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_pp.h"
  58 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  59 #include "gromacs/gmxlib/network.h"
  60 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  61 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  62 #include "gromacs/gmxlib/nrnb.h"
  63 #include "gromacs/gpu_utils/gpu_utils.h"
  64 #include "gromacs/imd/imd.h"
  65 #include "gromacs/listed_forces/disre.h"
  66 #include "gromacs/listed_forces/listed_forces_gpu.h"
  67 #include "gromacs/listed_forces/listed_forces.h"
  68 #include "gromacs/listed_forces/orires.h"
  69 #include "gromacs/math/arrayrefwithpadding.h"
  70 #include "gromacs/math/functions.h"
  71 #include "gromacs/math/units.h"
  72 #include "gromacs/math/vec.h"
  73 #include "gromacs/math/vecdump.h"
  74 #include "gromacs/mdlib/calcmu.h"
  75 #include "gromacs/mdlib/calcvir.h"
  76 #include "gromacs/mdlib/constr.h"
  77 #include "gromacs/mdlib/dispersioncorrection.h"
  78 #include "gromacs/mdlib/enerdata_utils.h"
  79 #include "gromacs/mdlib/force.h"
  80 #include "gromacs/mdlib/force_flags.h"
  81 #include "gromacs/mdlib/forcerec.h"
  82 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  83 #include "gromacs/mdlib/update.h"
  84 #include "gromacs/mdlib/vsite.h"
  85 #include "gromacs/mdlib/wall.h"
  86 #include "gromacs/mdlib/wholemoleculetransform.h"
  87 #include "gromacs/mdtypes/commrec.h"
  88 #include "gromacs/mdtypes/enerdata.h"
  89 #include "gromacs/mdtypes/forcebuffers.h"
  90 #include "gromacs/mdtypes/forceoutput.h"
  91 #include "gromacs/mdtypes/forcerec.h"
  92 #include "gromacs/mdtypes/iforceprovider.h"
  93 #include "gromacs/mdtypes/inputrec.h"
  94 #include "gromacs/mdtypes/md_enums.h"
  95 #include "gromacs/mdtypes/mdatom.h"
  96 #include "gromacs/mdtypes/multipletimestepping.h"
  97 #include "gromacs/mdtypes/simulation_workload.h"
  98 #include "gromacs/mdtypes/state.h"
  99 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
 100 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 101 #include "gromacs/nbnxm/nbnxm.h"
 102 #include "gromacs/nbnxm/nbnxm_gpu.h"
 103 #include "gromacs/pbcutil/ishift.h"
 104 #include "gromacs/pbcutil/pbc.h"
 105 #include "gromacs/pulling/pull.h"
 106 #include "gromacs/pulling/pull_rotation.h"
 107 #include "gromacs/timing/cyclecounter.h"
 108 #include "gromacs/timing/gpu_timing.h"
 109 #include "gromacs/timing/wallcycle.h"
 110 #include "gromacs/timing/wallcyclereporting.h"
 111 #include "gromacs/timing/walltime_accounting.h"
 112 #include "gromacs/topology/topology.h"
 113 #include "gromacs/utility/arrayref.h"
 114 #include "gromacs/utility/basedefinitions.h"
 115 #include "gromacs/utility/cstringutil.h"
 116 #include "gromacs/utility/exceptions.h"
 117 #include "gromacs/utility/fatalerror.h"
 118 #include "gromacs/utility/fixedcapacityvector.h"
 119 #include "gromacs/utility/gmxassert.h"
 120 #include "gromacs/utility/gmxmpi.h"
 121 #include "gromacs/utility/logger.h"
 122 #include "gromacs/utility/smalloc.h"
 123 #include "gromacs/utility/strconvert.h"
 124 #include "gromacs/utility/sysinfo.h"
 125
 126 #include "gpuforcereduction.h"
 127
 128 using gmx::ArrayRef;
 129 using gmx::AtomLocality;
 130 using gmx::DomainLifetimeWorkload;
 131 using gmx::ForceOutputs;
 132 using gmx::ForceWithShiftForces;
 133 using gmx::InteractionLocality;
 134 using gmx::RVec;
 135 using gmx::SimulationWorkload;
 136 using gmx::StepWorkload;
 137
 138 // TODO: this environment variable allows us to verify before release
 139 // that on less common architectures the total cost of polling is not larger than
 140 // a blocking wait (so polling does not introduce overhead when the static
 141 // PME-first ordering would suffice).
 142 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 143
 144 static void sum_forces(ArrayRef<RVec> f, ArrayRef<const RVec> forceToAdd)
 145 {
 146     GMX_ASSERT(f.size() >= forceToAdd.size(), "Accumulation buffer should be sufficiently large");
 147     const int end = forceToAdd.size();
 148
 149     int gmx_unused nt = gmx_omp_nthreads_get(ModuleMultiThread::Default);
 150 #pragma omp parallel for num_threads(nt) schedule(static)
 151     for (int i = 0; i < end; i++)
 152     {
 153         rvec_inc(f[i], forceToAdd[i]);
 154     }
 155 }
 156
 157 static void calc_virial(int                              start,
 158                         int                              homenr,
 159                         const rvec                       x[],
 160                         const gmx::ForceWithShiftForces& forceWithShiftForces,
 161                         tensor                           vir_part,
 162                         const matrix                     box,
 163                         t_nrnb*                          nrnb,
 164                         const t_forcerec*                fr,
 165                         PbcType                          pbcType)
 166 {
 167     /* The short-range virial from surrounding boxes */
 168     const rvec* fshift          = as_rvec_array(forceWithShiftForces.shiftForces().data());
 169     const rvec* shiftVecPointer = as_rvec_array(fr->shift_vec.data());
 170     calc_vir(gmx::c_numShiftVectors, shiftVecPointer, fshift, vir_part, pbcType == PbcType::Screw, box);
 171     inc_nrnb(nrnb, eNR_VIRIAL, gmx::c_numShiftVectors);
 172
 173     /* Calculate partial virial, for local atoms only, based on short range.
 174      * Total virial is computed in global_stat, called from do_md
 175      */
 176     const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
 177     f_calc_vir(start, start + homenr, x, f, vir_part, box);
 178     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 179
 180     if (debug)
 181     {
 182         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 183     }
 184 }
 185
 186 static void pull_potential_wrapper(const t_commrec*               cr,
 187                                    const t_inputrec&              ir,
 188                                    const matrix                   box,
 189                                    gmx::ArrayRef<const gmx::RVec> x,
 190                                    gmx::ForceWithVirial*          force,
 191                                    const t_mdatoms*               mdatoms,
 192                                    gmx_enerdata_t*                enerd,
 193                                    pull_t*                        pull_work,
 194                                    const real*                    lambda,
 195                                    double                         t,
 196                                    gmx_wallcycle*                 wcycle)
 197 {
 198     t_pbc pbc;
 199     real  dvdl;
 200
 201     /* Calculate the center of mass forces, this requires communication,
 202      * which is why pull_potential is called close to other communication.
 203      */
 204     wallcycle_start(wcycle, WallCycleCounter::PullPot);
 205     set_pbc(&pbc, ir.pbcType, box);
 206     dvdl = 0;
 207     enerd->term[F_COM_PULL] +=
 208             pull_potential(pull_work,
 209                            gmx::arrayRefFromArray(mdatoms->massT, mdatoms->nr),
 210                            pbc,
 211                            cr,
 212                            t,
 213                            lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Restraint)],
 214                            x,
 215                            force,
 216                            &dvdl);
 217     enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Restraint] += dvdl;
 218     wallcycle_stop(wcycle, WallCycleCounter::PullPot);
 219 }
 220
 221 static void pme_receive_force_ener(t_forcerec*           fr,
 222                                    const t_commrec*      cr,
 223                                    gmx::ForceWithVirial* forceWithVirial,
 224                                    gmx_enerdata_t*       enerd,
 225                                    bool                  useGpuPmePpComms,
 226                                    bool                  receivePmeForceToGpu,
 227                                    gmx_wallcycle*        wcycle)
 228 {
 229     real  e_q, e_lj, dvdl_q, dvdl_lj;
 230     float cycles_ppdpme, cycles_seppme;
 231
 232     cycles_ppdpme = wallcycle_stop(wcycle, WallCycleCounter::PpDuringPme);
 233     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 234
 235     /* In case of node-splitting, the PP nodes receive the long-range
 236      * forces, virial and energy from the PME nodes here.
 237      */
 238     wallcycle_start(wcycle, WallCycleCounter::PpPmeWaitRecvF);
 239     dvdl_q  = 0;
 240     dvdl_lj = 0;
 241     gmx_pme_receive_f(fr->pmePpCommGpu.get(),
 242                       cr,
 243                       forceWithVirial,
 244                       &e_q,
 245                       &e_lj,
 246                       &dvdl_q,
 247                       &dvdl_lj,
 248                       useGpuPmePpComms,
 249                       receivePmeForceToGpu,
 250                       &cycles_seppme);
 251     enerd->term[F_COUL_RECIP] += e_q;
 252     enerd->term[F_LJ_RECIP] += e_lj;
 253     enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Coul] += dvdl_q;
 254     enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] += dvdl_lj;
 255
 256     if (wcycle)
 257     {
 258         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 259     }
 260     wallcycle_stop(wcycle, WallCycleCounter::PpPmeWaitRecvF);
 261 }
 262
 263 static void print_large_forces(FILE*                fp,
 264                                const t_mdatoms*     md,
 265                                const t_commrec*     cr,
 266                                int64_t              step,
 267                                real                 forceTolerance,
 268                                ArrayRef<const RVec> x,
 269                                ArrayRef<const RVec> f)
 270 {
 271     real       force2Tolerance = gmx::square(forceTolerance);
 272     gmx::index numNonFinite    = 0;
 273     for (int i = 0; i < md->homenr; i++)
 274     {
 275         real force2    = norm2(f[i]);
 276         bool nonFinite = !std::isfinite(force2);
 277         if (force2 >= force2Tolerance || nonFinite)
 278         {
 279             fprintf(fp,
 280                     "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 281                     step,
 282                     ddglatnr(cr->dd, i),
 283                     x[i][XX],
 284                     x[i][YY],
 285                     x[i][ZZ],
 286                     std::sqrt(force2));
 287         }
 288         if (nonFinite)
 289         {
 290             numNonFinite++;
 291         }
 292     }
 293     if (numNonFinite > 0)
 294     {
 295         /* Note that with MPI this fatal call on one rank might interrupt
 296          * the printing on other ranks. But we can only avoid that with
 297          * an expensive MPI barrier that we would need at each step.
 298          */
 299         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 300     }
 301 }
 302
 303 //! When necessary, spreads forces on vsites and computes the virial for \p forceOutputs->forceWithShiftForces()
 304 static void postProcessForceWithShiftForces(t_nrnb*                   nrnb,
 305                                             gmx_wallcycle*            wcycle,
 306                                             const matrix              box,
 307                                             ArrayRef<const RVec>      x,
 308                                             ForceOutputs*             forceOutputs,
 309                                             tensor                    vir_force,
 310                                             const t_mdatoms&          mdatoms,
 311                                             const t_forcerec&         fr,
 312                                             gmx::VirtualSitesHandler* vsite,
 313                                             const StepWorkload&       stepWork)
 314 {
 315     ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
 316
 317     /* If we have NoVirSum forces, but we do not calculate the virial,
 318      * we later sum the forceWithShiftForces buffer together with
 319      * the noVirSum buffer and spread the combined vsite forces at once.
 320      */
 321     if (vsite && (!forceOutputs->haveForceWithVirial() || stepWork.computeVirial))
 322     {
 323         using VirialHandling = gmx::VirtualSitesHandler::VirialHandling;
 324
 325         auto                 f      = forceWithShiftForces.force();
 326         auto                 fshift = forceWithShiftForces.shiftForces();
 327         const VirialHandling virialHandling =
 328                 (stepWork.computeVirial ? VirialHandling::Pbc : VirialHandling::None);
 329         vsite->spreadForces(x, f, virialHandling, fshift, nullptr, nrnb, box, wcycle);
 330         forceWithShiftForces.haveSpreadVsiteForces() = true;
 331     }
 332
 333     if (stepWork.computeVirial)
 334     {
 335         /* Calculation of the virial must be done after vsites! */
 336         calc_virial(
 337                 0, mdatoms.homenr, as_rvec_array(x.data()), forceWithShiftForces, vir_force, box, nrnb, &fr, fr.pbcType);
 338     }
 339 }
 340
 341 //! Spread, compute virial for and sum forces, when necessary
 342 static void postProcessForces(const t_commrec*          cr,
 343                               int64_t                   step,
 344                               t_nrnb*                   nrnb,
 345                               gmx_wallcycle*            wcycle,
 346                               const matrix              box,
 347                               ArrayRef<const RVec>      x,
 348                               ForceOutputs*             forceOutputs,
 349                               tensor                    vir_force,
 350                               const t_mdatoms*          mdatoms,
 351                               const t_forcerec*         fr,
 352                               gmx::VirtualSitesHandler* vsite,
 353                               const StepWorkload&       stepWork)
 354 {
 355     // Extract the final output force buffer, which is also the buffer for forces with shift forces
 356     ArrayRef<RVec> f = forceOutputs->forceWithShiftForces().force();
 357
 358     if (forceOutputs->haveForceWithVirial())
 359     {
 360         auto& forceWithVirial = forceOutputs->forceWithVirial();
 361
 362         if (vsite)
 363         {
 364             /* Spread the mesh force on virtual sites to the other particles...
 365              * This is parallellized. MPI communication is performed
 366              * if the constructing atoms aren't local.
 367              */
 368             GMX_ASSERT(!stepWork.computeVirial || f.data() != forceWithVirial.force_.data(),
 369                        "We need separate force buffers for shift and virial forces when "
 370                        "computing the virial");
 371             GMX_ASSERT(!stepWork.computeVirial
 372                                || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
 373                        "We should spread the force with shift forces separately when computing "
 374                        "the virial");
 375             const gmx::VirtualSitesHandler::VirialHandling virialHandling =
 376                     (stepWork.computeVirial ? gmx::VirtualSitesHandler::VirialHandling::NonLinear
 377                                             : gmx::VirtualSitesHandler::VirialHandling::None);
 378             matrix virial = { { 0 } };
 379             vsite->spreadForces(x, forceWithVirial.force_, virialHandling, {}, virial, nrnb, box, wcycle);
 380             forceWithVirial.addVirialContribution(virial);
 381         }
 382
 383         if (stepWork.computeVirial)
 384         {
 385             /* Now add the forces, this is local */
 386             sum_forces(f, forceWithVirial.force_);
 387
 388             /* Add the direct virial contributions */
 389             GMX_ASSERT(
 390                     forceWithVirial.computeVirial_,
 391                     "forceWithVirial should request virial computation when we request the virial");
 392             m_add(vir_force, forceWithVirial.getVirial(), vir_force);
 393
 394             if (debug)
 395             {
 396                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 397             }
 398         }
 399     }
 400     else
 401     {
 402         GMX_ASSERT(vsite == nullptr || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
 403                    "We should have spread the vsite forces (earlier)");
 404     }
 405
 406     if (fr->print_force >= 0)
 407     {
 408         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 409     }
 410 }
 411
 412 static void do_nb_verlet(t_forcerec*                fr,
 413                          const interaction_const_t* ic,
 414                          gmx_enerdata_t*            enerd,
 415                          const StepWorkload&        stepWork,
 416                          const InteractionLocality  ilocality,
 417                          const int                  clearF,
 418                          const int64_t              step,
 419                          t_nrnb*                    nrnb,
 420                          gmx_wallcycle*             wcycle)
 421 {
 422     if (!stepWork.computeNonbondedForces)
 423     {
 424         /* skip non-bonded calculation */
 425         return;
 426     }
 427
 428     nonbonded_verlet_t* nbv = fr->nbv.get();
 429
 430     /* GPU kernel launch overhead is already timed separately */
 431     if (!nbv->useGpu())
 432     {
 433         /* When dynamic pair-list  pruning is requested, we need to prune
 434          * at nstlistPrune steps.
 435          */
 436         if (nbv->isDynamicPruningStepCpu(step))
 437         {
 438             /* Prune the pair-list beyond fr->ic->rlistPrune using
 439              * the current coordinates of the atoms.
 440              */
 441             wallcycle_sub_start(wcycle, WallCycleSubCounter::NonbondedPruning);
 442             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 443             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NonbondedPruning);
 444         }
 445     }
 446
 447     nbv->dispatchNonbondedKernel(
 448             ilocality,
 449             *ic,
 450             stepWork,
 451             clearF,
 452             fr->shift_vec,
 453             enerd->grpp.energyGroupPairTerms[fr->haveBuckingham ? NonBondedEnergyTerms::BuckinghamSR
 454                                                                 : NonBondedEnergyTerms::LJSR],
 455             enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR],
 456             nrnb);
 457 }
 458
 459 static inline void clearRVecs(ArrayRef<RVec> v, const bool useOpenmpThreading)
 460 {
 461     int nth = gmx_omp_nthreads_get_simple_rvec_task(ModuleMultiThread::Default, v.ssize());
 462
 463     /* Note that we would like to avoid this conditional by putting it
 464      * into the omp pragma instead, but then we still take the full
 465      * omp parallel for overhead (at least with gcc5).
 466      */
 467     if (!useOpenmpThreading || nth == 1)
 468     {
 469         for (RVec& elem : v)
 470         {
 471             clear_rvec(elem);
 472         }
 473     }
 474     else
 475     {
 476 #pragma omp parallel for num_threads(nth) schedule(static)
 477         for (gmx::index i = 0; i < v.ssize(); i++)
 478         {
 479             clear_rvec(v[i]);
 480         }
 481     }
 482 }
 483
 484 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 485  *
 486  * \param groupOptions  Group options, containing T-coupling options
 487  */
 488 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
 489 {
 490     real nrdfCoupled   = 0;
 491     real nrdfUncoupled = 0;
 492     real kineticEnergy = 0;
 493     for (int g = 0; g < groupOptions.ngtc; g++)
 494     {
 495         if (groupOptions.tau_t[g] >= 0)
 496         {
 497             nrdfCoupled += groupOptions.nrdf[g];
 498             kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * gmx::c_boltz;
 499         }
 500         else
 501         {
 502             nrdfUncoupled += groupOptions.nrdf[g];
 503         }
 504     }
 505
 506     /* This conditional with > also catches nrdf=0 */
 507     if (nrdfCoupled > nrdfUncoupled)
 508     {
 509         return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
 510     }
 511     else
 512     {
 513         return 0;
 514     }
 515 }
 516
 517 /*! \brief This routine checks that the potential energy is finite.
 518  *
 519  * Always checks that the potential energy is finite. If step equals
 520  * inputrec.init_step also checks that the magnitude of the potential energy
 521  * is reasonable. Terminates with a fatal error when a check fails.
 522  * Note that passing this check does not guarantee finite forces,
 523  * since those use slightly different arithmetics. But in most cases
 524  * there is just a narrow coordinate range where forces are not finite
 525  * and energies are finite.
 526  *
 527  * \param[in] step      The step number, used for checking and printing
 528  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to
 529  * enerd.term[F_EPOT] before calling this routine \param[in] inputrec  The input record
 530  */
 531 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
 532 {
 533     /* Threshold valid for comparing absolute potential energy against
 534      * the kinetic energy. Normally one should not consider absolute
 535      * potential energy values, but with a factor of one million
 536      * we should never get false positives.
 537      */
 538     constexpr real c_thresholdFactor = 1e6;
 539
 540     bool energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 541     real averageKineticEnergy = 0;
 542     /* We only check for large potential energy at the initial step,
 543      * because that is by far the most likely step for this too occur
 544      * and because computing the average kinetic energy is not free.
 545      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 546      * before they become NaN.
 547      */
 548     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 549     {
 550         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 551     }
 552
 553     if (energyIsNotFinite
 554         || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
 555     {
 556         gmx_fatal(
 557                 FARGS,
 558                 "Step %" PRId64
 559                 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
 560                 "contributions to the energy are %g and %g, respectively. A %s potential energy "
 561                 "can be caused by overlapping interactions in bonded interactions or very large%s "
 562                 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
 563                 "configuration, incorrect interactions or parameters in the topology.",
 564                 step,
 565                 enerd.term[F_EPOT],
 566                 energyIsNotFinite ? "not finite" : "extremely high",
 567                 enerd.term[F_LJ],
 568                 enerd.term[F_COUL_SR],
 569                 energyIsNotFinite ? "non-finite" : "very high",
 570                 energyIsNotFinite ? " or Nan" : "");
 571     }
 572 }
 573
 574 /*! \brief Return true if there are special forces computed this step.
 575  *
 576  * The conditionals exactly correspond to those in computeSpecialForces().
 577  */
 578 static bool haveSpecialForces(const t_inputrec&          inputrec,
 579                               const gmx::ForceProviders& forceProviders,
 580                               const pull_t*              pull_work,
 581                               const bool                 computeForces,
 582                               const gmx_edsam*           ed)
 583 {
 584
 585     return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
 586             (inputrec.bPull && pull_have_potential(*pull_work)) ||  // pull
 587             inputrec.bRot ||                                        // enforced rotation
 588             (ed != nullptr) ||                                      // flooding
 589             (inputrec.bIMD && computeForces));                      // IMD
 590 }
 591
 592 /*! \brief Compute forces and/or energies for special algorithms
 593  *
 594  * The intention is to collect all calls to algorithms that compute
 595  * forces on local atoms only and that do not contribute to the local
 596  * virial sum (but add their virial contribution separately).
 597  * Eventually these should likely all become ForceProviders.
 598  * Within this function the intention is to have algorithms that do
 599  * global communication at the end, so global barriers within the MD loop
 600  * are as close together as possible.
 601  *
 602  * \param[in]     fplog            The log file
 603  * \param[in]     cr               The communication record
 604  * \param[in]     inputrec         The input record
 605  * \param[in]     awh              The Awh module (nullptr if none in use).
 606  * \param[in]     enforcedRotation Enforced rotation module.
 607  * \param[in]     imdSession       The IMD session
 608  * \param[in]     pull_work        The pull work structure.
 609  * \param[in]     step             The current MD step
 610  * \param[in]     t                The current time
 611  * \param[in,out] wcycle           Wallcycle accounting struct
 612  * \param[in,out] forceProviders   Pointer to a list of force providers
 613  * \param[in]     box              The unit cell
 614  * \param[in]     x                The coordinates
 615  * \param[in]     mdatoms          Per atom properties
 616  * \param[in]     lambda           Array of free-energy lambda values
 617  * \param[in]     stepWork         Step schedule flags
 618  * \param[in,out] forceWithVirialMtsLevel0  Force and virial for MTS level0 forces
 619  * \param[in,out] forceWithVirialMtsLevel1  Force and virial for MTS level1 forces, can be nullptr
 620  * \param[in,out] enerd            Energy buffer
 621  * \param[in,out] ed               Essential dynamics pointer
 622  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 623  *
 624  * \todo Remove didNeighborSearch, which is used incorrectly.
 625  * \todo Convert all other algorithms called here to ForceProviders.
 626  */
 627 static void computeSpecialForces(FILE*                          fplog,
 628                                  const t_commrec*               cr,
 629                                  const t_inputrec&              inputrec,
 630                                  gmx::Awh*                      awh,
 631                                  gmx_enfrot*                    enforcedRotation,
 632                                  gmx::ImdSession*               imdSession,
 633                                  pull_t*                        pull_work,
 634                                  int64_t                        step,
 635                                  double                         t,
 636                                  gmx_wallcycle*                 wcycle,
 637                                  gmx::ForceProviders*           forceProviders,
 638                                  const matrix                   box,
 639                                  gmx::ArrayRef<const gmx::RVec> x,
 640                                  const t_mdatoms*               mdatoms,
 641                                  gmx::ArrayRef<const real>      lambda,
 642                                  const StepWorkload&            stepWork,
 643                                  gmx::ForceWithVirial*          forceWithVirialMtsLevel0,
 644                                  gmx::ForceWithVirial*          forceWithVirialMtsLevel1,
 645                                  gmx_enerdata_t*                enerd,
 646                                  gmx_edsam*                     ed,
 647                                  bool                           didNeighborSearch)
 648 {
 649     /* NOTE: Currently all ForceProviders only provide forces.
 650      *       When they also provide energies, remove this conditional.
 651      */
 652     if (stepWork.computeForces)
 653     {
 654         gmx::ForceProviderInput forceProviderInput(
 655                 x,
 656                 mdatoms->homenr,
 657                 gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->homenr),
 658                 gmx::arrayRefFromArray(mdatoms->massT, mdatoms->homenr),
 659                 t,
 660                 box,
 661                 *cr);
 662         gmx::ForceProviderOutput forceProviderOutput(forceWithVirialMtsLevel0, enerd);
 663
 664         /* Collect forces from modules */
 665         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 666     }
 667
 668     if (inputrec.bPull && pull_have_potential(*pull_work))
 669     {
 670         const int mtsLevel = forceGroupMtsLevel(inputrec.mtsLevels, gmx::MtsForceGroups::Pull);
 671         if (mtsLevel == 0 || stepWork.computeSlowForces)
 672         {
 673             auto& forceWithVirial = (mtsLevel == 0) ? forceWithVirialMtsLevel0 : forceWithVirialMtsLevel1;
 674             pull_potential_wrapper(
 675                     cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work, lambda.data(), t, wcycle);
 676         }
 677     }
 678     if (awh)
 679     {
 680         const int mtsLevel = forceGroupMtsLevel(inputrec.mtsLevels, gmx::MtsForceGroups::Pull);
 681         if (mtsLevel == 0 || stepWork.computeSlowForces)
 682         {
 683             const bool needForeignEnergyDifferences = awh->needForeignEnergyDifferences(step);
 684             std::vector<double> foreignLambdaDeltaH, foreignLambdaDhDl;
 685             if (needForeignEnergyDifferences)
 686             {
 687                 enerd->foreignLambdaTerms.finalizePotentialContributions(
 688                         enerd->dvdl_lin, lambda, *inputrec.fepvals);
 689                 std::tie(foreignLambdaDeltaH, foreignLambdaDhDl) = enerd->foreignLambdaTerms.getTerms(cr);
 690             }
 691
 692             auto& forceWithVirial = (mtsLevel == 0) ? forceWithVirialMtsLevel0 : forceWithVirialMtsLevel1;
 693             enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
 694                     inputrec.pbcType,
 695                     gmx::arrayRefFromArray(mdatoms->massT, mdatoms->nr),
 696                     foreignLambdaDeltaH,
 697                     foreignLambdaDhDl,
 698                     box,
 699                     forceWithVirial,
 700                     t,
 701                     step,
 702                     wcycle,
 703                     fplog);
 704         }
 705     }
 706     /* Add the forces from enforced rotation potentials (if any) */
 707     if (inputrec.bRot)
 708     {
 709         wallcycle_start(wcycle, WallCycleCounter::RotAdd);
 710         enerd->term[F_COM_PULL] +=
 711                 add_rot_forces(enforcedRotation, forceWithVirialMtsLevel0->force_, cr, step, t);
 712         wallcycle_stop(wcycle, WallCycleCounter::RotAdd);
 713     }
 714
 715     if (ed)
 716     {
 717         /* Note that since init_edsam() is called after the initialization
 718          * of forcerec, edsam doesn't request the noVirSum force buffer.
 719          * Thus if no other algorithm (e.g. PME) requires it, the forces
 720          * here will contribute to the virial.
 721          */
 722         do_flood(cr, inputrec, x, forceWithVirialMtsLevel0->force_, ed, box, step, didNeighborSearch);
 723     }
 724
 725     /* Add forces from interactive molecular dynamics (IMD), if any */
 726     if (inputrec.bIMD && stepWork.computeForces)
 727     {
 728         imdSession->applyForces(forceWithVirialMtsLevel0->force_);
 729     }
 730 }
 731
 732 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 733  *
 734  * \param[in]  pmedata              The PME structure
 735  * \param[in]  box                  The box matrix
 736  * \param[in]  stepWork             Step schedule flags
 737  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in the device memory.
 738  * \param[in]  lambdaQ              The Coulomb lambda of the current state.
 739  * \param[in]  wcycle               The wallcycle structure
 740  */
 741 static inline void launchPmeGpuSpread(gmx_pme_t*            pmedata,
 742                                       const matrix          box,
 743                                       const StepWorkload&   stepWork,
 744                                       GpuEventSynchronizer* xReadyOnDevice,
 745                                       const real            lambdaQ,
 746                                       gmx_wallcycle*        wcycle)
 747 {
 748     pme_gpu_prepare_computation(pmedata, box, wcycle, stepWork);
 749     pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle, lambdaQ);
 750 }
 751
 752 /*! \brief Launch the FFT and gather stages of PME GPU
 753  *
 754  * This function only implements setting the output forces (no accumulation).
 755  *
 756  * \param[in]  pmedata        The PME structure
 757  * \param[in]  lambdaQ        The Coulomb lambda of the current system state.
 758  * \param[in]  wcycle         The wallcycle structure
 759  * \param[in]  stepWork       Step schedule flags
 760  */
 761 static void launchPmeGpuFftAndGather(gmx_pme_t*               pmedata,
 762                                      const real               lambdaQ,
 763                                      gmx_wallcycle*           wcycle,
 764                                      const gmx::StepWorkload& stepWork)
 765 {
 766     pme_gpu_launch_complex_transforms(pmedata, wcycle, stepWork);
 767     pme_gpu_launch_gather(pmedata, wcycle, lambdaQ);
 768 }
 769
 770 /*! \brief
 771  *  Polling wait for either of the PME or nonbonded GPU tasks.
 772  *
 773  * Instead of a static order in waiting for GPU tasks, this function
 774  * polls checking which of the two tasks completes first, and does the
 775  * associated force buffer reduction overlapped with the other task.
 776  * By doing that, unlike static scheduling order, it can always overlap
 777  * one of the reductions, regardless of the GPU task completion order.
 778  *
 779  * \param[in]     nbv              Nonbonded verlet structure
 780  * \param[in,out] pmedata          PME module data
 781  * \param[in,out] forceOutputsNonbonded  Force outputs for the non-bonded forces and shift forces
 782  * \param[in,out] forceOutputsPme  Force outputs for the PME forces and virial
 783  * \param[in,out] enerd            Energy data structure results are reduced into
 784  * \param[in]     lambdaQ          The Coulomb lambda of the current system state.
 785  * \param[in]     stepWork         Step schedule flags
 786  * \param[in]     wcycle           The wallcycle structure
 787  */
 788 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
 789                                         gmx_pme_t*          pmedata,
 790                                         gmx::ForceOutputs*  forceOutputsNonbonded,
 791                                         gmx::ForceOutputs*  forceOutputsPme,
 792                                         gmx_enerdata_t*     enerd,
 793                                         const real          lambdaQ,
 794                                         const StepWorkload& stepWork,
 795                                         gmx_wallcycle*      wcycle)
 796 {
 797     bool isPmeGpuDone = false;
 798     bool isNbGpuDone  = false;
 799
 800     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 801
 802     while (!isPmeGpuDone || !isNbGpuDone)
 803     {
 804         if (!isPmeGpuDone)
 805         {
 806             GpuTaskCompletion completionType =
 807                     (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 808             isPmeGpuDone = pme_gpu_try_finish_task(
 809                     pmedata, stepWork, wcycle, &forceOutputsPme->forceWithVirial(), enerd, lambdaQ, completionType);
 810         }
 811
 812         if (!isNbGpuDone)
 813         {
 814             auto&             forceBuffersNonbonded = forceOutputsNonbonded->forceWithShiftForces();
 815             GpuTaskCompletion completionType =
 816                     (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 817             isNbGpuDone = Nbnxm::gpu_try_finish_task(
 818                     nbv->gpu_nbv,
 819                     stepWork,
 820                     AtomLocality::Local,
 821                     enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR].data(),
 822                     enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR].data(),
 823                     forceBuffersNonbonded.shiftForces(),
 824                     completionType,
 825                     wcycle);
 826
 827             if (isNbGpuDone)
 828             {
 829                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceBuffersNonbonded.force());
 830             }
 831         }
 832     }
 833 }
 834
 835 /*! \brief Set up the different force buffers; also does clearing.
 836  *
 837  * \param[in] forceHelperBuffers        Helper force buffers
 838  * \param[in] force                     force array
 839  * \param[in] domainWork                Domain lifetime workload flags
 840  * \param[in] stepWork                  Step schedule flags
 841  * \param[in] havePpDomainDecomposition Whether we have a PP domain decomposition
 842  * \param[out] wcycle                   wallcycle recording structure
 843  *
 844  * \returns                             Cleared force output structure
 845  */
 846 static ForceOutputs setupForceOutputs(ForceHelperBuffers*                 forceHelperBuffers,
 847                                       gmx::ArrayRefWithPadding<gmx::RVec> force,
 848                                       const DomainLifetimeWorkload&       domainWork,
 849                                       const StepWorkload&                 stepWork,
 850                                       const bool                          havePpDomainDecomposition,
 851                                       gmx_wallcycle*                      wcycle)
 852 {
 853     wallcycle_sub_start(wcycle, WallCycleSubCounter::ClearForceBuffer);
 854
 855     /* NOTE: We assume fr->shiftForces is all zeros here */
 856     gmx::ForceWithShiftForces forceWithShiftForces(
 857             force, stepWork.computeVirial, forceHelperBuffers->shiftForces());
 858
 859     if (stepWork.computeForces
 860         && (domainWork.haveCpuLocalForceWork || !stepWork.useGpuFBufferOps
 861             || (havePpDomainDecomposition && !stepWork.useGpuFHalo)))
 862     {
 863         /* Clear the short- and long-range forces */
 864         clearRVecs(forceWithShiftForces.force(), true);
 865
 866         /* Clear the shift forces */
 867         clearRVecs(forceWithShiftForces.shiftForces(), false);
 868     }
 869
 870     /* If we need to compute the virial, we might need a separate
 871      * force buffer for algorithms for which the virial is calculated
 872      * directly, such as PME. Otherwise, forceWithVirial uses the
 873      * the same force (f in legacy calls) buffer as other algorithms.
 874      */
 875     const bool useSeparateForceWithVirialBuffer =
 876             (stepWork.computeForces
 877              && (stepWork.computeVirial && forceHelperBuffers->haveDirectVirialContributions()));
 878     /* forceWithVirial uses the local atom range only */
 879     gmx::ForceWithVirial forceWithVirial(
 880             useSeparateForceWithVirialBuffer ? forceHelperBuffers->forceBufferForDirectVirialContributions()
 881                                              : force.unpaddedArrayRef(),
 882             stepWork.computeVirial);
 883
 884     if (useSeparateForceWithVirialBuffer)
 885     {
 886         /* TODO: update comment
 887          * We only compute forces on local atoms. Note that vsites can
 888          * spread to non-local atoms, but that part of the buffer is
 889          * cleared separately in the vsite spreading code.
 890          */
 891         clearRVecs(forceWithVirial.force_, true);
 892     }
 893
 894     wallcycle_sub_stop(wcycle, WallCycleSubCounter::ClearForceBuffer);
 895
 896     return ForceOutputs(
 897             forceWithShiftForces, forceHelperBuffers->haveDirectVirialContributions(), forceWithVirial);
 898 }
 899
 900
 901 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 902  */
 903 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec& inputrec,
 904                                                           const t_forcerec& fr,
 905                                                           const pull_t*     pull_work,
 906                                                           const gmx_edsam*  ed,
 907                                                           const t_mdatoms&  mdatoms,
 908                                                           bool havePPDomainDecomposition,
 909                                                           const SimulationWorkload& simulationWork,
 910                                                           const StepWorkload&       stepWork)
 911 {
 912     DomainLifetimeWorkload domainWork;
 913     // Note that haveSpecialForces is constant over the whole run
 914     domainWork.haveSpecialForces =
 915             haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 916     domainWork.haveCpuListedForceWork = false;
 917     domainWork.haveCpuBondedWork      = false;
 918     for (const auto& listedForces : fr.listedForces)
 919     {
 920         if (listedForces.haveCpuListedForces(*fr.fcdata))
 921         {
 922             domainWork.haveCpuListedForceWork = true;
 923         }
 924         if (listedForces.haveCpuBondeds())
 925         {
 926             domainWork.haveCpuBondedWork = true;
 927         }
 928     }
 929     domainWork.haveGpuBondedWork =
 930             ((fr.listedForcesGpu != nullptr) && fr.listedForcesGpu->haveInteractions());
 931     // Note that haveFreeEnergyWork is constant over the whole run
 932     domainWork.haveFreeEnergyWork =
 933             (fr.efep != FreeEnergyPerturbationType::No && mdatoms.nPerturbed != 0);
 934     // We assume we have local force work if there are CPU
 935     // force tasks including PME or nonbondeds.
 936     domainWork.haveCpuLocalForceWork =
 937             domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
 938             || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
 939             || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
 940     domainWork.haveLocalForceContribInCpuBuffer =
 941             domainWork.haveCpuLocalForceWork || havePPDomainDecomposition;
 942     domainWork.haveNonLocalForceContribInCpuBuffer =
 943             domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
 944
 945     return domainWork;
 946 }
 947
 948 /*! \brief Set up force flag stuct from the force bitmask.
 949  *
 950  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 951  * \param[in]      mtsLevels            The multiple time-stepping levels, either empty or 2 levels
 952  * \param[in]      step                 The current MD step
 953  * \param[in]      simulationWork       Simulation workload description.
 954  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 955  *
 956  * \returns New Stepworkload description.
 957  */
 958 static StepWorkload setupStepWorkload(const int                     legacyFlags,
 959                                       ArrayRef<const gmx::MtsLevel> mtsLevels,
 960                                       const int64_t                 step,
 961                                       const SimulationWorkload&     simulationWork,
 962                                       const bool                    rankHasPmeDuty)
 963 {
 964     GMX_ASSERT(mtsLevels.empty() || mtsLevels.size() == 2, "Expect 0 or 2 MTS levels");
 965     const bool computeSlowForces = (mtsLevels.empty() || step % mtsLevels[1].stepFactor == 0);
 966
 967     StepWorkload flags;
 968     flags.stateChanged                  = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 969     flags.haveDynamicBox                = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 970     flags.doNeighborSearch              = ((legacyFlags & GMX_FORCE_NS) != 0);
 971     flags.computeSlowForces             = computeSlowForces;
 972     flags.computeVirial                 = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 973     flags.computeEnergy                 = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 974     flags.computeForces                 = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 975     flags.useOnlyMtsCombinedForceBuffer = ((legacyFlags & GMX_FORCE_DO_NOT_NEED_NORMAL_FORCE) != 0);
 976     flags.computeListedForces           = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 977     flags.computeNonbondedForces =
 978             ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && simulationWork.computeNonbonded
 979             && !(simulationWork.computeNonbondedAtMtsLevel1 && !computeSlowForces);
 980     flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 981
 982     if (simulationWork.useGpuBufferOps)
 983     {
 984         GMX_ASSERT(simulationWork.useGpuNonbonded,
 985                    "Can only offload buffer ops if nonbonded computation is also offloaded");
 986     }
 987     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 988     // on virial steps the CPU reduction path is taken
 989     flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !flags.computeVirial;
 990     flags.useGpuPmeFReduction = flags.computeSlowForces && flags.useGpuFBufferOps && simulationWork.useGpuPme
 991                                 && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication);
 992     flags.useGpuXHalo          = simulationWork.useGpuHaloExchange;
 993     flags.useGpuFHalo          = simulationWork.useGpuHaloExchange && flags.useGpuFBufferOps;
 994     flags.haveGpuPmeOnThisRank = simulationWork.useGpuPme && rankHasPmeDuty && flags.computeSlowForces;
 995     flags.combineMtsForcesBeforeHaloExchange =
 996             (flags.computeForces && simulationWork.useMts && flags.computeSlowForces
 997              && flags.useOnlyMtsCombinedForceBuffer
 998              && !(flags.computeVirial || simulationWork.useGpuNonbonded || flags.haveGpuPmeOnThisRank));
 999
1000     return flags;
1001 }
1002
1003
1004 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
1005  *
1006  */
1007 static void launchGpuEndOfStepTasks(nonbonded_verlet_t*               nbv,
1008                                     gmx::ListedForcesGpu*             listedForcesGpu,
1009                                     gmx_pme_t*                        pmedata,
1010                                     gmx_enerdata_t*                   enerd,
1011                                     const gmx::MdrunScheduleWorkload& runScheduleWork,
1012                                     int64_t                           step,
1013                                     gmx_wallcycle*                    wcycle)
1014 {
1015     if (runScheduleWork.simulationWork.useGpuNonbonded && runScheduleWork.stepWork.computeNonbondedForces)
1016     {
1017         /* Launch pruning before buffer clearing because the API overhead of the
1018          * clear kernel launches can leave the GPU idle while it could be running
1019          * the prune kernel.
1020          */
1021         if (nbv->isDynamicPruningStepGpu(step))
1022         {
1023             nbv->dispatchPruneKernelGpu(step);
1024         }
1025
1026         /* now clear the GPU outputs while we finish the step on the CPU */
1027         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1028         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1029         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
1030         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1031         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1032     }
1033
1034     if (runScheduleWork.stepWork.haveGpuPmeOnThisRank)
1035     {
1036         pme_gpu_reinit_computation(pmedata, wcycle);
1037     }
1038
1039     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
1040     {
1041         // in principle this should be included in the DD balancing region,
1042         // but generally it is infrequent so we'll omit it for the sake of
1043         // simpler code
1044         listedForcesGpu->waitAccumulateEnergyTerms(enerd);
1045
1046         listedForcesGpu->clearEnergies();
1047     }
1048 }
1049
1050 //! \brief Data structure to hold dipole-related data and staging arrays
1051 struct DipoleData
1052 {
1053     //! Dipole staging for fast summing over MPI
1054     gmx::DVec muStaging[2] = { { 0.0, 0.0, 0.0 } };
1055     //! Dipole staging for states A and B (index 0 and 1 resp.)
1056     gmx::RVec muStateAB[2] = { { 0.0_real, 0.0_real, 0.0_real } };
1057 };
1058
1059
1060 static void reduceAndUpdateMuTot(DipoleData*                   dipoleData,
1061                                  const t_commrec*              cr,
1062                                  const bool                    haveFreeEnergy,
1063                                  gmx::ArrayRef<const real>     lambda,
1064                                  rvec                          muTotal,
1065                                  const DDBalanceRegionHandler& ddBalanceRegionHandler)
1066 {
1067     if (PAR(cr))
1068     {
1069         gmx_sumd(2 * DIM, dipoleData->muStaging[0], cr);
1070         ddBalanceRegionHandler.reopenRegionCpu();
1071     }
1072     for (int i = 0; i < 2; i++)
1073     {
1074         for (int j = 0; j < DIM; j++)
1075         {
1076             dipoleData->muStateAB[i][j] = dipoleData->muStaging[i][j];
1077         }
1078     }
1079
1080     if (!haveFreeEnergy)
1081     {
1082         copy_rvec(dipoleData->muStateAB[0], muTotal);
1083     }
1084     else
1085     {
1086         for (int j = 0; j < DIM; j++)
1087         {
1088             muTotal[j] = (1.0 - lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)])
1089                                  * dipoleData->muStateAB[0][j]
1090                          + lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)]
1091                                    * dipoleData->muStateAB[1][j];
1092         }
1093     }
1094 }
1095
1096 /*! \brief Combines MTS level0 and level1 force buffes into a full and MTS-combined force buffer.
1097  *
1098  * \param[in]     numAtoms        The number of atoms to combine forces for
1099  * \param[in,out] forceMtsLevel0  Input: F_level0, output: F_level0 + F_level1
1100  * \param[in,out] forceMts        Input: F_level1, output: F_level0 + mtsFactor * F_level1
1101  * \param[in]     mtsFactor       The factor between the level0 and level1 time step
1102  */
1103 static void combineMtsForces(const int      numAtoms,
1104                              ArrayRef<RVec> forceMtsLevel0,
1105                              ArrayRef<RVec> forceMts,
1106                              const real     mtsFactor)
1107 {
1108     const int gmx_unused numThreads = gmx_omp_nthreads_get(ModuleMultiThread::Default);
1109 #pragma omp parallel for num_threads(numThreads) schedule(static)
1110     for (int i = 0; i < numAtoms; i++)
1111     {
1112         const RVec forceMtsLevel0Tmp = forceMtsLevel0[i];
1113         forceMtsLevel0[i] += forceMts[i];
1114         forceMts[i] = forceMtsLevel0Tmp + mtsFactor * forceMts[i];
1115     }
1116 }
1117
1118 /*! \brief Setup for the local and non-local GPU force reductions:
1119  * reinitialization plus the registration of forces and dependencies.
1120  *
1121  * \param [in] runScheduleWork               Schedule workload flag structure
1122  * \param [in] cr                            Communication record object
1123  * \param [in] fr                            Force record object
1124  */
1125 static void setupGpuForceReductions(gmx::MdrunScheduleWorkload* runScheduleWork,
1126                                     const t_commrec*            cr,
1127                                     t_forcerec*                 fr)
1128 {
1129
1130     nonbonded_verlet_t*          nbv      = fr->nbv.get();
1131     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1132
1133     // (re-)initialize local GPU force reduction
1134     const bool accumulate =
1135             runScheduleWork->domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr);
1136     const int atomStart = 0;
1137     fr->gpuForceReduction[gmx::AtomLocality::Local]->reinit(stateGpu->getForces(),
1138                                                             nbv->getNumAtoms(AtomLocality::Local),
1139                                                             nbv->getGridIndices(),
1140                                                             atomStart,
1141                                                             accumulate,
1142                                                             stateGpu->fReducedOnDevice());
1143
1144     // register forces and add dependencies
1145     fr->gpuForceReduction[gmx::AtomLocality::Local]->registerNbnxmForce(Nbnxm::gpu_get_f(nbv->gpu_nbv));
1146
1147     if (runScheduleWork->simulationWork.useGpuPme
1148         && (thisRankHasDuty(cr, DUTY_PME) || runScheduleWork->simulationWork.useGpuPmePpCommunication))
1149     {
1150         DeviceBuffer<gmx::RVec> forcePtr =
1151                 thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1152                                               :                    // PME force buffer on same GPU
1153                         fr->pmePpCommGpu->getGpuForceStagingPtr(); // buffer received from other GPU
1154         fr->gpuForceReduction[gmx::AtomLocality::Local]->registerRvecForce(forcePtr);
1155
1156         GpuEventSynchronizer* const pmeSynchronizer =
1157                 (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1158                                                : // PME force buffer on same GPU
1159                          fr->pmePpCommGpu->getForcesReadySynchronizer()); // buffer received from other GPU
1160
1161         if (GMX_THREAD_MPI)
1162         {
1163             GMX_ASSERT(pmeSynchronizer != nullptr, "PME force ready cuda event should not be NULL");
1164             fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(pmeSynchronizer);
1165         }
1166     }
1167
1168     if (runScheduleWork->domainWork.haveCpuLocalForceWork && !runScheduleWork->simulationWork.useGpuHaloExchange)
1169     {
1170         // in the DD case we use the same stream for H2D and reduction, hence no explicit dependency needed
1171         if (!havePPDomainDecomposition(cr))
1172         {
1173             const bool useGpuForceBufferOps = true;
1174             fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1175                     stateGpu->getForcesReadyOnDeviceEvent(AtomLocality::All, useGpuForceBufferOps));
1176         }
1177     }
1178
1179     if (runScheduleWork->simulationWork.useGpuHaloExchange)
1180     {
1181         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1182                 cr->dd->gpuHaloExchange[0][0]->getForcesReadyOnDeviceEvent());
1183     }
1184
1185     if (havePPDomainDecomposition(cr))
1186     {
1187         // (re-)initialize non-local GPU force reduction
1188         const bool accumulate = runScheduleWork->domainWork.haveCpuBondedWork
1189                                 || runScheduleWork->domainWork.haveFreeEnergyWork;
1190         const int atomStart = dd_numHomeAtoms(*cr->dd);
1191         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->reinit(stateGpu->getForces(),
1192                                                                    nbv->getNumAtoms(AtomLocality::NonLocal),
1193                                                                    nbv->getGridIndices(),
1194                                                                    atomStart,
1195                                                                    accumulate);
1196
1197         // register forces and add dependencies
1198         // in the DD case we use the same stream for H2D and reduction, hence no explicit dependency needed
1199         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->registerNbnxmForce(
1200                 Nbnxm::gpu_get_f(nbv->gpu_nbv));
1201     }
1202 }
1203
1204
1205 /*! \brief Return the number of local atoms.
1206  */
1207 static int getLocalAtomCount(const gmx_domdec_t* dd, const t_mdatoms& mdatoms, bool havePPDomainDecomposition)
1208 {
1209     GMX_ASSERT(!(havePPDomainDecomposition && (dd == nullptr)),
1210                "Can't have PP decomposition with dd uninitialized!");
1211     return havePPDomainDecomposition ? dd_numAtomsZones(*dd) : mdatoms.homenr;
1212 }
1213
1214
1215 void do_force(FILE*                               fplog,
1216               const t_commrec*                    cr,
1217               const gmx_multisim_t*               ms,
1218               const t_inputrec&                   inputrec,
1219               gmx::Awh*                           awh,
1220               gmx_enfrot*                         enforcedRotation,
1221               gmx::ImdSession*                    imdSession,
1222               pull_t*                             pull_work,
1223               int64_t                             step,
1224               t_nrnb*                             nrnb,
1225               gmx_wallcycle*                      wcycle,
1226               const gmx_localtop_t*               top,
1227               const matrix                        box,
1228               gmx::ArrayRefWithPadding<gmx::RVec> x,
1229               const history_t*                    hist,
1230               gmx::ForceBuffersView*              forceView,
1231               tensor                              vir_force,
1232               const t_mdatoms*                    mdatoms,
1233               gmx_enerdata_t*                     enerd,
1234               gmx::ArrayRef<const real>           lambda,
1235               t_forcerec*                         fr,
1236               gmx::MdrunScheduleWorkload*         runScheduleWork,
1237               gmx::VirtualSitesHandler*           vsite,
1238               rvec                                muTotal,
1239               double                              t,
1240               gmx_edsam*                          ed,
1241               int                                 legacyFlags,
1242               const DDBalanceRegionHandler&       ddBalanceRegionHandler)
1243 {
1244     auto force = forceView->forceWithPadding();
1245     GMX_ASSERT(force.unpaddedArrayRef().ssize() >= fr->natoms_force_constr,
1246                "The size of the force buffer should be at least the number of atoms to compute "
1247                "forces for");
1248
1249     nonbonded_verlet_t*  nbv = fr->nbv.get();
1250     interaction_const_t* ic  = fr->ic.get();
1251
1252     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1253
1254     const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
1255
1256     runScheduleWork->stepWork = setupStepWorkload(
1257             legacyFlags, inputrec.mtsLevels, step, simulationWork, thisRankHasDuty(cr, DUTY_PME));
1258     const StepWorkload& stepWork = runScheduleWork->stepWork;
1259
1260     /* At a search step we need to start the first balancing region
1261      * somewhere early inside the step after communication during domain
1262      * decomposition (and not during the previous step as usual).
1263      */
1264     if (stepWork.doNeighborSearch)
1265     {
1266         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
1267     }
1268
1269     clear_mat(vir_force);
1270
1271     if (fr->pbcType != PbcType::No)
1272     {
1273         /* Compute shift vectors every step,
1274          * because of pressure coupling or box deformation!
1275          */
1276         if (stepWork.haveDynamicBox && stepWork.stateChanged)
1277         {
1278             calc_shifts(box, fr->shift_vec);
1279         }
1280
1281         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1282         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
1283         if (calcCGCM)
1284         {
1285             put_atoms_in_box_omp(fr->pbcType,
1286                                  box,
1287                                  x.unpaddedArrayRef().subArray(0, mdatoms->homenr),
1288                                  gmx_omp_nthreads_get(ModuleMultiThread::Default));
1289             inc_nrnb(nrnb, eNR_SHIFTX, mdatoms->homenr);
1290         }
1291     }
1292
1293     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1294
1295     const bool pmeSendCoordinatesFromGpu =
1296             simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1297     const bool reinitGpuPmePpComms =
1298             simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1299
1300     auto* localXReadyOnDevice = (stepWork.haveGpuPmeOnThisRank || simulationWork.useGpuBufferOps)
1301                                         ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1302                                                 AtomLocality::Local, simulationWork, stepWork)
1303                                         : nullptr;
1304
1305     // Copy coordinate from the GPU if update is on the GPU and there
1306     // are forces to be computed on the CPU, or for the computation of
1307     // virial, or if host-side data will be transferred from this task
1308     // to a remote task for halo exchange or PME-PP communication. At
1309     // search steps the current coordinates are already on the host,
1310     // hence copy is not needed.
1311     const bool haveHostPmePpComms =
1312             !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1313
1314     GMX_ASSERT(simulationWork.useGpuHaloExchange
1315                        == ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange[0].empty())),
1316                "The GPU halo exchange is active, but it has not been constructed.");
1317     const bool haveHostHaloExchangeComms =
1318             havePPDomainDecomposition(cr) && !simulationWork.useGpuHaloExchange;
1319
1320     bool gmx_used_in_debug haveCopiedXFromGpu = false;
1321     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1322         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1323             || haveHostPmePpComms || haveHostHaloExchangeComms || simulationWork.computeMuTot))
1324     {
1325         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1326         haveCopiedXFromGpu = true;
1327     }
1328
1329     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1330     // The local coordinates can be copied right away.
1331     // NOTE: Consider moving this copy to right after they are updated and constrained,
1332     //       if the later is not offloaded.
1333     if (stepWork.haveGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1334     {
1335         if (stepWork.doNeighborSearch)
1336         {
1337             // TODO refactor this to do_md, after partitioning.
1338             stateGpu->reinit(mdatoms->homenr,
1339                              getLocalAtomCount(cr->dd, *mdatoms, havePPDomainDecomposition(cr)));
1340             if (stepWork.haveGpuPmeOnThisRank)
1341             {
1342                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1343                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1344             }
1345         }
1346         // We need to copy coordinates when:
1347         // 1. Update is not offloaded
1348         // 2. The buffers were reinitialized on search step
1349         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1350         {
1351             GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1352             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1353         }
1354     }
1355
1356     if (!thisRankHasDuty(cr, DUTY_PME) && stepWork.computeSlowForces)
1357     {
1358         /* Send particle coordinates to the pme nodes */
1359         if (!pmeSendCoordinatesFromGpu && !stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1360         {
1361             GMX_ASSERT(haveCopiedXFromGpu,
1362                        "a wait should only be triggered if copy has been scheduled");
1363             stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1364         }
1365
1366         gmx_pme_send_coordinates(fr,
1367                                  cr,
1368                                  box,
1369                                  x.unpaddedArrayRef(),
1370                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1371                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)],
1372                                  (stepWork.computeVirial || stepWork.computeEnergy),
1373                                  step,
1374                                  simulationWork.useGpuPmePpCommunication,
1375                                  reinitGpuPmePpComms,
1376                                  pmeSendCoordinatesFromGpu,
1377                                  stepWork.useGpuPmeFReduction,
1378                                  localXReadyOnDevice,
1379                                  wcycle);
1380     }
1381
1382     if (stepWork.haveGpuPmeOnThisRank)
1383     {
1384         launchPmeGpuSpread(fr->pmedata,
1385                            box,
1386                            stepWork,
1387                            localXReadyOnDevice,
1388                            lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1389                            wcycle);
1390     }
1391
1392     const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1393
1394     /* do gridding for pair search */
1395     if (stepWork.doNeighborSearch)
1396     {
1397         if (fr->wholeMoleculeTransform && stepWork.stateChanged)
1398         {
1399             fr->wholeMoleculeTransform->updateForAtomPbcJumps(x.unpaddedArrayRef(), box);
1400         }
1401
1402         wallcycle_start(wcycle, WallCycleCounter::NS);
1403         if (!DOMAINDECOMP(cr))
1404         {
1405             const rvec vzero       = { 0.0_real, 0.0_real, 0.0_real };
1406             const rvec boxDiagonal = { box[XX][XX], box[YY][YY], box[ZZ][ZZ] };
1407             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSGridLocal);
1408             nbnxn_put_on_grid(nbv,
1409                               box,
1410                               0,
1411                               vzero,
1412                               boxDiagonal,
1413                               nullptr,
1414                               { 0, mdatoms->homenr },
1415                               -1,
1416                               fr->atomInfo,
1417                               x.unpaddedArrayRef(),
1418                               0,
1419                               nullptr);
1420             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSGridLocal);
1421         }
1422         else
1423         {
1424             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSGridNonLocal);
1425             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->atomInfo, x.unpaddedArrayRef());
1426             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSGridNonLocal);
1427         }
1428
1429         nbv->setAtomProperties(gmx::constArrayRefFromArray(mdatoms->typeA, mdatoms->nr),
1430                                gmx::constArrayRefFromArray(mdatoms->chargeA, mdatoms->nr),
1431                                fr->atomInfo);
1432
1433         wallcycle_stop(wcycle, WallCycleCounter::NS);
1434
1435         /* initialize the GPU nbnxm atom data and bonded data structures */
1436         if (simulationWork.useGpuNonbonded)
1437         {
1438             // Note: cycle counting only nononbondeds, GPU listed forces counts internally
1439             wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1440             wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1441             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1442             wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1443             wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1444
1445             if (fr->listedForcesGpu)
1446             {
1447                 /* Now we put all atoms on the grid, we can assign bonded
1448                  * interactions to the GPU, where the grid order is
1449                  * needed. Also the xq, f and fshift device buffers have
1450                  * been reallocated if needed, so the bonded code can
1451                  * learn about them. */
1452                 // TODO the xq, f, and fshift buffers are now shared
1453                 // resources, so they should be maintained by a
1454                 // higher-level object than the nb module.
1455                 fr->listedForcesGpu->updateInteractionListsAndDeviceBuffers(
1456                         nbv->getGridIndices(),
1457                         top->idef,
1458                         Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1459                         Nbnxm::gpu_get_f(nbv->gpu_nbv),
1460                         Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1461             }
1462         }
1463
1464         // Need to run after the GPU-offload bonded interaction lists
1465         // are set up to be able to determine whether there is bonded work.
1466         runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1467                 inputrec, *fr, pull_work, ed, *mdatoms, havePPDomainDecomposition(cr), simulationWork, stepWork);
1468
1469         wallcycle_start_nocount(wcycle, WallCycleCounter::NS);
1470         wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSSearchLocal);
1471         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1472         nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1473
1474         nbv->setupGpuShortRangeWork(fr->listedForcesGpu.get(), InteractionLocality::Local);
1475
1476         wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSSearchLocal);
1477         wallcycle_stop(wcycle, WallCycleCounter::NS);
1478
1479         if (stepWork.useGpuXBufferOps)
1480         {
1481             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1482         }
1483
1484         if (simulationWork.useGpuBufferOps)
1485         {
1486             setupGpuForceReductions(runScheduleWork, cr, fr);
1487         }
1488     }
1489     else if (!EI_TPI(inputrec.eI) && stepWork.computeNonbondedForces)
1490     {
1491         if (stepWork.useGpuXBufferOps)
1492         {
1493             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1494             nbv->convertCoordinatesGpu(AtomLocality::Local, stateGpu->getCoordinates(), localXReadyOnDevice);
1495         }
1496         else
1497         {
1498             if (simulationWork.useGpuUpdate)
1499             {
1500                 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1501                 GMX_ASSERT(haveCopiedXFromGpu,
1502                            "a wait should only be triggered if copy has been scheduled");
1503                 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1504             }
1505             nbv->convertCoordinates(AtomLocality::Local, x.unpaddedArrayRef());
1506         }
1507     }
1508
1509     if (simulationWork.useGpuNonbonded && (stepWork.computeNonbondedForces || domainWork.haveGpuBondedWork))
1510     {
1511         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1512
1513         wallcycle_start(wcycle, WallCycleCounter::LaunchGpu);
1514         wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1515         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1516         if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1517         {
1518             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1519         }
1520         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1521         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1522         // with X buffer ops offloaded to the GPU on all but the search steps
1523
1524         // bonded work not split into separate local and non-local, so with DD
1525         // we can only launch the kernel after non-local coordinates have been received.
1526         if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1527         {
1528             fr->listedForcesGpu->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1529         }
1530
1531         /* launch local nonbonded work on GPU */
1532         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1533         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1534         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1535         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1536         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1537     }
1538
1539     if (stepWork.haveGpuPmeOnThisRank)
1540     {
1541         // In PME GPU and mixed mode we launch FFT / gather after the
1542         // X copy/transform to allow overlap as well as after the GPU NB
1543         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1544         // the nonbonded kernel.
1545         launchPmeGpuFftAndGather(fr->pmedata,
1546                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1547                                  wcycle,
1548                                  stepWork);
1549     }
1550
1551     /* Communicate coordinates and sum dipole if necessary +
1552        do non-local pair search */
1553     if (havePPDomainDecomposition(cr))
1554     {
1555         if (stepWork.doNeighborSearch)
1556         {
1557             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1558             wallcycle_start_nocount(wcycle, WallCycleCounter::NS);
1559             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSSearchNonLocal);
1560             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1561             nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1562
1563             nbv->setupGpuShortRangeWork(fr->listedForcesGpu.get(), InteractionLocality::NonLocal);
1564             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSSearchNonLocal);
1565             wallcycle_stop(wcycle, WallCycleCounter::NS);
1566             // TODO refactor this GPU halo exchange re-initialisation
1567             // to location in do_md where GPU halo exchange is
1568             // constructed at partitioning, after above stateGpu
1569             // re-initialization has similarly been refactored
1570             if (simulationWork.useGpuHaloExchange)
1571             {
1572                 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1573             }
1574         }
1575         else
1576         {
1577             if (stepWork.useGpuXHalo)
1578             {
1579                 // The following must be called after local setCoordinates (which records an event
1580                 // when the coordinate data has been copied to the device).
1581                 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1582
1583                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1584                 {
1585                     // non-local part of coordinate buffer must be copied back to host for CPU work
1586                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1587                 }
1588             }
1589             else
1590             {
1591                 if (simulationWork.useGpuUpdate)
1592                 {
1593                     GMX_ASSERT(haveCopiedXFromGpu,
1594                                "a wait should only be triggered if copy has been scheduled");
1595                     stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1596                 }
1597                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1598             }
1599
1600             if (stepWork.useGpuXBufferOps)
1601             {
1602                 if (!stepWork.haveGpuPmeOnThisRank && !stepWork.useGpuXHalo)
1603                 {
1604                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1605                 }
1606                 nbv->convertCoordinatesGpu(AtomLocality::NonLocal,
1607                                            stateGpu->getCoordinates(),
1608                                            stateGpu->getCoordinatesReadyOnDeviceEvent(
1609                                                    AtomLocality::NonLocal, simulationWork, stepWork));
1610             }
1611             else
1612             {
1613                 nbv->convertCoordinates(AtomLocality::NonLocal, x.unpaddedArrayRef());
1614             }
1615         }
1616
1617         if (simulationWork.useGpuNonbonded)
1618         {
1619
1620             if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1621             {
1622                 wallcycle_start(wcycle, WallCycleCounter::LaunchGpu);
1623                 wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1624                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1625                 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1626                 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1627             }
1628
1629             if (domainWork.haveGpuBondedWork)
1630             {
1631                 fr->listedForcesGpu->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1632             }
1633
1634             /* launch non-local nonbonded tasks on GPU */
1635             wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1636             wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1637             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step, nrnb, wcycle);
1638             wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1639             wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1640         }
1641     }
1642
1643     if (simulationWork.useGpuNonbonded && stepWork.computeNonbondedForces)
1644     {
1645         /* launch D2H copy-back F */
1646         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1647         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1648
1649         if (havePPDomainDecomposition(cr))
1650         {
1651             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1652         }
1653         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1654         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1655
1656         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1657         {
1658             fr->listedForcesGpu->launchEnergyTransfer();
1659         }
1660         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1661     }
1662
1663     gmx::ArrayRef<const gmx::RVec> xWholeMolecules;
1664     if (fr->wholeMoleculeTransform)
1665     {
1666         xWholeMolecules = fr->wholeMoleculeTransform->wholeMoleculeCoordinates(x.unpaddedArrayRef(), box);
1667     }
1668
1669     // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1670     // this wait ensures that the D2H transfer is complete.
1671     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1672         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1673             || simulationWork.computeMuTot))
1674     {
1675         GMX_ASSERT(haveCopiedXFromGpu, "a wait should only be triggered if copy has been scheduled");
1676         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1677     }
1678
1679     DipoleData dipoleData;
1680
1681     if (simulationWork.computeMuTot)
1682     {
1683         const int start = 0;
1684
1685         /* Calculate total (local) dipole moment in a temporary common array.
1686          * This makes it possible to sum them over nodes faster.
1687          */
1688         gmx::ArrayRef<const gmx::RVec> xRef =
1689                 (xWholeMolecules.empty() ? x.unpaddedArrayRef() : xWholeMolecules);
1690         calc_mu(start,
1691                 mdatoms->homenr,
1692                 xRef,
1693                 mdatoms->chargeA ? gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->nr)
1694                                  : gmx::ArrayRef<real>{},
1695                 mdatoms->chargeB ? gmx::arrayRefFromArray(mdatoms->chargeB, mdatoms->nr)
1696                                  : gmx::ArrayRef<real>{},
1697                 mdatoms->nChargePerturbed != 0,
1698                 dipoleData.muStaging[0],
1699                 dipoleData.muStaging[1]);
1700
1701         reduceAndUpdateMuTot(
1702                 &dipoleData, cr, (fr->efep != FreeEnergyPerturbationType::No), lambda, muTotal, ddBalanceRegionHandler);
1703     }
1704
1705     /* Reset energies */
1706     reset_enerdata(enerd);
1707
1708     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1709     {
1710         wallcycle_start(wcycle, WallCycleCounter::PpDuringPme);
1711         dd_force_flop_start(cr->dd, nrnb);
1712     }
1713
1714     if (inputrec.bRot)
1715     {
1716         wallcycle_start(wcycle, WallCycleCounter::Rot);
1717         do_rotation(cr, enforcedRotation, box, x.unpaddedConstArrayRef(), t, step, stepWork.doNeighborSearch);
1718         wallcycle_stop(wcycle, WallCycleCounter::Rot);
1719     }
1720
1721     /* Start the force cycle counter.
1722      * Note that a different counter is used for dynamic load balancing.
1723      */
1724     wallcycle_start(wcycle, WallCycleCounter::Force);
1725
1726     /* Set up and clear force outputs:
1727      * forceOutMtsLevel0:  everything except what is in the other two outputs
1728      * forceOutMtsLevel1:  PME-mesh and listed-forces group 1
1729      * forceOutNonbonded: non-bonded forces
1730      * Without multiple time stepping all point to the same object.
1731      * With multiple time-stepping the use is different for MTS fast (level0 only) and slow steps.
1732      */
1733     ForceOutputs forceOutMtsLevel0 = setupForceOutputs(
1734             &fr->forceHelperBuffers[0], force, domainWork, stepWork, havePPDomainDecomposition(cr), wcycle);
1735
1736     // Force output for MTS combined forces, only set at level1 MTS steps
1737     std::optional<ForceOutputs> forceOutMts =
1738             (simulationWork.useMts && stepWork.computeSlowForces)
1739                     ? std::optional(setupForceOutputs(&fr->forceHelperBuffers[1],
1740                                                       forceView->forceMtsCombinedWithPadding(),
1741                                                       domainWork,
1742                                                       stepWork,
1743                                                       havePPDomainDecomposition(cr),
1744                                                       wcycle))
1745                     : std::nullopt;
1746
1747     ForceOutputs* forceOutMtsLevel1 =
1748             simulationWork.useMts ? (stepWork.computeSlowForces ? &forceOutMts.value() : nullptr)
1749                                   : &forceOutMtsLevel0;
1750
1751     const bool nonbondedAtMtsLevel1 = runScheduleWork->simulationWork.computeNonbondedAtMtsLevel1;
1752
1753     ForceOutputs* forceOutNonbonded = nonbondedAtMtsLevel1 ? forceOutMtsLevel1 : &forceOutMtsLevel0;
1754
1755     if (inputrec.bPull && pull_have_constraint(*pull_work))
1756     {
1757         clear_pull_forces(pull_work);
1758     }
1759
1760     /* We calculate the non-bonded forces, when done on the CPU, here.
1761      * We do this before calling do_force_lowlevel, because in that
1762      * function, the listed forces are calculated before PME, which
1763      * does communication.  With this order, non-bonded and listed
1764      * force calculation imbalance can be balanced out by the domain
1765      * decomposition load balancing.
1766      */
1767
1768     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1769
1770     if (!useOrEmulateGpuNb)
1771     {
1772         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1773     }
1774
1775     if (fr->efep != FreeEnergyPerturbationType::No && stepWork.computeNonbondedForces)
1776     {
1777         /* Calculate the local and non-local free energy interactions here.
1778          * Happens here on the CPU both with and without GPU.
1779          */
1780         nbv->dispatchFreeEnergyKernel(
1781                 InteractionLocality::Local,
1782                 x.unpaddedArrayRef(),
1783                 &forceOutNonbonded->forceWithShiftForces(),
1784                 fr->use_simd_kernels,
1785                 fr->ntype,
1786                 fr->rlist,
1787                 *fr->ic,
1788                 fr->shift_vec,
1789                 fr->nbfp,
1790                 fr->ljpme_c6grid,
1791                 mdatoms->chargeA ? gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->nr)
1792                                  : gmx::ArrayRef<real>{},
1793                 mdatoms->chargeB ? gmx::arrayRefFromArray(mdatoms->chargeB, mdatoms->nr)
1794                                  : gmx::ArrayRef<real>{},
1795                 mdatoms->typeA ? gmx::arrayRefFromArray(mdatoms->typeA, mdatoms->nr)
1796                                : gmx::ArrayRef<int>{},
1797                 mdatoms->typeB ? gmx::arrayRefFromArray(mdatoms->typeB, mdatoms->nr)
1798                                : gmx::ArrayRef<int>{},
1799                 inputrec.fepvals.get(),
1800                 lambda,
1801                 enerd,
1802                 stepWork,
1803                 nrnb);
1804
1805         if (havePPDomainDecomposition(cr))
1806         {
1807             nbv->dispatchFreeEnergyKernel(
1808                     InteractionLocality::NonLocal,
1809                     x.unpaddedArrayRef(),
1810                     &forceOutNonbonded->forceWithShiftForces(),
1811                     fr->use_simd_kernels,
1812                     fr->ntype,
1813                     fr->rlist,
1814                     *fr->ic,
1815                     fr->shift_vec,
1816                     fr->nbfp,
1817                     fr->ljpme_c6grid,
1818                     mdatoms->chargeA ? gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->nr)
1819                                      : gmx::ArrayRef<real>{},
1820                     mdatoms->chargeB ? gmx::arrayRefFromArray(mdatoms->chargeB, mdatoms->nr)
1821                                      : gmx::ArrayRef<real>{},
1822                     mdatoms->typeA ? gmx::arrayRefFromArray(mdatoms->typeA, mdatoms->nr)
1823                                    : gmx::ArrayRef<int>{},
1824                     mdatoms->typeB ? gmx::arrayRefFromArray(mdatoms->typeB, mdatoms->nr)
1825                                    : gmx::ArrayRef<int>{},
1826                     inputrec.fepvals.get(),
1827                     lambda,
1828                     enerd,
1829                     stepWork,
1830                     nrnb);
1831         }
1832     }
1833
1834     if (stepWork.computeNonbondedForces && !useOrEmulateGpuNb)
1835     {
1836         if (havePPDomainDecomposition(cr))
1837         {
1838             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step, nrnb, wcycle);
1839         }
1840
1841         if (stepWork.computeForces)
1842         {
1843             /* Add all the non-bonded force to the normal force array.
1844              * This can be split into a local and a non-local part when overlapping
1845              * communication with calculation with domain decomposition.
1846              */
1847             wallcycle_stop(wcycle, WallCycleCounter::Force);
1848             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All,
1849                                           forceOutNonbonded->forceWithShiftForces().force());
1850             wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
1851         }
1852
1853         /* If there are multiple fshift output buffers we need to reduce them */
1854         if (stepWork.computeVirial)
1855         {
1856             /* This is not in a subcounter because it takes a
1857                negligible and constant-sized amount of time */
1858             nbnxn_atomdata_add_nbat_fshift_to_fshift(
1859                     *nbv->nbat, forceOutNonbonded->forceWithShiftForces().shiftForces());
1860         }
1861     }
1862
1863     // TODO Force flags should include haveFreeEnergyWork for this domain
1864     if (stepWork.useGpuXHalo && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1865     {
1866         wallcycle_stop(wcycle, WallCycleCounter::Force);
1867         /* Wait for non-local coordinate data to be copied from device */
1868         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1869         wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
1870     }
1871
1872     // Compute wall interactions, when present.
1873     // Note: should be moved to special forces.
1874     if (inputrec.nwall && stepWork.computeNonbondedForces)
1875     {
1876         /* foreign lambda component for walls */
1877         real dvdl_walls = do_walls(inputrec,
1878                                    *fr,
1879                                    box,
1880                                    mdatoms->typeA ? gmx::arrayRefFromArray(mdatoms->typeA, mdatoms->nr)
1881                                                   : gmx::ArrayRef<int>{},
1882                                    mdatoms->typeB ? gmx::arrayRefFromArray(mdatoms->typeB, mdatoms->nr)
1883                                                   : gmx::ArrayRef<int>{},
1884                                    mdatoms->cENER ? gmx::arrayRefFromArray(mdatoms->cENER, mdatoms->nr)
1885                                                   : gmx::ArrayRef<unsigned short>{},
1886                                    mdatoms->homenr,
1887                                    mdatoms->nPerturbed,
1888                                    x.unpaddedConstArrayRef(),
1889                                    &forceOutMtsLevel0.forceWithVirial(),
1890                                    lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)],
1891                                    enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR],
1892                                    nrnb);
1893         enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] += dvdl_walls;
1894     }
1895
1896     if (stepWork.computeListedForces)
1897     {
1898         /* Check whether we need to take into account PBC in listed interactions */
1899         bool needMolPbc = false;
1900         for (const auto& listedForces : fr->listedForces)
1901         {
1902             if (listedForces.haveCpuListedForces(*fr->fcdata))
1903             {
1904                 needMolPbc = fr->bMolPBC;
1905             }
1906         }
1907
1908         t_pbc pbc;
1909
1910         if (needMolPbc)
1911         {
1912             /* Since all atoms are in the rectangular or triclinic unit-cell,
1913              * only single box vector shifts (2 in x) are required.
1914              */
1915             set_pbc_dd(&pbc, fr->pbcType, DOMAINDECOMP(cr) ? cr->dd->numCells : nullptr, TRUE, box);
1916         }
1917
1918         for (int mtsIndex = 0; mtsIndex < (simulationWork.useMts && stepWork.computeSlowForces ? 2 : 1);
1919              mtsIndex++)
1920         {
1921             ListedForces& listedForces = fr->listedForces[mtsIndex];
1922             ForceOutputs& forceOut     = (mtsIndex == 0 ? forceOutMtsLevel0 : *forceOutMtsLevel1);
1923             listedForces.calculate(wcycle,
1924                                    box,
1925                                    inputrec.fepvals.get(),
1926                                    cr,
1927                                    ms,
1928                                    x,
1929                                    xWholeMolecules,
1930                                    fr->fcdata.get(),
1931                                    hist,
1932                                    &forceOut,
1933                                    fr,
1934                                    &pbc,
1935                                    enerd,
1936                                    nrnb,
1937                                    lambda,
1938                                    mdatoms,
1939                                    DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr,
1940                                    stepWork);
1941         }
1942     }
1943
1944     if (stepWork.computeSlowForces)
1945     {
1946         calculateLongRangeNonbondeds(fr,
1947                                      inputrec,
1948                                      cr,
1949                                      nrnb,
1950                                      wcycle,
1951                                      mdatoms,
1952                                      x.unpaddedConstArrayRef(),
1953                                      &forceOutMtsLevel1->forceWithVirial(),
1954                                      enerd,
1955                                      box,
1956                                      lambda,
1957                                      dipoleData.muStateAB,
1958                                      stepWork,
1959                                      ddBalanceRegionHandler);
1960     }
1961
1962     wallcycle_stop(wcycle, WallCycleCounter::Force);
1963
1964     // VdW dispersion correction, only computed on master rank to avoid double counting
1965     if ((stepWork.computeEnergy || stepWork.computeVirial) && fr->dispersionCorrection && MASTER(cr))
1966     {
1967         // Calculate long range corrections to pressure and energy
1968         const DispersionCorrection::Correction correction = fr->dispersionCorrection->calculate(
1969                 box, lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)]);
1970
1971         if (stepWork.computeEnergy)
1972         {
1973             enerd->term[F_DISPCORR] = correction.energy;
1974             enerd->term[F_DVDL_VDW] += correction.dvdl;
1975             enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] += correction.dvdl;
1976         }
1977         if (stepWork.computeVirial)
1978         {
1979             correction.correctVirial(vir_force);
1980             enerd->term[F_PDISPCORR] = correction.pressure;
1981         }
1982     }
1983
1984     computeSpecialForces(fplog,
1985                          cr,
1986                          inputrec,
1987                          awh,
1988                          enforcedRotation,
1989                          imdSession,
1990                          pull_work,
1991                          step,
1992                          t,
1993                          wcycle,
1994                          fr->forceProviders,
1995                          box,
1996                          x.unpaddedArrayRef(),
1997                          mdatoms,
1998                          lambda,
1999                          stepWork,
2000                          &forceOutMtsLevel0.forceWithVirial(),
2001                          forceOutMtsLevel1 ? &forceOutMtsLevel1->forceWithVirial() : nullptr,
2002                          enerd,
2003                          ed,
2004                          stepWork.doNeighborSearch);
2005
2006     if (havePPDomainDecomposition(cr) && stepWork.computeForces && stepWork.useGpuFHalo
2007         && domainWork.haveCpuLocalForceWork)
2008     {
2009         stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(), AtomLocality::Local);
2010     }
2011
2012     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
2013                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
2014     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFHalo),
2015                "The schedule below does not allow for nonbonded MTS with GPU halo exchange");
2016     // Will store the amount of cycles spent waiting for the GPU that
2017     // will be later used in the DLB accounting.
2018     float cycles_wait_gpu = 0;
2019     if (useOrEmulateGpuNb && stepWork.computeNonbondedForces)
2020     {
2021         auto& forceWithShiftForces = forceOutNonbonded->forceWithShiftForces();
2022
2023         /* wait for non-local forces (or calculate in emulation mode) */
2024         if (havePPDomainDecomposition(cr))
2025         {
2026             if (simulationWork.useGpuNonbonded)
2027             {
2028                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
2029                         nbv->gpu_nbv,
2030                         stepWork,
2031                         AtomLocality::NonLocal,
2032                         enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR].data(),
2033                         enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR].data(),
2034                         forceWithShiftForces.shiftForces(),
2035                         wcycle);
2036             }
2037             else
2038             {
2039                 wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
2040                 do_nb_verlet(
2041                         fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes, step, nrnb, wcycle);
2042                 wallcycle_stop(wcycle, WallCycleCounter::Force);
2043             }
2044
2045             if (stepWork.useGpuFBufferOps)
2046             {
2047                 if (domainWork.haveNonLocalForceContribInCpuBuffer)
2048                 {
2049                     stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
2050                                               AtomLocality::NonLocal);
2051                 }
2052
2053
2054                 fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->execute();
2055
2056                 if (!stepWork.useGpuFHalo)
2057                 {
2058                     // copy from GPU input for dd_move_f()
2059                     stateGpu->copyForcesFromGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
2060                                                 AtomLocality::NonLocal);
2061                 }
2062             }
2063             else
2064             {
2065                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
2066             }
2067
2068             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
2069             {
2070                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
2071             }
2072         }
2073     }
2074
2075     /* Combining the forces for multiple time stepping before the halo exchange, when possible,
2076      * avoids an extra halo exchange (when DD is used) and post-processing step.
2077      */
2078     if (stepWork.combineMtsForcesBeforeHaloExchange)
2079     {
2080         combineMtsForces(getLocalAtomCount(cr->dd, *mdatoms, havePPDomainDecomposition(cr)),
2081                          force.unpaddedArrayRef(),
2082                          forceView->forceMtsCombined(),
2083                          inputrec.mtsLevels[1].stepFactor);
2084     }
2085
2086     if (havePPDomainDecomposition(cr))
2087     {
2088         /* We are done with the CPU compute.
2089          * We will now communicate the non-local forces.
2090          * If we use a GPU this will overlap with GPU work, so in that case
2091          * we do not close the DD force balancing region here.
2092          */
2093         ddBalanceRegionHandler.closeAfterForceComputationCpu();
2094
2095         if (stepWork.computeForces)
2096         {
2097
2098             if (stepWork.useGpuFHalo)
2099             {
2100                 // If there exist CPU forces, data from halo exchange should accumulate into these
2101                 bool accumulateForces = domainWork.haveCpuLocalForceWork;
2102                 if (!accumulateForces)
2103                 {
2104                     // Force halo exchange will set a subset of local atoms with remote non-local data
2105                     // First clear local portion of force array, so that untouched atoms are zero
2106                     stateGpu->clearForcesOnGpu(AtomLocality::Local);
2107                 }
2108                 communicateGpuHaloForces(*cr, accumulateForces);
2109             }
2110             else
2111             {
2112                 if (stepWork.useGpuFBufferOps)
2113                 {
2114                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
2115                 }
2116
2117                 // Without MTS or with MTS at slow steps with uncombined forces we need to
2118                 // communicate the fast forces
2119                 if (!simulationWork.useMts || !stepWork.combineMtsForcesBeforeHaloExchange)
2120                 {
2121                     dd_move_f(cr->dd, &forceOutMtsLevel0.forceWithShiftForces(), wcycle);
2122                 }
2123                 // With MTS we need to communicate the slow or combined (in forceOutMtsLevel1) forces
2124                 if (simulationWork.useMts && stepWork.computeSlowForces)
2125                 {
2126                     dd_move_f(cr->dd, &forceOutMtsLevel1->forceWithShiftForces(), wcycle);
2127                 }
2128             }
2129         }
2130     }
2131
2132     // With both nonbonded and PME offloaded a GPU on the same rank, we use
2133     // an alternating wait/reduction scheme.
2134     bool alternateGpuWait =
2135             (!c_disableAlternatingWait && stepWork.haveGpuPmeOnThisRank
2136              && simulationWork.useGpuNonbonded && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
2137     if (alternateGpuWait)
2138     {
2139         alternatePmeNbGpuWaitReduce(fr->nbv.get(),
2140                                     fr->pmedata,
2141                                     forceOutNonbonded,
2142                                     forceOutMtsLevel1,
2143                                     enerd,
2144                                     lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
2145                                     stepWork,
2146                                     wcycle);
2147     }
2148
2149     if (!alternateGpuWait && stepWork.haveGpuPmeOnThisRank)
2150     {
2151         pme_gpu_wait_and_reduce(fr->pmedata,
2152                                 stepWork,
2153                                 wcycle,
2154                                 &forceOutMtsLevel1->forceWithVirial(),
2155                                 enerd,
2156                                 lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)]);
2157     }
2158
2159     /* Wait for local GPU NB outputs on the non-alternating wait path */
2160     if (!alternateGpuWait && stepWork.computeNonbondedForces && simulationWork.useGpuNonbonded)
2161     {
2162         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
2163          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
2164          * but even with a step of 0.1 ms the difference is less than 1%
2165          * of the step time.
2166          */
2167         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
2168         const float waitCycles               = Nbnxm::gpu_wait_finish_task(
2169                 nbv->gpu_nbv,
2170                 stepWork,
2171                 AtomLocality::Local,
2172                 enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR].data(),
2173                 enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR].data(),
2174                 forceOutNonbonded->forceWithShiftForces().shiftForces(),
2175                 wcycle);
2176
2177         if (ddBalanceRegionHandler.useBalancingRegion())
2178         {
2179             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
2180             if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
2181             {
2182                 /* We measured few cycles, it could be that the kernel
2183                  * and transfer finished earlier and there was no actual
2184                  * wait time, only API call overhead.
2185                  * Then the actual time could be anywhere between 0 and
2186                  * cycles_wait_est. We will use half of cycles_wait_est.
2187                  */
2188                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
2189             }
2190             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
2191         }
2192     }
2193
2194     if (fr->nbv->emulateGpu())
2195     {
2196         // NOTE: emulation kernel is not included in the balancing region,
2197         // but emulation mode does not target performance anyway
2198         wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
2199         do_nb_verlet(fr,
2200                      ic,
2201                      enerd,
2202                      stepWork,
2203                      InteractionLocality::Local,
2204                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
2205                      step,
2206                      nrnb,
2207                      wcycle);
2208         wallcycle_stop(wcycle, WallCycleCounter::Force);
2209     }
2210
2211     // If on GPU PME-PP comms path, receive forces from PME before GPU buffer ops
2212     // TODO refactor this and unify with below default-path call to the same function
2213     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && stepWork.computeSlowForces
2214         && simulationWork.useGpuPmePpCommunication)
2215     {
2216         /* In case of node-splitting, the PP nodes receive the long-range
2217          * forces, virial and energy from the PME nodes here.
2218          */
2219         pme_receive_force_ener(fr,
2220                                cr,
2221                                &forceOutMtsLevel1->forceWithVirial(),
2222                                enerd,
2223                                simulationWork.useGpuPmePpCommunication,
2224                                stepWork.useGpuPmeFReduction,
2225                                wcycle);
2226     }
2227
2228
2229     /* Do the nonbonded GPU (or emulation) force buffer reduction
2230      * on the non-alternating path. */
2231     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
2232                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
2233     if (useOrEmulateGpuNb && !alternateGpuWait)
2234     {
2235         if (stepWork.useGpuFBufferOps)
2236         {
2237             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2238
2239             // TODO: move these steps as early as possible:
2240             // - CPU f H2D should be as soon as all CPU-side forces are done
2241             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
2242             //   before the next CPU task that consumes the forces: vsite spread or update)
2243             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
2244             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
2245             //   These should be unified.
2246             if (domainWork.haveLocalForceContribInCpuBuffer && !stepWork.useGpuFHalo)
2247             {
2248                 // Note: AtomLocality::All is used for the non-DD case because, as in this
2249                 // case copyForcesToGpu() uses a separate stream, it allows overlap of
2250                 // CPU force H2D with GPU force tasks on all streams including those in the
2251                 // local stream which would otherwise be implicit dependencies for the
2252                 // transfer and would not overlap.
2253                 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
2254
2255                 stateGpu->copyForcesToGpu(forceWithShift, locality);
2256             }
2257
2258             if (stepWork.computeNonbondedForces)
2259             {
2260                 fr->gpuForceReduction[gmx::AtomLocality::Local]->execute();
2261             }
2262
2263             // Copy forces to host if they are needed for update or if virtual sites are enabled.
2264             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
2265             // TODO: When the output flags will be included in step workload, this copy can be combined with the
2266             //       copy call done in sim_utils(...) for the output.
2267             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
2268             //       they should not be copied in do_md(...) for the output.
2269             if (!simulationWork.useGpuUpdate
2270                 || (simulationWork.useGpuUpdate && DOMAINDECOMP(cr) && haveHostPmePpComms) || vsite)
2271             {
2272                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
2273                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
2274             }
2275         }
2276         else if (stepWork.computeNonbondedForces)
2277         {
2278             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2279             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
2280         }
2281     }
2282
2283     launchGpuEndOfStepTasks(
2284             nbv, fr->listedForcesGpu.get(), fr->pmedata, enerd, *runScheduleWork, step, wcycle);
2285
2286     if (DOMAINDECOMP(cr))
2287     {
2288         dd_force_flop_stop(cr->dd, nrnb);
2289     }
2290
2291     const bool haveCombinedMtsForces = (stepWork.computeForces && simulationWork.useMts && stepWork.computeSlowForces
2292                                         && stepWork.combineMtsForcesBeforeHaloExchange);
2293     if (stepWork.computeForces)
2294     {
2295         postProcessForceWithShiftForces(
2296                 nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutMtsLevel0, vir_force, *mdatoms, *fr, vsite, stepWork);
2297
2298         if (simulationWork.useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2299         {
2300             postProcessForceWithShiftForces(
2301                     nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1, vir_force, *mdatoms, *fr, vsite, stepWork);
2302         }
2303     }
2304
2305     // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
2306     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
2307         && stepWork.computeSlowForces)
2308     {
2309         /* In case of node-splitting, the PP nodes receive the long-range
2310          * forces, virial and energy from the PME nodes here.
2311          */
2312         pme_receive_force_ener(fr,
2313                                cr,
2314                                &forceOutMtsLevel1->forceWithVirial(),
2315                                enerd,
2316                                simulationWork.useGpuPmePpCommunication,
2317                                false,
2318                                wcycle);
2319     }
2320
2321     if (stepWork.computeForces)
2322     {
2323         /* If we don't use MTS or if we already combined the MTS forces before, we only
2324          * need to post-process one ForceOutputs object here, called forceOutCombined,
2325          * otherwise we have to post-process two outputs and then combine them.
2326          */
2327         ForceOutputs& forceOutCombined = (haveCombinedMtsForces ? forceOutMts.value() : forceOutMtsLevel0);
2328         postProcessForces(
2329                 cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutCombined, vir_force, mdatoms, fr, vsite, stepWork);
2330
2331         if (simulationWork.useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2332         {
2333             postProcessForces(
2334                     cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1, vir_force, mdatoms, fr, vsite, stepWork);
2335
2336             combineMtsForces(mdatoms->homenr,
2337                              force.unpaddedArrayRef(),
2338                              forceView->forceMtsCombined(),
2339                              inputrec.mtsLevels[1].stepFactor);
2340         }
2341     }
2342
2343     if (stepWork.computeEnergy)
2344     {
2345         /* Compute the final potential energy terms */
2346         accumulatePotentialEnergies(enerd, lambda, inputrec.fepvals.get());
2347
2348         if (!EI_TPI(inputrec.eI))
2349         {
2350             checkPotentialEnergyValidity(step, *enerd, inputrec);
2351         }
2352     }
2353
2354     /* In case we don't have constraints and are using GPUs, the next balancing
2355      * region starts here.
2356      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
2357      * virial calculation and COM pulling, is not thus not included in
2358      * the balance timing, which is ok as most tasks do communication.
2359      */
2360     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
2361 }