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) 2013,2014,2015,2016,2017,2018,2019, 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 /*! \internal \file
  38  *
  39  * \brief This file contains function definitions necessary for
  40  * managing the offload of long-ranged PME work to separate MPI rank,
  41  * for computing energies and forces (Coulomb and LJ).
  42  *
  43  * \author Berk Hess <hess@kth.se>
  44  * \ingroup module_ewald
  45  */
  46
  47 #include "gmxpre.h"
  48
  49 #include "config.h"
  50
  51 #include <cstdio>
  52 #include <cstring>
  53
  54 #include "gromacs/domdec/domdec.h"
  55 #include "gromacs/domdec/domdec_struct.h"
  56 #include "gromacs/ewald/pme.h"
  57 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  58 #include "gromacs/gmxlib/network.h"
  59 #include "gromacs/math/vec.h"
  60 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  61 #include "gromacs/mdtypes/commrec.h"
  62 #include "gromacs/mdtypes/forceoutput.h"
  63 #include "gromacs/mdtypes/forcerec.h"
  64 #include "gromacs/mdtypes/interaction_const.h"
  65 #include "gromacs/mdtypes/md_enums.h"
  66 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  67 #include "gromacs/nbnxm/nbnxm.h"
  68 #include "gromacs/timing/wallcycle.h"
  69 #include "gromacs/utility/fatalerror.h"
  70 #include "gromacs/utility/gmxmpi.h"
  71 #include "gromacs/utility/smalloc.h"
  72
  73 #include "pme_internal.h"
  74 #include "pme_pp_communication.h"
  75
  76 /*! \brief Block to wait for communication to PME ranks to complete
  77  *
  78  * This should be faster with a real non-blocking MPI implementation
  79  */
  80 static constexpr bool c_useDelayedWait = false;
  81
  82 /*! \brief Wait for the pending data send requests to PME ranks to complete */
  83 static void gmx_pme_send_coeffs_coords_wait(gmx_domdec_t* dd)
  84 {
  85     if (dd->nreq_pme)
  86     {
  87 #if GMX_MPI
  88         MPI_Waitall(dd->nreq_pme, dd->req_pme, MPI_STATUSES_IGNORE);
  89 #endif
  90         dd->nreq_pme = 0;
  91     }
  92 }
  93
  94 /*! \brief Send data to PME ranks */
  95 static void gmx_pme_send_coeffs_coords(t_forcerec*      fr,
  96                                        const t_commrec* cr,
  97                                        unsigned int     flags,
  98                                        real gmx_unused* chargeA,
  99                                        real gmx_unused* chargeB,
 100                                        real gmx_unused* c6A,
 101                                        real gmx_unused* c6B,
 102                                        real gmx_unused* sigmaA,
 103                                        real gmx_unused* sigmaB,
 104                                        const matrix     box,
 105                                        const rvec gmx_unused* x,
 106                                        real                   lambda_q,
 107                                        real                   lambda_lj,
 108                                        int                    maxshift_x,
 109                                        int                    maxshift_y,
 110                                        int64_t                step,
 111                                        bool                   useGpuPmePpComms,
 112                                        bool                   reinitGpuPmePpComms,
 113                                        bool                   sendCoordinatesFromGpu,
 114                                        GpuEventSynchronizer*  coordinatesReadyOnDeviceEvent)
 115 {
 116     gmx_domdec_t*         dd;
 117     gmx_pme_comm_n_box_t* cnb;
 118     int                   n;
 119
 120     dd = cr->dd;
 121     n  = dd_numHomeAtoms(*dd);
 122
 123     if (debug)
 124     {
 125         fprintf(debug, "PP rank %d sending to PME rank %d: %d%s%s%s%s\n", cr->sim_nodeid, dd->pme_nodeid,
 126                 n, (flags & PP_PME_CHARGE) ? " charges" : "", (flags & PP_PME_SQRTC6) ? " sqrtC6" : "",
 127                 (flags & PP_PME_SIGMA) ? " sigma" : "", (flags & PP_PME_COORD) ? " coordinates" : "");
 128     }
 129
 130     if (useGpuPmePpComms)
 131     {
 132         flags |= PP_PME_GPUCOMMS;
 133     }
 134
 135     if (c_useDelayedWait)
 136     {
 137         /* We can not use cnb until pending communication has finished */
 138         gmx_pme_send_coeffs_coords_wait(dd);
 139     }
 140
 141     if (dd->pme_receive_vir_ener)
 142     {
 143         /* Peer PP node: communicate all data */
 144         if (dd->cnb == nullptr)
 145         {
 146             snew(dd->cnb, 1);
 147         }
 148         cnb = dd->cnb;
 149
 150         cnb->flags      = flags;
 151         cnb->natoms     = n;
 152         cnb->maxshift_x = maxshift_x;
 153         cnb->maxshift_y = maxshift_y;
 154         cnb->lambda_q   = lambda_q;
 155         cnb->lambda_lj  = lambda_lj;
 156         cnb->step       = step;
 157         if (flags & PP_PME_COORD)
 158         {
 159             copy_mat(box, cnb->box);
 160         }
 161 #if GMX_MPI
 162         MPI_Isend(cnb, sizeof(*cnb), MPI_BYTE, dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim,
 163                   &dd->req_pme[dd->nreq_pme++]);
 164 #endif
 165     }
 166     else if (flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
 167     {
 168 #if GMX_MPI
 169         /* Communicate only the number of atoms */
 170         MPI_Isend(&n, sizeof(n), MPI_BYTE, dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim,
 171                   &dd->req_pme[dd->nreq_pme++]);
 172 #endif
 173     }
 174
 175 #if GMX_MPI
 176     if (n > 0)
 177     {
 178         if (flags & PP_PME_CHARGE)
 179         {
 180             MPI_Isend(chargeA, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_ChargeA,
 181                       cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
 182         }
 183         if (flags & PP_PME_CHARGEB)
 184         {
 185             MPI_Isend(chargeB, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_ChargeB,
 186                       cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
 187         }
 188         if (flags & PP_PME_SQRTC6)
 189         {
 190             MPI_Isend(c6A, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_SQRTC6A,
 191                       cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
 192         }
 193         if (flags & PP_PME_SQRTC6B)
 194         {
 195             MPI_Isend(c6B, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_SQRTC6B,
 196                       cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
 197         }
 198         if (flags & PP_PME_SIGMA)
 199         {
 200             MPI_Isend(sigmaA, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_SigmaA,
 201                       cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
 202         }
 203         if (flags & PP_PME_SIGMAB)
 204         {
 205             MPI_Isend(sigmaB, n * sizeof(real), MPI_BYTE, dd->pme_nodeid, eCommType_SigmaB,
 206                       cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
 207         }
 208         if (flags & PP_PME_COORD)
 209         {
 210             if (reinitGpuPmePpComms)
 211             {
 212                 fr->pmePpCommGpu->reinit(n);
 213             }
 214
 215
 216             /* MPI_Isend does not accept a const buffer pointer */
 217             real* xRealPtr = const_cast<real*>(x[0]);
 218             if (useGpuPmePpComms && (fr != nullptr))
 219             {
 220                 void* sendPtr = sendCoordinatesFromGpu
 221                                         ? static_cast<void*>(fr->stateGpu->getCoordinates())
 222                                         : static_cast<void*>(xRealPtr);
 223                 fr->pmePpCommGpu->sendCoordinatesToPmeCudaDirect(sendPtr, n, sendCoordinatesFromGpu,
 224                                                                  coordinatesReadyOnDeviceEvent);
 225             }
 226             else
 227             {
 228                 MPI_Isend(xRealPtr, n * sizeof(rvec), MPI_BYTE, dd->pme_nodeid, eCommType_COORD,
 229                           cr->mpi_comm_mysim, &dd->req_pme[dd->nreq_pme++]);
 230             }
 231         }
 232     }
 233 #else
 234     GMX_UNUSED_VALUE(fr);
 235     GMX_UNUSED_VALUE(reinitGpuPmePpComms);
 236     GMX_UNUSED_VALUE(sendCoordinatesFromGpu);
 237     GMX_UNUSED_VALUE(coordinatesReadyOnDeviceEvent);
 238 #endif
 239     if (!c_useDelayedWait)
 240     {
 241         /* Wait for the data to arrive */
 242         /* We can skip this wait as we are sure x and q will not be modified
 243          * before the next call to gmx_pme_send_x_q or gmx_pme_receive_f.
 244          */
 245         gmx_pme_send_coeffs_coords_wait(dd);
 246     }
 247 }
 248
 249 void gmx_pme_send_parameters(const t_commrec*           cr,
 250                              const interaction_const_t* ic,
 251                              gmx_bool                   bFreeEnergy_q,
 252                              gmx_bool                   bFreeEnergy_lj,
 253                              real*                      chargeA,
 254                              real*                      chargeB,
 255                              real*                      sqrt_c6A,
 256                              real*                      sqrt_c6B,
 257                              real*                      sigmaA,
 258                              real*                      sigmaB,
 259                              int                        maxshift_x,
 260                              int                        maxshift_y)
 261 {
 262     unsigned int flags = 0;
 263
 264     if (EEL_PME(ic->eeltype))
 265     {
 266         flags |= PP_PME_CHARGE;
 267     }
 268     if (EVDW_PME(ic->vdwtype))
 269     {
 270         flags |= (PP_PME_SQRTC6 | PP_PME_SIGMA);
 271     }
 272     if (bFreeEnergy_q || bFreeEnergy_lj)
 273     {
 274         /* Assumes that the B state flags are in the bits just above
 275          * the ones for the A state. */
 276         flags |= (flags << 1);
 277     }
 278
 279     gmx_pme_send_coeffs_coords(nullptr, cr, flags, chargeA, chargeB, sqrt_c6A, sqrt_c6B, sigmaA,
 280                                sigmaB, nullptr, nullptr, 0, 0, maxshift_x, maxshift_y, -1, false,
 281                                false, false, nullptr);
 282 }
 283
 284 void gmx_pme_send_coordinates(t_forcerec*           fr,
 285                               const t_commrec*      cr,
 286                               const matrix          box,
 287                               const rvec*           x,
 288                               real                  lambda_q,
 289                               real                  lambda_lj,
 290                               gmx_bool              bEnerVir,
 291                               int64_t               step,
 292                               bool                  useGpuPmePpComms,
 293                               bool                  receiveCoordinateAddressFromPme,
 294                               bool                  sendCoordinatesFromGpu,
 295                               GpuEventSynchronizer* coordinatesReadyOnDeviceEvent,
 296                               gmx_wallcycle*        wcycle)
 297 {
 298     wallcycle_start(wcycle, ewcPP_PMESENDX);
 299
 300     unsigned int flags = PP_PME_COORD;
 301     if (bEnerVir)
 302     {
 303         flags |= PP_PME_ENER_VIR;
 304     }
 305     gmx_pme_send_coeffs_coords(fr, cr, flags, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr,
 306                                box, x, lambda_q, lambda_lj, 0, 0, step, useGpuPmePpComms,
 307                                receiveCoordinateAddressFromPme, sendCoordinatesFromGpu,
 308                                coordinatesReadyOnDeviceEvent);
 309
 310     wallcycle_stop(wcycle, ewcPP_PMESENDX);
 311 }
 312
 313 void gmx_pme_send_finish(const t_commrec* cr)
 314 {
 315     unsigned int flags = PP_PME_FINISH;
 316
 317     gmx_pme_send_coeffs_coords(nullptr, cr, flags, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr,
 318                                nullptr, nullptr, 0, 0, 0, 0, -1, false, false, false, nullptr);
 319 }
 320
 321 void gmx_pme_send_switchgrid(const t_commrec* cr, ivec grid_size, real ewaldcoeff_q, real ewaldcoeff_lj)
 322 {
 323 #if GMX_MPI
 324     gmx_pme_comm_n_box_t cnb;
 325
 326     /* Only let one PP node signal each PME node */
 327     if (cr->dd->pme_receive_vir_ener)
 328     {
 329         cnb.flags = PP_PME_SWITCHGRID;
 330         copy_ivec(grid_size, cnb.grid_size);
 331         cnb.ewaldcoeff_q  = ewaldcoeff_q;
 332         cnb.ewaldcoeff_lj = ewaldcoeff_lj;
 333
 334         /* We send this, uncommon, message blocking to simplify the code */
 335         MPI_Send(&cnb, sizeof(cnb), MPI_BYTE, cr->dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim);
 336     }
 337 #else
 338     GMX_UNUSED_VALUE(cr);
 339     GMX_UNUSED_VALUE(grid_size);
 340     GMX_UNUSED_VALUE(ewaldcoeff_q);
 341     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 342 #endif
 343 }
 344
 345 void gmx_pme_send_resetcounters(const t_commrec gmx_unused* cr, int64_t gmx_unused step)
 346 {
 347 #if GMX_MPI
 348     gmx_pme_comm_n_box_t cnb;
 349
 350     /* Only let one PP node signal each PME node */
 351     if (cr->dd->pme_receive_vir_ener)
 352     {
 353         cnb.flags = PP_PME_RESETCOUNTERS;
 354         cnb.step  = step;
 355
 356         /* We send this, uncommon, message blocking to simplify the code */
 357         MPI_Send(&cnb, sizeof(cnb), MPI_BYTE, cr->dd->pme_nodeid, eCommType_CNB, cr->mpi_comm_mysim);
 358     }
 359 #endif
 360 }
 361
 362 /*! \brief Receive virial and energy from PME rank */
 363 static void receive_virial_energy(const t_commrec*      cr,
 364                                   gmx::ForceWithVirial* forceWithVirial,
 365                                   real*                 energy_q,
 366                                   real*                 energy_lj,
 367                                   real*                 dvdlambda_q,
 368                                   real*                 dvdlambda_lj,
 369                                   float*                pme_cycles)
 370 {
 371     gmx_pme_comm_vir_ene_t cve;
 372
 373     if (cr->dd->pme_receive_vir_ener)
 374     {
 375         if (debug)
 376         {
 377             fprintf(debug, "PP rank %d receiving from PME rank %d: virial and energy\n",
 378                     cr->sim_nodeid, cr->dd->pme_nodeid);
 379         }
 380 #if GMX_MPI
 381         MPI_Recv(&cve, sizeof(cve), MPI_BYTE, cr->dd->pme_nodeid, 1, cr->mpi_comm_mysim, MPI_STATUS_IGNORE);
 382 #else
 383         memset(&cve, 0, sizeof(cve));
 384 #endif
 385
 386         forceWithVirial->addVirialContribution(cve.vir_q);
 387         forceWithVirial->addVirialContribution(cve.vir_lj);
 388         *energy_q  = cve.energy_q;
 389         *energy_lj = cve.energy_lj;
 390         *dvdlambda_q += cve.dvdlambda_q;
 391         *dvdlambda_lj += cve.dvdlambda_lj;
 392         *pme_cycles = cve.cycles;
 393
 394         if (cve.stop_cond != gmx_stop_cond_none)
 395         {
 396             gmx_set_stop_condition(cve.stop_cond);
 397         }
 398     }
 399     else
 400     {
 401         *energy_q   = 0;
 402         *energy_lj  = 0;
 403         *pme_cycles = 0;
 404     }
 405 }
 406
 407 /*! \brief Recieve force data from PME ranks */
 408 static void recvFFromPme(gmx::PmePpCommGpu* pmePpCommGpu,
 409                          void*              recvptr,
 410                          int                n,
 411                          const t_commrec*   cr,
 412                          bool               useGpuPmePpComms,
 413                          bool               receivePmeForceToGpu)
 414 {
 415     if (useGpuPmePpComms)
 416     {
 417         GMX_ASSERT(pmePpCommGpu != nullptr, "Need valid pmePpCommGpu");
 418         // Receive directly using CUDA memory copy
 419         pmePpCommGpu->receiveForceFromPmeCudaDirect(recvptr, n, receivePmeForceToGpu);
 420     }
 421     else
 422     {
 423         // Receive data using MPI
 424 #if GMX_MPI
 425         MPI_Recv(recvptr, n * sizeof(rvec), MPI_BYTE, cr->dd->pme_nodeid, 0, cr->mpi_comm_mysim,
 426                  MPI_STATUS_IGNORE);
 427 #else
 428         GMX_UNUSED_VALUE(cr);
 429 #endif
 430     }
 431 }
 432
 433
 434 void gmx_pme_receive_f(gmx::PmePpCommGpu*    pmePpCommGpu,
 435                        const t_commrec*      cr,
 436                        gmx::ForceWithVirial* forceWithVirial,
 437                        real*                 energy_q,
 438                        real*                 energy_lj,
 439                        real*                 dvdlambda_q,
 440                        real*                 dvdlambda_lj,
 441                        bool                  useGpuPmePpComms,
 442                        bool                  receivePmeForceToGpu,
 443                        float*                pme_cycles)
 444 {
 445     if (c_useDelayedWait)
 446     {
 447         /* Wait for the x request to finish */
 448         gmx_pme_send_coeffs_coords_wait(cr->dd);
 449     }
 450
 451     const int               natoms = dd_numHomeAtoms(*cr->dd);
 452     std::vector<gmx::RVec>& buffer = cr->dd->pmeForceReceiveBuffer;
 453     buffer.resize(natoms);
 454
 455     void* recvptr = reinterpret_cast<void*>(buffer.data());
 456     recvFFromPme(pmePpCommGpu, recvptr, natoms, cr, useGpuPmePpComms, receivePmeForceToGpu);
 457
 458     int nt = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, natoms);
 459
 460     gmx::ArrayRef<gmx::RVec> f = forceWithVirial->force_;
 461
 462     if (!receivePmeForceToGpu)
 463     {
 464         /* Note that we would like to avoid this conditional by putting it
 465          * into the omp pragma instead, but then we still take the full
 466          * omp parallel for overhead (at least with gcc5).
 467          */
 468         if (nt == 1)
 469         {
 470             for (int i = 0; i < natoms; i++)
 471             {
 472                 f[i] += buffer[i];
 473             }
 474         }
 475         else
 476         {
 477 #pragma omp parallel for num_threads(nt) schedule(static)
 478             for (int i = 0; i < natoms; i++)
 479             {
 480                 f[i] += buffer[i];
 481             }
 482         }
 483     }
 484
 485     receive_virial_energy(cr, forceWithVirial, energy_q, energy_lj, dvdlambda_q, dvdlambda_lj, pme_cycles);
 486 }