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] forceWithVirial  Force and virial buffers
 584  * \param[in,out] enerd            Energy buffer
 585  * \param[in,out] ed               Essential dynamics pointer
 586  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 587  *
 588  * \todo Remove didNeighborSearch, which is used incorrectly.
 589  * \todo Convert all other algorithms called here to ForceProviders.
 590  */
 591 static void computeSpecialForces(FILE*                          fplog,
 592                                  const t_commrec*               cr,
 593                                  const t_inputrec*              inputrec,
 594                                  gmx::Awh*                      awh,
 595                                  gmx_enfrot*                    enforcedRotation,
 596                                  gmx::ImdSession*               imdSession,
 597                                  pull_t*                        pull_work,
 598                                  int64_t                        step,
 599                                  double                         t,
 600                                  gmx_wallcycle_t                wcycle,
 601                                  gmx::ForceProviders*           forceProviders,
 602                                  const matrix                   box,
 603                                  gmx::ArrayRef<const gmx::RVec> x,
 604                                  const t_mdatoms*               mdatoms,
 605                                  gmx::ArrayRef<const real>      lambda,
 606                                  const StepWorkload&            stepWork,
 607                                  gmx::ForceWithVirial*          forceWithVirial,
 608                                  gmx_enerdata_t*                enerd,
 609                                  gmx_edsam*                     ed,
 610                                  bool                           didNeighborSearch)
 611 {
 612     /* NOTE: Currently all ForceProviders only provide forces.
 613      *       When they also provide energies, remove this conditional.
 614      */
 615     if (stepWork.computeForces)
 616     {
 617         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 618         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 619
 620         /* Collect forces from modules */
 621         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 622     }
 623
 624     if (inputrec->bPull && pull_have_potential(pull_work))
 625     {
 626         pull_potential_wrapper(cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work,
 627                                lambda.data(), t, wcycle);
 628     }
 629     if (awh)
 630     {
 631         const bool          needForeignEnergyDifferences = awh->needForeignEnergyDifferences(step);
 632         std::vector<double> foreignLambdaDeltaH, foreignLambdaDhDl;
 633         if (needForeignEnergyDifferences)
 634         {
 635             enerd->foreignLambdaTerms.finalizePotentialContributions(enerd->dvdl_lin, lambda,
 636                                                                      *inputrec->fepvals);
 637             std::tie(foreignLambdaDeltaH, foreignLambdaDhDl) = enerd->foreignLambdaTerms.getTerms(cr);
 638         }
 639
 640         enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
 641                 inputrec->pbcType, mdatoms->massT, foreignLambdaDeltaH, foreignLambdaDhDl, box,
 642                 forceWithVirial, t, step, wcycle, fplog);
 643     }
 644
 645     rvec* f = as_rvec_array(forceWithVirial->force_.data());
 646
 647     /* Add the forces from enforced rotation potentials (if any) */
 648     if (inputrec->bRot)
 649     {
 650         wallcycle_start(wcycle, ewcROTadd);
 651         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 652         wallcycle_stop(wcycle, ewcROTadd);
 653     }
 654
 655     if (ed)
 656     {
 657         /* Note that since init_edsam() is called after the initialization
 658          * of forcerec, edsam doesn't request the noVirSum force buffer.
 659          * Thus if no other algorithm (e.g. PME) requires it, the forces
 660          * here will contribute to the virial.
 661          */
 662         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
 663     }
 664
 665     /* Add forces from interactive molecular dynamics (IMD), if any */
 666     if (inputrec->bIMD && stepWork.computeForces)
 667     {
 668         imdSession->applyForces(f);
 669     }
 670 }
 671
 672 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 673  *
 674  * \param[in]  pmedata              The PME structure
 675  * \param[in]  box                  The box matrix
 676  * \param[in]  stepWork             Step schedule flags
 677  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in the device memory.
 678  * \param[in]  lambdaQ              The Coulomb lambda of the current state.
 679  * \param[in]  wcycle               The wallcycle structure
 680  */
 681 static inline void launchPmeGpuSpread(gmx_pme_t*            pmedata,
 682                                       const matrix          box,
 683                                       const StepWorkload&   stepWork,
 684                                       GpuEventSynchronizer* xReadyOnDevice,
 685                                       const real            lambdaQ,
 686                                       gmx_wallcycle_t       wcycle)
 687 {
 688     pme_gpu_prepare_computation(pmedata, box, wcycle, stepWork);
 689     pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle, lambdaQ);
 690 }
 691
 692 /*! \brief Launch the FFT and gather stages of PME GPU
 693  *
 694  * This function only implements setting the output forces (no accumulation).
 695  *
 696  * \param[in]  pmedata        The PME structure
 697  * \param[in]  lambdaQ        The Coulomb lambda of the current system state.
 698  * \param[in]  wcycle         The wallcycle structure
 699  * \param[in]  stepWork       Step schedule flags
 700  */
 701 static void launchPmeGpuFftAndGather(gmx_pme_t*               pmedata,
 702                                      const real               lambdaQ,
 703                                      gmx_wallcycle_t          wcycle,
 704                                      const gmx::StepWorkload& stepWork)
 705 {
 706     pme_gpu_launch_complex_transforms(pmedata, wcycle, stepWork);
 707     pme_gpu_launch_gather(pmedata, wcycle, lambdaQ);
 708 }
 709
 710 /*! \brief
 711  *  Polling wait for either of the PME or nonbonded GPU tasks.
 712  *
 713  * Instead of a static order in waiting for GPU tasks, this function
 714  * polls checking which of the two tasks completes first, and does the
 715  * associated force buffer reduction overlapped with the other task.
 716  * By doing that, unlike static scheduling order, it can always overlap
 717  * one of the reductions, regardless of the GPU task completion order.
 718  *
 719  * \param[in]     nbv              Nonbonded verlet structure
 720  * \param[in,out] pmedata          PME module data
 721  * \param[in,out] forceOutputsNonbonded  Force outputs for the non-bonded forces and shift forces
 722  * \param[in,out] forceOutputsPme  Force outputs for the PME forces and virial
 723  * \param[in,out] enerd            Energy data structure results are reduced into
 724  * \param[in]     lambdaQ          The Coulomb lambda of the current system state.
 725  * \param[in]     stepWork         Step schedule flags
 726  * \param[in]     wcycle           The wallcycle structure
 727  */
 728 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
 729                                         gmx_pme_t*          pmedata,
 730                                         gmx::ForceOutputs*  forceOutputsNonbonded,
 731                                         gmx::ForceOutputs*  forceOutputsPme,
 732                                         gmx_enerdata_t*     enerd,
 733                                         const real          lambdaQ,
 734                                         const StepWorkload& stepWork,
 735                                         gmx_wallcycle_t     wcycle)
 736 {
 737     bool isPmeGpuDone = false;
 738     bool isNbGpuDone  = false;
 739
 740     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 741
 742     while (!isPmeGpuDone || !isNbGpuDone)
 743     {
 744         if (!isPmeGpuDone)
 745         {
 746             GpuTaskCompletion completionType =
 747                     (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 748             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, stepWork, wcycle,
 749                                                    &forceOutputsPme->forceWithVirial(), enerd,
 750                                                    lambdaQ, completionType);
 751         }
 752
 753         if (!isNbGpuDone)
 754         {
 755             auto&             forceBuffersNonbonded = forceOutputsNonbonded->forceWithShiftForces();
 756             GpuTaskCompletion completionType =
 757                     (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 758             isNbGpuDone = Nbnxm::gpu_try_finish_task(
 759                     nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
 760                     enerd->grpp.ener[egCOULSR].data(), forceBuffersNonbonded.shiftForces(),
 761                     completionType, wcycle);
 762
 763             if (isNbGpuDone)
 764             {
 765                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceBuffersNonbonded.force());
 766             }
 767         }
 768     }
 769 }
 770
 771 /*! \brief Set up the different force buffers; also does clearing.
 772  *
 773  * \param[in] forceHelperBuffers  Helper force buffers
 774  * \param[in] force     force array
 775  * \param[in] stepWork  Step schedule flags
 776  * \param[out] wcycle   wallcycle recording structure
 777  *
 778  * \returns             Cleared force output structure
 779  */
 780 static ForceOutputs setupForceOutputs(ForceHelperBuffers*                 forceHelperBuffers,
 781                                       gmx::ArrayRefWithPadding<gmx::RVec> force,
 782                                       const StepWorkload&                 stepWork,
 783                                       gmx_wallcycle_t                     wcycle)
 784 {
 785     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 786
 787     /* NOTE: We assume fr->shiftForces is all zeros here */
 788     gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial,
 789                                                    forceHelperBuffers->shiftForces());
 790
 791     if (stepWork.computeForces)
 792     {
 793         /* Clear the short- and long-range forces */
 794         clearRVecs(forceWithShiftForces.force(), true);
 795
 796         /* Clear the shift forces */
 797         clearRVecs(forceWithShiftForces.shiftForces(), false);
 798     }
 799
 800     /* If we need to compute the virial, we might need a separate
 801      * force buffer for algorithms for which the virial is calculated
 802      * directly, such as PME. Otherwise, forceWithVirial uses the
 803      * the same force (f in legacy calls) buffer as other algorithms.
 804      */
 805     const bool useSeparateForceWithVirialBuffer =
 806             (stepWork.computeForces
 807              && (stepWork.computeVirial && forceHelperBuffers->haveDirectVirialContributions()));
 808     /* forceWithVirial uses the local atom range only */
 809     gmx::ForceWithVirial forceWithVirial(
 810             useSeparateForceWithVirialBuffer ? forceHelperBuffers->forceBufferForDirectVirialContributions()
 811                                              : force.unpaddedArrayRef(),
 812             stepWork.computeVirial);
 813
 814     if (useSeparateForceWithVirialBuffer)
 815     {
 816         /* TODO: update comment
 817          * We only compute forces on local atoms. Note that vsites can
 818          * spread to non-local atoms, but that part of the buffer is
 819          * cleared separately in the vsite spreading code.
 820          */
 821         clearRVecs(forceWithVirial.force_, true);
 822     }
 823
 824     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 825
 826     return ForceOutputs(forceWithShiftForces, forceHelperBuffers->haveDirectVirialContributions(),
 827                         forceWithVirial);
 828 }
 829
 830
 831 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 832  */
 833 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec&         inputrec,
 834                                                           const t_forcerec&         fr,
 835                                                           const pull_t*             pull_work,
 836                                                           const gmx_edsam*          ed,
 837                                                           const t_mdatoms&          mdatoms,
 838                                                           const SimulationWorkload& simulationWork,
 839                                                           const StepWorkload&       stepWork)
 840 {
 841     DomainLifetimeWorkload domainWork;
 842     // Note that haveSpecialForces is constant over the whole run
 843     domainWork.haveSpecialForces =
 844             haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 845     domainWork.haveCpuListedForceWork = false;
 846     domainWork.haveCpuBondedWork      = false;
 847     for (const auto& listedForces : fr.listedForces)
 848     {
 849         if (listedForces.haveCpuListedForces(*fr.fcdata))
 850         {
 851             domainWork.haveCpuListedForceWork = true;
 852         }
 853         if (listedForces.haveCpuBondeds())
 854         {
 855             domainWork.haveCpuBondedWork = true;
 856         }
 857     }
 858     domainWork.haveGpuBondedWork = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
 859     // Note that haveFreeEnergyWork is constant over the whole run
 860     domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
 861     // We assume we have local force work if there are CPU
 862     // force tasks including PME or nonbondeds.
 863     domainWork.haveCpuLocalForceWork =
 864             domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
 865             || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
 866             || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
 867
 868     return domainWork;
 869 }
 870
 871 /*! \brief Set up force flag stuct from the force bitmask.
 872  *
 873  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 874  * \param[in]      mtsLevels            The multiple time-stepping levels, either empty or 2 levels
 875  * \param[in]      step                 The current MD step
 876  * \param[in]      simulationWork       Simulation workload description.
 877  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 878  *
 879  * \returns New Stepworkload description.
 880  */
 881 static StepWorkload setupStepWorkload(const int                     legacyFlags,
 882                                       ArrayRef<const gmx::MtsLevel> mtsLevels,
 883                                       const int64_t                 step,
 884                                       const SimulationWorkload&     simulationWork,
 885                                       const bool                    rankHasPmeDuty)
 886 {
 887     GMX_ASSERT(mtsLevels.empty() || mtsLevels.size() == 2, "Expect 0 or 2 MTS levels");
 888     const bool computeSlowForces = (mtsLevels.empty() || step % mtsLevels[1].stepFactor == 0);
 889
 890     StepWorkload flags;
 891     flags.stateChanged        = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 892     flags.haveDynamicBox      = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 893     flags.doNeighborSearch    = ((legacyFlags & GMX_FORCE_NS) != 0);
 894     flags.computeSlowForces   = computeSlowForces;
 895     flags.computeVirial       = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 896     flags.computeEnergy       = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 897     flags.computeForces       = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 898     flags.computeListedForces = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 899     flags.computeNonbondedForces =
 900             ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && simulationWork.computeNonbonded
 901             && !(simulationWork.computeNonbondedAtMtsLevel1 && !computeSlowForces);
 902     flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 903
 904     if (simulationWork.useGpuBufferOps)
 905     {
 906         GMX_ASSERT(simulationWork.useGpuNonbonded,
 907                    "Can only offload buffer ops if nonbonded computation is also offloaded");
 908     }
 909     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 910     // on virial steps the CPU reduction path is taken
 911     flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !flags.computeVirial;
 912     flags.useGpuPmeFReduction = flags.computeSlowForces && flags.useGpuFBufferOps && simulationWork.useGpuPme
 913                                 && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication);
 914
 915     return flags;
 916 }
 917
 918
 919 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
 920  *
 921  * TODO: eliminate \p useGpuPmeOnThisRank when this is
 922  * incorporated in DomainLifetimeWorkload.
 923  */
 924 static void launchGpuEndOfStepTasks(nonbonded_verlet_t*               nbv,
 925                                     gmx::GpuBonded*                   gpuBonded,
 926                                     gmx_pme_t*                        pmedata,
 927                                     gmx_enerdata_t*                   enerd,
 928                                     const gmx::MdrunScheduleWorkload& runScheduleWork,
 929                                     bool                              useGpuPmeOnThisRank,
 930                                     int64_t                           step,
 931                                     gmx_wallcycle_t                   wcycle)
 932 {
 933     if (runScheduleWork.simulationWork.useGpuNonbonded && runScheduleWork.stepWork.computeNonbondedForces)
 934     {
 935         /* Launch pruning before buffer clearing because the API overhead of the
 936          * clear kernel launches can leave the GPU idle while it could be running
 937          * the prune kernel.
 938          */
 939         if (nbv->isDynamicPruningStepGpu(step))
 940         {
 941             nbv->dispatchPruneKernelGpu(step);
 942         }
 943
 944         /* now clear the GPU outputs while we finish the step on the CPU */
 945         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 946         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 947         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
 948         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 949         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 950     }
 951
 952     if (useGpuPmeOnThisRank)
 953     {
 954         pme_gpu_reinit_computation(pmedata, wcycle);
 955     }
 956
 957     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
 958     {
 959         // in principle this should be included in the DD balancing region,
 960         // but generally it is infrequent so we'll omit it for the sake of
 961         // simpler code
 962         gpuBonded->waitAccumulateEnergyTerms(enerd);
 963
 964         gpuBonded->clearEnergies();
 965     }
 966 }
 967
 968 //! \brief Data structure to hold dipole-related data and staging arrays
 969 struct DipoleData
 970 {
 971     //! Dipole staging for fast summing over MPI
 972     gmx::DVec muStaging[2] = { { 0.0, 0.0, 0.0 } };
 973     //! Dipole staging for states A and B (index 0 and 1 resp.)
 974     gmx::RVec muStateAB[2] = { { 0.0_real, 0.0_real, 0.0_real } };
 975 };
 976
 977
 978 static void reduceAndUpdateMuTot(DipoleData*                   dipoleData,
 979                                  const t_commrec*              cr,
 980                                  const bool                    haveFreeEnergy,
 981                                  gmx::ArrayRef<const real>     lambda,
 982                                  rvec                          muTotal,
 983                                  const DDBalanceRegionHandler& ddBalanceRegionHandler)
 984 {
 985     if (PAR(cr))
 986     {
 987         gmx_sumd(2 * DIM, dipoleData->muStaging[0], cr);
 988         ddBalanceRegionHandler.reopenRegionCpu();
 989     }
 990     for (int i = 0; i < 2; i++)
 991     {
 992         for (int j = 0; j < DIM; j++)
 993         {
 994             dipoleData->muStateAB[i][j] = dipoleData->muStaging[i][j];
 995         }
 996     }
 997
 998     if (!haveFreeEnergy)
 999     {
1000         copy_rvec(dipoleData->muStateAB[0], muTotal);
1001     }
1002     else
1003     {
1004         for (int j = 0; j < DIM; j++)
1005         {
1006             muTotal[j] = (1.0 - lambda[efptCOUL]) * dipoleData->muStateAB[0][j]
1007                          + lambda[efptCOUL] * dipoleData->muStateAB[1][j];
1008         }
1009     }
1010 }
1011
1012 /*! \brief Combines MTS level0 and level1 force buffes into a full and MTS-combined force buffer.
1013  *
1014  * \param[in]     numAtoms        The number of atoms to combine forces for
1015  * \param[in,out] forceMtsLevel0  Input: F_level0, output: F_level0 + F_level1
1016  * \param[in,out] forceMts        Input: F_level1, output: F_level0 + mtsFactor * F_level1
1017  * \param[in]     mtsFactor       The factor between the level0 and level1 time step
1018  */
1019 static void combineMtsForces(const int      numAtoms,
1020                              ArrayRef<RVec> forceMtsLevel0,
1021                              ArrayRef<RVec> forceMts,
1022                              const real     mtsFactor)
1023 {
1024     const int gmx_unused numThreads = gmx_omp_nthreads_get(emntDefault);
1025 #pragma omp parallel for num_threads(numThreads) schedule(static)
1026     for (int i = 0; i < numAtoms; i++)
1027     {
1028         const RVec forceMtsLevel0Tmp = forceMtsLevel0[i];
1029         forceMtsLevel0[i] += forceMts[i];
1030         forceMts[i] = forceMtsLevel0Tmp + mtsFactor * forceMts[i];
1031     }
1032 }
1033
1034 /*! \brief Setup for the local and non-local GPU force reductions:
1035  * reinitialization plus the registration of forces and dependencies.
1036  *
1037  * \param [in] runScheduleWork               Schedule workload flag structure
1038  * \param [in] cr                            Communication record object
1039  * \param [in] fr                            Force record object
1040  * \param [in] ddUsesGpuDirectCommunication  Whether GPU direct communication is in use
1041  */
1042 static void setupGpuForceReductions(gmx::MdrunScheduleWorkload* runScheduleWork,
1043                                     const t_commrec*            cr,
1044                                     t_forcerec*                 fr,
1045                                     bool                        ddUsesGpuDirectCommunication)
1046 {
1047
1048     nonbonded_verlet_t*          nbv      = fr->nbv.get();
1049     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1050
1051     // (re-)initialize local GPU force reduction
1052     const bool accumulate =
1053             runScheduleWork->domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr);
1054     const int atomStart = 0;
1055     fr->gpuForceReduction[gmx::AtomLocality::Local]->reinit(
1056             stateGpu->getForces(), nbv->getNumAtoms(AtomLocality::Local), nbv->getGridIndices(),
1057             atomStart, accumulate, stateGpu->fReducedOnDevice());
1058
1059     // register forces and add dependencies
1060     fr->gpuForceReduction[gmx::AtomLocality::Local]->registerNbnxmForce(nbv->getGpuForces());
1061
1062     if (runScheduleWork->simulationWork.useGpuPme
1063         && (thisRankHasDuty(cr, DUTY_PME) || runScheduleWork->simulationWork.useGpuPmePpCommunication))
1064     {
1065         void* forcePtr = thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1066                                                        : // PME force buffer on same GPU
1067                                  fr->pmePpCommGpu->getGpuForceStagingPtr(); // buffer received from other GPU
1068         fr->gpuForceReduction[gmx::AtomLocality::Local]->registerRvecForce(forcePtr);
1069
1070         GpuEventSynchronizer* const pmeSynchronizer =
1071                 (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1072                                                : // PME force buffer on same GPU
1073                          fr->pmePpCommGpu->getForcesReadySynchronizer()); // buffer received from other GPU
1074         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(pmeSynchronizer);
1075     }
1076
1077     if ((runScheduleWork->domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr))
1078         && !ddUsesGpuDirectCommunication)
1079     {
1080         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1081                 stateGpu->getForcesReadyOnDeviceEvent(AtomLocality::Local, true));
1082     }
1083
1084     if (ddUsesGpuDirectCommunication)
1085     {
1086         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1087                 cr->dd->gpuHaloExchange[0][0]->getForcesReadyOnDeviceEvent());
1088     }
1089
1090     if (havePPDomainDecomposition(cr))
1091     {
1092         // (re-)initialize non-local GPU force reduction
1093         const bool accumulate = runScheduleWork->domainWork.haveCpuBondedWork
1094                                 || runScheduleWork->domainWork.haveFreeEnergyWork;
1095         const int atomStart = dd_numHomeAtoms(*cr->dd);
1096         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->reinit(
1097                 stateGpu->getForces(), nbv->getNumAtoms(AtomLocality::NonLocal),
1098                 nbv->getGridIndices(), atomStart, accumulate);
1099
1100         // register forces and add dependencies
1101         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->registerNbnxmForce(nbv->getGpuForces());
1102         if (runScheduleWork->domainWork.haveCpuBondedWork || runScheduleWork->domainWork.haveFreeEnergyWork)
1103         {
1104             fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->addDependency(
1105                     stateGpu->getForcesReadyOnDeviceEvent(AtomLocality::NonLocal, true));
1106         }
1107     }
1108 }
1109
1110
1111 void do_force(FILE*                               fplog,
1112               const t_commrec*                    cr,
1113               const gmx_multisim_t*               ms,
1114               const t_inputrec*                   inputrec,
1115               gmx::Awh*                           awh,
1116               gmx_enfrot*                         enforcedRotation,
1117               gmx::ImdSession*                    imdSession,
1118               pull_t*                             pull_work,
1119               int64_t                             step,
1120               t_nrnb*                             nrnb,
1121               gmx_wallcycle_t                     wcycle,
1122               const gmx_localtop_t*               top,
1123               const matrix                        box,
1124               gmx::ArrayRefWithPadding<gmx::RVec> x,
1125               history_t*                          hist,
1126               gmx::ForceBuffersView*              forceView,
1127               tensor                              vir_force,
1128               const t_mdatoms*                    mdatoms,
1129               gmx_enerdata_t*                     enerd,
1130               gmx::ArrayRef<const real>           lambda,
1131               t_forcerec*                         fr,
1132               gmx::MdrunScheduleWorkload*         runScheduleWork,
1133               gmx::VirtualSitesHandler*           vsite,
1134               rvec                                muTotal,
1135               double                              t,
1136               gmx_edsam*                          ed,
1137               int                                 legacyFlags,
1138               const DDBalanceRegionHandler&       ddBalanceRegionHandler)
1139 {
1140     auto force = forceView->forceWithPadding();
1141     GMX_ASSERT(force.unpaddedArrayRef().ssize() >= fr->natoms_force_constr,
1142                "The size of the force buffer should be at least the number of atoms to compute "
1143                "forces for");
1144
1145     nonbonded_verlet_t*  nbv = fr->nbv.get();
1146     interaction_const_t* ic  = fr->ic;
1147
1148     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1149
1150     const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
1151
1152     runScheduleWork->stepWork    = setupStepWorkload(legacyFlags, inputrec->mtsLevels, step,
1153                                                   simulationWork, thisRankHasDuty(cr, DUTY_PME));
1154     const StepWorkload& stepWork = runScheduleWork->stepWork;
1155
1156     const bool useGpuPmeOnThisRank =
1157             simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME) && stepWork.computeSlowForces;
1158
1159     /* At a search step we need to start the first balancing region
1160      * somewhere early inside the step after communication during domain
1161      * decomposition (and not during the previous step as usual).
1162      */
1163     if (stepWork.doNeighborSearch)
1164     {
1165         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
1166     }
1167
1168     clear_mat(vir_force);
1169
1170     if (fr->pbcType != PbcType::No)
1171     {
1172         /* Compute shift vectors every step,
1173          * because of pressure coupling or box deformation!
1174          */
1175         if (stepWork.haveDynamicBox && stepWork.stateChanged)
1176         {
1177             calc_shifts(box, fr->shift_vec);
1178         }
1179
1180         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1181         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
1182         if (calcCGCM)
1183         {
1184             put_atoms_in_box_omp(fr->pbcType, box, x.unpaddedArrayRef().subArray(0, mdatoms->homenr),
1185                                  gmx_omp_nthreads_get(emntDefault));
1186             inc_nrnb(nrnb, eNR_SHIFTX, mdatoms->homenr);
1187         }
1188     }
1189
1190     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1191
1192     const bool pmeSendCoordinatesFromGpu =
1193             GMX_MPI && simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1194     const bool reinitGpuPmePpComms =
1195             GMX_MPI && simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1196
1197     const auto localXReadyOnDevice = (useGpuPmeOnThisRank || simulationWork.useGpuBufferOps)
1198                                              ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1199                                                        AtomLocality::Local, simulationWork, stepWork)
1200                                              : nullptr;
1201
1202     // If coordinates are to be sent to PME task from CPU memory, perform that send here.
1203     // Otherwise the send will occur after H2D coordinate transfer.
1204     if (GMX_MPI && !thisRankHasDuty(cr, DUTY_PME) && !pmeSendCoordinatesFromGpu && stepWork.computeSlowForces)
1205     {
1206         /* Send particle coordinates to the pme nodes */
1207         if (!stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1208         {
1209             GMX_RELEASE_ASSERT(false,
1210                                "GPU update and separate PME ranks are only supported with GPU "
1211                                "direct communication!");
1212             // TODO: when this code-path becomes supported add:
1213             // stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1214         }
1215
1216         gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1217                                  lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1218                                  step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1219                                  pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1220     }
1221
1222     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1223     // The local coordinates can be copied right away.
1224     // NOTE: Consider moving this copy to right after they are updated and constrained,
1225     //       if the later is not offloaded.
1226     if (useGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1227     {
1228         if (stepWork.doNeighborSearch)
1229         {
1230             // TODO refactor this to do_md, after partitioning.
1231             stateGpu->reinit(mdatoms->homenr,
1232                              cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1233             if (useGpuPmeOnThisRank)
1234             {
1235                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1236                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1237             }
1238         }
1239         // We need to copy coordinates when:
1240         // 1. Update is not offloaded
1241         // 2. The buffers were reinitialized on search step
1242         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1243         {
1244             GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1245             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1246         }
1247     }
1248
1249     // TODO Update this comment when introducing SimulationWorkload
1250     //
1251     // The conditions for gpuHaloExchange e.g. using GPU buffer
1252     // operations were checked before construction, so here we can
1253     // just use it and assert upon any conditions.
1254     const bool ddUsesGpuDirectCommunication =
1255             ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange[0].empty()));
1256     GMX_ASSERT(!ddUsesGpuDirectCommunication || stepWork.useGpuXBufferOps,
1257                "Must use coordinate buffer ops with GPU halo exchange");
1258     const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && stepWork.useGpuFBufferOps;
1259
1260     // Copy coordinate from the GPU if update is on the GPU and there
1261     // are forces to be computed on the CPU, or for the computation of
1262     // virial, or if host-side data will be transferred from this task
1263     // to a remote task for halo exchange or PME-PP communication. At
1264     // search steps the current coordinates are already on the host,
1265     // hence copy is not needed.
1266     const bool haveHostPmePpComms =
1267             !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1268     const bool haveHostHaloExchangeComms = havePPDomainDecomposition(cr) && !ddUsesGpuDirectCommunication;
1269
1270     bool gmx_used_in_debug haveCopiedXFromGpu = false;
1271     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1272         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1273             || haveHostPmePpComms || haveHostHaloExchangeComms))
1274     {
1275         GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1276         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1277         haveCopiedXFromGpu = true;
1278     }
1279
1280     // If coordinates are to be sent to PME task from GPU memory, perform that send here.
1281     // Otherwise the send will occur before the H2D coordinate transfer.
1282     if (!thisRankHasDuty(cr, DUTY_PME) && pmeSendCoordinatesFromGpu)
1283     {
1284         /* Send particle coordinates to the pme nodes */
1285         gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1286                                  lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1287                                  step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1288                                  pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1289     }
1290
1291     if (useGpuPmeOnThisRank)
1292     {
1293         launchPmeGpuSpread(fr->pmedata, box, stepWork, localXReadyOnDevice, lambda[efptCOUL], wcycle);
1294     }
1295
1296     const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1297
1298     /* do gridding for pair search */
1299     if (stepWork.doNeighborSearch)
1300     {
1301         if (fr->wholeMoleculeTransform && stepWork.stateChanged)
1302         {
1303             fr->wholeMoleculeTransform->updateForAtomPbcJumps(x.unpaddedArrayRef(), box);
1304         }
1305
1306         // TODO
1307         // - vzero is constant, do we need to pass it?
1308         // - box_diag should be passed directly to nbnxn_put_on_grid
1309         //
1310         rvec vzero;
1311         clear_rvec(vzero);
1312
1313         rvec box_diag;
1314         box_diag[XX] = box[XX][XX];
1315         box_diag[YY] = box[YY][YY];
1316         box_diag[ZZ] = box[ZZ][ZZ];
1317
1318         wallcycle_start(wcycle, ewcNS);
1319         if (!DOMAINDECOMP(cr))
1320         {
1321             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1322             nbnxn_put_on_grid(nbv, box, 0, vzero, box_diag, nullptr, { 0, mdatoms->homenr }, -1,
1323                               fr->cginfo, x.unpaddedArrayRef(), 0, nullptr);
1324             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1325         }
1326         else
1327         {
1328             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1329             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->cginfo, x.unpaddedArrayRef());
1330             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1331         }
1332
1333         nbv->setAtomProperties(gmx::constArrayRefFromArray(mdatoms->typeA, mdatoms->nr),
1334                                gmx::constArrayRefFromArray(mdatoms->chargeA, mdatoms->nr), fr->cginfo);
1335
1336         wallcycle_stop(wcycle, ewcNS);
1337
1338         /* initialize the GPU nbnxm atom data and bonded data structures */
1339         if (simulationWork.useGpuNonbonded)
1340         {
1341             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1342
1343             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1344             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1345             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1346
1347             if (fr->gpuBonded)
1348             {
1349                 /* Now we put all atoms on the grid, we can assign bonded
1350                  * interactions to the GPU, where the grid order is
1351                  * needed. Also the xq, f and fshift device buffers have
1352                  * been reallocated if needed, so the bonded code can
1353                  * learn about them. */
1354                 // TODO the xq, f, and fshift buffers are now shared
1355                 // resources, so they should be maintained by a
1356                 // higher-level object than the nb module.
1357                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(
1358                         nbv->getGridIndices(), top->idef, Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1359                         Nbnxm::gpu_get_f(nbv->gpu_nbv), Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1360             }
1361             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1362         }
1363
1364         // Need to run after the GPU-offload bonded interaction lists
1365         // are set up to be able to determine whether there is bonded work.
1366         runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1367                 *inputrec, *fr, pull_work, ed, *mdatoms, simulationWork, stepWork);
1368
1369         wallcycle_start_nocount(wcycle, ewcNS);
1370         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1371         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1372         nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1373
1374         nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1375
1376         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1377         wallcycle_stop(wcycle, ewcNS);
1378
1379         if (stepWork.useGpuXBufferOps)
1380         {
1381             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1382         }
1383
1384         if (simulationWork.useGpuBufferOps)
1385         {
1386             setupGpuForceReductions(runScheduleWork, cr, fr, ddUsesGpuDirectCommunication);
1387         }
1388     }
1389     else if (!EI_TPI(inputrec->eI) && stepWork.computeNonbondedForces)
1390     {
1391         if (stepWork.useGpuXBufferOps)
1392         {
1393             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1394             nbv->convertCoordinatesGpu(AtomLocality::Local, false, stateGpu->getCoordinates(),
1395                                        localXReadyOnDevice);
1396         }
1397         else
1398         {
1399             if (simulationWork.useGpuUpdate)
1400             {
1401                 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1402                 GMX_ASSERT(haveCopiedXFromGpu,
1403                            "a wait should only be triggered if copy has been scheduled");
1404                 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1405             }
1406             nbv->convertCoordinates(AtomLocality::Local, false, x.unpaddedArrayRef());
1407         }
1408     }
1409
1410     if (simulationWork.useGpuNonbonded && (stepWork.computeNonbondedForces || domainWork.haveGpuBondedWork))
1411     {
1412         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1413
1414         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1415
1416         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1417         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1418         if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1419         {
1420             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1421         }
1422         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1423         // with X buffer ops offloaded to the GPU on all but the search steps
1424
1425         // bonded work not split into separate local and non-local, so with DD
1426         // we can only launch the kernel after non-local coordinates have been received.
1427         if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1428         {
1429             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1430             fr->gpuBonded->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1431             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1432         }
1433
1434         /* launch local nonbonded work on GPU */
1435         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1436         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1437         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1438         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1439     }
1440
1441     if (useGpuPmeOnThisRank)
1442     {
1443         // In PME GPU and mixed mode we launch FFT / gather after the
1444         // X copy/transform to allow overlap as well as after the GPU NB
1445         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1446         // the nonbonded kernel.
1447         launchPmeGpuFftAndGather(fr->pmedata, lambda[efptCOUL], wcycle, stepWork);
1448     }
1449
1450     /* Communicate coordinates and sum dipole if necessary +
1451        do non-local pair search */
1452     if (havePPDomainDecomposition(cr))
1453     {
1454         if (stepWork.doNeighborSearch)
1455         {
1456             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1457             wallcycle_start_nocount(wcycle, ewcNS);
1458             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1459             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1460             nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1461
1462             nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1463             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1464             wallcycle_stop(wcycle, ewcNS);
1465             // TODO refactor this GPU halo exchange re-initialisation
1466             // to location in do_md where GPU halo exchange is
1467             // constructed at partitioning, after above stateGpu
1468             // re-initialization has similarly been refactored
1469             if (ddUsesGpuDirectCommunication)
1470             {
1471                 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1472             }
1473         }
1474         else
1475         {
1476             if (ddUsesGpuDirectCommunication)
1477             {
1478                 // The following must be called after local setCoordinates (which records an event
1479                 // when the coordinate data has been copied to the device).
1480                 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1481
1482                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1483                 {
1484                     // non-local part of coordinate buffer must be copied back to host for CPU work
1485                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1486                 }
1487             }
1488             else
1489             {
1490                 // Note: GPU update + DD without direct communication is not supported,
1491                 // a waitCoordinatesReadyOnHost() should be issued if it will be.
1492                 GMX_ASSERT(!simulationWork.useGpuUpdate,
1493                            "GPU update is not supported with CPU halo exchange");
1494                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1495             }
1496
1497             if (stepWork.useGpuXBufferOps)
1498             {
1499                 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1500                 {
1501                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1502                 }
1503                 nbv->convertCoordinatesGpu(AtomLocality::NonLocal, false, stateGpu->getCoordinates(),
1504                                            stateGpu->getCoordinatesReadyOnDeviceEvent(
1505                                                    AtomLocality::NonLocal, simulationWork, stepWork));
1506             }
1507             else
1508             {
1509                 nbv->convertCoordinates(AtomLocality::NonLocal, false, x.unpaddedArrayRef());
1510             }
1511         }
1512
1513         if (simulationWork.useGpuNonbonded)
1514         {
1515             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1516
1517             if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1518             {
1519                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1520                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1521                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1522             }
1523
1524             if (domainWork.haveGpuBondedWork)
1525             {
1526                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1527                 fr->gpuBonded->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1528                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1529             }
1530
1531             /* launch non-local nonbonded tasks on GPU */
1532             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1533             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1534                          nrnb, wcycle);
1535             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1536
1537             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1538         }
1539     }
1540
1541     if (simulationWork.useGpuNonbonded && stepWork.computeNonbondedForces)
1542     {
1543         /* launch D2H copy-back F */
1544         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1545         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1546
1547         if (havePPDomainDecomposition(cr))
1548         {
1549             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1550         }
1551         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1552         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1553
1554         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1555         {
1556             fr->gpuBonded->launchEnergyTransfer();
1557         }
1558         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1559     }
1560
1561     gmx::ArrayRef<const gmx::RVec> xWholeMolecules;
1562     if (fr->wholeMoleculeTransform)
1563     {
1564         xWholeMolecules = fr->wholeMoleculeTransform->wholeMoleculeCoordinates(x.unpaddedArrayRef(), box);
1565     }
1566
1567     DipoleData dipoleData;
1568
1569     if (simulationWork.computeMuTot)
1570     {
1571         const int start = 0;
1572
1573         /* Calculate total (local) dipole moment in a temporary common array.
1574          * This makes it possible to sum them over nodes faster.
1575          */
1576         gmx::ArrayRef<const gmx::RVec> xRef =
1577                 (xWholeMolecules.empty() ? x.unpaddedArrayRef() : xWholeMolecules);
1578         calc_mu(start, mdatoms->homenr, xRef, mdatoms->chargeA, mdatoms->chargeB,
1579                 mdatoms->nChargePerturbed, dipoleData.muStaging[0], dipoleData.muStaging[1]);
1580
1581         reduceAndUpdateMuTot(&dipoleData, cr, (fr->efep != efepNO), lambda, muTotal, ddBalanceRegionHandler);
1582     }
1583
1584     /* Reset energies */
1585     reset_enerdata(enerd);
1586
1587     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1588     {
1589         wallcycle_start(wcycle, ewcPPDURINGPME);
1590         dd_force_flop_start(cr->dd, nrnb);
1591     }
1592
1593     // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1594     // this wait ensures that the D2H transfer is complete.
1595     if ((simulationWork.useGpuUpdate)
1596         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial))
1597     {
1598         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1599     }
1600
1601     if (inputrec->bRot)
1602     {
1603         wallcycle_start(wcycle, ewcROT);
1604         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step,
1605                     stepWork.doNeighborSearch);
1606         wallcycle_stop(wcycle, ewcROT);
1607     }
1608
1609     /* Start the force cycle counter.
1610      * Note that a different counter is used for dynamic load balancing.
1611      */
1612     wallcycle_start(wcycle, ewcFORCE);
1613
1614     /* Set up and clear force outputs:
1615      * forceOutMtsLevel0:  everything except what is in the other two outputs
1616      * forceOutMtsLevel1:  PME-mesh and listed-forces group 1
1617      * forceOutNonbonded: non-bonded forces
1618      * Without multiple time stepping all point to the same object.
1619      * With multiple time-stepping the use is different for MTS fast (level0 only) and slow steps.
1620      */
1621     ForceOutputs forceOutMtsLevel0 =
1622             setupForceOutputs(&fr->forceHelperBuffers[0], force, stepWork, wcycle);
1623
1624     // Force output for MTS combined forces, only set at level1 MTS steps
1625     std::optional<ForceOutputs> forceOutMts =
1626             (fr->useMts && stepWork.computeSlowForces)
1627                     ? std::optional(setupForceOutputs(&fr->forceHelperBuffers[1],
1628                                                       forceView->forceMtsCombinedWithPadding(),
1629                                                       stepWork, wcycle))
1630                     : std::nullopt;
1631
1632     ForceOutputs* forceOutMtsLevel1 =
1633             fr->useMts ? (stepWork.computeSlowForces ? &forceOutMts.value() : nullptr) : &forceOutMtsLevel0;
1634
1635     const bool nonbondedAtMtsLevel1 = runScheduleWork->simulationWork.computeNonbondedAtMtsLevel1;
1636
1637     ForceOutputs* forceOutNonbonded = nonbondedAtMtsLevel1 ? forceOutMtsLevel1 : &forceOutMtsLevel0;
1638
1639     if (inputrec->bPull && pull_have_constraint(pull_work))
1640     {
1641         clear_pull_forces(pull_work);
1642     }
1643
1644     /* We calculate the non-bonded forces, when done on the CPU, here.
1645      * We do this before calling do_force_lowlevel, because in that
1646      * function, the listed forces are calculated before PME, which
1647      * does communication.  With this order, non-bonded and listed
1648      * force calculation imbalance can be balanced out by the domain
1649      * decomposition load balancing.
1650      */
1651
1652     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1653
1654     if (!useOrEmulateGpuNb)
1655     {
1656         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1657     }
1658
1659     if (fr->efep != efepNO && stepWork.computeNonbondedForces)
1660     {
1661         /* Calculate the local and non-local free energy interactions here.
1662          * Happens here on the CPU both with and without GPU.
1663          */
1664         nbv->dispatchFreeEnergyKernel(InteractionLocality::Local, fr,
1665                                       as_rvec_array(x.unpaddedArrayRef().data()),
1666                                       &forceOutNonbonded->forceWithShiftForces(), *mdatoms,
1667                                       inputrec->fepvals, lambda, enerd, stepWork, nrnb);
1668
1669         if (havePPDomainDecomposition(cr))
1670         {
1671             nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal, fr,
1672                                           as_rvec_array(x.unpaddedArrayRef().data()),
1673                                           &forceOutNonbonded->forceWithShiftForces(), *mdatoms,
1674                                           inputrec->fepvals, lambda, enerd, stepWork, nrnb);
1675         }
1676     }
1677
1678     if (stepWork.computeNonbondedForces && !useOrEmulateGpuNb)
1679     {
1680         if (havePPDomainDecomposition(cr))
1681         {
1682             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1683                          nrnb, wcycle);
1684         }
1685
1686         if (stepWork.computeForces)
1687         {
1688             /* Add all the non-bonded force to the normal force array.
1689              * This can be split into a local and a non-local part when overlapping
1690              * communication with calculation with domain decomposition.
1691              */
1692             wallcycle_stop(wcycle, ewcFORCE);
1693             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All,
1694                                           forceOutNonbonded->forceWithShiftForces().force());
1695             wallcycle_start_nocount(wcycle, ewcFORCE);
1696         }
1697
1698         /* If there are multiple fshift output buffers we need to reduce them */
1699         if (stepWork.computeVirial)
1700         {
1701             /* This is not in a subcounter because it takes a
1702                negligible and constant-sized amount of time */
1703             nbnxn_atomdata_add_nbat_fshift_to_fshift(
1704                     *nbv->nbat, forceOutNonbonded->forceWithShiftForces().shiftForces());
1705         }
1706     }
1707
1708     // TODO Force flags should include haveFreeEnergyWork for this domain
1709     if (ddUsesGpuDirectCommunication && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1710     {
1711         /* Wait for non-local coordinate data to be copied from device */
1712         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1713     }
1714
1715     // Compute wall interactions, when present.
1716     // Note: should be moved to special forces.
1717     if (inputrec->nwall && stepWork.computeNonbondedForces)
1718     {
1719         /* foreign lambda component for walls */
1720         real dvdl_walls = do_walls(*inputrec, *fr, box, *mdatoms, x.unpaddedConstArrayRef(),
1721                                    &forceOutMtsLevel0.forceWithVirial(), lambda[efptVDW],
1722                                    enerd->grpp.ener[egLJSR].data(), nrnb);
1723         enerd->dvdl_lin[efptVDW] += dvdl_walls;
1724     }
1725
1726     if (stepWork.computeListedForces)
1727     {
1728         /* Check whether we need to take into account PBC in listed interactions */
1729         bool needMolPbc = false;
1730         for (const auto& listedForces : fr->listedForces)
1731         {
1732             if (listedForces.haveCpuListedForces(*fr->fcdata))
1733             {
1734                 needMolPbc = fr->bMolPBC;
1735             }
1736         }
1737
1738         t_pbc pbc;
1739
1740         if (needMolPbc)
1741         {
1742             /* Since all atoms are in the rectangular or triclinic unit-cell,
1743              * only single box vector shifts (2 in x) are required.
1744              */
1745             set_pbc_dd(&pbc, fr->pbcType, DOMAINDECOMP(cr) ? cr->dd->numCells : nullptr, TRUE, box);
1746         }
1747
1748         for (int mtsIndex = 0; mtsIndex < (fr->useMts && stepWork.computeSlowForces ? 2 : 1); mtsIndex++)
1749         {
1750             ListedForces& listedForces = fr->listedForces[mtsIndex];
1751             ForceOutputs& forceOut     = (mtsIndex == 0 ? forceOutMtsLevel0 : *forceOutMtsLevel1);
1752             listedForces.calculate(
1753                     wcycle, box, inputrec->fepvals, cr, ms, x, xWholeMolecules, fr->fcdata.get(),
1754                     hist, &forceOut, fr, &pbc, enerd, nrnb, lambda.data(), mdatoms,
1755                     DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr, stepWork);
1756         }
1757     }
1758
1759     if (stepWork.computeSlowForces)
1760     {
1761         calculateLongRangeNonbondeds(fr, inputrec, cr, nrnb, wcycle, mdatoms,
1762                                      x.unpaddedConstArrayRef(), &forceOutMtsLevel1->forceWithVirial(),
1763                                      enerd, box, lambda.data(), as_rvec_array(dipoleData.muStateAB),
1764                                      stepWork, ddBalanceRegionHandler);
1765     }
1766
1767     wallcycle_stop(wcycle, ewcFORCE);
1768
1769     // VdW dispersion correction, only computed on master rank to avoid double counting
1770     if ((stepWork.computeEnergy || stepWork.computeVirial) && fr->dispersionCorrection && MASTER(cr))
1771     {
1772         // Calculate long range corrections to pressure and energy
1773         const DispersionCorrection::Correction correction =
1774                 fr->dispersionCorrection->calculate(box, lambda[efptVDW]);
1775
1776         if (stepWork.computeEnergy)
1777         {
1778             enerd->term[F_DISPCORR] = correction.energy;
1779             enerd->term[F_DVDL_VDW] += correction.dvdl;
1780             enerd->dvdl_lin[efptVDW] += correction.dvdl;
1781         }
1782         if (stepWork.computeVirial)
1783         {
1784             correction.correctVirial(vir_force);
1785             enerd->term[F_PDISPCORR] = correction.pressure;
1786         }
1787     }
1788
1789     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation, imdSession, pull_work, step, t,
1790                          wcycle, fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda, stepWork,
1791                          &forceOutMtsLevel0.forceWithVirial(), enerd, ed, stepWork.doNeighborSearch);
1792
1793     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
1794                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
1795     GMX_ASSERT(!(nonbondedAtMtsLevel1 && useGpuForcesHaloExchange),
1796                "The schedule below does not allow for nonbonded MTS with GPU halo exchange");
1797     // Will store the amount of cycles spent waiting for the GPU that
1798     // will be later used in the DLB accounting.
1799     float cycles_wait_gpu = 0;
1800     if (useOrEmulateGpuNb && stepWork.computeNonbondedForces)
1801     {
1802         auto& forceWithShiftForces = forceOutNonbonded->forceWithShiftForces();
1803
1804         /* wait for non-local forces (or calculate in emulation mode) */
1805         if (havePPDomainDecomposition(cr))
1806         {
1807             if (simulationWork.useGpuNonbonded)
1808             {
1809                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
1810                         nbv->gpu_nbv, stepWork, AtomLocality::NonLocal, enerd->grpp.ener[egLJSR].data(),
1811                         enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(), wcycle);
1812             }
1813             else
1814             {
1815                 wallcycle_start_nocount(wcycle, ewcFORCE);
1816                 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes,
1817                              step, nrnb, wcycle);
1818                 wallcycle_stop(wcycle, ewcFORCE);
1819             }
1820
1821             if (stepWork.useGpuFBufferOps)
1822             {
1823                 // TODO: move this into DomainLifetimeWorkload, including the second part of the
1824                 // condition The bonded and free energy CPU tasks can have non-local force
1825                 // contributions which are a dependency for the GPU force reduction.
1826                 bool haveNonLocalForceContribInCpuBuffer =
1827                         domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1828
1829                 if (haveNonLocalForceContribInCpuBuffer)
1830                 {
1831                     stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
1832                                               AtomLocality::NonLocal);
1833                 }
1834
1835                 fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->execute();
1836
1837                 if (!useGpuForcesHaloExchange)
1838                 {
1839                     // copy from GPU input for dd_move_f()
1840                     stateGpu->copyForcesFromGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
1841                                                 AtomLocality::NonLocal);
1842                 }
1843             }
1844             else
1845             {
1846                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
1847             }
1848
1849             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1850             {
1851                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
1852             }
1853         }
1854     }
1855
1856     /* Combining the forces for multiple time stepping before the halo exchange, when possible,
1857      * avoids an extra halo exchange (when DD is used) and post-processing step.
1858      */
1859     const bool combineMtsForcesBeforeHaloExchange =
1860             (stepWork.computeForces && fr->useMts && stepWork.computeSlowForces
1861              && (legacyFlags & GMX_FORCE_DO_NOT_NEED_NORMAL_FORCE) != 0
1862              && !(stepWork.computeVirial || simulationWork.useGpuNonbonded || useGpuPmeOnThisRank));
1863     if (combineMtsForcesBeforeHaloExchange)
1864     {
1865         const int numAtoms = havePPDomainDecomposition(cr) ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr;
1866         combineMtsForces(numAtoms, force.unpaddedArrayRef(), forceView->forceMtsCombined(),
1867                          inputrec->mtsLevels[1].stepFactor);
1868     }
1869
1870     if (havePPDomainDecomposition(cr))
1871     {
1872         /* We are done with the CPU compute.
1873          * We will now communicate the non-local forces.
1874          * If we use a GPU this will overlap with GPU work, so in that case
1875          * we do not close the DD force balancing region here.
1876          */
1877         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1878
1879         if (stepWork.computeForces)
1880         {
1881             if (useGpuForcesHaloExchange)
1882             {
1883                 if (domainWork.haveCpuLocalForceWork)
1884                 {
1885                     stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
1886                                               AtomLocality::Local);
1887                 }
1888                 communicateGpuHaloForces(*cr, domainWork.haveCpuLocalForceWork);
1889             }
1890             else
1891             {
1892                 if (stepWork.useGpuFBufferOps)
1893                 {
1894                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
1895                 }
1896
1897                 // Without MTS or with MTS at slow steps with uncombined forces we need to
1898                 // communicate the fast forces
1899                 if (!fr->useMts || !combineMtsForcesBeforeHaloExchange)
1900                 {
1901                     dd_move_f(cr->dd, &forceOutMtsLevel0.forceWithShiftForces(), wcycle);
1902                 }
1903                 // With MTS we need to communicate the slow or combined (in forceOutMtsLevel1) forces
1904                 if (fr->useMts && stepWork.computeSlowForces)
1905                 {
1906                     dd_move_f(cr->dd, &forceOutMtsLevel1->forceWithShiftForces(), wcycle);
1907                 }
1908             }
1909         }
1910     }
1911
1912     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1913     // an alternating wait/reduction scheme.
1914     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded
1915                              && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
1916     if (alternateGpuWait)
1917     {
1918         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, forceOutNonbonded,
1919                                     forceOutMtsLevel1, enerd, lambda[efptCOUL], stepWork, wcycle);
1920     }
1921
1922     if (!alternateGpuWait && useGpuPmeOnThisRank)
1923     {
1924         pme_gpu_wait_and_reduce(fr->pmedata, stepWork, wcycle,
1925                                 &forceOutMtsLevel1->forceWithVirial(), enerd, lambda[efptCOUL]);
1926     }
1927
1928     /* Wait for local GPU NB outputs on the non-alternating wait path */
1929     if (!alternateGpuWait && stepWork.computeNonbondedForces && simulationWork.useGpuNonbonded)
1930     {
1931         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1932          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1933          * but even with a step of 0.1 ms the difference is less than 1%
1934          * of the step time.
1935          */
1936         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1937         const float waitCycles               = Nbnxm::gpu_wait_finish_task(
1938                 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
1939                 enerd->grpp.ener[egCOULSR].data(),
1940                 forceOutNonbonded->forceWithShiftForces().shiftForces(), wcycle);
1941
1942         if (ddBalanceRegionHandler.useBalancingRegion())
1943         {
1944             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1945             if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
1946             {
1947                 /* We measured few cycles, it could be that the kernel
1948                  * and transfer finished earlier and there was no actual
1949                  * wait time, only API call overhead.
1950                  * Then the actual time could be anywhere between 0 and
1951                  * cycles_wait_est. We will use half of cycles_wait_est.
1952                  */
1953                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1954             }
1955             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1956         }
1957     }
1958
1959     if (fr->nbv->emulateGpu())
1960     {
1961         // NOTE: emulation kernel is not included in the balancing region,
1962         // but emulation mode does not target performance anyway
1963         wallcycle_start_nocount(wcycle, ewcFORCE);
1964         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local,
1965                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes, step, nrnb, wcycle);
1966         wallcycle_stop(wcycle, ewcFORCE);
1967     }
1968
1969     // If on GPU PME-PP comms or GPU update path, receive forces from PME before GPU buffer ops
1970     // TODO refactor this and unify with below default-path call to the same function
1971     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && stepWork.computeSlowForces
1972         && (simulationWork.useGpuPmePpCommunication || simulationWork.useGpuUpdate))
1973     {
1974         /* In case of node-splitting, the PP nodes receive the long-range
1975          * forces, virial and energy from the PME nodes here.
1976          */
1977         pme_receive_force_ener(fr, cr, &forceOutMtsLevel1->forceWithVirial(), enerd,
1978                                simulationWork.useGpuPmePpCommunication,
1979                                stepWork.useGpuPmeFReduction, wcycle);
1980     }
1981
1982
1983     /* Do the nonbonded GPU (or emulation) force buffer reduction
1984      * on the non-alternating path. */
1985     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
1986                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
1987     if (useOrEmulateGpuNb && !alternateGpuWait)
1988     {
1989         if (stepWork.useGpuFBufferOps)
1990         {
1991             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
1992
1993             // Flag to specify whether the CPU force buffer has contributions to
1994             // local atoms. This depends on whether there are CPU-based force tasks
1995             // or when DD is active the halo exchange has resulted in contributions
1996             // from the non-local part.
1997             const bool haveLocalForceContribInCpuBuffer =
1998                     (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
1999
2000             // TODO: move these steps as early as possible:
2001             // - CPU f H2D should be as soon as all CPU-side forces are done
2002             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
2003             //   before the next CPU task that consumes the forces: vsite spread or update)
2004             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
2005             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
2006             //   These should be unified.
2007             if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
2008             {
2009                 // Note: AtomLocality::All is used for the non-DD case because, as in this
2010                 // case copyForcesToGpu() uses a separate stream, it allows overlap of
2011                 // CPU force H2D with GPU force tasks on all streams including those in the
2012                 // local stream which would otherwise be implicit dependencies for the
2013                 // transfer and would not overlap.
2014                 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
2015
2016                 stateGpu->copyForcesToGpu(forceWithShift, locality);
2017             }
2018
2019             if (stepWork.computeNonbondedForces)
2020             {
2021                 fr->gpuForceReduction[gmx::AtomLocality::Local]->execute();
2022             }
2023
2024             // Copy forces to host if they are needed for update or if virtual sites are enabled.
2025             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
2026             // TODO: When the output flags will be included in step workload, this copy can be combined with the
2027             //       copy call done in sim_utils(...) for the output.
2028             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
2029             //       they should not be copied in do_md(...) for the output.
2030             if (!simulationWork.useGpuUpdate || vsite)
2031             {
2032                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
2033                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
2034             }
2035         }
2036         else if (stepWork.computeNonbondedForces)
2037         {
2038             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2039             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
2040         }
2041     }
2042
2043     launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd, *runScheduleWork,
2044                             useGpuPmeOnThisRank, step, wcycle);
2045
2046     if (DOMAINDECOMP(cr))
2047     {
2048         dd_force_flop_stop(cr->dd, nrnb);
2049     }
2050
2051     const bool haveCombinedMtsForces = (stepWork.computeForces && fr->useMts && stepWork.computeSlowForces
2052                                         && combineMtsForcesBeforeHaloExchange);
2053     if (stepWork.computeForces)
2054     {
2055         postProcessForceWithShiftForces(nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutMtsLevel0,
2056                                         vir_force, *mdatoms, *fr, vsite, stepWork);
2057
2058         if (fr->useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2059         {
2060             postProcessForceWithShiftForces(nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1,
2061                                             vir_force, *mdatoms, *fr, vsite, stepWork);
2062         }
2063     }
2064
2065     // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
2066     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
2067         && !simulationWork.useGpuUpdate && stepWork.computeSlowForces)
2068     {
2069         /* In case of node-splitting, the PP nodes receive the long-range
2070          * forces, virial and energy from the PME nodes here.
2071          */
2072         pme_receive_force_ener(fr, cr, &forceOutMtsLevel1->forceWithVirial(), enerd,
2073                                simulationWork.useGpuPmePpCommunication, false, wcycle);
2074     }
2075
2076     if (stepWork.computeForces)
2077     {
2078         /* If we don't use MTS or if we already combined the MTS forces before, we only
2079          * need to post-process one ForceOutputs object here, called forceOutCombined,
2080          * otherwise we have to post-process two outputs and then combine them.
2081          */
2082         ForceOutputs& forceOutCombined = (haveCombinedMtsForces ? forceOutMts.value() : forceOutMtsLevel0);
2083         postProcessForces(cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutCombined,
2084                           vir_force, mdatoms, fr, vsite, stepWork);
2085
2086         if (fr->useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2087         {
2088             postProcessForces(cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1,
2089                               vir_force, mdatoms, fr, vsite, stepWork);
2090
2091             combineMtsForces(mdatoms->homenr, force.unpaddedArrayRef(),
2092                              forceView->forceMtsCombined(), inputrec->mtsLevels[1].stepFactor);
2093         }
2094     }
2095
2096     if (stepWork.computeEnergy)
2097     {
2098         /* Compute the final potential energy terms */
2099         accumulatePotentialEnergies(enerd, lambda, inputrec->fepvals);
2100
2101         if (!EI_TPI(inputrec->eI))
2102         {
2103             checkPotentialEnergyValidity(step, *enerd, *inputrec);
2104         }
2105     }
2106
2107     /* In case we don't have constraints and are using GPUs, the next balancing
2108      * region starts here.
2109      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
2110      * virial calculation and COM pulling, is not thus not included in
2111      * the balance timing, which is ok as most tasks do communication.
2112      */
2113     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
2114 }