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                                                           const SimulationWorkload& simulationWork,
 909                                                           const StepWorkload&       stepWork)
 910 {
 911     DomainLifetimeWorkload domainWork;
 912     // Note that haveSpecialForces is constant over the whole run
 913     domainWork.haveSpecialForces =
 914             haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 915     domainWork.haveCpuListedForceWork = false;
 916     domainWork.haveCpuBondedWork      = false;
 917     for (const auto& listedForces : fr.listedForces)
 918     {
 919         if (listedForces.haveCpuListedForces(*fr.fcdata))
 920         {
 921             domainWork.haveCpuListedForceWork = true;
 922         }
 923         if (listedForces.haveCpuBondeds())
 924         {
 925             domainWork.haveCpuBondedWork = true;
 926         }
 927     }
 928     domainWork.haveGpuBondedWork =
 929             ((fr.listedForcesGpu != nullptr) && fr.listedForcesGpu->haveInteractions());
 930     // Note that haveFreeEnergyWork is constant over the whole run
 931     domainWork.haveFreeEnergyWork =
 932             (fr.efep != FreeEnergyPerturbationType::No && mdatoms.nPerturbed != 0);
 933     // We assume we have local force work if there are CPU
 934     // force tasks including PME or nonbondeds.
 935     domainWork.haveCpuLocalForceWork =
 936             domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
 937             || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
 938             || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
 939     domainWork.haveLocalForceContribInCpuBuffer =
 940             domainWork.haveCpuLocalForceWork || simulationWork.havePpDomainDecomposition;
 941     domainWork.haveNonLocalForceContribInCpuBuffer =
 942             domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
 943
 944     return domainWork;
 945 }
 946
 947 /*! \brief Set up force flag stuct from the force bitmask.
 948  *
 949  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 950  * \param[in]      mtsLevels            The multiple time-stepping levels, either empty or 2 levels
 951  * \param[in]      step                 The current MD step
 952  * \param[in]      simulationWork       Simulation workload description.
 953  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 954  *
 955  * \returns New Stepworkload description.
 956  */
 957 static StepWorkload setupStepWorkload(const int                     legacyFlags,
 958                                       ArrayRef<const gmx::MtsLevel> mtsLevels,
 959                                       const int64_t                 step,
 960                                       const SimulationWorkload&     simulationWork,
 961                                       const bool                    rankHasPmeDuty)
 962 {
 963     GMX_ASSERT(mtsLevels.empty() || mtsLevels.size() == 2, "Expect 0 or 2 MTS levels");
 964     const bool computeSlowForces = (mtsLevels.empty() || step % mtsLevels[1].stepFactor == 0);
 965
 966     StepWorkload flags;
 967     flags.stateChanged                  = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 968     flags.haveDynamicBox                = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 969     flags.doNeighborSearch              = ((legacyFlags & GMX_FORCE_NS) != 0);
 970     flags.computeSlowForces             = computeSlowForces;
 971     flags.computeVirial                 = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 972     flags.computeEnergy                 = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 973     flags.computeForces                 = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 974     flags.useOnlyMtsCombinedForceBuffer = ((legacyFlags & GMX_FORCE_DO_NOT_NEED_NORMAL_FORCE) != 0);
 975     flags.computeListedForces           = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 976     flags.computeNonbondedForces =
 977             ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && simulationWork.computeNonbonded
 978             && !(simulationWork.computeNonbondedAtMtsLevel1 && !computeSlowForces);
 979     flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 980
 981     if (simulationWork.useGpuBufferOps)
 982     {
 983         GMX_ASSERT(simulationWork.useGpuNonbonded,
 984                    "Can only offload buffer ops if nonbonded computation is also offloaded");
 985     }
 986     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 987     // on virial steps the CPU reduction path is taken
 988     flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !flags.computeVirial;
 989     flags.useGpuPmeFReduction = flags.computeSlowForces && flags.useGpuFBufferOps && simulationWork.useGpuPme
 990                                 && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication);
 991     flags.useGpuXHalo = simulationWork.useGpuHaloExchange;
 992     flags.useGpuFHalo = simulationWork.useGpuHaloExchange && flags.useGpuFBufferOps;
 993     // Note that rankHasPmeDuty is used confusingly due to the way cr->duty is set up (can be true even for non-PME runs),
 994     // but the haveGpuPmeOnThisRank still ends up correct as simulationWork.useGpuPme == false in such cases.
 995     // TODO: improve this when duty-reliance is eliminated
 996     flags.haveGpuPmeOnThisRank = simulationWork.useGpuPme && rankHasPmeDuty && flags.computeSlowForces;
 997     flags.combineMtsForcesBeforeHaloExchange =
 998             (flags.computeForces && simulationWork.useMts && flags.computeSlowForces
 999              && flags.useOnlyMtsCombinedForceBuffer
1000              && !(flags.computeVirial || simulationWork.useGpuNonbonded || flags.haveGpuPmeOnThisRank));
1001
1002     return flags;
1003 }
1004
1005
1006 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
1007  *
1008  */
1009 static void launchGpuEndOfStepTasks(nonbonded_verlet_t*               nbv,
1010                                     gmx::ListedForcesGpu*             listedForcesGpu,
1011                                     gmx_pme_t*                        pmedata,
1012                                     gmx_enerdata_t*                   enerd,
1013                                     const gmx::MdrunScheduleWorkload& runScheduleWork,
1014                                     int64_t                           step,
1015                                     gmx_wallcycle*                    wcycle)
1016 {
1017     if (runScheduleWork.simulationWork.useGpuNonbonded && runScheduleWork.stepWork.computeNonbondedForces)
1018     {
1019         /* Launch pruning before buffer clearing because the API overhead of the
1020          * clear kernel launches can leave the GPU idle while it could be running
1021          * the prune kernel.
1022          */
1023         if (nbv->isDynamicPruningStepGpu(step))
1024         {
1025             nbv->dispatchPruneKernelGpu(step);
1026         }
1027
1028         /* now clear the GPU outputs while we finish the step on the CPU */
1029         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1030         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1031         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
1032         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1033         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1034     }
1035
1036     if (runScheduleWork.stepWork.haveGpuPmeOnThisRank)
1037     {
1038         pme_gpu_reinit_computation(pmedata, wcycle);
1039     }
1040
1041     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
1042     {
1043         // in principle this should be included in the DD balancing region,
1044         // but generally it is infrequent so we'll omit it for the sake of
1045         // simpler code
1046         listedForcesGpu->waitAccumulateEnergyTerms(enerd);
1047
1048         listedForcesGpu->clearEnergies();
1049     }
1050 }
1051
1052 //! \brief Data structure to hold dipole-related data and staging arrays
1053 struct DipoleData
1054 {
1055     //! Dipole staging for fast summing over MPI
1056     gmx::DVec muStaging[2] = { { 0.0, 0.0, 0.0 } };
1057     //! Dipole staging for states A and B (index 0 and 1 resp.)
1058     gmx::RVec muStateAB[2] = { { 0.0_real, 0.0_real, 0.0_real } };
1059 };
1060
1061
1062 static void reduceAndUpdateMuTot(DipoleData*                   dipoleData,
1063                                  const t_commrec*              cr,
1064                                  const bool                    haveFreeEnergy,
1065                                  gmx::ArrayRef<const real>     lambda,
1066                                  rvec                          muTotal,
1067                                  const DDBalanceRegionHandler& ddBalanceRegionHandler)
1068 {
1069     if (PAR(cr))
1070     {
1071         gmx_sumd(2 * DIM, dipoleData->muStaging[0], cr);
1072         ddBalanceRegionHandler.reopenRegionCpu();
1073     }
1074     for (int i = 0; i < 2; i++)
1075     {
1076         for (int j = 0; j < DIM; j++)
1077         {
1078             dipoleData->muStateAB[i][j] = dipoleData->muStaging[i][j];
1079         }
1080     }
1081
1082     if (!haveFreeEnergy)
1083     {
1084         copy_rvec(dipoleData->muStateAB[0], muTotal);
1085     }
1086     else
1087     {
1088         for (int j = 0; j < DIM; j++)
1089         {
1090             muTotal[j] = (1.0 - lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)])
1091                                  * dipoleData->muStateAB[0][j]
1092                          + lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)]
1093                                    * dipoleData->muStateAB[1][j];
1094         }
1095     }
1096 }
1097
1098 /*! \brief Combines MTS level0 and level1 force buffers into a full and MTS-combined force buffer.
1099  *
1100  * \param[in]     numAtoms        The number of atoms to combine forces for
1101  * \param[in,out] forceMtsLevel0  Input: F_level0, output: F_level0 + F_level1
1102  * \param[in,out] forceMts        Input: F_level1, output: F_level0 + mtsFactor * F_level1
1103  * \param[in]     mtsFactor       The factor between the level0 and level1 time step
1104  */
1105 static void combineMtsForces(const int      numAtoms,
1106                              ArrayRef<RVec> forceMtsLevel0,
1107                              ArrayRef<RVec> forceMts,
1108                              const real     mtsFactor)
1109 {
1110     const int gmx_unused numThreads = gmx_omp_nthreads_get(ModuleMultiThread::Default);
1111 #pragma omp parallel for num_threads(numThreads) schedule(static)
1112     for (int i = 0; i < numAtoms; i++)
1113     {
1114         const RVec forceMtsLevel0Tmp = forceMtsLevel0[i];
1115         forceMtsLevel0[i] += forceMts[i];
1116         forceMts[i] = forceMtsLevel0Tmp + mtsFactor * forceMts[i];
1117     }
1118 }
1119
1120 /*! \brief Setup for the local and non-local GPU force reductions:
1121  * reinitialization plus the registration of forces and dependencies.
1122  *
1123  * \param [in] runScheduleWork               Schedule workload flag structure
1124  * \param [in] cr                            Communication record object
1125  * \param [in] fr                            Force record object
1126  */
1127 static void setupGpuForceReductions(gmx::MdrunScheduleWorkload* runScheduleWork,
1128                                     const t_commrec*            cr,
1129                                     t_forcerec*                 fr)
1130 {
1131
1132     nonbonded_verlet_t*          nbv      = fr->nbv.get();
1133     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1134
1135     // (re-)initialize local GPU force reduction
1136     const bool accumulate = runScheduleWork->domainWork.haveCpuLocalForceWork
1137                             || runScheduleWork->simulationWork.havePpDomainDecomposition;
1138     const int atomStart = 0;
1139     fr->gpuForceReduction[gmx::AtomLocality::Local]->reinit(stateGpu->getForces(),
1140                                                             nbv->getNumAtoms(AtomLocality::Local),
1141                                                             nbv->getGridIndices(),
1142                                                             atomStart,
1143                                                             accumulate,
1144                                                             stateGpu->fReducedOnDevice());
1145
1146     // register forces and add dependencies
1147     fr->gpuForceReduction[gmx::AtomLocality::Local]->registerNbnxmForce(Nbnxm::gpu_get_f(nbv->gpu_nbv));
1148
1149     if (runScheduleWork->simulationWork.useGpuPme
1150         && (!runScheduleWork->simulationWork.haveSeparatePmeRank
1151             || runScheduleWork->simulationWork.useGpuPmePpCommunication))
1152     {
1153         DeviceBuffer<gmx::RVec> forcePtr =
1154                 runScheduleWork->simulationWork.haveSeparatePmeRank
1155                         ? fr->pmePpCommGpu->getGpuForceStagingPtr() // buffer received from other GPU
1156                         : pme_gpu_get_device_f(fr->pmedata);        // PME force buffer on same GPU
1157         fr->gpuForceReduction[gmx::AtomLocality::Local]->registerRvecForce(forcePtr);
1158
1159         GpuEventSynchronizer* const pmeSynchronizer =
1160                 (runScheduleWork->simulationWork.haveSeparatePmeRank
1161                          ? fr->pmePpCommGpu->getForcesReadySynchronizer() // buffer received from other GPU
1162                          : pme_gpu_get_f_ready_synchronizer(fr->pmedata)); // PME force buffer on same GPU
1163         if (GMX_THREAD_MPI)
1164         {
1165             GMX_ASSERT(pmeSynchronizer != nullptr, "PME force ready cuda event should not be NULL");
1166             fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(pmeSynchronizer);
1167         }
1168     }
1169
1170     if (runScheduleWork->domainWork.haveCpuLocalForceWork && !runScheduleWork->simulationWork.useGpuHaloExchange)
1171     {
1172         // in the DD case we use the same stream for H2D and reduction, hence no explicit dependency needed
1173         if (!runScheduleWork->simulationWork.havePpDomainDecomposition)
1174         {
1175             const bool useGpuForceBufferOps = true;
1176             fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1177                     stateGpu->getForcesReadyOnDeviceEvent(AtomLocality::All, useGpuForceBufferOps));
1178         }
1179     }
1180
1181     if (runScheduleWork->simulationWork.useGpuHaloExchange)
1182     {
1183         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1184                 cr->dd->gpuHaloExchange[0][0]->getForcesReadyOnDeviceEvent());
1185     }
1186
1187     if (runScheduleWork->simulationWork.havePpDomainDecomposition)
1188     {
1189         // (re-)initialize non-local GPU force reduction
1190         const bool accumulate = runScheduleWork->domainWork.haveCpuBondedWork
1191                                 || runScheduleWork->domainWork.haveFreeEnergyWork;
1192         const int atomStart = dd_numHomeAtoms(*cr->dd);
1193         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->reinit(stateGpu->getForces(),
1194                                                                    nbv->getNumAtoms(AtomLocality::NonLocal),
1195                                                                    nbv->getGridIndices(),
1196                                                                    atomStart,
1197                                                                    accumulate);
1198
1199         // register forces and add dependencies
1200         // in the DD case we use the same stream for H2D and reduction, hence no explicit dependency needed
1201         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->registerNbnxmForce(
1202                 Nbnxm::gpu_get_f(nbv->gpu_nbv));
1203     }
1204 }
1205
1206
1207 /*! \brief Return the number of local atoms.
1208  */
1209 static int getLocalAtomCount(const gmx_domdec_t* dd, const t_mdatoms& mdatoms, bool havePPDomainDecomposition)
1210 {
1211     GMX_ASSERT(!(havePPDomainDecomposition && (dd == nullptr)),
1212                "Can't have PP decomposition with dd uninitialized!");
1213     return havePPDomainDecomposition ? dd_numAtomsZones(*dd) : mdatoms.homenr;
1214 }
1215
1216
1217 void do_force(FILE*                               fplog,
1218               const t_commrec*                    cr,
1219               const gmx_multisim_t*               ms,
1220               const t_inputrec&                   inputrec,
1221               gmx::Awh*                           awh,
1222               gmx_enfrot*                         enforcedRotation,
1223               gmx::ImdSession*                    imdSession,
1224               pull_t*                             pull_work,
1225               int64_t                             step,
1226               t_nrnb*                             nrnb,
1227               gmx_wallcycle*                      wcycle,
1228               const gmx_localtop_t*               top,
1229               const matrix                        box,
1230               gmx::ArrayRefWithPadding<gmx::RVec> x,
1231               const history_t*                    hist,
1232               gmx::ForceBuffersView*              forceView,
1233               tensor                              vir_force,
1234               const t_mdatoms*                    mdatoms,
1235               gmx_enerdata_t*                     enerd,
1236               gmx::ArrayRef<const real>           lambda,
1237               t_forcerec*                         fr,
1238               gmx::MdrunScheduleWorkload*         runScheduleWork,
1239               gmx::VirtualSitesHandler*           vsite,
1240               rvec                                muTotal,
1241               double                              t,
1242               gmx_edsam*                          ed,
1243               int                                 legacyFlags,
1244               const DDBalanceRegionHandler&       ddBalanceRegionHandler)
1245 {
1246     auto force = forceView->forceWithPadding();
1247     GMX_ASSERT(force.unpaddedArrayRef().ssize() >= fr->natoms_force_constr,
1248                "The size of the force buffer should be at least the number of atoms to compute "
1249                "forces for");
1250
1251     nonbonded_verlet_t*  nbv = fr->nbv.get();
1252     interaction_const_t* ic  = fr->ic.get();
1253
1254     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1255
1256     const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
1257
1258     runScheduleWork->stepWork = setupStepWorkload(
1259             legacyFlags, inputrec.mtsLevels, step, simulationWork, thisRankHasDuty(cr, DUTY_PME));
1260     const StepWorkload& stepWork = runScheduleWork->stepWork;
1261
1262     /* At a search step we need to start the first balancing region
1263      * somewhere early inside the step after communication during domain
1264      * decomposition (and not during the previous step as usual).
1265      */
1266     if (stepWork.doNeighborSearch)
1267     {
1268         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
1269     }
1270
1271     clear_mat(vir_force);
1272
1273     if (fr->pbcType != PbcType::No)
1274     {
1275         /* Compute shift vectors every step,
1276          * because of pressure coupling or box deformation!
1277          */
1278         if (stepWork.haveDynamicBox && stepWork.stateChanged)
1279         {
1280             calc_shifts(box, fr->shift_vec);
1281         }
1282
1283         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1284         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
1285         if (calcCGCM)
1286         {
1287             put_atoms_in_box_omp(fr->pbcType,
1288                                  box,
1289                                  x.unpaddedArrayRef().subArray(0, mdatoms->homenr),
1290                                  gmx_omp_nthreads_get(ModuleMultiThread::Default));
1291             inc_nrnb(nrnb, eNR_SHIFTX, mdatoms->homenr);
1292         }
1293     }
1294
1295     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1296
1297     const bool pmeSendCoordinatesFromGpu =
1298             simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1299     const bool reinitGpuPmePpComms =
1300             simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1301
1302     auto* localXReadyOnDevice = (stepWork.haveGpuPmeOnThisRank || simulationWork.useGpuBufferOps)
1303                                         ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1304                                                 AtomLocality::Local, simulationWork, stepWork)
1305                                         : nullptr;
1306
1307     GMX_ASSERT(simulationWork.useGpuHaloExchange
1308                        == ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange[0].empty())),
1309                "The GPU halo exchange is active, but it has not been constructed.");
1310
1311     bool gmx_used_in_debug haveCopiedXFromGpu = false;
1312     // Copy coordinate from the GPU if update is on the GPU and there
1313     // are forces to be computed on the CPU, or for the computation of
1314     // virial, or if host-side data will be transferred from this task
1315     // to a remote task for halo exchange or PME-PP communication. At
1316     // search steps the current coordinates are already on the host,
1317     // hence copy is not needed.
1318     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1319         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1320             || simulationWork.useCpuPmePpCommunication || simulationWork.useCpuHaloExchange
1321             || simulationWork.computeMuTot))
1322     {
1323         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1324         haveCopiedXFromGpu = true;
1325     }
1326
1327     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1328     // The local coordinates can be copied right away.
1329     // NOTE: Consider moving this copy to right after they are updated and constrained,
1330     //       if the later is not offloaded.
1331     if (stepWork.haveGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1332     {
1333         if (stepWork.doNeighborSearch)
1334         {
1335             // TODO refactor this to do_md, after partitioning.
1336             stateGpu->reinit(mdatoms->homenr,
1337                              getLocalAtomCount(cr->dd, *mdatoms, simulationWork.havePpDomainDecomposition));
1338             if (stepWork.haveGpuPmeOnThisRank)
1339             {
1340                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1341                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1342             }
1343         }
1344         // We need to copy coordinates when:
1345         // 1. Update is not offloaded
1346         // 2. The buffers were reinitialized on search step
1347         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1348         {
1349             GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1350             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1351         }
1352     }
1353
1354     if (simulationWork.haveSeparatePmeRank && stepWork.computeSlowForces)
1355     {
1356         /* Send particle coordinates to the pme nodes */
1357         if (!pmeSendCoordinatesFromGpu && !stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1358         {
1359             GMX_ASSERT(haveCopiedXFromGpu,
1360                        "a wait should only be triggered if copy has been scheduled");
1361             stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1362         }
1363
1364         gmx_pme_send_coordinates(fr,
1365                                  cr,
1366                                  box,
1367                                  x.unpaddedArrayRef(),
1368                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1369                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)],
1370                                  (stepWork.computeVirial || stepWork.computeEnergy),
1371                                  step,
1372                                  simulationWork.useGpuPmePpCommunication,
1373                                  reinitGpuPmePpComms,
1374                                  pmeSendCoordinatesFromGpu,
1375                                  stepWork.useGpuPmeFReduction,
1376                                  localXReadyOnDevice,
1377                                  wcycle);
1378     }
1379
1380     if (stepWork.haveGpuPmeOnThisRank)
1381     {
1382         launchPmeGpuSpread(fr->pmedata,
1383                            box,
1384                            stepWork,
1385                            localXReadyOnDevice,
1386                            lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1387                            wcycle);
1388     }
1389
1390     const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1391
1392     /* do gridding for pair search */
1393     if (stepWork.doNeighborSearch)
1394     {
1395         if (fr->wholeMoleculeTransform && stepWork.stateChanged)
1396         {
1397             fr->wholeMoleculeTransform->updateForAtomPbcJumps(x.unpaddedArrayRef(), box);
1398         }
1399
1400         wallcycle_start(wcycle, WallCycleCounter::NS);
1401         if (!DOMAINDECOMP(cr))
1402         {
1403             const rvec vzero       = { 0.0_real, 0.0_real, 0.0_real };
1404             const rvec boxDiagonal = { box[XX][XX], box[YY][YY], box[ZZ][ZZ] };
1405             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSGridLocal);
1406             nbnxn_put_on_grid(nbv,
1407                               box,
1408                               0,
1409                               vzero,
1410                               boxDiagonal,
1411                               nullptr,
1412                               { 0, mdatoms->homenr },
1413                               -1,
1414                               fr->atomInfo,
1415                               x.unpaddedArrayRef(),
1416                               0,
1417                               nullptr);
1418             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSGridLocal);
1419         }
1420         else
1421         {
1422             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSGridNonLocal);
1423             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->atomInfo, x.unpaddedArrayRef());
1424             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSGridNonLocal);
1425         }
1426
1427         nbv->setAtomProperties(gmx::constArrayRefFromArray(mdatoms->typeA, mdatoms->nr),
1428                                gmx::constArrayRefFromArray(mdatoms->chargeA, mdatoms->nr),
1429                                fr->atomInfo);
1430
1431         wallcycle_stop(wcycle, WallCycleCounter::NS);
1432
1433         /* initialize the GPU nbnxm atom data and bonded data structures */
1434         if (simulationWork.useGpuNonbonded)
1435         {
1436             // Note: cycle counting only nononbondeds, GPU listed forces counts internally
1437             wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1438             wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1439             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1440             wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1441             wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1442
1443             if (fr->listedForcesGpu)
1444             {
1445                 /* Now we put all atoms on the grid, we can assign bonded
1446                  * interactions to the GPU, where the grid order is
1447                  * needed. Also the xq, f and fshift device buffers have
1448                  * been reallocated if needed, so the bonded code can
1449                  * learn about them. */
1450                 // TODO the xq, f, and fshift buffers are now shared
1451                 // resources, so they should be maintained by a
1452                 // higher-level object than the nb module.
1453                 fr->listedForcesGpu->updateInteractionListsAndDeviceBuffers(
1454                         nbv->getGridIndices(),
1455                         top->idef,
1456                         Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1457                         Nbnxm::gpu_get_f(nbv->gpu_nbv),
1458                         Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1459             }
1460         }
1461
1462         // Need to run after the GPU-offload bonded interaction lists
1463         // are set up to be able to determine whether there is bonded work.
1464         runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1465                 inputrec, *fr, pull_work, ed, *mdatoms, simulationWork, stepWork);
1466
1467         wallcycle_start_nocount(wcycle, WallCycleCounter::NS);
1468         wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSSearchLocal);
1469         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1470         nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1471
1472         nbv->setupGpuShortRangeWork(fr->listedForcesGpu.get(), InteractionLocality::Local);
1473
1474         wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSSearchLocal);
1475         wallcycle_stop(wcycle, WallCycleCounter::NS);
1476
1477         if (stepWork.useGpuXBufferOps)
1478         {
1479             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1480         }
1481
1482         if (simulationWork.useGpuBufferOps)
1483         {
1484             setupGpuForceReductions(runScheduleWork, cr, fr);
1485         }
1486     }
1487     else if (!EI_TPI(inputrec.eI) && stepWork.computeNonbondedForces)
1488     {
1489         if (stepWork.useGpuXBufferOps)
1490         {
1491             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1492             nbv->convertCoordinatesGpu(AtomLocality::Local, stateGpu->getCoordinates(), localXReadyOnDevice);
1493         }
1494         else
1495         {
1496             if (simulationWork.useGpuUpdate)
1497             {
1498                 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1499                 GMX_ASSERT(haveCopiedXFromGpu,
1500                            "a wait should only be triggered if copy has been scheduled");
1501                 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1502             }
1503             nbv->convertCoordinates(AtomLocality::Local, x.unpaddedArrayRef());
1504         }
1505     }
1506
1507     if (simulationWork.useGpuNonbonded && (stepWork.computeNonbondedForces || domainWork.haveGpuBondedWork))
1508     {
1509         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1510
1511         wallcycle_start(wcycle, WallCycleCounter::LaunchGpu);
1512         wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1513         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1514         if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1515         {
1516             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1517         }
1518         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1519         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1520         // with X buffer ops offloaded to the GPU on all but the search steps
1521
1522         // bonded work not split into separate local and non-local, so with DD
1523         // we can only launch the kernel after non-local coordinates have been received.
1524         if (domainWork.haveGpuBondedWork && !simulationWork.havePpDomainDecomposition)
1525         {
1526             fr->listedForcesGpu->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1527         }
1528
1529         /* launch local nonbonded work on GPU */
1530         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1531         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1532         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1533         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1534         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1535     }
1536
1537     if (stepWork.haveGpuPmeOnThisRank)
1538     {
1539         // In PME GPU and mixed mode we launch FFT / gather after the
1540         // X copy/transform to allow overlap as well as after the GPU NB
1541         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1542         // the nonbonded kernel.
1543         launchPmeGpuFftAndGather(fr->pmedata,
1544                                  lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
1545                                  wcycle,
1546                                  stepWork);
1547     }
1548
1549     /* Communicate coordinates and sum dipole if necessary +
1550        do non-local pair search */
1551     if (simulationWork.havePpDomainDecomposition)
1552     {
1553         if (stepWork.doNeighborSearch)
1554         {
1555             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1556             wallcycle_start_nocount(wcycle, WallCycleCounter::NS);
1557             wallcycle_sub_start(wcycle, WallCycleSubCounter::NBSSearchNonLocal);
1558             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1559             nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1560
1561             nbv->setupGpuShortRangeWork(fr->listedForcesGpu.get(), InteractionLocality::NonLocal);
1562             wallcycle_sub_stop(wcycle, WallCycleSubCounter::NBSSearchNonLocal);
1563             wallcycle_stop(wcycle, WallCycleCounter::NS);
1564             // TODO refactor this GPU halo exchange re-initialisation
1565             // to location in do_md where GPU halo exchange is
1566             // constructed at partitioning, after above stateGpu
1567             // re-initialization has similarly been refactored
1568             if (simulationWork.useGpuHaloExchange)
1569             {
1570                 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1571             }
1572         }
1573         else
1574         {
1575             if (stepWork.useGpuXHalo)
1576             {
1577                 // The following must be called after local setCoordinates (which records an event
1578                 // when the coordinate data has been copied to the device).
1579                 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1580
1581                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1582                 {
1583                     // non-local part of coordinate buffer must be copied back to host for CPU work
1584                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1585                 }
1586             }
1587             else
1588             {
1589                 if (simulationWork.useGpuUpdate)
1590                 {
1591                     GMX_ASSERT(haveCopiedXFromGpu,
1592                                "a wait should only be triggered if copy has been scheduled");
1593                     stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1594                 }
1595                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1596             }
1597
1598             if (stepWork.useGpuXBufferOps)
1599             {
1600                 if (!stepWork.haveGpuPmeOnThisRank && !stepWork.useGpuXHalo)
1601                 {
1602                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1603                 }
1604                 nbv->convertCoordinatesGpu(AtomLocality::NonLocal,
1605                                            stateGpu->getCoordinates(),
1606                                            stateGpu->getCoordinatesReadyOnDeviceEvent(
1607                                                    AtomLocality::NonLocal, simulationWork, stepWork));
1608             }
1609             else
1610             {
1611                 nbv->convertCoordinates(AtomLocality::NonLocal, x.unpaddedArrayRef());
1612             }
1613         }
1614
1615         if (simulationWork.useGpuNonbonded)
1616         {
1617
1618             if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1619             {
1620                 wallcycle_start(wcycle, WallCycleCounter::LaunchGpu);
1621                 wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1622                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1623                 wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1624                 wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1625             }
1626
1627             if (domainWork.haveGpuBondedWork)
1628             {
1629                 fr->listedForcesGpu->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1630             }
1631
1632             /* launch non-local nonbonded tasks on GPU */
1633             wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1634             wallcycle_sub_start(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1635             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step, nrnb, wcycle);
1636             wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1637             wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1638         }
1639     }
1640
1641     if (simulationWork.useGpuNonbonded && stepWork.computeNonbondedForces)
1642     {
1643         /* launch D2H copy-back F */
1644         wallcycle_start_nocount(wcycle, WallCycleCounter::LaunchGpu);
1645         wallcycle_sub_start_nocount(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1646
1647         if (simulationWork.havePpDomainDecomposition)
1648         {
1649             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1650         }
1651         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1652         wallcycle_sub_stop(wcycle, WallCycleSubCounter::LaunchGpuNonBonded);
1653
1654         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1655         {
1656             fr->listedForcesGpu->launchEnergyTransfer();
1657         }
1658         wallcycle_stop(wcycle, WallCycleCounter::LaunchGpu);
1659     }
1660
1661     gmx::ArrayRef<const gmx::RVec> xWholeMolecules;
1662     if (fr->wholeMoleculeTransform)
1663     {
1664         xWholeMolecules = fr->wholeMoleculeTransform->wholeMoleculeCoordinates(x.unpaddedArrayRef(), box);
1665     }
1666
1667     // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1668     // this wait ensures that the D2H transfer is complete.
1669     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
1670     {
1671         const bool needCoordsOnHost  = (runScheduleWork->domainWork.haveCpuLocalForceWork
1672                                        || stepWork.computeVirial || simulationWork.computeMuTot);
1673         const bool haveAlreadyWaited = simulationWork.useCpuHaloExchange;
1674         if (needCoordsOnHost && !haveAlreadyWaited)
1675         {
1676             GMX_ASSERT(haveCopiedXFromGpu,
1677                        "a wait should only be triggered if copy has been scheduled");
1678             stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1679         }
1680     }
1681
1682     DipoleData dipoleData;
1683
1684     if (simulationWork.computeMuTot)
1685     {
1686         const int start = 0;
1687
1688         /* Calculate total (local) dipole moment in a temporary common array.
1689          * This makes it possible to sum them over nodes faster.
1690          */
1691         gmx::ArrayRef<const gmx::RVec> xRef =
1692                 (xWholeMolecules.empty() ? x.unpaddedArrayRef() : xWholeMolecules);
1693         calc_mu(start,
1694                 mdatoms->homenr,
1695                 xRef,
1696                 mdatoms->chargeA ? gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->nr)
1697                                  : gmx::ArrayRef<real>{},
1698                 mdatoms->chargeB ? gmx::arrayRefFromArray(mdatoms->chargeB, mdatoms->nr)
1699                                  : gmx::ArrayRef<real>{},
1700                 mdatoms->nChargePerturbed != 0,
1701                 dipoleData.muStaging[0],
1702                 dipoleData.muStaging[1]);
1703
1704         reduceAndUpdateMuTot(
1705                 &dipoleData, cr, (fr->efep != FreeEnergyPerturbationType::No), lambda, muTotal, ddBalanceRegionHandler);
1706     }
1707
1708     /* Reset energies */
1709     reset_enerdata(enerd);
1710
1711     if (DOMAINDECOMP(cr) && simulationWork.haveSeparatePmeRank)
1712     {
1713         wallcycle_start(wcycle, WallCycleCounter::PpDuringPme);
1714         dd_force_flop_start(cr->dd, nrnb);
1715     }
1716
1717     if (inputrec.bRot)
1718     {
1719         wallcycle_start(wcycle, WallCycleCounter::Rot);
1720         do_rotation(cr, enforcedRotation, box, x.unpaddedConstArrayRef(), t, step, stepWork.doNeighborSearch);
1721         wallcycle_stop(wcycle, WallCycleCounter::Rot);
1722     }
1723
1724     /* Start the force cycle counter.
1725      * Note that a different counter is used for dynamic load balancing.
1726      */
1727     wallcycle_start(wcycle, WallCycleCounter::Force);
1728
1729     /* Set up and clear force outputs:
1730      * forceOutMtsLevel0:  everything except what is in the other two outputs
1731      * forceOutMtsLevel1:  PME-mesh and listed-forces group 1
1732      * forceOutNonbonded: non-bonded forces
1733      * Without multiple time stepping all point to the same object.
1734      * With multiple time-stepping the use is different for MTS fast (level0 only) and slow steps.
1735      */
1736     ForceOutputs forceOutMtsLevel0 = setupForceOutputs(
1737             &fr->forceHelperBuffers[0], force, domainWork, stepWork, simulationWork.havePpDomainDecomposition, wcycle);
1738
1739     // Force output for MTS combined forces, only set at level1 MTS steps
1740     std::optional<ForceOutputs> forceOutMts =
1741             (simulationWork.useMts && stepWork.computeSlowForces)
1742                     ? std::optional(setupForceOutputs(&fr->forceHelperBuffers[1],
1743                                                       forceView->forceMtsCombinedWithPadding(),
1744                                                       domainWork,
1745                                                       stepWork,
1746                                                       simulationWork.havePpDomainDecomposition,
1747                                                       wcycle))
1748                     : std::nullopt;
1749
1750     ForceOutputs* forceOutMtsLevel1 =
1751             simulationWork.useMts ? (stepWork.computeSlowForces ? &forceOutMts.value() : nullptr)
1752                                   : &forceOutMtsLevel0;
1753
1754     const bool nonbondedAtMtsLevel1 = runScheduleWork->simulationWork.computeNonbondedAtMtsLevel1;
1755
1756     ForceOutputs* forceOutNonbonded = nonbondedAtMtsLevel1 ? forceOutMtsLevel1 : &forceOutMtsLevel0;
1757
1758     if (inputrec.bPull && pull_have_constraint(*pull_work))
1759     {
1760         clear_pull_forces(pull_work);
1761     }
1762
1763     /* We calculate the non-bonded forces, when done on the CPU, here.
1764      * We do this before calling do_force_lowlevel, because in that
1765      * function, the listed forces are calculated before PME, which
1766      * does communication.  With this order, non-bonded and listed
1767      * force calculation imbalance can be balanced out by the domain
1768      * decomposition load balancing.
1769      */
1770
1771     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1772
1773     if (!useOrEmulateGpuNb)
1774     {
1775         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1776     }
1777
1778     if (fr->efep != FreeEnergyPerturbationType::No && stepWork.computeNonbondedForces)
1779     {
1780         /* Calculate the local and non-local free energy interactions here.
1781          * Happens here on the CPU both with and without GPU.
1782          */
1783         nbv->dispatchFreeEnergyKernel(
1784                 InteractionLocality::Local,
1785                 x.unpaddedArrayRef(),
1786                 &forceOutNonbonded->forceWithShiftForces(),
1787                 fr->use_simd_kernels,
1788                 fr->ntype,
1789                 fr->rlist,
1790                 *fr->ic,
1791                 fr->shift_vec,
1792                 fr->nbfp,
1793                 fr->ljpme_c6grid,
1794                 mdatoms->chargeA ? gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->nr)
1795                                  : gmx::ArrayRef<real>{},
1796                 mdatoms->chargeB ? gmx::arrayRefFromArray(mdatoms->chargeB, mdatoms->nr)
1797                                  : gmx::ArrayRef<real>{},
1798                 mdatoms->typeA ? gmx::arrayRefFromArray(mdatoms->typeA, mdatoms->nr)
1799                                : gmx::ArrayRef<int>{},
1800                 mdatoms->typeB ? gmx::arrayRefFromArray(mdatoms->typeB, mdatoms->nr)
1801                                : gmx::ArrayRef<int>{},
1802                 inputrec.fepvals.get(),
1803                 lambda,
1804                 enerd,
1805                 stepWork,
1806                 nrnb);
1807
1808         if (simulationWork.havePpDomainDecomposition)
1809         {
1810             nbv->dispatchFreeEnergyKernel(
1811                     InteractionLocality::NonLocal,
1812                     x.unpaddedArrayRef(),
1813                     &forceOutNonbonded->forceWithShiftForces(),
1814                     fr->use_simd_kernels,
1815                     fr->ntype,
1816                     fr->rlist,
1817                     *fr->ic,
1818                     fr->shift_vec,
1819                     fr->nbfp,
1820                     fr->ljpme_c6grid,
1821                     mdatoms->chargeA ? gmx::arrayRefFromArray(mdatoms->chargeA, mdatoms->nr)
1822                                      : gmx::ArrayRef<real>{},
1823                     mdatoms->chargeB ? gmx::arrayRefFromArray(mdatoms->chargeB, mdatoms->nr)
1824                                      : gmx::ArrayRef<real>{},
1825                     mdatoms->typeA ? gmx::arrayRefFromArray(mdatoms->typeA, mdatoms->nr)
1826                                    : gmx::ArrayRef<int>{},
1827                     mdatoms->typeB ? gmx::arrayRefFromArray(mdatoms->typeB, mdatoms->nr)
1828                                    : gmx::ArrayRef<int>{},
1829                     inputrec.fepvals.get(),
1830                     lambda,
1831                     enerd,
1832                     stepWork,
1833                     nrnb);
1834         }
1835     }
1836
1837     if (stepWork.computeNonbondedForces && !useOrEmulateGpuNb)
1838     {
1839         if (simulationWork.havePpDomainDecomposition)
1840         {
1841             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step, nrnb, wcycle);
1842         }
1843
1844         if (stepWork.computeForces)
1845         {
1846             /* Add all the non-bonded force to the normal force array.
1847              * This can be split into a local and a non-local part when overlapping
1848              * communication with calculation with domain decomposition.
1849              */
1850             wallcycle_stop(wcycle, WallCycleCounter::Force);
1851             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All,
1852                                           forceOutNonbonded->forceWithShiftForces().force());
1853             wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
1854         }
1855
1856         /* If there are multiple fshift output buffers we need to reduce them */
1857         if (stepWork.computeVirial)
1858         {
1859             /* This is not in a subcounter because it takes a
1860                negligible and constant-sized amount of time */
1861             nbnxn_atomdata_add_nbat_fshift_to_fshift(
1862                     *nbv->nbat, forceOutNonbonded->forceWithShiftForces().shiftForces());
1863         }
1864     }
1865
1866     // TODO Force flags should include haveFreeEnergyWork for this domain
1867     if (stepWork.useGpuXHalo && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1868     {
1869         wallcycle_stop(wcycle, WallCycleCounter::Force);
1870         /* Wait for non-local coordinate data to be copied from device */
1871         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1872         wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
1873     }
1874
1875     // Compute wall interactions, when present.
1876     // Note: should be moved to special forces.
1877     if (inputrec.nwall && stepWork.computeNonbondedForces)
1878     {
1879         /* foreign lambda component for walls */
1880         real dvdl_walls = do_walls(inputrec,
1881                                    *fr,
1882                                    box,
1883                                    mdatoms->typeA ? gmx::arrayRefFromArray(mdatoms->typeA, mdatoms->nr)
1884                                                   : gmx::ArrayRef<int>{},
1885                                    mdatoms->typeB ? gmx::arrayRefFromArray(mdatoms->typeB, mdatoms->nr)
1886                                                   : gmx::ArrayRef<int>{},
1887                                    mdatoms->cENER ? gmx::arrayRefFromArray(mdatoms->cENER, mdatoms->nr)
1888                                                   : gmx::ArrayRef<unsigned short>{},
1889                                    mdatoms->homenr,
1890                                    mdatoms->nPerturbed,
1891                                    x.unpaddedConstArrayRef(),
1892                                    &forceOutMtsLevel0.forceWithVirial(),
1893                                    lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)],
1894                                    enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR],
1895                                    nrnb);
1896         enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] += dvdl_walls;
1897     }
1898
1899     if (stepWork.computeListedForces)
1900     {
1901         /* Check whether we need to take into account PBC in listed interactions */
1902         bool needMolPbc = false;
1903         for (const auto& listedForces : fr->listedForces)
1904         {
1905             if (listedForces.haveCpuListedForces(*fr->fcdata))
1906             {
1907                 needMolPbc = fr->bMolPBC;
1908             }
1909         }
1910
1911         t_pbc pbc;
1912
1913         if (needMolPbc)
1914         {
1915             /* Since all atoms are in the rectangular or triclinic unit-cell,
1916              * only single box vector shifts (2 in x) are required.
1917              */
1918             set_pbc_dd(&pbc, fr->pbcType, DOMAINDECOMP(cr) ? cr->dd->numCells : nullptr, TRUE, box);
1919         }
1920
1921         for (int mtsIndex = 0; mtsIndex < (simulationWork.useMts && stepWork.computeSlowForces ? 2 : 1);
1922              mtsIndex++)
1923         {
1924             ListedForces& listedForces = fr->listedForces[mtsIndex];
1925             ForceOutputs& forceOut     = (mtsIndex == 0 ? forceOutMtsLevel0 : *forceOutMtsLevel1);
1926             listedForces.calculate(wcycle,
1927                                    box,
1928                                    inputrec.fepvals.get(),
1929                                    cr,
1930                                    ms,
1931                                    x,
1932                                    xWholeMolecules,
1933                                    fr->fcdata.get(),
1934                                    hist,
1935                                    &forceOut,
1936                                    fr,
1937                                    &pbc,
1938                                    enerd,
1939                                    nrnb,
1940                                    lambda,
1941                                    mdatoms,
1942                                    DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr,
1943                                    stepWork);
1944         }
1945     }
1946
1947     if (stepWork.computeSlowForces)
1948     {
1949         calculateLongRangeNonbondeds(fr,
1950                                      inputrec,
1951                                      cr,
1952                                      nrnb,
1953                                      wcycle,
1954                                      mdatoms,
1955                                      x.unpaddedConstArrayRef(),
1956                                      &forceOutMtsLevel1->forceWithVirial(),
1957                                      enerd,
1958                                      box,
1959                                      lambda,
1960                                      dipoleData.muStateAB,
1961                                      stepWork,
1962                                      ddBalanceRegionHandler);
1963     }
1964
1965     wallcycle_stop(wcycle, WallCycleCounter::Force);
1966
1967     // VdW dispersion correction, only computed on master rank to avoid double counting
1968     if ((stepWork.computeEnergy || stepWork.computeVirial) && fr->dispersionCorrection && MASTER(cr))
1969     {
1970         // Calculate long range corrections to pressure and energy
1971         const DispersionCorrection::Correction correction = fr->dispersionCorrection->calculate(
1972                 box, lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)]);
1973
1974         if (stepWork.computeEnergy)
1975         {
1976             enerd->term[F_DISPCORR] = correction.energy;
1977             enerd->term[F_DVDL_VDW] += correction.dvdl;
1978             enerd->dvdl_lin[FreeEnergyPerturbationCouplingType::Vdw] += correction.dvdl;
1979         }
1980         if (stepWork.computeVirial)
1981         {
1982             correction.correctVirial(vir_force);
1983             enerd->term[F_PDISPCORR] = correction.pressure;
1984         }
1985     }
1986
1987     computeSpecialForces(fplog,
1988                          cr,
1989                          inputrec,
1990                          awh,
1991                          enforcedRotation,
1992                          imdSession,
1993                          pull_work,
1994                          step,
1995                          t,
1996                          wcycle,
1997                          fr->forceProviders,
1998                          box,
1999                          x.unpaddedArrayRef(),
2000                          mdatoms,
2001                          lambda,
2002                          stepWork,
2003                          &forceOutMtsLevel0.forceWithVirial(),
2004                          forceOutMtsLevel1 ? &forceOutMtsLevel1->forceWithVirial() : nullptr,
2005                          enerd,
2006                          ed,
2007                          stepWork.doNeighborSearch);
2008
2009     if (simulationWork.havePpDomainDecomposition && stepWork.computeForces && stepWork.useGpuFHalo
2010         && domainWork.haveCpuLocalForceWork)
2011     {
2012         stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(), AtomLocality::Local);
2013     }
2014
2015     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
2016                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
2017     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFHalo),
2018                "The schedule below does not allow for nonbonded MTS with GPU halo exchange");
2019     // Will store the amount of cycles spent waiting for the GPU that
2020     // will be later used in the DLB accounting.
2021     float cycles_wait_gpu = 0;
2022     if (useOrEmulateGpuNb && stepWork.computeNonbondedForces)
2023     {
2024         auto& forceWithShiftForces = forceOutNonbonded->forceWithShiftForces();
2025
2026         /* wait for non-local forces (or calculate in emulation mode) */
2027         if (simulationWork.havePpDomainDecomposition)
2028         {
2029             if (simulationWork.useGpuNonbonded)
2030             {
2031                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
2032                         nbv->gpu_nbv,
2033                         stepWork,
2034                         AtomLocality::NonLocal,
2035                         enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR].data(),
2036                         enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR].data(),
2037                         forceWithShiftForces.shiftForces(),
2038                         wcycle);
2039             }
2040             else
2041             {
2042                 wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
2043                 do_nb_verlet(
2044                         fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes, step, nrnb, wcycle);
2045                 wallcycle_stop(wcycle, WallCycleCounter::Force);
2046             }
2047
2048             if (stepWork.useGpuFBufferOps)
2049             {
2050                 if (domainWork.haveNonLocalForceContribInCpuBuffer)
2051                 {
2052                     stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
2053                                               AtomLocality::NonLocal);
2054                 }
2055
2056
2057                 fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->execute();
2058
2059                 if (!stepWork.useGpuFHalo)
2060                 {
2061                     // copy from GPU input for dd_move_f()
2062                     stateGpu->copyForcesFromGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
2063                                                 AtomLocality::NonLocal);
2064                 }
2065             }
2066             else
2067             {
2068                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
2069             }
2070
2071             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
2072             {
2073                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
2074             }
2075         }
2076     }
2077
2078     /* Combining the forces for multiple time stepping before the halo exchange, when possible,
2079      * avoids an extra halo exchange (when DD is used) and post-processing step.
2080      */
2081     if (stepWork.combineMtsForcesBeforeHaloExchange)
2082     {
2083         combineMtsForces(getLocalAtomCount(cr->dd, *mdatoms, simulationWork.havePpDomainDecomposition),
2084                          force.unpaddedArrayRef(),
2085                          forceView->forceMtsCombined(),
2086                          inputrec.mtsLevels[1].stepFactor);
2087     }
2088
2089     if (simulationWork.havePpDomainDecomposition)
2090     {
2091         /* We are done with the CPU compute.
2092          * We will now communicate the non-local forces.
2093          * If we use a GPU this will overlap with GPU work, so in that case
2094          * we do not close the DD force balancing region here.
2095          */
2096         ddBalanceRegionHandler.closeAfterForceComputationCpu();
2097
2098         if (stepWork.computeForces)
2099         {
2100
2101             if (stepWork.useGpuFHalo)
2102             {
2103                 // If there exist CPU forces, data from halo exchange should accumulate into these
2104                 bool accumulateForces = domainWork.haveCpuLocalForceWork;
2105                 if (!accumulateForces)
2106                 {
2107                     // Force halo exchange will set a subset of local atoms with remote non-local data
2108                     // First clear local portion of force array, so that untouched atoms are zero
2109                     stateGpu->clearForcesOnGpu(AtomLocality::Local);
2110                 }
2111                 communicateGpuHaloForces(*cr, accumulateForces);
2112             }
2113             else
2114             {
2115                 if (stepWork.useGpuFBufferOps)
2116                 {
2117                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
2118                 }
2119
2120                 // Without MTS or with MTS at slow steps with uncombined forces we need to
2121                 // communicate the fast forces
2122                 if (!simulationWork.useMts || !stepWork.combineMtsForcesBeforeHaloExchange)
2123                 {
2124                     dd_move_f(cr->dd, &forceOutMtsLevel0.forceWithShiftForces(), wcycle);
2125                 }
2126                 // With MTS we need to communicate the slow or combined (in forceOutMtsLevel1) forces
2127                 if (simulationWork.useMts && stepWork.computeSlowForces)
2128                 {
2129                     dd_move_f(cr->dd, &forceOutMtsLevel1->forceWithShiftForces(), wcycle);
2130                 }
2131             }
2132         }
2133     }
2134
2135     // With both nonbonded and PME offloaded a GPU on the same rank, we use
2136     // an alternating wait/reduction scheme.
2137     bool alternateGpuWait =
2138             (!c_disableAlternatingWait && stepWork.haveGpuPmeOnThisRank
2139              && simulationWork.useGpuNonbonded && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
2140     if (alternateGpuWait)
2141     {
2142         alternatePmeNbGpuWaitReduce(fr->nbv.get(),
2143                                     fr->pmedata,
2144                                     forceOutNonbonded,
2145                                     forceOutMtsLevel1,
2146                                     enerd,
2147                                     lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)],
2148                                     stepWork,
2149                                     wcycle);
2150     }
2151
2152     if (!alternateGpuWait && stepWork.haveGpuPmeOnThisRank)
2153     {
2154         pme_gpu_wait_and_reduce(fr->pmedata,
2155                                 stepWork,
2156                                 wcycle,
2157                                 &forceOutMtsLevel1->forceWithVirial(),
2158                                 enerd,
2159                                 lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)]);
2160     }
2161
2162     /* Wait for local GPU NB outputs on the non-alternating wait path */
2163     if (!alternateGpuWait && stepWork.computeNonbondedForces && simulationWork.useGpuNonbonded)
2164     {
2165         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
2166          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
2167          * but even with a step of 0.1 ms the difference is less than 1%
2168          * of the step time.
2169          */
2170         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
2171         const float waitCycles               = Nbnxm::gpu_wait_finish_task(
2172                 nbv->gpu_nbv,
2173                 stepWork,
2174                 AtomLocality::Local,
2175                 enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::LJSR].data(),
2176                 enerd->grpp.energyGroupPairTerms[NonBondedEnergyTerms::CoulombSR].data(),
2177                 forceOutNonbonded->forceWithShiftForces().shiftForces(),
2178                 wcycle);
2179
2180         if (ddBalanceRegionHandler.useBalancingRegion())
2181         {
2182             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
2183             if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
2184             {
2185                 /* We measured few cycles, it could be that the kernel
2186                  * and transfer finished earlier and there was no actual
2187                  * wait time, only API call overhead.
2188                  * Then the actual time could be anywhere between 0 and
2189                  * cycles_wait_est. We will use half of cycles_wait_est.
2190                  */
2191                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
2192             }
2193             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
2194         }
2195     }
2196
2197     if (fr->nbv->emulateGpu())
2198     {
2199         // NOTE: emulation kernel is not included in the balancing region,
2200         // but emulation mode does not target performance anyway
2201         wallcycle_start_nocount(wcycle, WallCycleCounter::Force);
2202         do_nb_verlet(fr,
2203                      ic,
2204                      enerd,
2205                      stepWork,
2206                      InteractionLocality::Local,
2207                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
2208                      step,
2209                      nrnb,
2210                      wcycle);
2211         wallcycle_stop(wcycle, WallCycleCounter::Force);
2212     }
2213
2214     // If on GPU PME-PP comms path, receive forces from PME before GPU buffer ops
2215     // TODO refactor this and unify with below default-path call to the same function
2216     if (PAR(cr) && simulationWork.haveSeparatePmeRank && simulationWork.useGpuPmePpCommunication
2217         && stepWork.computeSlowForces)
2218     {
2219         /* In case of node-splitting, the PP nodes receive the long-range
2220          * forces, virial and energy from the PME nodes here.
2221          */
2222         pme_receive_force_ener(fr,
2223                                cr,
2224                                &forceOutMtsLevel1->forceWithVirial(),
2225                                enerd,
2226                                simulationWork.useGpuPmePpCommunication,
2227                                stepWork.useGpuPmeFReduction,
2228                                wcycle);
2229     }
2230
2231
2232     /* Do the nonbonded GPU (or emulation) force buffer reduction
2233      * on the non-alternating path. */
2234     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
2235                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
2236     if (useOrEmulateGpuNb && !alternateGpuWait)
2237     {
2238         if (stepWork.useGpuFBufferOps)
2239         {
2240             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2241
2242             // TODO: move these steps as early as possible:
2243             // - CPU f H2D should be as soon as all CPU-side forces are done
2244             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
2245             //   before the next CPU task that consumes the forces: vsite spread or update)
2246             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
2247             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
2248             //   These should be unified.
2249             if (domainWork.haveLocalForceContribInCpuBuffer && !stepWork.useGpuFHalo)
2250             {
2251                 // Note: AtomLocality::All is used for the non-DD case because, as in this
2252                 // case copyForcesToGpu() uses a separate stream, it allows overlap of
2253                 // CPU force H2D with GPU force tasks on all streams including those in the
2254                 // local stream which would otherwise be implicit dependencies for the
2255                 // transfer and would not overlap.
2256                 auto locality = simulationWork.havePpDomainDecomposition ? AtomLocality::Local
2257                                                                          : AtomLocality::All;
2258
2259                 stateGpu->copyForcesToGpu(forceWithShift, locality);
2260             }
2261
2262             if (stepWork.computeNonbondedForces)
2263             {
2264                 fr->gpuForceReduction[gmx::AtomLocality::Local]->execute();
2265             }
2266
2267             // Copy forces to host if they are needed for update or if virtual sites are enabled.
2268             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
2269             // TODO: When the output flags will be included in step workload, this copy can be combined with the
2270             //       copy call done in sim_utils(...) for the output.
2271             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
2272             //       they should not be copied in do_md(...) for the output.
2273             if (!simulationWork.useGpuUpdate
2274                 || (simulationWork.useGpuUpdate && DOMAINDECOMP(cr) && simulationWork.useCpuPmePpCommunication)
2275                 || vsite)
2276             {
2277                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
2278                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
2279             }
2280         }
2281         else if (stepWork.computeNonbondedForces)
2282         {
2283             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2284             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
2285         }
2286     }
2287
2288     launchGpuEndOfStepTasks(
2289             nbv, fr->listedForcesGpu.get(), fr->pmedata, enerd, *runScheduleWork, step, wcycle);
2290
2291     if (DOMAINDECOMP(cr))
2292     {
2293         dd_force_flop_stop(cr->dd, nrnb);
2294     }
2295
2296     const bool haveCombinedMtsForces = (stepWork.computeForces && simulationWork.useMts && stepWork.computeSlowForces
2297                                         && stepWork.combineMtsForcesBeforeHaloExchange);
2298     if (stepWork.computeForces)
2299     {
2300         postProcessForceWithShiftForces(
2301                 nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutMtsLevel0, vir_force, *mdatoms, *fr, vsite, stepWork);
2302
2303         if (simulationWork.useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2304         {
2305             postProcessForceWithShiftForces(
2306                     nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1, vir_force, *mdatoms, *fr, vsite, stepWork);
2307         }
2308     }
2309
2310     // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
2311     if (PAR(cr) && simulationWork.haveSeparatePmeRank && simulationWork.useCpuPmePpCommunication
2312         && stepWork.computeSlowForces)
2313     {
2314         /* In case of node-splitting, the PP nodes receive the long-range
2315          * forces, virial and energy from the PME nodes here.
2316          */
2317         pme_receive_force_ener(fr,
2318                                cr,
2319                                &forceOutMtsLevel1->forceWithVirial(),
2320                                enerd,
2321                                simulationWork.useGpuPmePpCommunication,
2322                                false,
2323                                wcycle);
2324     }
2325
2326     if (stepWork.computeForces)
2327     {
2328         /* If we don't use MTS or if we already combined the MTS forces before, we only
2329          * need to post-process one ForceOutputs object here, called forceOutCombined,
2330          * otherwise we have to post-process two outputs and then combine them.
2331          */
2332         ForceOutputs& forceOutCombined = (haveCombinedMtsForces ? forceOutMts.value() : forceOutMtsLevel0);
2333         postProcessForces(
2334                 cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutCombined, vir_force, mdatoms, fr, vsite, stepWork);
2335
2336         if (simulationWork.useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2337         {
2338             postProcessForces(
2339                     cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1, vir_force, mdatoms, fr, vsite, stepWork);
2340
2341             combineMtsForces(mdatoms->homenr,
2342                              force.unpaddedArrayRef(),
2343                              forceView->forceMtsCombined(),
2344                              inputrec.mtsLevels[1].stepFactor);
2345         }
2346     }
2347
2348     if (stepWork.computeEnergy)
2349     {
2350         /* Compute the final potential energy terms */
2351         accumulatePotentialEnergies(enerd, lambda, inputrec.fepvals.get());
2352
2353         if (!EI_TPI(inputrec.eI))
2354         {
2355             checkPotentialEnergyValidity(step, *enerd, inputrec);
2356         }
2357     }
2358
2359     /* In case we don't have constraints and are using GPUs, the next balancing
2360      * region starts here.
2361      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
2362      * virial calculation and COM pulling, is not thus not included in
2363      * the balance timing, which is ok as most tasks do communication.
2364      */
2365     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
2366 }