src/gromacs/mdlib/sim_util.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) 2013-2019,2020,2021, 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 #include "gmxpre.h"
  38
  39 #include "config.h"
  40
  41 #include <cmath>
  42 #include <cstdint>
  43 #include <cstdio>
  44 #include <cstring>
  45
  46 #include <array>
  47 #include <optional>
  48
  49 #include "gromacs/applied_forces/awh/awh.h"
  50 #include "gromacs/domdec/dlbtiming.h"
  51 #include "gromacs/domdec/domdec.h"
  52 #include "gromacs/domdec/domdec_struct.h"
  53 #include "gromacs/domdec/gpuhaloexchange.h"
  54 #include "gromacs/domdec/partition.h"
  55 #include "gromacs/essentialdynamics/edsam.h"
  56 #include "gromacs/ewald/pme.h"
  57 #include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
  58 #include "gromacs/ewald/pme_pp.h"
  59 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  60 #include "gromacs/gmxlib/network.h"
  61 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  62 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  63 #include "gromacs/gmxlib/nrnb.h"
  64 #include "gromacs/gpu_utils/gpu_utils.h"
  65 #include "gromacs/imd/imd.h"
  66 #include "gromacs/listed_forces/disre.h"
  67 #include "gromacs/listed_forces/listed_forces_gpu.h"
  68 #include "gromacs/listed_forces/listed_forces.h"
  69 #include "gromacs/listed_forces/orires.h"
  70 #include "gromacs/math/arrayrefwithpadding.h"
  71 #include "gromacs/math/functions.h"
  72 #include "gromacs/math/units.h"
  73 #include "gromacs/math/vec.h"
  74 #include "gromacs/math/vecdump.h"
  75 #include "gromacs/mdlib/calcmu.h"
  76 #include "gromacs/mdlib/calcvir.h"
  77 #include "gromacs/mdlib/constr.h"
  78 #include "gromacs/mdlib/dispersioncorrection.h"
  79 #include "gromacs/mdlib/enerdata_utils.h"
  80 #include "gromacs/mdlib/force.h"
  81 #include "gromacs/mdlib/force_flags.h"
  82 #include "gromacs/mdlib/forcerec.h"
  83 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  84 #include "gromacs/mdlib/update.h"
  85 #include "gromacs/mdlib/vsite.h"
  86 #include "gromacs/mdlib/wall.h"
  87 #include "gromacs/mdlib/wholemoleculetransform.h"
  88 #include "gromacs/mdtypes/commrec.h"
  89 #include "gromacs/mdtypes/enerdata.h"
  90 #include "gromacs/mdtypes/forcebuffers.h"
  91 #include "gromacs/mdtypes/forceoutput.h"
  92 #include "gromacs/mdtypes/forcerec.h"
  93 #include "gromacs/mdtypes/iforceprovider.h"
  94 #include "gromacs/mdtypes/inputrec.h"
  95 #include "gromacs/mdtypes/md_enums.h"
  96 #include "gromacs/mdtypes/mdatom.h"
  97 #include "gromacs/mdtypes/multipletimestepping.h"
  98 #include "gromacs/mdtypes/simulation_workload.h"
  99 #include "gromacs/mdtypes/state.h"
 100 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
 101 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 102 #include "gromacs/nbnxm/nbnxm.h"
 103 #include "gromacs/nbnxm/nbnxm_gpu.h"
 104 #include "gromacs/pbcutil/ishift.h"
 105 #include "gromacs/pbcutil/pbc.h"
 106 #include "gromacs/pulling/pull.h"
 107 #include "gromacs/pulling/pull_rotation.h"
 108 #include "gromacs/timing/cyclecounter.h"
 109 #include "gromacs/timing/gpu_timing.h"
 110 #include "gromacs/timing/wallcycle.h"
 111 #include "gromacs/timing/wallcyclereporting.h"
 112 #include "gromacs/timing/walltime_accounting.h"
 113 #include "gromacs/topology/topology.h"
 114 #include "gromacs/utility/arrayref.h"
 115 #include "gromacs/utility/basedefinitions.h"
 116 #include "gromacs/utility/cstringutil.h"
 117 #include "gromacs/utility/exceptions.h"
 118 #include "gromacs/utility/fatalerror.h"
 119 #include "gromacs/utility/fixedcapacityvector.h"
 120 #include "gromacs/utility/gmxassert.h"
 121 #include "gromacs/utility/gmxmpi.h"
 122 #include "gromacs/utility/logger.h"
 123 #include "gromacs/utility/smalloc.h"
 124 #include "gromacs/utility/strconvert.h"
 125 #include "gromacs/utility/sysinfo.h"
 126
 127 #include "gpuforcereduction.h"
 128
 129 using gmx::ArrayRef;
 130 using gmx::AtomLocality;
 131 using gmx::DomainLifetimeWorkload;
 132 using gmx::ForceOutputs;
 133 using gmx::ForceWithShiftForces;
 134 using gmx::InteractionLocality;
 135 using gmx::RVec;
 136 using gmx::SimulationWorkload;
 137 using gmx::StepWorkload;
 138
 139 // TODO: this environment variable allows us to verify before release
 140 // that on less common architectures the total cost of polling is not larger than
 141 // a blocking wait (so polling does not introduce overhead when the static
 142 // PME-first ordering would suffice).
 143 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 144
 145 static void sum_forces(ArrayRef<RVec> f, ArrayRef<const RVec> forceToAdd)
 146 {
 147     GMX_ASSERT(f.size() >= forceToAdd.size(), "Accumulation buffer should be sufficiently large");
 148     const int end = forceToAdd.size();
 149
 150     int gmx_unused nt = gmx_omp_nthreads_get(ModuleMultiThread::Default);
 151 #pragma omp parallel for num_threads(nt) schedule(static)
 152     for (int i = 0; i < end; i++)
 153     {
 154         rvec_inc(f[i], forceToAdd[i]);
 155     }
 156 }
 157
 158 static void calc_virial(int                              start,
 159                         int                              homenr,
 160                         const rvec                       x[],
 161                         const gmx::ForceWithShiftForces& forceWithShiftForces,
 162                         tensor                           vir_part,
 163                         const matrix                     box,
 164                         t_nrnb*                          nrnb,
 165                         const t_forcerec*                fr,
 166                         PbcType                          pbcType)
 167 {
 168     /* The short-range virial from surrounding boxes */
 169     const rvec* fshift          = as_rvec_array(forceWithShiftForces.shiftForces().data());
 170     const rvec* shiftVecPointer = as_rvec_array(fr->shift_vec.data());
 171     calc_vir(gmx::c_numShiftVectors, shiftVecPointer, fshift, vir_part, pbcType == PbcType::Screw, box);
 172     inc_nrnb(nrnb, eNR_VIRIAL, gmx::c_numShiftVectors);
 173
 174     /* Calculate partial virial, for local atoms only, based on short range.
 175      * Total virial is computed in global_stat, called from do_md
 176      */
 177     const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
 178     f_calc_vir(start, start + homenr, x, f, vir_part, box);
 179     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 180
 181     if (debug)
 182     {
 183         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 184     }
 185 }
 186
 187 static void pull_potential_wrapper(const t_commrec*               cr,
 188                                    const t_inputrec&              ir,
 189                                    const matrix                   box,
 190                                    gmx::ArrayRef<const gmx::RVec> x,
 191                                    gmx::ForceWithVirial*          force,
 192                                    const t_mdatoms*               mdatoms,
 193                                    gmx_enerdata_t*                enerd,
 194                                    pull_t*                        pull_work,
 195                                    const real*                    lambda,
 196                                    double                         t,
 197                                    gmx_wallcycle*                 wcycle)
 198 {
 199     t_pbc pbc;
 200     real  dvdl;
 201
 202     /* Calculate the center of mass forces, this requires communication,
 203      * which is why pull_potential is called close to other communication.
 204      */
 205     wallcycle_start(wcycle, WallCycleCounter::PullPot);
 206     set_pbc(&pbc, ir.pbcType, box);
 207     dvdl = 0;
 208     enerd->term[F_COM_PULL] +=
 209             pull_potential(pull_work,
 210                            gmx::arrayRefFromArray(mdatoms->massT, mdatoms->nr),
 211                            pbc,
 212                            cr,
 213                            t,
 214                            lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Restraint)],
 215                            x,
 216                            force,
 217                            &dvdl);
 218     enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Restraint] += dvdl;
 219     wallcycle_stop(wcycle, WallCycleCounter::PullPot);
 220 }
 221
 222 static void pme_receive_force_ener(t_forcerec*           fr,
 223                                    const t_commrec*      cr,
 224                                    gmx::ForceWithVirial* forceWithVirial,
 225                                    gmx_enerdata_t*       enerd,
 226                                    bool                  useGpuPmePpComms,
 227                                    bool                  receivePmeForceToGpu,
 228                                    gmx_wallcycle*        wcycle)
 229 {
 230     real  e_q, e_lj, dvdl_q, dvdl_lj;
 231     float cycles_ppdpme, cycles_seppme;
 232
 233     cycles_ppdpme = wallcycle_stop(wcycle, WallCycleCounter::PpDuringPme);
 234     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 235
 236     /* In case of node-splitting, the PP nodes receive the long-range
 237      * forces, virial and energy from the PME nodes here.
 238      */
 239     wallcycle_start(wcycle, WallCycleCounter::PpPmeWaitRecvF);
 240     dvdl_q  = 0;
 241     dvdl_lj = 0;
 242     gmx_pme_receive_f(fr->pmePpCommGpu.get(),
 243                       cr,
 244                       forceWithVirial,
 245                       &e_q,
 246                       &e_lj,
 247                       &dvdl_q,
 248                       &dvdl_lj,
 249                       useGpuPmePpComms,
 250                       receivePmeForceToGpu,
 251                       &cycles_seppme);
 252     enerd->term[F_COUL_RECIP] += e_q;
 253     enerd->term[F_LJ_RECIP] += e_lj;
 254     enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Coul] += dvdl_q;
 255     enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] += dvdl_lj;
 256
 257     if (wcycle)
 258     {
 259         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 260     }
 261     wallcycle_stop(wcycle, WallCycleCounter::PpPmeWaitRecvF);
 262 }
 263
 264 static void print_large_forces(FILE*                fp,
 265                                const t_mdatoms*     md,
 266                                const t_commrec*     cr,
 267                                int64_t              step,
 268                                real                 forceTolerance,
 269                                ArrayRef<const RVec> x,
 270                                ArrayRef<const RVec> f)
 271 {
 272     real       force2Tolerance = gmx::square(forceTolerance);
 273     gmx::index numNonFinite    = 0;
 274     for (int i = 0; i < md->homenr; i++)
 275     {
 276         real force2    = norm2(f[i]);
 277         bool nonFinite = !std::isfinite(force2);
 278         if (force2 >= force2Tolerance || nonFinite)
 279         {
 280             fprintf(fp,
 281                     "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 282                     step,
 283                     ddglatnr(cr->dd, i),
 284                     x[i][XX],
 285                     x[i][YY],
 286                     x[i][ZZ],
 287                     std::sqrt(force2));
 288         }
 289         if (nonFinite)
 290         {
 291             numNonFinite++;
 292         }
 293     }
 294     if (numNonFinite > 0)
 295     {
 296         /* Note that with MPI this fatal call on one rank might interrupt
 297          * the printing on other ranks. But we can only avoid that with
 298          * an expensive MPI barrier that we would need at each step.
 299          */
 300         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 301     }
 302 }
 303
 304 //! When necessary, spreads forces on vsites and computes the virial for \p forceOutputs->forceWithShiftForces()
 305 static void postProcessForceWithShiftForces(t_nrnb*                   nrnb,
 306                                             gmx_wallcycle*            wcycle,
 307                                             const matrix              box,
 308                                             ArrayRef<const RVec>      x,
 309                                             ForceOutputs*             forceOutputs,
 310                                             tensor                    vir_force,
 311                                             const t_mdatoms&          mdatoms,
 312                                             const t_forcerec&         fr,
 313                                             gmx::VirtualSitesHandler* vsite,
 314                                             const StepWorkload&       stepWork)
 315 {
 316     ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
 317
 318     /* If we have NoVirSum forces, but we do not calculate the virial,
 319      * we later sum the forceWithShiftForces buffer together with
 320      * the noVirSum buffer and spread the combined vsite forces at once.
 321      */
 322     if (vsite && (!forceOutputs->haveForceWithVirial() || stepWork.computeVirial))
 323     {
 324         using VirialHandling = gmx::VirtualSitesHandler::VirialHandling;
 325
 326         auto                 f      = forceWithShiftForces.force();
 327         auto                 fshift = forceWithShiftForces.shiftForces();
 328         const VirialHandling virialHandling =
 329                 (stepWork.computeVirial ? VirialHandling::Pbc : VirialHandling::None);
 330         vsite->spreadForces(x, f, virialHandling, fshift, nullptr, nrnb, box, wcycle);
 331         forceWithShiftForces.haveSpreadVsiteForces() = true;
 332     }
 333
 334     if (stepWork.computeVirial)
 335     {
 336         /* Calculation of the virial must be done after vsites! */
 337         calc_virial(
 338                 0, mdatoms.homenr, as_rvec_array(x.data()), forceWithShiftForces, vir_force, box, nrnb, &fr, fr.pbcType);
 339     }
 340 }
 341
 342 //! Spread, compute virial for and sum forces, when necessary
 343 static void postProcessForces(const t_commrec*          cr,
 344                               int64_t                   step,
 345                               t_nrnb*                   nrnb,
 346                               gmx_wallcycle*            wcycle,
 347                               const matrix              box,
 348                               ArrayRef<const RVec>      x,
 349                               ForceOutputs*             forceOutputs,
 350                               tensor                    vir_force,
 351                               const t_mdatoms*          mdatoms,
 352                               const t_forcerec*         fr,
 353                               gmx::VirtualSitesHandler* vsite,
 354                               const StepWorkload&       stepWork)
 355 {
 356     // Extract the final output force buffer, which is also the buffer for forces with shift forces
 357     ArrayRef<RVec> f = forceOutputs->forceWithShiftForces().force();
 358
 359     if (forceOutputs->haveForceWithVirial())
 360     {
 361         auto& forceWithVirial = forceOutputs->forceWithVirial();
 362
 363         if (vsite)
 364         {
 365             /* Spread the mesh force on virtual sites to the other particles...
 366              * This is parallellized. MPI communication is performed
 367              * if the constructing atoms aren't local.
 368              */
 369             GMX_ASSERT(!stepWork.computeVirial || f.data() != forceWithVirial.force_.data(),
 370                        "We need separate force buffers for shift and virial forces when "
 371                        "computing the virial");
 372             GMX_ASSERT(!stepWork.computeVirial
 373                                || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
 374                        "We should spread the force with shift forces separately when computing "
 375                        "the virial");
 376             const gmx::VirtualSitesHandler::VirialHandling virialHandling =
 377                     (stepWork.computeVirial ? gmx::VirtualSitesHandler::VirialHandling::NonLinear
 378                                             : gmx::VirtualSitesHandler::VirialHandling::None);
 379             matrix virial = { { 0 } };
 380             vsite->spreadForces(x, forceWithVirial.force_, virialHandling, {}, virial, nrnb, box, wcycle);
 381             forceWithVirial.addVirialContribution(virial);
 382         }
 383
 384         if (stepWork.computeVirial)
 385         {
 386             /* Now add the forces, this is local */
 387             sum_forces(f, forceWithVirial.force_);
 388
 389             /* Add the direct virial contributions */
 390             GMX_ASSERT(
 391                     forceWithVirial.computeVirial_,
 392                     "forceWithVirial should request virial computation when we request the virial");
 393             m_add(vir_force, forceWithVirial.getVirial(), vir_force);
 394
 395             if (debug)
 396             {
 397                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 398             }
 399         }
 400     }
 401     else
 402     {
 403         GMX_ASSERT(vsite == nullptr || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
 404                    "We should have spread the vsite forces (earlier)");
 405     }
 406
 407     if (fr->print_force >= 0)
 408     {
 409         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 410     }
 411 }
 412
 413 static void do_nb_verlet(t_forcerec*                fr,
 414                          const interaction_const_t* ic,
 415                          gmx_enerdata_t*            enerd,
 416                          const StepWorkload&        stepWork,
 417                          const InteractionLocality  ilocality,
 418                          const int                  clearF,
 419                          const int64_t              step,
 420                          t_nrnb*                    nrnb,
 421                          gmx_wallcycle*             wcycle)
 422 {
 423     if (!stepWork.computeNonbondedForces)
 424     {
 425         /* skip non-bonded calculation */
 426         return;
 427     }
 428
 429     nonbonded_verlet_t* nbv = fr->nbv.get();
 430
 431     /* GPU kernel launch overhead is already timed separately */
 432     if (!nbv->useGpu())
 433     {
 434         /* When dynamic pair-list  pruning is requested, we need to prune
 435          * at nstlistPrune steps.
 436          */
 437         if (nbv->isDynamicPruningStepCpu(step))
 438         {
 439             /* Prune the pair-list beyond fr->ic->rlistPrune using
 440              * the current coordinates of the atoms.
 441              */
 442             wallcycle_sub_start(wcycle, WallCycleSubCounter::NonbondedPruning);
 443             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 444             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NonbondedPruning);
 445         }
 446     }
 447
 448     nbv->dispatchNonbondedKernel(
 449             ilocality,
 450             *ic,
 451             stepWork,
 452             clearF,
 453             fr->shift_vec,
 454             enerd->grpp.energyGroupPairTerms[fr->haveBuckingham ? NonBondedEnergyTerms::BuckinghamSR
 455                                                                 : NonBondedEnergyTerms::LJSR],
 456             enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR],
 457             nrnb);
 458 }
 459
 460 static inline void clearRVecs(ArrayRef<RVec> v, const bool useOpenmpThreading)
 461 {
 462     int nth = gmx_omp_nthreads_get_simple_rvec_task(ModuleMultiThread::Default, v.ssize());
 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 (!useOpenmpThreading || nth == 1)
 469     {
 470         for (RVec& elem : v)
 471         {
 472             clear_rvec(elem);
 473         }
 474     }
 475     else
 476     {
 477 #pragma omp parallel for num_threads(nth) schedule(static)
 478         for (gmx::index i = 0; i < v.ssize(); i++)
 479         {
 480             clear_rvec(v[i]);
 481         }
 482     }
 483 }
 484
 485 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 486  *
 487  * \param groupOptions  Group options, containing T-coupling options
 488  */
 489 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
 490 {
 491     real nrdfCoupled   = 0;
 492     real nrdfUncoupled = 0;
 493     real kineticEnergy = 0;
 494     for (int g = 0; g < groupOptions.ngtc; g++)
 495     {
 496         if (groupOptions.tau_t[g] >= 0)
 497         {
 498             nrdfCoupled += groupOptions.nrdf[g];
 499             kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * gmx::c_boltz;
 500         }
 501         else
 502         {
 503             nrdfUncoupled += groupOptions.nrdf[g];
 504         }
 505     }
 506
 507     /* This conditional with > also catches nrdf=0 */
 508     if (nrdfCoupled > nrdfUncoupled)
 509     {
 510         return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
 511     }
 512     else
 513     {
 514         return 0;
 515     }
 516 }
 517
 518 /*! \brief This routine checks that the potential energy is finite.
 519  *
 520  * Always checks that the potential energy is finite. If step equals
 521  * inputrec.init_step also checks that the magnitude of the potential energy
 522  * is reasonable. Terminates with a fatal error when a check fails.
 523  * Note that passing this check does not guarantee finite forces,
 524  * since those use slightly different arithmetics. But in most cases
 525  * there is just a narrow coordinate range where forces are not finite
 526  * and energies are finite.
 527  *
 528  * \param[in] step      The step number, used for checking and printing
 529  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to
 530  * enerd.term[F_EPOT] before calling this routine \param[in] inputrec  The input record
 531  */
 532 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
 533 {
 534     /* Threshold valid for comparing absolute potential energy against
 535      * the kinetic energy. Normally one should not consider absolute
 536      * potential energy values, but with a factor of one million
 537      * we should never get false positives.
 538      */
 539     constexpr real c_thresholdFactor = 1e6;
 540
 541     bool energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 542     real averageKineticEnergy = 0;
 543     /* We only check for large potential energy at the initial step,
 544      * because that is by far the most likely step for this too occur
 545      * and because computing the average kinetic energy is not free.
 546      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 547      * before they become NaN.
 548      */
 549     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 550     {
 551         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 552     }
 553
 554     if (energyIsNotFinite
 555         || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
 556     {
 557         gmx_fatal(
 558                 FARGS,
 559                 "Step %" PRId64
 560                 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
 561                 "contributions to the energy are %g and %g, respectively. A %s potential energy "
 562                 "can be caused by overlapping interactions in bonded interactions or very large%s "
 563                 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
 564                 "configuration, incorrect interactions or parameters in the topology.",
 565                 step,
 566                 enerd.term[F_EPOT],
 567                 energyIsNotFinite ? "not finite" : "extremely high",
 568                 enerd.term[F_LJ],
 569                 enerd.term[F_COUL_SR],
 570                 energyIsNotFinite ? "non-finite" : "very high",
 571                 energyIsNotFinite ? " or Nan" : "");
 572     }
 573 }
 574
 575 /*! \brief Return true if there are special forces computed this step.
 576  *
 577  * The conditionals exactly correspond to those in computeSpecialForces().
 578  */
 579 static bool haveSpecialForces(const t_inputrec&          inputrec,
 580                               const gmx::ForceProviders& forceProviders,
 581                               const pull_t*              pull_work,
 582                               const bool                 computeForces,
 583                               const gmx_edsam*           ed)
 584 {
 585
 586     return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
 587             (inputrec.bPull && pull_have_potential(*pull_work)) ||  // pull
 588             inputrec.bRot ||                                        // enforced rotation
 589             (ed != nullptr) ||                                      // flooding
 590             (inputrec.bIMD && computeForces));                      // IMD
 591 }
 592
 593 /*! \brief Compute forces and/or energies for special algorithms
 594  *
 595  * The intention is to collect all calls to algorithms that compute
 596  * forces on local atoms only and that do not contribute to the local
 597  * virial sum (but add their virial contribution separately).
 598  * Eventually these should likely all become ForceProviders.
 599  * Within this function the intention is to have algorithms that do
 600  * global communication at the end, so global barriers within the MD loop
 601  * are as close together as possible.
 602  *
 603  * \param[in]     fplog            The log file
 604  * \param[in]     cr               The communication record
 605  * \param[in]     inputrec         The input record
 606  * \param[in]     awh              The Awh module (nullptr if none in use).
 607  * \param[in]     enforcedRotation Enforced rotation module.
 608  * \param[in]     imdSession       The IMD session
 609  * \param[in]     pull_work        The pull work structure.
 610  * \param[in]     step             The current MD step
 611  * \param[in]     t                The current time
 612  * \param[in,out] wcycle           Wallcycle accounting struct
 613  * \param[in,out] forceProviders   Pointer to a list of force providers
 614  * \param[in]     box              The unit cell
 615  * \param[in]     x                The coordinates
 616  * \param[in]     mdatoms          Per atom properties
 617  * \param[in]     lambda           Array of free-energy lambda values
 618  * \param[in]     stepWork         Step schedule flags
 619  * \param[in,out] forceWithVirialMtsLevel0  Force and virial for MTS level0 forces
 620  * \param[in,out] forceWithVirialMtsLevel1  Force and virial for MTS level1 forces, can be nullptr
 621  * \param[in,out] enerd            Energy buffer
 622  * \param[in,out] ed               Essential dynamics pointer
 623  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 624  *
 625  * \todo Remove didNeighborSearch, which is used incorrectly.
 626  * \todo Convert all other algorithms called here to ForceProviders.
 627  */
 628 static void computeSpecialForces(FILE*                          fplog,
 629                                  const t_commrec*               cr,
 630                                  const t_inputrec&              inputrec,
 631                                  gmx::Awh*                      awh,
 632                                  gmx_enfrot*                    enforcedRotation,
 633                                  gmx::ImdSession*               imdSession,
 634                                  pull_t*                        pull_work,
 635                                  int64_t                        step,
 636                                  double                         t,
 637                                  gmx_wallcycle*                 wcycle,
 638                                  gmx::ForceProviders*           forceProviders,
 639                                  const matrix                   box,
 640                                  gmx::ArrayRef<const gmx::RVec> x,
 641                                  const t_mdatoms*               mdatoms,
 642                                  gmx::ArrayRef<const real>      lambda,
 643                                  const StepWorkload&            stepWork,
 644                                  gmx::ForceWithVirial*          forceWithVirialMtsLevel0,
 645                                  gmx::ForceWithVirial*          forceWithVirialMtsLevel1,
 646                                  gmx_enerdata_t*                enerd,
 647                                  gmx_edsam*                     ed,
 648                                  bool                           didNeighborSearch)
 649 {
 650     /* NOTE: Currently all ForceProviders only provide forces.
 651      *       When they also provide energies, remove this conditional.
 652      */
 653     if (stepWork.computeForces)
 654     {
 655         gmx::ForceProviderInput forceProviderInput(
 656                 x,
 657                 mdatoms->homenr,
 658                 gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->homenr),
 659                 gmx::arrayRefFromArray(mdatoms->massT, mdatoms->homenr),
 660                 t,
 661                 box,
 662                 *cr);
 663         gmx::ForceProviderOutput forceProviderOutput(forceWithVirialMtsLevel0, enerd);
 664
 665         /* Collect forces from modules */
 666         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 667     }
 668
 669     if (inputrec.bPull && pull_have_potential(*pull_work))
 670     {
 671         const int mtsLevel = forceGroupMtsLevel(inputrec.mtsLevels, gmx::MtsForceGroups::Pull);
 672         if (mtsLevel == 0 || stepWork.computeSlowForces)
 673         {
 674             auto& forceWithVirial = (mtsLevel == 0) ? forceWithVirialMtsLevel0 : forceWithVirialMtsLevel1;
 675             pull_potential_wrapper(
 676                     cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work, lambda.data(), t, wcycle);
 677         }
 678     }
 679     if (awh)
 680     {
 681         const int mtsLevel = forceGroupMtsLevel(inputrec.mtsLevels, gmx::MtsForceGroups::Pull);
 682         if (mtsLevel == 0 || stepWork.computeSlowForces)
 683         {
 684             const bool needForeignEnergyDifferences = awh->needForeignEnergyDifferences(step);
 685             std::vector<double> foreignLambdaDeltaH, foreignLambdaDhDl;
 686             if (needForeignEnergyDifferences)
 687             {
 688                 enerd->foreignLambdaTerms.finalizePotentialContributions(
 689                         enerd->dvdl_lin, lambda, *inputrec.fepvals);
 690                 std::tie(foreignLambdaDeltaH, foreignLambdaDhDl) = enerd->foreignLambdaTerms.getTerms(cr);
 691             }
 692
 693             auto& forceWithVirial = (mtsLevel == 0) ? forceWithVirialMtsLevel0 : forceWithVirialMtsLevel1;
 694             enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
 695                     inputrec.pbcType,
 696                     gmx::arrayRefFromArray(mdatoms->massT, mdatoms->nr),
 697                     foreignLambdaDeltaH,
 698                     foreignLambdaDhDl,
 699                     box,
 700                     forceWithVirial,
 701                     t,
 702                     step,
 703                     wcycle,
 704                     fplog);
 705         }
 706     }
 707     /* Add the forces from enforced rotation potentials (if any) */
 708     if (inputrec.bRot)
 709     {
 710         wallcycle_start(wcycle, WallCycleCounter::RotAdd);
 711         enerd->term[F_COM_PULL] +=
 712                 add_rot_forces(enforcedRotation, forceWithVirialMtsLevel0->force_, cr, step, t);
 713         wallcycle_stop(wcycle, WallCycleCounter::RotAdd);
 714     }
 715
 716     if (ed)
 717     {
 718         /* Note that since init_edsam() is called after the initialization
 719          * of forcerec, edsam doesn't request the noVirSum force buffer.
 720          * Thus if no other algorithm (e.g. PME) requires it, the forces
 721          * here will contribute to the virial.
 722          */
 723         do_flood(cr, inputrec, x, forceWithVirialMtsLevel0->force_, ed, box, step, didNeighborSearch);
 724     }
 725
 726     /* Add forces from interactive molecular dynamics (IMD), if any */
 727     if (inputrec.bIMD && stepWork.computeForces)
 728     {
 729         imdSession->applyForces(forceWithVirialMtsLevel0->force_);
 730     }
 731 }
 732
 733 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 734  *
 735  * \param[in]  pmedata              The PME structure
 736  * \param[in]  box                  The box matrix
 737  * \param[in]  stepWork             Step schedule flags
 738  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in the device memory.
 739  * \param[in]  lambdaQ              The Coulomb lambda of the current state.
 740  * \param[in]  wcycle               The wallcycle structure
 741  */
 742 static inline void launchPmeGpuSpread(gmx_pme_t*            pmedata,
 743                                       const matrix          box,
 744                                       const StepWorkload&   stepWork,
 745                                       GpuEventSynchronizer* xReadyOnDevice,
 746                                       const real            lambdaQ,
 747                                       gmx_wallcycle*        wcycle)
 748 {
 749     pme_gpu_prepare_computation(pmedata, box, wcycle, stepWork);
 750     bool                           useGpuDirectComm         = false;
 751     gmx::PmeCoordinateReceiverGpu* pmeCoordinateReceiverGpu = nullptr;
 752     pme_gpu_launch_spread(
 753             pmedata, xReadyOnDevice, wcycle, lambdaQ, useGpuDirectComm, pmeCoordinateReceiverGpu);
 754 }
 755
 756 /*! \brief Launch the FFT and gather stages of PME GPU
 757  *
 758  * This function only implements setting the output forces (no accumulation).
 759  *
 760  * \param[in]  pmedata        The PME structure
 761  * \param[in]  lambdaQ        The Coulomb lambda of the current system state.
 762  * \param[in]  wcycle         The wallcycle structure
 763  * \param[in]  stepWork       Step schedule flags
 764  */
 765 static void launchPmeGpuFftAndGather(gmx_pme_t*               pmedata,
 766                                      const real               lambdaQ,
 767                                      gmx_wallcycle*           wcycle,
 768                                      const gmx::StepWorkload& stepWork)
 769 {
 770     pme_gpu_launch_complex_transforms(pmedata, wcycle, stepWork);
 771     pme_gpu_launch_gather(pmedata, wcycle, lambdaQ);
 772 }
 773
 774 /*! \brief
 775  *  Polling wait for either of the PME or nonbonded GPU tasks.
 776  *
 777  * Instead of a static order in waiting for GPU tasks, this function
 778  * polls checking which of the two tasks completes first, and does the
 779  * associated force buffer reduction overlapped with the other task.
 780  * By doing that, unlike static scheduling order, it can always overlap
 781  * one of the reductions, regardless of the GPU task completion order.
 782  *
 783  * \param[in]     nbv              Nonbonded verlet structure
 784  * \param[in,out] pmedata          PME module data
 785  * \param[in,out] forceOutputsNonbonded  Force outputs for the non-bonded forces and shift forces
 786  * \param[in,out] forceOutputsPme  Force outputs for the PME forces and virial
 787  * \param[in,out] enerd            Energy data structure results are reduced into
 788  * \param[in]     lambdaQ          The Coulomb lambda of the current system state.
 789  * \param[in]     stepWork         Step schedule flags
 790  * \param[in]     wcycle           The wallcycle structure
 791  */
 792 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
 793                                         gmx_pme_t*          pmedata,
 794                                         gmx::ForceOutputs*  forceOutputsNonbonded,
 795                                         gmx::ForceOutputs*  forceOutputsPme,
 796                                         gmx_enerdata_t*     enerd,
 797                                         const real          lambdaQ,
 798                                         const StepWorkload& stepWork,
 799                                         gmx_wallcycle*      wcycle)
 800 {
 801     bool isPmeGpuDone = false;
 802     bool isNbGpuDone  = false;
 803
 804     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 805
 806     while (!isPmeGpuDone || !isNbGpuDone)
 807     {
 808         if (!isPmeGpuDone)
 809         {
 810             GpuTaskCompletion completionType =
 811                     (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 812             isPmeGpuDone = pme_gpu_try_finish_task(
 813                     pmedata, stepWork, wcycle, &forceOutputsPme->forceWithVirial(), enerd, lambdaQ, completionType);
 814         }
 815
 816         if (!isNbGpuDone)
 817         {
 818             auto&             forceBuffersNonbonded = forceOutputsNonbonded->forceWithShiftForces();
 819             GpuTaskCompletion completionType =
 820                     (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 821             isNbGpuDone = Nbnxm::gpu_try_finish_task(
 822                     nbv->gpu_nbv,
 823                     stepWork,
 824                     AtomLocality::Local,
 825                     enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR].data(),
 826                     enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR].data(),
 827                     forceBuffersNonbonded.shiftForces(),
 828                     completionType,
 829                     wcycle);
 830
 831             if (isNbGpuDone)
 832             {
 833                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceBuffersNonbonded.force());
 834             }
 835         }
 836     }
 837 }
 838
 839 /*! \brief Set up the different force buffers; also does clearing.
 840  *
 841  * \param[in] forceHelperBuffers        Helper force buffers
 842  * \param[in] force                     force array
 843  * \param[in] domainWork                Domain lifetime workload flags
 844  * \param[in] stepWork                  Step schedule flags
 845  * \param[in] havePpDomainDecomposition Whether we have a PP domain decomposition
 846  * \param[out] wcycle                   wallcycle recording structure
 847  *
 848  * \returns                             Cleared force output structure
 849  */
 850 static ForceOutputs setupForceOutputs(ForceHelperBuffers*                 forceHelperBuffers,
 851                                       gmx::ArrayRefWithPadding<gmx::RVec> force,
 852                                       const DomainLifetimeWorkload&       domainWork,
 853                                       const StepWorkload&                 stepWork,
 854                                       const bool                          havePpDomainDecomposition,
 855                                       gmx_wallcycle*                      wcycle)
 856 {
 857     wallcycle_sub_start(wcycle, WallCycleSubCounter::ClearForceBuffer);
 858
 859     /* NOTE: We assume fr->shiftForces is all zeros here */
 860     gmx::ForceWithShiftForces forceWithShiftForces(
 861             force, stepWork.computeVirial, forceHelperBuffers->shiftForces());
 862
 863     if (stepWork.computeForces
 864         && (domainWork.haveCpuLocalForceWork || !stepWork.useGpuFBufferOps
 865             || (havePpDomainDecomposition && !stepWork.useGpuFHalo)))
 866     {
 867         /* Clear the short- and long-range forces */
 868         clearRVecs(forceWithShiftForces.force(), true);
 869
 870         /* Clear the shift forces */
 871         clearRVecs(forceWithShiftForces.shiftForces(), false);
 872     }
 873
 874     /* If we need to compute the virial, we might need a separate
 875      * force buffer for algorithms for which the virial is calculated
 876      * directly, such as PME. Otherwise, forceWithVirial uses the
 877      * the same force (f in legacy calls) buffer as other algorithms.
 878      */
 879     const bool useSeparateForceWithVirialBuffer =
 880             (stepWork.computeForces
 881              && (stepWork.computeVirial && forceHelperBuffers->haveDirectVirialContributions()));
 882     /* forceWithVirial uses the local atom range only */
 883     gmx::ForceWithVirial forceWithVirial(
 884             useSeparateForceWithVirialBuffer ? forceHelperBuffers->forceBufferForDirectVirialContributions()
 885                                              : force.unpaddedArrayRef(),
 886             stepWork.computeVirial);
 887
 888     if (useSeparateForceWithVirialBuffer)
 889     {
 890         /* TODO: update comment
 891          * We only compute forces on local atoms. Note that vsites can
 892          * spread to non-local atoms, but that part of the buffer is
 893          * cleared separately in the vsite spreading code.
 894          */
 895         clearRVecs(forceWithVirial.force_, true);
 896     }
 897
 898     wallcycle_sub_stop(wcycle, WallCycleSubCounter::ClearForceBuffer);
 899
 900     return ForceOutputs(
 901             forceWithShiftForces, forceHelperBuffers->haveDirectVirialContributions(), forceWithVirial);
 902 }
 903
 904
 905 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 906  */
 907 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec&         inputrec,
 908                                                           const t_forcerec&         fr,
 909                                                           const pull_t*             pull_work,
 910                                                           const gmx_edsam*          ed,
 911                                                           const t_mdatoms&          mdatoms,
 912                                                           const SimulationWorkload& simulationWork,
 913                                                           const StepWorkload&       stepWork)
 914 {
 915     DomainLifetimeWorkload domainWork;
 916     // Note that haveSpecialForces is constant over the whole run
 917     domainWork.haveSpecialForces =
 918             haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 919     domainWork.haveCpuListedForceWork = false;
 920     domainWork.haveCpuBondedWork      = false;
 921     for (const auto& listedForces : fr.listedForces)
 922     {
 923         if (listedForces.haveCpuListedForces(*fr.fcdata))
 924         {
 925             domainWork.haveCpuListedForceWork = true;
 926         }
 927         if (listedForces.haveCpuBondeds())
 928         {
 929             domainWork.haveCpuBondedWork = true;
 930         }
 931     }
 932     domainWork.haveGpuBondedWork =
 933             ((fr.listedForcesGpu != nullptr) && fr.listedForcesGpu->haveInteractions());
 934     // Note that haveFreeEnergyWork is constant over the whole run
 935     domainWork.haveFreeEnergyWork =
 936             (fr.efep != FreeEnergyPerturbationType::No && mdatoms.nPerturbed != 0);
 937     // We assume we have local force work if there are CPU
 938     // force tasks including PME or nonbondeds.
 939     domainWork.haveCpuLocalForceWork =
 940             domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
 941             || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
 942             || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
 943     domainWork.haveLocalForceContribInCpuBuffer =
 944             domainWork.haveCpuLocalForceWork || simulationWork.havePpDomainDecomposition;
 945     domainWork.haveNonLocalForceContribInCpuBuffer =
 946             domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
 947
 948     return domainWork;
 949 }
 950
 951 /*! \brief Set up force flag stuct from the force bitmask.
 952  *
 953  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 954  * \param[in]      mtsLevels            The multiple time-stepping levels, either empty or 2 levels
 955  * \param[in]      step                 The current MD step
 956  * \param[in]      simulationWork       Simulation workload description.
 957  *
 958  * \returns New Stepworkload description.
 959  */
 960 static StepWorkload setupStepWorkload(const int                     legacyFlags,
 961                                       ArrayRef<const gmx::MtsLevel> mtsLevels,
 962                                       const int64_t                 step,
 963                                       const SimulationWorkload&     simulationWork)
 964 {
 965     GMX_ASSERT(mtsLevels.empty() || mtsLevels.size() == 2, "Expect 0 or 2 MTS levels");
 966     const bool computeSlowForces = (mtsLevels.empty() || step % mtsLevels[1].stepFactor == 0);
 967
 968     StepWorkload flags;
 969     flags.stateChanged                  = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 970     flags.haveDynamicBox                = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 971     flags.doNeighborSearch              = ((legacyFlags & GMX_FORCE_NS) != 0);
 972     flags.computeSlowForces             = computeSlowForces;
 973     flags.computeVirial                 = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 974     flags.computeEnergy                 = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 975     flags.computeForces                 = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 976     flags.useOnlyMtsCombinedForceBuffer = ((legacyFlags & GMX_FORCE_DO_NOT_NEED_NORMAL_FORCE) != 0);
 977     flags.computeListedForces           = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 978     flags.computeNonbondedForces =
 979             ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && simulationWork.computeNonbonded
 980             && !(simulationWork.computeNonbondedAtMtsLevel1 && !computeSlowForces);
 981     flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 982
 983     if (simulationWork.useGpuBufferOps)
 984     {
 985         GMX_ASSERT(simulationWork.useGpuNonbonded,
 986                    "Can only offload buffer ops if nonbonded computation is also offloaded");
 987     }
 988     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 989     // on virial steps the CPU reduction path is taken
 990     flags.useGpuFBufferOps       = simulationWork.useGpuBufferOps && !flags.computeVirial;
 991     const bool rankHasGpuPmeTask = simulationWork.useGpuPme && !simulationWork.haveSeparatePmeRank;
 992     flags.useGpuPmeFReduction    = flags.computeSlowForces && flags.useGpuFBufferOps
 993                                 && (rankHasGpuPmeTask || simulationWork.useGpuPmePpCommunication);
 994     flags.useGpuXHalo          = simulationWork.useGpuHaloExchange && !flags.doNeighborSearch;
 995     flags.useGpuFHalo          = simulationWork.useGpuHaloExchange && flags.useGpuFBufferOps;
 996     flags.haveGpuPmeOnThisRank = rankHasGpuPmeTask && flags.computeSlowForces;
 997     flags.combineMtsForcesBeforeHaloExchange =
 998             (flags.computeForces && simulationWork.useMts && flags.computeSlowForces
 999              && flags.useOnlyMtsCombinedForceBuffer
1000              && !(flags.computeVirial || simulationWork.useGpuNonbonded || flags.haveGpuPmeOnThisRank));
1001
1002     return flags;
1003 }
1004
1005
1006 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
1007  *
1008  */
1009 static void launchGpuEndOfStepTasks(nonbonded_verlet_t*               nbv,
1010                                     gmx::ListedForcesGpu*             listedForcesGpu,
1011                                     gmx_pme_t*                        pmedata,
1012                                     gmx_enerdata_t*                   enerd,
1013                                     const gmx::MdrunScheduleWorkload& runScheduleWork,
1014                                     int64_t                           step,
1015                                     gmx_wallcycle*                    wcycle)
1016 {
1017     if (runScheduleWork.simulationWork.useGpuNonbonded && runScheduleWork.stepWork.computeNonbondedForces)
1018     {
1019         /* Launch pruning before buffer clearing because the API overhead of the
1020          * clear kernel launches can leave the GPU idle while it could be running
1021          * the prune kernel.
1022          */
1023         if (nbv->isDynamicPruningStepGpu(step))
1024         {
1025             nbv->dispatchPruneKernelGpu(step);
1026         }
1027
1028         /* now clear the GPU outputs while we finish the step on the CPU */
1029         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1030         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1031         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
1032         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1033         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1034     }
1035
1036     if (runScheduleWork.stepWork.haveGpuPmeOnThisRank)
1037     {
1038         pme_gpu_reinit_computation(pmedata, wcycle);
1039     }
1040
1041     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
1042     {
1043         // in principle this should be included in the DD balancing region,
1044         // but generally it is infrequent so we'll omit it for the sake of
1045         // simpler code
1046         listedForcesGpu->waitAccumulateEnergyTerms(enerd);
1047
1048         listedForcesGpu->clearEnergies();
1049     }
1050 }
1051
1052 //! \brief Data structure to hold dipole-related data and staging arrays
1053 struct DipoleData
1054 {
1055     //! Dipole staging for fast summing over MPI
1056     gmx::DVec muStaging[2] = { { 0.0, 0.0, 0.0 } };
1057     //! Dipole staging for states A and B (index 0 and 1 resp.)
1058     gmx::RVec muStateAB[2] = { { 0.0_real, 0.0_real, 0.0_real } };
1059 };
1060
1061
1062 static void reduceAndUpdateMuTot(DipoleData*                   dipoleData,
1063                                  const t_commrec*              cr,
1064                                  const bool                    haveFreeEnergy,
1065                                  gmx::ArrayRef<const real>     lambda,
1066                                  rvec                          muTotal,
1067                                  const DDBalanceRegionHandler& ddBalanceRegionHandler)
1068 {
1069     if (PAR(cr))
1070     {
1071         gmx_sumd(2 * DIM, dipoleData->muStaging[0], cr);
1072         ddBalanceRegionHandler.reopenRegionCpu();
1073     }
1074     for (int i = 0; i < 2; i++)
1075     {
1076         for (int j = 0; j < DIM; j++)
1077         {
1078             dipoleData->muStateAB[i][j] = dipoleData->muStaging[i][j];
1079         }
1080     }
1081
1082     if (!haveFreeEnergy)
1083     {
1084         copy_rvec(dipoleData->muStateAB[0], muTotal);
1085     }
1086     else
1087     {
1088         for (int j = 0; j < DIM; j++)
1089         {
1090             muTotal[j] = (1.0 - lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)])
1091                                  * dipoleData->muStateAB[0][j]
1092                          + lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)]
1093                                    * dipoleData->muStateAB[1][j];
1094         }
1095     }
1096 }
1097
1098 /*! \brief Combines MTS level0 and level1 force buffers into a full and MTS-combined force buffer.
1099  *
1100  * \param[in]     numAtoms        The number of atoms to combine forces for
1101  * \param[in,out] forceMtsLevel0  Input: F_level0, output: F_level0 + F_level1
1102  * \param[in,out] forceMts        Input: F_level1, output: F_level0 + mtsFactor * F_level1
1103  * \param[in]     mtsFactor       The factor between the level0 and level1 time step
1104  */
1105 static void combineMtsForces(const int      numAtoms,
1106                              ArrayRef<RVec> forceMtsLevel0,
1107                              ArrayRef<RVec> forceMts,
1108                              const real     mtsFactor)
1109 {
1110     const int gmx_unused numThreads = gmx_omp_nthreads_get(ModuleMultiThread::Default);
1111 #pragma omp parallel for num_threads(numThreads) schedule(static)
1112     for (int i = 0; i < numAtoms; i++)
1113     {
1114         const RVec forceMtsLevel0Tmp = forceMtsLevel0[i];
1115         forceMtsLevel0[i] += forceMts[i];
1116         forceMts[i] = forceMtsLevel0Tmp + mtsFactor * forceMts[i];
1117     }
1118 }
1119
1120 /*! \brief Setup for the local and non-local GPU force reductions:
1121  * reinitialization plus the registration of forces and dependencies.
1122  *
1123  * \param [in] runScheduleWork               Schedule workload flag structure
1124  * \param [in] cr                            Communication record object
1125  * \param [in] fr                            Force record object
1126  */
1127 static void setupGpuForceReductions(gmx::MdrunScheduleWorkload* runScheduleWork,
1128                                     const t_commrec*            cr,
1129                                     t_forcerec*                 fr)
1130 {
1131
1132     nonbonded_verlet_t*          nbv      = fr->nbv.get();
1133     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1134
1135     // (re-)initialize local GPU force reduction
1136     const bool accumulate = runScheduleWork->domainWork.haveCpuLocalForceWork
1137                             || runScheduleWork->simulationWork.havePpDomainDecomposition;
1138     const int atomStart = 0;
1139     fr->gpuForceReduction[gmx::AtomLocality::Local]->reinit(
1140             stateGpu->getForces(),
1141             nbv->getNumAtoms(AtomLocality::Local),
1142             nbv->getGridIndices(),
1143             atomStart,
1144             accumulate,
1145             stateGpu->fReducedOnDevice(AtomLocality::Local));
1146
1147     // register forces and add dependencies
1148     fr->gpuForceReduction[gmx::AtomLocality::Local]->registerNbnxmForce(Nbnxm::gpu_get_f(nbv->gpu_nbv));
1149
1150     if (runScheduleWork->simulationWork.useGpuPme
1151         && (!runScheduleWork->simulationWork.haveSeparatePmeRank
1152             || runScheduleWork->simulationWork.useGpuPmePpCommunication))
1153     {
1154         DeviceBuffer<gmx::RVec> forcePtr =
1155                 runScheduleWork->simulationWork.haveSeparatePmeRank
1156                         ? fr->pmePpCommGpu->getGpuForceStagingPtr() // buffer received from other GPU
1157                         : pme_gpu_get_device_f(fr->pmedata);        // PME force buffer on same GPU
1158         fr->gpuForceReduction[gmx::AtomLocality::Local]->registerRvecForce(forcePtr);
1159
1160         if (runScheduleWork->simulationWork.haveSeparatePmeRank)
1161         {
1162             // PME force buffer on remote GPU -
1163             // event synchronizer received from other GPU only in case of thread-mpi
1164             if (GMX_THREAD_MPI)
1165             {
1166                 GpuEventSynchronizer* const pmeSynchronizer =
1167                         fr->pmePpCommGpu->getForcesReadySynchronizer();
1168                 GMX_ASSERT(pmeSynchronizer != nullptr,
1169                            "PME force ready cuda event should not be NULL");
1170                 fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(pmeSynchronizer);
1171             }
1172         }
1173         else
1174         {
1175             // PME force buffer on same GPU - add dependency on PME force computation
1176             GpuEventSynchronizer* const pmeSynchronizer = pme_gpu_get_f_ready_synchronizer(fr->pmedata);
1177             GMX_ASSERT(pmeSynchronizer != nullptr, "PME force ready cuda event should not be NULL");
1178             fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(pmeSynchronizer);
1179         }
1180     }
1181
1182     if (runScheduleWork->domainWork.haveCpuLocalForceWork
1183         || (runScheduleWork->simulationWork.havePpDomainDecomposition
1184             && !runScheduleWork->simulationWork.useGpuHaloExchange))
1185     {
1186         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1187                 stateGpu->fReadyOnDevice(AtomLocality::Local));
1188     }
1189
1190     if (runScheduleWork->simulationWork.useGpuHaloExchange)
1191     {
1192         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1193                 cr->dd->gpuHaloExchange[0][0]->getForcesReadyOnDeviceEvent());
1194     }
1195
1196     if (runScheduleWork->simulationWork.havePpDomainDecomposition)
1197     {
1198         // (re-)initialize non-local GPU force reduction
1199         const bool accumulate = runScheduleWork->domainWork.haveCpuBondedWork
1200                                 || runScheduleWork->domainWork.haveFreeEnergyWork;
1201         const int atomStart = dd_numHomeAtoms(*cr->dd);
1202         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->reinit(
1203                 stateGpu->getForces(),
1204                 nbv->getNumAtoms(AtomLocality::NonLocal),
1205                 nbv->getGridIndices(),
1206                 atomStart,
1207                 accumulate,
1208                 stateGpu->fReducedOnDevice(AtomLocality::NonLocal));
1209
1210         // register forces and add dependencies
1211         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->registerNbnxmForce(
1212                 Nbnxm::gpu_get_f(nbv->gpu_nbv));
1213
1214         if (runScheduleWork->domainWork.haveNonLocalForceContribInCpuBuffer)
1215         {
1216             fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->addDependency(
1217                     stateGpu->fReadyOnDevice(AtomLocality::NonLocal));
1218         }
1219     }
1220 }
1221
1222
1223 /*! \brief Return the number of local atoms.
1224  */
1225 static int getLocalAtomCount(const gmx_domdec_t* dd, const t_mdatoms& mdatoms, bool havePPDomainDecomposition)
1226 {
1227     GMX_ASSERT(!(havePPDomainDecomposition && (dd == nullptr)),
1228                "Can't have PP decomposition with dd uninitialized!");
1229     return havePPDomainDecomposition ? dd_numAtomsZones(*dd) : mdatoms.homenr;
1230 }
1231
1232
1233 void do_force(FILE*                               fplog,
1234               const t_commrec*                    cr,
1235               const gmx_multisim_t*               ms,
1236               const t_inputrec&                   inputrec,
1237               gmx::Awh*                           awh,
1238               gmx_enfrot*                         enforcedRotation,
1239               gmx::ImdSession*                    imdSession,
1240               pull_t*                             pull_work,
1241               int64_t                             step,
1242               t_nrnb*                             nrnb,
1243               gmx_wallcycle*                      wcycle,
1244               const gmx_localtop_t*               top,
1245               const matrix                        box,
1246               gmx::ArrayRefWithPadding<gmx::RVec> x,
1247               const history_t*                    hist,
1248               gmx::ForceBuffersView*              forceView,
1249               tensor                              vir_force,
1250               const t_mdatoms*                    mdatoms,
1251               gmx_enerdata_t*                     enerd,
1252               gmx::ArrayRef<const real>           lambda,
1253               t_forcerec*                         fr,
1254               gmx::MdrunScheduleWorkload*         runScheduleWork,
1255               gmx::VirtualSitesHandler*           vsite,
1256               rvec                                muTotal,
1257               double                              t,
1258               gmx_edsam*                          ed,
1259               CpuPpLongRangeNonbondeds*           longRangeNonbondeds,
1260               int                                 legacyFlags,
1261               const DDBalanceRegionHandler&       ddBalanceRegionHandler)
1262 {
1263     auto force = forceView->forceWithPadding();
1264     GMX_ASSERT(force.unpaddedArrayRef().ssize() >= fr->natoms_force_constr,
1265                "The size of the force buffer should be at least the number of atoms to compute "
1266                "forces for");
1267
1268     nonbonded_verlet_t*  nbv = fr->nbv.get();
1269     interaction_const_t* ic  = fr->ic.get();
1270
1271     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1272
1273     const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
1274
1275     runScheduleWork->stepWork = setupStepWorkload(legacyFlags, inputrec.mtsLevels, step, simulationWork);
1276     const StepWorkload& stepWork = runScheduleWork->stepWork;
1277
1278     if (stepWork.useGpuFHalo && !runScheduleWork->domainWork.haveCpuLocalForceWork)
1279     {
1280         // GPU Force halo exchange will set a subset of local atoms with remote non-local data
1281         // First clear local portion of force array, so that untouched atoms are zero.
1282         // The dependency for this is that forces from previous timestep have been consumed,
1283         // which is satisfied when getCoordinatesReadyOnDeviceEvent has been marked.
1284         stateGpu->clearForcesOnGpu(AtomLocality::Local,
1285                                    stateGpu->getCoordinatesReadyOnDeviceEvent(
1286                                            AtomLocality::Local, simulationWork, stepWork));
1287     }
1288
1289     /* At a search step we need to start the first balancing region
1290      * somewhere early inside the step after communication during domain
1291      * decomposition (and not during the previous step as usual).
1292      */
1293     if (stepWork.doNeighborSearch)
1294     {
1295         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
1296     }
1297
1298     clear_mat(vir_force);
1299
1300     if (fr->pbcType != PbcType::No)
1301     {
1302         /* Compute shift vectors every step,
1303          * because of pressure coupling or box deformation!
1304          */
1305         if (stepWork.haveDynamicBox && stepWork.stateChanged)
1306         {
1307             calc_shifts(box, fr->shift_vec);
1308         }
1309
1310         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1311         const bool calcCGCM = (fillGrid && !haveDDAtomOrdering(*cr));
1312         if (calcCGCM)
1313         {
1314             put_atoms_in_box_omp(fr->pbcType,
1315                                  box,
1316                                  x.unpaddedArrayRef().subArray(0, mdatoms->homenr),
1317                                  gmx_omp_nthreads_get(ModuleMultiThread::Default));
1318             inc_nrnb(nrnb, eNR_SHIFTX, mdatoms->homenr);
1319         }
1320     }
1321
1322     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1323
1324     const bool pmeSendCoordinatesFromGpu =
1325             simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1326     const bool reinitGpuPmePpComms =
1327             simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1328
1329     auto* localXReadyOnDevice = (stepWork.haveGpuPmeOnThisRank || simulationWork.useGpuBufferOps)
1330                                         ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1331                                                 AtomLocality::Local, simulationWork, stepWork)
1332                                         : nullptr;
1333
1334     GMX_ASSERT(simulationWork.useGpuHaloExchange
1335                        == ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange[0].empty())),
1336                "The GPU halo exchange is active, but it has not been constructed.");
1337
1338     bool gmx_used_in_debug haveCopiedXFromGpu = false;
1339     // Copy coordinate from the GPU if update is on the GPU and there
1340     // are forces to be computed on the CPU, or for the computation of
1341     // virial, or if host-side data will be transferred from this task
1342     // to a remote task for halo exchange or PME-PP communication. At
1343     // search steps the current coordinates are already on the host,
1344     // hence copy is not needed.
1345     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1346         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1347             || simulationWork.useCpuPmePpCommunication || simulationWork.useCpuHaloExchange
1348             || simulationWork.computeMuTot))
1349     {
1350         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1351         haveCopiedXFromGpu = true;
1352     }
1353
1354     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1355     // The local coordinates can be copied right away.
1356     // NOTE: Consider moving this copy to right after they are updated and constrained,
1357     //       if the later is not offloaded.
1358     if (stepWork.haveGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1359     {
1360         if (stepWork.doNeighborSearch)
1361         {
1362             // TODO refactor this to do_md, after partitioning.
1363             stateGpu->reinit(mdatoms->homenr,
1364                              getLocalAtomCount(cr->dd, *mdatoms, simulationWork.havePpDomainDecomposition));
1365             if (stepWork.haveGpuPmeOnThisRank)
1366             {
1367                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1368                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1369             }
1370         }
1371         // We need to copy coordinates when:
1372         // 1. Update is not offloaded
1373         // 2. The buffers were reinitialized on search step
1374         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1375         {
1376             GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1377             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1378         }
1379     }
1380
1381     if (simulationWork.haveSeparatePmeRank && stepWork.computeSlowForces)
1382     {
1383         /* Send particle coordinates to the pme nodes */
1384         if (!pmeSendCoordinatesFromGpu && !stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1385         {
1386             GMX_ASSERT(haveCopiedXFromGpu,
1387                        "a wait should only be triggered if copy has been scheduled");
1388             stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1389         }
1390
1391         gmx_pme_send_coordinates(fr,
1392                                  cr,
1393                                  box,
1394                                  x.unpaddedArrayRef(),
1395                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1396                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)],
1397                                  (stepWork.computeVirial || stepWork.computeEnergy),
1398                                  step,
1399                                  simulationWork.useGpuPmePpCommunication,
1400                                  reinitGpuPmePpComms,
1401                                  pmeSendCoordinatesFromGpu,
1402                                  stepWork.useGpuPmeFReduction,
1403                                  localXReadyOnDevice,
1404                                  wcycle);
1405     }
1406
1407     if (stepWork.haveGpuPmeOnThisRank)
1408     {
1409         launchPmeGpuSpread(fr->pmedata,
1410                            box,
1411                            stepWork,
1412                            localXReadyOnDevice,
1413                            lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1414                            wcycle);
1415     }
1416
1417     const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1418
1419     /* do gridding for pair search */
1420     if (stepWork.doNeighborSearch)
1421     {
1422         if (fr->wholeMoleculeTransform && stepWork.stateChanged)
1423         {
1424             fr->wholeMoleculeTransform->updateForAtomPbcJumps(x.unpaddedArrayRef(), box);
1425         }
1426
1427         wallcycle_start(wcycle, WallCycleCounter::NS);
1428         if (!haveDDAtomOrdering(*cr))
1429         {
1430             const rvec vzero       = { 0.0_real, 0.0_real, 0.0_real };
1431             const rvec boxDiagonal = { box[XX][XX], box[YY][YY], box[ZZ][ZZ] };
1432             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSGridLocal);
1433             nbnxn_put_on_grid(nbv,
1434                               box,
1435                               0,
1436                               vzero,
1437                               boxDiagonal,
1438                               nullptr,
1439                               { 0, mdatoms->homenr },
1440                               -1,
1441                               fr->atomInfo,
1442                               x.unpaddedArrayRef(),
1443                               0,
1444                               nullptr);
1445             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSGridLocal);
1446         }
1447         else
1448         {
1449             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSGridNonLocal);
1450             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->atomInfo, x.unpaddedArrayRef());
1451             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSGridNonLocal);
1452         }
1453
1454         nbv->setAtomProperties(gmx::constArrayRefFromArray(mdatoms->typeA, mdatoms->nr),
1455                                gmx::constArrayRefFromArray(mdatoms->chargeA, mdatoms->nr),
1456                                fr->atomInfo);
1457
1458         wallcycle_stop(wcycle, WallCycleCounter::NS);
1459
1460         /* initialize the GPU nbnxm atom data and bonded data structures */
1461         if (simulationWork.useGpuNonbonded)
1462         {
1463             // Note: cycle counting only nononbondeds, GPU listed forces counts internally
1464             wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1465             wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1466             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1467             wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1468             wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1469
1470             if (fr->listedForcesGpu)
1471             {
1472                 /* Now we put all atoms on the grid, we can assign bonded
1473                  * interactions to the GPU, where the grid order is
1474                  * needed. Also the xq, f and fshift device buffers have
1475                  * been reallocated if needed, so the bonded code can
1476                  * learn about them. */
1477                 // TODO the xq, f, and fshift buffers are now shared
1478                 // resources, so they should be maintained by a
1479                 // higher-level object than the nb module.
1480                 fr->listedForcesGpu->updateInteractionListsAndDeviceBuffers(
1481                         nbv->getGridIndices(),
1482                         top->idef,
1483                         Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1484                         Nbnxm::gpu_get_f(nbv->gpu_nbv),
1485                         Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1486             }
1487         }
1488
1489         // Need to run after the GPU-offload bonded interaction lists
1490         // are set up to be able to determine whether there is bonded work.
1491         runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1492                 inputrec, *fr, pull_work, ed, *mdatoms, simulationWork, stepWork);
1493
1494         wallcycle_start_nocount(wcycle, WallCycleCounter::NS);
1495         wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSSearchLocal);
1496         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1497         nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1498
1499         nbv->setupGpuShortRangeWork(fr->listedForcesGpu.get(), InteractionLocality::Local);
1500
1501         wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSSearchLocal);
1502         wallcycle_stop(wcycle, WallCycleCounter::NS);
1503
1504         if (stepWork.useGpuXBufferOps)
1505         {
1506             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1507         }
1508
1509         if (simulationWork.useGpuBufferOps)
1510         {
1511             setupGpuForceReductions(runScheduleWork, cr, fr);
1512         }
1513     }
1514     else if (!EI_TPI(inputrec.eI) && stepWork.computeNonbondedForces)
1515     {
1516         if (stepWork.useGpuXBufferOps)
1517         {
1518             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1519             nbv->convertCoordinatesGpu(AtomLocality::Local, stateGpu->getCoordinates(), localXReadyOnDevice);
1520         }
1521         else
1522         {
1523             if (simulationWork.useGpuUpdate)
1524             {
1525                 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1526                 GMX_ASSERT(haveCopiedXFromGpu,
1527                            "a wait should only be triggered if copy has been scheduled");
1528                 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1529             }
1530             nbv->convertCoordinates(AtomLocality::Local, x.unpaddedArrayRef());
1531         }
1532     }
1533
1534     if (simulationWork.useGpuNonbonded && (stepWork.computeNonbondedForces || domainWork.haveGpuBondedWork))
1535     {
1536         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1537
1538         wallcycle_start(wcycle, WallCycleCounter::LaunchGpu);
1539         wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1540         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1541         if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1542         {
1543             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1544         }
1545         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1546         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1547         // with X buffer ops offloaded to the GPU on all but the search steps
1548
1549         // bonded work not split into separate local and non-local, so with DD
1550         // we can only launch the kernel after non-local coordinates have been received.
1551         if (domainWork.haveGpuBondedWork && !simulationWork.havePpDomainDecomposition)
1552         {
1553             fr->listedForcesGpu->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1554         }
1555
1556         /* launch local nonbonded work on GPU */
1557         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1558         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1559         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1560         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1561         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1562     }
1563
1564     if (stepWork.haveGpuPmeOnThisRank)
1565     {
1566         // In PME GPU and mixed mode we launch FFT / gather after the
1567         // X copy/transform to allow overlap as well as after the GPU NB
1568         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1569         // the nonbonded kernel.
1570         launchPmeGpuFftAndGather(fr->pmedata,
1571                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1572                                  wcycle,
1573                                  stepWork);
1574     }
1575
1576     /* Communicate coordinates and sum dipole if necessary +
1577        do non-local pair search */
1578     if (simulationWork.havePpDomainDecomposition)
1579     {
1580         if (stepWork.doNeighborSearch)
1581         {
1582             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1583             wallcycle_start_nocount(wcycle, WallCycleCounter::NS);
1584             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSSearchNonLocal);
1585             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1586             nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1587
1588             nbv->setupGpuShortRangeWork(fr->listedForcesGpu.get(), InteractionLocality::NonLocal);
1589             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSSearchNonLocal);
1590             wallcycle_stop(wcycle, WallCycleCounter::NS);
1591             // TODO refactor this GPU halo exchange re-initialisation
1592             // to location in do_md where GPU halo exchange is
1593             // constructed at partitioning, after above stateGpu
1594             // re-initialization has similarly been refactored
1595             if (simulationWork.useGpuHaloExchange)
1596             {
1597                 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1598             }
1599         }
1600         else
1601         {
1602             GpuEventSynchronizer* gpuCoordinateHaloLaunched = nullptr;
1603             if (stepWork.useGpuXHalo)
1604             {
1605                 // The following must be called after local setCoordinates (which records an event
1606                 // when the coordinate data has been copied to the device).
1607                 gpuCoordinateHaloLaunched = communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1608
1609                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1610                 {
1611                     // non-local part of coordinate buffer must be copied back to host for CPU work
1612                     stateGpu->copyCoordinatesFromGpu(
1613                             x.unpaddedArrayRef(), AtomLocality::NonLocal, gpuCoordinateHaloLaunched);
1614                 }
1615             }
1616             else
1617             {
1618                 if (simulationWork.useGpuUpdate)
1619                 {
1620                     GMX_ASSERT(haveCopiedXFromGpu,
1621                                "a wait should only be triggered if copy has been scheduled");
1622                     stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1623                 }
1624                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1625             }
1626
1627             if (stepWork.useGpuXBufferOps)
1628             {
1629                 if (!stepWork.useGpuXHalo)
1630                 {
1631                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1632                 }
1633                 nbv->convertCoordinatesGpu(
1634                         AtomLocality::NonLocal,
1635                         stateGpu->getCoordinates(),
1636                         stateGpu->getCoordinatesReadyOnDeviceEvent(
1637                                 AtomLocality::NonLocal, simulationWork, stepWork, gpuCoordinateHaloLaunched));
1638             }
1639             else
1640             {
1641                 nbv->convertCoordinates(AtomLocality::NonLocal, x.unpaddedArrayRef());
1642             }
1643         }
1644
1645         if (simulationWork.useGpuNonbonded)
1646         {
1647
1648             if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1649             {
1650                 wallcycle_start(wcycle, WallCycleCounter::LaunchGpu);
1651                 wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1652                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1653                 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1654                 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1655             }
1656
1657             if (domainWork.haveGpuBondedWork)
1658             {
1659                 fr->listedForcesGpu->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1660             }
1661
1662             /* launch non-local nonbonded tasks on GPU */
1663             wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1664             wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1665             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step, nrnb, wcycle);
1666             wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1667             wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1668         }
1669     }
1670
1671     // With FEP we set up the reduction over threads for local+non-local simultaneously,
1672     // so we need to do that here after the local and non-local pairlist construction.
1673     if (stepWork.doNeighborSearch && fr->efep != FreeEnergyPerturbationType::No)
1674     {
1675         wallcycle_sub_start(wcycle, WallCycleSubCounter::NonbondedFep);
1676         nbv->setupFepThreadedForceBuffer(fr->natoms_force_constr);
1677         wallcycle_sub_stop(wcycle, WallCycleSubCounter::NonbondedFep);
1678     }
1679
1680     if (simulationWork.useGpuNonbonded && stepWork.computeNonbondedForces)
1681     {
1682         /* launch D2H copy-back F */
1683         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1684         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1685
1686         if (simulationWork.havePpDomainDecomposition)
1687         {
1688             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1689         }
1690         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1691         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1692
1693         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1694         {
1695             fr->listedForcesGpu->launchEnergyTransfer();
1696         }
1697         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1698     }
1699
1700     gmx::ArrayRef<const gmx::RVec> xWholeMolecules;
1701     if (fr->wholeMoleculeTransform)
1702     {
1703         xWholeMolecules = fr->wholeMoleculeTransform->wholeMoleculeCoordinates(x.unpaddedArrayRef(), box);
1704     }
1705
1706     // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1707     // this wait ensures that the D2H transfer is complete.
1708     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
1709     {
1710         const bool needCoordsOnHost  = (runScheduleWork->domainWork.haveCpuLocalForceWork
1711                                        || stepWork.computeVirial || simulationWork.computeMuTot);
1712         const bool haveAlreadyWaited = simulationWork.useCpuHaloExchange;
1713         if (needCoordsOnHost && !haveAlreadyWaited)
1714         {
1715             GMX_ASSERT(haveCopiedXFromGpu,
1716                        "a wait should only be triggered if copy has been scheduled");
1717             stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1718         }
1719     }
1720
1721     DipoleData dipoleData;
1722
1723     if (simulationWork.computeMuTot)
1724     {
1725         const int start = 0;
1726
1727         /* Calculate total (local) dipole moment in a temporary common array.
1728          * This makes it possible to sum them over nodes faster.
1729          */
1730         gmx::ArrayRef<const gmx::RVec> xRef =
1731                 (xWholeMolecules.empty() ? x.unpaddedArrayRef() : xWholeMolecules);
1732         calc_mu(start,
1733                 mdatoms->homenr,
1734                 xRef,
1735                 mdatoms->chargeA ? gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->nr)
1736                                  : gmx::ArrayRef<real>{},
1737                 mdatoms->chargeB ? gmx::arrayRefFromArray(mdatoms->chargeB, mdatoms->nr)
1738                                  : gmx::ArrayRef<real>{},
1739                 mdatoms->nChargePerturbed != 0,
1740                 dipoleData.muStaging[0],
1741                 dipoleData.muStaging[1]);
1742
1743         reduceAndUpdateMuTot(
1744                 &dipoleData, cr, (fr->efep != FreeEnergyPerturbationType::No), lambda, muTotal, ddBalanceRegionHandler);
1745     }
1746
1747     /* Reset energies */
1748     reset_enerdata(enerd);
1749
1750     if (haveDDAtomOrdering(*cr) && simulationWork.haveSeparatePmeRank)
1751     {
1752         wallcycle_start(wcycle, WallCycleCounter::PpDuringPme);
1753         dd_force_flop_start(cr->dd, nrnb);
1754     }
1755
1756     if (inputrec.bRot)
1757     {
1758         wallcycle_start(wcycle, WallCycleCounter::Rot);
1759         do_rotation(cr, enforcedRotation, box, x.unpaddedConstArrayRef(), t, step, stepWork.doNeighborSearch);
1760         wallcycle_stop(wcycle, WallCycleCounter::Rot);
1761     }
1762
1763     /* Start the force cycle counter.
1764      * Note that a different counter is used for dynamic load balancing.
1765      */
1766     wallcycle_start(wcycle, WallCycleCounter::Force);
1767
1768     /* Set up and clear force outputs:
1769      * forceOutMtsLevel0:  everything except what is in the other two outputs
1770      * forceOutMtsLevel1:  PME-mesh and listed-forces group 1
1771      * forceOutNonbonded: non-bonded forces
1772      * Without multiple time stepping all point to the same object.
1773      * With multiple time-stepping the use is different for MTS fast (level0 only) and slow steps.
1774      */
1775     ForceOutputs forceOutMtsLevel0 = setupForceOutputs(
1776             &fr->forceHelperBuffers[0], force, domainWork, stepWork, simulationWork.havePpDomainDecomposition, wcycle);
1777
1778     // Force output for MTS combined forces, only set at level1 MTS steps
1779     std::optional<ForceOutputs> forceOutMts =
1780             (simulationWork.useMts && stepWork.computeSlowForces)
1781                     ? std::optional(setupForceOutputs(&fr->forceHelperBuffers[1],
1782                                                       forceView->forceMtsCombinedWithPadding(),
1783                                                       domainWork,
1784                                                       stepWork,
1785                                                       simulationWork.havePpDomainDecomposition,
1786                                                       wcycle))
1787                     : std::nullopt;
1788
1789     ForceOutputs* forceOutMtsLevel1 =
1790             simulationWork.useMts ? (stepWork.computeSlowForces ? &forceOutMts.value() : nullptr)
1791                                   : &forceOutMtsLevel0;
1792
1793     const bool nonbondedAtMtsLevel1 = runScheduleWork->simulationWork.computeNonbondedAtMtsLevel1;
1794
1795     ForceOutputs* forceOutNonbonded = nonbondedAtMtsLevel1 ? forceOutMtsLevel1 : &forceOutMtsLevel0;
1796
1797     if (inputrec.bPull && pull_have_constraint(*pull_work))
1798     {
1799         clear_pull_forces(pull_work);
1800     }
1801
1802     /* We calculate the non-bonded forces, when done on the CPU, here.
1803      * We do this before calling do_force_lowlevel, because in that
1804      * function, the listed forces are calculated before PME, which
1805      * does communication.  With this order, non-bonded and listed
1806      * force calculation imbalance can be balanced out by the domain
1807      * decomposition load balancing.
1808      */
1809
1810     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1811
1812     if (!useOrEmulateGpuNb)
1813     {
1814         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1815     }
1816
1817     if (fr->efep != FreeEnergyPerturbationType::No && stepWork.computeNonbondedForces)
1818     {
1819         /* Calculate the local and non-local free energy interactions here.
1820          * Happens here on the CPU both with and without GPU.
1821          */
1822         nbv->dispatchFreeEnergyKernels(
1823                 x,
1824                 &forceOutNonbonded->forceWithShiftForces(),
1825                 fr->use_simd_kernels,
1826                 fr->ntype,
1827                 fr->rlist,
1828                 *fr->ic,
1829                 fr->shift_vec,
1830                 fr->nbfp,
1831                 fr->ljpme_c6grid,
1832                 mdatoms->chargeA ? gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->nr)
1833                                  : gmx::ArrayRef<real>{},
1834                 mdatoms->chargeB ? gmx::arrayRefFromArray(mdatoms->chargeB, mdatoms->nr)
1835                                  : gmx::ArrayRef<real>{},
1836                 mdatoms->typeA ? gmx::arrayRefFromArray(mdatoms->typeA, mdatoms->nr)
1837                                : gmx::ArrayRef<int>{},
1838                 mdatoms->typeB ? gmx::arrayRefFromArray(mdatoms->typeB, mdatoms->nr)
1839                                : gmx::ArrayRef<int>{},
1840                 inputrec.fepvals.get(),
1841                 lambda,
1842                 enerd,
1843                 stepWork,
1844                 nrnb);
1845     }
1846
1847     if (stepWork.computeNonbondedForces && !useOrEmulateGpuNb)
1848     {
1849         if (simulationWork.havePpDomainDecomposition)
1850         {
1851             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step, nrnb, wcycle);
1852         }
1853
1854         if (stepWork.computeForces)
1855         {
1856             /* Add all the non-bonded force to the normal force array.
1857              * This can be split into a local and a non-local part when overlapping
1858              * communication with calculation with domain decomposition.
1859              */
1860             wallcycle_stop(wcycle, WallCycleCounter::Force);
1861             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All,
1862                                           forceOutNonbonded->forceWithShiftForces().force());
1863             wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
1864         }
1865
1866         /* If there are multiple fshift output buffers we need to reduce them */
1867         if (stepWork.computeVirial)
1868         {
1869             /* This is not in a subcounter because it takes a
1870                negligible and constant-sized amount of time */
1871             nbnxn_atomdata_add_nbat_fshift_to_fshift(
1872                     *nbv->nbat, forceOutNonbonded->forceWithShiftForces().shiftForces());
1873         }
1874     }
1875
1876     // TODO Force flags should include haveFreeEnergyWork for this domain
1877     if (stepWork.useGpuXHalo && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1878     {
1879         wallcycle_stop(wcycle, WallCycleCounter::Force);
1880         /* Wait for non-local coordinate data to be copied from device */
1881         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1882         wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
1883     }
1884
1885     // Compute wall interactions, when present.
1886     // Note: should be moved to special forces.
1887     if (inputrec.nwall && stepWork.computeNonbondedForces)
1888     {
1889         /* foreign lambda component for walls */
1890         real dvdl_walls = do_walls(inputrec,
1891                                    *fr,
1892                                    box,
1893                                    mdatoms->typeA ? gmx::arrayRefFromArray(mdatoms->typeA, mdatoms->nr)
1894                                                   : gmx::ArrayRef<int>{},
1895                                    mdatoms->typeB ? gmx::arrayRefFromArray(mdatoms->typeB, mdatoms->nr)
1896                                                   : gmx::ArrayRef<int>{},
1897                                    mdatoms->cENER ? gmx::arrayRefFromArray(mdatoms->cENER, mdatoms->nr)
1898                                                   : gmx::ArrayRef<unsigned short>{},
1899                                    mdatoms->homenr,
1900                                    mdatoms->nPerturbed,
1901                                    x.unpaddedConstArrayRef(),
1902                                    &forceOutMtsLevel0.forceWithVirial(),
1903                                    lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)],
1904                                    enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR],
1905                                    nrnb);
1906         enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] += dvdl_walls;
1907     }
1908
1909     if (stepWork.computeListedForces)
1910     {
1911         /* Check whether we need to take into account PBC in listed interactions */
1912         bool needMolPbc = false;
1913         for (const auto& listedForces : fr->listedForces)
1914         {
1915             if (listedForces.haveCpuListedForces(*fr->fcdata))
1916             {
1917                 needMolPbc = fr->bMolPBC;
1918             }
1919         }
1920
1921         t_pbc pbc;
1922
1923         if (needMolPbc)
1924         {
1925             /* Since all atoms are in the rectangular or triclinic unit-cell,
1926              * only single box vector shifts (2 in x) are required.
1927              */
1928             set_pbc_dd(&pbc, fr->pbcType, haveDDAtomOrdering(*cr) ? cr->dd->numCells : nullptr, TRUE, box);
1929         }
1930
1931         for (int mtsIndex = 0; mtsIndex < (simulationWork.useMts && stepWork.computeSlowForces ? 2 : 1);
1932              mtsIndex++)
1933         {
1934             ListedForces& listedForces = fr->listedForces[mtsIndex];
1935             ForceOutputs& forceOut     = (mtsIndex == 0 ? forceOutMtsLevel0 : *forceOutMtsLevel1);
1936             listedForces.calculate(wcycle,
1937                                    box,
1938                                    inputrec.fepvals.get(),
1939                                    cr,
1940                                    ms,
1941                                    x,
1942                                    xWholeMolecules,
1943                                    fr->fcdata.get(),
1944                                    hist,
1945                                    &forceOut,
1946                                    fr,
1947                                    &pbc,
1948                                    enerd,
1949                                    nrnb,
1950                                    lambda,
1951                                    mdatoms,
1952                                    haveDDAtomOrdering(*cr) ? cr->dd->globalAtomIndices.data() : nullptr,
1953                                    stepWork);
1954         }
1955     }
1956
1957     if (stepWork.computeSlowForces)
1958     {
1959         longRangeNonbondeds->calculate(fr->pmedata,
1960                                        cr,
1961                                        x.unpaddedConstArrayRef(),
1962                                        &forceOutMtsLevel1->forceWithVirial(),
1963                                        enerd,
1964                                        box,
1965                                        lambda,
1966                                        dipoleData.muStateAB,
1967                                        stepWork,
1968                                        ddBalanceRegionHandler);
1969     }
1970
1971     wallcycle_stop(wcycle, WallCycleCounter::Force);
1972
1973     // VdW dispersion correction, only computed on master rank to avoid double counting
1974     if ((stepWork.computeEnergy || stepWork.computeVirial) && fr->dispersionCorrection && MASTER(cr))
1975     {
1976         // Calculate long range corrections to pressure and energy
1977         const DispersionCorrection::Correction correction = fr->dispersionCorrection->calculate(
1978                 box, lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)]);
1979
1980         if (stepWork.computeEnergy)
1981         {
1982             enerd->term[F_DISPCORR] = correction.energy;
1983             enerd->term[F_DVDL_VDW] += correction.dvdl;
1984             enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] += correction.dvdl;
1985         }
1986         if (stepWork.computeVirial)
1987         {
1988             correction.correctVirial(vir_force);
1989             enerd->term[F_PDISPCORR] = correction.pressure;
1990         }
1991     }
1992
1993     computeSpecialForces(fplog,
1994                          cr,
1995                          inputrec,
1996                          awh,
1997                          enforcedRotation,
1998                          imdSession,
1999                          pull_work,
2000                          step,
2001                          t,
2002                          wcycle,
2003                          fr->forceProviders,
2004                          box,
2005                          x.unpaddedArrayRef(),
2006                          mdatoms,
2007                          lambda,
2008                          stepWork,
2009                          &forceOutMtsLevel0.forceWithVirial(),
2010                          forceOutMtsLevel1 ? &forceOutMtsLevel1->forceWithVirial() : nullptr,
2011                          enerd,
2012                          ed,
2013                          stepWork.doNeighborSearch);
2014
2015     if (simulationWork.havePpDomainDecomposition && stepWork.computeForces && stepWork.useGpuFHalo
2016         && domainWork.haveCpuLocalForceWork)
2017     {
2018         stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(), AtomLocality::Local);
2019     }
2020
2021     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
2022                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
2023     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFHalo),
2024                "The schedule below does not allow for nonbonded MTS with GPU halo exchange");
2025     // Will store the amount of cycles spent waiting for the GPU that
2026     // will be later used in the DLB accounting.
2027     float cycles_wait_gpu = 0;
2028     if (useOrEmulateGpuNb && stepWork.computeNonbondedForces)
2029     {
2030         auto& forceWithShiftForces = forceOutNonbonded->forceWithShiftForces();
2031
2032         /* wait for non-local forces (or calculate in emulation mode) */
2033         if (simulationWork.havePpDomainDecomposition)
2034         {
2035             if (simulationWork.useGpuNonbonded)
2036             {
2037                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
2038                         nbv->gpu_nbv,
2039                         stepWork,
2040                         AtomLocality::NonLocal,
2041                         enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR].data(),
2042                         enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR].data(),
2043                         forceWithShiftForces.shiftForces(),
2044                         wcycle);
2045             }
2046             else
2047             {
2048                 wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
2049                 do_nb_verlet(
2050                         fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes, step, nrnb, wcycle);
2051                 wallcycle_stop(wcycle, WallCycleCounter::Force);
2052             }
2053
2054             if (stepWork.useGpuFBufferOps)
2055             {
2056                 if (domainWork.haveNonLocalForceContribInCpuBuffer)
2057                 {
2058                     stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
2059                                               AtomLocality::NonLocal);
2060                 }
2061
2062
2063                 fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->execute();
2064
2065                 if (!stepWork.useGpuFHalo)
2066                 {
2067                     // copy from GPU input for dd_move_f()
2068                     stateGpu->copyForcesFromGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
2069                                                 AtomLocality::NonLocal);
2070                 }
2071             }
2072             else
2073             {
2074                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
2075             }
2076
2077             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
2078             {
2079                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
2080             }
2081         }
2082     }
2083
2084     /* Combining the forces for multiple time stepping before the halo exchange, when possible,
2085      * avoids an extra halo exchange (when DD is used) and post-processing step.
2086      */
2087     if (stepWork.combineMtsForcesBeforeHaloExchange)
2088     {
2089         combineMtsForces(getLocalAtomCount(cr->dd, *mdatoms, simulationWork.havePpDomainDecomposition),
2090                          force.unpaddedArrayRef(),
2091                          forceView->forceMtsCombined(),
2092                          inputrec.mtsLevels[1].stepFactor);
2093     }
2094
2095     if (simulationWork.havePpDomainDecomposition)
2096     {
2097         /* We are done with the CPU compute.
2098          * We will now communicate the non-local forces.
2099          * If we use a GPU this will overlap with GPU work, so in that case
2100          * we do not close the DD force balancing region here.
2101          */
2102         ddBalanceRegionHandler.closeAfterForceComputationCpu();
2103
2104         if (stepWork.computeForces)
2105         {
2106
2107             if (stepWork.useGpuFHalo)
2108             {
2109                 // If there exist CPU forces, data from halo exchange should accumulate into these
2110                 bool accumulateForces = domainWork.haveCpuLocalForceWork;
2111                 gmx::FixedCapacityVector<GpuEventSynchronizer*, 2> gpuForceHaloDependencies;
2112                 gpuForceHaloDependencies.push_back(stateGpu->fReadyOnDevice(AtomLocality::Local));
2113                 gpuForceHaloDependencies.push_back(stateGpu->fReducedOnDevice(AtomLocality::NonLocal));
2114
2115                 communicateGpuHaloForces(*cr, accumulateForces, &gpuForceHaloDependencies);
2116             }
2117             else
2118             {
2119                 if (stepWork.useGpuFBufferOps)
2120                 {
2121                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
2122                 }
2123
2124                 // Without MTS or with MTS at slow steps with uncombined forces we need to
2125                 // communicate the fast forces
2126                 if (!simulationWork.useMts || !stepWork.combineMtsForcesBeforeHaloExchange)
2127                 {
2128                     dd_move_f(cr->dd, &forceOutMtsLevel0.forceWithShiftForces(), wcycle);
2129                 }
2130                 // With MTS we need to communicate the slow or combined (in forceOutMtsLevel1) forces
2131                 if (simulationWork.useMts && stepWork.computeSlowForces)
2132                 {
2133                     dd_move_f(cr->dd, &forceOutMtsLevel1->forceWithShiftForces(), wcycle);
2134                 }
2135             }
2136         }
2137     }
2138
2139     // With both nonbonded and PME offloaded a GPU on the same rank, we use
2140     // an alternating wait/reduction scheme.
2141     bool alternateGpuWait =
2142             (!c_disableAlternatingWait && stepWork.haveGpuPmeOnThisRank && simulationWork.useGpuNonbonded
2143              && !simulationWork.havePpDomainDecomposition && !stepWork.useGpuFBufferOps);
2144
2145     if (alternateGpuWait)
2146     {
2147         alternatePmeNbGpuWaitReduce(fr->nbv.get(),
2148                                     fr->pmedata,
2149                                     forceOutNonbonded,
2150                                     forceOutMtsLevel1,
2151                                     enerd,
2152                                     lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
2153                                     stepWork,
2154                                     wcycle);
2155     }
2156
2157     if (!alternateGpuWait && stepWork.haveGpuPmeOnThisRank)
2158     {
2159         pme_gpu_wait_and_reduce(fr->pmedata,
2160                                 stepWork,
2161                                 wcycle,
2162                                 &forceOutMtsLevel1->forceWithVirial(),
2163                                 enerd,
2164                                 lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)]);
2165     }
2166
2167     /* Wait for local GPU NB outputs on the non-alternating wait path */
2168     if (!alternateGpuWait && stepWork.computeNonbondedForces && simulationWork.useGpuNonbonded)
2169     {
2170         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
2171          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
2172          * but even with a step of 0.1 ms the difference is less than 1%
2173          * of the step time.
2174          */
2175         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
2176         const float waitCycles               = Nbnxm::gpu_wait_finish_task(
2177                 nbv->gpu_nbv,
2178                 stepWork,
2179                 AtomLocality::Local,
2180                 enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR].data(),
2181                 enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR].data(),
2182                 forceOutNonbonded->forceWithShiftForces().shiftForces(),
2183                 wcycle);
2184
2185         if (ddBalanceRegionHandler.useBalancingRegion())
2186         {
2187             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
2188             if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
2189             {
2190                 /* We measured few cycles, it could be that the kernel
2191                  * and transfer finished earlier and there was no actual
2192                  * wait time, only API call overhead.
2193                  * Then the actual time could be anywhere between 0 and
2194                  * cycles_wait_est. We will use half of cycles_wait_est.
2195                  */
2196                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
2197             }
2198             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
2199         }
2200     }
2201
2202     if (fr->nbv->emulateGpu())
2203     {
2204         // NOTE: emulation kernel is not included in the balancing region,
2205         // but emulation mode does not target performance anyway
2206         wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
2207         do_nb_verlet(fr,
2208                      ic,
2209                      enerd,
2210                      stepWork,
2211                      InteractionLocality::Local,
2212                      haveDDAtomOrdering(*cr) ? enbvClearFNo : enbvClearFYes,
2213                      step,
2214                      nrnb,
2215                      wcycle);
2216         wallcycle_stop(wcycle, WallCycleCounter::Force);
2217     }
2218
2219     // If on GPU PME-PP comms path, receive forces from PME before GPU buffer ops
2220     // TODO refactor this and unify with below default-path call to the same function
2221     if (PAR(cr) && simulationWork.haveSeparatePmeRank && simulationWork.useGpuPmePpCommunication
2222         && stepWork.computeSlowForces)
2223     {
2224         /* In case of node-splitting, the PP nodes receive the long-range
2225          * forces, virial and energy from the PME nodes here.
2226          */
2227         pme_receive_force_ener(fr,
2228                                cr,
2229                                &forceOutMtsLevel1->forceWithVirial(),
2230                                enerd,
2231                                simulationWork.useGpuPmePpCommunication,
2232                                stepWork.useGpuPmeFReduction,
2233                                wcycle);
2234     }
2235
2236
2237     /* Do the nonbonded GPU (or emulation) force buffer reduction
2238      * on the non-alternating path. */
2239     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
2240                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
2241     if (useOrEmulateGpuNb && !alternateGpuWait)
2242     {
2243         if (stepWork.useGpuFBufferOps)
2244         {
2245             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2246
2247             // TODO: move these steps as early as possible:
2248             // - CPU f H2D should be as soon as all CPU-side forces are done
2249             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
2250             //   before the next CPU task that consumes the forces: vsite spread or update)
2251             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
2252             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
2253             //   These should be unified.
2254             if (domainWork.haveLocalForceContribInCpuBuffer && !stepWork.useGpuFHalo)
2255             {
2256                 stateGpu->copyForcesToGpu(forceWithShift, AtomLocality::Local);
2257             }
2258
2259             if (stepWork.computeNonbondedForces)
2260             {
2261                 fr->gpuForceReduction[gmx::AtomLocality::Local]->execute();
2262             }
2263
2264             // Copy forces to host if they are needed for update or if virtual sites are enabled.
2265             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
2266             // TODO: When the output flags will be included in step workload, this copy can be combined with the
2267             //       copy call done in sim_utils(...) for the output.
2268             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
2269             //       they should not be copied in do_md(...) for the output.
2270             if (!simulationWork.useGpuUpdate
2271                 || (simulationWork.useGpuUpdate && haveDDAtomOrdering(*cr) && simulationWork.useCpuPmePpCommunication)
2272                 || vsite)
2273             {
2274                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
2275                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
2276             }
2277         }
2278         else if (stepWork.computeNonbondedForces)
2279         {
2280             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2281             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
2282         }
2283     }
2284
2285     launchGpuEndOfStepTasks(
2286             nbv, fr->listedForcesGpu.get(), fr->pmedata, enerd, *runScheduleWork, step, wcycle);
2287
2288     if (haveDDAtomOrdering(*cr))
2289     {
2290         dd_force_flop_stop(cr->dd, nrnb);
2291     }
2292
2293     const bool haveCombinedMtsForces = (stepWork.computeForces && simulationWork.useMts && stepWork.computeSlowForces
2294                                         && stepWork.combineMtsForcesBeforeHaloExchange);
2295     if (stepWork.computeForces)
2296     {
2297         postProcessForceWithShiftForces(
2298                 nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutMtsLevel0, vir_force, *mdatoms, *fr, vsite, stepWork);
2299
2300         if (simulationWork.useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2301         {
2302             postProcessForceWithShiftForces(
2303                     nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1, vir_force, *mdatoms, *fr, vsite, stepWork);
2304         }
2305     }
2306
2307     // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
2308     if (PAR(cr) && simulationWork.haveSeparatePmeRank && simulationWork.useCpuPmePpCommunication
2309         && stepWork.computeSlowForces)
2310     {
2311         /* In case of node-splitting, the PP nodes receive the long-range
2312          * forces, virial and energy from the PME nodes here.
2313          */
2314         pme_receive_force_ener(fr,
2315                                cr,
2316                                &forceOutMtsLevel1->forceWithVirial(),
2317                                enerd,
2318                                simulationWork.useGpuPmePpCommunication,
2319                                false,
2320                                wcycle);
2321     }
2322
2323     if (stepWork.computeForces)
2324     {
2325         /* If we don't use MTS or if we already combined the MTS forces before, we only
2326          * need to post-process one ForceOutputs object here, called forceOutCombined,
2327          * otherwise we have to post-process two outputs and then combine them.
2328          */
2329         ForceOutputs& forceOutCombined = (haveCombinedMtsForces ? forceOutMts.value() : forceOutMtsLevel0);
2330         postProcessForces(
2331                 cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutCombined, vir_force, mdatoms, fr, vsite, stepWork);
2332
2333         if (simulationWork.useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2334         {
2335             postProcessForces(
2336                     cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1, vir_force, mdatoms, fr, vsite, stepWork);
2337
2338             combineMtsForces(mdatoms->homenr,
2339                              force.unpaddedArrayRef(),
2340                              forceView->forceMtsCombined(),
2341                              inputrec.mtsLevels[1].stepFactor);
2342         }
2343     }
2344
2345     if (stepWork.computeEnergy)
2346     {
2347         /* Compute the final potential energy terms */
2348         accumulatePotentialEnergies(enerd, lambda, inputrec.fepvals.get());
2349
2350         if (!EI_TPI(inputrec.eI))
2351         {
2352             checkPotentialEnergyValidity(step, *enerd, inputrec);
2353         }
2354     }
2355
2356     /* In case we don't have constraints and are using GPUs, the next balancing
2357      * region starts here.
2358      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
2359      * virial calculation and COM pulling, is not thus not included in
2360      * the balance timing, which is ok as most tasks do communication.
2361      */
2362     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
2363 }