src/gromacs/ewald/pme_pp.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 /*! \internal \file
  39  *
  40  * \brief This file contains function definitions necessary for
  41  * managing the offload of long-ranged PME work to separate MPI rank,
  42  * for computing energies and forces (Coulomb and LJ).
  43  *
  44  * \author Berk Hess <hess@kth.se>
  45  * \ingroup module_ewald
  46  */
  47
  48 #include "gmxpre.h"
  49
  50 #include "pme_pp.h"
  51
  52 #include "config.h"
  53
  54 #include <cstdio>
  55 #include <cstring>
  56
  57 #include "gromacs/domdec/domdec.h"
  58 #include "gromacs/domdec/domdec_struct.h"
  59 #include "gromacs/ewald/pme.h"
  60 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  61 #include "gromacs/gmxlib/network.h"
  62 #include "gromacs/math/vec.h"
  63 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  64 #include "gromacs/mdtypes/commrec.h"
  65 #include "gromacs/mdtypes/forceoutput.h"
  66 #include "gromacs/mdtypes/forcerec.h"
  67 #include "gromacs/mdtypes/interaction_const.h"
  68 #include "gromacs/mdtypes/md_enums.h"
  69 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  70 #include "gromacs/nbnxm/nbnxm.h"
  71 #include "gromacs/timing/wallcycle.h"
  72 #include "gromacs/utility/fatalerror.h"
  73 #include "gromacs/utility/gmxmpi.h"
  74 #include "gromacs/utility/smalloc.h"
  75
  76 #include "pme_pp_communication.h"
  77
  78 /*! \brief Block to wait for communication to PME ranks to complete
  79  *
  80  * This should be faster with a real non-blocking MPI implementation
  81  */
  82 static constexpr bool c_useDelayedWait = false;
  83
  84 /*! \brief Wait for the pending data send requests to PME ranks to complete */
  85 static void gmx_pme_send_coeffs_coords_wait(gmx_domdec_t* dd)
  86 {
  87     if (dd->nreq_pme)
  88     {
  89 #if GMX_MPI
  90         MPI_Waitall(dd->nreq_pme, dd->req_pme, MPI_STATUSES_IGNORE);
  91 #endif
  92         dd->nreq_pme = 0;
  93     }
  94 }
  95
  96 /*! \brief Send data to PME ranks */
  97 static void gmx_pme_send_coeffs_coords(t_forcerec*         fr,
  98                                        const t_commrec*    cr,
  99                                        unsigned int        flags,
 100                                        gmx::ArrayRef<real> chargeA,
 101                                        gmx::ArrayRef<real> chargeB,
 102                                        gmx::ArrayRef<real> c6A,
 103                                        gmx::ArrayRef<real> c6B,
 104                                        gmx::ArrayRef<real> sigmaA,
 105                                        gmx::ArrayRef<real> sigmaB,
 106                                        const matrix        box,
 107                                        const rvec gmx_unused* x,
 108                                        real                   lambda_q,
 109                                        real                   lambda_lj,
 110                                        int                    maxshift_x,
 111                                        int                    maxshift_y,
 112                                        int64_t                step,
 113                                        bool                   useGpuPmePpComms,
 114                                        bool                   reinitGpuPmePpComms,
 115                                        bool                   sendCoordinatesFromGpu,
 116                                        GpuEventSynchronizer*  coordinatesReadyOnDeviceEvent)
 117 {
 118     gmx_domdec_t*         dd;
 119     gmx_pme_comm_n_box_t* cnb;
 120     int                   n;
 121
 122     dd = cr->dd;
 123     n  = dd_numHomeAtoms(*dd);
 124
 125     if (debug)
 126     {
 127         fprintf(debug,
 128                 "PP rank %d sending to PME rank %d: %d%s%s%s%s\n",
 129                 cr->sim_nodeid,
 130                 dd->pme_nodeid,
 131                 n,
 132                 (flags & PP_PME_CHARGE) ? " charges" : "",
 133                 (flags & PP_PME_SQRTC6) ? " sqrtC6" : "",
 134                 (flags & PP_PME_SIGMA) ? " sigma" : "",
 135                 (flags & PP_PME_COORD) ? " coordinates" : "");
 136     }
 137
 138     if (useGpuPmePpComms)
 139     {
 140         flags |= PP_PME_GPUCOMMS;
 141     }
 142
 143     if (c_useDelayedWait)
 144     {
 145         /* We can not use cnb until pending communication has finished */
 146         gmx_pme_send_coeffs_coords_wait(dd);
 147     }
 148
 149     if (dd->pme_receive_vir_ener)
 150     {
 151         /* Peer PP node: communicate all data */
 152         if (dd->cnb == nullptr)
 153         {
 154             snew(dd->cnb, 1);
 155         }
 156         cnb = dd->cnb;
 157
 158         cnb->flags      = flags;
 159         cnb->natoms     = n;
 160         cnb->maxshift_x = maxshift_x;
 161         cnb->maxshift_y = maxshift_y;
 162         cnb->lambda_q   = lambda_q;
 163         cnb->lambda_lj  = lambda_lj;
 164         cnb->step       = step;
 165         if (flags & PP_PME_COORD)
 166         {
 167             copy_mat(box, cnb->box);
 168         }
 169 #if GMX_MPI
 170         MPI_Isend(cnb,
 171                   sizeof(*cnb),
 172                   MPI_BYTE,
 173                   dd->pme_nodeid,
 174                   eCommType_CNB,
 175                   cr->mpi_comm_mysim,
 176                   &dd->req_pme[dd->nreq_pme++]);
 177 #endif
 178     }
 179     else if (flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
 180     {
 181 #if GMX_MPI
 182         /* Communicate only the number of atoms */
 183         MPI_Isend(&n,
 184                   sizeof(n),
 185                   MPI_BYTE,
 186                   dd->pme_nodeid,
 187                   eCommType_CNB,
 188                   cr->mpi_comm_mysim,
 189                   &dd->req_pme[dd->nreq_pme++]);
 190 #endif
 191     }
 192
 193 #if GMX_MPI
 194     if (n > 0)
 195     {
 196         if (flags & PP_PME_CHARGE)
 197         {
 198             MPI_Isend(chargeA.data(),
 199                       n * sizeof(real),
 200                       MPI_BYTE,
 201                       dd->pme_nodeid,
 202                       eCommType_ChargeA,
 203                       cr->mpi_comm_mysim,
 204                       &dd->req_pme[dd->nreq_pme++]);
 205         }
 206         if (flags & PP_PME_CHARGEB)
 207         {
 208             MPI_Isend(chargeB.data(),
 209                       n * sizeof(real),
 210                       MPI_BYTE,
 211                       dd->pme_nodeid,
 212                       eCommType_ChargeB,
 213                       cr->mpi_comm_mysim,
 214                       &dd->req_pme[dd->nreq_pme++]);
 215         }
 216         if (flags & PP_PME_SQRTC6)
 217         {
 218             MPI_Isend(c6A.data(),
 219                       n * sizeof(real),
 220                       MPI_BYTE,
 221                       dd->pme_nodeid,
 222                       eCommType_SQRTC6A,
 223                       cr->mpi_comm_mysim,
 224                       &dd->req_pme[dd->nreq_pme++]);
 225         }
 226         if (flags & PP_PME_SQRTC6B)
 227         {
 228             MPI_Isend(c6B.data(),
 229                       n * sizeof(real),
 230                       MPI_BYTE,
 231                       dd->pme_nodeid,
 232                       eCommType_SQRTC6B,
 233                       cr->mpi_comm_mysim,
 234                       &dd->req_pme[dd->nreq_pme++]);
 235         }
 236         if (flags & PP_PME_SIGMA)
 237         {
 238             MPI_Isend(sigmaA.data(),
 239                       n * sizeof(real),
 240                       MPI_BYTE,
 241                       dd->pme_nodeid,
 242                       eCommType_SigmaA,
 243                       cr->mpi_comm_mysim,
 244                       &dd->req_pme[dd->nreq_pme++]);
 245         }
 246         if (flags & PP_PME_SIGMAB)
 247         {
 248             MPI_Isend(sigmaB.data(),
 249                       n * sizeof(real),
 250                       MPI_BYTE,
 251                       dd->pme_nodeid,
 252                       eCommType_SigmaB,
 253                       cr->mpi_comm_mysim,
 254                       &dd->req_pme[dd->nreq_pme++]);
 255         }
 256         if (flags & PP_PME_COORD)
 257         {
 258             if (reinitGpuPmePpComms)
 259             {
 260                 fr->pmePpCommGpu->reinit(n);
 261             }
 262
 263
 264             /* MPI_Isend does not accept a const buffer pointer */
 265             real* xRealPtr = const_cast<real*>(x[0]);
 266             if (useGpuPmePpComms && (fr != nullptr))
 267             {
 268                 void* sendPtr = sendCoordinatesFromGpu
 269                                         ? static_cast<void*>(fr->stateGpu->getCoordinates())
 270                                         : static_cast<void*>(xRealPtr);
 271                 fr->pmePpCommGpu->sendCoordinatesToPmeCudaDirect(
 272                         sendPtr, n, sendCoordinatesFromGpu, coordinatesReadyOnDeviceEvent);
 273             }
 274             else
 275             {
 276                 MPI_Isend(xRealPtr,
 277                           n * sizeof(rvec),
 278                           MPI_BYTE,
 279                           dd->pme_nodeid,
 280                           eCommType_COORD,
 281                           cr->mpi_comm_mysim,
 282                           &dd->req_pme[dd->nreq_pme++]);
 283             }
 284         }
 285     }
 286 #else
 287     GMX_UNUSED_VALUE(fr);
 288     GMX_UNUSED_VALUE(reinitGpuPmePpComms);
 289     GMX_UNUSED_VALUE(sendCoordinatesFromGpu);
 290     GMX_UNUSED_VALUE(coordinatesReadyOnDeviceEvent);
 291 #endif
 292     if (!c_useDelayedWait)
 293     {
 294         /* Wait for the data to arrive */
 295         /* We can skip this wait as we are sure x and q will not be modified
 296          * before the next call to gmx_pme_send_x_q or gmx_pme_receive_f.
 297          */
 298         gmx_pme_send_coeffs_coords_wait(dd);
 299     }
 300 }
 301
 302 void gmx_pme_send_parameters(const t_commrec*           cr,
 303                              const interaction_const_t& interactionConst,
 304                              bool                       bFreeEnergy_q,
 305                              bool                       bFreeEnergy_lj,
 306                              gmx::ArrayRef<real>        chargeA,
 307                              gmx::ArrayRef<real>        chargeB,
 308                              gmx::ArrayRef<real>        sqrt_c6A,
 309                              gmx::ArrayRef<real>        sqrt_c6B,
 310                              gmx::ArrayRef<real>        sigmaA,
 311                              gmx::ArrayRef<real>        sigmaB,
 312                              int                        maxshift_x,
 313                              int                        maxshift_y)
 314 {
 315     unsigned int flags = 0;
 316
 317     if (EEL_PME(interactionConst.eeltype))
 318     {
 319         flags |= PP_PME_CHARGE;
 320     }
 321     if (EVDW_PME(interactionConst.vdwtype))
 322     {
 323         flags |= (PP_PME_SQRTC6 | PP_PME_SIGMA);
 324     }
 325     if (bFreeEnergy_q || bFreeEnergy_lj)
 326     {
 327         /* Assumes that the B state flags are in the bits just above
 328          * the ones for the A state. */
 329         flags |= (flags << 1);
 330     }
 331
 332     gmx_pme_send_coeffs_coords(nullptr,
 333                                cr,
 334                                flags,
 335                                chargeA,
 336                                chargeB,
 337                                sqrt_c6A,
 338                                sqrt_c6B,
 339                                sigmaA,
 340                                sigmaB,
 341                                nullptr,
 342                                nullptr,
 343                                0,
 344                                0,
 345                                maxshift_x,
 346                                maxshift_y,
 347                                -1,
 348                                false,
 349                                false,
 350                                false,
 351                                nullptr);
 352 }
 353
 354 void gmx_pme_send_coordinates(t_forcerec*           fr,
 355                               const t_commrec*      cr,
 356                               const matrix          box,
 357                               const rvec*           x,
 358                               real                  lambda_q,
 359                               real                  lambda_lj,
 360                               bool                  computeEnergyAndVirial,
 361                               int64_t               step,
 362                               bool                  useGpuPmePpComms,
 363                               bool                  receiveCoordinateAddressFromPme,
 364                               bool                  sendCoordinatesFromGpu,
 365                               GpuEventSynchronizer* coordinatesReadyOnDeviceEvent,
 366                               gmx_wallcycle*        wcycle)
 367 {
 368     wallcycle_start(wcycle, ewcPP_PMESENDX);
 369
 370     unsigned int flags = PP_PME_COORD;
 371     if (computeEnergyAndVirial)
 372     {
 373         flags |= PP_PME_ENER_VIR;
 374     }
 375     gmx_pme_send_coeffs_coords(fr,
 376                                cr,
 377                                flags,
 378                                {},
 379                                {},
 380                                {},
 381                                {},
 382                                {},
 383                                {},
 384                                box,
 385                                x,
 386                                lambda_q,
 387                                lambda_lj,
 388                                0,
 389                                0,
 390                                step,
 391                                useGpuPmePpComms,
 392                                receiveCoordinateAddressFromPme,
 393                                sendCoordinatesFromGpu,
 394                                coordinatesReadyOnDeviceEvent);
 395
 396     wallcycle_stop(wcycle, ewcPP_PMESENDX);
 397 }
 398
 399 void gmx_pme_send_finish(const t_commrec* cr)
 400 {
 401     unsigned int flags = PP_PME_FINISH;
 402
 403     gmx_pme_send_coeffs_coords(
 404             nullptr, cr, flags, {}, {}, {}, {}, {}, {}, nullptr, nullptr, 0, 0, 0, 0, -1, false, false, false, nullptr);
 405 }
 406
 407 void gmx_pme_send_switchgrid(const t_commrec* cr, ivec grid_size, real ewaldcoeff_q, real ewaldcoeff_lj)
 408 {
 409 #if GMX_MPI
 410     gmx_pme_comm_n_box_t cnb;
 411
 412     /* Only let one PP node signal each PME node */
 413     if (cr->dd->pme_receive_vir_ener)
 414     {
 415         cnb.flags = PP_PME_SWITCHGRID;
 416         copy_ivec(grid_size, cnb.grid_size);
 417         cnb.ewaldcoeff_q  = ewaldcoeff_q;
 418         cnb.ewaldcoeff_lj = ewaldcoeff_lj;
 419
 420         /* We send this, uncommon, message blocking to simplify the code */
 421         MPI_Send(&cnb, sizeof(cnb), MPI_BYTE, cr->dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim);
 422     }
 423 #else
 424     GMX_UNUSED_VALUE(cr);
 425     GMX_UNUSED_VALUE(grid_size);
 426     GMX_UNUSED_VALUE(ewaldcoeff_q);
 427     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 428 #endif
 429 }
 430
 431 void gmx_pme_send_resetcounters(const t_commrec gmx_unused* cr, int64_t gmx_unused step)
 432 {
 433 #if GMX_MPI
 434     gmx_pme_comm_n_box_t cnb;
 435
 436     /* Only let one PP node signal each PME node */
 437     if (cr->dd->pme_receive_vir_ener)
 438     {
 439         cnb.flags = PP_PME_RESETCOUNTERS;
 440         cnb.step  = step;
 441
 442         /* We send this, uncommon, message blocking to simplify the code */
 443         MPI_Send(&cnb, sizeof(cnb), MPI_BYTE, cr->dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim);
 444     }
 445 #endif
 446 }
 447
 448 /*! \brief Receive virial and energy from PME rank */
 449 static void receive_virial_energy(const t_commrec*      cr,
 450                                   gmx::ForceWithVirial* forceWithVirial,
 451                                   real*                 energy_q,
 452                                   real*                 energy_lj,
 453                                   real*                 dvdlambda_q,
 454                                   real*                 dvdlambda_lj,
 455                                   float*                pme_cycles)
 456 {
 457     gmx_pme_comm_vir_ene_t cve;
 458
 459     if (cr->dd->pme_receive_vir_ener)
 460     {
 461         if (debug)
 462         {
 463             fprintf(debug,
 464                     "PP rank %d receiving from PME rank %d: virial and energy\n",
 465                     cr->sim_nodeid,
 466                     cr->dd->pme_nodeid);
 467         }
 468 #if GMX_MPI
 469         MPI_Recv(&cve, sizeof(cve), MPI_BYTE, cr->dd->pme_nodeid, 1, cr->mpi_comm_mysim, MPI_STATUS_IGNORE);
 470 #else
 471         memset(&cve, 0, sizeof(cve));
 472 #endif
 473
 474         forceWithVirial->addVirialContribution(cve.vir_q);
 475         forceWithVirial->addVirialContribution(cve.vir_lj);
 476         *energy_q  = cve.energy_q;
 477         *energy_lj = cve.energy_lj;
 478         *dvdlambda_q += cve.dvdlambda_q;
 479         *dvdlambda_lj += cve.dvdlambda_lj;
 480         *pme_cycles = cve.cycles;
 481
 482         if (cve.stop_cond != gmx_stop_cond_none)
 483         {
 484             gmx_set_stop_condition(cve.stop_cond);
 485         }
 486     }
 487     else
 488     {
 489         *energy_q   = 0;
 490         *energy_lj  = 0;
 491         *pme_cycles = 0;
 492     }
 493 }
 494
 495 /*! \brief Recieve force data from PME ranks */
 496 static void recvFFromPme(gmx::PmePpCommGpu* pmePpCommGpu,
 497                          void*              recvptr,
 498                          int                n,
 499                          const t_commrec*   cr,
 500                          bool               useGpuPmePpComms,
 501                          bool               receivePmeForceToGpu)
 502 {
 503     if (useGpuPmePpComms)
 504     {
 505         GMX_ASSERT(pmePpCommGpu != nullptr, "Need valid pmePpCommGpu");
 506         // Receive directly using CUDA memory copy
 507         pmePpCommGpu->receiveForceFromPmeCudaDirect(recvptr, n, receivePmeForceToGpu);
 508     }
 509     else
 510     {
 511         // Receive data using MPI
 512 #if GMX_MPI
 513         MPI_Recv(recvptr, n * sizeof(rvec), MPI_BYTE, cr->dd->pme_nodeid, 0, cr->mpi_comm_mysim, MPI_STATUS_IGNORE);
 514 #else
 515         GMX_UNUSED_VALUE(cr);
 516 #endif
 517     }
 518 }
 519
 520
 521 void gmx_pme_receive_f(gmx::PmePpCommGpu*    pmePpCommGpu,
 522                        const t_commrec*      cr,
 523                        gmx::ForceWithVirial* forceWithVirial,
 524                        real*                 energy_q,
 525                        real*                 energy_lj,
 526                        real*                 dvdlambda_q,
 527                        real*                 dvdlambda_lj,
 528                        bool                  useGpuPmePpComms,
 529                        bool                  receivePmeForceToGpu,
 530                        float*                pme_cycles)
 531 {
 532     if (c_useDelayedWait)
 533     {
 534         /* Wait for the x request to finish */
 535         gmx_pme_send_coeffs_coords_wait(cr->dd);
 536     }
 537
 538     const int               natoms = dd_numHomeAtoms(*cr->dd);
 539     std::vector<gmx::RVec>& buffer = cr->dd->pmeForceReceiveBuffer;
 540     buffer.resize(natoms);
 541
 542     void* recvptr = reinterpret_cast<void*>(buffer.data());
 543     recvFFromPme(pmePpCommGpu, recvptr, natoms, cr, useGpuPmePpComms, receivePmeForceToGpu);
 544
 545     int nt = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, natoms);
 546
 547     gmx::ArrayRef<gmx::RVec> f = forceWithVirial->force_;
 548
 549     if (!receivePmeForceToGpu)
 550     {
 551         /* Note that we would like to avoid this conditional by putting it
 552          * into the omp pragma instead, but then we still take the full
 553          * omp parallel for overhead (at least with gcc5).
 554          */
 555         if (nt == 1)
 556         {
 557             for (int i = 0; i < natoms; i++)
 558             {
 559                 f[i] += buffer[i];
 560             }
 561         }
 562         else
 563         {
 564 #pragma omp parallel for num_threads(nt) schedule(static)
 565             for (int i = 0; i < natoms; i++)
 566             {
 567                 f[i] += buffer[i];
 568             }
 569         }
 570     }
 571
 572     receive_virial_energy(cr, forceWithVirial, energy_q, energy_lj, dvdlambda_q, dvdlambda_lj, pme_cycles);
 573 }