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) 2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 /* IMPORTANT FOR DEVELOPERS:
  38  *
  39  * Triclinic pme stuff isn't entirely trivial, and we've experienced
  40  * some bugs during development (many of them due to me). To avoid
  41  * this in the future, please check the following things if you make
  42  * changes in this file:
  43  *
  44  * 1. You should obtain identical (at least to the PME precision)
  45  *    energies, forces, and virial for
  46  *    a rectangular box and a triclinic one where the z (or y) axis is
  47  *    tilted a whole box side. For instance you could use these boxes:
  48  *
  49  *    rectangular       triclinic
  50  *     2  0  0           2  0  0
  51  *     0  2  0           0  2  0
  52  *     0  0  6           2  2  6
  53  *
  54  * 2. You should check the energy conservation in a triclinic box.
  55  *
  56  * It might seem an overkill, but better safe than sorry.
  57  * /Erik 001109
  58  */
  59
  60 #include "gmxpre.h"
  61
  62 #include "config.h"
  63
  64 #include <cassert>
  65 #include <cmath>
  66 #include <cstdio>
  67 #include <cstdlib>
  68 #include <cstring>
  69
  70 #include <memory>
  71 #include <numeric>
  72 #include <vector>
  73
  74 #include "gromacs/domdec/domdec.h"
  75 #include "gromacs/ewald/pme.h"
  76 #include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
  77 #include "gromacs/ewald/pme_force_sender_gpu.h"
  78 #include "gromacs/fft/parallel_3dfft.h"
  79 #include "gromacs/fileio/pdbio.h"
  80 #include "gromacs/gmxlib/network.h"
  81 #include "gromacs/gmxlib/nrnb.h"
  82 #include "gromacs/gpu_utils/hostallocator.h"
  83 #include "gromacs/math/gmxcomplex.h"
  84 #include "gromacs/math/units.h"
  85 #include "gromacs/math/vec.h"
  86 #include "gromacs/mdtypes/commrec.h"
  87 #include "gromacs/mdtypes/forceoutput.h"
  88 #include "gromacs/mdtypes/inputrec.h"
  89 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  90 #include "gromacs/timing/cyclecounter.h"
  91 #include "gromacs/timing/wallcycle.h"
  92 #include "gromacs/utility/fatalerror.h"
  93 #include "gromacs/utility/futil.h"
  94 #include "gromacs/utility/gmxmpi.h"
  95 #include "gromacs/utility/gmxomp.h"
  96 #include "gromacs/utility/smalloc.h"
  97
  98 #include "pme_gpu_internal.h"
  99 #include "pme_internal.h"
 100 #include "pme_pp_communication.h"
 101
 102 /*! \brief environment variable to enable GPU P2P communication */
 103 static const bool c_enableGpuPmePpComms = (getenv("GMX_GPU_PME_PP_COMMS") != nullptr)
 104     && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA);
 105
 106 /*! \brief Master PP-PME communication data structure */
 107 struct gmx_pme_pp {
 108     MPI_Comm             mpi_comm_mysim; /**< MPI communicator for this simulation */
 109     std::vector<PpRanks> ppRanks;        /**< The PP partner ranks                 */
 110     int                  peerRankId;     /**< The peer PP rank id                  */
 111     //@{
 112     /**< Vectors of A- and B-state parameters used to transfer vectors to PME ranks  */
 113     gmx::PaddedHostVector<real> chargeA;
 114     std::vector<real>           chargeB;
 115     std::vector<real>           sqrt_c6A;
 116     std::vector<real>           sqrt_c6B;
 117     std::vector<real>           sigmaA;
 118     std::vector<real>           sigmaB;
 119     //@}
 120     gmx::HostVector<gmx::RVec>  x; /**< Vector of atom coordinates to transfer to PME ranks */
 121     std::vector<gmx::RVec>      f; /**< Vector of atom forces received from PME ranks */
 122     //@{
 123     /**< Vectors of MPI objects used in non-blocking communication between multiple PP ranks per PME rank */
 124     std::vector<MPI_Request> req;
 125     std::vector<MPI_Status>  stat;
 126     //@}
 127
 128     /*! \brief object for receiving coordinates using communications operating on GPU memory space */
 129     std::unique_ptr<gmx::PmeCoordinateReceiverGpu> pmeCoordinateReceiverGpu;
 130     /*! \brief object for sending PME force using communications operating on GPU memory space */
 131     std::unique_ptr<gmx::PmeForceSenderGpu>        pmeForceSenderGpu;
 132
 133     /*! \brief whether GPU direct communications are active for PME-PP transfers */
 134     bool useGpuDirectComm = false;
 135 };
 136
 137 /*! \brief Initialize the PME-only side of the PME <-> PP communication */
 138 static std::unique_ptr<gmx_pme_pp> gmx_pme_pp_init(const t_commrec *cr)
 139 {
 140     auto pme_pp = std::make_unique<gmx_pme_pp>();
 141
 142 #if GMX_MPI
 143     int rank;
 144
 145     pme_pp->mpi_comm_mysim = cr->mpi_comm_mysim;
 146     MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
 147     auto ppRanks = get_pme_ddranks(cr, rank);
 148     pme_pp->ppRanks.reserve(ppRanks.size());
 149     for (const auto &ppRankId : ppRanks)
 150     {
 151         pme_pp->ppRanks.push_back({ppRankId, 0});
 152     }
 153     // The peer PP rank is the last one.
 154     pme_pp->peerRankId = pme_pp->ppRanks.back().rankId;
 155     pme_pp->req.resize(eCommType_NR*pme_pp->ppRanks.size());
 156     pme_pp->stat.resize(eCommType_NR*pme_pp->ppRanks.size());
 157 #else
 158     GMX_UNUSED_VALUE(cr);
 159 #endif
 160
 161     return pme_pp;
 162 }
 163
 164 static void reset_pmeonly_counters(gmx_wallcycle_t           wcycle,
 165                                    gmx_walltime_accounting_t walltime_accounting,
 166                                    t_nrnb                   *nrnb,
 167                                    int64_t                   step,
 168                                    bool                      useGpuForPme)
 169 {
 170     /* Reset all the counters related to performance over the run */
 171     wallcycle_stop(wcycle, ewcRUN);
 172     wallcycle_reset_all(wcycle);
 173     *nrnb = { 0 };
 174     wallcycle_start(wcycle, ewcRUN);
 175     walltime_accounting_reset_time(walltime_accounting, step);
 176
 177     if (useGpuForPme)
 178     {
 179         resetGpuProfiler();
 180     }
 181 }
 182
 183 static gmx_pme_t *gmx_pmeonly_switch(std::vector<gmx_pme_t *> *pmedata,
 184                                      const ivec grid_size,
 185                                      real ewaldcoeff_q, real ewaldcoeff_lj,
 186                                      const t_commrec *cr, const t_inputrec *ir)
 187 {
 188     GMX_ASSERT(pmedata, "Bad PME tuning list pointer");
 189     for (auto &pme : *pmedata)
 190     {
 191         GMX_ASSERT(pme, "Bad PME tuning list element pointer");
 192         if (pme->nkx == grid_size[XX] &&
 193             pme->nky == grid_size[YY] &&
 194             pme->nkz == grid_size[ZZ])
 195         {
 196             /* Here we have found an existing PME data structure that suits us.
 197              * However, in the GPU case, we have to reinitialize it - there's only one GPU structure.
 198              * This should not cause actual GPU reallocations, at least (the allocated buffers are never shrunk).
 199              * So, just some grid size updates in the GPU kernel parameters.
 200              * TODO: this should be something like gmx_pme_update_split_params()
 201              */
 202             gmx_pme_reinit(&pme, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 203             return pme;
 204         }
 205     }
 206
 207     const auto &pme          = pmedata->back();
 208     gmx_pme_t  *newStructure = nullptr;
 209     // Copy last structure with new grid params
 210     gmx_pme_reinit(&newStructure, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 211     pmedata->push_back(newStructure);
 212     return newStructure;
 213 }
 214
 215 /*! \brief Called by PME-only ranks to receive coefficients and coordinates
 216  *
 217  * \param[in] pme           PME data structure.
 218  * \param[in,out] pme_pp    PME-PP communication structure.
 219  * \param[out] natoms       Number of received atoms.
 220  * \param[out] box        System box, if received.
 221  * \param[out] maxshift_x        Maximum shift in X direction, if received.
 222  * \param[out] maxshift_y        Maximum shift in Y direction, if received.
 223  * \param[out] lambda_q         Free-energy lambda for electrostatics, if received.
 224  * \param[out] lambda_lj         Free-energy lambda for Lennard-Jones, if received.
 225  * \param[out] bEnerVir          Set to true if this is an energy/virial calculation step, otherwise set to false.
 226  * \param[out] step              MD integration step number.
 227  * \param[out] grid_size         PME grid size, if received.
 228  * \param[out] ewaldcoeff_q         Ewald cut-off parameter for electrostatics, if received.
 229  * \param[out] ewaldcoeff_lj         Ewald cut-off parameter for Lennard-Jones, if received.
 230  * \param[in] useGpuForPme      flag on whether PME is on GPU
 231  * \param[in] stateGpu          GPU state propagator object
 232  * \param[in] runMode           PME run mode
 233  *
 234  * \retval pmerecvqxX             All parameters were set, chargeA and chargeB can be NULL.
 235  * \retval pmerecvqxFINISH        No parameters were set.
 236  * \retval pmerecvqxSWITCHGRID    Only grid_size and *ewaldcoeff were set.
 237  * \retval pmerecvqxRESETCOUNTERS *step was set.
 238  */
 239 static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t            *pme,
 240                                       gmx_pme_pp                  *pme_pp,
 241                                       int                         *natoms,
 242                                       matrix                       box,
 243                                       int                         *maxshift_x,
 244                                       int                         *maxshift_y,
 245                                       real                        *lambda_q,
 246                                       real                        *lambda_lj,
 247                                       gmx_bool                    *bEnerVir,
 248                                       int64_t                     *step,
 249                                       ivec                        *grid_size,
 250                                       real                        *ewaldcoeff_q,
 251                                       real                        *ewaldcoeff_lj,
 252                                       bool                         useGpuForPme,
 253                                       gmx::StatePropagatorDataGpu *stateGpu,
 254                                       PmeRunMode gmx_unused        runMode)
 255 {
 256     int status = -1;
 257     int nat    = 0;
 258
 259 #if GMX_MPI
 260     unsigned int flags          = 0;
 261     int          messages       = 0;
 262     bool         atomSetChanged = false;
 263
 264     do
 265     {
 266         gmx_pme_comm_n_box_t cnb;
 267         cnb.flags = 0;
 268
 269         /* Receive the send count, box and time step from the peer PP node */
 270         MPI_Recv(&cnb, sizeof(cnb), MPI_BYTE,
 271                  pme_pp->peerRankId, eCommType_CNB,
 272                  pme_pp->mpi_comm_mysim, MPI_STATUS_IGNORE);
 273
 274         /* We accumulate all received flags */
 275         flags |= cnb.flags;
 276
 277         *step  = cnb.step;
 278
 279         if (debug)
 280         {
 281             fprintf(debug, "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
 294         if (cnb.flags & PP_PME_FINISH)
 295         {
 296             status = pmerecvqxFINISH;
 297         }
 298
 299         if (cnb.flags & PP_PME_SWITCHGRID)
 300         {
 301             /* Special case, receive the new parameters and return */
 302             copy_ivec(cnb.grid_size, *grid_size);
 303             *ewaldcoeff_q  = cnb.ewaldcoeff_q;
 304             *ewaldcoeff_lj = cnb.ewaldcoeff_lj;
 305
 306             status         = pmerecvqxSWITCHGRID;
 307         }
 308
 309         if (cnb.flags & PP_PME_RESETCOUNTERS)
 310         {
 311             /* Special case, receive the step (set above) and return */
 312             status = pmerecvqxRESETCOUNTERS;
 313         }
 314
 315         if (cnb.flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
 316         {
 317             atomSetChanged = true;
 318
 319             /* Receive the send counts from the other PP nodes */
 320             for (auto &sender : pme_pp->ppRanks)
 321             {
 322                 if (sender.rankId == pme_pp->peerRankId)
 323                 {
 324                     sender.numAtoms = cnb.natoms;
 325                 }
 326                 else
 327                 {
 328                     MPI_Irecv(&sender.numAtoms, sizeof(sender.numAtoms),
 329                               MPI_BYTE,
 330                               sender.rankId, eCommType_CNB,
 331                               pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 332                 }
 333             }
 334             MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 335             messages = 0;
 336
 337             nat = 0;
 338             for (const auto &sender : pme_pp->ppRanks)
 339             {
 340                 nat += sender.numAtoms;
 341             }
 342
 343             if (cnb.flags & PP_PME_CHARGE)
 344             {
 345                 pme_pp->chargeA.resizeWithPadding(nat);
 346             }
 347             if (cnb.flags & PP_PME_CHARGEB)
 348             {
 349                 pme_pp->chargeB.resize(nat);
 350             }
 351             if (cnb.flags & PP_PME_SQRTC6)
 352             {
 353                 pme_pp->sqrt_c6A.resize(nat);
 354             }
 355             if (cnb.flags & PP_PME_SQRTC6B)
 356             {
 357                 pme_pp->sqrt_c6B.resize(nat);
 358             }
 359             if (cnb.flags & PP_PME_SIGMA)
 360             {
 361                 pme_pp->sigmaA.resize(nat);
 362             }
 363             if (cnb.flags & PP_PME_SIGMAB)
 364             {
 365                 pme_pp->sigmaB.resize(nat);
 366             }
 367             pme_pp->x.resize(nat);
 368             pme_pp->f.resize(nat);
 369
 370             /* maxshift is sent when the charges are sent */
 371             *maxshift_x = cnb.maxshift_x;
 372             *maxshift_y = cnb.maxshift_y;
 373
 374             /* Receive the charges in place */
 375             for (int q = 0; q < eCommType_NR; q++)
 376             {
 377                 real *bufferPtr;
 378
 379                 if (!(cnb.flags & (PP_PME_CHARGE<<q)))
 380                 {
 381                     continue;
 382                 }
 383                 switch (q)
 384                 {
 385                     case eCommType_ChargeA: bufferPtr = pme_pp->chargeA.data();  break;
 386                     case eCommType_ChargeB: bufferPtr = pme_pp->chargeB.data();  break;
 387                     case eCommType_SQRTC6A: bufferPtr = pme_pp->sqrt_c6A.data(); break;
 388                     case eCommType_SQRTC6B: bufferPtr = pme_pp->sqrt_c6B.data(); break;
 389                     case eCommType_SigmaA:  bufferPtr = pme_pp->sigmaA.data();   break;
 390                     case eCommType_SigmaB:  bufferPtr = pme_pp->sigmaB.data();   break;
 391                     default: gmx_incons("Wrong eCommType");
 392                 }
 393                 nat = 0;
 394                 for (const auto &sender : pme_pp->ppRanks)
 395                 {
 396                     if (sender.numAtoms > 0)
 397                     {
 398                         MPI_Irecv(bufferPtr+nat,
 399                                   sender.numAtoms*sizeof(real),
 400                                   MPI_BYTE,
 401                                   sender.rankId, q,
 402                                   pme_pp->mpi_comm_mysim,
 403                                   &pme_pp->req[messages++]);
 404                         nat += sender.numAtoms;
 405                         if (debug)
 406                         {
 407                             fprintf(debug, "Received from PP rank %d: %d %s\n",
 408                                     sender.rankId, sender.numAtoms,
 409                                     (q == eCommType_ChargeA ||
 410                                      q == eCommType_ChargeB) ? "charges" : "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());
 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, "GPU Direct PME-PP communication has been enabled, "
 430                                "but PME run mode is not PmeRunMode::GPU\n");
 431
 432                     // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses.
 433                     rvec* d_x = nullptr;
 434                     rvec* d_f = nullptr;
 435 #if (GMX_GPU == GMX_GPU_CUDA) //avoid invalid cast for OpenCL
 436                     d_x = reinterpret_cast<rvec*> (pme_gpu_get_device_x(pme));
 437                     d_f = reinterpret_cast<rvec*> (pme_gpu_get_device_f(pme));
 438 #endif
 439                     pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(d_x);
 440                     pme_pp->pmeForceSenderGpu->sendForceBufferAddressToPpRanks(d_f);
 441                 }
 442             }
 443
 444
 445             /* The box, FE flag and lambda are sent along with the coordinates
 446              *  */
 447             copy_mat(cnb.box, box);
 448             *lambda_q       = cnb.lambda_q;
 449             *lambda_lj      = cnb.lambda_lj;
 450             *bEnerVir       = ((cnb.flags & PP_PME_ENER_VIR) != 0U);
 451             *step           = cnb.step;
 452
 453             /* Receive the coordinates in place */
 454             nat = 0;
 455             for (const auto &sender : pme_pp->ppRanks)
 456             {
 457                 if (sender.numAtoms > 0)
 458                 {
 459                     if (pme_pp->useGpuDirectComm)
 460                     {
 461                         pme_pp->pmeCoordinateReceiverGpu->receiveCoordinatesFromPpCudaDirect(sender.rankId);
 462                     }
 463                     else
 464                     {
 465                         MPI_Irecv(pme_pp->x[nat],
 466                                   sender.numAtoms*sizeof(rvec),
 467                                   MPI_BYTE,
 468                                   sender.rankId, eCommType_COORD,
 469                                   pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 470                     }
 471                     nat += sender.numAtoms;
 472                     if (debug)
 473                     {
 474                         fprintf(debug, "Received from PP rank %d: %d "
 475                                 "coordinates\n",
 476                                 sender.rankId, sender.numAtoms);
 477                     }
 478                 }
 479             }
 480
 481             status = pmerecvqxX;
 482         }
 483
 484         /* Wait for the coordinates and/or charges to arrive */
 485         MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 486         messages = 0;
 487     }
 488     while (status == -1);
 489 #else
 490     GMX_UNUSED_VALUE(pme_pp);
 491     GMX_UNUSED_VALUE(box);
 492     GMX_UNUSED_VALUE(maxshift_x);
 493     GMX_UNUSED_VALUE(maxshift_y);
 494     GMX_UNUSED_VALUE(lambda_q);
 495     GMX_UNUSED_VALUE(lambda_lj);
 496     GMX_UNUSED_VALUE(bEnerVir);
 497     GMX_UNUSED_VALUE(step);
 498     GMX_UNUSED_VALUE(grid_size);
 499     GMX_UNUSED_VALUE(ewaldcoeff_q);
 500     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 501
 502     status = pmerecvqxX;
 503 #endif
 504
 505     if (status == pmerecvqxX)
 506     {
 507         *natoms   = nat;
 508     }
 509
 510     return status;
 511 }
 512
 513 /*! \brief Send force data to PP ranks */
 514 static void sendFToPP(void* sendbuf, PpRanks receiver, gmx_pme_pp *pme_pp, int *messages)
 515 {
 516
 517     if (pme_pp->useGpuDirectComm)
 518     {
 519         GMX_ASSERT((pme_pp->pmeForceSenderGpu != nullptr),
 520                    "The use of GPU direct communication for PME-PP is enabled, "
 521                    "but the PME GPU force reciever object does not exist" );
 522
 523         pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(receiver.rankId);
 524     }
 525     else
 526     {
 527 #if GMX_MPI
 528         //Send using MPI
 529         MPI_Isend(sendbuf, receiver.numAtoms*sizeof(rvec), MPI_BYTE, receiver.rankId,
 530                   0, pme_pp->mpi_comm_mysim, &pme_pp->req[*messages]);
 531         *messages = *messages+1;
 532 #endif
 533     }
 534
 535 }
 536
 537 /*! \brief Send the PME mesh force, virial and energy to the PP-only ranks. */
 538 static void gmx_pme_send_force_vir_ener(const gmx_pme_t &pme,
 539                                         gmx_pme_pp *pme_pp,
 540                                         const PmeOutput &output,
 541                                         real dvdlambda_q, 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         void *sendbuf = const_cast<void *>(static_cast<const void *>(output.forces_[ind_start]));
 557         if (pme_pp->useGpuDirectComm)
 558         {
 559             //Data will be transferred directly from GPU.
 560             rvec* d_f = reinterpret_cast<rvec*> (pme_gpu_get_device_f(&pme));
 561             sendbuf = reinterpret_cast<void*> (&d_f[ind_start]);
 562         }
 563         sendFToPP(sendbuf, receiver, pme_pp, &messages);
 564     }
 565
 566     /* send virial and energy to our last PP node */
 567     copy_mat(output.coulombVirial_, cve.vir_q);
 568     copy_mat(output.lennardJonesVirial_, cve.vir_lj);
 569     cve.energy_q     = output.coulombEnergy_;
 570     cve.energy_lj    = output.lennardJonesEnergy_;
 571     cve.dvdlambda_q  = dvdlambda_q;
 572     cve.dvdlambda_lj = dvdlambda_lj;
 573     /* check for the signals to send back to a PP node */
 574     cve.stop_cond = gmx_get_stop_condition();
 575
 576     cve.cycles = cycles;
 577
 578     if (debug)
 579     {
 580         fprintf(debug, "PME rank sending to PP rank %d: virial and energy\n",
 581                 pme_pp->peerRankId);
 582     }
 583     MPI_Isend(&cve, sizeof(cve), MPI_BYTE,
 584               pme_pp->peerRankId, 1,
 585               pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 586
 587     /* Wait for the forces to arrive */
 588     MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 589 #else
 590     gmx_call("MPI not enabled");
 591     GMX_UNUSED_VALUE(pme_pp);
 592     GMX_UNUSED_VALUE(output);
 593     GMX_UNUSED_VALUE(dvdlambda_q);
 594     GMX_UNUSED_VALUE(dvdlambda_lj);
 595     GMX_UNUSED_VALUE(cycles);
 596 #endif
 597 }
 598
 599 int gmx_pmeonly(struct gmx_pme_t *pme,
 600                 const t_commrec *cr, t_nrnb *mynrnb,
 601                 gmx_wallcycle  *wcycle,
 602                 gmx_walltime_accounting_t walltime_accounting,
 603                 t_inputrec *ir, PmeRunMode runMode)
 604 {
 605     int                ret;
 606     int                natoms = 0;
 607     matrix             box;
 608     real               lambda_q   = 0;
 609     real               lambda_lj  = 0;
 610     int                maxshift_x = 0, maxshift_y = 0;
 611     real               dvdlambda_q, dvdlambda_lj;
 612     float              cycles;
 613     int                count;
 614     gmx_bool           bEnerVir = FALSE;
 615     int64_t            step;
 616
 617     /* This data will only use with PME tuning, i.e. switching PME grids */
 618     std::vector<gmx_pme_t *> pmedata;
 619     pmedata.push_back(pme);
 620
 621     auto        pme_pp       = gmx_pme_pp_init(cr);
 622     //TODO the variable below should be queried from the task assignment info
 623     const bool  useGpuForPme   = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
 624     const void *commandStream  = useGpuForPme ? pme_gpu_get_device_stream(pme) : nullptr;
 625     const void *deviceContext  = useGpuForPme ? pme_gpu_get_device_context(pme) : nullptr;
 626     const int   paddingSize    = pme_gpu_get_padding_size(pme);
 627     if (useGpuForPme)
 628     {
 629         changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
 630         changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
 631         if (c_enableGpuPmePpComms)
 632         {
 633             pme_pp->pmeCoordinateReceiverGpu =
 634                 std::make_unique<gmx::PmeCoordinateReceiverGpu>(pme_gpu_get_device_stream(pme),
 635                                                                 pme_pp->mpi_comm_mysim,
 636                                                                 pme_pp->ppRanks);
 637             pme_pp->pmeForceSenderGpu =
 638                 std::make_unique<gmx::PmeForceSenderGpu>(pme_gpu_get_device_stream(pme),
 639                                                          pme_pp->mpi_comm_mysim,
 640                                                          pme_pp->ppRanks);
 641         }
 642     }
 643
 644     std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
 645     if (useGpuForPme)
 646     {
 647         // TODO: Special PME-only constructor is used here. There is no mechanism to prevent from using the other constructor here.
 648         //       This should be made safer.
 649         stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(commandStream, deviceContext, GpuApiCallBehavior::Async, paddingSize);
 650     }
 651
 652
 653     clear_nrnb(mynrnb);
 654
 655     count = 0;
 656     do /****** this is a quasi-loop over time steps! */
 657     {
 658         /* The reason for having a loop here is PME grid tuning/switching */
 659         do
 660         {
 661             /* Domain decomposition */
 662             ivec newGridSize;
 663             real ewaldcoeff_q   = 0, ewaldcoeff_lj = 0;
 664             ret = gmx_pme_recv_coeffs_coords(pme,
 665                                              pme_pp.get(),
 666                                              &natoms,
 667                                              box,
 668                                              &maxshift_x, &maxshift_y,
 669                                              &lambda_q, &lambda_lj,
 670                                              &bEnerVir,
 671                                              &step,
 672                                              &newGridSize,
 673                                              &ewaldcoeff_q,
 674                                              &ewaldcoeff_lj,
 675                                              useGpuForPme,
 676                                              stateGpu.get(),
 677                                              runMode);
 678
 679             if (ret == pmerecvqxSWITCHGRID)
 680             {
 681                 /* Switch the PME grid to newGridSize */
 682                 pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
 683             }
 684
 685             if (ret == pmerecvqxRESETCOUNTERS)
 686             {
 687                 /* Reset the cycle and flop counters */
 688                 reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
 689             }
 690         }
 691         while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
 692
 693         if (ret == pmerecvqxFINISH)
 694         {
 695             /* We should stop: break out of the loop */
 696             break;
 697         }
 698
 699         if (count == 0)
 700         {
 701             wallcycle_start(wcycle, ewcRUN);
 702             walltime_accounting_start_time(walltime_accounting);
 703         }
 704
 705         wallcycle_start(wcycle, ewcPMEMESH);
 706
 707         dvdlambda_q  = 0;
 708         dvdlambda_lj = 0;
 709
 710         // TODO Make a struct of array refs onto these per-atom fields
 711         // of pme_pp (maybe box, energy and virial, too; and likewise
 712         // from mdatoms for the other call to gmx_pme_do), so we have
 713         // fewer lines of code and less parameter passing.
 714         const int pmeFlags = GMX_PME_DO_ALL_F | (bEnerVir ? GMX_PME_CALC_ENER_VIR : 0);
 715         PmeOutput output;
 716         if (useGpuForPme)
 717         {
 718             const bool boxChanged              = false;
 719             const bool useGpuPmeForceReduction = pme_pp->useGpuDirectComm;
 720             //TODO this should be set properly by gmx_pme_recv_coeffs_coords,
 721             // or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
 722             pme_gpu_prepare_computation(pme, boxChanged, box, wcycle, pmeFlags, useGpuPmeForceReduction);
 723             if (!pme_pp->useGpuDirectComm)
 724             {
 725                 stateGpu->copyCoordinatesToGpu(gmx::ArrayRef<gmx::RVec>(pme_pp->x), gmx::StatePropagatorDataGpu::AtomLocality::All);
 726             }
 727             // On the separate PME rank we do not need a synchronizer as we schedule everything in a single stream
 728             // TODO: with pme on GPU the receive should make a list of synchronizers and pass it here #3157
 729             auto xReadyOnDevice = nullptr;
 730
 731             pme_gpu_launch_spread(pme, xReadyOnDevice, wcycle);
 732             pme_gpu_launch_complex_transforms(pme, wcycle);
 733             pme_gpu_launch_gather(pme, wcycle, PmeForceOutputHandling::Set);
 734             output = pme_gpu_wait_finish_task(pme, pmeFlags, wcycle);
 735             pme_gpu_reinit_computation(pme, wcycle);
 736         }
 737         else
 738         {
 739             GMX_ASSERT(pme_pp->x.size() == static_cast<size_t>(natoms), "The coordinate buffer should have size natoms");
 740
 741             gmx_pme_do(pme, pme_pp->x, pme_pp->f,
 742                        pme_pp->chargeA.data(), pme_pp->chargeB.data(),
 743                        pme_pp->sqrt_c6A.data(), pme_pp->sqrt_c6B.data(),
 744                        pme_pp->sigmaA.data(), pme_pp->sigmaB.data(), box,
 745                        cr, maxshift_x, maxshift_y, mynrnb, wcycle,
 746                        output.coulombVirial_, output.lennardJonesVirial_,
 747                        &output.coulombEnergy_, &output.lennardJonesEnergy_,
 748                        lambda_q, lambda_lj, &dvdlambda_q, &dvdlambda_lj,
 749                        pmeFlags);
 750             output.forces_ = pme_pp->f;
 751         }
 752
 753         cycles = wallcycle_stop(wcycle, ewcPMEMESH);
 754         gmx_pme_send_force_vir_ener(*pme, pme_pp.get(), output,
 755                                     dvdlambda_q, dvdlambda_lj, cycles);
 756
 757         count++;
 758     } /***** end of quasi-loop, we stop with the break above */
 759     while (TRUE);
 760
 761     walltime_accounting_end_time(walltime_accounting);
 762
 763     return 0;
 764 }