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