief object for receiving coordinates using communications operating on GPU memory space */
 133     std::unique_ptr<gmx::PmeCoordinateReceiverGpu> pmeCoordinateReceiverGpu;
 134     /*! \brief object for sending PME force using communications operating on GPU memory space */
 135     std::unique_ptr<gmx::PmeForceSenderGpu> pmeForceSenderGpu;
 136
 137     /*! \brief whether GPU direct communications are active for PME-PP transfers */
 138     bool useGpuDirectComm = 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_t           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, ewcRUN);
 176     wallcycle_reset_all(wcycle);
 177     *nrnb = { 0 };
 178     wallcycle_start(wcycle, ewcRUN);
 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] bEnerVir          Set to true if this is an energy/virial calculation step, otherwise
 230  * set to false. \param[out] step              MD integration step number. \param[out] grid_size PME
 231  * grid size, if received. \param[out] ewaldcoeff_q         Ewald cut-off parameter for
 232  * electrostatics, if received. \param[out] ewaldcoeff_lj         Ewald cut-off parameter for
 233  * Lennard-Jones, if received. \param[in] useGpuForPme      flag on whether PME is on GPU \param[in]
 234  * stateGpu          GPU state propagator object \param[in] runMode           PME run mode
 235  *
 236  * \retval pmerecvqxX             All parameters were set, chargeA and chargeB can be NULL.
 237  * \retval pmerecvqxFINISH        No parameters were set.
 238  * \retval pmerecvqxSWITCHGRID    Only grid_size and *ewaldcoeff were set.
 239  * \retval pmerecvqxRESETCOUNTERS *step was set.
 240  */
 241 static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t*            pme,
 242                                       gmx_pme_pp*                  pme_pp,
 243                                       int*                         natoms,
 244                                       matrix                       box,
 245                                       int*                         maxshift_x,
 246                                       int*                         maxshift_y,
 247                                       real*                        lambda_q,
 248                                       real*                        lambda_lj,
 249                                       gmx_bool*                    bEnerVir,
 250                                       int64_t*                     step,
 251                                       ivec*                        grid_size,
 252                                       real*                        ewaldcoeff_q,
 253                                       real*                        ewaldcoeff_lj,
 254                                       bool                         useGpuForPme,
 255                                       gmx::StatePropagatorDataGpu* stateGpu,
 256                                       PmeRunMode gmx_unused runMode)
 257 {
 258     int status = -1;
 259     int nat    = 0;
 260
 261 #if GMX_MPI
 262     unsigned int flags          = 0;
 263     int          messages       = 0;
 264     bool         atomSetChanged = false;
 265
 266     do
 267     {
 268         gmx_pme_comm_n_box_t cnb;
 269         cnb.flags = 0;
 270
 271         /* Receive the send count, box and time step from the peer PP node */
 272         MPI_Recv(&cnb, sizeof(cnb), MPI_BYTE, pme_pp->peerRankId, eCommType_CNB,
 273                  pme_pp->mpi_comm_mysim, MPI_STATUS_IGNORE);
 274
 275         /* We accumulate all received flags */
 276         flags |= cnb.flags;
 277
 278         *step = cnb.step;
 279
 280         if (debug)
 281         {
 282             fprintf(debug, "PME only rank receiving:%s%s%s%s%s\n",
 283                     (cnb.flags & PP_PME_CHARGE) ? " charges" : "",
 284                     (cnb.flags & PP_PME_COORD) ? " coordinates" : "",
 285                     (cnb.flags & PP_PME_FINISH) ? " finish" : "",
 286                     (cnb.flags & PP_PME_SWITCHGRID) ? " switch grid" : "",
 287                     (cnb.flags & PP_PME_RESETCOUNTERS) ? " reset counters" : "");
 288         }
 289
 290         pme_pp->useGpuDirectComm = ((cnb.flags & PP_PME_GPUCOMMS) != 0);
 291         GMX_ASSERT(!pme_pp->useGpuDirectComm || (pme_pp->pmeForceSenderGpu != nullptr),
 292                    "The use of GPU direct communication for PME-PP is enabled, "
 293                    "but the PME GPU force reciever object does not exist");
 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, sizeof(sender.numAtoms), MPI_BYTE, sender.rankId,
 330                               eCommType_CNB, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 331                 }
 332             }
 333             MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 334             messages = 0;
 335
 336             nat = 0;
 337             for (const auto& sender : pme_pp->ppRanks)
 338             {
 339                 nat += sender.numAtoms;
 340             }
 341
 342             if (cnb.flags & PP_PME_CHARGE)
 343             {
 344                 pme_pp->chargeA.resizeWithPadding(nat);
 345             }
 346             if (cnb.flags & PP_PME_CHARGEB)
 347             {
 348                 pme_pp->chargeB.resize(nat);
 349             }
 350             if (cnb.flags & PP_PME_SQRTC6)
 351             {
 352                 pme_pp->sqrt_c6A.resize(nat);
 353             }
 354             if (cnb.flags & PP_PME_SQRTC6B)
 355             {
 356                 pme_pp->sqrt_c6B.resize(nat);
 357             }
 358             if (cnb.flags & PP_PME_SIGMA)
 359             {
 360                 pme_pp->sigmaA.resize(nat);
 361             }
 362             if (cnb.flags & PP_PME_SIGMAB)
 363             {
 364                 pme_pp->sigmaB.resize(nat);
 365             }
 366             pme_pp->x.resize(nat);
 367             pme_pp->f.resize(nat);
 368
 369             /* maxshift is sent when the charges are sent */
 370             *maxshift_x = cnb.maxshift_x;
 371             *maxshift_y = cnb.maxshift_y;
 372
 373             /* Receive the charges in place */
 374             for (int q = 0; q < eCommType_NR; q++)
 375             {
 376                 real* bufferPtr;
 377
 378                 if (!(cnb.flags & (PP_PME_CHARGE << q)))
 379                 {
 380                     continue;
 381                 }
 382                 switch (q)
 383                 {
 384                     case eCommType_ChargeA: bufferPtr = pme_pp->chargeA.data(); break;
 385                     case eCommType_ChargeB: bufferPtr = pme_pp->chargeB.data(); break;
 386                     case eCommType_SQRTC6A: bufferPtr = pme_pp->sqrt_c6A.data(); break;
 387                     case eCommType_SQRTC6B: bufferPtr = pme_pp->sqrt_c6B.data(); break;
 388                     case eCommType_SigmaA: bufferPtr = pme_pp->sigmaA.data(); break;
 389                     case eCommType_SigmaB: bufferPtr = pme_pp->sigmaB.data(); break;
 390                     default: gmx_incons("Wrong eCommType");
 391                 }
 392                 nat = 0;
 393                 for (const auto& sender : pme_pp->ppRanks)
 394                 {
 395                     if (sender.numAtoms > 0)
 396                     {
 397                         MPI_Irecv(bufferPtr + nat, sender.numAtoms * sizeof(real), MPI_BYTE,
 398                                   sender.rankId, q, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 399                         nat += sender.numAtoms;
 400                         if (debug)
 401                         {
 402                             fprintf(debug, "Received from PP rank %d: %d %s\n", sender.rankId,
 403                                     sender.numAtoms,
 404                                     (q == eCommType_ChargeA || q == eCommType_ChargeB) ? "charges"
 405                                                                                        : "params");
 406                         }
 407                     }
 408                 }
 409             }
 410         }
 411
 412         if (cnb.flags & PP_PME_COORD)
 413         {
 414             if (atomSetChanged)
 415             {
 416                 gmx_pme_reinit_atoms(pme, nat, pme_pp->chargeA.data());
 417                 if (useGpuForPme)
 418                 {
 419                     stateGpu->reinit(nat, nat);
 420                     pme_gpu_set_device_x(pme, stateGpu->getCoordinates());
 421                 }
 422                 if (pme_pp->useGpuDirectComm)
 423                 {
 424                     GMX_ASSERT(runMode == PmeRunMode::GPU,
 425                                "GPU Direct PME-PP communication has been enabled, "
 426                                "but PME run mode is not PmeRunMode::GPU\n");
 427
 428                     // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses.
 429                     pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(
 430                             pme_gpu_get_device_x(pme));
 431                     pme_pp->pmeForceSenderGpu->sendForceBufferAddressToPpRanks(
 432                             reinterpret_cast<rvec*>(pme_gpu_get_device_f(pme)));
 433                 }
 434             }
 435
 436
 437             /* The box, FE flag and lambda are sent along with the coordinates
 438              *  */
 439             copy_mat(cnb.box, box);
 440             *lambda_q  = cnb.lambda_q;
 441             *lambda_lj = cnb.lambda_lj;
 442             *bEnerVir  = ((cnb.flags & PP_PME_ENER_VIR) != 0U);
 443             *step      = cnb.step;
 444
 445             /* Receive the coordinates in place */
 446             nat = 0;
 447             for (const auto& sender : pme_pp->ppRanks)
 448             {
 449                 if (sender.numAtoms > 0)
 450                 {
 451                     if (pme_pp->useGpuDirectComm)
 452                     {
 453                         pme_pp->pmeCoordinateReceiverGpu->launchReceiveCoordinatesFromPpCudaDirect(
 454                                 sender.rankId);
 455                     }
 456                     else
 457                     {
 458                         MPI_Irecv(pme_pp->x[nat], sender.numAtoms * sizeof(rvec), MPI_BYTE, sender.rankId,
 459                                   eCommType_COORD, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 460                     }
 461                     nat += sender.numAtoms;
 462                     if (debug)
 463                     {
 464                         fprintf(debug,
 465                                 "Received from PP rank %d: %d "
 466                                 "coordinates\n",
 467                                 sender.rankId, sender.numAtoms);
 468                     }
 469                 }
 470             }
 471
 472             if (pme_pp->useGpuDirectComm)
 473             {
 474                 pme_pp->pmeCoordinateReceiverGpu->enqueueWaitReceiveCoordinatesFromPpCudaDirect();
 475             }
 476
 477             status = pmerecvqxX;
 478         }
 479
 480         /* Wait for the coordinates and/or charges to arrive */
 481         MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 482         messages = 0;
 483     } while (status == -1);
 484 #else
 485     GMX_UNUSED_VALUE(pme);
 486     GMX_UNUSED_VALUE(pme_pp);
 487     GMX_UNUSED_VALUE(box);
 488     GMX_UNUSED_VALUE(maxshift_x);
 489     GMX_UNUSED_VALUE(maxshift_y);
 490     GMX_UNUSED_VALUE(lambda_q);
 491     GMX_UNUSED_VALUE(lambda_lj);
 492     GMX_UNUSED_VALUE(bEnerVir);
 493     GMX_UNUSED_VALUE(step);
 494     GMX_UNUSED_VALUE(grid_size);
 495     GMX_UNUSED_VALUE(ewaldcoeff_q);
 496     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 497     GMX_UNUSED_VALUE(useGpuForPme);
 498     GMX_UNUSED_VALUE(stateGpu);
 499
 500     status = pmerecvqxX;
 501 #endif
 502
 503     if (status == pmerecvqxX)
 504     {
 505         *natoms = nat;
 506     }
 507
 508     return status;
 509 }
 510
 511 #if GMX_MPI
 512 /*! \brief Send force data to PP ranks */
 513 static void sendFToPP(void* sendbuf, PpRanks receiver, gmx_pme_pp* pme_pp, int* messages)
 514 {
 515
 516     if (pme_pp->useGpuDirectComm)
 517     {
 518         GMX_ASSERT((pme_pp->pmeForceSenderGpu != nullptr),
 519                    "The use of GPU direct communication for PME-PP is enabled, "
 520                    "but the PME GPU force reciever object does not exist");
 521
 522         pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(receiver.rankId);
 523     }
 524     else
 525     {
 526         // Send using MPI
 527         MPI_Isend(sendbuf, receiver.numAtoms * sizeof(rvec), MPI_BYTE, receiver.rankId, 0,
 528                   pme_pp->mpi_comm_mysim, &pme_pp->req[*messages]);
 529         *messages = *messages + 1;
 530     }
 531 }
 532 #endif
 533
 534 /*! \brief Send the PME mesh force, virial and energy to the PP-only ranks. */
 535 static void gmx_pme_send_force_vir_ener(const gmx_pme_t& pme,
 536                                         gmx_pme_pp*      pme_pp,
 537                                         const PmeOutput& output,
 538                                         real             dvdlambda_q,
 539                                         real             dvdlambda_lj,
 540                                         float            cycles)
 541 {
 542 #if GMX_MPI
 543     gmx_pme_comm_vir_ene_t cve;
 544     int                    messages, ind_start, ind_end;
 545     cve.cycles = cycles;
 546
 547     /* Now the evaluated forces have to be transferred to the PP nodes */
 548     messages = 0;
 549     ind_end  = 0;
 550     for (const auto& receiver : pme_pp->ppRanks)
 551     {
 552         ind_start     = ind_end;
 553         ind_end       = ind_start + receiver.numAtoms;
 554         void* sendbuf = const_cast<void*>(static_cast<const void*>(output.forces_[ind_start]));
 555         if (pme_pp->useGpuDirectComm)
 556         {
 557             // Data will be transferred directly from GPU.
 558             rvec* d_f = reinterpret_cast<rvec*>(pme_gpu_get_device_f(&pme));
 559             sendbuf   = reinterpret_cast<void*>(&d_f[ind_start]);
 560         }
 561         sendFToPP(sendbuf, receiver, pme_pp, &messages);
 562     }
 563
 564     /* send virial and energy to our last PP node */
 565     copy_mat(output.coulombVirial_, cve.vir_q);
 566     copy_mat(output.lennardJonesVirial_, cve.vir_lj);
 567     cve.energy_q     = output.coulombEnergy_;
 568     cve.energy_lj    = output.lennardJonesEnergy_;
 569     cve.dvdlambda_q  = dvdlambda_q;
 570     cve.dvdlambda_lj = dvdlambda_lj;
 571     /* check for the signals to send back to a PP node */
 572     cve.stop_cond = gmx_get_stop_condition();
 573
 574     cve.cycles = cycles;
 575
 576     if (debug)
 577     {
 578         fprintf(debug, "PME rank sending to PP rank %d: virial and energy\n", pme_pp->peerRankId);
 579     }
 580     MPI_Isend(&cve, sizeof(cve), MPI_BYTE, pme_pp->peerRankId, 1, pme_pp->mpi_comm_mysim,
 581               &pme_pp->req[messages++]);
 582
 583     /* Wait for the forces to arrive */
 584     MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 585 #else
 586     gmx_call("MPI not enabled");
 587     GMX_UNUSED_VALUE(pme);
 588     GMX_UNUSED_VALUE(pme_pp);
 589     GMX_UNUSED_VALUE(output);
 590     GMX_UNUSED_VALUE(dvdlambda_q);
 591     GMX_UNUSED_VALUE(dvdlambda_lj);
 592     GMX_UNUSED_VALUE(cycles);
 593 #endif
 594 }
 595
 596 int gmx_pmeonly(struct gmx_pme_t*         pme,
 597                 const t_commrec*          cr,
 598                 t_nrnb*                   mynrnb,
 599                 gmx_wallcycle*            wcycle,
 600                 gmx_walltime_accounting_t walltime_accounting,
 601                 t_inputrec*               ir,
 602                 PmeRunMode                runMode)
 603 {
 604     int      ret;
 605     int      natoms = 0;
 606     matrix   box;
 607     real     lambda_q   = 0;
 608     real     lambda_lj  = 0;
 609     int      maxshift_x = 0, maxshift_y = 0;
 610     real     dvdlambda_q, dvdlambda_lj;
 611     float    cycles;
 612     int      count;
 613     gmx_bool bEnerVir = FALSE;
 614     int64_t  step;
 615
 616     /* This data will only use with PME tuning, i.e. switching PME grids */
 617     std::vector<gmx_pme_t*> pmedata;
 618     pmedata.push_back(pme);
 619
 620     auto pme_pp = gmx_pme_pp_init(cr);
 621     // TODO the variable below should be queried from the task assignment info
 622     const bool  useGpuForPme  = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
 623     const void* commandStream = useGpuForPme ? pme_gpu_get_device_stream(pme) : nullptr;
 624     const void* deviceContext = useGpuForPme ? pme_gpu_get_device_context(pme) : nullptr;
 625     const int   paddingSize   = pme_gpu_get_padding_size(pme);
 626     if (useGpuForPme)
 627     {
 628         changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
 629         changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
 630         if (c_enableGpuPmePpComms)
 631         {
 632             pme_pp->pmeCoordinateReceiverGpu = std::make_unique<gmx::PmeCoordinateReceiverGpu>(
 633                     pme_gpu_get_device_stream(pme), pme_pp->mpi_comm_mysim, pme_pp->ppRanks);
 634             pme_pp->pmeForceSenderGpu = std::make_unique<gmx::PmeForceSenderGpu>(
 635                     pme_gpu_get_device_stream(pme), pme_pp->mpi_comm_mysim, pme_pp->ppRanks);
 636         }
 637     }
 638
 639     std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
 640     if (useGpuForPme)
 641     {
 642         // TODO: Special PME-only constructor is used here. There is no mechanism to prevent from using the other constructor here.
 643         //       This should be made safer.
 644         stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
 645                 commandStream, deviceContext, GpuApiCallBehavior::Async, paddingSize, wcycle);
 646     }
 647
 648
 649     clear_nrnb(mynrnb);
 650
 651     count = 0;
 652     do /****** this is a quasi-loop over time steps! */
 653     {
 654         /* The reason for having a loop here is PME grid tuning/switching */
 655         do
 656         {
 657             /* Domain decomposition */
 658             ivec newGridSize;
 659             real ewaldcoeff_q = 0, ewaldcoeff_lj = 0;
 660             ret = gmx_pme_recv_coeffs_coords(pme, pme_pp.get(), &natoms, box, &maxshift_x,
 661                                              &maxshift_y, &lambda_q, &lambda_lj, &bEnerVir, &step,
 662                                              &newGridSize, &ewaldcoeff_q, &ewaldcoeff_lj,
 663                                              useGpuForPme, stateGpu.get(), runMode);
 664
 665             if (ret == pmerecvqxSWITCHGRID)
 666             {
 667                 /* Switch the PME grid to newGridSize */
 668                 pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
 669             }
 670
 671             if (ret == pmerecvqxRESETCOUNTERS)
 672             {
 673                 /* Reset the cycle and flop counters */
 674                 reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
 675             }
 676         } while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
 677
 678         if (ret == pmerecvqxFINISH)
 679         {
 680             /* We should stop: break out of the loop */
 681             break;
 682         }
 683
 684         if (count == 0)
 685         {
 686             wallcycle_start(wcycle, ewcRUN);
 687             walltime_accounting_start_time(walltime_accounting);
 688         }
 689
 690         wallcycle_start(wcycle, ewcPMEMESH);
 691
 692         dvdlambda_q  = 0;
 693         dvdlambda_lj = 0;
 694
 695         // TODO Make a struct of array refs onto these per-atom fields
 696         // of pme_pp (maybe box, energy and virial, too; and likewise
 697         // from mdatoms for the other call to gmx_pme_do), so we have
 698         // fewer lines of code and less parameter passing.
 699         const int pmeFlags = GMX_PME_DO_ALL_F | (bEnerVir ? GMX_PME_CALC_ENER_VIR : 0);
 700         PmeOutput output;
 701         if (useGpuForPme)
 702         {
 703             const bool boxChanged              = false;
 704             const bool useGpuPmeForceReduction = pme_pp->useGpuDirectComm;
 705             // TODO this should be set properly by gmx_pme_recv_coeffs_coords,
 706             // or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
 707             pme_gpu_prepare_computation(pme, boxChanged, box, wcycle, pmeFlags, useGpuPmeForceReduction);
 708             if (!pme_pp->useGpuDirectComm)
 709             {
 710                 stateGpu->copyCoordinatesToGpu(gmx::ArrayRef<gmx::RVec>(pme_pp->x), gmx::AtomLocality::All);
 711             }
 712             // On the separate PME rank we do not need a synchronizer as we schedule everything in a single stream
 713             // TODO: with pme on GPU the receive should make a list of synchronizers and pass it here #3157
 714             auto xReadyOnDevice = nullptr;
 715
 716             pme_gpu_launch_spread(pme, xReadyOnDevice, wcycle);
 717             pme_gpu_launch_complex_transforms(pme, wcycle);
 718             pme_gpu_launch_gather(pme, wcycle, PmeForceOutputHandling::Set);
 719             output = pme_gpu_wait_finish_task(pme, pmeFlags, wcycle);
 720             pme_gpu_reinit_computation(pme, wcycle);
 721         }
 722         else
 723         {
 724             GMX_ASSERT(pme_pp->x.size() == static_cast<size_t>(natoms),
 725                        "The coordinate buffer should have size natoms");
 726
 727             gmx_pme_do(pme, pme_pp->x, pme_pp->f, pme_pp->chargeA.data(), pme_pp->chargeB.data(),
 728                        pme_pp->sqrt_c6A.data(), pme_pp->sqrt_c6B.data(), pme_pp->sigmaA.data(),
 729                        pme_pp->sigmaB.data(), box, cr, maxshift_x, maxshift_y, mynrnb, wcycle,
 730                        output.coulombVirial_, output.lennardJonesVirial_, &output.coulombEnergy_,
 731                        &output.lennardJonesEnergy_, lambda_q, lambda_lj, &dvdlambda_q,
 732                        &dvdlambda_lj, pmeFlags);
 733             output.forces_ = pme_pp->f;
 734         }
 735
 736         cycles = wallcycle_stop(wcycle, ewcPMEMESH);
 737         gmx_pme_send_force_vir_ener(*pme, pme_pp.get(), output, dvdlambda_q, dvdlambda_lj, cycles);
 738
 739         count++;
 740     } /***** end of quasi-loop, we stop with the break above */
 741     while (TRUE);
 742
 743     walltime_accounting_end_time(walltime_accounting);
 744
 745     return 0;
 746 }